Datasets:

timaeus

/

pile-arxiv

like 0

Follow

Timaeus 10

--- abstract: 'In this paper we formulate the pure spinor superstring theory on $AdS_4\times \mathbb{CP}^3$. By recasting the pure spinor action as a topological A-model on the fermionic supercoset $Osp(6|4)/SO(6)\times Sp(4)$ plus a BRST exact term, we prove the exactness of the $\sigma$-model. We then give a gauged linear $\sigma$-model which reduces to the superstring in the limit of large volume and we study its branch geometry in different phases. Moreover, we discuss possible D-brane boundary conditions and the principal chiral model for the fermionic supercoset.' --- **** **** Giulio Bonelli$^{~a}$, Pietro Antonio Grassi$^{~b}$, and Houman Safaai$^{~a}$ .7cm $a)$ International School of Advanced Studies (SISSA) via Beirut 2-4, 34014 Trieste, Italy INFN - Sezione di Trieste .2cm $b)$ DISTA, Università del Piemonte Orientale, via Bellini 25/G, Alessandria, 15100, Italy INFN - Sezione di Torino, Gruppo Collegato di Alessandria Introduction ============ The pure spinor formulation of superstrings [@purespinors] is a powerfull method to tame superconformal exactness, the string loop expansion and RR-background $\sigma$-model couplings at the same time. In particular, let us remark that, although conjectured because of maximal supersymmetry, the quantum exactness of type IIB string on $AdS_5\times S^5$ was proved in [@quantumex] by making use of the pure spinor formulation. Recently, the M-theory analog $AdS_4\times S^7$ is receiving large attentions [@AharonyUG; @disastro] because of its conjectured duality with the Bagger-Lambert-Gustavsson [@BLG] theory of multiple M2-branes. As a superstring theory, because of the circle fibration $$\begin{matrix} S^1 & \hookrightarrow & S^7 \\ {} & {} & \downarrow \\ {} & {} & {\mathbb P}^3 \end{matrix}$$ this is represented as type IIA superstring on $AdS_4\times \mathbb{CP}^3$ with appropriate RR-fluxes turned on. This superstring background is undergoing an intense study [@frolov; @stefanski; @antonio; @AdS-4strings], but its exact superconformal invariance has not been established so far. This is one motivation[^1] to study the pure spinor formulation of superstrings on $AdS_4\times \mathbb{CP}^3$ and in this paper we actually prove its superconformal exactness. This is done by writing the pure spinor action as a manifestly superconformal term plus a BRST trivial term. Under the assumption that only mild non-locality arises in the BRST trivial term, we establish superconformal exactness. Let us point out some further considerations about the system for $AdS_4 \times \mathbb{CP}^3$. In the present background there is a RR flux balancing the spacetime curvature which can be tuned to reach opposite limits: the strong coupling limit where the RR fields become dominant and the weak coupling limit where the supergravity approximation is valid. In the case of $AdS_5 \times S^5$, the two limits where covered by the same theory, namely $N=4, d=4$ SYM and the two limits of string theory where represented by the opposite limits of the gauge theory side. In that perspective, the supergravity computations in $AdS_5 \times S^5$ background leads to strong coupling correlation functions, whereas the strong coupling limit in string theory (where the supergravity approximation is no longer a good one) corresponds to perturbative SYM at weak coupling. Recently, in [@Berkovits1] and in [@BV], it has been conjectured that this limit can be achieved by constructing the pure spinor sigma model on the coset $PSU(2,2|4)/ SO(1,4) \times SO(5)$, by taking the limit where the coupling constant goes to infinity and, finally, by adding a BRST trivial term, one can recast the $\sigma$-model into a non-linear sigma model which is a topological A-model that can be proved to be conformal to all orders (given the fact that the sigma model is based on a symmetric coset and the supergroup $PSU(2,2|4)$ is a super-Calabi-Yau). This construction has been tested in [@BS] where, by using a mirror symmetry argument, this program was realized for circular planar 1/2-BPS Wilson loops. See also [@brenno] for further developments. Actually, in [@BerkovitsQC], the relation between perturbative SYM $N=4$ and the topological sigma model has been further developed. There, it has been noticed that given a suitable measure for integrating the pure spinors in tree level amplitudes and identifying the vertex operators of the topological sigma model with the states of the fundamental representation of $PSU(2,2|4)$ (known as singleton), one can define some correlation functions with the properties of tree level SYM amplitudes. However, several checks and computations should be performed to test this new idea. Nevertheless, it has been shown that the topological $\sigma$-model can be viewed as a gauged linear $\sigma$-model of the G/G type with a suitable gauge fixing. On another side, we would like to perform the same analysis with another gauged linear sigma model. We consider the gauged sigma model with the supergroup $Osp(6|4)$. It has been already shown that this model leads to a pure spinor string theory model and its action has been constructed. In the present work we construct the gauged sigma model by gauge fixing the gauge symmetries of $Osp(6|4)$. Again, there are two limits. One limit is the weak coupling limit where the supergravity approximation is valid (see for example [@AharonyUG]) and this leads to multiple M2 brane interpretation of its dual theory (based on Bagger-Lambert-Gustavsson theory). This model is a superconformal Chern-Simons theory in $d=3$ with $N=6$. Notice that in this model the kinetic term of the gauge field is neglected since it has a dimensionful coupling, corresponding to the inverse of radius of $AdS$ which, in the limit of small RR flux $e$, tends to zero. The only remaining gauge field dynamics is described by a Chern-Simons model. However, being supersymmetric, it has some auxiliary fields which have to be integrated leading to the potential for the matter fields. So, the limit of strong coupling is the limit explored in the weak coupling limit of string theory. On the other side, by considering the opposite limit, namely for large RR flux $e$, one should see the opposite limit of perturbative SYM d=3 model (with the kinetic term). This model has been constructed in [@FabbriAY]. So, we expect that in this limit the theory is no longer superconformal and contains some dynamical gauge field. The perturbative computation, which can be performed using the singleton conjecture of [@BerkovitsQC] should not correspond to the strong coupling limit of N=6 d=3 model, but to the weak coupling limit of $N=6, d=3, SYM$ (the symmetry of this model could be maybe enhanced to $N=8$). The plan of the paper is as follows. In the next section we implement and solve the pure spinor constraints on $AdS_4\times \mathbb{CP}^3$ and therefore we formulate the pure spinor superstring theory on this background. In section 3 we study the A-model topological string on the fermionic supercoset $Osp(6|4)/SO(6)\times Sp(4)$ and we establish its superconformal invariance. In section 4 we show that the pure spinor action for the superstring on $AdS_4\times \mathbb{CP}^3$ can be recasted up to an additive BRST exact term as the topological A-model on the fermionic supercoset $Osp(6|4)/SO(6)\times Sp(4)$. Actually, this is obtained as a particular case of a more general construction for supercosets admitting a compatible $\mathbb Z_4$ grading. In section 5 we then give a gauged linear $\sigma$-model which reduces to the topological string in the limit of large Fayet-Illiopoulos coupling and we study its Coulomb branch geometry. In section 6 we formulate a principal chiral model of $G/G$ type which upon gauge fixing reduces to the A-model on the supercoset. Moreover, in Section 7 we discuss possible D-brane boundary conditions and discuss their geometric structure. We indicate further directions to explore in Section 8 where we collect some open issues too. Pure spinor superstring in $AdS_4\times \mathbb{CP}^3$ background ================================================================= As it was shown in [@antonio; @frolov; @stefanski], the $AdS_4\times \mathbb{CP}^3$ background can be derived from a supercoset element $g\in\frac{Osp(6|4)}{U(3)\times SO(1,3)}$. Its Maurer-Cartan left invariant 1-form can be expanded into the generators of $Osp(6|4)$ as follows $$J=J^a \gamma_a+J_{IJ} T^{IJ}+J^{IJ} T_{IJ}+H^{ab}\gamma_{ab}+H_I^{\;J} T_J^{\;I}+J_I^{\;\alpha} Q^{\;I}_\alpha+J_I^{\;\dot\alpha} Q^{\;I}_{\dot\alpha}+J^{I\alpha} Q_{I\alpha}+J^{I\dot\alpha} Q_{I\dot\alpha}\,,$$ where $(T_{IJ},T^{IJ},T_J^{\;I})$ are the generators of $SO(6)$, $T_{[AB]}$ with $A,B=1\ldots 6$ decomposes according to irreducible reperesentations of $U(3)$ as it will be explained later, and $T_J^{\;I}$ are the generators of $U(3)$. Then, $J_{IJ}$ and $J^{IJ}$ are the Maurer-Cartan forms associated to the generators of the coset $\frac{SU(4)}{U(3)}$ and $H_I^{\;J}$ are the corresponding spin connections of the coset. Similarly, $(\gamma_a,\gamma_{ab})$ with $a,b=1\ldots 4$ are the generators of the anti de Sitter group $SO(2,3)$ which as is shown in [@antonio] they all turn out to be given by real symplectic matrices and $\gamma_{ab}$ are the generators of the Lorentz group $SO(1,3)$. The matrices $Q_I^{\;\alpha},Q_I^{\;\dot\alpha},Q^{\;I}_\alpha$ and $Q^{\;I}_{\dot\alpha}$ are the 24 fermionic generators where we split the symplectic indices $x=1\ldots 4$ into $SO(1,3)$ spinorial indices $\alpha,\dot\alpha=1,2$. The Maurer-Cartan 1-forms of the symplectic group $Sp(4,\mathbb{R})$ are related to the Maurer-Cartan of $SO(2,3)$ with the relation $J^{xy}=J^a \gamma_a^{xy}+H^{ab} \gamma_{ab}^{xy}$. The fermionic 1-forms $J_A^x$ are real and transform in the fundamental 4-dimensional representation of $\mathfrak{sp}(4,\mathbb{R})$ and in the fundamental 6-dimensional representation of $\mathfrak{so}(6)$ with the symplectic invariant antisymmetric metric $\epsilon_{xy}=i\sigma_1\otimes \ma$. Notice that $\eta^{ab}$ is the invariant metric on $AdS_4$ and $g_{I\bar J}$ is the $U(3)$ invariant metric on $\mathbb{P}^3$ and we denote by $k_{I\bar J}$ as the Kähler form on $\mathbb{P}^3$. The index $I$ can be raised and lowered with the inverse metric $g^{\bar I J}$ as $J^{\bar I \bar J}=g^{\bar I K}g^{\bar J L} J_{KL}$ which is independent of $J^{IJ}$, similarly we can make $J_{\bar I \bar J}$ out of $J_{IJ}$. The $\mathfrak{osp}(6|4)$ algebra $\mathcal{H}$ admits a $\mathbb{Z}_4$ grading with decomposition $\mathcal{H}=\sum_{i=0}^3 \mathcal{H}_i$ as follows[^2] $$\begin{aligned} \label{Z4} {\mathcal H}_0 &=& \Big\{ H_{\alpha\beta},H_{\dot\alpha\dot\beta},H_{I}^{\;J} \Big\}\,, \hspace{0.93cm} {\mathcal H}_1 = \Big\{ J^{\alpha I},J^{\dot\alpha \bar I} \Big\}\,, \nonumber \\ {\mathcal H}_2 &=& \Big\{ J_{\alpha\dot\alpha}, J_{IJ}, J^{IJ} \Big\}\,, ~~~~~~~\hspace{1.05cm} {\mathcal H}_3 = \Big\{ J^{\,\,\alpha}_{I},J^{\,\,\dot\alpha}_{\bar I } \Big\}\,.\end{aligned}$$ satisfying $$\left[ {\mathcal H}_m,{\mathcal H}_n\right]\subset {\mathcal H}_{m+n\;(\text{mod }4)}$$ We can check that the bilinear metric is also ${\mathbb Z}_4$ invariant. Recall that the invariant supermetric for $Osp(6|4)$ is given by $$\begin{aligned} \label{bil1} {\rm Str}( T_{AB} T_{CD} ) &=& \delta_{A C} \delta_{D B} - \delta_{A D} \delta_{C B} \,, \\ \nonumber {\rm Str}(T_{xy}\, T_{zt}) &=& \epsilon_{x z} \epsilon_{t y} + \epsilon_{x t} \epsilon_{z y} \,, \\ \nonumber {\rm Str}(T_{x}\, T_{y}) &=& \epsilon_{x y} \,, \\ \nonumber {\rm Str}( Q^x_A Q^y_B ) &=& \delta_{AB} \epsilon^{xy}\,.\end{aligned}$$ where $T_{AB}$ and $T_{xy}$ are the generators of the bosonic subgroups $SO(6)$ and $Sp(4,\mathbb{R})$, and $Q^x_A$ are the fermionic generators of the supergroup. It is convenient to adopt a complex basis for the generators of $SO(6)$ and we define $T_{AB} = U^{IJ}_{AB} T_{IJ} + U^{I}_{J, AB} T_{I}^{~J}+ U_{IJ, AB} T^{IJ}$ where $U^{IJ}_{AB}, U^{I}_{J, AB}, U_{IJ, AB}$ are the Clebsh-Gordon matrices mapping from ${\underline{15}}$ of $SO(6)$ to the representations ${\underline 3}(-1)$, ${\underline 8}(0)$, ${\underline 3}^*(+1)$ of $U(3)$, respectively. In the same way, we decompose the fermionic generators into $Q^x_I$ and $Q^{x I}$ of ${\underline 3}(-1)$ and ${\underline 3}^*(1)$, respectively. The metric becomes $$\begin{aligned} \label{bil2} {\rm Str}( T_{IJ} T^{KL} ) &=& \delta_{I}^{~K} \delta_{J}^{~L} - \delta_{J}^{~K} \delta_{I}^{~L}\,, \\ \nonumber {\rm Str}( T_{I}^{~J} T_K^{~L} ) &=& \delta_{I}^{~L} \delta^{~J}_{K}\,,~~~~~ \\ \nonumber {\rm Str}( Q^x_I Q^{y J} ) &=& \delta_{I}^J \epsilon^{xy}\,.\end{aligned}$$ while the other traces vanish. Which all these mean that the bilinear metric is ${\mathbb Z}_4$ invariant, satisfying $$<{\mathcal H}_m,{\mathcal H}_n>=\text{Str}({\mathcal H}_m{\mathcal H}_n)=0,\;\;\text{unless }m+n=0\;\text{mod }4$$ Using this ${\mathbb Z}_4$ automorphism, it was shown that the pure spinor sigma model action can be decomposed in the following way $$S=S_{GS}+S_{GF}+S_{\text{ghost}} \,,$$ where $S_{GS}$ is the Green-Schwarz action was shown in [@antonio; @frolov; @stefanski] to exhibit the usual quadratic form after using the important feature of the possibility of writing the Wess-Zumino term as a total derivative in this background $$S_{GS}=R^2\displaystyle\int d^2 z \text{Str}\left[ \frac{1}{2}J_{2}{\bar J}_{2}+\frac{1}{4}\left( J _{1}{\bar J} _{3}-{ J} _{3}{\bar J} _{1}\right) \right]\,,$$ where $J_{i}=J|_{\mathcal{H}_i}$ are the projections of the MC left invariant currents into different subclasses according to $\mathbb{Z}_4$ automorphism as it was given in (\[Z4\]). The action can be written in terms of the left-invariant supercurrents of the coset in the following form $$S_{GS}=R^2\displaystyle\int d^2z \left[ \epsilon_{xy} J^x \bar{J}^y+\frac{1}{2} J_{IJ} \bar{J}^{IJ} + \frac{1}{4}\left( J_{\alpha I}{\bar J}^{\alpha I}+J_{\dot\alpha\bar I}{\bar J}^{\dot\alpha\bar I}-J_{\alpha \bar I}{\bar J}^{\alpha \bar I}-J_{\dot\alpha I}{\bar J}^{\dot\alpha I} \right)\right]\,.$$ To this, one has to add a term which breaks $\kappa-$symmetry and adds kinetic terms for the target-space fermions and the coupling to the RR flux. This gauge fixing action $S_{GF}$ was shown to be given by [@antonio] $$S_{GF}= R^2\displaystyle\int d^2 z \left( J_{\alpha \bar I}{\bar J}^{\alpha \bar I}+J_{\dot\alpha I}{\bar J}^{\dot\alpha I} \right)\,,$$ which gives $$S_{GS}+S_{GF}=R^2\displaystyle\int d^2z \left[ \epsilon_{xy} J^x \bar{J}^y+\frac{1}{2} J_{IJ} \bar{J}^{IJ} + \frac{1}{4}\left( J_{\alpha I}{\bar J}^{\alpha I}+J_{\dot\alpha\bar I}{\bar J}^{\dot\alpha\bar I}\right) +\frac{3}{4} \left( J_{\alpha \bar I}{\bar J}^{\alpha \bar I}+J_{\dot\alpha I}{\bar J}^{\dot\alpha I} \right)\right]\,.$$ In order to write the pure spinor ghost part of the action, we introduce the pure spinors $(\lambda_I^{\,\,\alpha},\lambda_{\bar I}^{\,\,\dot\alpha})$, $({\hat\lambda}^{\alpha I},{\hat\lambda}^{\dot\alpha\bar I})$ and their conjugate momenta $(w_{\alpha}^{\,\, I},w^{\,\,\bar I}_{\dot\alpha})$, $({\hat w}_{\alpha I},{\hat w}_{\dot\alpha\bar I})$, belonging to the ${\mathcal{H}}_1$ and ${\mathcal{H}}_3$ respectively. The pure spinor constraints can be written as follows $$\begin{aligned} \label{pures} {\left\lbrace \begin{array}{lllllllllllllllllllllllllllllllll} \lambda_I^\alpha \lambda^{\dot\alpha I}=0\\\lambda_I^\alpha \epsilon_{\alpha\beta}\lambda_J^{\beta}=0\\\lambda^{\dot\alpha I}\epsilon_{\dot\alpha\dot\beta}\lambda^{\dot\beta J}=0\\ \end{array} \right.} ,\hspace{2cm} {\left\lbrace \begin{array}{lllllllllllllllllllllllllllllllll} {\hat\lambda}^{\alpha I} {\hat\lambda}^{\dot\alpha}_I=0\\{\hat\lambda}^{I\alpha} \epsilon_{\alpha\beta}{\hat\lambda}^{\beta J}=0\\{\hat\lambda}^{\dot\alpha}_I \epsilon_{\dot\alpha\dot\beta}{\hat\lambda}^{\dot\beta}_J=0\\ \end{array} \right.}\end{aligned}$$ to solve this constraint, we can use the following ansatz $$\begin{aligned} \lambda^{\,\,\alpha}_I&=&\lambda^\alpha u_I,\;\;\;\;\;\;\;\lambda^{\dot\alpha I}=\lambda^{\dot\alpha}v^I\,, \\ \nonumber {\hat\lambda}^{\alpha I}&=&{\hat\lambda}^\alpha {\hat u}^I,\;\;\;\;\;\;\;{\hat\lambda}^{\,\,\dot\alpha}_I={\hat\lambda}^{\dot\alpha}{\hat v}_I\,,\end{aligned}$$ subject to the following gauge transformations $$\begin{aligned} \lambda^\alpha\rightarrow \frac{1}{\rho}\lambda^\alpha,\;\;\;\;\; \lambda^{\dot\alpha}\rightarrow \frac{1}{\sigma} \lambda^{\dot\alpha},\;\;\;\;\; u_I\rightarrow \rho u_I,\;\;\;\;\; v^I\rightarrow \sigma v^I\,, \\ \nonumber {\hat\lambda}^\alpha\rightarrow \frac{1}{\hat\rho}{\hat\lambda}^\alpha,\;\;\;\;\; {\hat\lambda}^{\dot\alpha}\rightarrow \frac{1}{\hat\sigma} {\hat\lambda}^{\dot\alpha},\;\;\;\;\; {\hat u}^I\rightarrow \hat\rho {\hat u}^I,\;\;\;\;\; {\hat v}_I\rightarrow \hat\sigma {\hat v}_I\,,\end{aligned}$$ where $\rho,\sigma,\hat\rho,\hat\sigma\in\mathbb{C}^*$. Inserting these factorization into (\[pures\]), we arrive to the following constraints $$\begin{aligned} u_Iv^I=0,\;\;\;\;\;\;\; {\hat v}_I{\hat u}^I=0\,.\end{aligned}$$ So, the counting of the degrees of freedom gives $2\times(2+3-1)-1=7$ complex for $\lambda$ and the same for $\hat\lambda$. The geometry of the pure spinor space can be easily described. Using the gauge symmetries $\rho$ and $\sigma$ we can fix the norm of $u_I$ and $v^I$ as such $u_I \bar u^I = 1$ and $v^I \bar v_I = 1$. Then, together the constraint $u_I v^I =0$, the matrix $(u_I, \bar v_I, \epsilon_{IJK} \bar u^J v^K)$ is an $SU(3)$ matrix. In addition, using the remaining phases of the gauge symmetries $\rho$ and $\sigma$, we see that the variables $u_I$ and $v^I$ parametrize the space $SU(3)/ U(1) \times U(1)$ which is the space of the harmonic variables of the $N=3$ harmonic superspace (it is also known as the flag manifold $F(1,2,3)$.[^3] The pure spinor constraints are first class constraints and they commute with the Hamiltonian, therefore they generat the gauge symmetries on the antighost fields $w$’s. In particular if we denote by $\eta_{\a\dot\a}, \eta^{IJ}, \eta_{IJ}$ and by $\kappa_{\a\dot\a}, \kappa^{IJ}, \kappa_{IJ}$ the infinitesimal parameters of the gauge symmetries we have that $$\begin{aligned} \label{gautra} &&\delta w_\a^I = \eta_{\a\dot\a} \lambda^{\dot\a I} + 2 \eta^{IJ} \e_{\a\b} \, \lambda^\b_J\,, \hspace{2cm}\delta w_{\dot \a I} = \eta_{\a\dot\a} \lambda^{\a}_I + 2 \eta_{IJ} \e_{\dot\a\dot\b} \, \lambda^{\b J}\,, \nonumber \\ &&\delta \hat w_{\a I} = \kappa_{\a\dot\a} \hat\lambda^{\dot\a}_I + 2 \kappa^{IJ} \e_{\a\b} \, \hat\lambda^{\b J}\,, \hspace{2cm} \delta \hat w_{\dot \a}^I = \kappa_{\a\dot\a} \hat\lambda^{\a I} + 2 \eta^{IJ} \e_{\dot\a\dot\b} \, \hat\lambda^{\b}_I\,, \end{aligned}$$ We can also introduce the pure spinor Lorentz generators $(N=-\{w,\lambda\},\hat N=-\{\hat w,\hat\lambda\})\in {\mathcal{H}}_0$, which are needed in the action and determine the couplings between the pure spinor fields and matter fields, as follows $$\begin{aligned} \label{lollo} N_{\alpha\beta} &=& w^I_{(\alpha} \lambda_{\beta)I},\;\;\;\;\;\;\; {\hat N}_{\alpha\beta}= w_{I (\alpha} \lambda_{\beta)}^I\,, \\ \nonumber { N}_{\dot\alpha\dot\beta} &=& { w}_{(\dot\alpha I} {\lambda}_{\dot\beta)}^I,\;\;\;\;\;\;\; {\hat N}_{\dot\alpha\dot\beta}= {\hat w}_{(\dot\alpha}^I {\hat\lambda}_{\dot\beta) I}\,, \\\nonumber N_I^{\,\, J} &=& w^I_\alpha \lambda^\alpha_I + w_{I \dot\alpha} \lambda^{I \dot\alpha}\,, \\ \nonumber \hat N_I^{\,\,J} &=& {\hat w}_I^\alpha {\hat\lambda}_\alpha^I + {\hat w}^{I \dot\alpha} {\hat\lambda}_{I \dot\alpha}\,.\end{aligned}$$ They are gauge invariant under the transformations (\[gautra\]). Finally, we can write the pure spinor ghost piece of the action $$\begin{aligned} S_{\text{ghost}}&=&R^2\displaystyle\int d^2 z \Big( {w}^{I}_\alpha \bar\nabla \lambda_{I}^\alpha+ {w}_{\dot\alpha I} \bar\nabla \lambda^{\dot\alpha I}+{\hat w}_{\alpha I} {\nabla} {\hat\lambda}^{ I \alpha}+{\hat w}_{\dot\alpha}^I {\nabla} {\hat\lambda}_{I}^{\dot\alpha} \\ \nonumber &-&\eta^{(\alpha\beta)(\gamma\delta)}N_{\alpha\beta}{\hat N}_{\alpha\delta} - \eta^{(\dot\alpha\dot\beta)(\dot\gamma\dot\delta)}N_{\dot\alpha\dot\beta}{\hat N}_{\dot\gamma\dot\delta}-\eta^{I\,\,K}_{\,\,J\,\,L} N_I^{\,\, J}{\hat N}_K^{\,\, L} \Big)\,,\end{aligned}$$ where the bilinear metrics $\eta$ are given from (\[bil1\]) and (\[bil2\]) as $$\begin{aligned} \eta^{(\alpha\beta)(\gamma\delta)}=\epsilon^{\alpha \gamma} \epsilon^{\beta \delta} + \epsilon^{\alpha \delta} \epsilon^{\beta\gamma},\;\;\;\;\;\eta^{I\,\,K}_{\,\,J\,\,L}=\delta^{~I}_{L}\delta_{J}^{~K} \,.\end{aligned}$$ Putting everything together we get the pure spinor action for $AdS_4\times \mathbb{CP}^3$ $$\begin{aligned} \label{pure1} S&=&R^2\displaystyle\int d^2z \Big[ \epsilon_{xy} J^x \bar{J}^y+\frac{1}{2} J_{IJ} \bar{J}^{IJ} + \frac{1}{4}\left( J_{\alpha I}{\bar J}^{\alpha I}+J_{\dot\alpha\bar I}{\bar J}^{\dot\alpha\bar I}\right) +\frac{3}{4} \left( J_{\alpha \bar I}{\bar J}^{\alpha \bar I}+J_{\dot\alpha I}{\bar J}^{\dot\alpha I} \right) \\ \nonumber &&\;\;\;\;\;\;\;\;\;+ {w}^{I}_\alpha \bar\nabla \lambda_{I}^\alpha+ {w}_{\dot\alpha I} \bar\nabla \lambda^{\dot\alpha I}+{\hat w}_{\alpha I} {\nabla} {\hat\lambda}^{ I \alpha}+{\hat w}_{\dot\alpha}^I {\nabla} {\hat\lambda}_{I}^{\dot\alpha} \\ \nonumber &&\;\;\;\;\;\;\;\;\;-\eta^{(\alpha\beta)(\gamma\delta)}N_{\alpha\beta}{\hat N}_{\gamma\delta}- \eta^{(\dot\alpha\dot\beta)(\dot\gamma\dot\delta)}N_{\dot\alpha\dot\beta}{\hat N}_{\dot\gamma\dot\delta}-\eta^{I\,\,K}_{\,\,J\,\,L} N_I^{\,\, J}{\hat N}_K^{\,\, L} \Big]\,,\end{aligned}$$ The theory admits a BRST transformation with the following BRST charge $$\begin{aligned} {Q}+\bar{{Q}}&=&\displaystyle\oint \left< dz \lambda J_3+d\bar z \hat\lambda {\bar J}_1 \right> \\ \nonumber &=&\displaystyle\oint dz \left( \lambda_{I \alpha} \hat J^{\alpha I}+ \lambda^{\dot\alpha I} \hat J_{\dot\alpha I} \right) + \displaystyle\oint d\bar z \left( {\hat\lambda}^{\alpha I} {\bar J}_{\alpha I}+ {\hat\lambda}^{\dot\alpha}_I {\bar J}_{\dot\alpha}^I \right)\,.\end{aligned}$$ The general pure spinor action with ${\mathbb{Z}}_4$ discrete symmetry is invariant under the following BRST variations $$\begin{aligned} \delta_B (J_0)&=&[J_3,\lambda]+[{\bar J}_1,\hat\lambda] \\ \nonumber \delta_B (J_1)&=&\nabla \lambda +[J_2,\hat\lambda] \\ \nonumber \delta_B (J_2)&=&[J_1,\lambda]+[J_3,\hat\lambda] \\ \nonumber \delta_B (J_3)&=&\nabla\hat\lambda+[J_2,\lambda] \\ \nonumber \delta_B (\lambda) &=& 0,\;\;\;\;\;\;\;\;\;\;\;\; \delta_B (\hat\lambda) = 0\ \\ \nonumber \delta_B (\omega) &=& -J_3,\;\;\;\;\;\;\,\, \delta_B (\hat\omega) = -{\bar J}_1\ \\ \nonumber \delta_B (N) &=& [J_3,\lambda], \;\;\;\; \delta_B (\hat N) = [{\bar J}_1,\hat\lambda]\end{aligned}$$ where $\nabla Y=\partial Y+[J_0,Y]$ and $\bar\nabla Y=\bar\partial Y+[{\bar J}_0,Y]$. These can be written in the following form for the $AdS_4\times \mathbb{CP}^3$, $$\begin{aligned} \label{BRST} \delta_B J_{\alpha\beta}&=& - 2 \lambda_{(\a I} \hat J^{I}_{\b)} - 2 J_{(\a I} \hat \lambda^{I}_{\b)} \,, \hspace{1cm} \delta_B J_{\dot\alpha\dot\beta}= - 2 \lambda_{(\dot \a}^I \hat J_{\dot \b)I } - 2 J_{(\dot\a}^I \hat \lambda_{\b)I} \,, \\ \nonumber \delta_B \hat J^{\alpha I}&=& (\nabla \hat\lambda)^{\alpha I} + J^{IJ} \lambda^\a_{J} + J^\a_{~\dot\a} \lambda^{\dot\a I} ,\hspace{.3cm} \delta_B \hat J^{\dot\alpha}_I= (\nabla \hat\lambda)^{\dot\alpha}_I + J_{IJ} \lambda^{\dot\alpha J} + J_\a^{~\dot\a} \lambda^{\a}_I \\ \nonumber \delta_B J^{\alpha}_I&=& (\nabla \lambda)^{\alpha}_I + J_{IJ} \hat\lambda^{\a J} + J^\a_{~\dot\a} \hat\lambda^{\dot\a}_I ,\hspace{.4cm} \delta_B J^{\dot\alpha I}= (\nabla \lambda)^{\dot \alpha I} + J^{IJ} \hat\lambda^{\dot\a}_J + J^\a_{~\dot\a} \hat\lambda^{\dot\a I}\,, \\ \nonumber \delta_B J_{\alpha\dot\beta}&=& \lambda_{\a I} J^{I}_{\dot\b} + J_{\a I} \lambda^{I}_{\dot\b} + \hat J_{\dot \beta I} \hat \lambda^{I}_{\a } + \hat \lambda_{\dot \beta I} \hat J^{I}_{\a } \,, \hspace{1cm} \\\nonumber \delta_B J_{IJ}&=& 2\, \e^{\a\b} \lambda_{\a [I} J_{J] \b} + 2\, \e^{\dot\a\dot\b} \hat J_{\dot\a [I} \hat\lambda_{J]\dot\b} \,, \\\nonumber \delta_B J^{IJ} &=& 2\, \e^{\a\b} \lambda_{\a}^{[I} J^{J]}_{\b} + 2\, \e^{\dot\a\dot\b} \hat J_{\dot\a}^{[I} \hat\lambda^{J]}_{\dot\b} \,, \\\nonumber \delta_B \omega_{\alpha}^I&=& - \hat J_\a^I\,, \hspace{4cm} \delta_B \omega_{\dot\alpha I}= - \hat J_{\dot \a I}\,, \\ \nonumber \delta_B {\hat\omega}_{\alpha I}&=& - J_{\a I}\,, \hspace{3.8cm} \delta_B {\hat\omega}^{I}_{\dot\alpha}= - J_{\dot\a}^I\,,\end{aligned}$$ the variations of $N_{\a\b}, N_{\dot\a\dot\b}, \hat N_{\a\b}, \hat N_{\dot\a\dot\b}$ can be easily derived by their definitions (\[lollo\]). Using this notation, we can assign a further quantum number by assigning $0$ to $J_{\a\dot\a}$, $+1$ to $J^{IJ}$, $-1$ to $J_{IJ}$, $-1/2$ to $J_{\a I}, \hat J_{\dot \a, I}$ and $+1/2$ to $\hat J_{\a I}, J_{\dot \a, I}$. This is the center of $U(1)$ inside of $U(3)$. Notice that the symmetry is a $\mathbb Z_5$ symmetry. The action, the BRST transformations and the pure spinor conditions respect such a symmetry. Kähler potential for the Grassmannian and the A-model ===================================================== Let’s consider the Grassmannian coset $\frac{Ops(6|4)}{SO(6)\times Sp(4)}$ which is obtained out of the similar twisted coordinates $\Theta_A^x$ which was introduced by Berkovits for $AdS_5\times S^5$ [@Berkovits1]. A general Kähler potential on a coset $G/H$ was shown in [@itoh] to have the form $$\label{kah} K(\Theta,\bar\Theta)=\frac{1}{2}\ln \det \left( \bar\xi(\bar\Theta) \xi(\Theta) \right)\,,$$ where $\xi(\Theta)\in G/H$ is a representative of the coset $G/H$ where for any $h\in H$ and $g\in G$ satisfies $$g \xi(\Theta)=\xi(\Theta')h(\Theta,g)\,,$$ Like in the case of the $G/H=\frac{PU(2,2|4)}{SU(4)\times SU(2,2)}$ coset [@Berkovits1], for $G/H=\frac{Ops(6|4)}{SO(6)\times Sp(4)}$ also, there exists a gauging in which the coset representative can be written in the following form $$\xi={\left( \begin{array}{ccccccccccccccccccccccccccccccccccccccccc} \ma_{4\times 4}&\Theta\\\bar\Theta&\ma_{6\times 6}\\ \end{array} \right)},\;\;\; \bar\xi={\left( \begin{array}{ccccccccccccccccccccccccccccccccccccccccc} \ma_{4\times 4}&{\Theta}\\-\bar\Theta&\ma_{6\times 6}\\ \end{array} \right)}\,,$$ where here, $\Theta_A^x$ and $\bar\Theta_x^A$ are $4\times 6$ and $6\times 4$ fermionic matrices respectively. Using the convention $i\bar\Theta=\Theta^\dagger$, the Kähler potential (\[kah\]) can be written as $$\begin{aligned} \label{kah1} \nonumber K(\Theta,\bar\Theta)&=&\frac{1}{2}\ln\det \left[ {\left( \begin{array}{ccccccccccccccccccccccccccccccccccccccccc} \ma_{4\times 4}&\Theta\\\bar\Theta&\ma_{6\times 6}\\ \end{array} \right)} {\left( \begin{array}{ccccccccccccccccccccccccccccccccccccccccc} \ma_{4\times 4}&{\Theta}\\-\bar\Theta&\ma_{6\times 6}\\ \end{array} \right)} \right] \\ \nonumber &=&\frac{1}{2}\ln\det \left[ {\left( \begin{array}{ccccccccccccccccccccccccccccccccccccccccc} \ma_{4\times 4}-\Theta\bar\Theta& 0\\ 0 & \ma_{6\times 6}+\bar\Theta\Theta\\ \end{array} \right)} \right] \\ \nonumber &=&\frac{1}{2}\ln\left[\det(\ma_{4\times 4}-\Theta\bar\Theta)\times\det(\ma_{6\times 6}+\bar\Theta\Theta) \right] \\ &=& \,\Tr \,\ln (\ma_{6\times 6}+\bar\Theta\Theta)\end{aligned}$$ which in the last line we used the fact that $$\begin{aligned} \Tr (\Theta\bar\Theta)^n=-\Tr (\bar\Theta\Theta)^n, \;\;\text{for }n>0\,,\end{aligned}$$ One can easily show, for such a Kähler potential, exactly in the same way as it was shown in section (4.3) of [@Berkovits1], that this N=2 action is conformal invariant, namely by computing the one-loop beta function $$R=\ln \det (\partial_\Theta\partial_{\bar\Theta} K)=0\,,$$ which then the $N=2$ supersymmetry non-renormalization theorem implies its conformal invariance to all loops. The worldsheet variables for this Kähler N=2 sigma-model on $\frac{Osp(6|4)}{SO(6)\times Sp(4)}$ are fermionic superfields $\Theta_A^x$ and $\bar\Theta_x^A$ where $A=1,\ldots ,6$ and $x=1,\ldots ,4$ label fundamental representations of $SO(6)$ and $Sp(4)$ respectively. These $N=2$ chiral and anti-chiral superfields can be expanded in terms of the fields of the pure spinor superstring theory on the target $AdS_4\times \mathbb{CP}^3$ as follows $$\begin{aligned} \Theta_A^x(\kappa_{+},\kappa_{-})&=&\theta_A^x+\kappa_{+}Z_A^x+\kappa_{-}\bar{Y}_A^x+\kappa_{+}\kappa_{-}f_A^x\,, \\ \nonumber \bar\Theta^A_x(\bar\kappa_{+},\bar\kappa_{-})&=&\bar\theta^A_x+\bar\kappa_{+}\bar{Z}^A_x+\bar\kappa_{-}{Y}^A_x+ \bar\kappa_{+}\bar\kappa_{-}\bar{f}^A_x\,,\end{aligned}$$ where $(\kappa_{+},\bar\kappa_{+})$ are left-moving and $(\kappa_{-},\bar\kappa_{-})$ are right-moving Grassmannian parameters of the worldsheet N=2 supersymmetry. In this expansion, the 24 lowest components $\theta_A^x$ and $\bar\theta_x^A$ are 24 fermionic coordinates of the $\frac{Osp(6|4)}{U(3)\times SO(1,3)}$ supercoset which parametrizes the $AdS_4\times \mathbb{CP}^3$ superspace together with the 24 bosonic variables $Z_A^x$ and $\bar{Z}_x^A$ which are twistor-like variables combining the 10 spacetime coordinates of $AdS_4$ and $\mathbb{CP}^3$ with pure spinors $(\lambda_A^x,\bar\lambda^A_x)$ which their number was obtained in [@antonio] to be 14. They can be expressed explicitly as follows $$\begin{aligned} Z_A^x&=&H^x_{x'}(x_A)(\tilde{H}^{-1}({x_P}))_A^{A'}\,\lambda_{A'}^{x'}\,, \\ \nonumber \bar{Z}^A_x&=&(H^{-1}(x_A))_{x}^{x'}\tilde{H}_{A'}^A({x_P})\,\bar\lambda^{A'}_{x'}\,,\end{aligned}$$ Here $H^x_{x'}(x_A)$ is a coset representative for the $AdS_4$ coset $\frac{Sp(4)}{SO(1,3)}$ and $\tilde{H}_{A'}^A(x_P)$ is a coset representative for the $\mathbb{CP}^3$ coset $\frac{SO(6)}{U(3)}$. Similarly, the conjugate twistor-like variables $Y_J^A$ and $\bar{Y}^J_A$ are constructed from the conjugate momenta to the pure spinors and $f_A^x$ and $\bar{f}_x^A$ are auxiliary fields. From pure spinor to A-model =========================== Here we show that the same way Berkovits and Vafa [@BV] showed the equivalence of the A-model and the pure spinor superstrig for $AdS_5\times S^5$, we can show the existence of such an equivalence for any superscoset admitting a $\mathbb{Z}_4$ automorphism, as is the case also for the $AdS_4\times \mathbb{CP}^3$ supercoset. Pure spinor with $\mathbb{Z}_4$ automorphism and “bonus“ symmetry ----------------------------------------------------------------- Consider a supercoset $G/H$ which admits a $\mathbb{Z}_4$ automorphism under which its generators can be decomposed into invariant subspaces $\mathcal{H}_i,i=0\cdot 3$. The matter fields of the sigma model can be written in terms of the left-invariant currents $J=g^{-1}\partial g,\,\bar{J}=g^{-1}\bar\partial g$, where $g\in G$. The left-invariant currents are decomposed according to the invariant subspaces of the $\mathbb{Z}_4$ into $J=J_0+J_1+J_2+J_3$ as follows $$\begin{matrix} \mathcal{H}_0&\mathcal{H}_1&\mathcal{H}_2&\mathcal{H}_3\\J^{[AB]}&J^\alpha&J^M&J^{\hat\alpha} \end{matrix}$$ where the left-invariant current $J=g^{-1}\partial g$ is expanded by the generators of the superalgebra as $$J=\sum_{i=0}^3 J_i=J^{[AB]}T_{[AB]}+J^m T_m+J^\alpha T_\alpha+J^{\hat\alpha} T_{\hat\alpha}\,,$$ here, $J^{[AB]}\in H$ are the spin connections of the supercoset and $J^m$ and ($J^\alpha,J^{\hat\alpha}$) are the bosonic and fermionic components of the supervielbein respectively. The generators of the supercoset are $(T_{[AB]},T_m,T_\alpha,T_{\hat\alpha})$ which are the Lorentz generators, translations and fermionic generators respectively with the following non-zero structure constants $$\begin{aligned} f_{mn}^{\;\;\;\; p},\;\;f_{mn}^{\;\;\;[AB]},\;\;f_{{[AB]}{[CD]}}^{\;\;\;\;\;\;\;\;\;[EF]},\;\;f_{\alpha\hat\beta}^{\;\;\;[AB]},\;\;f_{\alpha\beta}^{\;\;\; m}\,,\end{aligned}$$ Besides the matter fields, the pure spinor action has a ghost sector consisting of the pure spinors and their conjugate momenta $$\lambda=\lambda^\alpha T_\alpha,\;\;\hat\lambda={\hat\lambda}^{\hat\alpha}T_{\hat\alpha},\;\;\omega=\eta^{\alpha\hat\alpha} \omega_\alpha T_{\hat\alpha},\;\;\hat\omega=\eta^{\alpha\hat\alpha}{\hat\omega}_{\hat\alpha} T_{\alpha}\,,$$ and the corresponding pure spinor currents $N=-\{\omega,\lambda\},\hat{N}=\{\hat\omega,\hat\lambda\}\in \mathcal{H}_0$ which generate the Lorentz transformations in the pure spinor variables. The theory admits a BRST transformation with the following operator $${Q}+\bar{{Q}}=\displaystyle\oint \left< dz \lambda J_3+d\bar z \hat\lambda {\bar J}_1 \right>\,,$$ under which the fields transform as follows $$\begin{aligned} \delta_B (J_0)&=&[J_3,\lambda]+[J_1,\hat\lambda]\,, \\ \nonumber \delta_B (J_1)&=&\nabla \lambda +[J_2,\hat\lambda]\,, \\ \nonumber \delta_B (J_2)&=&[J_1,\lambda]+[J_3,\hat\lambda]\,, \\ \nonumber \delta_B (J_3)&=&\nabla\hat\lambda+[J_2,\lambda]\,, \\ \nonumber \delta_B (\lambda) &=& 0,\;\;\;\;\;\;\;\;\;\;\;\; \delta_B (\hat\lambda) = 0\,, \ \\ \nonumber \delta_B (\omega) &=& -J_3,\;\;\;\;\;\;\,\, \delta_B (\hat\omega) = -{\bar J}_1 \,, \ \\ \nonumber \delta_B (N) &=& [J_3,\lambda], \;\;\;\; \delta_B (\hat N) = [{\bar J}_1,\hat\lambda]\,,\end{aligned}$$ where $\nabla Y=\partial Y+[J_0,Y]$ and $\bar\nabla Y=\bar\partial Y+[{\bar J}_0,Y]$. These can also be written in the expanded form, $$\begin{aligned} \label{BRST1} \delta_B (J^{[AB]}) &=& J^{\hat\alpha} \lambda^\beta f_{\hat\alpha\beta}^{\;\;\;[AB]}+J^{\alpha}{\hat\lambda}^{\hat\beta} f_{\alpha\hat\beta}^{\;\;\;[AB]}\,, \\ \nonumber \delta_B (J^m)&=&J^\alpha \lambda^\beta f_{\alpha\beta}^{\;\;\;m}+J^{\hat\alpha}{\hat\lambda}^{\hat\beta} f_{\hat\alpha\hat\beta}^{\;\;\;m}\,, \\ \nonumber \delta_B (J^\alpha)&=& \nabla \lambda^\alpha + J^m {\hat\lambda}^{\hat\alpha} f_{m \hat\alpha}^{\;\;\;\alpha}\,, \\ \nonumber \delta_B (J^{\hat\alpha})&=& \nabla\hat\lambda^{\hat\alpha}+J^m\lambda^\alpha f_{m\alpha}^{\;\;\;\hat\alpha}\,.\end{aligned}$$ The sigma model is invariant under the global transformations $\delta g=\Sigma g,\; \Sigma\in \cal G$ and under the BRST transformations, using the fact that $\left<AB\right>\neq 0$ only for $A\in \mathcal{H}_i$ and $B\in \mathcal{H}_{4-i}$. It can be written in the following form $$\label{z4} S=R^2\displaystyle\int d^2 z \left< \frac{1}{2} J_2 {\bar J}_2+\frac{1}{4} J_1{\bar J}_3+\frac{3}{4} J_3 {\bar J}_1 +w \bar\partial \lambda + \hat w\partial\hat\lambda+N{\bar J}_0+{\hat N}J_0 - N\hat N\right>\,,$$ for any supercoset admitting a ${\mathbb Z}_4$ automorphism including $AdS_5\times S^5$ and $AdS_4\times \mathbb{CP}^3$ examples (see also [@Adam:2006bt; @Adam:2007ws] for non-critical examples based on different sets of pure spinor variables). On top of the global bosonic isometry group $G_b$ of the supergroup $G$, the A-model action has a ’bonus’ chiral symmetry exchanging left and right movers which appears in the sigma model as a symmetry between left and right moving fermions $J^\alpha$ and $J^{\hat\alpha}$. Apparently (\[z4\]) does not have such a symmetry because of the different coefficients of $J_1{\bar J}_3$ and $ J_3 {\bar J}_1$ terms. To promote the symmetry of (\[z4\]), one can add an additional term to the action including a $-\frac{1}{2} J_3 {\bar J}_1$ to cancel the asymmetry of the fermionic currents together with its appropriate companion in order that the whole term stays a BRST-closed term, $$\begin{aligned} \label{trivial} S_{trivial}&=&S_{m}+S_g \nonumber \\ &=& \frac{R^2}{2}\displaystyle\int d^2z \left( C_{mn}J^m{\bar J}^n-<J_3 {\bar J}_1>+<\omega \bar\nabla \lambda + \hat\omega\nabla\hat\lambda - N\hat N>\right) \\ \nonumber &=&\frac{R^2}{2}\displaystyle\int d^2z \left( C_{mn}J^m{\bar J}^n+\eta_{\alpha\hat\beta} J^{\hat\beta} {\bar J}^\alpha+\omega_\alpha \bar\nabla \lambda^\alpha + \hat\omega_{\hat\alpha}\nabla\hat\lambda^{\hat\alpha} - \eta_{[AB][CD]} N^{[AB]}\hat N^{[CD]} \right)\,,\end{aligned}$$ where $S_g=\frac{R^2}{2} \int d^2z (\omega \bar\nabla \lambda + \hat\omega\nabla\hat\lambda - N\hat N)$ is exactly the ghost part of the original action (\[z4\]) and $\eta_{XY}=<T_XT_Y>=Str(T_X T_Y)$. The requirement of BRST invariance of the $S_{trivial}$ will determine the unknown tensor $C_{mn}$. Using the classical equations of motion $$\begin{aligned} \nabla \hat\lambda-[N,\hat\lambda]=0,\;\;\;\;\bar\nabla\lambda-[\hat N,\lambda]=0\,,\end{aligned}$$ and the identities $[ N,\lambda ] = [ \hat N,\hat\lambda ] =0 $ coming from the pure spinor constraints, it can be shown that under the BRST transformations (\[BRST1\]), $S_{g}$ and $S_m$ vary as follows $$\begin{aligned} \label{1add} \delta_B(S_{g})&=&\frac{R^2}{2}\displaystyle\int d^2 z <-J_3\bar\partial\lambda-{\bar J}_1\partial\hat\lambda-J_3[{\bar J}_0,\lambda]-{\bar J}_1[J_0,\hat\lambda]> \\ \nonumber &=&\frac{R^2}{2}\displaystyle\int d^2 z \;\eta_{\alpha\hat\beta}(-J^{\hat\beta}\bar\nabla\lambda^\alpha+{\bar J}^\alpha\nabla {\hat\lambda}^{\hat\beta})\,, \\ \label{2add} \delta_B(L_{m})&=&\frac{R^2}{2}\left[C_{mn}\left( J^\alpha \lambda^\beta f_{\alpha\beta}^{\;\;\;\;m}+J^{\hat\alpha}{\hat\lambda}^{\hat\beta}f_{\hat\alpha \hat\beta}^{\;\;\;\;m} \right){\bar J}^n+C_{mn} J^m\left( {\bar J}^\alpha \lambda^\beta f_{\alpha\beta}^{\;\;\;\;n} + {\bar J}^{\hat\alpha}{\hat\lambda}^{\hat\beta}f_{\hat\alpha\hat\beta}^{\;\;\;\;n}\right)\,\right. \\ \nonumber &-&\left. \eta_{\alpha\hat\beta}\left( \nabla\hat\lambda^{\hat\beta}+J^m\lambda^\beta f_{m\beta}^{\;\;\;\;\hat\beta} \right){\bar J}^\alpha+\eta_{\alpha\hat\beta} J^{\hat\beta}\left( \bar\nabla\lambda^{\alpha}+{\bar J}^n\hat\lambda^{\hat\alpha} f_{n\hat\alpha}^{\;\;\;\;\alpha} \right)\,\right],\end{aligned}$$ which gives $$\begin{aligned} \label{delsb} \frac{1}{R^2} \delta_B(S_{trivial})&=&\frac{1}{2} C_{mn} J^m {\bar J}^\alpha\lambda^\beta f_{\alpha\beta}^{\;\;\;\; n}+\frac{1}{2}\eta_{\alpha\hat\beta} J^m{\bar J}^\alpha\lambda^\beta f_{m\beta}^{\;\;\;\;\hat\beta}\,, \\ \nonumber &+&\frac{1}{2}C_{mn}{\bar J}^n J^{\hat\beta}{\hat\lambda}^{\hat\alpha}f_{\hat\beta\hat\alpha}^{\;\;\;\;m}+\frac{1}{2}\eta_{\alpha\hat\beta} {\bar J}^n J^{\hat\beta}{\hat\lambda}^{\hat\alpha}f_{n\hat\alpha}^{\;\;\;\;\alpha} \\ \nonumber &+&\frac{1}{2}C_{mn}{\bar J}^n J^\alpha\lambda^\beta f_{\alpha\beta}^{\;\;\;\;m}+\frac{1}{2}C_{mn} J^m {\bar J}^{\hat\alpha}{\hat\lambda}^{\hat\beta} f_{\hat\alpha\hat\beta}^{\;\;\;\;n} \\\nonumber &=&0\,,\end{aligned}$$ which admits the following solution for $\delta_B(S_{trivial})=0$ after using the Jacobi identities for the structural constants $$\label{cmn} C_{mn}=\frac{1}{2}\frac{\eta_{\alpha\hat\beta}({\hat\lambda}^{\hat\alpha}f_{n\hat\alpha}^{\;\;\;\;\alpha})(\lambda^\beta f_{m\beta}^{\;\;\;\;\hat\beta})}{\eta_{\alpha\hat\beta}\lambda^\alpha{\hat\lambda}^{\hat\beta}}\,.$$ The first and the second lines of (\[delsb\]) vanish because of the identity $\eta_{\beta\hat\alpha}=Str\left(T_\beta T_{\hat\alpha}\right)= f_{\alpha\beta}^{\,\,\,n} f_{n\hat\alpha}^{\,\,\alpha}$ and the terms in the last line vanish because of the following Jacobi identity, $$\label{jac1} f_{\alpha\gamma}^{\,\,m}f_{m\beta}^{\,\,\hat\beta}+f_{\alpha\beta}^{\,\,m}f_{m\gamma}^{\,\,\hat\beta}=f_{\beta\gamma}^{\,\,m}f_{m\alpha}^{\,\,\hat\beta}\,,$$ which implies $$\lambda^\beta\lambda^\gamma\left(f_{\alpha\gamma}^{\,\,m}f_{m\beta}^{\,\,\hat\beta}+f_{\alpha\beta}^{\,\,m}f_{m\gamma}^{\,\,\hat\beta}\right)=0\,. \label{jac2}$$ So $S_{trivial}$ of (\[trivial\]) with $C_{mn}$ given in (\[cmn\]) is BRST-closed. We should also show that it is really a BRST-trivial term satisfying $S_{trivial}=Q\bar Q X$, up to the equations of motion. In order to do that, we introduce the antifields $w^*_\alpha$ and ${{\hat w}^*}_{\hat\alpha}$ which after adding the term $$R^2\displaystyle\int d^2 z \eta^{\alpha\hat\beta}{w^*}_\alpha{\hat w}^*_{\hat\beta}\,,$$ the full action stay invariant under the new BRST transformations, $$\begin{aligned} Q' w_\alpha&=&-\eta_{\alpha\hat\alpha}J^{\hat\alpha},\hspace{24.5mm} {\bar Q}' w_\alpha=w^*_\alpha\,, \\\nonumber Q'{\hat w}_{\hat\alpha}&=&{\hat w}^*_{\hat\alpha},\hspace{34mm} {\bar Q}'{\hat w}_{\hat\alpha}=-\eta_{\hat\alpha\alpha}{\bar J}^\alpha\,, \\\nonumber Q' w^*_\alpha&=&\eta_{\alpha\hat\alpha}(\nabla{\hat\lambda}^{\hat\alpha}-[N,\hat\lambda]^{\hat\alpha}),\hspace{4.5mm}\bar Q' w^*_\alpha=0\,, \\\nonumber Q' {\hat w}^*_{\hat\alpha}&=&0,\hspace{37mm} {\bar Q}' {\hat w}^*_{\hat\alpha}=\eta_{\hat\alpha\alpha}(\bar\nabla{\lambda}^{\alpha}-[\hat N,\lambda]^\alpha)\,, \\\nonumber Q' N&=&[J_3,\lambda],\hspace{30mm} {\bar Q}' N=[w^*,\lambda]\,, \\\nonumber Q' \hat N&=&[{\bar J}_1,\hat\lambda],\hspace{30mm} {\bar Q}' \hat N=[{\hat w}^*,\hat\lambda]\,,\end{aligned}$$ which this BRST transformation is nilpotent off-shell instead of being nilpotent up to the equations of motion. Now consider the following identities $$\begin{aligned} \label{qq1} Q'{\bar Q}' \left(C_{mn}J^m{\bar J}^n\right)&=&C_{mn} \left\lbrace Q'{\bar Q}' (J^m){\bar J}^n +Q'(J^m){\bar Q}'({\bar J}^n) +{\bar Q}'(J^m) Q'({\bar J}^n) + J^mQ'{\bar Q}'({\bar J}^n) \right\rbrace \nonumber \\ \nonumber &=&C_{mn} \left\lbrace \nabla {\hat\lambda}^{\hat\alpha}\,{\hat\lambda}^{\hat\beta} f_{\hat\alpha\hat\beta}^{\;\;\;\; m}{\bar J}^n+ J^m \bar\nabla {\lambda}^{\alpha}\,{\lambda}^{\beta} f_{\alpha\beta}^{\;\;\;\; n} \right\rbrace \\ \nonumber &+& C_{mn} \left\lbrace J^p{\bar J}^n\lambda^\alpha{\hat\lambda}^{\hat\beta}f_{p\alpha}^{\;\;\;\;\hat\alpha}f_{\hat\alpha\hat\beta}^{\;\;\;\;m}+J^m{\bar J}^p \lambda^\alpha {\hat\lambda}^{\hat\beta}f_{p\alpha}^{\;\;\;\;\hat\alpha}f_{\hat\alpha\hat\beta}^{\;\;\;\;n} \right\rbrace \\ \nonumber &+& C_{mn} \left\lbrace J^\alpha{\bar J}^{\hat\alpha} \lambda^\beta{\hat\lambda}^{\hat\beta} f_{\alpha\beta}^{\;\;\;\;m}f_{\hat\alpha\hat\beta}^{\;\;\;\;n} +J^{\hat\alpha}{\bar J}^{\alpha} {\hat\lambda}^{\hat\beta}{\lambda}^{\beta} f_{\hat\alpha\hat\beta}^{\;\;\;\;m} f_{\alpha\beta}^{\;\;\;\;n} \right\rbrace \\ &=&2C_{mn}J^m{\bar J}^n \left(\eta \lambda\hat\lambda\right)\, ,\end{aligned}$$ $$\begin{aligned} \label{qq2} Q'{\bar Q}' \left(N\hat N\right) &=&Q'{\bar Q}'(N) \hat N+Q'(N){\bar Q}'(\hat N)+{\bar Q}'(N)Q' (\hat N)+NQ'{\bar Q}'(\hat N) \\ \nonumber &=&[(\nabla\hat\lambda-[N,\hat\lambda]),\lambda]\hat N+[J_3,\lambda][{\bar J}_1,\hat\lambda]+[w^*,\lambda][{\hat w}^*,\hat\lambda]+N[(\bar\nabla\lambda-[\hat N,\lambda]),\hat\lambda]\,, $$ and, $$\begin{aligned} \label{qq3}\nonumber Q'{\bar Q}' \left((\omega\lambda)(\hat\omega\hat\lambda)\right)&=&Q'{\bar Q}'(\omega\lambda)(\hat\omega\hat\lambda)+ Q'(\omega\lambda){\bar Q}'(\hat\omega\hat\lambda)+{\bar Q}'(\omega\lambda){ Q}'(\hat\omega\hat\lambda)+(\omega\lambda){ Q}'{\bar Q}'(\hat\omega\hat\lambda) \\ \nonumber &=&\frac{1}{2}[(\nabla\hat\lambda-[N,\hat\lambda]),\lambda](\hat w\hat\lambda)+\frac{1}{4}[J_3,\lambda][{\bar J}_1,\hat\lambda]+\frac{1}{4}[w^*,\lambda][{\hat w}^*,\hat\lambda] \nonumber\\ &+&\frac{1}{2}(w\lambda)[(\bar\nabla\lambda-[\hat N,\lambda]),\hat\lambda]\,, $$ to get these identities, we used the equation of motions, (\[delsb\]), (\[jac1\]) and (\[jac2\]) together with the following Jacobi identity, $$\begin{aligned} \label{JI} f_{M\underline\alpha}^{\;\;\;\;\underline\beta}f_{N\underline\beta}^{\;\;\;\;\underline\gamma}-f_{N\underline\alpha}^{\;\;\;\;\underline\beta}f_{M\underline\beta}^{\;\;\;\;\underline\gamma}=f_{MN}^{\;\;\;\;P}f_{P \underline{\alpha}}^{\;\;\;\;\underline{\beta}}\,,\end{aligned}$$ where $M,N,\cdots=\lbrace m,[mn]\rbrace$ and $\underline\alpha,\underline\beta,\cdots=\lbrace \alpha,\hat\alpha\rbrace$. From (\[qq1\]), (\[qq2\]) and (\[qq3\]) one can see that there exists a linear combination of them such that $S_{trivial}=Q\bar Q X$ up to the anti-ghost term, that is up to the momenta equations of motion, $$X=\frac{1}{2}\displaystyle\int d^2 z \frac{1}{\eta_{\alpha\hat\alpha}\lambda^\alpha{\hat\lambda}^{\hat\alpha}}\left[\frac{1}{4}C_{mn}J^m{\bar J}^n+\frac{1}{4}(\omega\lambda)(\hat\omega\hat\lambda)-\frac{1}{8}N\hat N \right]\,.$$ The sigma model action after adding $S_{trivial}$ becomes $$\begin{aligned} \label{ps1} S_b&=&\frac{R^2}{2}\displaystyle\int d^2 z \Big[\left(\frac{1}{2}\frac{\eta_{\alpha\hat\beta}({\hat\lambda}^{\hat\alpha}f_{n\hat\alpha}^{\;\;\;\;\alpha})(\lambda^\beta f_{m\beta}^{\;\;\;\;\hat\beta})}{\eta_{\alpha\hat\beta}\lambda^\alpha{\hat\lambda}^{\hat\beta}}+\eta_{mn}\right) j^m{\bar J}^n \\\nonumber &&\;\;\;\;\;\;\;\;\;\;\;\;\;\;+\frac{1}{2} < J_3 {\bar J}_1-J_1 {\bar J}_3 +\omega\bar\nabla\lambda+\hat\omega\nabla\hat\lambda-N\hat N > \Big]\,\end{aligned}$$ The analysis follows the considerations in the literature, but it is derived in a very general way. Mapping pure spinor to A-model ------------------------------ In order to relate $S_b$ and the A-model action, we should write the supercoset element $g(x,\theta,\bar\theta)\in\frac{G}{H}$ in terms of the Grassmannian coset element $G(\theta,\bar\theta)\in \frac{G}{G_b}$. We can define the following bosonic twisted variables out of the bosonic coset elements $H(x)\in\frac{G_b}{H}$ and the pure spinors in this way $$\begin{aligned} \label{twisted} Z^\alpha&=&[H,\lambda]=H^{[AB]}(x)\lambda^\beta f_{[AB]\beta}^{\;\;\;\;\;\;\;\;\alpha} \\ \nonumber {\bar Z}^{\hat\alpha}&=&[H^{-1},\hat\lambda]=(H^{-1})^{[AB]}(x){\hat\lambda}^{\hat\beta} f_{[AB]\hat\beta}^{\;\;\;\;\;\;\;\;\hat\alpha} \\ \nonumber { Y}^{\hat\alpha}&=&[H^{-1}, w]=(H^{-1})^{[AB]}(x)\eta^{\beta\hat\beta}{w}_{\beta} f_{[AB]\hat\beta}^{\;\;\;\;\;\;\;\;\hat\alpha} \\ \nonumber {\bar Y}^\alpha&=&[H,\hat w]=H^{[AB]}(x) \eta^{\beta\hat\beta}{\hat w}_{\hat\beta} f_{[AB]\beta}^{\;\;\;\;\;\;\;\;\alpha}\end{aligned}$$ Supercoset element $g$ can be parametrized as follows $$\label{gG} g(x,\theta,\bar\theta)=G(\theta,\bar\theta)H(x)$$ where $G(\theta,\bar\theta)=e^{\theta^\alpha T_\alpha+{\bar\theta}^{\hat\alpha} T_{\hat\alpha}}$ and $H(x)=e^{x^m T_m}$ in which $(T_m,T_\alpha,T_{\hat\alpha})$ are the generators of the supercoset $G/H$. According to (\[gG\]), we can decompose the left-invariant currents $J=g^{-1} \partial g$. The pure spinor action can be written into $H$ and $G$ components, corresponding to the purely bosonic part and purely fermionic part of the supercoset as follows $$\begin{aligned} J=H^{-1}\partial H+H^{-1}(G^{-1}\partial G)H\end{aligned}$$ Its componets $J=J^m T_m+J^{[AB]}T_{[AB]}+J^\alpha T_\alpha+J^{\hat\alpha}T_{\hat\alpha}$ can be written as $$\begin{aligned} J^M&=&(H^{-1}\partial H)^M+(H^{-1})^M(G^{-1}\partial G)^P H^{Q}f_{NP}^{\;\;\;\;R}f_{RQ}^{\;\;\;\;M} \\ J^{\underline\alpha}&=&(H^{-1})^M(G^{-1}\partial G)^{\underline\beta} H^N f_{M\underline\beta}^{\;\;\;\;\underline\gamma}f_{\underline\gamma N}^{\;\;\;\;\underline\alpha}\end{aligned}$$ where $M,N,\cdots=\{m,[AB]\}$ and $\underline{\alpha},\underline{\beta},\cdots=\{\alpha,\hat\alpha\}$. The A-model action can be written in terms of the fermionic superfields $(\Theta^\alpha,{\bar\Theta}^{\hat\alpha})$ which was defined before as $S=\int \Tr \ln[1+\bar\Theta\Theta]$. Here we assume that for the Grassmannian supercoset $G/G_b$, there exist a gauging in which the supercoset elements $G$ can be written in the following form $$G^m=\ma,\;\;\;G^{[AB]}=\ma,\;\;\;G^{\alpha}=\theta^\alpha,\;\;\;G^{\hat\alpha}={\bar\theta}^{\hat\alpha}$$ Finally, the A-model action, after integration over the auxiliary fields can be written in this form $$\begin{aligned} \label{amodel} S_A&=&t \displaystyle\int d^2 z\Big[(G^{-1}\partial G)(G^{-1}\bar\partial G)+Y\bar\nabla Z+\bar Y\nabla\bar Z-(YZ)(\bar Z\bar Y)\Big] \\\nonumber &=&t\displaystyle\int d^2 z [\eta_{\alpha\hat\alpha}(G^{-1}\partial G)^\alpha(G^{-1}\bar\partial G)^{\hat\alpha}+\eta_{MN}(G^{-1}\partial G)^M(G^{-1}\bar\partial G)^{N} \\ \nonumber &+& \eta_{\alpha\hat\alpha}Y^{\hat\alpha}(\bar\nabla Z)^\alpha+\eta_{\alpha\hat\alpha}{\bar Y}^\alpha(\nabla\bar Z)^{\hat\alpha}-\eta_{mn}f_{\alpha\hat\alpha}^{\;\;\;\;m}f_{\beta\hat\beta}^{\;\;\;\;n}\left[(Y^{\hat\alpha}Z^{\alpha})({\bar Z}^{\hat\beta}{\bar Y}^{\beta})+(Z^\alpha Y^{\hat\alpha})({\bar Y}^{\beta}{\bar Z}^{\hat\beta})\right]]\end{aligned}$$ where, $$\begin{aligned} \label{cdev} (\bar\nabla Z)^\alpha&=& \bar\partial Z+[G^{-1}\bar\partial G,Z] \\\nonumber &=&\bar\partial Z^\alpha+(G^{-1}\bar\partial G)^{[AB]} Z^\beta f_{[AB]\beta}^{\;\;\;\;\;\;\alpha} \\ \nonumber (\nabla \bar Z)^{\hat\alpha}&=& \partial \bar Z+[G^{-1}\partial G,\bar Z] \\\nonumber &=&\partial {\bar Z}^{\hat\alpha}+(G^{-1}\partial G)^{[AB]} Z^{\hat\beta} f_{[AB]\hat\beta}^{\;\;\;\;\;\;\hat\alpha}\end{aligned}$$ To relate the pure spinor action (\[ps1\]) and the A-model action (\[amodel\]), we use the explicit form of the twisted variables (\[twisted\]). Using (\[twisted\]) and Jacobi identity (\[JI\]), one can write $$\begin{aligned} Y\bar\partial Z&=&[H^{-1},w]\bar\partial\left([H,\lambda]\right) \\\nonumber &=&[H^{-1},w]\left([\bar\partial H,\lambda]+[H,\bar\partial\lambda]\right) \\\nonumber &=&w\bar\partial\lambda+[H^{-1}\bar\partial H,w\lambda] \\\nonumber &=&w\bar\partial\lambda+[H^{-1}\bar\partial H,w\lambda]+[H^{-1}(G^{-1}\bar\partial G)H,w \lambda]-[H^{-1}(G^{-1}\bar\partial G)H,w \lambda] \\\nonumber &=&w\bar\partial\lambda+[\bar J,w\lambda]-[(G^{-1}\bar\partial G),YZ]\end{aligned}$$ which after using (\[cdev\]), we get $$\begin{aligned} Y\bar\nabla Z&=&w\bar\partial\lambda+[\bar J,w\lambda] \\\nonumber &=&w_\alpha\bar\partial\lambda^\alpha+{\bar J}^{[AB]}w_\alpha\lambda^\beta f_{[AB] \beta}^{\,\,\,\,\,\,\,\,\,\,\,\alpha}+\eta_{mn}\eta^{\alpha\beta} {\bar J}^m w_\alpha \lambda^{\gamma} f_{\gamma\beta}^{\,\,\,\,n} \\\nonumber &=&w_\alpha\bar\nabla\lambda^\alpha+\eta_{mn}\eta^{\alpha\beta} {\bar J}^m w_\alpha \lambda^{\gamma} f_{\gamma\beta}^{\,\,\,\,n}\end{aligned}$$ similarly, one can see that $$\begin{aligned} \bar Y\nabla \bar Z&=&\hat w\partial\hat\lambda+[ J,\hat w\hat\lambda] \\\nonumber &=&{\hat w}_{\hat\alpha}\partial{\hat\lambda}^{\hat\alpha}+{J}^{[AB]}{\hat w}_{\hat\alpha}{\hat\lambda}^{\hat\beta} f_{[AB] \hat\beta}^{\,\,\,\,\,\,\,\,\,\,\,\hat\alpha}+\eta_{mn}\eta^{\hat\alpha\hat\beta} { J}^m {\hat w}_{\hat\alpha} {\hat\lambda}^{\hat\gamma} f_{\hat\gamma\hat\beta}^{\,\,\,\,n} \\\nonumber &=&{\hat w}_{\hat\alpha}\nabla{\hat\lambda}^{\hat\alpha}+\eta_{mn}\eta^{\hat\alpha\hat\beta} { J}^m {\hat w}_{\hat\alpha} {\hat\lambda}^{\hat\gamma} f_{\hat\gamma\hat\beta}^{\,\,\,\,n}\end{aligned}$$ the last term simplifies as follows $$\begin{aligned} (YZ)(\bar Z\bar Y)&=&\left([H^{-1},w][H,\lambda]\right)\left([H^{-1},\hat\lambda][H,\hat w]\right) \\\nonumber &=&(w\lambda)(\hat w\hat\lambda) \\\nonumber &=&\eta^{[AB][CD]}\left(f_{\alpha [AB]}^{\;\;\;\;\;\;\;\beta}w_\beta\lambda^\alpha\right)\left({\hat w}_{\hat\beta}{\hat\lambda}^{\hat\alpha}f_{\hat\alpha [CD]}^{\;\;\;\;\;\;\;\hat\beta}\right) -\eta_{mn}\left(\eta^{\alpha\gamma}f_{\alpha\beta}^{\,\,\,\,m}w_\gamma \lambda^\beta \right)\left(\eta^{\hat\alpha\hat\gamma}f_{\hat\alpha\hat\beta}^{\,\,\,\,n}{\hat w}_{\hat\gamma}{\hat\lambda}^{\hat\beta}\right) $$ Putting everything together, we obtain the A-model action in terms of the pure spinor fields $$\begin{aligned} \label{eqs1} S_A&=&t\displaystyle\int d^2 z \Big[\frac{1}{2}\eta_{\alpha\hat\beta}(J^{\hat\beta}{\bar J}^\alpha-J^\alpha{\bar J}^{\hat\beta}) +w \bar\nabla\lambda+{\hat w}\nabla{\hat\lambda}-N\hat N \\\nonumber &&\;\;\;\;\;\;\;\;\;\;\;\;\;\;+\eta^{\alpha\beta} {\bar J}^m w_\alpha \lambda^{\gamma} f_{m\gamma}^{\,\,\,\alpha}+\eta^{\hat\alpha\hat\beta} { J}^m {\hat w}_{\hat\alpha} {\hat\lambda}^{\hat\gamma} f_{m\hat\gamma}^{\,\,\,\hat\alpha}-\eta_{mn}\left(\eta^{\alpha\gamma}f_{\alpha\beta}^{\,\,\,\,m}w_\gamma \lambda^\beta \right)\left(\eta^{\hat\alpha\hat\gamma}f_{\hat\alpha\hat\beta}^{\,\,\,\,n}{\hat w}_{\hat\gamma}{\hat\lambda}^{\hat\beta}\right)\Big]\end{aligned}$$ The equations of motion for $w$ and $\hat w$ comes from the variation of the action under the transformations $\delta w_\alpha=f_{\alpha\beta}^{\,\,\,\,m} \lambda^\beta \Lambda_m$ and $\delta{\hat w}_{\hat\alpha}=f_{\hat\alpha\hat\beta}^{\,\,\,\,m}{\hat\lambda}^{\hat\beta}{\tilde\Lambda}_m$, as follows $$\begin{aligned} \label{eqw} (f_{m\alpha}^{\,\,\,\,\hat\delta}\lambda^\alpha)\left({\bar J}^m-\eta^{\hat\beta\hat\gamma}f_{\hat\beta\hat\alpha}^{\,\,\,\,m}{\hat w}_{\hat\gamma}{\hat\lambda}^{\hat\alpha}\right)&=&0 \\\nonumber (f_{m\hat\alpha}^{\,\,\,\,\delta}{\hat\lambda}^{\hat\alpha})\left({ J}^m-\eta^{\beta\gamma}f_{\beta\alpha}^{\,\,\,\,m}{ w}_{\gamma}{\lambda}^{\alpha}\right)&=&0\end{aligned}$$ After inserting these equations of motion into (\[eqs1\]), the second line of (\[eqs1\]) produces the kinetic term for the bosonic Maurer-Cartan currents, $$t\displaystyle\int d^2 z \Big[\frac{1}{2}\frac{\eta_{\alpha\hat\beta}({\hat\lambda}^{\hat\alpha}f_{n\hat\alpha}^{\;\;\;\;\alpha})(\lambda^\beta f_{m\beta}^{\;\;\;\;\hat\beta})}{\eta_{\alpha\hat\beta}\lambda^\alpha{\hat\lambda}^{\hat\beta}}+\eta_{mn}\Big] J^n{\bar J}^m$$ Then the action (\[eqs1\]), becomes $$\begin{aligned} S&=&t\displaystyle\int d^2 z \Big[\left(\frac{1}{2}\frac{\eta_{\alpha\hat\beta}({\hat\lambda}^{\hat\alpha}f_{n\hat\alpha}^{\;\;\;\;\alpha})(\lambda^\beta f_{m\beta}^{\;\;\;\;\hat\beta})}{\eta_{\alpha\hat\beta}\lambda^\alpha{\hat\lambda}^{\hat\beta}}+\eta_{mn}\right) J^n{\bar J}^m +\frac{1}{2}\eta_{\alpha\hat\beta}(J^{\hat\beta}{\bar J}^\alpha-J^\alpha{\bar J}^{\hat\beta} ) \\ \nonumber &&\;\;\;\;\;\;\;\;+w \bar\nabla\lambda+{\hat w}\nabla{\hat\lambda}-N\hat N \Big]\end{aligned}$$ which coincides with the action (\[ps1\]) after identifying $t=\frac{1}{2}R^2$. Linear gauged $\sigma-$model for $AdS_4\times \mathbb{CP}^3$ ============================================================ Similarly to the non-linear sigma model of the $AdS_5\times S^5$ which was studied by Berkovits and Vafa in [@BV], we can write a linear gauged sigma model for the non-linear sigma model for $AdS_4\times \mathbb{CP}^3$ which was given in the previous section. The 2-dimensional N=(2,2) linear gauged sigma model can be described by a set of matter fields which are chiral and antichiral superfields $\Phi_R^\Sigma$ and ${\bar\Phi}_\Sigma^R$ gauged under the real worldsheet superfield $V_S^R$ taking value in the $SO(6)$ gauge group where $R,S,...=1,\ldots ,6$ are gauge field indices and $\Sigma=(x,A)$ is a global $Osp(6|4)$ index. We can take $\Phi_R^x$ to be fermionic while $\Phi_R^A$ are bosonic superfields. The gauged linear sigma model action can be written in a $Osp(6|4)$ invariant way as $$S=\displaystyle\int d^2z \displaystyle\int d^4\kappa \left[ {\bar\Phi}_\Sigma^S (e^V)_S^R {\Phi}^\Sigma_R+t \,\Tr V+\frac{1}{e^2}\Sigma^2 \right]$$ where $\Sigma=\bar{D} D V$ is the field strength of the gauge field $V$ and is a twisted chiral superfield. As it is clear from the matter content of the theory, it contains 24 fermions and 36 bosons and so the theory actually has conformal anomaly if we ask the bosons and fermions to be gauged in the same representation of the gauge group as we did. But still the theory has a conformal IR fixed point corresponding to the large volume and gauge coupling limit which after integrating out the auxiliary equations of motion for the gauge field we obtain the non-linear sigma model (when $e\rightarrow \infty$) $$S=t\displaystyle\int d^2 z \displaystyle\int d^4\kappa \Tr \left[ {\bar\Phi}_\Sigma^R \Phi_S^\Sigma \right]$$ which can be rewritten in terms of the meson fields $\Theta_A^x$ and $\bar\Theta_x^A$ defined as $$\begin{aligned} \Theta_A^x\equiv \Phi_R^x (\Phi^{-1})^R_A,\;\;\;\;\;\;\bar\Theta^A_x\equiv ({\bar\Phi}^{-1})_R^A\bar\Phi^R_x \end{aligned}$$ which gives exactly the A-model sigma model which was obtained from the pure spinor string for $AdS_4\times \mathbb{CP}^3$ as $$S=t \displaystyle\int d^2z \displaystyle\int d^4\kappa \Tr\,\ln\left[1+\bar\Theta\Theta\right]$$ The FI parameter corresponds to the Kähler parameter of the supercoset Grassmannian target space $\frac{Osp(6|4)}{SO(6)\times Sp(4)}$. Vacua of the gauged linear sigma model and zero radius limit ------------------------------------------------------------ The small radius limit of the gauged linear sigma-model is convenient to study the perturbative regime of the gauge theory since the introduction of the Coulomb branch, because of the presence of the gauge group which is an additional degree of freedom in the gauged linear sigma model with respect to non-linear sigma model, resolves the singularity of the non-linear sigma-model in the small radius limit. To study different phases of the theory, we should solve the D-term equations comming from the gauged linear sigma-model. It is enough to focus on the fields which have conformal weight zero because they are the only fields which can get non-zero expectation value. We analyze the gauged linear $\sigma$-model following the standard techniques of [@Witten:1993yc] and [@Seki:2006cj]. The gauge superfield $V_S^R$ in Wess-Zumino gauge can be expanded as $$V_S^R=\sigma_S^R \kappa_+{\bar\kappa}_-+{\bar\sigma}_S^R {\kappa}_+{\bar\kappa}_++\ldots+\kappa_+\kappa_-{\bar\kappa}_+{\bar\kappa}_-D_S^R$$ similarly we can expand the fermionic and bosonic superfields as follows $$\begin{aligned} \Phi_R^\Sigma=\phi_R^\Sigma+\kappa_+ \psi_R^\Sigma+\ldots,\;\;\;\;\; {\bar\Phi}^R_\Sigma={\bar\phi}^R_\Sigma+{\bar\kappa}_- {\bar\psi}^R_\Sigma+\ldots\end{aligned}$$ where we just keep the components which will have zero conformal weight after the A-twist because they are the only fields which can attain nonzero expectation value and so can appear in the D-term equations. Here the index $\Sigma$ refers to both $x$ and $A$ indices. Note that $(\phi_R^A,\psi_R^x,{\bar\phi}_A^R,{\bar\psi}_x^R)$ are bosonic and $(\phi_R^x,\psi_R^A,{\bar\phi}_x^R,{\bar\psi}_A^R)$ are fermionic fields. Using the vector superfield and the usual superderivatives $D_{\pm}$ and ${\bar D}_{\pm}$, one can define the covariant superderivatives as follows $$\begin{aligned} {\mathcal{D}}_{\pm}=e^{-V}D_{\pm}e^{+V},\;\;\;{\bar{\mathcal{D}}}_{\pm}=e^{+V}{\bar D}_{\pm}e^{-V}\end{aligned}$$ Then the field strength $\Sigma$ which is a twisted chiral superfield is constructed as follows $$\begin{aligned} \Sigma&=&\{{\bar{\mathcal{D}}}_+,{\mathcal{D}}_-\} \\ \nonumber &=& \sigma+\ldots+\kappa_+\kappa_-{\bar\kappa}_+{\bar\kappa}_- (D^m D_m \sigma+[\sigma,[\sigma,\bar\sigma]]+i[\partial^m v_m,\sigma])\end{aligned}$$ which produces the following gauge field kinetic term in the Lagrangian $$\begin{aligned} L_{gauge}&=&-\frac{1}{e^2}\displaystyle\int d^4\kappa \Tr \bar\Sigma\Sigma \\ \nonumber &=& \frac{1}{e^2} \Tr\left(-D_i \bar\sigma D^i \sigma -\frac{1}{2} [\sigma,\bar\sigma]^2 +\ldots \right)\end{aligned}$$ and also we have the FI term $L_{D,\theta}$, $$\begin{aligned} L_{D,\theta}&=& {it} \left.\displaystyle\int d\kappa_+ d{\bar\kappa}_- \Tr \Sigma\right|_{\kappa_-={\bar\kappa}_+=0}-{i\bar t} \left.\displaystyle\int d\kappa_- d{\bar\kappa}_+ \Tr \bar\Sigma\right|_{\kappa_+={\bar\kappa}_-=0} \\ \nonumber &=&\Tr \left( -rD +\frac{\theta}{2\pi} v_{01} \right)\end{aligned}$$ Now we can consider the matter part of the gauged linear sigma model consisting of the kinetic terms for the fermionic and bosonic superfields which carries the kinetic and interaction terms for the bosonic and fermionic fields, $$\begin{aligned} L_{kin}^b&=& \displaystyle\int d^4 \kappa {\bar\Phi}_A^R e^V \Phi_R^A \\ \nonumber &=& -({\bar D}_j {\bar\phi}_A^R)(D^j \phi_R^A)+{\bar F}^R_A F_R^A-{\bar\phi}_A^S \{\sigma,\bar\sigma \}^R_S \phi_R^A+{\bar\phi}_A^S D_S^R \phi^A_R + \ldots \end{aligned}$$ Similarly we can write the kinetic term for the fermionic chiral superfields, $$\begin{aligned} L_{kin}^f&=& \displaystyle\int d^4 \kappa {\bar\Phi}_x^R e^V \Phi_R^x \\ \nonumber &=& -({\bar D}_j {\bar\phi}_x^R)(D^j \phi_R^x)+{\bar F}^R_x F_R^x-{\bar\phi}_x^S \{\sigma,\bar\sigma \}^R_S \phi_R^x+{\bar\phi}_x^S D_S^R \phi^x_R + \ldots \end{aligned}$$ We can see that $\{\sigma,\bar\sigma\}$ appears as the mass for the matter fields and so whenever $\sigma$ gets VEV, the matter fields become massive and can be integrated out in the effective theory as is happening in the Coulomb phase. The potential of the theory can be written as, $$\begin{aligned} L_V&=&\frac{1}{2e^2} \Tr D^2-r\Tr D+{\bar\phi}_x^S D_S^R \phi^x_R +{\bar\phi}_A^S D_S^R \phi^A_R \\ \nonumber &-&\frac{1}{2e^2} \Tr [\sigma,\bar\sigma]^2-{\bar\phi}_x^S \{\sigma,\bar\sigma \}^R_S \phi_R^x-{\bar\phi}_A^S \{\sigma,\bar\sigma \}^R_S \phi_R^A\end{aligned}$$ which after eliminating the D-field by using the following D-term equation $$D_R^S={\bar\phi}_x^S \phi^x_R +{\bar\phi}_A^S \phi^A_R - r \delta^S_R$$ one obtains the potential $$\begin{aligned} V&=&\frac{e^2}{2} \left[{\bar\phi}_x^S \phi^x_R +{\bar\phi}_A^S \phi^A_R - r \delta^S_R\right]\left[{\bar\phi}_x^R \phi^x_S +{\bar\phi}_A^R \phi^A_S - r \delta^R_S \right] \\ \nonumber &+& \frac{1}{2e^2} \Tr [\sigma,\bar\sigma]^2+{\bar\phi}_x^S \{\sigma,\bar\sigma \}^R_S \phi_R^x+{\bar\phi}_A^S \{\sigma,\bar\sigma \}^R_S \phi_R^A\end{aligned}$$ The space of the classical vacua is given by putting the potential to zero up to gauge transformations. We can study the vacua in two regimes, when $r>0$ and not small, the constraint $V=0$ implies that $\sigma=0$ which implies the following condition as the classical vacua for the matter fields $$D_R^S={\bar\phi}_x^S \phi^x_R +{\bar\phi}_A^S \phi^A_R - r \delta^S_R=0$$ It means actually that the vectors $(\phi^x_R,\psi_R^A)$ for any $R=1,...,4$ are orthonormal. Any such vector, after diagonalization, is subject to the constraint $$\sum_{A=1}^6{\bar\phi}_A \phi^A + \sum_{x=1}^4{\bar\phi}_x \phi^x =r$$ which defines a supersphere ${\mathbb S}^{(5|4)}$.[^4] The space of classical vacua is the gauge invariant subspace of the product of such vectors [@cv] giving the orbit space $({\mathbb S}^{(5|4)})^3//S_3\times \mathbb{Z}_2$ obtained by dividing the action of $S_3\times\mathbb{Z}_2$ on the three copies, where $\mathbb{Z}_2$ is the simultaneous reflection. This phase corresponds to the Higgs phase of the theory because the gauge symmetry completely breaks. If one looks into $r\rightarrow 0$ limit, on top of the above Higgs phase, one can have another possibility as it is explained in [@OV] and [@BV]. In this phase, the $\sigma_R^S$ is unconstrained but the matter variables are constrained to satisfy $$\label{C1} {\mathcal{O}}_R^S={\bar\phi}_x^S \phi^x_R +{\bar\phi}_A^S \phi^A_R=0$$ The mass term for the fermions and bosons are written as $$\begin{aligned} {\bar\phi}_x^S \{\sigma,\bar\sigma \}^R_S \phi_R^x+{\bar\phi}_A^S \{\sigma,\bar\sigma \}^R_S \phi_R^A\end{aligned}$$ And so whenever the $\sigma$ gets expectation value the matter fields become massive and one can integrate them out from the theory. One can easily compute the 1-loop correction to the condition (\[C1\]) which should be proportional to $r$ by doing the path integral with a cut-off $\mu$, $$\begin{aligned} \left<{\mathcal{O}}\right>_{\text{1-loop}}&=&-\sum_{A=1}^6 \displaystyle\int d^2 p \frac{1}{p^2+\{\sigma,\bar\sigma\}}+\sum_{x=1}^4 \displaystyle\int d^2 p \frac{1}{p^2+\{\sigma,\bar\sigma\}} \\ \nonumber &=&-\frac{1}{2\pi}\log \left( \frac{\{\sigma,\bar\sigma\}}{2\mu^2}\right)=r\end{aligned}$$ which has a solution as $$\label{C2} \{\sigma,\bar\sigma\}=2\mu^2 \exp \left(-2\pi r\right)$$ After integrating over all the matter fields, the classical vacua $V=0$ is given by condition $\Tr[\sigma,\bar\sigma]^2=0$ which together with (\[C2\]) gives the following solution, $$\sigma=\sigma_0 \mu \exp \left( -2\pi r\right)$$ where here $\sigma_0$ is an orthogonal $6\times 6$ constant matrix. This means that $\sigma$ can be diagonalized and for each diagonal component of the $\sigma$ in the small radius regime, one gets a copy of the ${\mathbb S}^{(5|4)}$ as it was seen before. Principal chiral model ====================== In this section, we derive the PCM (principal chiral model) for $Osp(6|4)$. We analyze the differences. The model is based on gauging the coset ${Osp(6|4) \over SO(6) \times Sp(4)}$. This is a purely Grassmannian coset manifold with 24 fermionic coordinates $\Theta^x_A$. There are other gaugings leading to ${Osp(6|4) \over Osp(4|2) \times Osp(2|2)}$ and to ${Osp(6|4) \over Osp(4|2) \times SU(1|1,1)}$, but we do not discuss them in the present paper. Notice that unlike ${Osp(6|4) \over SO(6) \times Sp(4)}$, which has 24 fermions, the other two spaces have 12 bosons and 12 fermions. The worldsheet action is $$S = r^2_{AdS} \int d^2z \, {\rm Str} \, \left(( g^{-1} \p g - A) ( g^{-1} \bar\p g - \bar A) \right)$$ where the indices $x,y$ are raised and lowered with $\e^{xy}$. It is invariant under the local symmetry $Osp(6|4)$ under the transformations $$\delta g = g \, \Omega\,, \delta A = d \Omega + [A, \Omega]\,,$$ where $\Omega \in Osp(6|4)$. We can gauge-fix the subgroup $SO(6) \times Sp(4)$ by choosing the gauge $g = G(\theta, \hat\theta) = {\rm exp}( \theta^x_I Q_x^I + \theta^{x I} Q_{x I})$. Furthermore, we can gauge-fix the rest of the symmetries by choosing the gauge $$A^x_I = 0\,, \bar A_x^I = 0\,.$$ This second gauge fixing requires the ghost fields $(\overline Z^x_I, Z_x^I)$ and the antighosts $(\overline Y_x^I, Y^x_I)$ with the action $$S_{ghost} = r^2_{AdS}\int d^2z \, \Big[ - Y^x_I (\overline\nabla Z)_x^I + \overline Y_x^I (\nabla \overline Z)^x_I \Big]\,,$$ where $$(\overline\nabla Z)_x^I = \bar\p Z_x^I + \overline A^{~y}_{x} Z_y^I + \overline A^{I}_{~J} Z_x^J\,, ~~~ (\nabla \overline Z)_{I}^x = \p \overline Z_{I}^x + A^{x}_{~y} \overline Z_{I}^y + A_{I}^{~J} \overline Z_{J}^x \,.$$ Assuming that the kinetic term for the remaining gauge fields $A^I_x, \bar A^x_I$ vanishes in the limit of large RR fluxes, we can integrate out these fields leading to get the complete action $$\begin{aligned} S &=& r^2_{AdS} \int d^2z \, \Big[( G^{-1} \p G )^x_I ( G^{-1} \bar\p G)^{I}_x - Y^x_I (\overline\nabla Z)_x^I + \overline Y_x^I (\nabla \overline Z)^x_I \\ \nonumber &+& ( G^{-1} \p G - A)_{xy} ( G^{-1} \bar\p G - \bar A)^{xy} + ( G^{-1} \p G - A)_{IJ} ( G^{-1} \bar\p G - \bar A)^{IJ} \\ \nonumber &+& ( G^{-1} \p G - A)_{I}^{~J} ( G^{-1} \bar\p G - \bar A)_{J}^{~I} + ( G^{-1} \p G - A)^{IJ} ( G^{-1} \bar\p G - \bar A)_{IJ} \Big]\,.\end{aligned}$$ Notice that the action has the gauge symmetry $SO(6) \times Sp(4)$. Eliminating the gauge fields $A^{xy}\,, \dots\,, \overline A^{IJ}$, one gets a non-linear sigma model which corresponds to the pure spinor sigma model with the addition of a BRST exact term (\[amodel\]). D-branes and gauge theories =========================== In order to discuss open strings and D-branes we have to see how to put the boundary conditions. We start from the supercoset $Osp(6|4)/ SO(6) \times Sp(4)$. We reduce it as follows: the bosonic subcoset: $SO(6) \times Sp(4)$ is reduced to $U(3)\times Sp(2)$ and the fermionic part is halved. This achieved by using the boundary conditions \^[I]{} = \_\^ [J]{}\^I\_J |\^[J]{},       |\_[I]{}\^[ ]{} = \^\_ [J]{}\_I\^J \_[J]{}\^[ ]{}, where ${\cal J}^I_j$ is the complex structure on ${\mathbb P}^3$. The tensor $\delta_\a^{\dot\a}$ reduce the subgroup $Sp(4)$ to $Sp(2)$. We recall that using the symplectic matrices $\Lambda$ of $Sp(4, {\mathbb R})$ as the $4\times 4$ matrices satisfying $\Lambda^T \e \Lambda = \e$ where $\e = {\rm i} \, \sigma_2 \otimes \ma$, we can see immediately the two subgroups $Sp(2, {\mathbb R}) \times Sp(2, {\mathbb R})$. In the above equation, we have selected the diagonal subgroup $Sp(2, {\mathbb R})$. The above equations are invariant under $Sp(2, {\mathbb R}) \times U(3)$. Notice that we have identified on the boundary of the Riemann surface the fermionic variables of the subset ${\cal H}_1 = \{\Theta^{\a I}, \Theta_{I}^{\dot\a}\}$ with those of the other subset ${\cal H}_3 = \{{\bar\Theta}_{I}^{ \a}, {\bar\Theta}^{\dot\a I}\}$. This simply reduces the 24 fermions to 12 ones. The new set of states can be represented in terms of the supercoset (where we have used the isomorphism $Sp(2, {\mathbb R}) \simeq SL(2,{\mathbb R}) \simeq SU(1,1)$). The 6 fermions are in the $(3, 2)$ or in the $(\bar 3, 2)$ representation of the bosonic subgroup. In addition, we have to recall $SL(2, {\mathbb R}) \simeq AdS_3$, which can be seen by parameterizing a group element of $SL(2,{\mathbb R})$ as follows g = ( X\_[-1]{} + X\_1 & X\_0 -X\_2\ - X\_0 - X\_2 & X\_[-1]{} - X\_1 ) with the condition $\det{g} = X_{-1}^2 - X^2_1 + X_0^2 - X^2_2 = 1$ which shows that the $SL(2, \mathbb R)$ group manifold is a 3-dimensional hyperboloid. The metric on $AdS_3$ is given by $ds^2 = - d X_{-1}^2 + d X_{1}^2 - d X_{0}^2 + d X_{2}^2$, which is the invariant metric on the group manifold. Then, we have that these boundary conditions imply a boundary theory of the type $N=6$ super-YM/Chern-Simons on $AdS_3$ space. There is another possibility which is given by the following boundary conditions \^[I]{} = \^\_ \^I\_J |\^[J]{},       |\_[I]{}\^[ ]{} = \_\^ \_I\^J \_[J]{}\^[ ]{}, In this case the supergroup $Osp(6|4)$ is broken to $Osp(6|2) \times SO(2)$. Notice that using the delta $\delta_I^J$ in place of ${\cal J}_I^J$ we do not break the $SO(6)$. In addition, the subgroup $Sp(4)$ is broken to $Sp(2) \times SO(2)$. Now, using the isomorphism $SU(4) \simeq SO(6)$, we can see the coset $SO(6) \times SO(2) / SU(3) \times U(1) \simeq {\mathbb S}^7/{\mathbb Z}_p$ where $p$ defines how the $U(1)$ is embedded in the groups of the numerator. This observation would help us to lift the D-branes solution to KK monopoles of M-theory. The fermions are halved by the boundary conditions. So, the boundary open topological model can be described as the Grassmannian . This solution deserves more attention and the study will be postponed in future publications. Further directions ================== There are several open questions to answer in the framework of gauge/string correspondence and in particular for this peculiar case given by $AdS_4\times \mathbb{CP}^3$. Here we list some of them and we hope to report on them in the near future. To complete the program presented here, one needs to explore the cohomology of the BRST operator in order to check if the bulk and and the boundary theory describe at least at the linearized level the supergravity states we expect. In addition, using the analysis performed in [@Berkovits:2006ik], it should be possible to devise a way to define a pure spinor measure for tree level and higher loop computations. Once this has been established, one of the problems is to prescribe quantum amplitudes for the pure spinor superstring which could be compared with super Chern-Simons amplitudes. It would be interesting to single out a subclass of BPS protected amplitudes whose string counterpart is therefore calculable via the point particle limit and first quantized Chern-Simons theory. Having noticed that the vacuum of the target space theory has a Coulomb branch and the relation with the supersphere $\mathbb{S}^{(5|4)}$, one is tempted to put the gauge amplitude in relation with a topological/twistor string theory on that superspace similarly to [@Witten:2003nn]. Regarding the boundary field theory, we recall that, using the oscillator technique, the UIR of ${\mathrm {Osp}}(6|4)$ are decomposed into representations of its maximal subgroup $SU(3|1,1)$ [@sing]. The singleton is generated out of the vacuum $| 0 \rangle$ and its superpartner $K^{I\alpha} |0\rangle$ where $K^{I \alpha}$ is a fermionic oscillator in the fundamental representation of $SU(3) \times SU(2)$. The quantum numbers of the vacuum are $$\begin{aligned} |0\rangle &=& | j_0 = 0\,, Q_2 =1 \,, \underline{1} \,, Q_3 = -2 \rangle\,, \\ \nonumber K^{i \a} |0\rangle &=& | j_0 = 1/2 \,, Q_2= 2 \,, \underline{3}\,, Q_3 = -1 \rangle\,, \end{aligned}$$ where its energy is given by $E_0 = Q_2/2$. These are the only two states annihilated by the annihilation operators of the subgroup $SU(3|1,1)$. Acting repeatedly with a single-oscillator creation operator ($\a^I$) of $U(3)$ denoted by $L^+ = \a^{[I} \a^{J]}$ we get the states $$\begin{aligned} |0\rangle \,, ~~ L^+ |0\rangle &\longrightarrow& \underline{1}(-2) \oplus \underline{3}^*(0) \\ \nonumber K^{i \a} |0\rangle \,, ~~ L^+ K^{i \a} |0\rangle &\longrightarrow& \underline{3}(-1) \oplus \underline{1}(+1)\end{aligned}$$ The first set is a scalar multiplet that can be recast into a spinorial representation of $SO(6)$, namely the fundamental rep $\underline{4}$ of $SU(4)$. The second set of states forms a multiplet of spin $1/2$ fermions in the $\underline{4}^*$ rep of $SU(4)$. The number of fields coincides exactly with the content of $D2$ brane counting. So, it would be interesting to study the relation between the supersingleton representation and the dual theory [@DallAgataWZ]. Of course the relation with M-theory and the membrane theory should be explored also in the pure spinor context. We thank N. Berkovits, P. Fré, A. Tanzini, M. Trigiante and R. D’Auria for very useful discussions. H.S. is grateful to DISTA where part of this work was done. This research has been supported by the Italian MIUR under the program “Teoria dei Campi, Superstringhe e Gravità”. The work of G.B. and of P.A.G. is supported by the European Commission RTN Program MRTN-CT-2004-005104. [99]{} N. Berkovits, [*Super-Poincare covariant quantization of the superstring,*]{} JHEP [**0004**]{} (2000) 018 \[arXiv:hep-th/0001035\]. N. Berkovits, [*Quantum consistency of the superstring in $AdS_5 \times S^5$ background,*]{} JHEP [**0503**]{} (2005) 041 \[arXiv:hep-th/0411170\]. O. Aharony, O. Bergman, D. L. Jafferis and J. Maldacena, [*N=6 superconformal Chern-Simons-matter theories, M2-branes and their*]{} gravity duals,” arXiv:0806.1218 \[hep-th\]. K. Hosomichi, K. M. Lee, S. Lee, S. Lee and J. Park, JHEP [**0807**]{} (2008) 091 \[arXiv:0805.3662 \[hep-th\]\]. K. Hosomichi, K. M. Lee, S. Lee, S. Lee and J. Park, arXiv:0806.4977 \[hep-th\]. C. Ahn, arXiv:0806.1420 \[hep-th\]. M. Benna, I. Klebanov, T. Klose and M. Smedback, arXiv:0806.1519 \[hep-th\]. J. Bhattacharya and S. Minwalla, arXiv:0806.3251 \[hep-th\]. J. A. Minahan and K. Zarembo, arXiv:0806.3951 \[hep-th\]. A. Armoni and A. Naqvi, arXiv:0806.4068 \[hep-th\]. A. Hanany, N. Mekareeya and A. Zaffaroni, arXiv:0806.4212 \[hep-th\]. D. Gaiotto, S. Giombi and X. Yin, arXiv:0806.4589 \[hep-th\]. C. Ahn, JHEP [**0807**]{} (2008) 101 \[arXiv:0806.4807 \[hep-th\]\]. G. Grignani, T. Harmark and M. Orselli, arXiv:0806.4959 \[hep-th\]. N. Gromov and P. Vieira, arXiv:0807.0437 \[hep-th\]. N. Gromov and P. Vieira, arXiv:0807.0777 \[hep-th\]. M. A. Bandres, A. E. Lipstein and J. H. Schwarz, arXiv:0807.0880 \[hep-th\]. M. Schnabl and Y. Tachikawa, arXiv:0807.1102 \[hep-th\]. M. R. Garousi, A. Ghodsi and M. Khosravi, arXiv:0807.1478 \[hep-th\]. A. Hashimoto and P. Ouyang, arXiv:0807.1500 \[hep-th\]. D. Bak and S. J. Rey, arXiv:0807.2063 \[hep-th\]. E. A. Bergshoeff, O. Hohm, D. Roest, H. Samtleben and E. Sezgin, arXiv:0807.2841 \[hep-th\]. I. Shenderovich, arXiv:0807.2861 \[hep-th\]. C. Krishnan, arXiv:0807.4561 \[hep-th\]. N. Gromov and V. Mikhaylov, arXiv:0807.4897 \[hep-th\]. O. Aharony, O. Bergman and D. L. Jafferis, arXiv:0807.4924 \[hep-th\]. H. Singh, arXiv:0807.5016 \[hep-th\]. G. Bonelli, A. Tanzini and M. Zabzine, arXiv:0807.5113 \[hep-th\]. D. Bak, D. Gang and S. J. Rey, arXiv:0808.0170 \[hep-th\]. A. Giveon and D. Kutasov, arXiv:0808.0360 \[hep-th\]. M. R. Garousi and A. Ghodsi, arXiv:0808.0411 \[hep-th\]. H. Ooguri and C. S. Park, arXiv:0808.0500 \[hep-th\]. D. L. Jafferis and A. Tomasiello, arXiv:0808.0864 \[hep-th\]. D. Martelli and J. Sparks, arXiv:0808.0904 \[hep-th\]. J. Bagger and N. Lambert, [*Comments On Multiple M2-branes,*]{} JHEP [**0802**]{} (2008) 105 \[arXiv:0712.3738 \[hep-th\]\]; [*Gauge Symmetry and Supersymmetry of Multiple M2-Branes,*]{} Phys. Rev.  D [**77**]{} (2008) 065008 \[arXiv:0711.0955 \[hep-th\]\] [*Modeling multiple M2’s,*]{} Phys. Rev.  D [**75**]{} (2007) 045020 \[arXiv:hep-th/0611108\]. A. Gustavsson, [*Algebraic structures on parallel M2-branes,*]{} arXiv:0709.1260 \[hep-th\]. G. Arutyunov and S. Frolov, [*Superstrings on $AdS_4 \times CP^3$ as a Coset Sigma-model,* ]{} arXiv:0806.4940 \[hep-th\]. B. J. Stefanski, [*Green-Schwarz action for Type IIA strings on $AdS_4\times CP^3$,* ]{} arXiv:0806.4948 \[hep-th\]. P. Fré and P. A. Grassi, [*Pure Spinor Formalism for ${Osp}(N|4)$ backgrounds,* ]{} arXiv:0807.0044 \[hep-th\]. B. Chen and J. B. Wu, arXiv:0807.0802 \[hep-th\]. D. Astolfi, V. G. M. Puletti, G. Grignani, T. Harmark and M. Orselli, arXiv:0807.1527 \[hep-th\]. B. H. Lee, K. L. Panigrahi and C. Park, arXiv:0807.2559 \[hep-th\]. C. Ahn, P. Bozhilov and R. C. Rashkov, arXiv:0807.3134 \[hep-th\]. T. McLoughlin and R. Roiban, arXiv:0807.3965 \[hep-th\]. L. F. Alday, G. Arutyunov and D. Bykov, arXiv:0807.4400 \[hep-th\]. A. Tomasiello, [*New string vacua from twistor spaces,*]{} arXiv:0712.1396 \[hep-th\]. N. Berkovits, [*A New Limit of the $AdS_5 \times S^5$ Sigma Model,*]{} JHEP [**0708**]{}, 011 (2007) \[arXiv:hep-th/0703282\]. N. Berkovits and C. Vafa, [*Towards a Worldsheet Derivation of the Maldacena Conjecture,*]{} JHEP [**0803**]{} (2008) 031 \[arXiv:0711.1799 \[hep-th\]\]. G. Bonelli and H. Safaai, [*On gauge/string correspondence and mirror symmetry,*]{} JHEP [**0806**]{} (2008) 050 \[arXiv:0804.2629 \[hep-th\]\]. W. D. . Linch and B. C. Vallilo, [*Integrability of the Gauged Linear Sigma Model for $AdS_5xS^5$,*]{} arXiv:0804.4507 \[hep-th\]. J. Kluson, [*D-brane Description of New Open String Solutions in AdS(5),*]{} arXiv:0805.4719 \[hep-th\]. N. Berkovits, [*Perturbative Super-Yang-Mills from the Topological $AdS_5 \times S^5$ Sigma Model,*]{} arXiv:0806.1960 \[hep-th\]. D. Fabbri, P. Fre, L. Gualtieri and P. Termonia, [*$Osp(N|4)$ supermultiplets as conformal superfields on d(AdS(4)) and the*]{} generic form of N = 2, D = 3 gauge theories,” Class. Quant. Grav.  [**17**]{}, 55 (2000) \[arXiv:hep-th/9905134\]. K. Itoh, T. Kugo and H. Kunitomo, [*Supersymmetric Nonlinear Realization For Arbitrary Kahlerian Coset Space G/H,*]{} Nucl. Phys.  B [**263**]{} (1986) 295. I. Adam, P. A. Grassi, L. Mazzucato, Y. Oz and S. Yankielowicz, [*Non-critical pure spinor superstrings,*]{} JHEP [**0703**]{}, 091 (2007) \[arXiv:hep-th/0605118\]. I. Adam, A. Dekel, L. Mazzucato and Y. Oz, [*Integrability of type II superstrings on Ramond-Ramond backgrounds in various dimensions,*]{} JHEP [**0706**]{}, 085 (2007) \[arXiv:hep-th/0702083\]. E. Witten, [*Phases of N = 2 theories in two dimensions,*]{} Nucl. Phys.  B [**403**]{}, 159 (1993) \[arXiv:hep-th/9301042\]. S. Seki, K. Sugiyama and T. Tokunaga, [*Superconformal symmetry in linear sigma model on supermanifolds,*]{} Nucl. Phys.  B [**753**]{}, 295 (2006) \[arXiv:hep-th/0605021\]. S. Seki and K. Sugiyama, [*Gauged linear sigma model on supermanifold,*]{} arXiv:hep-th/0503074. P. A. Grassi and M. Marescotti, [*Flux vacua and supermanifolds,*]{} JHEP [**0701**]{}, 068 (2007) \[arXiv:hep-th/0607243\]. S. Cecotti and C. Vafa, [*On classification of N=2 supersymmetric theories,*]{} Commun. Math. Phys.  [**158**]{}, 569 (1993) \[arXiv:hep-th/9211097\]. R. Gopakumar and C. Vafa, [*On the gauge theory/geometry correspondence,*]{} Adv. Theor. Math. Phys.  [**3**]{}, 1415 (1999) \[arXiv:hep-th/9811131\]. H. Ooguri and C. Vafa, [*Worldsheet derivation of a large N duality,*]{} Nucl. Phys.  B [**641**]{}, 3 (2002) \[arXiv:hep-th/0205297\]. N. Berkovits, [*Explaining pure spinor superspace,*]{} arXiv:hep-th/0612021. E. Witten, [*Perturbative gauge theory as a string theory in twistor space,*]{} Commun. Math. Phys.  [**252**]{}, 189 (2004) \[arXiv:hep-th/0312171\]. H. Nicolai and E. Sezgin, [*Singleton Representations Of Osp(N,4),*]{} Phys. Lett.  B [**143**]{}, 389 (1984). M. P. Blencowe and M. J. Duff, [*SUPERSINGLETONS,*]{} Phys. Lett.  B [**203**]{} (1988) 229. H. Nicolai, E. Sezgin and Y. Tanii, [*Conformally Invariant Supersymmetric Field Theories On S\*\*P X S\*\*1 And Super P-Branes,*]{} Nucl. Phys.  B [**305**]{} (1988) 483. C. Fronsdal, [*THE SUPERSINGLETON. 1. FREE DIPOLE AND INTERACTIONS AT INFINITY,*]{} Lett. Math. Phys.  [**16**]{} (1988) 163. M. Gunaydin, [*Singleton And Doubleton Supermultiplets Of Space-Time Supergroups And Infinite Spin Superalgebras,*]{} Published in Trieste Supermembr. 1989:0442-456. M. Gunaydin and D. Minic, [*Singletons, doubletons and M-theory,*]{} Nucl. Phys.  B [**523**]{}, 145 (1998) \[arXiv:hep-th/9802047\]. S. Ferrara and C. Fronsdal, [*Conformal Maxwell theory as a singleton field theory on AdS(5), IIB three branes and duality,*]{} Class. Quant. Grav.  [**15**]{}, 2153 (1998) \[arXiv:hep-th/9712239\]. G. Dall’Agata, D. Fabbri, C. Fraser, P. Fre, P. Termonia and M. Trigiante, [*The $Osp(8|4)$ singleton action from the supermembrane,*]{} Nucl. Phys.  B [**542**]{}, 157 (1999) \[arXiv:hep-th/9807115\]. [^1]: For further motivations and results on the $AdS_4\times \mathbb{CP}^3$ from a supergravity point of view, see also [@Tomasiello]. [^2]: In the paper, also the notation $\hat J$ will be used to denote the currents of the subset $\mathcal{H}_3$. [^3]: Another way to solve the constraints (\[pures\]) is decomposing the pure spinor into $\lambda^\a_I = ( \lambda^\a_a, \lambda^\a)$ and $\lambda^{\dot\a I} = (\lambda^{\dot \a a}, \lambda^{\dot \a})$ where $a=1,2$. It is easy to show that the pure spinor constrains become $\lambda^\a_a \lambda^{\dot\a a} + \lambda^\a \lambda^{\dot\a} =0$, $\det(\lambda^\a_a) =0$, $\det(\lambda^{\dot\a}_a) =0$, $\lambda^\a_a \e_{\a\b} \lambda^\b=0$ and $\lambda^{\dot\a}_a \e_{\dot\a\dot\b} \lambda^{\dot\b}=0$. The first set of constraints implies that we can solve 3 parameters in terms of the rest and we get a consistency condition $\det(\lambda^\a_a) \det(\lambda^{\dot\a}_a) =0$. This is solved by imposing the second and the third conditions. The latter also imply the existence of a solution for the forth and for the fifth constraints. Again the counting of the parameters gives 7 complex numbers. [^4]: The conditions for a supermanifold of being a super-Ricci flat are discussed in [@Grassi:2006cd].

--- abstract: 'Previous work in the literature has studied gravitational radiation in black-hole collisions at the speed of light. In particular, it had been proved that the perturbative field equations may all be reduced to equations in only two independent variables, by virtue of a conformal symmetry at each order in perturbation theory. The Green function for the perturbative field equations is here analyzed by studying the corresponding second-order hyperbolic operator with variable coefficients, instead of using the reduction method from the retarded flat-space Green function in four dimensions. After reduction to canonical form of this hyperbolic operator, the integral representation of the solution in terms of the Riemann function is obtained. The Riemann function solves a characteristic initial-value problem for which analytic formulae leading to the numerical solution are derived.' address: | ${ }^{1}$Istituto Nazionale di Fisica Nucleare, Sezione di Napoli, Complesso Universitario di Monte S. Angelo, Via Cintia, Edificio N’, 80126 Napoli, Italy\ ${ }^{2}$Università di Napoli Federico II, Dipartimento di Scienze Fisiche, Complesso Universitario di Monte S. Angelo, Via Cintia, Edificio N’, 80126 Napoli, Italy author: - 'Giampiero Esposito$^{1,2}$' title: 'An application of Green-function methods to gravitational radiation theory' --- \#1 [^1] Introduction ============ In the theory of gravitational radiation, it remains true what R. Sachs stressed in his 1963 Les Houches lectures, i.e. that we understand the following three main features [@sach64]: 0.3cm (i) We can give a description of radiation at large distances from its sources in an asymptotically flat universe; this description is geometrically elegant and sufficiently detailed to analyze all conceptual experiments concerning the behaviour of test particles or test absorbers in the far field. 0.3cm (ii) How the exact theory relates the far field to the near field and the sources or non-gravitational fields. 0.3cm (iii) We have approximation methods that make it possible to obtain numerical results for the amount of radiation emitted in a particular situation, or for scattering cross-sections, etc.; these approximation methods do not have a profound geometrical nature, but are very important in comparing theory with the experiment, in case the latter succeeds in finding observational evidence in favour of gravitational waves. 0.3cm Within this framework, it is often desirable to use Green-function methods, since the construction of suitable inverses of differential operators lies still at the very heart of many profound properties in classical and quantum field theory. For example, the theory of small disturbances in local field theory can only be built if suitable invertible operators are considered [@dewi65]. In a functional-integral formulation, these correspond to the gauge-field and ghost operators, respectively [@dewi84; @espo01a]. Moreover, the Peierls bracket on the space of physical observables, which is a Poisson bracket preserving the invariance under the full infinite-dimensional symmetry group of the theory, is obtained from the advanced and retarded Green functions of the theory via the supercommutator function [@dewi65; @dewi84; @espo01a; @cart00] and leads possibly to a deeper approach to quantization. Last, but not least, a perturbation approach to classical general relativity relies heavily on a careful construction of Green functions of operators of hyperbolic [@deat92a; @deat92b; @deat92c; @deat96] and elliptic [@espo00a; @espo00b] type. In particular, following [@deat92a; @deat92b; @deat92c; @deat96], we shall be concerned with the axisymmetric collision of two black holes travelling at the speed of light, each described in the centre-of-mass frame before the collision by an impulsive plane-fronted shock wave with energy $\mu$. One then passes to a new frame to which a large Lorentz boost is applied. There the energy $\nu=\mu {\rm e}^{\alpha}$ of the incoming shock $1$ obeys $\nu >> \lambda$, where $\lambda=\mu {\rm e}^{-\alpha}$ is the energy of the incoming shock $2$ and ${\rm e}^{\alpha} \equiv \sqrt{{1+\beta}\over {1-\beta}}$ ($\beta$ being the usual relativistic parameter). In the boosted frame, to the future of the strong shock $1$, the metric can be expanded in the form [@deat92b; @deat96] $$g_{ab} \sim \nu^{2} \left[\eta_{ab}+\sum_{i=1}^{\infty} \left({\lambda \over \nu}\right)^{i}h_{ab}^{(i)}\right], \label{(1.1)}$$ where $\eta_{ab}$ is the standard notation for the Minkowski metric. The task of solving the Einstein field equations becomes then a problem in singular perturbation theory, having to find $h_{ab}^{(1)},h_{ab}^{(2)},...$ by solving the linearized field equations at first, second, ... order respectively in ${\lambda \over \nu}$, once that characteristic initial data are given just to the future of the strong shock $1$. The perturbation series (1.1) is physically relevant because, on boosting back to the centre-of-mass frame, it is found to give an accurate description of space-time geometry where gravitational radiation propagates at small angles away from the forward symmetry axis ${\hat \theta}=0$. The news function $c_{0}$ (see appendix), which describes gravitational radiation arriving at future null infinity in the centre-of-mass frame, is expected to have the convergent series expansion [@deat92b; @deat96] $$c_{0}({\hat \tau},{\hat \theta})=\sum_{n=0}^{\infty} a_{2n}({\hat \tau}/\mu)(\sin {\hat \theta})^{2n}, \label{(1.2)}$$ with $\hat \tau$ a suitable retarded time coordinate, and $\mu$ the energy of each incoming black hole in the centre-of-mass frame. In [@deat92b; @deat96] a very useful analytic expression of $a_{2}({\hat \tau}/\mu)$ was derived, exploiting the property that perturbative field equations may all be reduced to equations in only two independent variables, by virtue of a remarkable conformal symmetry at each order in perturbation theory. The Green function for perturbative field equations was then found by reduction from the retarded flat-space Green function in four dimensions. However, a [*direct*]{} approach to the evaluation of Green functions appears both desirable and helpful in general, and it has been our aim to pursue such a line of investigation. For this purpose, following hereafter our work in [@espo01], reduction to two dimensions with the associated hyperbolic operator is studied again in section 2. Section 3 performs reduction to canonical form with the associated Riemann function. Equations for the Goursat problem obeyed by the Riemann function are derived in section 4, while the corresponding numerical algorithm is discussed in section 5. Some backgound material is described in the appendix. Reduction to two dimensions and the associated operator ======================================================= As is well known from the work in [@deat92b] and [@deat96], the field equations for the first-order correction $h_{ab}^{(1)}$ in the expansion (1.1) are particular cases of the general system given by the flat-space wave equation (here $u \equiv {1\over \sqrt{2}}(z+t), v \equiv {1\over \sqrt{2}}(z-t)$) $$\cstok{\ }\psi=2{\partial^{2}\psi \over \partial u \partial v} +{1\over \rho}{\partial \over \partial \rho} \left( \rho {\partial \psi \over \partial \rho}\right) +{1\over \rho^{2}}{\partial^{2}\psi \over \partial \phi^{2}}=0, \label{(2.1)}$$ supplemented by the boundary condition $$\psi(u=0)={\rm e}^{im \phi}\rho^{-n}f[8 \log(v \rho)-\sqrt{2}v], \label{(2.2)}$$ $$f(x)=0 \; \; \forall x < 0. \label{(2.3)}$$ Moreover, $\psi$ should be of the form ${\rm e}^{i m \phi}\rho^{-n}\chi(q,r)$ for $u \geq 0$, where $$q \equiv u \rho^{-2}, \label{(2.4)}$$ $$r \equiv 8 \log(\nu \rho)-\sqrt{2}v. \label{(2.5)}$$ For the homogeneous wave equation (2.1) there is no advantage in eliminating $\rho$ and $\phi$ from the differential equation. However, the higher-order metric perturbations turn out to obey inhomogeneous flat-space wave equations of the form $$\cstok{\ }\psi=S, \label{(2.6)}$$ where $S$ is a source term equal to ${\rm e}^{im \phi}\rho^{-(n+2)}H(q,r)$. This leads to the following equation for $\chi \equiv {\rm e}^{-im \phi}\rho^{n}\psi$: $${\cal L}_{m,n}\chi(q,r)=H(q,r), \label{(2.7)}$$ where ${\cal L}_{m,n}$ is an hyperbolic operator in the independent variables $q$ and $r$, and takes the form [@deat92b; @deat96] $$\begin{aligned} {\cal L}_{m,n}&=&-(2\sqrt{2}+32q){\partial^{2}\over \partial q \partial r} +4q^{2}{\partial^{2}\over \partial q^{2}} +64{\partial^{2}\over \partial r^{2}} \nonumber \\ &+& 4(n+1)q{\partial \over \partial q}-16n{\partial \over \partial r} +n^{2}-m^{2}. \label{(2.8)}\end{aligned}$$ The proof of hyperbolicity of ${\cal L}_{m,n}$, with the associated normal hyperbolic form, can be found in section 3 of [@deat92b], and in [@deat96]. The advantage of studying Eq. (2.7) is twofold: to evaluate the solution at some space-time point one has simply to integrate the product of $H$ and the Green function $G_{m,n}$ of ${\cal L}_{m,n}$: $$\chi(q,r)=\int G_{m,n}(q,r;q_{0},r_{0})H(q_{0},r_{0})dq_{0}dr_{0}, \label{(2.9)}$$ and the resulting numerical calculation of the solution is now feasible [@deat92c; @deat96]. If one defines the variables $$X \equiv \log(q)+{r\over 4}, \; Y \equiv \log(q)-{r\over 4}, \label{(2.10)}$$ the operator ${\cal L}_{m,n}$ is turned into $$T_{m,n} \equiv 16{\partial^{2}\over \partial Y^{2}} +8n{\partial \over \partial Y}+n^{2}-m^{2} -{1\over \sqrt{2}}{\rm e}^{-(X+Y)/2} \left({\partial^{2}\over \partial X^{2}} -{\partial^{2}\over \partial Y^{2}}\right). \label{(2.11)}$$ The operator $T_{m,n}$ is the ‘sum’ of an elliptic operator in the $Y$ variable and a two-dimensional wave operator ‘weighted’ with the exponential ${\rm e}^{-(X+Y)/2}$, which is the main source of technical complications in these variables. Reduction to canonical form and the Riemann function ==================================================== It is therefore more convenient, in our general analysis, to reduce first Eq. (2.7) to canonical form, and then find an integral representation of the solution. Reduction to canonical form means that new coordinates $x=x(q,r)$ and $y=y(q,r)$ are introduced such that the coefficients of ${\partial^{2}\over \partial x^{2}}$ and ${\partial^{2}\over \partial y^{2}}$ vanish. As is shown in [@deat92b; @deat96], this is achieved if $${\partial x \over \partial r}={\partial y \over \partial r}=1, \label{(3.1)}$$ $${\partial x \over \partial q}={1+8q \sqrt{2}+\sqrt{1+16q\sqrt{2}} \over 2\sqrt{2}q^{2}}, \label{(3.2)}$$ $${\partial y \over \partial q}={1+8q \sqrt{2}-\sqrt{1+16q\sqrt{2}} \over 2\sqrt{2}q^{2}}. \label{(3.3)}$$ The resulting formulae are considerably simplified if one defines $$t \equiv \sqrt{1+16q \sqrt{2}}=t(x,y). \label{(3.4)}$$ The dependence of $t$ on $x$ and $y$ is obtained implicitly by solving the system [@deat92b; @deat96] $$x=r+\log \left({{t-1}\over 2}\right)-{8\over (t-1)}-4, \label{(3.5)}$$ $$y=r+\log \left({{t+1}\over 2}\right)+{8\over (t+1)}-4. \label{(3.6)}$$ This leads to the equation $$\log{(t-1)\over (t+1)}-{2t \over (t^{2}-1)} ={(x-y)\over 8}, \label{(3.7)}$$ which can be cast in the form $${(t-1)\over (t+1)}{\rm e}^{2t \over (1-t^{2})} ={\rm e}^{(x-y) \over 8}. \label{(3.8)}$$ This suggests defining $$w \equiv {(t-1)\over (t+1)}, \label{(3.9)}$$ so that one first has to solve the transcendental equation $$w{\rm e}^{(w^{2}-1)\over 2w}={\rm e}^{(x-y)\over 8}, \label{(3.10)}$$ to obtain $w=w(x-y)$, from which one gets $$t={(1+w)\over (1-w)}=t(x-y). \label{(3.11)}$$ On denoting by $g(w)$ the left-hand side of Eq. (3.10), one finds that, in the plane $(w,g(w))$, the right-hand side of Eq. (3.10) is a line parallel to the $w$-axis, which intersects $g(w)$ at no more than one point for each value of $x-y$. For example, when $w=1$, $g(w)$ intersects the line taking the constant value $1$, for which $x-y=0$. The function $$g:w \rightarrow g(w)=w{\rm e}^{(w^{2}-1)\over 2w}$$ is asymmetric and has the limiting behaviour described by $$\lim_{w \to 0^{-}}g(w)=-\infty, \; \; \lim_{w \to 0^{+}}g(w)=0, \label{(3.12)}$$ $$\lim_{w \to -\infty}g(w)=0, \; \; \lim_{w \to +\infty}g(w)=\infty . \label{(3.13)}$$ Thus, in the lower half-plane, $g$ has an horizontal asymptote given by the $w$-axis, and a vertical asymptote given by the line $w=0$, while it has no asymptotes in the upper half-plane, since $$\lim_{w \to \infty}{g(w)\over w}=\infty$$ in addition to (3.13). The first derivative of $g$ reads as $$g'(w)={(w+1)^{2}\over 2w}{\rm e}^{(w^{2}-1)\over 2w}. \label{(3.14)}$$ One therefore has $g'(w)>0$ for all $w>0$, and $g'(w)<0$ for all $w \in (-\infty,0) - \left \{ -1 \right \}$, and $g$ is monotonically decreasing for negative $w$ and monotonically increasing for positive $w$. The point $w=-1$, at which $g'(w)$ vanishes, is neither a maximum nor a minimum point, because $$g''(w)=\left({1\over 4w^{3}}+{1\over 2w}+1+{w\over 4}\right) {\rm e}^{(w^{2}-1)\over 2w}, \label{(3.15)}$$ $$g'''(w)=\left({1\over 8w^{5}}-{3\over 4w^{4}} +{3\over 8 w^{3}}+{3\over 8w}+{3\over 4}+{w\over 8}\right) {\rm e}^{(w^{2}-1)\over 2w}. \label{(3.16)}$$ These formulae imply that $g''(-1)=0$ but $g'''(-1)=-1 \not = 0$, and hence $w=-1$ yields a flex of $g(w)$ (see Fig. 1). In the $(x,y)$ variables, the operator ${\cal L}_{m,n}$ therefore reads $${\cal L}_{m,n}=f(x,y){\partial^{2}\over \partial x \partial y} +g(x,y){\partial \over \partial x}+h(x,y){\partial \over \partial y} +n^{2}-m^{2}, \label{(3.17)}$$ where, exploiting the formulae $${\partial x \over \partial q}={64 \sqrt{2}\over (t-1)^{2}}, \label{(3.18)}$$ $${\partial y \over \partial q}={64 \sqrt{2}\over (t+1)^{2}}, \label{(3.19)}$$ one finds $$\begin{aligned} f(x,y)&=& -(2\sqrt{2}+32q)\left({\partial x \over \partial q} +{\partial y \over \partial q}\right) +8q^{2}{\partial x \over \partial q}{\partial y \over \partial q} +128 \nonumber \\ &=& 256 \left[1-{2t^{2}(t^{2}+1)\over (t-1)^{2}(t+1)^{2}}\right], \label{(3.20)}\end{aligned}$$ $$g(x,y)=4(n+1)q{\partial x \over \partial q}-16n =16 \left[1+{2(n+1)\over (t-1)}\right], \label{(3.21)}$$ $$h(x,y)=4(n+1)q{\partial y \over \partial q}-16n =16 \left[1-{2(n+1)\over (t+1)}\right]. \label{(3.22)}$$ The resulting canonical form of Eq. (2.7) is $$\begin{aligned} L[\chi]&=& \left({\partial^{2}\over \partial x \partial y} +a(x,y){\partial \over \partial x} +b(x,y){\partial \over \partial y}+c(x,y) \right)\chi(x,y) \nonumber \\ &=& {\widetilde H}(x,y) \label{(3.23)}\end{aligned}$$ where $$a(x,y) \equiv {g(x,y)\over f(x,y)} ={1\over 16}{(1-t)(t+1)^{2}(2n+1+t)\over (t^{4}+4t^{2}-1)}, \label{(3.24)}$$ $$b(x,y) \equiv {h(x,y)\over f(x,y)} ={1\over 16}{(t+1)(t-1)^{2}(2n+1-t)\over (t^{4}+4t^{2}-1)}, \label{(3.25)}$$ $$c(x,y) \equiv {n^{2}-m^{2}\over f(x,y)} ={(m^{2}-n^{2})\over 256} {(t-1)^{2}(t+1)^{2}\over (t^{4}+4t^{2}-1)}, \label{(3.26)}$$ $${\widetilde H}(x,y) \equiv {H(x,y)\over f(x,y)} =-{H(x,y)\over 256} {(t-1)^{2}(t+1)^{2}\over (t^{4}+4t^{2}-1)}. \label{(3.27)}$$ Note that $a(-t)=b(t), b(-t)=a(t), c(-t)=c(t), {\widetilde H}(-t)={\widetilde H}(t)$. For an hyperbolic equation in the form (3.23), we can use the Riemann integral representation of the solution. For this purpose, recall from [@cour61] that, on denoting by $L^{\dagger}$ the adjoint of the operator $L$ in (3.23), which acts according to $$L^{\dagger}[\chi]=\chi_{xy}-(a\chi)_{x}-(b \chi)_{y}+c \chi, \label{(3.28)}$$ one has to find a ‘function’ $R(x,y;\xi,\eta)$ (actually a kernel) subject to the following conditions ($(\xi,\eta)$ being the coordinates of a point $P$ such that characteristics through it intersect a curve $C$ at points $A$ and $B$, $AP$ being a segment with constant $y$, and $BP$ being a segment with constant $x$, as is shown in Fig. 2): 0.3cm (i) As a function of $x$ and $y$, $R$ satisfies the adjoint equation $$L^{\dagger}_{(x,y)}[R]=0, \label{(3.29)}$$ (ii) $R_{x}=bR$ on $AP$, i.e. $$R_{x}(x,y;\xi,\eta)=b(x,\eta)R(x,y;\xi,\eta) \; {\rm on} \; y=\eta, \label{(3.30)}$$ and $R_{y}=aR$ on $BP$, i.e. $$R_{y}(x,y;\xi,\eta)=a(\xi,y)R(x,y;\xi,\eta) \; {\rm on} \; x=\xi, \label{(3.31)}$$ (iii) $R$ equals $1$ at $P$, i.e. $$R(\xi,\eta;\xi,\eta)=1. \label{(3.32)}$$ It is then possible to express the solution of Eq. (3.23) in the form $$\begin{aligned} \chi(P)&=& {1\over 2}[\chi(A)R(A)+\chi(B)R(B)] +\int_{AB}\biggr( \left[{R\over 2}\chi_{x} +\left(bR-{1\over 2}R_{x}\right)\chi \right]dx \nonumber \\ & - & \left[{R\over 2}\chi_{y} +\left(aR-{1\over 2}R_{y}\right)\chi \right]dy \biggr) + \int \int_{\Omega}R(x,y;\xi,\eta){\widetilde H}(x,y)dx dy, \label{(3.33)}\end{aligned}$$ where $\Omega$ is a domain with boundary. Note that Eqs. (3.30) and (3.31) are ordinary differential equations for the Riemann function $R(x,y;\xi,\eta)$ along the characteristics parallel to the coordinate axes. By virtue of (3.32), their integration yields $$R(x,\eta;\xi,\eta)={\rm exp} \int_{\xi}^{x}b(\lambda,\eta)d\lambda, \label{(3.34)}$$ $$R(\xi,y;\xi,\eta)={\rm exp} \int_{\eta}^{y}a(\lambda,\xi)d\lambda, \label{(3.35)}$$ which are the values of $R$ along the characteristics through $P$. Equation (3.33) yields instead the solution of Eq. (3.23) for arbitrary initial values given along an arbitrary non-characteristic curve $C$, by means of a solution $R$ of the adjoint equation (3.29) which depends on $x,y$ and two parameters $\xi,\eta$. Unlike $\chi$, the Riemann function $R$ solves a characteristic initial-value problem. Goursat problem for the Riemann function ======================================== By fully exploiting the reduction to canonical form of Eq. (2.7) we have considered novel features with respect to the analysis in [@deat92b; @deat96], because the Riemann formula (3.33) also contains the integral along the piece of curve $C$ from $A$ to $B$, and the term ${1\over 2}[\chi(A)R(A)+\chi(B)R(B)]$. This representation of the solution might be more appropriate for the numerical purposes considered in [@deat92c], but the task of finding the Riemann function $R$ remains extremely difficult. One can however use approximate methods for solving Eq. (3.29). For this purpose, we first point out that, by virtue of Eq. (3.28), Eq. (3.29) is a canonical hyperbolic equation of the form $$\left({\partial^{2}\over \partial x \partial y} +A{\partial \over \partial x}+B{\partial \over \partial y} +C \right)R(x,y;\xi,\eta)=0, \label{(4.1)}$$ where $$A \equiv -a, \label{(4.2)}$$ $$B \equiv -b, \label{(4.3)}$$ $$C \equiv c-a_{x}-b_{y}. \label{(4.4)}$$ Thus, on defining $$U \equiv R, \label{(4.5)}$$ $$V \equiv R_{x}+BR, \label{(4.6)}$$ the equation (4.1) for the Riemann function is equivalent to the hyperbolic canonical system [@cour61] $$U_{x}=f_{1}(x,y)U+f_{2}(x,y)V, \label{(4.7)}$$ $$V_{y}=g_{1}(x,y)U+g_{2}(x,y)V, \label{(4.8)}$$ where $$f_{1} \equiv -B=b, \label{(4.9)}$$ $$f_{2}=1, \label{(4.10)}$$ $$g_{1} \equiv AB-C+B_{y}=ab-c+a_{x}, \label{(4.11)}$$ $$g_{2} \equiv -A=a. \label{(4.12)}$$ For the system described by Eqs. (4.7) and (4.8) with boundary data (3.34) and (3.35) an existence and uniqueness theorem holds (see [@cour61] for the Lipschitz conditions on boundary data), and we can therefore exploit the finite differences method to find approximate solutions for the Riemann function $R(x,y;\xi,\eta)$, and eventually $\chi(P)$ with the help of the integral representation (3.33). Concluding remarks ================== The inverses of hyperbolic operators [@lera53] and the Cauchy problem for hyperbolic equations with polynomial coefficients [@lera56] have always been the object of intensive investigation in the mathematical literature. We have here considered the application of such issues to axisymmetric black hole collisions at the speed of light, relying on the work in [@deat92a; @deat92b; @deat92c; @deat96]. We have pointed out that, for the inhomogeneous equations (2.7) occurring in the perturbative analysis, the task of inverting the operator (2.8) can be accomplished with the help of the Riemann integral representation (3.33), after solving Eq. (4.1) for the Riemann function. One has then to solve a characteristic initial-value problem for a homogeneous hyperbolic equation in canonical form in two independent variables, for which we have developed formulae to be used for the numerical solution with the help of a finite differences scheme. For this purpose one studies the canonical system (cf (4.7) and (4.8)) $$U_{x}=F(x,y,U,V), \label{(5.1)}$$ $$V_{y}=G(x,y,U,V), \label{(5.2)}$$ in the rectangle ${\cal R} \equiv \left \{ x,y: x \in [x_{0},x_{0}+a], y \in [y_{0},y_{0}+b] \right \}$ with known values of $U$ on the vertical side $AD$ where $x=x_{0}$, and known values of $V$ on the horizontal side $AB$ where $y=y_{0}$. The segments $AB$ and $AD$ are then divided into $m$ and $n$ equal parts, respectively. On setting ${a\over m} \equiv h$ and ${b \over n} \equiv k$, the original differential equations become equations relating values of $U$ and $V$ at three intersection points of the resulting lattice, i.e. $${U(P_{r,s+1})-U(P_{rs})\over h}=F, \label{(5.3)}$$ $${V(P_{r+1,s})-V(P_{rs})\over k}=G. \label{(5.4)}$$ It is now convenient to set $U_{rs} \equiv U(P_{rs}), V_{rs} \equiv V(P_{rs})$, so that these equations read as $$U_{r,s+1}=U_{rs}+hF(P_{rs},U_{rs},V_{rs}), \label{(5.5)}$$ $$V_{r+1,s}=V_{rs}+kG(P_{rs},U_{rs},V_{rs}). \label{(5.6)}$$ Thus, if both $U$ and $V$ are known at $P_{rs}$, one can evaluate $U$ at $P_{r,s+1}$ and $V$ at $P_{r+1,s}$. The evaluation at subsequent intersection points of the lattice goes on along horizontal or vertical segments. In the former case, the resulting algorithm is $$U_{rs}=U_{r0}+h \sum_{i=1}^{s-1}F(P_{ri},U_{ri},V_{ri}), \label{(5.7)}$$ $$V_{rs}=V_{r-1,s}+kG(P_{r-1,s},U_{r-1,s},V_{r-1,s}), \label{(5.8)}$$ while in the latter case one obtains the algorithm expressed by the equations $$V_{rs}=V_{0s}+\sum_{i=1}^{r-1}G(P_{is},U_{is},V_{is}), \label{(5.9)}$$ $$U_{rs}=U_{r,s-1}+hF(P_{r,s-1},U_{r,s-1},V_{r,s-1}). \label{(5.10)}$$ Stability of such solutions is closely linked with the geometry of the associated characteristics, and the criteria to be fulfilled are studied in section 13.2 of [@gara64] (stability depends crucially on whether or not ${h\over k} \leq 1$). To sum up, one solves numerically Eq. (3.10) for $w=w(x,y)=w(x-y)$, from which one gets $t(x-y)$ with the help of (3.11), which is a fractional linear transformation. This yields $a,b,c$ and $\widetilde H$ as functions of $(x,y)$ according to (3.24)–(3.27), and hence $A,B$ and $C$ in the equation for the Riemann function are obtained according to (4.2)–(4.4), where derivatives with respect to $x$ and $y$ are evaluated numerically. Eventually, the system given by (4.7) and (4.8) is solved according to the finite-differences scheme of the present section, with $$F=f_{1}U+f_{2}V=f_{1}R+f_{2}(R_{x}+BR), \label{(5.11)}$$ $$G=g_{1}U+g_{2}V=g_{1}R+g_{2}(R_{x}+BR). \label{(5.12)}$$ Once the Riemann function $R=U$ is obtained with the desired accuracy, numerical evaluation of the integral (3.33) yields $\chi(P)$, and $\chi(q,r)$ is obtained upon using Eqs. (3.5) and (3.6) for the characteristic coordinates. Our steps are conceptually desirable since they rely on well established techniques for the solution of hyperbolic equations in two independent variables [@cour61; @gara64], and provide a viable alternative to the numerical analysis performed in [@deat92c], because all functions should be evaluated numerically. Our method is not obviously more powerful than the one used in [@deat92a; @deat92b; @deat92c; @deat96], but is well suited for a systematic and lengthy numerical analysis, while its analytic side provides an interesting alternative for the evaluation of Green functions both in black hole physics and in other problems where hyperbolic operators with variable coefficients might occur. This task remains very important because a strong production of gravitational radiation is mainly expected in the extreme events studied in [@deat92a; @deat92b; @deat92c; @deat96] and which motivated our paper. Any viable way of looking at mathematical and numerical aspects of the problem is therefore of physical interest for research planned in the years to come [@alle00]. The characteristic initial-value problem ======================================== In our expository article, we find it appropriate to include some background material, following, for example, the presentation in section IV of [@sach64]. We therefore consider some four-dimensional region of space-time and choose in it a set of null hypersurfaces $u={\rm constant}$; the corresponding ray congruence with tangent vector $k_{a}=u_{,a}$ is assumed to have expansion $\rho \equiv {1\over 2}k_{\; ;a}^{a} \not=0$, which can always be arranged in a space-time patch, whereas outside of some patch the rays start to cross and hence our construction breaks down globally. On completing the $k^{a}$ direction to a quasi-normal tetrad $(k,m,t)$, one finds the following split of the vacuum Einstein equations with Einstein tensor $G_{ab}$: Main equations (6 equations) $$k^{a}G_{ab}=0, \; G_{ab}t^{a}t^{b}=0, \label{(A1)}$$ trivial equation $$G_{ab}t^{a}{\overline t}^{b}=0, \label{(A2)}$$ and 3 supplementary conditions $$G_{ab}m^{a}t^{b}=0, \; G_{ab}m^{a}m^{b}=0, \label{(A3)}$$ where a single complex equation has been counted as two real equations. Remarkably, if the main equations hold everywhere, then the trivial equation holds everywhere and the supplementary conditions hold everywhere if they hold at one point on each ray. The fulfillment of the trivial equation is proved by writing, from the vacuum Einstein equations, that $$G_{\; b;a}^{a}=0, \label{(A4)}$$ and then exploiting the main equations (A1) jointly with the split of $k_{a;b}$ as given in [@sach64]: $$\begin{aligned} k_{a;b}&=& zt_{a}{\overline t}_{b}+\sigma t_{a}t_{b} +\Omega t_{a}k_{b} +\zeta k_{a}t_{b}+{\rm c.c.} \nonumber \\ &+& \xi k_{a}k_{b}. \label{(A5)}\end{aligned}$$ Hence one gets $$0=k^{a}G_{\; a;b}^{b}=-k_{a;b}G^{ab}=-2\rho G_{ab} t^{a}{\overline t}^{b}. \label{(A6)}$$ By hypothesis the expansion $\rho$ does not vanish, so that the trivial equation is, indeed, identically satisfied. The fulfillment of (A3) everywhere is proved along similar lines. Thus, one can again integrate the main equations (A1) first and worry about the supplementary conditions (A3) later. Choose now the coordinate $x^{0}$ as the retarded time: $x^{0}=u$. Let $r=x^{1}$ be a luminosity distance along the rays; let $x^{\alpha}$ (with $\alpha=2,3$) be any other pair of coordinates constant along the rays. The line element in these coordinates takes therefore the form (no confusion should arise with the $\beta$ of section 1) $$ds^{2}=W\; du^{2}+2{\rm e}^{2\beta} \; dudr -r^{2}h_{\alpha \gamma}\Bigr(dx^{\alpha}-U^{\alpha}du \Bigr) \Bigr(dx^{\gamma}-U^{\gamma} \; du \Bigr), \label{(A7)}$$ where $W,\beta,h_{\alpha \gamma},U^{\alpha}$ depend on the $x^{\alpha}$ coordinates. Since $r$ is a luminosity distance, the determinant of $h_{\nu \mu}$ is independent of $r$. Bearing in mind that the luminosity distance is defined only up to a factor constant along each ray, one can demand without loss of generality that (here $\theta \equiv x^{2}, \; \phi \equiv x^{3}$) $$\begin{aligned} 2h_{\mu \nu}dx^{\mu}dx^{\nu}&=& \Bigr({\rm e}^{2\gamma} +{\rm e}^{2\delta}\Bigr)d\theta^{2}+4 \sinh(\gamma-\delta) d\theta d\phi \; \sin \theta \nonumber \\ &+& \sin^{2}\theta \Bigr({\rm e}^{-2\gamma} +{\rm e}^{-2\delta}\Bigr)d\phi^{2}. \label{(A8)}\end{aligned}$$ The metric corresponding to the line element (A7) contains only six unknown functions of four variables, and our coordinate system is ‘rigid’ enough for our purposes [@sach64]. One can either analyze the field in the neighbourhood of some point, or the field near infinity in an asymptotically flat space-time. Indeed, if in Minkowski space-time one uses a retarded time $u=t-r$ and spherical coordinates $r,\theta,\phi$ one finds for the line element $$ds^{2}=du^{2}+2dudr-r^{2}\Bigr(d\theta^{2}+\sin^{2}\theta \; d\phi^{2}\Bigr), \label{(A9)}$$ hence one is led to require that, if asymptotic flatness holds, $$\lim_{r \to \infty}W=1, \; \lim_{r \to \infty}(rU^{\alpha})=\lim_{r \to \infty}\beta =\lim_{r \to \infty}\delta =\lim_{r \to \infty}\gamma=0, \label{(A10)}$$ where all limits are taken as $r$ approaches infinity with $u,\theta,\phi$ fixed. The second requirement in (A10), i.e. that all quantities of interest admit a power-series expansion in ${1\over r}$, e.g. $$(1-i)(\delta+i \gamma)/2={c(u,\theta,\phi)\over r} +{d(u,\theta,\phi)\over r^{2}}+{\ldots} , \label{(A11)}$$ is indeed restrictive. Such a requirement can be drastically weakened but not fully eliminated; moreover, it is closely related to an outgoing radiation condition of the Sommerfeld type. It should be stressed that all these requirements no longer hold when $r$ becomes small to the extent that rays start to cross each other. In the axially- and reflection-symmetric case considered by Bondi et al. [@bond62], one has $$\delta=\gamma, \; U^{3}=0, \; {\partial g_{ab}\over \partial \phi}=0, \label{(A12)}$$ and the $\phi$-direction is a hypersurface-orthogonal Killing direction. The line element acquires the simpler form $$ds^{2}={V{\rm e}^{2\beta} \over r}du^{2} +2{\rm e}^{2\beta}\; dudr -r^{2}\Bigr[{\rm e}^{2\gamma} (d\theta-U \; du)^{2}+{\rm e}^{-2\gamma} \sin^{2}\theta \; d\phi^{2}\Bigr], \label{(A13)}$$ where the peculiar form of the first coefficient is chosen to simplify the resulting calculations. Interestingly, two main equations are found to be identically satisfied by virtue of axial symmetry, whereas the other four turn out to be linear combinations of [@bond62] $$R_{11}=-{4\over r}\Bigr(\beta_{1}-{1\over 2}r\gamma_{1}^{2}\Bigr)=0, \label{(A14)}$$ $$-2r^{2}R_{12}= {\Bigr[r^{4}{\rm e}^{2(\gamma-\beta)}U_{1}\Bigr]}_{1} -2r^{2}\biggr(-\gamma_{12}+2\gamma_{1}\gamma_{2}-2\gamma_{1} \cot \theta +\beta_{12}-2{\beta_{2}\over r}\biggr)=0, \label{(A15)}$$ $$\begin{aligned} \; & \; & R_{22}{\rm e}^{2(\beta-\gamma)}-r^{2}R_{3}^{3} {\rm e}^{2\beta}=2V_{1}+{1\over 2}r^{4} {\rm e}^{2(\gamma-\beta)}U_{1}^{2} \nonumber \\ &-& r^{2}U_{12}-4rU_{2}-r^{2}U_{1} \cot \theta -4r U \cot \theta \nonumber \\ &+& 2 {\rm e}^{2(\beta-\gamma)} \Bigr[\beta_{22}+\beta_{2}^{2}-1-(3\gamma_{2}-\beta_{2}) \cot \theta -\gamma_{22} +2\gamma_{2}(\gamma_{2}-\beta_{2})\Bigr]=0, \label{(A16)}\end{aligned}$$ $$\begin{aligned} \; & \; & -r^{2}R_{3}^{3}{\rm e}^{2\beta}= 2r(r\gamma)_{01}+(1-r \gamma_{1})V_{1} -(r\gamma_{11}+\gamma_{1})V-r(1-r \gamma_{1})U_{2} \nonumber \\ &-& r^{2}(\cot \theta-\gamma_{2})U_{1} +r(2r \gamma_{12}+2\gamma_{2}+r \gamma_{1}\cot \theta -3 \cot \theta)U \nonumber \\ &+& {\rm e}^{2(\beta-\gamma)}\Bigr[-1 -(3 \gamma_{2}-2 \beta_{2})\cot \theta -\gamma_{22} +2 \gamma_{2} (\gamma_{2}-\beta_{2})\Bigr]=0. \label{(A17)}\end{aligned}$$ Equations (A14)–(A16) are called [*hypersurface equations*]{} because they contain no $u$ derivatives, while Eq. (A17) is called the [*standard equation*]{}. Now if $\gamma$ is given for one value of $u$, Eq. (A14) and the boundary conditions (A10) determine $\beta$ uniquely. Next Eq. (A15) and the boundary conditions determine $U$ up to a function of integration $-6N(u,\theta)$ that can be added to $r^{4}{\rm e}^{2(\gamma-\beta)}U_{1}$. Equation (A16) determines $V$ up to the additive function $-2M(u,\theta)$; last, Eq. (A17) determines $\gamma_{0}$ up to an additive function ${c_{0}(u,\theta)\over r}$. One can then differentiate Eqs. (A14)–(A17) with respect to $u$ and repeat the whole procedure. To sum up, given $\gamma$ at one moment the main equations determine the future or past up to the three integration functions just mentioned. In the general case, the results are completely similar. One has to assign at one value of $u$ the two functions $\gamma$ and $\delta$. The future is then determined up to five integration functions: a term $-2{M(u,\theta,\phi)\over r}$ to be added to $W$; two functions $N^{\alpha}(u,\theta,\phi)$ which occur in the $r^{-3}$ term for $U$; and two ‘news functions’ $c_{0}(u,\theta,\phi)$, where the complex function $c$ is given by Eq. (A11). As far as the supplementary conditions are concerned, the lemma just given makes it clear that they should only involve the functions $M,N$ and $c_{0}$, while a long calculation yields [@sach64] $$M_{0}=-|c_{0}|^{2}+{1\over 2}(\sin \theta)^{-1}{\rm Re} \left \{ {\overline \nabla}\Bigr[(1/\sin \theta) {\overline \nabla}(c_{0}\sin^{2}\theta)\Bigr] \right \}, \label{(A18)}$$ $$3(N^{2}+i\sin \theta N^{3})=-\nabla M -[4c \; \cot \theta +(\nabla c)+3c \nabla] {\overline c}_{0}, \label{(A19)}$$ where $\nabla \equiv {\partial \over \partial \theta} +i(\sin \theta)^{-1}{\partial \over \partial \phi}$. The desired $M$ and $N^{\alpha}$ can be determined once that $c(u,\theta,\phi)$ and some initial values are given. In the axially symmetric case, Eqs. (A18) and (A19) take the simpler form [@bond62] $$M_{0}=-c_{0}^{2}+{1\over 2}(c_{22}+3c_{2} \; \cot \theta -2c)_{0}, \label{(A20)}$$ $$-3N_{0}=M_{2}+3c \; c_{02}+4c c_{0} \; \cot \theta +c_{0}c_{2}. \label{(A21)}$$ The functions $\gamma$ and $\delta$ given on the initial hypersurface $u={\rm constant}$, jointly with the two news functions $c_{0}$ given at $r=\infty$ describe the two transverse degrees of freedom. Moreover, one should specify $M$ and $N$ at the initial (or final) retarded time, and these three functions of two variables must be related to the longitudinal-timelike degrees of freedom of the gravitational field. In the characteristic value problem for general relativity, the independent data appear therefore in a very explicit form [@sach64]. The work of G. Esposito has been partially supported by PRIN [*SINTESI*]{}. He is grateful to Decio Cocolicchio and Sorin Dragomir for warm encouragement. IOP Publishing has kindly granted permission to republish material from the author’s paper in [@espo01]. R.K. Sachs, in [*Relativity, Groups and Topology*]{}, edited by C. DeWitt and B.S. DeWitt (Gordon & Breach, New York, 1964). B.S. DeWitt, [*Dynamical Theory of Groups and Fields*]{} (Gordon and Breach, New York, 1965). B.S. DeWitt, in [*Relativity, Groups and Topology II*]{}, edited by B.S. DeWitt and R. Stora (North–Holland, Amsterdam, 1984). G. Esposito, [*Quantum Gravity in Four Dimensions*]{} (Nova Science, New York, 2001). P. Cartier and C. DeWitt–Morette, [*J. Math. Phys.*]{} [**41**]{}, 4154 (2000). P.D. D’Eath and P.N. Payne, [*Phys. Rev.*]{} D [**46**]{}, 658 (1992). P.D. D’Eath and P.N. Payne, [*Phys. Rev.*]{} D [**46**]{}, 675 (1992). P.D. D’Eath and P.N. Payne, [*Phys. Rev.*]{} D [**46**]{}, 694 (1992). P.D. D’Eath, [*Black Holes: Gravitational Interactions*]{} (Clarendon Press, Oxford, 1996). G. Esposito and C. Stornaiolo, [*Class. Quantum Grav.*]{} [**17**]{}, 1989 (2000). G. Esposito and C. Stornaiolo, [*Found. Phys. Lett.*]{} [**13**]{}, 279 (2000). G. Esposito, [*Class. Quantum Grav.*]{} [**18**]{}, 1997 (2001). R. Courant and D. Hilbert, [*Methods of Mathematical Physics. II. Partial Differential Equations*]{} (Interscience, New York, 1961). J. Leray, [*Hyperbolic Differential Equations*]{} (Princeton University Press, Princeton, 1953). J. Leray, [*C. R. Acad. Sci. Paris*]{} [**242**]{}, 953 (1956). P.R. Garabedian, [*Partial Differential Equations*]{} (Chelsea, New York, 1964). B. Allen and A. Ottewill, [*Gen. Rel. Grav.*]{} [**32**]{}, 385 (2000). H. Bondi, M.G.J. van der Burg, A.W.K. Metzner, [*Proc. Roy. Soc. Lond.*]{} [**A 269**]{}, 21 (1962). [^1]: Submitted to [*Lecture Notes of S.I.M.*]{}, volume edited by D. Cocolicchio and S. Dragomir, with kind permission by IOP to use material in [@espo01].

--- abstract: 'We present a cross-correlation of the European Large Area [*ISO*]{} survey (ELAIS) with the ELAIS Deep X-ray Survey of the N1 and N2 fields. There are 7 [*Chandra*]{} point sources with matches in the ELAIS Final Analysis 15$\mu$m catalogue, out of a total of 28 extragalactic [*ISO*]{} sources present in the [*Chandra*]{} fields. Five of these are consistent with AGN giving an AGN fraction of $\sim 19$ per cent in the $15\mu$m flux range $0.8 - 6$ mJy. We have co-added the hard X-ray fluxes of the individually-undetected [*ISO*]{} sources and find a low significance detection consistent with star formation in the remaining population. We combine our point source cross-correlation fraction with the [*XMM-Newton*]{} observations of the Lockman Hole and [*Chandra*]{} observations of the Hubble Deep Field North to constrain source count models of the mid-infrared galaxy population. The low dust-enshrouded AGN fraction in ELAIS implied by the number of cross-identifications between the ELAIS mid-infrared sample and the [*Chandra*]{} point sources is encouraging for the use of mid-infrared surveys to constrain the cosmic star formation history, provided there are not further large undetected populations of Compton-thick AGN.' date: MNRAS in press title: 'Mid-infrared sources in the ELAIS Deep X-ray Survey' --- \[firstpage\] surveys - X-rays: general - X-rays: galaxies - galaxies: active - quasars: general Introduction ============ Enormous progress has recently been made in resolving the sources that comprise the extragalactic hard X-ray background (e.g. Mushotzky [*et al.*]{} 2000, Cowie [*et al.*]{} 2002, Moretti [*et al.*]{} 2003). Studies are now focusing on characterising the nature of these sources through multi-waveband imaging and spectroscopy. Surveys with the [*ISO*]{} satellite have recently discovered strong evolution in the mid-infrared galaxy population (e.g. Elbaz [*et al.*]{} 1999, Serjeant [*et al.*]{} 2000, Chary & Elbaz 2001, Gruppioni [*et al.*]{} 2002) which is thought to be mainly due to a strongly evolving obscured cosmic star formation history (e.g. Aussel [*et al.*]{} 1999), although there is also a contribution from dust-enshrouded AGN. Optical spectroscopic follow-ups of these samples are underway (e.g. La Franca [*et al.*]{} 2003, Gonzalez-Solares [*et al.*]{} 2004, Perez-Fournon [*et al.*]{} in preparation) but as the sources might not be optically-thin at optical wavelengths, nor emitting isotropically, it is possible that heavily dust-enshrouded populations may be mis-classified by this approach. The [*Chandra*]{} observations of the HDF-North (e.g. Hornschmeier [*et al.*]{} 2001, Alexander [*et al.*]{} 2002) placed constraints on the fraction of dust-enshrouded AGN at the faintest end of the $15\mu$m source counts. Most of the known $15\mu$m galaxies are at somewhat higher flux densities, however. Some inroads were made on the AGN fraction for brighter mid-infrared sources by Alexander et al. (2001) using [*BeppoSAX*]{} observations of the ELAIS survey (described below), but the [*BeppoSAX*]{} depth was not sufficient to detect most Compton-thin Seyfert II galaxies in the targeted ELAIS field. Fadda et al. (2002) provided the first study with a reasonably significant number of sources with hard X-ray and mid-infrared emission. They combined [*XMM-Newton*]{} observations of the Lockman Hole and [*Chandra*]{} observations of the HDF-N with coincident ISOCAM data. This paper will provide further statistics using 2 [*Chandra*]{} pointings in the ELAIS northern survey regions N1 and N2. An upcoming analysis of the [*Spitzer*]{} Wide-area Infrared Extragalactic Survey (SWIRE, Lonsdale [*et al.*]{} 2003) observations in the N1 region will provide enhanced statistics in complementary wavebands. The European Large Area [*ISO*]{} Survey, ELAIS, was the largest open time project on [*ISO*]{}, covering wavelengths from $6.7\mu$m to $175\mu$m (Oliver [*et al.*]{} 2000, Rowan-Robinson [*et al.*]{} 2004). The survey, and its follow-ups, have many ambitious aims, including tracing the cosmic star formation history to $z\sim1$ and the discovery of ultra- and hyper-luminous galaxies at high redshift. The mid-infrared source counts (Serjeant [*et al.*]{} 2000, Gruppioni [*et al.*]{} 2002, Elbaz [*et al.*]{} in preparation) and far-infrared source counts (Efstathiou [*et al.*]{} 2000) both show evidence for strong evolution, as does the far-infrared luminosity function (Serjeant [*et al.*]{} 2001). The ELAIS mid-infrared source counts cover the transition from Euclidean slope to steep evolution; by virtue of the large area and large investment in observing time, ELAIS represents by far the largest sample of galaxies from this strongly evolving mid-infrared population. A high proportion have been revealed as ultra-luminous infrared galaxies (14% of $15\mu$m galaxies with known z) including 9 hyper-luminous infrared galaxies (Morel [*et al.*]{} 2001, Rowan-Robinson [*et al.*]{} 2004). Thanks to extensive multi-wavelength coverage, the ELAIS fields have now arguably become the most well studied regions of their size, and natural targets for on-going or planned large-area surveys with the most powerful ground and space-based facilities. Further details on ELAIS multi-wavelength observations and catalogues are presented in Rowan-Robinson [*et al.*]{} (2004). In particular, ELAIS 15 observations will complement the SWIRE survey in three areas (N1, N2 and S1) by covering the 8-24um gap in [*Spitzer*]{}’s filters. In the ELAIS Deep X-ray survey we made deep [*Chandra*]{} pointings of the ELAIS N1 and N2 fields to a limiting depth of $\sim10^{-15}$ erg s$^{-1}$ cm$^{-2}$ in the $0.5-8$keV band, with a total exposure of $\sim75$ks in each field. Manners [*et al.*]{} (2003) present the data and source counts. Gonzalez-Solares [*et al.*]{} (in preparation) present follow-up imaging and optical spectroscopy of the sample and Willott [*et al.*]{} (2003) present Subaru infrared spectroscopy. Almaini [*et al.*]{} (2003) present a cross-correlation of the [*Chandra*]{} sources with the sub-mm sources of Scott [*et al.*]{} (2002) and Fox [*et al.*]{} (2002), and measure the clustering of the [*Chandra*]{} population. Here we present a cross-correlation of the [*Chandra*]{} X-ray sources with the mid-infrared sources from the ELAIS survey. Section \[sec:observations\] summarises the data acquisition in the mid-infrared and X-ray, and provides references to more exhaustive descriptions for the interested reader. Section \[sec:method\] describes the cross-correlation between the [*ISO*]{} and [*Chandra*]{} data, and section \[sec:results\] discusses the significance of the results. Observations {#sec:observations} ============ =8.5 truecm --------- ---------------------------- ------------------------ ------------ ------------------ ------------------------- ------------------- ---------------- ------------------ ID$^1$ [*Chandra*]{} position$^2$ [*ISO*]{} position$^3$ Offset$^4$ Prob.$^5$ $f_{X}$ 0.5 - 8 keV$^6$ $f_{15\mu m}$$^7$ z$^8$ Class$^9$ (J2000.0) (J2000.0) $''$ $\times 10^{-3}$ erg cm$^{-2}$ s$^{-1}$ mJy N1\_20 16:10:46.57+54:35:38.8 16:10:46.66+54:35:39.0 2.26 1.26 2.31 $\times 10^{-15}$ 1.844 0.0634 SB N1\_49 16:10:20.88+54:39:00.9 16:10:20.80+54:39:01.7 2.14 0.32 11.1 $\times 10^{-15}$ 2.918 AGN 2 N1\_69 16:10:03.18+54:36:28.4 16:10:03.01+54:36:30.1 4.75 0.13 59.5 $\times 10^{-15}$ 1.453 0.2675 AGN 1 N2\_25 16:36:55.79+40:59:10.5 16:36:55.83+40:59:09.2 1.46 0.12 10.2 $\times 10^{-15}$ 1.017 2.61 AGN 1$^\ddagger$ N2\_33 16:36:51.69+40:56:00.4 16:36:51.61+40:55:59.0 2.12 0.79 3.05 $\times 10^{-15}$ 1.732 0.4762 AGN 2$^\ddagger$ N2\_52 16:36:29.71+41:02:22.7 16:36:29.78+41:02:23.0 1.44 0.03 37.8 $\times 10^{-15}$ 1.009 0.02$^\dagger$ AGN 1 N2\_107 16:36:08.41+41:05:07.0 16:36:08.18+41:05:07.2 4.58 4.79 1.76 $\times 10^{-15}$ 8.935 0.1683 SB --------- ---------------------------- ------------------------ ------------ ------------------ ------------------------- ------------------- ---------------- ------------------ [l|ccccccc]{} & N1\_20 & N1\_49 & N1\_69 & N2\_25 & N2\_33 & N2\_52 & N2\_107\ $f_{1.4 {\rm GHz}}$$^a$ & 0.260.01 & 5.080.01 & 0.150.01 & 0.130.01$^3$ & 57.620.02$^4$ & & 3.090.02$^4$\ [*ISO*]{} $f_{175\mu{\rm m}}$$^a$ & & & & & & & 803102\ [*IRAS*]{} $f_{100\mu {\rm m}}$$^a$ & & & & & & & 898126$^1$\ [*ISO*]{} $f_{90\mu {\rm m}}$$^a$ & & & & & & & 61437\ [*IRAS*]{} $f_{60\mu {\rm m}}$$^a$ & & & & & & & 35139$^1$\ [*ISO*]{} $f_{15\mu {\rm m}}$$^a$ & 1.840.21 & 2.920.27 & 1.450.19 & 1.020.14 & 1.730.11 & 1.010.14 & 8.940.10\ [*ISO*]{} $f_{6.7\mu {\rm m}}$ $^a$ & & & & & & & 2.300.12\ K 2.2$\mu$m$^b$ & 14.570.08 & & & 19.10.1$^3$ & & 13.040.07 & 14.050.12\ H 1.65$\mu$m$^b$ & 14.870.02$^2$ & 16.410.05$^\dagger$ & 15.910.03$^2$ & 21.200.03$^2$ & 16.230.03$^2$ & 13.330.05 & 14.610.14\ J 1.25$\mu$m $^b$ & & & & & 17.210.13 & 13.990.04 & 15.520.13\ i$'$ 775nm$^b$ & 16.640.02$^2$ & & 17.860.02$^2$ & 22.550.03$^2$ & 18.150.03$^2$ & 15.010.03$^2$ & 16.560.03$^2$\ r$'$ 623nm$^b$ & 17.120.02$^2$ & & 18.460.02$^2$ & 22.950.03$^2$ & 19.390.03$^2$ & 15.680.03$^2$ & 17.250.03$^2$\ g$'$ 486nm$^b$ & 17.680.02$^2$ & 23.880.13$^\dagger$ & 19.210.02$^2$ & 22.870.03$^2$ & 20.450.03$^2$ & 16.500.03$^2$ & 18.140.03$^2$\ U 361nm $^b$ & 17.610.02$^2$ & 23.780.16$^\dagger$ & 19.110.03$^2$ & 22.940.03$^2$ & 20.620.03$^2$ & 17.330.03$^2$ & 18.990.03$^2$\ $f_{{\rm 0.5 - 2 keV}}$$^c$ & 0.80.3 & 1.00.3 & 23.01.2 & 2.10.4 & $<$ 0.5 & 14.21.0 & 1.00.3\ $f_{{\rm 2 - 8 keV}}$$^c$ & $<$ 2.4 & 22.12.7 & 39.83.5 & 14.72.1 & 5.21.4 & 26.02.8 & $<$ 3.9\ \ $\alpha_{IX}$ & 1.530.03 & 1.320.01 & 1.200.01 & 1.260.02 & 1.410.03 & 1.210.02 & 1.700.03\ HR & -0.340.24 & 0.690.09 & -0.400.04 & 0.260.11 & 0.370.19 & -0.380.05 & -0.260.22\ =17 truecm =17 truecm As part of ELAIS, ISOCAM observations were carried out covering 10.3 deg$^2$ at 15 $\mu$m. The observations were carried out in raster mode, with most of the survey area confined to three northern fields (N1: 2.67 deg$^2$, N2: 2.67 deg$^2$, N3: 0.88 deg$^2$) and one southern field (S1: 3.96 deg$^2$). Further details on the observation strategy can be found in Oliver [*et al.*]{} (2000). Data reduction of ELAIS 15 $\mu$m observations was recently completed by Vaccari [*et al.*]{} (in preparation). Data reduction was carried out using the LARI method (Lari [*et al.*]{} 2001, Lari [*et al.*]{} 2003), a new technique devised for the reduction of ISOCAM and ISOPHOT imaging data. Based on a physical model of the [*ISO*]{} detectors’ behaviour, the method is particularly suited for the reliable detection of faint sources in [*ISO*]{} surveys, allowing sensitivity to be pushed to the instrumental limits. A sample of 1056 sources (490 in N1 and 566 in N2) were detected with $S/N > 5$, spanning the 0.5 – 100 mJy flux range and thus filling the gap between the ISOCAM deep surveys (e.g. Elbaz [*et al.*]{} 1999) and the [*IRAS*]{} Faint Source Catalogue (Moshir [*et al.*]{} 1990). The ELAIS Deep X-ray Survey was centred on regions in the ELAIS N1 and N2 fields selected to have low cirrus and HI column density (see Oliver [*et al.*]{} 2000 for more details). The [*Chandra*]{} observations are described in detail elsewhere (Manners [*et al.*]{} 2003), though we summarise the main points here. The observations were taken with the [*Chandra*]{} ACIS (Advanced CCD Imaging Spectrometer) array. Integrations of $75$ks were taken in each of N1 and N2. The pointing centroids are 16:10:20.11 +54:33:22.3 in N1 and 16:36:46.99 +41:01:33.7 in N2, with an area of $16.9' \times 16.9'$ covered by the ACIS-I chips in each case, giving a total area of 571 sq. arcmin. The limiting flux levels are $4.6 \times 10^{-16}$ erg s$^{-1}$ cm$^{-2}$ in the 0.5 - 2 keV band, and $2.2 \times 10^{-15}$ erg s$^{-1}$ cm$^{-2}$ in the 2 - 8 keV band. The N1 region contains 125 [*Chandra*]{} sources from the 4 ACIS-I chips and 5 from the ACIS-S2 chip. The N2 region has 99 sources on the ACIS-I chips and 4 on the ACIS-S2 chip. Only sources from the ACIS-I chips were used for this analysis due to the poor resolution of the off-axis ACIS-S chips. Positional correlation between the [*Chandra*]{} and [*ISO*]{} sources {#sec:method} ====================================================================== In the N1 region there are 9 $15\mu$m Final Analysis ELAIS sources within the [*Chandra*]{} region, three of which are identified with stars and the remainder of which are galaxy IDs. In the N2 region there are 24 $15\mu$m sources, two of which are identified with stars and the rest with galaxies or blank fields in the r$'$ band image. The positions of the 28 extragalactic sources are plotted in Fig. \[fig:sources\] superimposed on the smoothed [*Chandra*]{} images. We performed a simple near-neighbour search to cross-correlate the extragalactic [*Chandra*]{} and [*ISO*]{} sources within the area of the [*Chandra*]{} ACIS-I chips, using a $5''$ search radius. Astrometric $1\sigma$ errors for the [*Chandra*]{} sources are $\sim 1''$ (Manners [*et al.*]{} 2003), while the nominal astrometric accuracy for the ELAIS 15 $\mu$m FA catalogue sources ranges from $\sim 0.8''$ to $2.0''$ (Vaccari [*et al.*]{} in preparation). The search radius of $5''$ was chosen as the approximate sum of the $2\sigma$ astrometric errors. Three matches were found in the N1 region and four in the N2 region (table \[tab:ids\], Fig. \[fig:sources\]). All seven matches are with high reliability [*ISO*]{} sources (5$\sigma$). To ensure the associations were real we calculated the probability of a random association between each mid-IR source and its X-ray counterpart. Following Fadda [*et al.*]{} (2002), we assume the X-ray counterpart belongs to a Poissonian distributed population so that $$P = 1 - e^{-{\rm N(}>{\rm S)} \pi d^2}$$ where P is the probability of a random association within an offset distance $d$. N($>$S) is the number density of sources with flux greater than the possible X-ray counterpart (S). We calculate this probability for each source (table \[tab:ids\]) with reference to the log(N)-log(S) relation for these regions (Manners [*et al.*]{} 2003). The chances of random associations are found to be very low. Figs. \[fig:postage1\] & \[fig:postage2\] display the cross-correlations overlaid on optical postage stamps. Results and discussion {#sec:results} ====================== Identifications {#sec:ids} --------------- Properties of the 7 cross-correlated sources are reported in table \[tab:ids\]. Multi-waveband data are reported and referenced in table \[tab:seds\]. Also reported in table \[tab:seds\] are the mid-infrared to X-ray spectral indices ($\alpha_{IX}$, described in section \[sec:IRX\]) and the X-ray hardness ratios (HR, from Manners [*et al.*]{} 2003). r$'$ band postage stamps to a depth of r$' \sim$26 are shown in Figs.  \[fig:postage1\] & \[fig:postage2\] together with the spectral energy distribution (SED) of each source. For the 6 sources with available redshifts, SEDs are displayed in the rest frame in units of luminosity assuming a cosmology with $\Omega_\Lambda = 0.73$, $\Omega_M = 0.27$, and H$_0 = 71$ km s$^{-1}$ Mpc$^{-1}$. For the source without a redshift (N1\_49), the SED is displayed in the observed frame in units of flux. In order to derive flux densities for the SEDs, magnitudes were converted using the following zero points for each photometric band: K 657 Jy; H 1020 Jy; J 1600 Jy; i$'$ 2491 Jy; r$'$ 3133 Jy; g$'$ 3876 Jy; U 1810 Jy. Flux densities at 1 keV and 5 keV were derived from the 0.5 - 2 keV and 2 - 8 keV [*Chandra*]{} bands respectively, assuming a power-law spectrum with photon index $\Gamma = 1.7$ within each band. A description follows of the properties and most probable identifications for each source: [**N1\_20 (CXOEN1 J161046.5+543538)**]{} : A complex r$'$ band morphology indicative of a recent merger. This low redshift ($z=0.0634$) object has a soft X-ray spectrum and is below the detection threshold in the 2 - 8 keV band image. The mid-infrared to X-ray spectral index is also quite steep ($\alpha_{IX}=1.51$). These characteristics are consistent with identification as a starburst galaxy (SB, table \[tab:ids\]). [**N1\_49 (CXOEN1 J161020.8+543900)**]{} : Spectroscopic identification of this object is hampered by the presence of a nearby bright star (as can be seen in the r$'$ band postage stamp). No redshift is available, however a very hard X-ray spectrum combined with a relatively flat mid-infrared to X-ray spectral index indicates likely identification with an AGN. The X-ray spectrum is consistent with an apparent absorbing column of N$_H \sim 6 \pm 2 \times 10^{22}$ cm$^{-2}$ (assuming an underlying power law of $\alpha = 0.7$ and $z = 0$). Depending on the redshift of this source the actual N$_H$ is likely to be higher with apparent absorbing column scaling as $(1+z)^{2.6}$. The mid-infrared to X-ray spectral index ($\alpha_{IX}=1.32$) is also consistent with an absorbed AGN (see section  \[sec:IRX\]). Given an AGN 2 classification in table \[tab:ids\]. [**N1\_69 (CXOEN1 J161003.1+543628)**]{} : Spectroscopic redshift of 0.2675. The flat mid-infrared to X-ray spectral index, luminosity and X-ray hardness ratio are all typical of a Seyfert galaxy (AGN 1, table \[tab:ids\]). [**N2\_25 (CXOEN2 J163655.7+405910)**]{} : Spectroscopically confirmed quasar at a redshift of 2.61 (AGN 1, table \[tab:ids\]). Its properties are extensively reported in Willott [*et al.*]{} (2003). [**N2\_33 (CXOEN2 J163651.6+405600)**]{} : Spectroscopically identified as a Seyfert 2 galaxy at a redshift of 0.4762 (AGN 2, table \[tab:ids\]). The r$'$ band image indicates an interaction is occurring with a smaller neighbour. Radio morphology suggests this to be an FR-II type AGN. [**N2\_52 (CXOEN2 J163629.7+410222)**]{} : Elliptical galaxy with no obvious point-like nuclear source. A photometric redshift of $\sim 0.02$ suggests an X-ray luminosity that is low for an active galaxy ($1\times 10^{40}$ erg s$^{-1}$ at 1 keV, $2\times 10^{40}$ erg s$^{-1}$ at 5 keV). However, mid-infrared to X-ray spectral index and X-ray hardness ratio are both consistent with an AGN (AGN 1, table \[tab:ids\]). [**N2\_107 (CXOEN2 J163608.4+410507)**]{} : Interesting r$'$ band morphology displaying merging galaxies with a double nucleus. Spectroscopic redshift of 0.1683. The X-ray spectrum is soft and of low-luminosity ($6\times 10^{40}$ erg s$^{-1}$ at 1 keV) with the source undetected in the 2 - 8 keV band image. The relatively luminous infra-red spectrum supports identification with a starburst galaxy (SB, table \[tab:ids\]). Of the 7 matched sources, 1 is a spectroscopically confirmed quasar, 1 is spectroscopically identified as a Seyfert 2, a further 3 display properties of AGN, and 2 have properties consistent with starburst galaxies. The mid-infrared to X-ray spectral indices {#sec:IRX} ------------------------------------------ =8.5 truecm Values for the mid-infrared to X-ray spectral index ($\alpha_{IX}$) are calculated using the flux density observed at 15$\mu$m and 5 keV, assuming a power-law spectrum of the form F$_\nu \propto \nu^{-\alpha_{IX}}$. The flux density at 5 keV is calculated from flux in the 2 - 8 keV [*Chandra*]{} band. For the 2 objects undetected in this band, the full 0.5 - 8 keV band is used. The mid-infrared to X-ray spectral index can be a useful indicator to distinguish between AGN and starburst galaxies (e.g. Alexander [*et al.*]{} 2001, Fadda [*et al.*]{} 2002). Starburst galaxies are found to have high values of $\alpha_{IX}$, while type-1 AGN have low values ($\alpha_{IX} < 1.2$). Type-2 AGN have values in between depending on the amount of obscuration. Fig. \[fig:xir\] plots the $\alpha_{IX}$ values for our 6 matched sources with available redshifts. The figure displays values of $\alpha_{IX}$ as a function of redshift for template SEDs compiled by Alexander [*et al.*]{} (2001). Source N2\_33, spectroscopically identified as a Seyfert 2 lies squarely within the region of the Seyfert 2 templates. N2\_25, spectroscopically identified as a QSO has a value slightly higher than the QSO templates although still lower than expected for a Seyfert 2. Of the remaining objects studied here, N1\_69 and N2\_52 lie in the region of type-1 AGN, N1\_20 and N2\_107 are consistent with starbursts or highly obscured AGN, and N1\_49 is consistent with a type-2 AGN. Constraints on models of mid-infrared source counts --------------------------------------------------- In the N1 region there are 6 extragalactic $15\mu$m Final Analysis ELAIS sources within the [*Chandra*]{} region. In the N2 region conversely there are 22 extragalactic $15\mu$m sources. The [*Chandra*]{} N2 region falls within an area of repeated [*ISO*]{} observations while the [*Chandra*]{} N1 region does not. The difference in source counts, however, is larger than expected and may indicate clustering on scales larger than the field size. This highlights the need for larger areas to be surveyed before sufficient count statistics can be gained. Here we consider our results along with number counts from the Hubble Deep Field - North (HDF-N) and the Lockman Hole (Fadda [*et al.*]{} 2002) in order to provide a comparison with recent models of mid-infrared source counts. The $15\mu$m flux range $0.8 - 6.0$ mJy is well covered by the [*ISO*]{} observations in our [*Chandra*]{} regions. In this flux range there are 26 extragalactic $15\mu$m sources in our sample. We take 5 of these objects to contain evidence of AGN from the [*Chandra*]{} and multi-waveband data. This gives an AGN fraction of 5/26 extragalactic sources, or $0.19\pm0.09$ over the given flux range, in a total area of $\sim 571$ sq. arcmin. This fraction will be a lower limit if any heavily obscured, Compton-thick AGN are present in our sample. Fadda [*et al.*]{} (2002) report AGN number counts for mid-infrared sources in the Lockman Hole and HDF-N. They find for the Lockman Hole an AGN fraction of 13/103 extragalactic sources ($0.13\pm0.04$) within the $15\mu$m flux range $0.5 - 3.0$ mJy in a total area of 218 sq. arcmin. For the HDF-N the AGN fraction is 5/42 ($0.12\pm0.05$) over a $15\mu$m flux range of $0.1 - 0.5$ mJy in a total area of 24.3 sq. arcmin. We compare these results with models for the AGN contribution to mid-infrared surveys by Pearson (2001) and King & Rowan-Robinson (2003). These models primarily attempt to model the star formation history by fitting infrared source count observations with contributions from normal galaxies, starbursts, ultra-luminous infrared galaxies (ULIRGs) and AGN. The Pearson (2001) model uses local luminosity functions and pure luminosity evolution to describe the normal galaxy, starburst and AGN populations while the ULIRG component is evolved in both density and luminosity. The $15\mu$m and $850\mu$m source counts are used to constrain the model. The best fit to the observations is obtained where the ULIRG component undergoes two major phases of evolution, rapid merging to $z \sim 1$ and an exponential evolution in luminosity to higher redshifts. This has an effect on the AGN fraction of the source counts which will be tested here. The King & Rowan-Robinson (2003) models are updated versions of models developed by Rowan-Robinson (2001) and are calculated for lambda and Einstein de Sitter cosmologies. They allow for both density and luminosity evolution in all four populations. The AGN contribution in both models is found using a $12\mu$m luminosity function from Rush, Malkan & Spinoglio (1993) which does not include Compton-thick AGN and can therefore be directly compared with these observations. Fig. \[fig:agn\_fraction\] displays the AGN fraction predicted by these models against source flux. All 3 models shown are consistent with AGN fractions reported here for the ELAIS regions and for the HDF-N. The AGN fraction in the Lockman Hole is slightly lower than the predictions of these models. =8.5 truecm =8.5 truecm We also compare the observations to modelled predictions for the cumulative AGN source counts at $15\mu$m (Fig. \[fig:iragn\]). To make a reliable comparison it was necessary to correct our AGN source counts for completeness. While the 15 $\mu$m data reduction has been completed by Vaccari [*et al.*]{} (in preparation), completeness estimates are to be published in Lari [*et al.*]{} (in preparation) and are not yet available. Therefore, values from Gruppioni [*et al.*]{} (2002) were taken as representative. By estimating the completeness at the flux level of our sources we calculate the expected number of AGN above the flux limit to be $\sim 7.7$ (ELAIS data point, Fig. \[fig:iragn\]). We can now compare a further model from Manners (2002) which assumes AGN are responsible for the entire hard X-ray background and predicts the emission of these AGN in the mid-infrared. Observations in the HDF-N agree with models by King & Rowan-Robinson (2003) and Manners (2002), while the Pearson (2001) model over-predicts the number of sources at this depth. AGN counts from the Lockman Hole are consistent with the Pearson (2001) and King & Rowan-Robinson (2003, Lambda) models, whilst the AGN counts reported here for the ELAIS survey agree well with all the model predictions. Statistical hard X-ray limits of ELAIS [*ISO*]{} sources -------------------------------------------------------- One further method of constraining the source count models is the statistical detection of [*ISO*]{} sources in the [*Chandra*]{} map. By co-adding the [*Chandra*]{} fluxes at the positions of [*ISO*]{} sources not detected individually, we can obtain a constraint on the mean hard X-ray flux and hard X-ray to mid-infrared flux ratio for the remaining population. In the N1 region there are 3 extragalactic sources undetected by [*Chandra*]{} whilst in the N2 region there are 18 (see Fig. \[fig:sources\]). In order to obtain limits on the X-ray flux for these sources it was necessary to define source regions on the [*Chandra*]{} images. These regions were positioned accurately using the astrometry obtained by cross-correlating [*Chandra*]{} and r$'$-band sources (Manners [*et al.*]{} 2003). The size of each region was defined as a circle containing 95% encircled energy for a monoenergetic [*Chandra*]{} PSF at 4.51 keV at the relevant source position. For each field, around 6 circular background regions of radii ranging from $\sim 1' - 2'$ were selected. These were chosen carefully to avoid contamination from known [*Chandra*]{} sources and to cover regions of typical effective exposure. The background counts expected in each source region were then calculated by correcting for the difference in effective exposure. This method was preferred over the use of individual local background estimates due to the slow change in background counts over the image and the better statistics gained from using the larger background regions. Counts were extracted from the hard (2 - 8 keV) and soft (0.5 - 2 keV) band [*Chandra*]{} images. Source regions were individually corrected for effective exposure, background subtracted, and then co-added to obtain an estimate of the total flux. In the hard band we obtain a total of 58 counts with 49.3 background counts expected. The Poisson probability of obtaining 58 or more counts with 49.3 expected is 0.123, equivalent to a tentative detection at a confidence level of 87.7%. This is equivalent to a mean flux of 2.2 $\pm$ 1.5 $\times 10^{-16}$ erg cm$^{-2}$ s$^{-1}$ for each of the 21 sources. In the soft band we obtain a total of 43 counts with 24.1 background counts expected. The Poisson probability of obtaining 43 or more counts with 24.1 expected is 3 $\times 10^{-4}$, equivalent to a reliable detection at a confidence level of 99.97%. These counts are equivalent to a mean flux of 6.7 $\pm$ 2.5 $\times 10^{-17}$ erg cm$^{-2}$ s$^{-1}$. In order to check the validity of the stacking analysis we repeated the procedure with the positions randomised over the areas of the [*Chandra*]{} field that are free of known sources. This was repeated 10 times for each band. The distributions of the counts obtained were consistent with a Gaussian of mean and variance equal to the expected number of background counts in each case. The stronger detection in the soft band is indicative of the higher efficiency of [*Chandra*]{} in this band and does not imply these sources have a particularly soft X-ray spectrum. The hardness ratio (defined in Manners [*et al.*]{} 2003) for the co-added regions is -0.09 $\pm$ 0.38, somewhat harder than expected for starbursts. However, the low significance of the hard X-ray detection means this must be used with caution. Seven of the unmatched sources have spectroscopic redshifts available, ranging from $z = 0.10$ to 0.24. At a typical redshift of 0.2 the mean fluxes given above would equate to luminosities of $5.4 \times 10^{39}$ at 1 keV and $1.9 \times 10^{40}$ at 5 keV. This is entirely consistent with a starburst origin for the X-ray emission. The mean $15\mu$m flux for the 21 unmatched sources is 1.489 mJy. This leads to a mean spectral index of $\alpha_{IX} = 1.68$, also consistent with purely starburst galaxies. It should be noted that ‘Compton thick’ AGN will not appear in the X-ray data and constitute an unknown fraction of the unmatched sources. The cosmic star formation history from mid-infrared flux limited samples ------------------------------------------------------------------------ Of the sources detected in field surveys by the [*ISO*]{} satellite, by far the largest fraction and largest number of moderate-redshift ($z>0.5$) systems are found in mid-infrared ISOCAM surveys, as opposed to the far-infrared ($90-175\mu$m) surveys conducted by the ISOPHOT instrument. Also, the surveys now being performed by the [*Spitzer*]{} satellite, such as SWIRE (Lonsdale [*et al.*]{} 2003), will have high-redshift objects preferentially detected in the mid-infrared passbands. The AGN torus emission peaks in the infrared (e.g. Haas [*et al.*]{} 1998) so it might be expected that a large population of AGN would occur in these surveys in addition to high-$z$ star forming galaxies. Our AGN fractions, together with other studies (Fadda [*et al.*]{} 2002, Alexander [*et al.*]{} 2002), show this population to be a fairly well-determined minority and are encouraging for the use of mid-infrared samples for constraining the cosmic star formation history. The mid-infrared luminosity is a reasonably good star formation rate indicator, albeit affected by complicated K-correction effects (e.g. Xu [*et al.*]{} 1998). The main caveat is that our [*Chandra*]{} data still do not exclude the possibility of a large population of Compton-thick objects at moderate redshifts ($z\sim1$). Alexander [*et al.*]{} (2002), however, argue that this fraction should be low. They performed an X-ray stacking analysis on those infrared galaxies that were not clearly AGN at X-ray energies but were still individually detected in the 1 Ms Chandra Deep Field North Survey. They find an average X-ray spectral slope of $\Gamma = 2.0$, indicating a low contribution from the much flatter spectra of obscured AGN. Conclusions =========== We performed a cross-correlation of X-ray and mid-infrared point sources in the ELAIS N1 and N2 fields, using data from the [*Chandra*]{} and [*ISO*]{} satellites. 7 extragalactic matches are found (out of a total of 28 [*ISO*]{} sources) within the area of the [*Chandra*]{} ACIS-I chips. 2 of these are spectroscopically identified as AGN. Based on X-ray to IR flux ratios, X-ray hardness ratios, and luminosities, 3 of the remaining 5 are also consistent with AGN while the other 2 are consistent with starburst galaxies. This provides an AGN fraction of $\sim 19$ per cent in the $15\mu$m flux range $0.8 - 6$ mJy. We have co-added the hard X-ray flux at the positions of the 21 undetected [*ISO*]{} sources providing only a 1.4 sigma detection. This translates to a mean hard X-ray to mid-IR flux ratio consistent with star formation in these objects. Our cross-correlations, when compared with [*XMM-Newton*]{} observations of the Lockman Hole and [*Chandra*]{} observations of the northern Hubble Deep Field, allow us to place constraints on source count models of the $15\mu$m source population. The AGN fractions and number counts are broadly consistent with models by Pearson (2001) and King & Rowan-Robinson (2003). We argue our data is encouraging for the use of mid-infrared samples to constrain the cosmic star formation history, provided there is not a large contribution from Compton-thick AGN. Acknowledgements {#acknowledgements .unnumbered} ================ This research has made use of the NASA Extra-galactic Database (NED), operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. This work was partly funded under PPARC grant number GR/K98728. JCM would like to thank the referee for detailed and useful comments. Alexander D.M., La Franca F., Fiore F., Barcons X., et al. 2001, ApJ, 554, 18 Alexander D.M., Aussel H., Bauer F.E., Brandt W.N., et al. 2002, ApJ, 568, L85 Almaini O., Scott S.E., Dunlop J.S., Manners J.C., et al. 2003, MNRAS, 338, 303 Aussel H., Vigroux L., Franceschini A., Elbaz D., et al. 1999, AAS, 195, 0917 Chary R. & Elbaz D. 2001, ApJ, 556, 562 Ciliegi P., McMahon R.G., Miley G., Gruppioni C., et al. 1999, VizieR Online Data Catalog, 730, 20222 Comastri A., Mignoli M., Ciliegi P., Severgnini P., et al. 2002, ApJ, 571, 771 Cowie L.L., Garmire G.P., Bautz M.W., Barger A.J., et al. 2002, ApJ, 566, L5 Elbaz D., Cesarsky C.J., Fadda D., Aussel H., et al. 1999, A&A, 351, L37 Efstathiou A., Oliver S., Rowan-Robinson M., Surace C. 2000, MNRAS, 319, 1169 Fadda D., Flores H., Hasinger G., Franceschini A., et al. 2002, A&A, 383, 838 Fox M.J., Efstathiou A., Rowan-Robinson M., Dunlop J.S., et al. 2002, MNRAS, 331, 839 Gonzalez-Solares E.A., Perez-Fournon I., Rowan-Robinson M., Oliver S., et al. 2004 MNRAS submitted, astro-ph/0402406 Gruppioni C., Lari C., Pozzi F., Zamorani G., et al. 2002, MNRAS, 335, 831 Haas M., Chini R., Meisenheimer K., Stickel M., et al. 1998, ApJ, 503, L109 Hornschemeier A.E., Brandt W.N., Garmire G.P., Schneider D.P., et al. 2001, ApJ, 554, 742 King A.J. & Rowan-Robinson M. 2003, MNRAS, 339, 260 La Franca F., Gruppioni C., Matute I., Pozzi F., et al. 2003, AJ submitted Lari C., Pozzi F., Gruppioni C., Aussel H., et al. 2001, MNRAS, 325, 1173 Lari C., Vaccari M., Rodighiero G., Fadda D., et al. 2003, ESA SP-511: Exploiting the [*ISO*]{} Data Archive. Infrared Astronomy in the Internet Age, 349 Lonsdale C.J., Smith H.E., Rowan-Robinson M., Surace J., et al. 2003, PASP, 115, 897 Manners J.C. 2002, Ph.D. thesis, University of Edinburgh Manners J.C., Johnson O., Almaini O., Willott C.J., et al. 2003, MNRAS, 343, 293 Morel T., Efstathiou A., Serjeant S., Marquez I. 2001, MNRAS, 327, 1187 Moretti A., Campana S., Lazzati D., Tagliaferri G. 2003, ApJ, 588, 696 Moshir M., Kopan G., Conrow T., McCallon H., et al. 1990, IRAS Faint Source Catalogue, version 2.0 Mushotzky R.F., Cowie L.L., Barger A.J., & Arnaud K.A. 2000, Nature, 404, 459 Oliver S., Rowan-Robinson M., Alexander D.M., Almaini O., et al. 2000, MNRAS, 316, 749 Pearson C.P. 2001, MNRAS, 325, 1511 Rowan-Robinson M. 2001, ApJ, 549, 745 Rowan-Robinson M., Lari C., Perez-Fournon I., Gonzalez-Solares E.A., et al. 2004, MNRAS 351, 1290 Rush B., Malkan M.A., & Spinoglio L. 1993, ApJS, 89, 1 Scott S.E., Fox M.J., Dunlop J.S., Serjeant S., et al. 2002, MNRAS, 331, 817 Serjeant S., Oliver S., Rowan-Robinson M., Crockett H., et al. 2000, MNRAS, 316, 768 Serjeant S., Efstathiou A., Oliver S., Surace C. 2001, MNRAS, 322, 262 Willott C.J., Simpson C., Almaini O., Manners J.C., et al. 2003, MNRAS, 339, 397 Xu C., Hacking P.B., Fang F., Shupe D.L., et al. 1998, ApJ, 508, 576

--- abstract: 'A class of quantum superintegrable Hamiltonians defined on a two-dimensional hyperboloid is considered together with a set of intertwining operators connecting them. It is shown that such intertwining operators close a $su(2,1)$ Lie algebra and determine the Hamiltonians through the Casimir operators. By means of discrete symmetries a broader set of operators is obtained closing a $so(4,2)$ algebra. The physical states corresponding to the discrete spectrum of bound states as well as the degeneration are characterized in terms of unitary representations of $su(2,1)$ and $so(4,2)$.' address: - '$^1$Departamento de Matemática Aplicada, Escuela Superior de Ingenieros Industriales, Universidad de Valladolid, 47011 Valladolid, Spain' - '$^2$Department of Physics, Faculty of Science, Ankara University, 06100 Ankara, Turkey' - '$^3$Departamento de Física Teórica, Atómica y Optica, Facultad de Ciencias, Universidad de Valladolid, 47011 Valladolid, Spain' author: - 'J A Calzada$^1$, Ş Kuru$^2$, J Negro$^3$ and M A del Olmo$^3$' title: Intertwining symmetry algebras of quantum superintegrable systems on the hyperboloid --- Introduction ============ In this work we will consider a quantum superintegrable system living in a two-dimensional hyperboloid of two-sheets. Although this system is well known in the literature [@winternitz]-[@contracciones] and can be dealt with standard procedures [@ranada]-[@evans], it will be studied here under a different point of view based on the properties of intertwining operators (IO), a form of Darboux transformations [@darboux]. We will see how this approach can give a simple explanation of the main features of this physical system. The intertwining operators and integrable Hamiltonians have been studied in previous references [@kuru1]-[@fernandez], but we will supply here a thorough non-trivial application by means of this example. Besides, there are several points of interest for the specific case here considered because of the non-compact character. The intertwining operators are first order differential operators connecting different Hamiltonians in the same class (called hierarchy) and they are associated to separable coordinates of the Hamiltonians. We will obtain just a complete set of such intertwining operators, in the sense that any of the Hamiltonians of the hierarchy can be expressed in terms of these operators. In our case the initial IO’s close an algebraic structure which is the non-compact Lie algebra $su(2,1)$ (see [@olmo106] for a compact case). In a second step we will get a larger $so(4,2)$ Lie algebra of operators. This structure allows us to characterize the discrete spectrum and the corresponding eigenfunctions of the system by means of (infinite dimensional) irreducible unitary representations (iur). The construction of such representations, as it is known, is not so standard as for compact Lie algebras. We will compute the ground state and characterize the representation space of the wave-functions which share the same energy. Notice that these systems include also a continuum spectrum, but we will not go into this point here. The organization of the paper is as follows. Section 2 introduces the superintegrable Hamiltonians and Section 3 shows how to build the IO’s connecting hierarchies of these kind of Hamiltonians. In Section 4 it is seen that these operators close a $su(2,1)$ algebra. The Hamiltonians are related to the Casimirs of such an algebra, while the discrete spectrum of the Hamiltonians is related to unitary representations (iur’s) of $su(2,1)$. Next, in Section 5 a broader class of IO’s is defined leading to the $so(4,2)$ Lie algebra, and it is shown how this new structure helps to understand better the Hamiltonians in the new hierarchies. Finally, some remarks and conclusions in Section 6 will end the paper. Parametrizations of the two-sheet hyperboloid {#parametrizations} ============================================= Let us consider the two-dimensional two-sheet hyperboloid $s_{0}^2+s_{1}^2-s_{2}^2=-1$, where we define the following Hamiltonian $$\label{hamiltonian} H_{\ell}=J_{2}^2-J_{1}^2-J_{0}^2- \frac{l_{2}^2-\frac{1}{4}}{s_{2}^2}+\frac{l_{1}^2-\frac{1}{4}}{s_{1}^2}+\frac{l_{0}^2-\frac{1}{4}}{s_{0}^2},$$ where $\ell=(l_{0},\,l_{1},\,l_{2}) \in \mathbb{R}^{3}$, and the differential operators $$\label{diffgenerators} J_0=s_1\partial_2+ s_2\partial_1,\quad J_1=s_2\partial_0+ s_0\partial_2,\quad J_2=s_0\partial_1- s_1\partial_0,$$ constitute a realization of the $so(2,1)$ Lie algebra with Lie commutators $$[J_{0},J_{1}]=-J_{2}, \qquad [J_{2},J_{0}]=J_{1},\qquad [J_{1},J_{2}]=J_{0}.$$ The generator $J_2$ corresponds to a rotation around the axis $s_2$, while the generators $J_0$ and $J_1$ give pseudo-rotations (i.e., non-compact rotations) around the axes $s_0$ and $s_1$, respectively. The Casimir operator $$C=J_{0}^2+J_{1}^2-J_{2}^2$$ gives the ‘kinetic’ part of the Hamiltonian. We can parametrize the hyperbolic surface by means of the ‘analogue’ of the spherical coordinates $$\label{sa} s_{0}={\sinh\xi}\,\cos{\theta},\quad s_{1}={\sinh\xi}\,\sin{\theta},\quad s_{2}={\cosh{\xi}},$$ where $0\leq\theta<2\pi$ and $0\leq\xi<\infty$. In these coordinates, the infinitesimal generators (\[diffgenerators\]) take the following expressions $$\label{ja} \fl J_{0}=\sin{\theta}\,\partial_{\xi}+\cos{\theta}\,{\coth\xi} \,\partial_{\theta} ,\quad J_{1}=\cos{\theta}\,\partial_{\xi}-\sin{\theta}\,{\coth\xi} \,\partial_{\theta} ,\quad J_{2}=\partial_{\theta}\, .$$ It is easy to check that the generators $J_i$, $i=1,2,3$, are anti-Hermitian inside the space of square-integrable functions with the invariant measure $d\mu(\theta,\xi)= \sinh\xi\, d\theta d\xi$. Using the coordinates (\[sa\]), the Hamiltonian (\[hamiltonian\]) has the expression $$\label{ha} \fl H_{\ell}=-\partial_{\xi}^2-{\coth\xi}\,\partial_{\xi} -\frac{l_{2}^2-\frac{1}{4}}{\cosh^{2}\xi} +\frac{1}{\sinh^{2}\xi}\left[-\partial_{\theta}^2+\frac{l_{1}^2 -\frac{1}{4}}{\sin^{2}\theta}+\frac{l_{0}^2 -\frac{1}{4}}{\cos^{2}\theta}\right].$$ Therefore, $H_{\ell}$ can be separated in the variables $\xi$ and $\theta$. Choosing its eigenfunctions $\Phi_{\ell}$, $H_{\ell}\Phi_{\ell}=E\Phi_{\ell}$, in the form $$\label{wave} \Phi_{\ell}(\theta,\xi)=f(\theta)\,g(\xi),$$ we get the separated equations $$\label{ha1} H_{l_{0},l_{1}}^{\theta} f(\theta)\equiv \left[-\partial_{\theta}^2+\frac{l_{1}^2-\frac{1}{4}}{\sin^{2}\theta} +\frac{l_{0}^2-\frac{1}{4}}{\cos^{2}\theta}\right]f(\theta) =\alpha\,f(\theta)$$ and $$\label{ha1b} \left[-\partial_{\xi}^2-{\coth\xi}\,\partial_{\xi} -\frac{l_{2}^2-\frac{1}{4}}{\cosh^{2}\xi} +\frac{\alpha}{\sinh^{2}\xi}\right]g(\xi)=E\,g(\xi),$$ where $\alpha>0$ is a separation constant. A complete set of intertwining operators {#intertwiningoperators} ======================================== The second order operator at the l.h.s. of (\[ha1\]) in the variable $\theta$ can be factorized in terms of first order operators [@barut; @quesne] $$H_{l_{0},l_{1}}^{\theta}=A^{+}_{l_{0},l_{1}}A^{-}_{l_{0},l_{1}}+\lambda_{l_{0},l_{1}},$$ being $$\label{apm} \fl A_{l_{0},l_{1}}^{\pm}=\pm\partial_{\theta}-(l_{0}+1/2)\,{\tan{\theta}} +(l_{1}+1/2)\,{\cot{\theta}},\quad \lambda_{l_{0},l_{1}}=(1+l_{0}+l_{1})^2.$$ The Hamiltonian can also be rewritten in terms of the triplet ($A^{\pm}_{l_{0}-1,l_{1}-1}, \lambda_{l_{0}-1,l_{1-1}}$) $$\label{ah} H_{l_{0},l_{1}}^{\theta}= A_{l_{0}-1,l_{1}-1}^{-}A_{l_{0}-1,l_{1}-1}^{+}+\lambda_{l_{0}-1,l_{1}-1} =A_{l_{0},l_{1}}^{+}A_{l_{0},l_{1}}^{-} +\lambda_{l_{0},l_{1}} .$$ In this way we get a hierarchy of Hamiltonians $$\label{hierarchy} \cdots ,H_{{l_{0}-1},{l_{1}-1}}^{\theta},H_{{l_{0}},{l_{1}}}^{\theta},H_{{l_{0}+1},{l_{1}+1}}^{\theta}, \cdots, H_{{l_{0}+n},{l_{1}+n}}^{\theta},\cdots$$ satisfying the following recurrence relations $$\begin{array}{l} A_{l_{0}-1,l_{1}-1}^{-}H_{l_{0}-1,l_{1}-1}^{\theta}=H_{l_{0},l_{1}}^{\theta}A_{l_{0}-1,l_{1}-1}^{-},\\[1.5ex] A_{l_{0}-1,l_{1}-1}^{+}H_{l_{0},l_{1}}^{\theta}=H_{l_{0}-1,l_{1}-1}^{\theta}A_{l_{0}-1,l_{1}-1}^{+}\, . \end{array}$$ Hence, the operators $\{ A^\pm _{l_{0}+n,l_{1}+n}\}_{n\in\mathbb{Z}}$ are intertwining operators and they act as transformations between the eigenfunctions of the Hamiltonians in the hierarchy (\[hierarchy\]), $$A_{l_{0}-1,l_{1}-1}^{-}:f_{l_{0}-1,l_{1}-1}\rightarrow f_{l_{0},l_{1}},\qquad A_{l_{0}-1,l_{1}-1}^{+}:f_{l_{0},l_{1}}\rightarrow f_{l_{0}-1,l_{1}-1},$$ where the subindex refers to the corresponding Hamiltonian. We can define new operators in terms of $A_{l_{0},l_{1}}^{\pm}$, together with a diagonal operator $A_{l_{0},l_{1}}=(l_{0}+l_{1})\mathbb{I}$, acting in the following way in the space of eigenfunctions $$\label{noindex} \fl \hat{A}^{-} f_{l_{0},l_{1}}:=\frac{1}{2}\,A_{l_{0},l_{1}}^{-} \,f_{l_{0},l_{1}},\ \hat{A}^{+} f_{l_{0},l_{1}}:=\frac{1}{2}\,A_{l_{0},l_{1}}^{+} \,f_{l_{0},l_{1}},\ \hat{A}\,f_{l_{0},l_{1}}:=-\frac{1}{2}\,(l_{0}+l_{1})\,f_{l_{0},l_{1}}.$$ It can be shown from (\[ah\]) that $\{\hat{A}^{-}, \hat{A}^{+}, \hat{A}\}$ satisfy the commutation relations of a $su(2)$ Lie algebra, i.e., $$\label{a} [\hat{A}^{-}, \hat{A}^{+}]=-2\,\hat{A},\qquad [\hat{A}, \hat{A}^{\pm}]=\pm \hat{A}^{\pm}.$$ The ‘fundamental’ states, $f_{l_{0},l_{1}}^{0}$, of the $su(2)$ representations are determined by the relation $\hat A^{-}\,f_{l_{0},l_{1}}^{0}(\theta)=0$. They are $$f_{l_{0},l_{1}}^{0}(\theta)=N\,(\cos{\theta})^{l_{0}+1/2}\,(\sin{\theta})^{l_{1}+1/2},$$ where $N$ is a normalization constant. These functions are regular and square-integrable when $$\label{des1} l_0,l_1\geq -1/2\, .$$ Since $\hat{A}\,f_{l_{0},l_{1}}\equiv -\frac{1}{2}\,(l_{0}+l_{1})\,f_{l_{0},l_{1}}$, then the label of the $j$-representation is $j=\frac{1}{2}\,(l_{0}+l_{1})$ and the dimension of the iur will be $2j+1=l_{0}+l_{1}+1$. Now, observe that because the IO’s $A^{\pm}_{l_{0},l_{1}}$ depend only on the $\theta$-variable, they can act also as IO’s of the total Hamiltonians $H_{\ell}$ (\[ha\]) and its global eigenfunctions $\Phi_{\ell}$ (\[wave\]), leaving the parameter $l_2$ unchanged (in this framework we will use three-fold indexes) $$A_{\ell'}^{-}H_{\ell'}=H_{\ell}A_{\ell'}^{-},\qquad A_{\ell'}^{+}H_{\ell}=H_{\ell'}A_{\ell'}^{+},$$ where $\ell=(l_{0},l_{1},l_2)$ and $\ell'=(l_{0}-1,l_{1}-1,l_2)$. In this sense, many of the above relations can be straightforwardly extended under this global point of view. Second set of pseudo-spherical coordinates ------------------------------------------ A second coordinate set is obtained from the non-compact rotations about the axes $s_{0}$ and $s_{1}$, respectively. In this way we obtain the following parametrization of the hyperboloid $$\label{sb} s_{0}={\cosh\psi}\,{\sinh\chi},\qquad s_{1}={\sinh\psi},\qquad s_{2}={\cosh\psi}\,{\cosh\chi}\, .$$ The expressions of the $so(2,1)$ generators in these coordinates are $$\fl J_{0}=-{\tanh{\psi}}\,{\sinh{\chi}}\partial_{\chi} +{\cosh{\chi}}\;\partial_{\psi},\quad J_{1}=\partial_{\chi} ,\quad J_{2}={\sinh{\chi}}\partial_{\psi}-{\tanh{\psi}} \,{\cosh{\chi}}\;\partial_{\chi}\,$$ and the explicit expression of the Hamiltonian (\[hamiltonian\]) is now $$\fl H_{\ell}=-\partial_{\psi}^2-{\tanh\psi}\,\partial_{\psi} +\frac{l_{1}^2-\frac{1}{4}}{\sinh^{2}\psi} +\frac{1}{\cosh^{2}\psi}\left[-\partial_{\chi}^2+\frac{l_{0}^2 -\frac{1}{4}}{\sinh^{2}\chi}-\frac{l_{2}^2-\frac{1}{4}}{\cosh^{2}\chi}\right].$$ This Hamiltonian can be separated in the variables $\psi$ and $\chi$ considering the eigenfunctions $\Phi$ of $H_\ell$ ($H_{\ell}\,\Phi=E\,\Phi$) as $\Phi(\chi,\psi)=f(\chi)\,g(\psi)$. We obtain the two folllowing (separated) equations $$\label{hb1} H_{l_{0},l_{2}}^{\chi}f(\chi)\equiv \left[-\partial_{\chi}^2+\frac{l_{0}^2 -\frac{1}{4}}{\sinh^{2}\chi}-\frac{l_{2}^2 -\frac{1}{4}}{\cosh^{2}\chi}\right]f(\chi) =\alpha\,f(\chi),$$ $$\left[-\partial_{\psi}^2 -{\tanh\psi}\,\partial_{\psi}+\frac{l_{1}^2 -\frac{1}{4}}{\sinh^{2}\psi}+\frac{\alpha}{\cosh^{2}\psi}\right]g(\psi) =E\,g(\psi),$$ with $\alpha$ a separation constant. The second order operator in the variable $\chi$ at the l.h.s. of (\[hb1\]) can be factorized as a product of first order operators $$\label{hbb1} H_{l_{0},l_{2}}^{\chi}=B^{+}_{l_{0},l_{2}}B^{-}_{l_{0},l_{2}} +\lambda_{l_{0},l_{2}}= B^{-}_{l_{0}-1,l_{2}-1}B^{+}_{l_{0}-1,l_{2}-1}+\lambda_{l_{0}-1,l_{2}-1},$$ being $$\fl \label{bpm} B_{l_{0},l_{2}}^{\pm}=\pm\partial_{\chi}+(l_{2}+1/2)\, \tanh{\chi}+(l_{0}+1/2) \,{\coth{\chi}}, \quad\lambda_{l_{0},l_{2}}=-(1+l_{0}+l_{2})^2.$$ In this case the intertwining relations take the form $$\begin{array}{l} B^{-}_{l_{0}-1,l_{2}-1}H_{l_{0}-1,l_{2}-1}^{\chi}=H_{l_{0},l_{2}}^{\chi} B^{-}_{l_{0}-1,l_{2}-1},\\[1.5ex] B^{+}_{l_{0}-1,l_{2}-1}H_{l_{0},l_{2}}^{\chi}= H_{l_{0}-1,l_{2}-1}^{\chi}B^{+}_{l_{0}-1,l_{2}-1}, \end{array}$$ and imply that these operators $B^{\pm}$ connect eigenfunctions in the following way $$B^{-}_{l_{0}-1,l_{2}-1}:f_{l_{0}-1,l_{2}-1}\rightarrow f_{l_{0},l_{2}},\qquad B^{+}_{l_{0}-1,l_{2}-1}:f_{l_{0},l_{2}}\rightarrow f_{l_{0}-1,l_{2}-1}\,.$$ The operators $B^{\pm}_{l_{0},l_{2}}$ can be expressed in terms of $\xi$ and $\theta$ using relations (\[sa\]) and (\[sb\]) $$B^{\pm}_{l_{0},l_{2}}=\pm J_{1}+(l_{2}+1/2) \,{\tanh{\xi}}\,\cos{\theta}+(l_{0}+1/2)\,{\coth{\xi}}\,\sec{\theta},$$ where $J_{1}$ is given by (\[ja\]). We define new free-index operators in the following way $$\fl \hat{B}^{-}\,f_{l_{0},l_{2}}:=\frac{1}{2}\,B_{l_{0},l_{2}}^{-} \,f_{l_{0},l_{2}},\quad \hat{B}^{+}\,f_{l_{0},l_{2}}:=\frac{1}{2}\,B_{l_{0},l_{2}}^{+} \,f_{l_{0},l_{2}},\quad \hat{B}\,f_{l_{0},l_{2}}:=-\frac{1}{2}\,(l_{0}+l_{2})\,f_{l_{0},l_{2}},$$ and, having in mind the expressions (\[hbb1\]) and (\[bpm\]), we can prove than they close the $su(1,1)$ Lie algebra $$\label{b} [\hat{B}^{-}, \hat{B}^{+}]=2\,\hat{B},\qquad [\hat{B},\hat{B}^{\pm}]=\pm \hat{B}^{\pm}.$$ Since the Lie algebra $su(1,1)$ is non-compact, its iur’s are infinite dimensional. In particular, we will be interested in the discrete series, that is, in those having a fundamental state annihilated by the lowering operator, i.e., $B^{-} \,f^0_{l_{0},l_{2}}=0$. The explicit expression of these states is $$%\begin{array}{ll} f^0_{l_{0},l_{2}}(\chi)%& = N (\cosh \chi)^{l_2+1/2}(\sinh \chi)^{l_0+1/2} %&=(\cos\theta)^{1/2+l_0}(\sin\theta)^{1/2-l_0} %(\cosh\xi)^{1/2+l_2}\sinh\xi \, K(2 \sin\theta %\sinh \xi) %\end{array}$$ where $N$ is a normalization constant. In order to have a regular and square-integrable function we must have $$\label{des2} l_0\geq -1/2,\quad -k_1\equiv l_0+l_2<-1 \, .$$ Since $ \hat{B}\,f^0_{l_{0},l_{2}}= -\frac{1}{2}\,(l_{0}+l_{2})\,f^0_{l_{0},l_{2}} $, we can say that the lowest weight of this unitary $su(1,1)$ infinite representation is $j_1'= k_1/2>1/2$. The IO’s $\hat B^{\pm}$ can be considered also as intertwining operators of the Hamiltonians $H_{\ell}$ linking their eigenfunctions $\Phi_{\ell}$, similarly to the IO’s $\hat A^{\pm}$ described before (in this situation we will also use three-fold indexes but now with $l_1$ remaining unchanged). Third set of pseudo-spherical coordinates ----------------------------------------- A third set of coordinates is obtained from the noncompact rotations about the axes $s_{1}$ and $s_{0}$, respectively. They give rise to the parametrization $$\label{sc} s_{0}={\sinh\phi},\quad s_{1}={\cosh\phi}\,{\sinh\beta},\quad s_{2}={\cosh\phi}\,{\cosh\beta},$$ and the generators have the expressions $$\fl J_{0}=\partial_{\beta},\quad J_{1}={\cosh{\beta}}\;\partial_{\phi}-{\tanh{\phi}}\,{\sinh{\beta}} \;\partial_{\beta} ,\quad J_{2}=-{\sinh{\beta}}\;\partial_{\phi} +{\tanh{\phi}}\,{\cosh{\beta}}\;\partial_{\beta} .$$ Now, the Hamiltonian takes the form $$\fl H_{\ell}=-\partial_{\phi}^2-{\tanh\phi}\,\partial_{\phi} +\frac{l_{0}^2-\frac{1}{4}}{\sinh^{2}\phi} +\frac{1}{\cosh^{2}\phi}\left[-\partial_{\beta}^2+\frac{l_{1}^2 -\frac{1}{4}}{\sinh^{2}\beta}-\frac{l_{2}^2 -\frac{1}{4}}{\cosh^{2}\beta}\right],$$ and can be separated in the variables $\phi,\beta$ in terms of its eigenfunctions $\Phi$ $(H_{\ell}\,\Phi=E\,\Phi)$ such that $\Phi(\beta,\phi)=f(\beta)\,g(\phi)$ in the following way $$\label{hc1} H_{l_{1},l_{2}}^{\beta}f(\beta)\equiv \left[-\partial_{\beta}^2+\frac{l_{1}^2 -\frac{1}{4}}{\sinh^{2}\beta}-\frac{l_{2}^2 -\frac{1}{4}}{\cosh^{2}\beta}\right]f(\beta)=\alpha\,f(\beta),$$ $$\left[-\partial_{\phi}^2 -{\tanh\phi}\,\partial_{\phi}+\frac{l_{0}^2-\frac{1}{4}}{\sinh^{2}\phi} +\frac{\alpha}{\cosh^{2}\phi}\right]g(\phi)= E\,g(\phi),$$ with the separation constant $\alpha$. The second order operator in $\beta$ at the l.h.s. of expression (\[hc1\]) can be factorized as a product of first order operators $$H_{l_{1},l_{2}}^{\beta}={C}^{+}_{l_{1},l_{2}}{C}^{-}_{l_{1},l_{2}}+{\lambda}_{l_{1},l_{2}} ={C}^{-}_{l_{1}+1,l_{2}-1}{C}^{+}_{l_{1}+1,l_{2}-1}+ {\lambda}_{l_{1}+1,l_{2}-1},$$ being $$\fl C_{l_{1},l_{2}}^{\pm}= \pm\partial_{\beta}+(l_{2}+1/2)\,{\tanh{\beta}}+(-l_{1}+1/2) \,{\coth{\beta}},\quad \lambda_{l_{1},l_{2}}=-(1-l_{1}+l_{2})^2 .$$ These operators ${C}^{\pm}_{l_{1},l_{2}}$ give rise to the intertwining relations $$\begin{array}{l} {C}^{+}_{l_{1}+1,l_{2}-1}H_{l_{1},l_{2}}^{\beta}= H_{l_{1}+1,l_{2}-1}^{\beta}C^{+}_{l_{1}+1,l_{2}-1} \\[1.5ex] {C}^{-}_{l_{1}+1,l_{2}-1}H_{l_{1}+1,l_{2}-1}^{\beta}= H_{l_{1},l_{2}}^{\beta}{C}^{-}_{l_{1}+1,l_{2}-1}, \end{array}$$ which imply the connection among eigenfunctions $${C}^{-}_{l_{1}+1,l_{2}-1}:f_{l_{1}+1,l_{2}-1}\rightarrow f_{l_{1},l_{2}},\qquad {C}^{+}_{l_{1}+1,l_{2}-1}:f_{l_{1},l_{2}}\rightarrow f_{l_{1}+1,l_{2}-1}\,.$$ In this case $C^{\pm}_{l_{1},l_{2}}$ can also be expressed in terms of $\xi$ and $\theta$ using relations (\[sa\]) and (\[sc\]) $$C^{\pm}_{l_ {1},l_{2}}=\pm J_{0}+(l_{2}+1/2) \,{\tanh{\xi}}\,\sin{\theta}+(-l_{1}+1/2)\,{\coth{\xi}}\, \csc{\theta},$$ where $J_{0}$ is given by (\[ja\]). Now, the new operators are defined as $$\fl \hat{C}^{-}\,f_{l_{1},l_{2}}:=\frac{1}{2}\,{C}_{l_{1},l_{2}}^{-} \,f_{l_{1},l_{2}}\quad \hat{C}^{+}\,f_{l_{1},l_{2}}:=\frac{1}{2}\,{C}_{l_{1},l_{2}}^{+} \,f_{l_{1},l_{2}}\quad \hat{C}\,f_{l_{1},l_{2}}:=-\frac{1}{2}\,(l_{2}-l_{1})f_{l_{1},l_{2}}.$$ and satisfy the commutation relations of the $su(1,1)$ algebra $$\label{c} [\hat{C}^{-}, \hat{C}^{+}]=2\,\hat{C},\qquad [\hat{C}, \hat{C}^{\pm}]=\pm \hat{C}^{\pm}.$$ The fundamental state for the $su(1,1)$ representation, in this case, given by $\hat {C}^{-}\,f^0_{l_{1},l_{2}}=0$, has the expression $$f^0_{l_{1},l_{2}}(\beta)= N (\cosh \beta)^{l_2 +1/2} (\sinh \beta)^{-l_1 +1/2} %(\cos\theta )^{\frac{1}{2}+l_1}(\sin\theta )^{\frac{1}{2}-l_1} %(\cosh \xi )^{\frac{1}{2}+ l_2}\sinh\xi \, K(2 \cos\theta \sinh \xi)$$ where $N$ is a normalization constant. In order to get a iur from this function, we impose it to be regular and normalizable, therefore $$\label{des3} l_1\leq 1/2,\quad -k_2\equiv l_2-l_1< -1 \, .$$ Since $ \hat{C} f^0_{l_{1},l_{2}}= -\frac{1}{2}\,(l_{2}-l_{1})f^0_{l_{1},l_{2}} $ the lowest weight of the iur is given by $j'_2= k_2/2>1/2$. As in the previous cases, we can consider the IO’s $C^{\pm}$ as connecting global Hamiltonians $H_{\ell}$ and their eigenfunctions, having in mind that now the parameter $l_0$ is unaltered. Algebraic structure of the intertwining operators ================================================= If we consider together all the IO’s $\{\hat{A}^{\pm},\hat{A},\hat{B}^{\pm}, \hat{B},\hat{C}^{\pm},\hat{C}\}$ that have appeared in section \[intertwiningoperators\], then, we find that they close the Lie algebra $su(2,1)$ since they satisfy, besides (\[a\]), (\[b\]), (\[c\]), the following commutation relations $$\label{su21} \fl \label{abc} \begin{array}{llll} [\hat{A}^{+}, \hat{B}^{+}]=0\quad &[\hat{A}^{-},\hat{B}^{-}]=0\quad &[\hat{A}^{+},\hat{B}^{-}]=-\hat{C}^{-}\quad &[\hat{A}^{-},\hat{B}^{+}]=\hat{C}^{+}\nonumber\\[2.ex] [\hat{C}^{+}, \hat{B}^{+}]=0\quad &[\hat{C}^{-}, \hat{B}^{-}]=0\quad &[\hat{C}^{+}, \hat{A}^{+}]=-\hat{B}^{+}\quad &[\hat{C}^{-},\hat{A}^{-}]=\hat{B}^{-}\nonumber\\[2.ex] [\hat{C}^{+}, \hat{B}^{-}]=-\hat{A}^{-}\quad &[\hat{C}^{-}, \hat{B}^{+}]=\hat{A}^{+}\quad &[\hat{C}^{+},\hat{A}^{-}]=0\quad &[\hat{C}^{-},\hat{A}^{+}]=0\nonumber\\[2.ex] [\hat{A}, \hat{B}^{+}]=\frac{1}{2} \hat{B}^{+}\quad &[\hat{A},\hat{B}^{-}]=-\frac{1}{2} \hat{B}^{-}\quad &[\hat{B},\hat{A}^{+}]=\frac{1}{2} \hat{A}^{+}\quad &[\hat{B},\hat{A}^{-}]=-\frac{1}{2} \hat{A}^{-}\nonumber\\[2.ex] [\hat{C}, \hat{B}^{+}]=\frac{1}{2} \hat{B}^{+}\quad &[\hat{C},\hat{B}^{-}]=-\frac{1}{2} \hat{B}^{-}\quad &[\hat{C},\hat{A}^{+}]=-\frac{1}{2} \hat{A}^{+}\quad &[\hat{C},\hat{A}^{-}]=\frac{1}{2} \hat{A}^{-}\nonumber\\[2.ex] [\hat{A}, \hat{C}^{-}]=\frac{1}{2} \hat{C}^{-}\quad &[\hat{A}, \hat{C}^{+}]=-\frac{1}{2} \hat{C}^{+}\quad &[\hat{B}, \hat{C}^{-}]=-\frac{1}{2} \hat{C}^{-}\quad &[\hat{B},\hat{C}^{+}]=\frac{1}{2} \hat{C}^{+}\nonumber\\[2.ex] [\hat{A}, \hat{B}]=0\quad &[\hat{A}, \hat{C}]=0\quad &[\hat{B}, \hat{C}]=0\, .& \qquad\nonumber \end{array}$$ Obviously $su(2,1)$ includes as subalgebras the Lie algebras $su(2)$ and $su(1,1)$ defined in the previous section \[intertwiningoperators\]. The second order Casimir operator of $su(2,1)$ can be written as follows $$\label{cas} \fl {\cal C} = \hat A^+\hat A^- - \hat B^+\hat B^- - \hat C^+\hat C^- + \frac23\left( \hat A^2+ \hat B^2 + \hat C^2\right) - (\hat A+\hat B+\hat C)\, .$$ It is worthy noticing that in our differential realization we have $\hat A - \hat B + \hat C =0$, and that there is another generator $$\label{cas2} {\cal C}'= l_1+l_2-l_0$$ commuting with the rest of generators. Hence, adding this new generator ${\cal C}'$ to the other ones we get the Lie algebra $u(2,1)$. The eigenfunctions of the Hamiltonians $H_{\ell}$ that have the same energy support unitary representations of $su(2,1)$ characterized by a value of ${\cal C}$ and other of ${\cal C}'$. In fact, we can show that $$\label{hcc} H_{\ell} = -4\, {\cal C}+ \frac13\, {{\cal C}'}^2 - \frac{15}{4}.$$ These representations can be obtained, as usual, starting with a fundamental state simultaneously annihilated by the lowering operators $\hat A^-, \hat C^-$ and $\hat B^-$ $$\label{equation1} A^-_{\ell}\Phi^0_{\ell} = C^-_{\ell} \Phi^0_{\ell} = B^-_{\ell} \Phi^0_{\ell} =0\, .$$ Solving equations (\[equation1\]) we find $$\label{fs} \Phi^0_{\ell}(\xi,\theta) = N (\cos \theta)^{l_0+1/2} (\sin \theta)^{1/2} (\cosh \xi)^{l_2+1/2}(\sinh \xi)^{l_0+1},$$ where $\ell= (l_0, 0, l_2)$. From the inequalities (\[des1\]) and (\[des3\]) the parameters of $\Phi^0_{\ell}$ must satisfy $(l_0+l_2)<-1$ and $l_0\geq -1/2$. In this particular case to guarantee the normalization of $\Phi^0_{\ell}$ using the invariant measure we must impose $(l_0+l_2)<-5/2$. Thus, the above state supports also iur’s of the subalgebras $su(2)$ (generated by $\hat A^{\pm}$) with the weight $j=l_0/2$ and $su(1,1)$ (generated by $\hat C^{\pm}$) with $j'_2=-l_2/2$. The energies of the fundamental states of the form (\[fs\]) are obtained by applying $H_{\ell}$ as given in (\[hcc\]) taking into account with the expressions for the Casimir operators (\[cas\]) and (\[cas2\]), $$\label{energy} H_{\ell} \Phi^0_{\ell} = - (l_0+l_2+3/2)(l_0+l_2+5/2) \Phi^0_{\ell} \equiv E^0_{\ell} \Phi^0_{\ell} \, .$$ From $\Phi^0_{\ell}$ we can get the rest of eigenfunctions in the $su(2,1)$ representation using the raising operators $\hat A^+,\hat B^+$ and $\hat C^+$, all of them sharing the same energy eigenvalue $E^0_{\ell}$ (\[energy\]). ![The states of iur’s of $su(2,1)$ represented by points in the three dark planes corresponding to $\Phi^0_{\ell}$ with $\ell=(0,0,-3)$, $\ell=(1,0,-4)$ and $\ell=(2,0,-5)$. All of them share the same energy. []{data-label="fig1"}](fig1.epsf) Since the expression (\[energy\]) for $E^0_{\ell}$ depends on $l_0+l_2$, it means that states in the family of iur’s derived from fundamental states (\[fs\]) such that have the same value of $l_0+l_2$ will also have the same energy eigenvalue. It is worth to remark that the energy in (\[energy\]) corresponding to bound states is negative, which is consistent with the expressions (\[ha\]) and (\[ha1b\]) for the Hamiltonians, and that the set of such bound states for each Hamiltonian is finite. In Fig. \[fig1\], by means of an example, we represent the states of some iur’s by points $(l_0,l_1,l_2)\in \mathbb{R}^3$ linked to the ground state $\Phi^0_{\ell}$, represented by the point $(l_0,0,l_2)$, through the raising operators $\hat A^+,\hat C^+$. The points belonging to a iurare in a 2-dimensional plane (corresponding to the particular value of ${\cal C}'$), and other iur’s are described by points in parallel 2D planes. These parallel planes are closed inside a tetrahedral unbounded pyramid whose basis extends towards $l_2\to -\infty$. As in the case of $su(3)$ representations[@olmo106], in the above $su(2,1)$ iur’s we have some points (in the parameter space) which are degenerated, that is, they correspond to an eigen-space whose dimension is greater than one. For example, let us consider first the representation based on the fundamental state $\Phi^0_{\ell}$ with values $\ell=(0,0,-3)$. From this state we can build a iur made of points in a triangle, where each point represents a non-degenerate 1-dimensional (1D) eigenspace. Now, consider the iur corresponding to the ground state with $\ell_0=(1,0,-4)$, which has eigenstates with the same energy $E= -(-3+3/2)(-3+5/2)$ as the previous one (they have the same value of $l_0+l_2=-3$). Now the eigenstates corresponding to $\ell_1=(0,0,-5)$, inside this representation, can be obtained in two ways: $$\Phi_{\ell_1} = \hat C^+ \hat A^+ \Phi_{\ell_0},\quad \tilde\Phi_{\ell_1} = \hat A^+ \hat C^+ \Phi_{\ell_0} \, .$$ It can be shown that these states are independent and that they span the 2D eigenspace of the corresponding Hamiltonian $H_{\ell_1}$ for that eigenvalue. ![Plot of the orthogonal normalized eigenfunctions $\Phi(\theta,\xi)$, $0<\theta<\pi/2$, $0<\xi<\infty$, of $H_{\ell}$, for $\ell=(0,0,-5)$ corresponding to the ground state (left) and two independent states generating the eigenspace of the second (and last) excited energy level.[]{data-label="fig2"}](fund5.epsf "fig:") -0.05cm ![Plot of the orthogonal normalized eigenfunctions $\Phi(\theta,\xi)$, $0<\theta<\pi/2$, $0<\xi<\infty$, of $H_{\ell}$, for $\ell=(0,0,-5)$ corresponding to the ground state (left) and two independent states generating the eigenspace of the second (and last) excited energy level.[]{data-label="fig2"}](fig2a.epsf "fig:") -0.05cm ![Plot of the orthogonal normalized eigenfunctions $\Phi(\theta,\xi)$, $0<\theta<\pi/2$, $0<\xi<\infty$, of $H_{\ell}$, for $\ell=(0,0,-5)$ corresponding to the ground state (left) and two independent states generating the eigenspace of the second (and last) excited energy level.[]{data-label="fig2"}](fig2b.epsf "fig:") Remark that the ground state for the Hamiltonian $H_{\ell_1}$, $\ell_1=(0,0,-5)$, is given by the wavefunction (\[fs\]) and its ground energy is $E^0_{\ell_1}= - (-5+3/2)(-5+5/2)$. The plot of the ground wavefunction and two independent excited wavefunctions are shown in Fig. \[fig2\]. Following the same pattern it can be obtained the degeneration of higher excited levels in the discrete spectrum: the $n$ excited level, when it exists, has associated an eigenspace with dimension $n$. The complete symmetry algebra ${so(4,2)}$ ========================================= As it is explicit from its expression (\[hamiltonian\]) the Hamiltonian $H_{\ell}$ is invariant under reflections $$\begin{array}{l} I_0: (l_0,l_1,l_2)\to (-l_0,l_1,l_2)\\[1.5ex] I_1: (l_0,l_1,l_2)\to (l_0,-l_1,l_2)\\[1.5ex] I_2: (l_0,l_1,l_2)\to (l_0,l_1,-l_2) \, . \end{array}$$ These operators can generate, by means of conjugation, other sets of intertwining operators from the ones already defined. For instance, $$I_0: \{ \hat A^{\pm},\hat A \} \to \{ \tilde A^{\pm}= I_0 \hat A^{\pm}I_0, \tilde A = I_0 \hat A I_0 \}$$ where, from (\[apm\]) we get $$\label{apmt} \fl \tilde A_{l_{0},l_{1}}^{\pm}=\pm\partial_{\theta}-(-l_{0}+1/2)\,{\tan{\theta}} +(l_{1}+1/2)\,{\cot{\theta}},\quad \tilde \lambda_{l_{0},l_{1}}=(1-l_{0}+l_{1})^2$$ such that $$\tilde A_{l_{0},l_{1}}^{-}:f_{l_{0},l_{1}}\rightarrow f_{l_{0}-1,l_{1}+1},\quad \tilde A_{l_{0},l_{1}}^{+}:f_{l_{0-1},l_{1}+1}\rightarrow f_{l_{0},l_{1}} \, .$$ Thus, we can define the operators $\tilde A^\pm$ as in (\[noindex\]) that together with $\tilde{A}\,f_{l_{0},l_{1}}:=-\frac{1}{2}\,(-l_{0}+l_{1})\,f_{l_{0},l_{1}}$ close a second $\widetilde{su}(2)$. Other sets of operators $\{\tilde B^{\pm},\tilde B\}$ and $\{\tilde C^{\pm},\tilde C\}$, closing $\widetilde{su}(1,1)$ Lie algebras, can also be defined with the help of these reflections in the following way (the choice is non unique) $$\label{tildes} \begin{array}{l} I_0:\{\hat A^\pm, \hat A;\ \hat B^\pm,\hat B;\ \hat C^\pm, \hat C \} \to \{\tilde A^\pm, \tilde A;\ \tilde B^\pm, \tilde B;\ C^\pm, C \} \\[1.5ex] I_1:\{\hat A^\pm, \hat A;\ \hat B^\pm,\hat B;\ \hat C^\pm, \hat C \} \to \{\tilde A^\mp, -\tilde A;\ B^\pm, B;\ \tilde C^\pm, \tilde C \} \\[1.5ex] I_2:\{\hat A^\pm, \hat A;\ \hat B^\pm,\hat B;\ \hat C^\pm, \hat C \} \to \{ A^\pm, A;\ \tilde B^\mp,-\tilde B;\ -\tilde C^\mp, -\tilde C \} \, . \end{array}$$ The whole set of the operators $$\label{so42} \{\hat A^\pm,\tilde A^\pm, \hat B^\pm, \tilde B^\pm, \hat C^\pm,\tilde C^\pm, L_0, L_1, L_2\}$$ where the diagonal operators $L_i$ are defined as $$L_i \Psi_\ell = l_i \Psi_\ell$$ generate an $o(4,2)$ Lie algebra of rank three with commutations rules that can be easily derived from those of $su(2,1)$ given in (\[su21\]) and the action of the reflections (\[tildes\]). These generators link eigenstates of the Hamiltonians $H_\ell$ with the same eigenvalue. Now, consider a fundamental state $\Psi^0_\ell$ for the $so(4,2)$ algebra annihilated by the lowering operators, $$\label{equation2} A^-_{\ell}\Psi^0_{\ell} =\tilde A^-_{\ell}\Psi^0_{\ell}= C^-_{\ell} \Psi^0_{\ell} =\tilde C^-_{\ell} \Psi^0_{\ell}= B^-_{\ell} \Psi^0_{\ell} = \tilde B^-_{\ell} \Psi^0_{\ell} = 0\, .$$ This state should be a particular case of (\[fs\]) invariant also under the $l_0$-reflection: $$\label{fs2} \Phi^0_{\ell}(\xi,\theta) = N (\cos \theta)^{1/2} (\sin \theta)^{1/2} (\cosh \xi)^{l_2+1/2}\sinh \xi,$$ thus, it has the label $\ell=(l_0=0,l_1=0,l_2)$, where $l_2<-5/2$. This point in the parameter space, for the example of Fig. \[fig1\], corresponds to the top vertex of the pyramid, from which all the other points displayed in the figure can be obtained with the help of raising operators. Such points correspond to a iur of the $so(4,2)$ algebra, including the series of iur’s of the $su(2,1)$ algebra mentioned in the previous section. Fixed the iur corresponding to $\ell=(0,0,l_2)$ such that $-7/2\leq l_2 <-5/2$, then, the points on the surface of the associated pyramid in the parameter space correspond to non-degenerated ground levels of their respective Hamiltonians. This ‘top’ pyramid includes other ‘inner’ pyramids, see Fig. 3, with vertexes $\ell_{n} = (0,0,l_2-2n)$. Each point on the surface of an inner pyramid associated to $\ell_n$ represents an $n$-excited level $(n+1)$-fold degenerated of the iur associated to $\ell$. Finally, we must remark that the same set of Hamiltonians and eigenstates can be described by ‘dual’ representations of $so(4,2)$ (or $su(2,1)$) by means of inverted pyramids with positive values of $l_2$ fixing the inverted vertex. ![It is shown two pyramids associated to the same iur of $so(4,2)$. The exterior with vertex $(0,0,-3)$ has points on its exterior faces which represent non-degenerated levels. The inner one has exterior faces corresponding to first excited double-degenerated levels.[]{data-label="fig3"}](fig3.epsf) Concluding remarks ================== In this work we have built a set of intertwining operators for a superintegrable system defined on a two-sheet hyperboloid and we have found that they close a non-compact $su(2,1)$ Lie algebra structure. By using the reflections operators of the system we have implemented these IO’s obtaining an $so(4,2)$ algebra. These IO’s lead to hierarchies of Hamiltonians described by points on planes ($su(2,1)$) or in the 3D space ($so(4,2)$), corresponding to the rank of the respective Lie algebra. We have shown how these IO’s can be very helpful in the characterization of the physical system by selecting separable coordinates, determining the eigenvalues and building eigenfunctions. We have also displayed the relation of eigenstates and eigenvalues with unitary representations of the $su(2,1)$ and $so(4,2)$ Lie algebras. In particular we have studied the degeneration problem as well as the number of bound states. Remark that such a detailed study of a ‘non-compact’ superintegrable system has not been realized till now, up to our knowledge. We have restricted to iur’s, but a wider analysis can be done for hierarchies associated to representations with a not well defined unitary character. The IO’s can also be used to find the second order integrals of motion for a Hamiltonian $H_{\ell}$ and their algebraic relations, which is the usual approach to (super) integrable systems. However, we see that it is much easier to deal directly with the IO’s, which are more elementary and simpler, than with constants of motion. Our program in the near future is the application of this method to wider situations. Besides, in principle, we can also adapt the method to classical versions of such systems. On this aspect we must remark that some symmetry procedures usually considered only for quantum systems can be extended in an appropriate way to classical ones [@KN07]. Acknowledgements. {#acknowledgements. .unnumbered} ================= Partial financial support is acknowledged to Junta de Castilla y León (Spain) under project VA013C05 and the Ministerio de Educación y Ciencia of Spain under project FIS2005-03989. References {#references .unnumbered} ========== [9]{} del Olmo M A, Rodríguez M A and Winternitz P 1993 [*J. Math. Phys.*]{} [**34**]{} 5118 Kalnins E G, Miller W and Pogosyan G S 1996 [*J. Math. Phys*]{}. [**37**]{} 6439 del Olmo M A, Rodríguez M A and Winternitz P 1996 [*Fortschritte der Physik*]{}, [**44**]{}, 91 Calzada J A, del Olmo M A and Rodríguez M A 1997 [*J. Geom. Phys.*]{} [**23**]{} 14 Calzada J A, del Olmo M A and Rodriguez M A 1999 [*J. Math. Phys.*]{} [**40**]{} 88 Calzada J A, Negro J, del Olmo M A and Rodríguez M A 1999 [*J. Math. Phys.*]{} [**41**]{} 317 Rañada M F 2000 [*J. Math. Phys.*]{} [**41**]{} 2121 Rañada M F and Santander M 2003 [*J. Math. Phys.*]{} [**44**]{} 2149 Evans N W 1990 [*Phys. Rev.*]{} [**41**]{} 5666 1990 [*Phys. Lett.*]{} [**147A**]{} 483 1991 [*J. Math. Phys.*]{} [**32**]{} 3369 Matveev V B and Salle M A 1991 [*Darboux Transformations and Solitons*]{} (Berlin: Springer) Kuru Ş, Teǧmen A and Verçin A 2001 [*J. Math. Phys.*]{} [**42**]{} 3344 Demircioǧlu B, Kuru Ş, Önder M and Verçin A 2002 [*J. Math. Phys.*]{} [**43**]{} 2133 Samani K A and Zarei M 2005 [*Ann. Phys.*]{} [**316**]{} 466 Fernández D J, Negro J and del Olmo M A 1996 [*Ann. Phys.*]{} [**252**]{} 386 Calzada J A, Negro J and del Olmo M A 2006 J. Math. Phys. [**47**]{} 043511 Barut A O, Inomata A and Wilson R 1987 [*J. Phys. A*]{} [**20**]{} 4075 1987 [*J. Phys. A*]{} [**20**]{} 4083 del Sol Mesa A, Quesne C and Smirnov Yu F 1998 [*J. Phys. A*]{} [**31**]{} 321 Kuru Ş and Negro J 2008 [*Ann. Phys.*]{} [**323**]{} 413

--- abstract: 'We discuss variance reduced simulations for an individual-based model of chemotaxis of bacteria with internal dynamics. The variance reduction is achieved via a coupling of this model with a simpler process in which the internal dynamics has been replaced by a direct gradient sensing of the chemoattractants concentrations. In the companion paper [@limits], we have rigorously shown, using a pathwise probabilistic technique, that both processes converge towards the same advection-diffusion process in the diffusive asymptotics. In this work, a direct coupling is achieved between paths of individual bacteria simulated by both models, by using the same sets of random numbers in both simulations. This coupling is used to construct a hybrid scheme with reduced variance. We first compute a deterministic solution of the kinetic density description of the direct gradient sensing model; the deviations due to the presence of internal dynamics are then evaluated via the coupled individual-based simulations. We show that the resulting variance reduction is *asymptotic*, in the sense that, in the diffusive asymptotics, the difference between the two processes has a variance which vanishes according to the small parameter.' author: - 'Mathias Rousset[^1]' - 'Giovanni Samaey[^2]' bibliography: - 'bib-papers.bib' title: 'Simulating individual-based models of bacterial chemotaxis with asymptotic variance reduction' --- [**MSC**]{}: 35Q80, 92B05, 65C35.\ [**Key words**]{}: asymptotic variance reduction, bacterial chemotaxis, velocity-jump process. Introduction ============ The motion of flagellated bacteria can be modeled as a sequence of run phases, during which a bacterium moves in a straight line at constant speed. Between two run phases, the bacterium changes direction in a tumble phase, which takes much less time than the run phase and acts as a reorientation. To bias movement towards regions with high concentration of chemoattractant, the bacterium adjusts its turning rate by increasing, resp. decreasing, the probability of tumbling when moving in an unfavorable, resp. favorable, direction [@Alt:1980p8992; @Stock:1999p8984]. Since many species are unable to sense chemoattractant gradients reliably due to their small size, this adjustment is often done via an intracellular mechanism that allows the bacterium to retain information on the history of the chemoattractant concentrations along its path [@Bren:2000p7499]. The resulting model, which will be called the “internal state” or “fine-scale” model in this text, can be formulated as a velocity-jump process, combined with an ordinary differential equation (ODE) that describes the evolution of an internal state that incorporates this memory effect [@Erban:2005p4247]. The probability density distribution of the velocity-jump process evolves according to a kinetic equation, in which the internal variables appear as additional dimensions. A direct deterministic simulation of this equation is therefore prohibitively expensive, and one needs to resort to a stochastic particle method. Unfortunately, a direct fine-scale simulation using stochastic particles presents a large statistical variance, even in the diffusive asymptotic regime, where the behavior of the bacterial density is known explicitly to satisfy a Keller-Segel advection-diffusion equation. Consequently, it is difficult to assess accurately how the solutions of the fine-scale model differ from their advection-diffusion limit in intermediate regimes. We refer to [@Horst1; @Horst2; @KelSeg70] for numerous historical references on the Keller-Segel equation, to [@ErbOthm04; @Erban:2005p4247] for formal derivations (based on moment closures) of convergence of the the velocity-jump process with internal state to the Keller-Segel equation, and to the companion paper [@limits] for a proof of this convergence based on probabilistic arguments. In this paper, we propose and analyze a numerical method to simulate individual-based models for chemotaxis of bacteria with internal dynamics with reduced variance. The variance reduction is based on a coupling technique (control variate): the main idea is to simultaneously simulate, using the same random numbers, a simpler, “coarse” process where the internal dynamics is replaced by a direct “gradient sensing” mechanism (see [@Alt:1980p7984; @Othmer:1988p7986; @Patlak:1953p7738] for references on such gradient sensing models). The probability density of the latter satisfies a kinetic equation without the additional dimensions of the internal state, and converges to a similar advection-diffusion limit, see e.g. [@Chalub:2004p7641; @Hillen:2000p4751; @Othmer:2002p4752; @limits]. More precisely, we consider the coupling of two velocity-jump processes with exactly the same advection-diffusion limit (for both processes, a proof of the latter convergence has been obtained in the companion paper [@limits]). The first one is the process with internal dynamics described by the system of equations , with jump rates satisfying and internal dynamics satisfying the two main assumptions and as well as other technical assumptions contained in Section \[sec:asympt\]. The second one is the process with direct gradient sensing , with jump rates satisfying  with . The same random numbers are used for the two processes in the definition of both the jump times and the velocity directions. The main contributions of the present paper are then twofold : - From a numerical point of view, we couple two systems of many particles consisting of different realizations of the fine-scale (internal state) and coarse (gradient sensing) processes, simulated with similar discretization schemes (see Section \[sec:coupl\_princ\]) and the same random numbers. Additionally, for the coarse model, also the continuum description is simulated on a spatial grid. Evolution of the fine-scale model is then evaluated on the grid at hand by adding the difference between the two coupled particle descriptions to the evolution of the coarse continuum description. This can be seen as using the coarse process as a control variate. As coupling is lost over time, a regular reinitialization of the two coupled particle systems is included in the algorithm. The asymptotic variance reduction method can be depicted using Table \[tab:table\] (see also Figure \[fig:pic\] in Section \[sec:method\]). \[tab:table\] Model Numerics ----------------------------------------------- ------------------------------ Internal state model–Individual description Monte-Carlo method Gradient sensing model–Individual description Monte-Carlo method (coupled) Gradient sensing model–Density description Grid method : Models and numerical methods used in the variance reduction algorithm. - From an analysis point of view, we rigorously prove, using probabilistic arguments, $L^p$-bounds on the position difference at diffusive times of both processes in terms of appropriate powers of the scaling parameter (Theorem \[thm:coupling\]). This analysis shows that the numerical variance reduction obtained by the method is *asymptotic* (the remaining statistical variance vanishes with the small scaling parameter), and provides upper bounds on the rate of the latter convergence. We argue in Section \[sec:sharp\] that this bound should be sharp in relevant regimes. The idea of asymptotic variance reduction is a general and very recent idea in scientific computing that appears when using hybrid Monte-Carlo/PDE (partial differential equation) methods. While the use of control variates is fairly common in Monte Carlo simulation, the only explicit attempt to develop asymptotic variance reduction in hybrid methods, up to our knowledge, can be found in [@DimPar08] and related papers in the context of the Boltzmann equation, and is based on an importance sampling approach (see also [@10.1063/1.1899210] and related papers). This paper is organized as follows. In Section \[sec:model\], we briefly review the mathematical models that we will consider, and summarize the results on their advection-diffusion limit that were obtained in the companion paper [@limits]. In Section \[sec:method\], we proceed to describe the coupled numerical method in detail. Section \[sec:variance\] is dedicated to a detailed analysis of the variance reduction of the coupled method. The main result is the following: up to a time shift perturbation, the variance of the difference between the two coupled process scales according to the small parameter of the diffusive asymptotics. We proceed with numerical illustrations in Sections \[sec:illustr\] and \[sec:appl\], and conclude in Section \[sec:concl\] with an outlook to future research. Particle-based models for bacterial chemotaxis \[sec:model\] ============================================================ In this paper, we directly present the models of interest in a nondimensional form, incorporating the appropriate space and time scales. For a discussion on the chosen scaling, we refer to the companion paper [@limits]. Model with internal state\[sec:model-internal\] ----------------------------------------------- We consider bacteria that are sensitive to the concentration of $m$ chemoattractants ${ \left(\rho_i(x)\right) }_{i=1}^m$, with $\rho_i(x) \geq 0$ for $x \in {\mathbb{R}}^d$. While we do not consider time dependence of chemoattractant via production or consumption by the bacteria, a generalization to this situation is straightforward, at least for the definition of the models and the numerical method. Bacteria move with a constant speed $v$ (run), and change direction at random instances in time (tumble), in an attempt to move towards regions with high chemoattractant concentrations. As in [@ErbOthm04], we describe this behavior by a velocity-jump process driven by some internal state $y \in \mathbb{Y}\subset{\mathbb{R}}^{n}$ of each individual bacterium. The internal state models the memory of the bacterium and is subject to an evolution mechanism attracted by a function $\psi: {\mathbb{R}}^m \to {\mathbb{R}}^n$ of the chemoattractants concentrations, $$S(x):= \psi(\rho_1(x),\ldots,\rho_m(x)) \in {\mathbb{R}}^n,$$ where $x$ is the present position of the bacterium. $S$ is assumed to be smooth with bounded derivatives up to order $2$. We refer to Section 2.2 in [@limits] for some typical choices; in the concrete example in this paper, we choose $m=n=1$, and $S(x)=\rho_1(x)$. By convention, the gradient of $S(x):{\mathbb{R}}^d \to {\mathbb{R}}^n$ is a matrix with dimension $$\nabla S(x) \in {\mathbb{R}}^{n\times d}.$$ We then denote the evolution of each individual bacterium position by  $ t \mapsto X_t, $ with normalized velocity $$\frac{{\mathrm{d}}X_t}{{\mathrm{d}}t} = \epsilon V_t, \qquad V_t\in \mathbb{V}={\mathbb{S}}^{d-1},$$ with ${\mathbb{S}}^{d-1}$ the unit sphere in ${\mathbb{R}}^d$. Hence, $V_t$ represents the direction and the parameter $\epsilon$ represents the size of the velocity. The evolution of the internal state is denoted by $t \mapsto Y_t$. The internal state adapts to the local chemoattractant concentration through an ordinary differential equation (ODE), $$\label{eq:internal} \frac{{\mathrm{d}}Y_t}{{\mathrm{d}}t}= F_{\epsilon}(Y_t,S(X_t)),$$ which is required to have a unique fixed point $y^*=S(x^*)$ for every fixed value $x^* \in {\mathbb{R}}^d$. We also introduce the deviations from equilibrium $z=S(x)-y$. The evolution of these deviations is denoted as $$t \mapsto Z_t = S(X_t) -Y_t.$$ The velocity of each bacterium is switched at random jump times $(T_n)_{n \geq 1}$ that are generated via a Poisson process with a time dependent rate given by $\lambda(Z_t)$, where $z \mapsto \lambda(z)$ is a smooth function satisfying $$\label{eq:ratebound} 0 < \lambda_{\rm min} \leq \lambda { \left(z\right) } \leq \lambda_{\rm max},$$ as well as, $$\label{eq:lin_rate} \lambda(z) = \lambda_0 - b^T z + c_\lambda{\mathcal{O}}{ \left( {\left | z\right |}^k \right) } ,$$ with $b \in {\mathbb{R}}^ n $, $k \geq 2$, and $c_\lambda >0$ is used to keep track of the non-linearity in the analysis. The new velocity at time $T_n$ is generated at random according to a centered probability distribution ${\mathcal M}(dv)$ with $\int v \, {\mathcal M}(dv) =0 $, typically $${\mathcal M}(dv) = \sigma_{{\mathbb{S}}^{d-1}}(dv),$$ where $\sigma_{{\mathbb{S}}^{d-1}}$ is the uniform distribution on the unit sphere. The resulting fine-scale stochastic evolution of a bacterium is then described by a left continuous with right limits (lcrl) process $$t \mapsto { \left(X_t, V_t, Y_t\right) } ,$$ which satisfies the following differential velocity-jump equation : $$\label{eq:process_noscale} \begin{cases} {\displaystyle}\dfrac{{\mathrm{d}}X_t}{{\mathrm{d}}t} = \epsilon V_t, \\[8pt] {\displaystyle}\dfrac{{\mathrm{d}}Y_t}{{\mathrm{d}}t} = F_{\epsilon}(Y_t,S(X_t)), \\[8pt] {\displaystyle}\int_{T_n}^{T_{n+1}} \lambda(Z_t) {\mathrm{d}}t = \theta_{n+1}, \qquad \text {with } Z_t := S(X_t)-Y_t ,\\[8pt] {\displaystyle}V_{t} = {\mathcal{V}}_{n} \quad \text{ for $t \in [T_{n},T_{n+1}]$ } , \end{cases}$$ with initial condition $X_0, {\mathcal{V}}_0 \in {\mathbb{R}}^d$, $Y_0 \in {\mathbb{R}}^n$ and $T_0 = 0$. In , ${ \left(\theta _n \right) }_{n \geq 1}$ denote i.i.d. random variables with normalized exponential distribution, and ${ \left( {\mathcal{V}}_n \right) }_{n \geq 1}$ denote i.i.d. random variables with distribution $ \mathcal M (dv)$. For concreteness, we provide a specific example, adapted from [@ErbOthm04], which will also be used later on to illustrate our results numerically. We consider $m=1$, i.e., there is only one chemoattractant $S(x)$ and the internal dynamics  reduces to a scalar equation $$\label{e:scalar-y} \frac{{\mathrm{d}}y(t)}{{\mathrm{d}}t} = \frac{S(x)-y}{\tau}=\frac{z}{\tau}.$$ For the turning rate $z\mapsto \lambda(z)$, we choose the following nonlinear strictly decreasing smooth function $$\label{eq:rate} \lambda(z)=2\lambda_0 { \left(\frac{1}{2}-\frac{1}{\pi}\arctan { \left(\frac{\pi}{2\lambda_0}z\right) }\right) }.$$ Asymptotic regimes and linearizations\[sec:asympt\] --------------------------------------------------- In the adimensional models -, a small parameter ${\epsilon}$ is introduced, which can be interpreted as the ratio of the typical time of random change of velocity direction to the typical time associated with chemoattractant variations as seen by the bacteria. In the present section, we present the main scaling assumptions on the internal dynamics that are necessary to consider the asymptotic regime ${\epsilon}\ll 1$, as well as the corresponding notation. Most of these assumptions were introduced in the companion paper [@limits], to which we refer for more details and motivation. Throughout the text, the Landau symbol ${\mathcal{O}}$ denotes a *deterministic* and globally Lipschitz function satisfying ${\mathcal{O}}(0)=0$. Its precise value *may vary* from line to line, may depend on all parameters of the model, but its Lipschitz constant is uniform in ${\epsilon}$. In the same way, we will denote generically by $C > 0$ a deterministic constant that may depend on all the parameters of the model except ${\epsilon}$. We assume that the ODE  driving the internal state is well approximated by a near equilibrium evolution equation in following sense : \[a:1\] We have $$\label{eq:tau_eps_def} {{F_{{\epsilon}}}}(y,s) = - {{\tau_{{\epsilon}}}}^{-1} (y-s) + {\epsilon}^{1-\delta}c_F{\mathcal{O}}({\left | s-y\right |}^2),$$ where $\delta >0$ and ${{\tau_{{\epsilon}}}}\in{\mathbb{R}}^{n\times n}$ is an invertible constant matrix. The constant $c_F$ is used to keep track of the dependance on the non-linearity of $F_{\epsilon}$ in the analysis. For technical reasons that are specific to the analysis of the time shift in the coupling (Lemma \[lem:DTestimbis\]), we need the following assumption on the scale of the non-linearity (this assumption was not necessary in the companion paper [@limits]): \[a:1bis\] Assume that ${\left | y-s\right |} = {\mathcal{O}}({\epsilon}^\delta)$, then there is a $\gamma > 0$ such that: $$\label{eq:taueps_hypbis} {\left | (y-s) - {{\tau_{{\epsilon}}}}{{F_{{\epsilon}}}}(y,s) \right |} = c_F{\mathcal{O}}( {\epsilon}^{1 + \gamma} ).$$ We now formulate the necessary assumptions on the scale of the matrix ${{\tau_{{\epsilon}}}}$ : \[a:2\] There is a constant $C >0$ such that for any $t \geq 0$, one has $$\label{eq:tau_scales} {\left\Vert{{\rm e}}^{- t {{\tau_{{\epsilon}}}}^{-1} }\right\Vert} \leq C {{\rm e}}^{- t {\epsilon}^{1-\delta}/C}.$$ \[a:3\] There is a constant $C >0$ such that $$\sup_{t \geq 0} {\left\Vertt{{\tau_{{\epsilon}}}}^{-1}{{\rm e}}^{- t {{\tau_{{\epsilon}}}}^{-1} } \right\Vert} \leq C {\epsilon}^{-1}.$$ Assumption \[a:2\] is used to ensure exponential convergence of the linear ODE with time scale at least of order ${\epsilon}^{1-\delta}$. Assumption \[a:3\] is necessary for technical reasons in the proof of Lemma \[lem:DTestim\] (which is given in [@limits]). Note that if ${{\tau_{{\epsilon}}}}^{-1}$ is symmetric and strictly positive with a lower bound of order ${\mathcal{O}}({\epsilon}^{1-\delta})$ on the spectrum, then Assumption \[a:2\] and Assumption \[a:3\] are satisfied. This implies that the slowest time scale of the ODE  is at worse of order ${\epsilon}^{1-\delta}$. Indeed, integration over time in Assumption \[a:2\] implies ${\left\Vert{{\tau_{{\epsilon}}}}\right\Vert} \leq C {\epsilon}^{\delta-1}$. Moreover, we assume that the solution of the ODE driving the internal state in satisfies the following long time behavior : \[ass:ODE\] Consider a path $t \mapsto S_t$ such that $ \sup_{t \in [0,+\infty] } {\left | \dfrac{ d S_t}{ dt} \right |}= {\mathcal{O}}( {\epsilon})$, and assume ${\left | Y_0 - S_0\right |} = {\mathcal{O}}({\epsilon}^\delta)$. Then the solution of $$\frac{ d Y_t}{ dt} = {{F_{{\epsilon}}}}(Y_t,s_t),$$ satisfies $$\sup_{t \in [0,+\infty] } {\left | Y_t - S_t \right |} = {\mathcal{O}}{ \left( {\epsilon}^\delta \right) }.$$ Assumption \[ass:ODE\] is motivated by the case of linear ODEs; indeed, in that case, Assumption \[ass:ODE\] is a consequence of Assumption \[a:2\] (see Section 3.3 in [@limits]). Recalling that $Z_t=S_t-Y_t$, we will assume in the remainder of the paper that the solution of  satisfies, $${\left | Z_0\right |} = {\mathcal{O}}({\epsilon}^\delta),$$ as well as $${\left | {{\tau_{{\epsilon}}}}^{-1} Z_0\right |} = {\mathcal{O}}( 1 ).$$ Then, under Assumption \[a:1\], Assumption \[a:2\], Assumption \[a:3\], and Assumption \[ass:ODE\], we obtain the bounds (see Section 3.3 in [@limits]) $$\label{eq:techbound} \sup_{t \in [0,+\infty] } {\left | {{\tau_{{\epsilon}}}}^{-1}{{\rm e}}^{- {{\tau_{{\epsilon}}}}^{-1} t } (S_t - Y_t) \right |} = {\mathcal{O}}{ \left( 1 \right) } ,$$ $$\sup_{t \in [0,+\infty]} { \left( {\left | Z_t\right |} \right) } = {\mathcal{O}}({\epsilon}^{\delta}),$$ as well as $$\sup_{t \in [0,+\infty]} { \left( {\left | {{\tau_{{\epsilon}}}}^{-1} {{\rm e}}^{- {{\tau_{{\epsilon}}}}^{-1} t} Z_t\right |} \right) } = {\mathcal{O}}(1) .$$ Finally, we will make use of the following notation. Denoting $c_S = {\left\Vert\nabla^{2} S\right\Vert}_{\infty},$ with $\nabla^2$ the Hessian, the error terms due to non-linearities of the problem will be handled through $$\label{eq:nl_err} \begin{cases} { {\rm nl} }({\epsilon}) := c_{F} {\epsilon}^{\delta} + c_S {\epsilon}+ c_{\lambda} {\epsilon}^{ k \delta-1} = {o}(1),\\ { {\rm nl_{2}} }({\epsilon}) := c_{F}( {\epsilon}^{\delta} + {\epsilon}^\gamma)+ c_S( {\epsilon}+ {\epsilon}^\delta) + c_{\lambda} {\epsilon}^{ k \delta-1} = {o}(1) . \end{cases}$$ Note that $ { {\rm nl} }({\epsilon}) \leq { {\rm nl_{2}} }({\epsilon})$ and $ { {\rm nl} }({\epsilon}) \sim { {\rm nl_{2}} }({\epsilon})$ if $\delta \leq 1$ and $\delta \leq \gamma$. Recall that $k \geq 2$ is defined in . Note that all the assumptions of the present section hold in the following case: (i) the ODE is linear ($F_{\epsilon}$ is linear), (ii) the associated matrix $\tau_{{\epsilon}}$ is symmetric positive and satisfies $$\tau_{{\epsilon}}^{-1} \geq C {\epsilon}^{1-\delta},$$ for some $\delta > 0$, (iii) the initial internal state is close to equilibrium in the sense that ${{\tau_{{\epsilon}}}}^{-1} Z_0={\mathcal{O}}(1)$. The assumptions of the present section can be seen as technical generalizations to non-linear ODEs with non-symmetric linear part. Model with direct gradient sensing (control process) ---------------------------------------------------- We now turn to a simplified, “coarse” model, in which the internal process (\[eq:internal\]), and the corresponding state variables, are eliminated. Instead, the turning rate depends directly on the chemoattractant gradient. This process will be called the *control process* since it will be used in Section \[sec:method\] as a control variate to perform variance reduced simulations of (\[eq:process\_noscale\]). The control process is a Markov process in position-velocity variables $$t \mapsto (X^c_t,V^c_t) ,$$ which evolves according to the following differential velocity-jump equations : $$\label{eq:cprocess} \system{ & \frac{{\mathrm{d}}X_t^c}{{\mathrm{d}}t} = {\epsilon}V_t^c , &\\ & \int_{T_n^c}^{T_{n+1}^c} \lambda^c_{\epsilon}(X_t^c , V_t^c) {\mathrm{d}}t = \theta_{n+1} , & \\ & V_{t}^c = {\mathcal{V}}_n \quad \text{ for $t \in [T_{n}^c,T^c_{n+1}]$ } , & }$$ with initial condition $X_0, {\mathcal{V}}_0 \in {\mathbb{R}}^d$. In (\[eq:cprocess\]), ${ \left(\theta _n \right) }_{n \geq 1}$ denote i.i.d. random variables with normalized exponential distribution, and ${ \left( {\mathcal{V}}_n \right) }_{n \geq 1}$ denote i.i.d. random variables with distribution $ {\mathcal M}(dv)$. The turning rate of the control process is assumed to satisfy $$\label{eq:rateboundc} 0 < \lambda_{\rm min} \leq \lambda^c_{{\epsilon}}(x,v) \leq \lambda_{\rm max},$$ as well as $$\label{eq:control_rate} \lambda^c_{\epsilon}(x,v) := \lambda_0 - {\epsilon}\; A^T_{\epsilon}(x) v + {\mathcal{O}}({\epsilon}^2).$$ Typically, $\lambda^c_{{\epsilon}}(x,v)$ is a function of $\nabla S(x)$, so that the model may describe a large bacterium that is able to directly sense chemoattractant gradients. When $m=n=1$, and the turning rate (\[eq:control\_rate\]) is proportional to $ \nabla S(x) v \in {\mathbb{R}}$, it can be interpreted as follows: the rate at which a bacterium will change its velocity direction depends on the alignment of the velocity with the gradient of the chemoattractant concentration $\nabla S(x)$, resulting in a transport towards areas with higher chemoattractant concentrations. In the companion paper [@limits], it is shown that, when the vector field $A_{\epsilon}(x) \in {\mathbb{R}}^{d}$ is given by the formula $$\label{eq:Adef} A_{\epsilon}(x) = b^T \frac{{{\tau_{{\epsilon}}}}}{\operatorname{{\rm Id}}+ \lambda_0 {{\tau_{{\epsilon}}}}} \nabla S(x),$$ asymptotic consistency with the internal state model is obtained in the limit when ${\epsilon}\to 0$ (see Section \[sec:model-hydro\]). Kinetic formulation and advection-diffusion limits\[sec:model-hydro\] --------------------------------------------------------------------- We now turn to the kinetic description of the probability distribution of the processes introduced above. The probability distribution density of the fine-scale process with internal state at time $t$ with respect to the measure ${\mathrm{d}}x \, {\mathcal M}( {\mathrm{d}}v ) \, {\mathrm{d}}y$ is denoted as $p(x,v,y,t)$, suppressing the dependence on ${\epsilon}$ for notational convenience, and evolves according to the Kolomogorov forward evolution equation (or master equation). In the present context, the latter is the following kinetic equation $$\label{e:kinetic} \partial_t p + {\epsilon}v \cdot \nabla_x p + { {\rm div}}_y { \left( F_{\epsilon}(x,y) p \right) } = \lambda { \left(S(x)-y\right) } { \left( R(p) - p \right) },$$ where $$R(p) := \int_{v \in {\mathbb{S}}^{d-1}} p(\cdot,v,\cdot) \, {\mathcal M}(dv)$$ is the operator integrating velocities with respect to ${\mathcal M}$, and $F(x,y)$ and $\lambda(z)$ are defined as in Section \[sec:model-internal\]. Similarly, the distribution density of the control process is denoted as $p^c(x,v,t)$, and evolves according to the kinetic/master equation: $$\label{e:kinetic_control} \partial_t p^c + {\epsilon}v \cdot \nabla_x p^c = { \left( R(\lambda^c_{\epsilon}p) - \lambda^c_{\epsilon}p \right) },$$ with $\lambda^c_{\epsilon}$ defined as in . The reader is referred to [@EthKur86] for the derivation of master equations associated to Markov jump processes. In the companion paper [@limits], we show, using probabilistic arguments, that, in the limit ${\epsilon}\to 0$, both the equation for the control process  and the equation for the process with internal state  converge to an advection-diffusion limit on diffusive time scales. Convergence is to be understood *pathwise*, in the sense of convergence of probability distribution on paths endowed with the uniform topology. Denote diffusive times by $${{\bar{t}}}:=t {\epsilon}^2,$$ and the processes considered on diffusive time scales as $$X^{{\epsilon}}_{{\bar{t}}}:=X_{{{\bar{t}}}/{\epsilon}^2}, \qquad X^{c,{\epsilon}}_{{\bar{t}}}:=X^c_{{{\bar{t}}}/{\epsilon}^2}.$$ For the control process, we then have the following theorem  : For ${\epsilon}\to 0$, the process ${{\bar{t}}}\mapsto X^{c,{\epsilon}}_{{\bar{t}}}$, solution of , converges towards an advection-diffusion process, satisfying the stochastic differential equation (SDE) $$\label{eq:cprocess_hydro} {\mathrm{d}}X_{\bar{t}}^{c,0} ={ \left(\frac{D A_0(X_{\bar{t}}^{c,0})}{\lambda_0} {\mathrm{d}}{{\bar{t}}}+ { \left(\frac{2D}{\lambda_0}\right) }^{1/2} {\mathrm{d}}W_{\bar{t}}\right) },$$ where ${{\bar{t}}}\mapsto W_{{\bar{t}}}$ is a standard Brownian motion, the parameters $\lambda_0$ and $A_0:=\lim_{{\epsilon}\to 0} A_{\epsilon}$ originate from the turning rate , and the diffusion matrix is given by the covariance of the Maxwellian distribution : $$\label{eq:D} D = \int_{{\mathbb{S}}^{d-1}} v \otimes v \, {\mathcal M}(dv) \in {\mathbb{R}}^{d \times d}.$$ In particular, this result implies that, at the level of the Kolomogorov/master evolution equation, the evolution of the position bacterial density, $$\label{eq:cdensity} n^{c,{\epsilon}}(x,{{\bar{t}}}):= n^c(x,{{\bar{t}}}/{\epsilon}^2):=\int_{{\mathbb{V}}}p^c(x,v,{{\bar{t}}}/{\epsilon}^2) \, {\mathcal M}({\mathrm{d}}v)$$ converges to the advection-diffusion equation $$\label{eq:cdens_hydro} \partial_{{\bar{t}}}n^{c,0} = \frac{1}{\lambda_0}{ {\rm div}}_x { \left( D \nabla_x n^{c,0} - D A_0(x) n^{c,0}\right) }$$ on diffusive time scales as $\epsilon \to 0$. In the same way, a standard probabilistic diffusion approximation argument can be used to derive the pathwise diffusive limit of the process with internal state  : For ${\epsilon}\to 0$, the process ${{\bar{t}}}\mapsto X^{{\epsilon}}_{{\bar{t}}}$, solution of , converges towards an advection-diffusion process, satisfying the stochastic differential equation (SDE) , where $A_0$ originates from $$\label{eq:control_field} A_0(x) = b^T \lim_{{\epsilon}\to 0} \frac{{{\tau_{{\epsilon}}}}}{\lambda_0 {{\tau_{{\epsilon}}}}+ \operatorname{{\rm Id}}} \nabla S(x),$$ in which, $b$, ${{\tau_{{\epsilon}}}}$, and $\lambda_0$ were introduced in - as parameters of the process with internal state, and $\operatorname{{\rm Id}}\in {\mathbb{R}}^{n\times n}$ is the identity matrix. Again, the diffusion matrix $D$ is given by the covariance of the Maxwellian distribution . Also here, convergence needs to be understood in terms of probability distribution on paths endowed with the uniform topology. Introducing the bacterial density of the process with internal state as $$n(x,t)= \int_{{\mathbb{Y}}}\int_{{\mathbb{V}}}p(x,v,y,t) {\mathcal M}({\mathrm{d}}v) {\mathrm{d}}y,$$ this implies that the evolution of $n$ converges to on diffusive time scales in the limit of $\epsilon \to 0$. Numerical method\[sec:method\] ============================== To simulate the process with internal state, solving the kinetic equation over diffusive time scales can be cumbersome, due to the additional dimensions associated with the internal state. The alternative is to use to stochastic particles. However, a particle-based simulation of equation is subject to a large statistical variance of the order $\mathcal{O}(N^{-1/2})$, where $N$ is the number of simulated particles. The asymptotic analysis shows that the position bacterial density approaches an advection-diffusion limit when ${\epsilon}\to 0$. Consequently, to accurately assess the deviations of the process with internal state as compared to its advection-diffusion limit (for small but non-vanishing (intermediate) values of ${\epsilon}$), the required number of particles needs to increase substantially with decreasing ${\epsilon}$, which may become prohibitive from a computational point of view. In this section, we therefore propose a hybrid method, based on the principle of control variates, that couples the process with internal dynamics with the control process, which is simulated simultaneously using a grid-based method. We first describe the variance reduction technique (Section \[sec:coupl\_princ\]). The analysis in Section \[sec:variance\] will reveal that the variance reduction is *asymptotic*, in the sense that the variance vanishes in the diffusion limit. To ensure this asymptotic variance reduction during actual simulations, one needs to ensure that the time discretization preserves the diffusion limits of the time-continuous process. An appropriate time discretization is discussed in Section \[sec:discr\]. Coupling and asymptotic variance reduction\[sec:coupl\_princ\] -------------------------------------------------------------- The proposed variance reduction technique is based on the introduction of a control variate that exploits a coupling between the process with internal state and the control process. While the idea of control variates for Monte Carlo simulation is already well known, see e.g. [@KloPla92] and references therein, the coupling that is proposed here is particular (we call it *asymptotic*), since the difference of the coupled processes, and hence the variance, vanishes with an estimable rate in the diffusion limit $\epsilon\to 0$, as will be shown in Section \[sec:variance\]. ### The control variates Let us first assume that we are able to compute the exact solution of the kinetic equation for the control process, , with infinite precision in space and time. The algorithm of asymptotic variance reduction is based on a coupling between an ensemble of realizations evolving according to the process with internal state , denoted as $$\left\{X^i_t,V^i_t,Y^i_t\right\}_{i=1}^N,$$ and an ensemble of realization of the control process , denoted as $$\left\{X^{i,c}_t,V^{i,c}_{t}\right\}_{i=1}^N.$$ We denote the empirical measure of the particles with internal state in position-velocity space as $$\mu_{{\bar{t}}}^{N}(x,v) = \frac{1}{N} \sum_{i=1}^{N} \delta_{X^i_{{{\bar{t}}}/{\epsilon}^2},V^i_{{{\bar{t}}}/{\epsilon}^2}},$$ and, correspondingly, the empirical measure of the control particles as $$\mu_{{\bar{t}}}^{c,N} = \frac{1}{N} \sum_{i=1}^{N} \delta_{X^{c,i}_{{{\bar{t}}}/{\epsilon}^2},V^{c,i}_{{{\bar{t}}}/{\epsilon}^2}}.$$ A coupling between the two ensembles is obtained by ensuring that both simulations use *the same random numbers* $(\theta_n)_{n\ge 1}$ and $({\mathcal{V}}_n)_{n\ge 0}$, which results in a strong correlation between $(X^i_{t},V^{i}_t)$ and $(X_t^{i,c},V^{c,i}_t)$ for each realization. Simultaneously, the kinetic equation for the control process is also solved using a deterministic method (which, for now, is assumed to be exact). We formally denote the corresponding semi-group evolution as $${{\rm e}}^{{{\bar{t}}}L^c}, \qquad \text{ with } L^c(p^c)= - {\epsilon}v \cdot \nabla_x p^c + { \left( R(\lambda^c_{\epsilon}p^c) - \lambda^c_{\epsilon}p^c \right) }.$$ Besides the two particle measures $\mu^N_{\bar{t}}$ and $\mu^{c,N}_{\bar{t}}$, we denote by $\overline{\mu}^N_{\bar{t}}$ the variance reduced measure, which will be defined by the algorithm below. Since, with increasing diffusive time, the variance of the algorithm increases due to a loss of coupling between the particles with internal state and the control particles, the variance reduced algorithm will also make use of a reinitialization time step ${\overline{\delta t}_{ri}}$, which is defined on the diffusive time scale. The corresponding time instances are denoted as $\bar{t}_n=n{\overline{\delta t}_{ri}}$ on the diffusive time scale, or equivalently, on the original time scale as $t_n=n{\overline{\delta t}_{ri}}/{\epsilon}^2$. Starting from an initial probability measure $\mu_0$ at time $t=0$, we sample $\mu_0$ to obtain the ensemble $\left\{X^i_t,V^i_t,Y^i_t\right\}_{i=1}^N$, corresponding to $\mu^N_0$, and then set $\mu^{c,N}_0:=\mu^N_0$, i.e., $X^{i,c}_{0}=X^{i}_{0}$ and $V^{i,c}_{0}=V^{i}_{0}$ for all $i=1,\ldots,N$. Furthermore, we set the variance reduced estimator as $\overline{\mu}_0^N:=\mu_0={ {\mathbb E} }(\mu_0^N)$. We then use the following algorithm to advance from $\bar{t}_n$ to $\bar{t}_{n+1}$, (see also Figure ) : ![\[fig:pic\] A schematic description of Algorithm \[algo\]. The dashed line represent the evolution of $N$ bacteria with internal state. The dotted line represent the coupled evolution of $N$ bacteria with gradient sensing, subject to regular reinitializations. The dashed-dotted line is computed according to a deterministic method simulating the density of the model with gradient sensing, and subject to regular reinitializations. The solid line is the variance reduced simulation of the internal state dynamics, and is computed by adding the difference between the particle computation with internal state, and the particle simulation with gradient sensing to the deterministic gradient sensing simulation. At each reinitialization step, the two simulations (deterministic and particles) of the gradient sensing dynamics are reinitialized to the values of their internal state simulation counterpart (as represented by the arrows).](figuren/fig-pic.pdf) \[algo\] At time $t_n$, we have that the particle measure $\mu_{\bar{t}_n}^{c,N}= \mu_{\bar{t}_n}^{N}$, and the variance reduced measure is given by $\overline{\mu}^N_{\bar{t}_n}$. To advance from time $ \bar{t}_n$ to $\bar{t}_{n+1}$, we perform the following steps : - Evolve the particles $\left\{X^i_t,V^i_t,Y^i_t\right\}_{i=1}^N$ from $t_n$ to $t_{n+1}$, according to , - Evolve the particles $\left\{X^{i,c}_t,V^{i,c}_{t}\right\}_{i=1}^N$ according to , using the same random numbers as for the process with internal state, - Compute the variance reduced evolution $$\label{eq:var_reduced_estimation} \overline{\mu}^N_{\bar{t}_{n+1}} = \overline{\mu}^N_{\bar{t}_n} {{\rm e}}^{{\overline{\delta t}_{ri}}/ {\epsilon}^2 L^c} + { \left(\mu_{\bar{t}_{n+1}}^{N} - \mu_{\bar{t}^{-}_{n+1}}^{c,N} \right) }.$$ - Reinitialize the control particles by setting $$X^{i,c}_{t_{n+1}}=X^{i}_{t_{n+1}}, \qquad V^{i,c}_{t_{n+1}}=V^{i}_{t_{n+1}}, \qquad i=1,\ldots,N,$$ i.e., we set the state of the control particles to be identical to the state of the particles with internal state. In formula , we use the symbol $\bar{t}^{-}_{n+1}$ to emphasize that the involved particle positions and velocities are those obtained *before* the reinitialization. An easy computation shows that the algorithm is unbiased in the sense that for any $n \geq 0$, $${ {\mathbb E} }{ \left( \overline{\mu}^N_{\bar{t}_n}\right) } = { {\mathbb E} }{ \left(\mu^N_{\bar{t}_n}\right) },$$ since the particles with internal dynamics are unaffected by the reinitialization, and, additionally, $${ {\mathbb E} }\mu^{c,N}_{\bar{t}_{n+1}}= { {\mathbb E} }\overline{\mu}^N_{{{\bar{t}}}_n}{{\rm e}}^{{\overline{\delta t}_{ri}}/ {\epsilon}^2 L^c}.$$ Moreover, the variance is controlled by the coupling between the two processes. Indeed, using the independence of the random numbers between two steps of Algorithm \[algo\], and introducing ${\varphi}$ as a position and velocity dependent test function), we get (according to the main result of Theorem \[thm:coupling\]) $$\begin{aligned} \operatorname{{\rm var}}( \overline{\mu}^N_{\bar{t}_n}({\varphi}) ) &= \sum_{k=1}^n { {\mathbb E} }{ \left(\mu_{}^{N}({\varphi}) - \mu_{\bar{t}_k^{-}}^{c,N}({\varphi}) \right) }^2 \nonumber \\ & \leq \sum_{k=1}^n \frac{{\left\Vert\nabla {\varphi}\right\Vert}_\infty}{N} { {\mathbb E} }{ \left( {\left | X_{\bar{t}_k/{\epsilon}^2} - X_{\bar{t}_k/{\epsilon}^2}^c \right |}^2 \right) } , \nonumber\end{aligned}$$ and thus $$\begin{aligned} \label{eq:MC_bound} \operatorname{{\rm var}}( \overline{\mu}^N_{\bar{t}_n}({\varphi}) ) & \leq C n\frac{ {\epsilon}+ {\epsilon}^\delta + { {\rm nl_{2}} }({\epsilon}) }{N},\end{aligned}$$ where in the last line, $C$ is independant of $n$, ${\epsilon}$, and $N$. In some generic situations (see Section \[sec:sharp\]), we can argue that the statistical error in Algorithm \[algo\] coming from the coupling is “sharp” with respect to the order in ${\epsilon}$. This means that the difference between the probability distribution of the model with internal state and the probability distribution of the model with gradient sensing is of the same order. This would imply that, with the asymptotic variance reduction technique, one is able to reliably assess the true deviation of the process with internal variables from the control process using a number of particles $N$ that is independent of ${\epsilon}$. ### Time-space discretization of the kinetic equation For both processes, the probability density distribution of the position and velocity can then be computed via binning in a histogram, or via standard kernel density estimation [@WScott:1992p8096; @WSilverman:1986p8151] using a kernel $K_h$, in which the chosen bandwidth $h$ can be based on the grid used for the deterministic simulation and on the data, for instance using the Silverman heuristic [@WSilverman:1986p8151]. This yields $$\label{eq:kde} \hat{p}_N(x,v,t) = \frac{1}{N h} \sum_{i=1}^{N} K_h(x - X^i_{t},v-V^i_{t}),$$ and similarly $$\label{eq:kdec} \hat{p}^c_N(x,v,t) = \frac{1}{N h} \sum_{i=1}^{N} K_h(x - X^{c,i}_{t},v-V^{c,i}_{t}).$$ The important point is that the solution $p^c(x,v,t)$ for the control process may be accurately approximated with a deterministic grid-based method, which ensures that $${\left | p^c(x,v,{{\bar{t}}}/ {\epsilon}^2)-p^c_{\rm grid}(x,v,{{\bar{t}}}/ {\epsilon}^2)\right |} = {\mathcal{O}}(\delta x^l) +{\mathcal{O}}(\delta t^k),$$ for some integers $k,l\ge 1$. In this text, we will perform the simulations in one space dimension using a third-order upwind-biased scheme, and perform time integration using the standard fourth order Runge–Kutta method, i.e., $l=3$ and $k=4$. Then, the unbiased nature of the variance-reduced estimator is conserved up to ${\mathcal{O}}(\delta x^l)+{\mathcal{O}}(\delta t^k)+{\mathcal{O}}(h)$ discretization errors. Time discretization of velocity-jump processes \[sec:discr\] ------------------------------------------------------------ When simulating equation , a time discretization error originates from the fact that the equation for the evolution of the internal state, and hence the evolution of the fluctuations $Z_t$, is discretized in time. This results in an approximation of the jump times $(T_n)_{n\ge 1}$, and hence of $X_t$. To retain the diffusion limits of the time-continuous process, special care is needed. We now briefly recall the time discretization procedure that was proposed and analyzed in the companion paper [@limits]. For ease of exposition, we consider the scalar equation for the internal state; generalization to nonlinear systems of equations is briefly discussed in [@limits]. Consider first the linear turning rate . In that case, we define a numerical solution $({X^{\delta t}}_t,{V^{\delta t}}_t,{Y^{\delta t}}_t)$ as follows. Between jumps, we discretize the simulation in steps of size $\delta t$ and denote by $({X^{\delta t}}_{n,k},{Z^{\delta t}}_{n,k})$ the solution at $t_{n,k}={T^{\delta t}}_n+k\delta t$. The numerical solution for $t\in[t_{n,k},t_{n,k+1}]$ is given by $$\label{e:discr-nonlin} \begin{cases} {X^{\delta t}}_t = {X^{\delta t}}_{n,k}+{\epsilon}{\mathcal{V}}_{n}\;(t-t_{n,k}) \\ {Z^{\delta t}}_t = \exp(-(t-t_{n,k}) {{\tau_{{\epsilon}}}}^{-1}) {Z^{\delta t}}_{n,k} + {\epsilon}{{\tau_{{\epsilon}}}}{ \left(\operatorname{{\rm Id}}-\exp(-(t-t_{n,k}){{\tau_{{\epsilon}}}}^{-1})\right) } \nabla S({X^{\delta t}}_{n,k}) \, {\mathcal{V}}_n. \end{cases}$$ We denote by $K\geq 0$ the integer such that the simulated jump time ${T^{\delta t}}_{n+1}\in [t_{n,K},t_{n,K+1}]$. To find ${T^{\delta t}}_{n+1}$, we first approximate the integral $ \int_{{T^{\delta t}}_n}^{{T^{\delta t}}_{n+1}}\lambda({Z^{\delta t}}_t){\mathrm{d}}t$ using $$\begin{aligned} \label{eq:time-disc-lin} \int_{{T^{\delta t}}_n}^{{T^{\delta t}}_{n+1}}\lambda({Z^{\delta t}}_t){\mathrm{d}}t &=\sum_{k=0}^{K-1}\int_{t_{n,k}}^{t_{n,k+1}}\lambda({Z^{\delta t}}_t){\mathrm{d}}t + \int_{t_{n,K}}^{{T^{\delta t}}_{n+1}}\lambda({Z^{\delta t}}_t){\mathrm{d}}t ,\end{aligned}$$ and then compute : $$\begin{gathered} \int_{t_{n,k}}^{t_{n,k+1}}\lambda({Z^{\delta t}}_t){\mathrm{d}}t= \lambda_0 \delta t - b^T { \left(\operatorname{{\rm Id}}- {{\rm e}}^{- {\delta t}{{{\tau_{{\epsilon}}}}^{-1}}} \right) }{{{\tau_{{\epsilon}}}}} {Z^{\delta t}}_{T_n} \\ - {\epsilon}b^T { \left(\delta t {{\tau_{{\epsilon}}}}-(\operatorname{{\rm Id}}-{{\rm e}}^{-\delta t {{\tau_{{\epsilon}}}}^{-1} } ){{\tau_{{\epsilon}}}}^2 \right) } \nabla S({X^{\delta t}}_{n,k}) {\mathcal{V}}_n .\end{gathered}$$ The jump time ${T^{\delta t}}_{n+1}$ can then be computed as the solution of $$\label{e:lin-Newton} \int_{t_{n,K}}^{{T^{\delta t}}_{n+1}}\lambda({Z^{\delta t}}_t){\mathrm{d}}t=\theta_{n+1}- \sum_{k=0}^{K-1}\int_{t_{n,k}}^{t_{n,k+1}}\lambda({Z^{\delta t}}_t){\mathrm{d}}t,$$ using a Newton procedure. It is shown in [@limits] that the results on the diffusion limit (as outlined in Section \[sec:model-hydro\]) are not affected by the discretization. We now consider a general nonlinear turning rate. We again discretize in time to obtain the time-discrete solution . The jump time ${T^{\delta t}}_{n+1}$ is now computed by linearizing in each time step, $$\label{eq:nonlinrate_disc} {\lambda^{\delta t}}{ \left({Z^{\delta t}}_t,{Z^{\delta t}}_{n,k}\right) }=\lambda({Z^{\delta t}}_{n,k})+ \frac{{\mathrm{d}}\lambda({Z^{\delta t}}_{n,k})}{{\mathrm{d}}z}{ \left({Z^{\delta t}}_t-{Z^{\delta t}}_{n,k}\right) }.$$ One then replaces equation by $$\begin{aligned} \int_{{T^{\delta t}}_n}^{{T^{\delta t}}_{n+1}}{\lambda^{\delta t}}({Z^{\delta t}}_t){\mathrm{d}}t &=\sum_{k=0}^{K-1}\int_{t_{n,k}}^{t_{n,k+1}}{\lambda^{\delta t}}({Z^{\delta t}}_t,{Z^{\delta t}}_{n,k}){\mathrm{d}}t + \int_{t_{n,K}}^{{T^{\delta t}}_{n+1}}{\lambda^{\delta t}}({Z^{\delta t}}_t,{Z^{\delta t}}_{n,K}){\mathrm{d}}t, \end{aligned}$$ and proceeds in the same way as for the linear case. Also in this case, the diffusive limit is recovered in an exact fashion for the time-discretized process. For the control process, there is no internal state. Hence, the only time dependence of turning rate $\lambda^c$ is due to the spatial variation of $\nabla S(x)$, which can be treated by discretizing the integral of $\lambda$ in the same way as above. Asymptotic variance reduction of the coupling\[sec:variance\] ============================================================= In this section, we show how the difference between the two coupled processes on diffusive time scales $\bar{t}=t/{\epsilon}^2$ behaves in the limit of ${\epsilon}\to 0$. We first recall some notation from the companion paper [@limits] (Section \[sec:not\]), after which we state and prove the main theorem (Section \[sec:proof\]). Notations and asymptotic estimates\[sec:not\] --------------------------------------------- The main theorem in Section \[sec:proof\] relies on following auxiliary definitions and lemmas that were given in the companion paper [@limits]: We denote by $m:{\mathbb{R}}\to {\mathbb{R}}^{n \times n}$ the function $$\label{eq:m(t)} m(t) := t {{\tau_{{\epsilon}}}}- { \left(\operatorname{{\rm Id}}- {\rm e}^{-t {{\tau_{{\epsilon}}}}^{-1}}\right) } {{\tau_{{\epsilon}}}}^2,$$ whose derivative is given by $$m'(t) = {{\tau_{{\epsilon}}}}{ \left(\operatorname{{\rm Id}}- {\rm e}^{-t {{\tau_{{\epsilon}}}}^{-1}} \right) }.$$ This function satisfies the following lemmas : Let $\theta$ be an exponential random variable of mean $1$. Then: $$\label{eq:av_m} { {\mathbb E} }{ \left( m\left( \frac{\theta}{\lambda_0} \right) \right) } = \frac{1}{\lambda_0} \frac{{{\tau_{{\epsilon}}}}}{\operatorname{{\rm Id}}+ \lambda_0 {{\tau_{{\epsilon}}}}}=A_{\epsilon}(x),$$ see equation . For all $t \in {\mathbb{R}}$, we have ${\left\Vert m'(t) \right\Vert} \leq t$, as well as ${\left\Vert m(t) \right\Vert} \leq t^2/2$. The difference between jump times is denoted as $$\Delta T_{n+1}^{c} := T_{n+1}^{c}- T_{n}^{c}, \qquad \Delta T_{n+1} := T_{n+1}- T_{n}.$$ The proofs of the asymptotic variance reduction make use of the following asymptotic expansions of the jump time differences of both processes : \[lem:dt-control\] The difference between two jump times of the control process satisfies $$\label{eq:deltaTc} \Delta T_{n+1}^{c} = \frac{\theta_{n+1}}{\lambda_0} + {\epsilon}\frac{\theta_{n+1}}{\lambda_0^2} A_{\epsilon}^T(X_{T^c_n}^c) {\mathcal{V}}_n + \theta_{n+1} {\mathcal{O}}({\epsilon}^2 ).$$ and \[lem:DTestim\] The jump time variations of the process with internal state can be written in the following form (${ {\rm nl} }({\epsilon})$ being defined by ) : $$\label{eq:deltaT} \Delta T_{n+1} = \Delta T_{n+1}^0 + {\epsilon}\Delta T_{n+1}^1 + (\theta_{n+1}^6+\theta_{n+1}){\mathcal{O}}{ \left({\epsilon}^2+ {\epsilon}\,{ {\rm nl} }({\epsilon})\right) } ,$$ where $$\begin{aligned} \label{eq:estimDT0} \Delta T_{n+1}^0 &= \frac{\theta_{n+1}}{\lambda_0} + \frac{b^T}{\lambda_0} m'(\Delta T_{n+1}^0 ) Z_{T_n} \nonumber \\ &= \frac{\theta_{n+1}}{\lambda_0} + \theta_{n+1} {\mathcal{O}}( {\epsilon}^\delta) \end{aligned}$$ and, correspondingly, $$\begin{aligned} \label{eq:estimDT1} \Delta T_{n+1}^1 &= \frac{1}{\lambda_0 - b^T {{\rm e}}^{- {\Delta T_{n+1}^0}{\tau^{-1}}} Z_{T_n} } b^T m(\Delta T_{n+1}^0) \nabla S(X_{T_n}) {\mathcal{V}}_{n}\nonumber \\ & = \frac{1}{\lambda_0} b^T m{ \left(\frac{\theta_{n+1}}{\lambda_0}\right) } \nabla S(X_{T_n}) {\mathcal{V}}_{n} + \theta_{n+1} {\mathcal{O}}({\epsilon}^\delta) .\end{aligned}$$ It is also useful to recall that according to -, the following hold : $${\left | \Delta T_{n+1}\right |} \leq C \theta_{n+1}, \qquad {\left | \Delta T_{n+1} ^c\right |} \leq C \theta_{n+1}.$$ Finally, we will also need a different expansion of the jump times of the process with internal state : \[lem:DTestimbis\] When using ${ {\rm nl} }_2({\epsilon})$ as defined by , the jump time variations of the process with internal state can be written in the following form : $$\label{eq:deltaTbis} \begin{split} & \Delta T_{n+1} = \frac{\theta_{n+1}}{\lambda_0} - \frac{b^T {{\tau_{{\epsilon}}}}}{\lambda_0}(Z_{T_{n+1}}-Z_{T_{n}})+ {\epsilon}\Delta T _{0 } b^T {{\tau_{{\epsilon}}}}\nabla S (X_{T_n}) {\mathcal{V}}_n \\ & + { \left(\theta_{n+1}^3+\theta_{n+1}\right) }{\mathcal{O}}{ \left( {\epsilon}\,{ {\rm nl_{2}} }({\epsilon}) + {\epsilon}^{1+\delta} \right) }. \end{split}$$ The proof is based on Duhamel integration of the ODE  on $[T_n,T_{n+1}]$, as in the proof of Lemma 4.7 of the companion paper [@limits]. Following equation (4.24) in [@limits], we get for $t\in [T_n,T_{n+1}]$ : $$\begin{aligned} \label{eq:key_Duhamel} Z_t = {{\rm e}}^{- { \left(t-T_n\right) }{{{\tau_{{\epsilon}}}}^{-1}}} Z_{T_n} + {\epsilon}{ \left(\operatorname{{\rm Id}}- {{\rm e}}^{- { \left(t-T_n\right) }{{{\tau_{{\epsilon}}}}^{-1}}} \right) }{{{\tau_{{\epsilon}}}}} \nabla S(X_{T_n}) {\mathcal{V}}_{n} + (\theta_{n+1}^2+\theta_{n+1}){\mathcal{O}}{ \left(c_S {\epsilon}^2+ c_F {\epsilon}^{1+\delta} \right) }.\end{aligned}$$ Multiplying both sides by $\tau_{\epsilon}$, recalling that Assumption \[a:2\] implies $\|\tau_{\epsilon}\|={\mathcal{O}}({\epsilon}^{\delta-1})$, and using Assumption \[a:1bis\], we get for $t\in [T_n,T_{n+1}]$ : $$\begin{aligned} \label{e:help} {{\tau_{{\epsilon}}}}Z_t = {{\tau_{{\epsilon}}}}{{\rm e}}^{- { \left(t-T_n\right) }{{{\tau_{{\epsilon}}}}^{-1}}} Z_{T_n} + {\epsilon}{ \left(\operatorname{{\rm Id}}- {{\rm e}}^{- { \left(t-T_n\right) }{{{\tau_{{\epsilon}}}}^{-1}}} \right) }{{{\tau_{{\epsilon}}}}^2} \nabla S(X_{T_n}) {\mathcal{V}}_{n} + (\theta_{n+1}^2+\theta_{n+1}){\mathcal{O}}{ \left(c_S {\epsilon}^{1+\delta}+ c_F {\epsilon}^{1+\gamma} \right) }.\end{aligned}$$ Moreover, integrating  on $t\in [T_n,T_{n+1}]$ yields $$\begin{aligned} \label{eq:key_Duhamelint} \begin{split} \int_{T_n}^{T_{n+1}} Z_t {\mathrm{d}}t &= { \left(\operatorname{{\rm Id}}- {{\rm e}}^{- { \left(T_{n+1}-T_n\right) }{{{\tau_{{\epsilon}}}}^{-1}}} \right) }{{{\tau_{{\epsilon}}}}} { \left(Z_{T_n} -{\epsilon}{{\tau_{{\epsilon}}}}\nabla S(X_{T_n}) {\mathcal{V}}_{n} \right) } \\ & \quad +{\epsilon}(T_{n+1}-T_n){{\tau_{{\epsilon}}}}\nabla S(X_{T_n}) {\mathcal{V}}_{n} + (\theta_{n+1}^3+\theta_{n+1}^2){\mathcal{O}}{ \left(c_S {\epsilon}^2+ c_F {\epsilon}^{1+\delta} \right) }. \end{split}\end{aligned}$$ Evaluating equation  at $t=T_{n+1}$, add adding equation , we get $$\begin{aligned} \int_{T_n}^{T_{n+1}} Z_t {\mathrm{d}}t + {{\tau_{{\epsilon}}}}Z_{T_{n+1}} = & {{\tau_{{\epsilon}}}}Z_{T_n} +{\epsilon}(T_{n+1}-T_n){{\tau_{{\epsilon}}}}\nabla S(X_{T_n}) {\mathcal{V}}_{n} \\ & \quad + { \left(\theta_{n+1}^3+\theta_{n+1}\right) }{\mathcal{O}}{ \left(c_S({\epsilon}^{1+\delta}+{\epsilon}^2 ) + c_F({\epsilon}^{1+\delta}+{\epsilon}^{1+\gamma}) \right) } . \end{aligned}$$ The estimates in Lemma $4.8$ from [@limits] (i.e., ${\left | \Delta T_{n+1}-\Delta T_{n+1}^0\right |} = (\theta_{n+1}^4+\theta_{n+1}){\mathcal{O}}({\epsilon}) $) yields $$\begin{aligned} \int_{T_n}^{T_{n+1}} Z_t {\mathrm{d}}t + {{\tau_{{\epsilon}}}}Z_{T_{n+1}} = & {{\tau_{{\epsilon}}}}Z_{T_n} +{\epsilon}\Delta T_{n+1}^0 {{\tau_{{\epsilon}}}}\nabla S(X_{T_n}) {\mathcal{V}}_{n} \\ & \quad + { \left(\theta_{n+1}^3+\theta_{n+1}\right) }{\mathcal{O}}{ \left(c_S({\epsilon}^{1+\delta}+{\epsilon}^2 ) + (c_F+1){\epsilon}^{1+\delta}+c_F{\epsilon}^{1+\gamma}) \right) } . \end{aligned}$$ Finally, plugging the last equation in the estimate  yields the result. Analysis of variance of the coupling\[sec:proof\] ------------------------------------------------- The present section is devoted to the proof of the following theorem: \[thm:coupling\] Assume the assumptions of Section \[sec:asympt\] hold, and that $k \delta > 1$ where $k\geq 2$ is defined in . Then the difference between the process with internal state and the coupling process satisfies : $$\label{eq:coupl_estim} { {\mathbb E} }{ \left({ \left( X_{{{\bar{t}}}/{\epsilon}^2} - X^c_{{{\bar{t}}}/{\epsilon}^2} \right) } ^ p\right) }^{1/p} = {\mathcal{O}}( {\epsilon}+ {\epsilon}^\delta + { {\rm nl_{2}} }({\epsilon}) ), \qquad p\ge 1,$$ where ${ {\rm nl_{2}} }({\epsilon})$ is defined in . The proof of Theorem \[thm:coupling\] relies on a number of steps that will be detailed by a series of lemmas. We will make use of the following random sequences, defined as the position of both processes after $n$ jumps, $$\label{eq:Xi} \Xi_n := X_{T_n}= X_0 + {\epsilon}\sum_{m=0}^{n-1}\Delta T _{m+1} {\mathcal{V}}_m, \qquad n\geq 0,$$ as well as $$\label{eq:Xic} \Xi^c_n := X_{T^c_n}^c= X_0 + {\epsilon}\sum_{m=0}^{n-1}\Delta T _{m+1}^c {\mathcal{V}}_m, \qquad n\geq 0.$$ We will also make use of the following random integers : \[def:N\_Nc\] The random integers $N \geq 1$ and $N^c \geq 1$ are uniquely defined by: $$T_N \leq {{\bar{t}}}/ {\epsilon}^2 < T_{N+1}, \qquad T^c_{N^c} \leq {{\bar{t}}}/ {\epsilon}^2 < T^c_{N^c+1} .$$ The first lemma bounds the difference at time ${{\bar{t}}}/ {\epsilon}^2$ between the two processes $X_{{{\bar{t}}}/{\epsilon}^2}$ and $X^c_{{{\bar{t}}}/{\epsilon}^2}$ by expressing it in terms of differences of positions and jump times of both processes after the same random number of jumps, $\Xi_n$ and $\Xi_n^c$ : The difference between the rescaled process with internal state $X^{\epsilon}_{{{\bar{t}}}}:= X_{{{\bar{t}}}/{\epsilon}^2}$ and the rescaled coupling process $X^{c,{\epsilon}}_{{{\bar{t}}}}:= X^c_{{{\bar{t}}}/{\epsilon}^2}$ satisfies: $$\begin{aligned} {\left | X^{\epsilon}_{{{\bar{t}}}} - X^{{\epsilon},c}_{{{\bar{t}}}}\right |} & \leq {\left | \Xi_{N}-\Xi_{N}^c\right |} + {\epsilon}{\left | T_{N}-T_{N}^c\right |}+{ \left(\theta_{N+1}+\theta_{N_c+1}\right) } {\mathcal{O}}({\epsilon}) \label{eq:coupl_11} \\ & \leq {\left | \Xi_{N^c}-\Xi_{N^c}^c\right |} + {\epsilon}{\left | T_{N^c}-T_{N^c}^c\right |}+{ \left(\theta_{N+1}+\theta_{N_c+1}\right) } {\mathcal{O}}({\epsilon}) \label{eq:coupl_12} \end{aligned}$$ Only one of the two estimates - is necessary in the remainder of the proof, but we detail both to highlight the symmetry. By definition, we have $$\begin{cases} X^{\epsilon}_{{{\bar{t}}}}= \Xi_N+ {\epsilon}({{\bar{t}}}/{\epsilon}^2-T_N){\mathcal{V}}_{N}, \\ X^{c,{\epsilon}}_{{{\bar{t}}}}= \Xi_{N^c}^c+ {\epsilon}({{\bar{t}}}/{\epsilon}^2-T_{N^c}^c){\mathcal{V}}_{N}, \end{cases}$$ so that by Definition \[def:N\_Nc\] of $N$ and $N^c$, and by realizing that ${\left | {{\bar{t}}}/{\epsilon}^2-T_{N^c}^c\right |} \leq C \theta_{N_c+1}$ and ${\left | {{\bar{t}}}/{\epsilon}^2-T_{N}\right |} \leq C \theta_{N+1}$, we get $$\label{eq:diffX} \begin{cases} {\left | X^{\epsilon}_{{{\bar{t}}}}- X^{c,{\epsilon}}_{{{\bar{t}}}}\right |}\leq {\left | \Xi_N-\Xi^c_N\right |}+\underbrace{{\left | \Xi^c_{N^c}-\Xi^c_{N}\right |}}_{(a)}+{ \left(\theta_{N+1}+\theta_{N_c+1}\right) } {\mathcal{O}}({\epsilon}) , \\ {\left | X^{\epsilon}_{{{\bar{t}}}}- X^{c,{\epsilon}}_{{{\bar{t}}}}\right |} \leq {\left | \Xi_{N^c}-\Xi^c_{N^c}\right |}+\underbrace{{\left | \Xi_{N^c}-\Xi_{N}\right |}}_{(b)}+ { \left(\theta_{N+1}+\theta_{N_c+1}\right) } {\mathcal{O}}({\epsilon}). \end{cases}$$ To analyze $(a)$, we consider three cases: - Then $(a)=0$. - Then $$\begin{aligned} \frac{(a)}{{\epsilon}} & = {\left | \sum_{n = N_c}^{N-1} \Delta T _{n+1}^c {\mathcal{V}}_n \right |} \\ & \leq T _{N}^c-T_{N_c}^c=\Delta T^c_{N^c+1} + T^c_ N - T^c_{N^c+1} \\ &\leq T^c_ N - T_{N} + C \theta_{N^c+1}, \end{aligned}$$ where in the last line we have used that, $\Delta T^c_{N^c+1}\le C\theta_{N^c+1}$, and, by definition, $ T_{N} \leq {{\bar{t}}}/ {\epsilon}^2 < T^c_{N^c+1}$. - Then $$\begin{aligned} \frac{(a)}{{\epsilon}} & = {\left | \sum_{n = N}^{N_c-1} \Delta T _{n+1}^c {\mathcal{V}}_n \right |} \\ & \leq T _{N_c}^c-T_{N}^c \\ &\leq T_{N+1} - T_{N}^c = \Delta T_{N+1} + T_N-T_N^c \leq T_ N - T^c_{N} + C \theta_{N+1} , \end{aligned}$$ where in the last line we have used that by definition $ T^c_{N^c} \leq {{\bar{t}}}/ {\epsilon}^2 \leq T_{N+1}$. This proves . By symmetry, we get: $$\frac{(b)}{{\epsilon}} \leq {\left | T_{N^c} - T^c_{N^c}\right |} + C{ \left(\theta_{N+1}+\theta_{N^c+1}\right) },$$ which yields . In the next lemma, we estimate the supremum of the difference of the position of both processes after $n$ jumps for $n\in [0,n_{\epsilon}]$. Let $n_{\epsilon}\geq 1$ be a deterministic integer verifying $n_{\epsilon}= {\mathcal{O}}({\epsilon}^{-2})$. Then $$\label{eq:estim_jumps} { {\mathbb E} }{ \left( \sup_{0 \leq n \leq n_{\epsilon}}{\left | \Xi_n-\Xi_n^c\right |}^p \right) } ^{1/p} = {\mathcal{O}}({\epsilon}+ {\epsilon}^\delta + { {\rm nl} }({\epsilon}) ).$$ First, using the estimates of jump times -, we can decompose the differences of positions of both processes as follows : $$\begin{aligned} \Xi_{n}-\Xi_{n}^c &= \sum_{m=0}^{n-1}{\epsilon}(\Delta T _ {m+1}-\Delta T _ {m+1} ^c) {\mathcal{V}}_{m} \\ & = \sum_{m=0}^{n-1} \alpha_{m+1} + d_{m+1} + r_{m+1}, \end{aligned}$$ where, by definition, $$\begin{cases} \alpha_{m+1} := {\epsilon}^2 \frac{{\mathcal{V}}_m}{\lambda_0^2} { \left(A_{\epsilon}(\Xi_{m})^T-A_{\epsilon}(\Xi_{m}^c)^T \right) } {\mathcal{V}}_m, \\ d_{m+1} :={\epsilon}^2\frac{{\mathcal{V}}_m}{\lambda_0^2} { \left(1-\theta_{m+1}\right) }A_{\epsilon}(\Xi_{m}^c)^T {\mathcal{V}}_m \\ \hspace{1.5cm} + {\epsilon}^2 \frac{{\mathcal{V}}_{m}}{\lambda_0^2} { \left( \lambda_0 b^T m{ \left(\frac{\theta_{m+1}}{\lambda_0}\right) } \nabla S(\Xi_m) - A_{\epsilon}(\Xi_{m})^T \right) } {\mathcal{V}}_{m} \\ \hspace{1.5cm} + {\epsilon}b^T m'(\Delta T_{m+1}^0 ) Z_{T_m} {\mathcal{V}}_m , \\ r_{m+1}:= { \left(\theta_{m+1}^6+\theta_{n+1}\right) } {\mathcal{O}}({\epsilon}^{3} + {\epsilon}^2 { {\rm nl} }({\epsilon}) ). \end{cases}$$ Since $A_{\epsilon}$ is Lipschitz, we have that $\alpha_{m+1} \leq C {\epsilon}^2 {\left | \Xi_{m}-\Xi_{m}^c\right |}$, and we can write $${\left | \Xi_{n}-\Xi_{n}^c\right |} \leq C{\epsilon}^2 \sum_{m=0}^{n-1} {\left | \Xi_{m}-\Xi_{m}^c\right |} + \sup_{0 \leq l \leq n_{{\epsilon}}}{\left | \sum_{m=1}^{l} d_{m}\right |}+ \sum_{m=1}^{n_{\epsilon}}{\left | r_{m}\right |},$$ for $n \leq n_{{\epsilon}}$, which yields $$\begin{aligned} {\left | \Xi_{n}-\Xi_{n}^c\right |} &\leq { \left(\sup_{0 \leq l \leq n_{{\epsilon}}}{\left | \sum_{m=1}^{l} d_{m}\right |}+ \sum_{m=1}^{n_{\epsilon}}{\left | r_{m}\right |}\right) } { \left(1+C {\epsilon}^2+ \dots + (C {\epsilon}^2)^{n_{\epsilon}}\right) }, \label{eq:key_xi}.\end{aligned}$$ Note that $${ \left(1+C {\epsilon}^2+ \dots + (C {\epsilon}^2)^{n_{\epsilon}}\right) } \leq C .$$ Now, we can remark that the discrete time process $ { \left(M_l\right) }_{l \geq 0}:={ \left( \sum_{m=1}^{l} d_{m} \right) }_{l \geq 0} $ is a martingale with respect to the filtration $\mathcal F _{l} = \sigma{ \left((\theta_{m},{\mathcal{V}}_{m-1}) \vert 1 \leq m \leq l \right) }$ with $M_0=0$, so that we can apply the Burkholder-Davies-Gundy upper bound (which is a simple consequence of Doob maximal inequality here, see [@Karatzasbook]), for some $p \geq 1$: $$\label{eq:BDG} { {\mathbb E} }{ \left(\sup_{0 \leq l \leq n_{\epsilon}} {\left | M_l\right |}^p\right) } \leq C_p n_{{\epsilon}}^{p/2 - 1} \sum_{l=1}^{n_{{\epsilon}}} { {\mathbb E} }{ \left({\left | M_l-M_{l-1}\right |}^p\right) }.$$ In the present case, since ${\left | Z_{T_n}\right |}={\mathcal{O}}({\epsilon}^\delta)$: $${ {\mathbb E} }{ \left({\left | d_l\right |}^p\right) } \leq C_p {\epsilon}^{p(1+\delta)}.$$ Moreover, $$\begin{aligned} { {\mathbb E} }{ \left({\left | \sum_{m=1}^{n_{\epsilon}}{\left | r_{m}\right |}\right |}^p\right) } &\leq n_{{\epsilon}}^{p - 1} \sum_{m=1}^{n_{\epsilon}}{ {\mathbb E} }{\left | r_{m}\right |}^p \\ & \leq C { \left( {\epsilon}+ { {\rm nl} }({\epsilon}) \right) }^p,\end{aligned}$$ and the result follows from . In the same way, we can estimate the difference of the $n_{\epsilon}$-th jump time of both processes : Let $n_{\epsilon}\geq 1$ a deterministic integer. Then $$\label{eq:estim_jumps_T} {\epsilon}\, { {\mathbb E} }{ \left( \sup_{0 \leq n \leq n_{\epsilon}}{\left | T_n-T_n^c\right |}^p \right) }^{1/p} \leq C_p { \left( {\epsilon}^\delta + { {\rm nl_{2}} }({\epsilon}) \right) }$$ First, using the estimates of jump times -, we can decompose the differences of jump times of both processes as follows : $$\begin{aligned} T_{n}-T_{n}^c &= \sum_{m=0}^{n-1} \Delta T _ {m+1}-\Delta T _ {m+1} ^c\\ & = \sum_{m=0}^{n-1} \delta_{m+1} + d_{m+1} + r_{m+1}, \end{aligned}$$ where, by definition, $$\begin{cases} \delta_{m+1} := - \frac{b^T {{\tau_{{\epsilon}}}}}{\lambda_0}(Z_{T_{m+1}}-Z_{T_{m}}), \\ d_{m+1} :={\epsilon}\Delta T _{0 } b^T {{\tau_{{\epsilon}}}}\nabla S (X_{T_n}) {\mathcal{V}}_n \\ \hspace{1.5cm} - {\epsilon}\frac{\theta_{n+1}}{\lambda_0^2} A_{\epsilon}^T(X_{T^c_n}^c) {\mathcal{V}}_n \\ r_{m+1}:= { \left(\theta_{m+1}^3+ \theta_{n+1} \right) } {\epsilon}{\mathcal{O}}({\epsilon}^\delta + { {\rm nl_{2}} }({\epsilon})). \end{cases}$$ We can write for $n \leq n_{{\epsilon}}$: $$\begin{aligned} {\left | T_{n}-T_{n}^c\right |} \leq {\left | \frac{b^T {{\tau_{{\epsilon}}}}}{\lambda_0}(Z_{T_{n}}-Z_{T_{0}}) \right |} + \sup_{0 \leq l \leq n_{{\epsilon}}}{\left | \sum_{m=1}^{l} d_{m}\right |} + \sum_{m=1}^{n_{\epsilon}}{\left | r_{m}\right |}.\end{aligned}$$ Now, we can remark that the discrete time process $ { \left(M_l\right) }_{l \geq 0}:={ \left( \sum_{m=1}^{l} d_{m} \right) }_{l \geq 0} $ is a martingale with respect to the filtration $\mathcal F _{l} = \sigma{ \left((\theta_{m},{\mathcal{V}}_{m-1}) \vert 1 \leq m \leq l \right) }$ with $M_0=0$. Using the Burkholder-Davies-Gundy inequality , together with: $${ {\mathbb E} }{ \left({\left | m_{m}\right |}^p\right) } \leq C {\epsilon}^{p\delta},$$ yields $${ {\mathbb E} }{ \left( \sup_{0 \leq l \leq n_{{\epsilon}}}{\left | \sum_{m=1}^{l} d_{m}\right |}^p \right) }^{1/p} \leq C {\epsilon}^{\delta-1}.$$ Using: $${\left | \frac{b^T {{\tau_{{\epsilon}}}}}{\lambda_0}(Z_{T_{n}}-Z_{T_{0}}) \right |} = {\mathcal{O}}({\epsilon}^{2 \delta-1}) < {\mathcal{O}}({\epsilon}^{\delta-1}),$$ and plugging in the rest term, we finally get . We can now conclude with the proof of Theorem \[thm:coupling\] : We consider a given integer $n_{{\epsilon}} \geq 1$. We decompose the difference between both processes as $$\label{eq:proof1} {\left | X^{\epsilon}_{{{\bar{t}}}}- X^{c,{\epsilon}}_{{{\bar{t}}}}\right |} = {\left | X^{\epsilon}_{{{\bar{t}}}}- X^{c,{\epsilon}}_{{{\bar{t}}}}\right |} {{\bf 1 }}_{N\leq n_{{\epsilon}}}+{\left | X^{\epsilon}_{{{\bar{t}}}}- X^{c,{\epsilon}}_{{{\bar{t}}}}\right |} {{\bf 1 }}_{N > n_{{\epsilon}}},$$ and use  to write $$\label{eq:proof2} {\left | X^{\epsilon}_{{{\bar{t}}}}- X^{c,{\epsilon}}_{{{\bar{t}}}}\right |} {{\bf 1 }}_{N\leq n_{{\epsilon}}} \leq \sup_{0 \leq n \leq n_{\epsilon}}{\left | \Xi_n-\Xi_n^c\right |} +{\epsilon}\sup_{0 \leq n \leq n_{\epsilon}}{\left | T_n-T_n^c\right |} + { \left(\theta_{N+1}+\theta_{N_c+1}\right) } {\mathcal{O}}({\epsilon}).$$ Using now the estimates  and in , we get $$\begin{aligned} \label{eq:estim_jumps_tot} { {\mathbb E} }{ \left( {\left | X^{\epsilon}_{{{\bar{t}}}}- X^{c,{\epsilon}}_{{{\bar{t}}}}\right |}^p {{\bf 1 }}_{N \leq n_{{\epsilon}}} \right) }^{1/p}& \leq C ( {\epsilon}+ {\epsilon}^\delta + { {\rm nl_{2}} }({\epsilon}) )\\ & \qquad + C{\epsilon}{ {\mathbb E} }(\theta_{N}^p+\theta_{N^c}^p)^{1/p} \nonumber \\ & = {\mathcal{O}}( {\epsilon}+ {\epsilon}^\delta + { {\rm nl_{2}} }({\epsilon}) ),\end{aligned}$$ where in the last line we have used a technical lemma (Lemma \[lem:tech1\]). It remains to control the probability of the event $\{ N > n_{\epsilon}\}$. To this purpose, we consider $t \mapsto N_t \in \mathbb N$ the Poisson process uniquely defined by: $$\sum_{n=1}^{N_t} \theta_n \leq t < \sum_{n=1}^{N_t +1} \theta_n.$$ By definition of $N$ and $N_c$, there is a constant $C$ such that $N_{C{{\bar{t}}}/ {\epsilon}^2} \geq N$ and $N_{C{{\bar{t}}}/ {\epsilon}^2} \geq N^c$, so that we can write: $$\label{eq:proof3} {\left | X^{\epsilon}_{{{\bar{t}}}}- X^{c,{\epsilon}}_{{{\bar{t}}}}\right |} \leq {\epsilon}C { \left(\sum_{n=1}^{N_{C{{\bar{t}}}/ {\epsilon}^2}+1}\theta_{n}\right) }{{\bf 1 }}_{N_{C{{\bar{t}}}/ {\epsilon}^2} > n_{{\epsilon}}}.$$ Let us choose $n_{{\epsilon}} \geq 4 {{\rm e}}C{{\bar{t}}}/ {\epsilon}^2 $, a choice that satisfies $n_{{\epsilon}}={\mathcal{O}}({\epsilon}^{-2})$. A second technical lemma (Lemma \[lem:tech2\]) implies that $$\label{eq:proof4} { {\mathbb E} }{ \left( {\left | X^{\epsilon}_{{{\bar{t}}}}- X^{c,{\epsilon}}_{{{\bar{t}}}}\right |}^p {{\bf 1 }}_{N > n_{{\epsilon}}} \right) } \leq {\epsilon}C_p 2^{-1/(C{\epsilon}^2)},$$ so that the contribution of the latter to the final estimate  is negligible, since it is exponentially small with respect to ${\epsilon}^{-2}$. The proof is complete. Discussion about the sharpness of the coupling {#sec:sharp} ---------------------------------------------- One may ask about the “sharpness” of the precise estimates of the coupling, given mainly by Theorem \[thm:coupling\]; (see also  (position shift), and (time shift)). Let us discuss the case where $c_F=0$, $ \delta \in [1/k , 1/(k-1)]$. In this case the dominant error term in the jump times expansion of the internal state model in Lemma $4.8$ in [@limits] is due to the non-linearity of the turning rate: $c_\lambda {\epsilon}^{k \delta -1}$. Since this term is due to the internal state mechanism, it can be conjectured that the difference between the density of the internal state model and gradient sensing model will be at least of this order. On the other hand, the coupling estimate in Theorem \[thm:coupling\] is controlled by the same term: $c_\lambda {\epsilon}^{k \delta -1}$ , suggesting the sharpness of the latter. Numerical illustration\[sec:illustr\] ===================================== In this section, we demonstrate the validity of the analysis above. For the numerical experiments, we restrict ourselves to one space dimension. For the process with internal state, we use internal dynamics that are given by the scalar cartoon model . In all experiments, $\tau$ is chosen independently of ${\epsilon}$ so that $\delta=1$. The corresponding turning rate is given by with $\lambda_0=1$. For the control process, we choose the linear turning rate , with parameters such that the two processes have the same diffusion limit; in particular, $b=1$. The physical domain is $x\in [0,20]$, and we use reflecting boundary conditions, [*i.e*]{}. the bacterial velocity is reversed when $x=0$ or $x=20$. We fix a scalar bimodal chemoattractant concentration field $$\label{eq:chemoattractant} S(x)=\alpha\left(\exp\left(-\beta\left(x-\xi\right)^2\right)+\exp\left(-\beta\left(x-\eta\right)^2\right)\right),$$ in which $\alpha=5$, $\beta=1$, $\xi=7.5$ and $\eta=12.5$. #### Single bacterium as a function of time. In a first experiment, we simulate a single bacterium evolving according to the fine-scale model (\[eq:process\_noscale\]), in which we set the parameters to ${\epsilon}=0.2$ and $\tau=1$. We compare this evolution to that of a bacterium that satisfies the corresponding control process . Both simulations are performed using the same random numbers starting from the initial condition $X_0=X_0^c=8$ and $V_0=V_0^c=+1$; the bacterium with internal state starts from $Y_0=S(X_0)$. The time step $\delta t=0.1$. The results are shown in figure \[fig:one-particle\]. ![\[fig:one-particle\]Evolution of a single bacterium evolving according to the fine-scale process (\[eq:process\_noscale\]) (solid line) and the corresponding control process (\[eq:cprocess\]) (dashed) on short (left) and long (right) time-scales.](figuren/fig-one-particle) We see a very good coupling initially, which degrades over time. Note the time shift in the short time picture. In the long time picture, coupling is completely lost at some point. #### Expectation and variance as a function of $\epsilon$ and ${{\bar{t}}}$. Next, we repeat the experiment using $N=10000$ particles and compute the empirical mean and variance of the coupling, i.e.,  $$E\left(\left|X_{{{\bar{t}}}}-X_{{{\bar{t}}}}^c\right|\right)=\frac{1}{N}\sum_i^N \left(\left|X^i_{{{\bar{t}}}}-X^{i,c}_{{{\bar{t}}}}\right|\right), \;\;\; \text{resp., } \;\;\; E\left((X_{{{\bar{t}}}}-X_{{{\bar{t}}}}^c)^2\right)=\frac{1}{N}\sum_i^N \left(X^i_{{{\bar{t}}}}-X^{i,c}_{{{\bar{t}}}}\right)^2.$$ As fine-scale parameters, we choose $\tau=1$, $\lambda_0=b=1$ and several values of ${\epsilon}$. The chemoattractant concentration is again given as , now with $\alpha=\beta=1$, $\xi=7.5$ and $\eta=12.5$. The time interval for the computation is $t\in [0,30/{\epsilon}^2]$. ![\[fig:eps-long\] Empirical mean (left) and variance (right) of the difference between the fine-scale process (\[eq:process\_noscale\]) and the corresponding control process (\[eq:cprocess\]) as a function of ${\epsilon}$ for different values of the reporting time $T$. The theoretical slope is indicated with a dashdotted line. The sample size is $N=10000$.](figuren/fig-eps-long) Figure \[fig:eps-long\] shows the dependence in ${\epsilon}$ of the coupling, by plotting the empirical mean and variance defined above as a function of ${\epsilon}$ for different values of the reporting time $T$. The results shown in figure \[fig:eps-long\] are in clear accordance with the theoretical slope predicted by the asymptotic analysis. ![\[fig:T-long\]Evolution of the empirical mean (left) and variance (right) of the difference between the fine-scale process (\[eq:process\_noscale\]) and the corresponding control process (\[eq:cprocess\]) as a function of ${{\bar{t}}}=t/{\epsilon}^2$ for different values of ${\epsilon}$. The sample size is $N=10000$. ](figuren/fig-T-long) Figure \[fig:T-long\] shows the mean and variance of the coupling difference as a function of time. The time dependence of the variance has not been analyzed mathematically in Section \[sec:variance\]; the specific behaviour viewed in figure \[fig:T-long\] and is probably due to: (i) sufficiently short diffusive times; (ii) the specific double-well form of the chemoattractant potential. #### Expectation and variance as a function of $\tau$. Finally, in a last experiment, we illustrate the dependence on $\tau$. To this end, we again simulate $N=10000$ particles choosing $\lambda_0=1$, ${\epsilon}=0.1$ and $X_0=7.5$, for different values of $\tau$. The results in figure \[fig:tau-long\] show that the variance quickly increases with $\tau$ until it reaches a plateau for $\tau>1$. ![\[fig:tau-long\] Empirical mean (left) and variance (right) of the difference between the fine-scale process (\[eq:process\_noscale\]) and the corresponding control process (\[eq:cprocess\]) as a function of $\tau$ for different values of the reporting time $T$. The sample size is $N=10000$.](figuren/fig-tau-long) Simulation on diffusive time scales\[sec:appl\] =============================================== In this section, we consider a simulation of the density of an ensemble of particles, with and without variance reduction and/or reinitialization, as described in Section \[sec:coupl\_princ\]. We again restrict ourselves to one space dimension, with domain $x\in [0,20]$ and periodic boundary conditions. In this case, the kinetic equation corresponding to the control process reduces to the system $$\label{eq:cprocess_hydro1d} \system{ & \partial_t p^c_++\epsilon\partial_x p^c_+=-\frac{\lambda^c(x,+1)}{2}p_+^c+ \frac{\lambda^c(x,-1)}{2}p_-^c &\\ & \partial_t p^c_--\epsilon\partial_x p^c_-=\frac{\lambda^c(x,+1)}{2}p_+^c -\frac{\lambda^c(x,-1)}{2}p_-^c & }.$$ of two PDEs, which is straightforward to simulate using finite differences. We fix the chemoattractant concentration field as , with parameters $\alpha=2$, $\beta=1$, $\xi=7.5$ and $\eta=12.5$. For the internal dynamics, the same model (-, -) is used. The parameters are ${\epsilon}=0.5$, $\lambda_0=1$, $\tau=1$, $\delta t=0.1$. All simulations are performed with $N=5000$ particles. The initial positions are uniformly distributed in the interval $x\in[13,15]$; the initial velocities are chosen uniformly, i.e., each particle has an equal probability of having an initial velocity of $\pm{\epsilon}$. The initial condition for the internal variable is chosen to be in local equilibrium, i.e., $Y_0^i=S(X^i_0)$. The initial positions and velocities of the control particles are chosen to be identical. We discretize the continuum description on a mesh with $\Delta x = 0.1$ using a third-order upwind-biased scheme, and perform time integration using the standard fourth order Runge–Kutta method with time step $\delta t_{pde}= 10^{-1}$. The initial condition is given as $$p^+(x,0)=p^-(x,0)=\begin{cases} 0.25 , & {x\in [13,15],} \\ 0, & \text{otherwise.} \end{cases}$$ #### Simulation without variance reduction. First, we simulate both stochastic processes up to time $\bar{t}=50$ ($t=50/{\epsilon}^2$) and estimate the density of each of these processes $\hat{n}_N(x,{{\bar{t}}})$, resp. $\hat{n}^c_N(x,{{\bar{t}}})$, without variance reduction. The density is obtained via binning in a histogram, in which the grid points of the deterministic simulation are the centers of the bins. Figure \[fig:no-variance-reduction\] (left) ![\[fig:no-variance-reduction\] Bacterial density as a function of space at $t=50/{\epsilon}^2$ without variance reduction. Left: one realization. Right: mean over $100$ realizations and $95\%$ confidence interval. The solid line is the estimated density from a particle simulation using the process with internal state; the dashed line is estimated from a particle simulation using the control process. Both used $N=5000$ particles. The dotted line is the solution of the deterministic PDE (\[eq:cdensity\]). ](figuren/fig-no-variance-reduction) shows the results for a single realization. We see that, given the fluctuations on the obtained density, it is impossible to conclude on differences between the two models. This observation is confirmed by computing the average density of both processes over $100$ realizations. The mean densities are shown in figure \[fig:no-variance-reduction\] (right), which also reveals that the mean density of the control process is within the $95\%$ confidence interval of the process with internal state. Both figures also show the density that is computed using the continuum description, which coincides with the mean of the density of the control particles. #### Simulation with variance reduction. Next, we compare the variance reduced estimation with the density of the control PDE. We reinitialize the control particles after each coarse-scale step, i.e., each $k$ steps of the particle scheme, where $k \delta t = \delta t_{pde}$, (here $k=1$). The results are shown in figure \[fig:var-red-reinit\]. ![\[fig:var-red-reinit\] Bacterial density as a function of space at $t=50/{\epsilon}^2$ with variance reduction and reinitialization. Left: variance reduced density estimation of one realization with $N=5000$ particles (solid) and deterministic solution for the control process (\[eq:cdensity\]) (dashed). Right: mean over $100$ realization and $95\%$ confidence interval (solid) and the deterministic solution for the control process (\[eq:cdensity\]) (dashed).](figuren/fig-variance-reduction) We see that, using this reinitialization, the difference between the behaviour of the two processes is visually clear from one realization (left figure). Also, the resulting variance is such that the density of the control PDE is no longer within the $95\%$ confidence interval of the variance reduced density estimation (right figure). We see that there is a significant difference between both models: the density corresponding to the control process is more peaked, indicating that bacteria that follow the control process are more sensitive to sudden changes in chemoattractant gradient. This difference can be interpreted from the fact that the bacteria with internal state do not adjust themselves instantaneously to their environment, but instead with a time constant $\tau$. #### Simulation with variance reduction but without reinitialization. We also compare the variance reduced estimation with the density of the control PDE without performing any reinitialization of the control process to restore the coupling. ![\[fig:var-red-noreinit\] Bacterial density as a function of space at $t=50/{\epsilon}^2$ with variance reduction. Left: variance reduced density estimation of one realization with $N=5000$ particles (solid) and deterministic solution for the control process (\[eq:cdensity\]) (dashed). Right: mean over $200$ realization and $95\%$ confidence interval (solid) and the deterministic solution for the control process (\[eq:cdensity\]) (dashed).](figuren/fig-var-red-noreinit) Figure \[fig:var-red-noreinit\] (right) shows the mean estimation of the density over $200$ realizations, as well as the density of the control process. As evidenced by the error bars, simulating a single realization of the process, even with variance reduction, is not able to reliably reveal this difference. This phenomenon is due to the degeneracy of the coupling on long diffusive times. This is also illustrated in figure \[fig:var-red-noreinit\] (left), which compares the variance reduced density estimation of a single realization with the density of the control PDE. Note that, as predicted by the analysis, the variance is much larger in regions where $\nabla S(x)$ is large. Finally, we repeat the experiment with a larger chemoattractant gradient. We again choose the chemoattractant field as , with the same parameters as above, except that we now take $\alpha=5$. We also use the same discretization parameters as above. The variance reduced estimation , as well as the density of the control PDE, are shown in figure \[fig:var-red-reinit-alpha5\]. ![\[fig:var-red-reinit-alpha5\] Bacterial density as a function of space at $t=50/{\epsilon}^2$ with variance reduction and reinitialization. Left: variance reduced density estimation of one realization with $N=5000$ particles (solid) and deterministic solution for the control process (\[eq:cdensity\]) (dashed). Chemoattractant given by with $\alpha=5$. Right: mean over $100$ realization and $95\%$ confidence interval (solid) and the deterministic solution for the control process (\[eq:cdensity\]) (dashed).](figuren/fig-variance-reduction-alpha5) We see that, although the difference between the process with internal state and the control process becomes much larger, the variance is still very well controlled by the coupling. Conclusions and discussion\[sec:concl\] ======================================= In this paper, we studied the simulation of stochastic individual-based models for chemotaxis of bacteria with internal state. We have used a coupling with a simpler, direct gradient sensing model with the same diffusion limit to obtain an “asymptotic” variance reduction. Throughout this work, we have assumed that the computation of the bacterial density using the kinetic description of the direct gradient sensing model can be performed accurately with a grid-based method. Note that in higher spatial dimensions, this may become cumbersome, and the grid-based computation should be carried out at the level of the advection-diffusion description (or any desired moment system) only, using a second level of coupling between the velocity-jump process associated with the kinetic description, and the stochastic differential equation associated with the advection-diffusion limit. This is left for future work. The current algorithm allows to explore the differences between the fine-scale process with internal state and the simpler, coarse process *using a fixed number of particles, independently of the small parameter ${\epsilon}$*. However, the computational cost to simulated an individual particle over diffusive time-scales becomes very large for ${\epsilon}\to 0$. In future work, we will therefore also study the development of truly “asymptotic preserving” schemes in the diffusion asymptotics, in the sense that the computational cost of the simulation of processes is independent of the small parameter ${\epsilon}$. This will require to deal with two kinds of difficulties: (i) We will need to use an asymptotic preserving method to solve the density evolution of the control model on the grid; and (ii) We will need to extrapolate forward in time the state of the fine-scale simulation. The first difficulty implies that, instead of solving the full kinetic equation associated with the control process, we only solve its diffusion limit. This may imply the use of a second level of variance reduction, coupling the control velocity jump process and its limiting drift-diffusion process. The second difficulty, extrapolation in time, is related to the equation-free [@KevrGearHymKevrRunTheo03; @Kevrekidis:2009p7484] and HMM [@EEng03; @E:2007p3747] types of methodologies; ideas of this type can be traced back to Erhenfest [@Ehrenfest:1990p9192]. One approach is to use a *coarse projective integration* method [@GearKevrTheo02; @KevrGearHymKevrRunTheo03]. One then extrapolates the bacterial density over a projective time step on diffusive time scales, after which the projected density needs to be *lifted* to an ensemble of individual bacteria. For such methods, besides the effect of extrapolation on the variance of the obtained results, also the effects of reconstructing the velocities and internal variables have to be systematically studied. Acknowledgements {#acknowledgements .unnumbered} ================ The authors thank Radek Erban, Thierry Goudon, Yannis Kevrekidis and Tony Lelièvre for interesting discussions that eventually led to this work. This work was performed during a research stay of GS at SIMPAF (INRIA - Lille). GS warmly thanks the whole SIMPAF team for its hospitality. GS is a Postdoctoral Fellow of the Research Foundation – Flanders. This work was partially supported by the Research Foundation – Flanders through Research Project G.0130.03 and by the Interuniversity Attraction Poles Programme of the Belgian Science Policy Office through grant IUAP/V/22 (GS). The scientific responsibility rests with its authors. Technical lemmas, used in proof of Theorem \[thm:coupling\] =========================================================== \[lem:tech1\] Let $(\theta_n)_{n\geq 1}$ a sequence of independent exponential random variables with mean $1$, and $(T_{n})_{n \geq 1}$ a strictly increasing sequence of random times with $T_0=0$ such that: - For any $m \geq 1$, the sequence $(\theta_n)_{n\geq m+1}$ is independent of the past $(\theta_n,T_n)_{n\leq m}$. - There is a constant $C$ such that for all $n\geq 0 $, $$\label{eq:h1} \frac{1}{C} \theta_{n+1} \leq T_{n+1} - T_n \leq C \theta_{n+1} .$$ Let $N_t$ a random integer such that: $$T_{N_t} \leq t \leq T_{N_t +1}.$$ Then for any integer $ k \geq 0$, there is a constant $C_k$ independent of $t$ such that: $${ {\mathbb E} }{ \left( \theta_{N_t+1} \right) }^k \leq C_k.$$ The result may seem rather intuitive, but the proof is slightly tricky. First, we consider sub-intervals of $[0,t]$ of the form: $$I_p = [\frac{p-1}{P} t, \frac{p}{P} t],$$ for some $P\geq 1$ and $p = 1, \dots , P$. Step (i). The first step consists in proving the following estimate: $$\label{eq:app1} { {\mathbb E} }{ \left( \theta_{N_t+1} ^k \right) } \leq \sum_{p=0}^{P-1} { {\mathbb E} }{ \left(\theta^k {{\bf 1 }}{ \left( \theta \geq \frac{p}{C} \frac{t}{P} \right) } \right) } \sup_{p=1 \dots P} \sum_{n=0}^{+\infty} {\mathbb{P}}(T_n \in I_p)$$ Indeed, $$\begin{aligned} { {\mathbb E} }{ \left( \theta_{N_t+1} ^k \right) } & = \sum_{n=0}^{+ \infty} { {\mathbb E} }{ \left( \theta^k_{n+1} {{\bf 1 }}{ \left( T_{n} \leq t \leq T_{n +1} \right) } \right) } \\ & \leq \sum_{n=0}^{+ \infty} \sum_{p=1}^P { {\mathbb E} }{ \left( \theta^k_{n+1} {{\bf 1 }}{ \left( T_{n +1} -T_n \geq \frac{P-p}{P} t \vert T_n \in I_p \right) } \right) }{\mathbb{P}}{ \left(T_n \in I_p\right) }.\end{aligned}$$ Using the independence of $\theta_{n+1}$ with $T_n$, and the assumption , it yields, $${ {\mathbb E} }{ \left( \theta^k_{n+1} {{\bf 1 }}{ \left( T_{n +1} -T_n \geq \frac{P-p}{P} t \big \vert T_n \in I_p \right) } \right) }\leq { {\mathbb E} }{ \left( \theta^k {{\bf 1 }}_{ { \left( \theta \geq \frac{P-p}{C P} t \right) } } \right) }.$$ and finally changing the sum index $p$ to $P-p+1$ yields . Step (ii). The second step consists in the following decomposition. Let $I=[a,b]$ be a finite interval of ${\mathbb{R}}^+$, $T_{-1}=-\infty$. Then, $$\label{eq:app2} \sum_{n=0}^{+\infty} {\mathbb{P}}(T_n \in I) = \sum_{n=0}^{+\infty} { \left( {\mathbb{P}}( T_n \in I, T_{n-1} \notin I ) \sum_{m=0}^{+\infty} {\mathbb{P}}( T_{n+1},\dots,T_{n+m} \in I \vert T_n \in I, T_{n-1} \notin I ) \right) } .$$ The key is to write $$\begin{aligned} \underbrace{\sum_{n=0}^{+\infty} {\mathbb{P}}(T_n \in I)}_{(a_0)} & = \sum_{n=0}^{+\infty} {\mathbb{P}}(T_{n-1} \notin I,T_n \in I) + \sum_{n=0}^{+\infty} {\mathbb{P}}(T_{n-1},T_n \in I) \\ & = \sum_{n=0}^{+\infty} {\mathbb{P}}(T_{n-1} \notin I,T_n \in I) + \underbrace{\sum_{n=0}^{+\infty} {\mathbb{P}}(T_{n} ,T_{n+1} \in I)}_{(a_1)} .\end{aligned}$$ We wish to iterate the decomposition of $(a_0)$ to $(a_1)$, $(a_2)$, etc... For this purpose, remark that $$\begin{aligned} (a_m) &= \sum_{n=0}^{+\infty} {\mathbb{P}}(T_{n}, \dots ,T_{n+m} \in I) \\ & = \sum_{n=0}^{+\infty} {\mathbb{P}}(T_{n} \in I ) {\mathbb{P}}(T_{n+1}, \dots ,T_{n+m} \in I \vert T_{n} \in I ) \\ & \leq { \left( \sum_{n=0}^{+\infty} {\mathbb{P}}(T_{n} \in I ) \right) } {\mathbb{P}}( \theta_1+ \dots + \theta_m \leq C (b-a) ) \\ & \xrightarrow{m \to + \infty} 0 ,\end{aligned}$$ so that we get in the end $$(a_0) = \sum_{n=0}^{+\infty} \sum_{m=0}^{+\infty} {\mathbb{P}}(T_{n-1} \notin I,T_n, \dots T_{n+m} \in I).$$ Factorizing ${\mathbb{P}}( T_n \in I, T_{n-1} \notin I ) $ with Bayes’ formula yields . Step (iii). The estimate  yields $$\label{eq:app4} { {\mathbb E} }{ \left( \theta_{N_t+1} ^k \right) } \leq \sum_{p=0}^{+\infty} { {\mathbb E} }{ \left(\theta^k {{\bf 1 }}_{ { \left( \theta \geq \frac{p}{C} \frac{t}{P} \right) } } \right) } \sum_{m=0}^{+\infty} {\mathbb{P}}{ \left(\theta_1+\dots +\theta_{m} \leq \frac{t}{CP} \right) }.$$ Indeed for $I=[a,b]$, using the independence of $(\theta_{n+1}, \dots,\theta_{n+m})$ from $(T_n,T_{n-1})$: $$\sum_{m=0}^{+\infty} {\mathbb{P}}( T_{n+1},\dots,T_{n+m} \in I \vert T_n \in I, T_{n-1} \notin I ) \leq \sum_{m=0}^{+\infty} {\mathbb{P}}{ \left( \theta_{1} + \dots + \theta_{m} \leq C(b-a) \right) },$$ and $$\sum_{n=0}^{+\infty} {\mathbb{P}}( T_n \in I, T_{n-1} \notin I ) = {\mathbb{P}}(\exists n \geq 0 \vert T_n \in I ) \leq 1.$$ The two last estimates in  yields: $$\sup_{p=1 \dots P} \sum_{n=0}^{+\infty} {\mathbb{P}}(T_n \in I_p) \leq \sum_{m=0}^{+\infty} {\mathbb{P}}{ \left(\theta_1+\dots +\theta_{m} \leq \frac{t}{CP} \right) },$$ and  follows. Finally, we choose $P= \lfloor t +1\rfloor$ so that the right hand side of  does not depend on $t$ any more. Finally we use the exponential convergence of the terms of the two series in the right hand side of  (using for instnace large deviation estimates for the second) one, to conclude that the two series indeed converge, and conclude the proof. \[lem:tech2\] Let ${ \left(\theta_n\right) }_{n \geq 0}$ be a sequence of exponential random variables of mean $1$, and $t \mapsto N_t \in \mathbb N$ the Poisson process uniquely defined by: $$\sum_{n=1}^{N_t} \theta_n \leq t < \sum_{n=1}^{N_t +1} \theta_n.$$ Let $n_0 \in \mathbb N$ such that $n_0 \geq 4 {{\rm e}}t$, $p_1 \geq 1$, and $p_2 \geq 1$. Then there is a constant $C_{p_1,p_2}$ independent of $(n_0,t)$ such that: $${ {\mathbb E} }{ \left( {\left | \sum_{n=1}^{N_t+1}\theta_n^{p_1} \right |}^{p_2} {{\bf 1 }}_{N_t > n_0}\right) } \leq C_{p_1,p_2} 2^{- n_0} .$$ Step (i).\ Let us condition on the different possible values of $N_t > n_0$: $$\begin{aligned} { {\mathbb E} }{ \left( {\left | \sum_{n=1}^{N_t+1}\theta_n^{p_1} \right |}^{p_2} {{\bf 1 }}_{N_t > n_0}\right) } & = \sum_{m=n_0+1}^{+\infty}{ {\mathbb E} }{ \left( {\left | \sum_{n=1}^{N_t+1}\theta_n^{p_1} \right |}^{p_2} \Big \vert N_t = m \right) } {\mathbb{P}}{ \left(N_t = m\right) } \\ & = \sum_{m=n_0+1}^{+\infty}{ {\mathbb E} }{ \left( {\left | \sum_{n=1}^{N_t+1}\theta_n^{p_1} \right |}^{p_2} \Big \vert N_t = m \right) } {{\rm e}}^{-t}\frac{t^{m}}{{m} ! } . \end{aligned}$$ Step (ii).\ Let us denote $({\mathbb{P}}_{t,m},{ {\mathbb E} }_{t,m}) = ({\mathbb{P}}{ \left(\, . \, \vert N_t = m\right) },{ {\mathbb E} }{ \left(\, . \, \vert N_t = m\right) })$ the probability/expectation conditionally on the event ${\left\{N_t = m\right\}}$. Conditionally on ${\left\{N_t = m\right\}}$, the sequence $(\theta_1, \dots,\theta_{m+1})$ has the same distribution as $(t U_{(1)}, t U_{(2)}-t U_{(1)}, \dots, t U_{(m)}- t U_{(m-1)}, t- t U_{(m)}+\theta)$, where $(U_{(1)} < \dots < U_{(m)})$ is the order statistics of $m$ independent random variables $(U_1, \dots , U_m)$ uniformly distributed on the interval $[0,1]$, and $\theta$ is an independent exponential random variable of mean $1$ (see Section 2.4 of [@Nor98] for classical properties of Poisson processes). Thus denoting $U_{(0)}=0$ we can write: $$\begin{aligned} \label{eq:app11} & { {\mathbb E} }{ \left( {\left | \sum_{n=1}^{N_t+1}\theta_n^{p_1} \right |}^{p_2} {{\bf 1 }}_{N_t > n_0}\right) } = \nonumber\\ & \qquad \sum_{m=n_0+1}^{+\infty}{ {\mathbb E} }_{t,m} { \left( {\left | { \left(t- t U_{(m)}+\theta\right) }^{p_1} + t^{p_1}\sum_{n=1}^{m} { \left(U_{(n)}-U_{(n-1)}\right) }^{p_1} \right |}^{p_2} \right) } {{\rm e}}^{-t}\frac{t^{m}}{{m} ! } . \end{aligned}$$ Using the inequality $(a+b)^p \leq C_p a^p + b^b$ for any $a,b > 0$, Jensen inequality, and denoting $U_{(m+1)}= 1$, there is a constant $C_{p_1,p_2}$ such that: $$\begin{aligned} \label{eq:app12} & {\left | { \left(t- t U_{(m)}+\theta\right) }^{p_1} + t^{p_1}\sum_{n=1}^{m} { \left(U_{(n)}-U_{(n-1)}\right) }^{p_1} \right |}^{p_2} \leq \nonumber \\ & \quad C_{p_1,p2}{ \left(\theta^{p_2p_1} + t^{p_1 p_2}(m+1)^{p_2-1} \sum_{n=1}^{m+1} { \left(U_{(n)}-U_{(n-1)}\right) }^{p_1 p_2} \right) }.\end{aligned}$$ Step (iii).\ Let us compute ${ {\mathbb E} }_{t,m} { \left( { \left(U_{(n)}-U_{(n-1)}\right) }^{p_1 p_2} \right) }$ for $n=1, \dots, m+1$. First, let us recall the probability density of a couple $(U_{(i)},U_{(j)})$ for $1 \leq i<j \leq m$ in an order statistics of size $m$: $$\begin{aligned} \label{eq:orderstat} & {\mathbb{P}}{ \left(U_{(i)}\in [u_i,u_i+d u_i], U_{(j)} \in [u_j,u_j + d u_j] \right) } \nonumber \\ & \qquad = m! {\mathbb{P}}{ \left(U_1 < \dots < U_{i-1} < u_{i}\right) } \times {\mathbb{P}}{ \left( U_{i} \in [u_{i},u_{i}+d u_{i}] \right) } \nonumber \\ & \qquad \quad \times {\mathbb{P}}{ \left( u_i < U_{i+1} < \dots < U_{j-1}<u_{j}\right) } \times {\mathbb{P}}{ \left( U_{j} \in [u_{j},u_{j}+d u_{j}]\right) } \times {\mathbb{P}}{ \left( u_j < U_{j+1} < \dots < U_m\right) } , \nonumber \\ & \qquad = m! \frac{u_{i}^{i-1}}{(i-1)!} \frac{(u_j-u_i)^{j-i-1}}{(j-i-1)!} \frac{(1-u_j)^{m-j}}{(m-j)!} d u_{i} \, d u_{j} .\end{aligned}$$ As a consequence, $$\begin{aligned} { {\mathbb E} }_{t,m} { \left( { \left(U_{(n)}-U_{(n-1)}\right) }^{p_1 p_2} \right) } & = \int_{0< u_{n-1} < u_n < 1} m! \frac{u_{n-1}^{n-2}}{(n-2)!} (u_n - u_{n-1} )^{p_1 p_2} \frac{(1-u_n)^{m-n}}{(m-n)!} d u_{n-1} \, d u_{n} \\ & = m! (p_1 p_2 ) ! \int_{0< u_{n-1} < u_n < 1} \frac{u_{n-1}^{n-2}}{(n-2)!} \frac{(u_n - u_{n-1} )^{p_1 p_2}}{(p_1 p_2 ) !} \frac{(1-u_n)^{m-n}}{(m-n)!} d u_{n-1} \, d u_{n} \\ & = \frac{m! (p_1 p_2 ) !}{( m+p_1p_2 )!},\end{aligned}$$ where in the last line we have used  with appropriate $(i,j,m)$ to compute the integral. In the same way, the probability density of $U_{(i)}$ for $1 \leq i \leq m$ in an order statistics of size $m$ is given by: $${\mathbb{P}}{ \left(U_{(i)}\in [u_i,u_i+d u_i]\right) }=m! \frac{u_{i}^{i-1}}{(i-1)!} \frac{(1-u_i)^{m-i}}{(m-i)!} d u_i,$$ so that $$\begin{aligned} { {\mathbb E} }_{t,m} { \left( U_{(1)} ^{p_1 p_2} \right) } & = m! (p_1 p_2 ) ! \int_{0< u_{1} < 1} \frac{u_{1}^{p_1 p_2}}{(p_1 p_2)!} \frac{(1-u_1)^{m-1}}{(m-1)!} d u_1 \\ & = \frac{m! (p_1 p_2 ) !}{( m+p_1p_2 )!},\end{aligned}$$ and in the same way, $${ {\mathbb E} }_{t,m} { \left( (1-U_{(m)})^{p_1 p_2} \right) } = \frac{m! (p_1 p_2 ) !}{( m+p_1p_2 )!} .$$ Step (iv).\ Remarking that ${ {\mathbb E} }_m(\theta^{p_1p_2})=(p_1p_2)!$, and inserting the computations of Step (ii) in - yields: $$\begin{aligned} & { {\mathbb E} }{ \left( {\left | \sum_{n=1}^{N_t+1}\theta_n^{p_1} \right |}^{p_2} {{\bf 1 }}_{N_t > n_0}\right) } \leq \nonumber\\ & \qquad C_{p_1p_2}\sum_{m=n_0+1}^{+\infty} { \left( (p_1p_2)! + t^{p_1 p_2}(m+1)^{p_2-1} \frac{m! (p_1 p_2 ) !}{( m+p_1p_2 )!} \right) }{{\rm e}}^{-t}\frac{t^{m}}{{m} ! } .\end{aligned}$$ Using the inequalities $ \frac{(m+1)^{p_2-1} m !}{( m+p_1p_2 )!} \leq 1$, ${{\rm e}}^{-t}\leq 1$ and $$\sum_{m=n_0+1}^{+\infty} \frac{t^{m}}{{m} ! } \leq \frac{t^{n_0}}{ n_0 ! },$$ it yields $${ {\mathbb E} }{ \left( {\left | \sum_{n=1}^{N_t+1}\theta_n^{p_1} \right |}^{p_2} {{\bf 1 }}_{N_t > n_0}\right) } \leq C_{p_1p_2} \frac{t^{n_0+p_1p_2}}{n_0 !},$$ so that in the end, using the assumption $n_0 \geq 4 {{\rm e}}t$ and the Stirling formula, $$\begin{aligned} & { {\mathbb E} }{ \left( {\left | \sum_{n=1}^{N_t+1}\theta_n^{p_1} \right |}^{p_2} {{\bf 1 }}_{N_t > n_0}\right) } \leq C_{p_1p_2} (\frac{n_0}{4 {{\rm e}}})^{n_0+p_1p_2} \frac{1}{\sqrt{n_0}} (\frac{{{\rm e}}}{n_0})^{n_0} \\ & \qquad \leq C_{p_1p_2} \frac{n_0^{p_1 p_2}}{\sqrt{n_0} 2 ^{n_0}} 2^{- n _0} \leq C_{p_1p_2} 2^{- n _0} ,\end{aligned}$$ which yields the result. [^1]: SIMPAF, INRIA Lille - Nord Europe, Lille, France, (mathias.rousset@inria.fr). [^2]: Department of Computer Science, K.U. Leuven, Celestijnenlaan 200A, 3001 Leuven, Belgium, (giovanni.samaey@cs.kuleuven.be).

--- abstract: 'Recent arguments show that some curvaton field may generate the cosmological curvature perturbation. As the curvaton is independent of the inflaton field, there is a hope that the fine-tunings of inflation models can be cured by the curvaton scenario. More recently, however, D.H.Lyth discussed that there is a strong bound for the Hubble parameter during inflation even if one assumes the curvaton scenario. Although the most serious constraint was evaded, the bound seems rather crucial for many models of a low inflation scale. In this paper we try to remove the constraint. We show that the bound is drastically modified if there were multiple stages of inflation.' --- [SIT-HEP/TM-18]{} 1.5 truecm .75 truecm **Tomohiro Matsuda [^1]** .4 truecm *Laboratory of Physics, Saitama Institute of Technology,* *Fusaiji, Okabe-machi, Saitama 369-0293, Japan* 1\. truecm addtoreset[equation]{}[section]{} 1. truecm Introduction ============ It has recently been proposed that the energy density perturbations could originate in a scalar field other than the conventional inflaton field[@curvaton_1; @curvaton_2], which is called curvaton. After inflation, the curvaton starts to oscillate in a radiation background. During this period the energy density of the curvaton grows to account for the cosmological curvature perturbation when it decays. This curvaton paradigm has attracted a lot of attention because it has an obvious advantage. The inflaton energy scale is decoupled from the magnitude of the cosmic microwave background(CMB) temperature fluctuations, which may cure the fine-tunings of inflation models. Along this line of thought, it seems attractive to suppose that the curvaton paradigm can also save the models of a low inflation scale[@curvaton_liberate]. The construction of a realistic model of low inflation scale is an interesting problem especially when the fundamental scale is (much) lower than the Planck scale, because in these models the Hubble parameter cannot become so large as the one in the conventional models of inflation. In the history of the string theory, originally the inverse of the size of extra dimensions as well as the fundamental scale was assumed to be as large as $M_p$. However, later observations showed that there is no reason to require such a tiny compactification radius[@Extra_1]. In models of large extra dimensions, the observed Planck mass is obtained by the relation $M_p^2=M^{n+2}_{*}V_n$, where $M_{*}$ and $V_n$ denote the fundamental scale of gravity and the volume of the $n$-dimensional compact space. In the new scenario of the string theory, the compactification radius and the fundamental scale are unknown parameters that should be determined by observations. However, in many cases the cosmology of the large extra dimension must be quite different from the conventional ones.[^2] In spite of the above expectations, D.H.Lyth showed recently that there will be a strong bound for the Hubble parameter during inflation even if one assumes curvaton scenario[@Lyth_constraint]. At the same time, a mechanism that evades the most serious constraint is also suggested in ref.[@Lyth_constraint]. However, to achieve the bound $H_I > 10 TeV$, a huge curvaton mass ($m_\sigma\sim M_p$) is required after the phase transition that takes place at the end of inflation. Moreover, since the obtained bound is rather restrictive ($H_I> 10 TeV$), it should be fair to say that many models of a low inflation scale are still not liberated by the curvaton[@curvaton_liberate]. In this paper we show that the constraints obtained in [@Lyth_constraint] are relaxed to a satisfactory level if there were multiple stages of inflation. No large hierarchy between the scales of each inflation is required. At least two stages of inflation are required in our scenario. Denoting the vacuum energy during the two kinds of inflation by $V_1$ and $V_2$, the original constraint is reduced by the factor of $\epsilon^6$, where $\epsilon$ is the ratio of the scales, $\epsilon\equiv (V_2/V_1)^{\frac{1}{4}}$. As a result, for example, if the original bound for single inflation is $H_I>10^7 GeV$, we can reduce it as $H_1 >10^{-5} GeV$ for $\epsilon=10^{-2}$, where $H_1$ is the Hubble parameter during the first inflation. No sensible bound is obtained for the Hubble parameter during the second inflation. Constraint on inflation scales ============================== Here it may be helpful to begin with the review of the discussion in ref.[@Lyth_constraint]. First we start with more general settings and try to reproduce the constraint showing how the original model in [@Lyth_constraint] is realized in the boundary condition of our settings. For simplicity, we assume that the curvaton field $\sigma$ is frozen at $\sigma=\sigma_{osc}$ from the epoch of horizon exit during first inflation to the epoch when the curvaton start to oscillate. At the time when the curvaton starts to oscillate, its density is $\rho_\sigma \sim m_\sigma^2 \sigma_{osc}^2$. Denoting the total density of other fields by $\rho_{tot}$, the ratio $r$ at the time of curvaton oscillation is $$\left.\frac{\rho_{\sigma}}{\rho_{tot}}\right|_{H=H_{osc}} \sim \frac{m_{\sigma}^2 \sigma_{osc}^2}{H_{osc}^2 M_p^2}.$$ During the period when the total density $\rho_{tot}$ is radiation-dominated, the ratio $r$ grows and reaches finally at $$\label{rfirst} r \le \frac{\sqrt{H_{osc} M_p}}{T_d } \frac{m_\sigma^2 \sigma_{osc}^2}{H_{osc}^2 M_p^2}$$ when the curvaton decay. Here $T_d$ is the temperature just after curvaton decay. Now the curvature perturbation $$\zeta \simeq \frac{r}{3}\frac{\delta \rho_{\sigma} }{\rho_\sigma}$$ is generated by the curvaton. Using the spectrum of the perturbation $<\delta \sigma_{osc}^2>=(\frac{H_I}{2\pi})^2$, the spectrum of the curvature perturbation is given by $$\label{Pzeta} {\cal P}_{\zeta}^{\frac{1}{2}}\simeq \frac{2r}{3}\frac{H_I}{2\pi\sigma_{osc}}.$$ Using the required value from the observations ${\cal P}_{\zeta}^{\frac{1}{2}}=5\times 10^{-5}$, one obtains that $$\label{pert1} \frac{2r}{3}\frac{H_I}{2\pi\sigma_{osc}}\simeq 5\times 10^{-5}.$$ Using (\[rfirst\]) and (\[pert1\]), one finds the following constraint; $$\label{start} \frac{2}{3}\frac{H_I}{2\pi\sigma_{osc}} \frac{\sqrt{H_{osc} M_p}}{T_d } \frac{m_\sigma^2 \sigma_{osc}^2}{H_{osc}^2 M_p^2} \ge 5\times 10^{-5}.$$ Since the naive bound from nucleosynthesis is $T_d >1 MeV$, one can obtain $$\label{constraint1} \frac{2}{3} \frac{H_I \sigma_{osc} m_{\sigma}^2} { 2\pi H_{osc}^{\frac{3}{2}} M_p ^{\frac{5}{2}}} \ge 5\times 10^{-26}.$$ On the other hand, one can use the lower bound for the curvaton decay rate $\Gamma_{\sigma}\ge \frac{m_{\sigma}^3}{M_p^2}$ to obtain $T_d \simeq \sqrt{M_p \Gamma}\ge M_p (m_{\sigma}/M_p)^{\frac{3}{2}}$, which implies that $$\label{constraint2} \frac{2}{3} \frac{H_I \sigma_{osc} m_{\sigma}^\frac{1}{2}} { 2\pi H_{osc}^{\frac{3}{2}} M_p } \ge 5\times 10^{-5}.$$ The four parameters ($H_I, \sigma_{osc}, H_{osc}, m_{\sigma}$), are the boundary condition that depends on which model one may choose. Here the Hubble parameter during inflation when the observable University leaves the horizon is denoted by $H_I$. Considering (\[pert1\]) and $r<1$, one finds $$\label{ineq1} \frac{2}{3}\frac{H_I}{2\pi \sigma_{osc}} > 5\times 10^{-5}.$$ From eq.(\[ineq1\]) and (\[constraint1\]), one finds $$\label{constraint1a} H_I > 10^{-15} \times \frac{H_{osc}^{\frac{3}{4}} M_p^{\frac{5}{4}}}{m_\sigma}.$$ On the other hand, from eq.(\[ineq1\]) and (\[constraint2\]) one obtains the following constraint; $$\label{constraint2a} H_I > 10^{-4} \times \frac{H_{osc}^{\frac{3}{4}} M_p^\frac{1}{2}}{m_\sigma^{\frac{1}{4}}}.$$ First let us consider a simple example of the boundary condition. We will assume that the mass of the curvaton $m_\sigma$ is a constant during the period we are interested in. Then the oscillation of the curvaton field starts when the Hubble parameter falls below the curvaton mass, $H< m_\sigma$, which means $H_{osc} \simeq m_\sigma$. As the inequality $H_I>H_{osc}$ always holds, we set a new parameter $\epsilon_H$, which is defined as $H_{osc} \equiv \epsilon_H^2 H_{I}$. Now one can replace $H_{osc}$ and $m_\sigma$ in eq.(\[constraint1a\]) by $H_I$ to find $$\label{constraint1b} H_I > M_p \times 10^{-12} \times \epsilon_H^{-\frac{2}{5}} \simeq 10^{6} \epsilon_H^{-\frac{2}{5}} GeV.$$ Following the same arguments, one can find from (\[constraint2a\]), $$\label{constraint2b} H_I > 10^{-8} M_p \epsilon_H^2 \simeq 10^{10} \epsilon_H^2 GeV$$ As is discussed in ref.[@Lyth_constraint], the bound from eq.(\[constraint2b\]) becomes more strict than (\[constraint1b\]) when $\epsilon_H > 10^{-2}$. Note that a tiny $\epsilon_H$ does not relax the bound. The obtained bound $H_I >10^7 GeV$ corresponds to the first constraint that was obtained in ref.[@Lyth_constraint]. In the above arguments we showed that the bound obtained in ref.[@Lyth_constraint] is derived from an initial condition that fixes the four parameters ($H_I, \sigma_{osc}, H_{osc}, m_{\sigma}$). Because some part of the initial condition is still model dependent, it is interesting to find models that evade the above constraint without introducing fine-tunings or large hierarchy. A way to relax the bound is already suggested in [@Lyth_constraint]. We think it is instructive to reproduce the argument within our setups and show explicitly how the initial condition is modified in the model. In the “heavy curvaton” scenario, two differences appear in the initial condition. Assuming that there was a phase transition just after inflation, which makes the curvaton mass much larger than the one during inflation, the conditions are modified as follows; $$\begin{aligned} H_{osc}\simeq m_{\sigma} \, \, &\rightarrow& \, \, H_{osc} \ll m_{\sigma}\nonumber\\ H_{osc}<H_I \, \, &\rightarrow& \, \, H_{osc} \simeq H_I.\end{aligned}$$ Using the modified initial conditions and (\[constraint1a\]), one can easily find $$\label{constraint1c} H_I > 10^{-56} \frac{M_p^5}{m_{\sigma}^4}.$$ One can also find another bound from (\[constraint2a\]), $$\label{constraint2c} H_I > 10^{-14}\frac{M_p^2}{m_\sigma}.$$ As is suggested in [@Lyth_constraint], it might be possible to obtain a preferable bound $H_I> O(TeV)$ if the curvaton mass becomes as large as the Planck mass. A huge curvaton mass may be allowed in a conventional supergravity, but is not preferable in the models of large or intermediate extra dimensions. Thus it is still an interesting problem to find models in which the bound is more relaxed. Multiple inflation ================== In the followings we try to find another model whose initial condition significantly lowers the above bound for $H_I$. The most promising example will be to assume that the phase transition that induces the curvaton oscillation starts independently of the inflation that produces the spectrum of the perturbation $\delta \sigma$. The easiest way to realize the model is to introduce secondary weak inflation that triggers the required phase transition.[^3] Here we introduce the parameter $H_e$ that denotes the Hubble parameter during the last inflation. The spectrum of the perturbation $\delta \sigma$ is produced during the first inflation, the Hubble parameter during which is denoted by $H_I$. More explicitly, the parameters follow the conditions, $H_I \gg H_e$, $m_\sigma \gg H_{osc}$ and $H_{osc}\simeq H_e$. The bound (\[constraint1a\]) now becomes $$\label{constraint1d} H_I > 10^{-15} \frac{H_{e}^\frac{3}{4} M_p^\frac{5}{4}}{m_\sigma}.$$ Of course, setting $H_I=H_e=H_{osc}$, one obtains (\[constraint1c\]) again. The bound (\[constraint2a\]) becomes $$\label{constraint2d} H_I > 10^{-4} \frac{H_{e}^\frac{3}{4} M_p^\frac{1}{2}}{m_\sigma^{\frac{1}{4}}}.$$ Now let us consider a case when $H_{e}=H_I \times 10^{-4}$. In this case, (\[constraint2d\]) becomes $$\label{constraint2d2} H_I > 10^{-26} \frac{M_p^2}{m_\sigma}$$ which is about $10^{-12}$ times smaller than (\[constraint2c\]). To see exactly what happened in the bound, here we consider a more generic situation. Introducing mass scales $M_I \equiv V_I^{\frac{1}{4}}=H_I^2M_p^2$ and $M_e \equiv V_e^{\frac{1}{4}}=H_e^2M_p^2$, and denoting the ratio by $\epsilon \equiv \frac{M_e}{M_I}$, one finds from (\[constraint1d\]), $$\label{constraint1e} H_I > 10^{-56} \epsilon^6 \frac{M_p^5}{m_\sigma^4}.$$ One can also find from (\[constraint2d\]), $$\label{constraint2e} H_I > 10^{-14} \epsilon^6 \frac{M_p^2}{m_\sigma}.$$ From eq.(\[constraint1e\]) and eq.(\[constraint2e\]), it is easy to understand why the bound is so sensitive to the ratio $\epsilon$. If the scale of the vacuum energy during the first inflation is only $10^{2}$ times larger than the one during the second inflation, the bound is significantly reduced by the factor of $(10^{-2})^6$. Conclusions and Discussions =========================== In this paper we have examined the constraint on the Hubble parameter during inflation when the curvaton hypothesis is used to explain the observed density perturbation of the Universe. The constraint (\[start\]) that we have started from produces bounds for the Hubble parameter once the initial conditions are determined by the models. We have found an example of the model that significantly modifies the bound. At least two stages of inflation are required in our scenario. The first produces the spectrum of the perturbation $\delta \sigma$ while the other triggers the phase transition that produces the curvaton mass. Then the ratio of the vacuum energy $\epsilon$ reduces the constraint by the factor of $\epsilon^6$, which evades the bound that was discussed in [@Lyth_constraint]. Acknowledgment ============== We wish to thank K.Shima for encouragement, and our colleagues in Tokyo University for their kind hospitality. [1]{} S. Mollerach, Phys.Rev.D42(1990)313; A. D. Linde, V. Mukhanov, Phys.Rev.D56(1997)535; K. Enqvist, M. S. Sloth, Nucl.Phys.B626(2002)395. T. Moroi, T. Takahashi, Phys.Lett.B522(2001)215; D. H. Lyth, D. Wands, Phys.Lett.B524(2002)5. “Models of inflation liberated by the curvaton hypothesis”, K. Dimopoulos, D. H. Lyth, hep-ph/0209180; M. Giovannini, Phys.Rev.D67(2003)123512, “Tracking curvaton(s)?”, hep-ph/0310024 I.Antoniadis, N.A-Hamed, S.Dimopoulos, and G.R.Dvali, Phys.Lett.B436(1998)257; I.Antoniadis, Phys.Lett.B246(1990)377; N.A-Hamed, S.Dimopoulos and G.R.Dvali, Phys.Lett.B429(1998)263. N. Arkani-Hamed, S. Dimopoulos, N. Kaloper, and J. March-Russell, Nucl.Phys.B567(2000)189; R. N. Mohapatra, A. Perez-Lorenzana, and C. A. de S. Pires, Phys.Rev.D62(2000)105030; A. M. Green and A. Mazumdar, Phys.Rev.D65(2002)105022; D. H. Lyth, Phys.Lett.B448(1999)191; Phys.Lett.B466(1999)85; T. Matsuda, Phys.Rev.D67(2003)083519; JCAP0306(2003)007; Phys.Rev.D68:047702,2003; “ F-term, D-term and hybrid brane inflation”, hep-ph/0302078; “Q ball inflation”, hep-ph/0309339; M. Axenides, K. Dimopoulos, “Inflation without flat directions”, hep-ph/0310194 T.Matsuda, Phys.Lett.B423(1998)35. G.R. Dvali, G. Gabadadze, Phys.Lett.B460(1999)47; T. Matsuda, Phys.Rev.D66(2002)107301; Phys.Rev.D66(2002)023508; Phys.Rev.D65(2002)107302; Phys.Rev.D64(2001)083512; J.Phys.G27(2001)L103; A. Masiero, M. Peloso, L. Sorbo, and R. Tabbash, Phys.Rev.D62(2000)063515; A.Pilaftsis, Phys.Rev.D60(1999)105023; R.Allahverdi, K.Enqvist, A.Mazumdar and A.P-Lorenzana, Nucl.Phys. B618(2001)377; S. Davidson, M. Losada, and A. Riotto, Phys.Rev.Lett.84(2000)4284. A.Mazumdar, Nucl.Phys.B597(2001)561, Phys.Rev.D64(2001)027304; A. Mazumdar and A. Perez-Lorenzana, Phys.Rev.D65(2002)107301; R. Allahverdi, K. Enqvist, A. Mazumdar, and A. Perez-Lorenzana, Nucl.Phys.B618(2001)277; T.Matsuda, Phys.Rev.D65(2002)103502. T.Matsuda, Phys.Rev.D65(2002)103501; Phys.Rev.D67(2003)127302. “Can the curvaton paradigm accommodate a low inflation scale?” D. H. Lyth,hep-th/0308110 D.H. Lyth and E. D. Stewart, Phys.Rev.D53(1996)1784; G.R. Dvali, Phys.Lett.B459(1999)489. [^1]: matsuda@sit.ac.jp [^2]: Finding crucial constraints on the compactification, or constructing successful models of the inflation with low (or intermediate) fundamental scale is a challenging issue[@low_scale_inflation]. In the models of large extra dimensions, the difficulties of inflation are sometimes related to the stability of the compactified space[@low_scale_inflation; @stability] or the mechanism of baryogenesis that must take place after inflation[@low_baryo; @low_AD; @ADafterThermal]. [^3]: In this paper we do not mention how to realize the multiple inflation. Some arguments are given in [@low_scale_inflation; @thermal].

--- abstract: '*Delusional bias* is a fundamental source of error in approximate Q-learning. To date, the only techniques that explicitly address delusion require comprehensive search using tabular value estimates. In this paper, we develop efficient methods to mitigate delusional bias by training Q-approximators with labels that are “consistent” with the underlying greedy policy class. We introduce a simple penalization scheme that encourages Q-labels used *across training batches* to remain (jointly) consistent with the expressible policy class. We also propose a search framework that allows multiple Q-approximators to be generated and tracked, thus mitigating the effect of premature (implicit) policy commitments. Experimental results demonstrate that these methods can improve the performance of Q-learning in a variety of Atari games, sometimes dramatically.' bibliography: - 'long.bib' - 'standard.bib' --- Introduction {#sec:intro} ============ *Q-learning* [@watkins:mlj92; @sutton:rlbook_2ed] lies at the heart of many of the recent successes of deep reinforcement learning (RL) [@dqn-atari; @alphaGo:nature2016], with recent advancements (e.g., @hasselt:2010 [@bellemare:icml17; @dueling:icml16; @rainbow:2017]) helping to make it among the most widely used methods in applied RL. Despite these successes, many properties of Q-learning are poorly understood, and it is challenging to successfully apply deep Q-learning in practice. Various modifications have been proposed to improve convergence or approximation error [@gordon:ml95; @Gordon99; @szepesvarismart04; @melo07; @maeietal:icml2010; @munosetal:nips16]; but it remains difficult to reliably attain both robustness and scalability. Recently, @LuEtAl_Qdelusion:nips18 identified a source of error in Q-learning with function approximation known as *delusional bias*. This bias arises because Q-learning updates the value of state-action pairs using estimates of (sampled) successor-state values that can be *mutually inconsistent given the policy class induced by the approximator*. This can result in unbounded approximation error, divergence, policy cycling, and other undesirable behavior. To handle delusion, the authors propose a *policy-consistent backup* operator that maintains multiple Q-value estimates organized into *information sets*. Each information set has its own backed-up Q-values and corresponding “policy commitments” responsible for inducing these values. Systematic management of these sets ensures that only *consistent* choices of maximizing actions are used to update Q-values. All potential solutions are tracked to prevent premature convergence on specific policy commitments. Unfortunately, the proposed algorithms use tabular representations of Q-functions, so while this establishes foundations for delusional bias, the function approximator is used neither for generalization nor to manage the size of the state/action space. Consequently, this approach is not scalable to practical RL problems. In this work, we develop [<span style="font-variant:small-caps;">ConQUR</span>]{} (*CONsistent Q-Update Regression*), a general framework for integrating policy-consistent backups with regression-based function approximation for Q-learning and for managing the search through the space of possible regressors (i.e., information sets). With suitable search heuristics, the proposed framework provides a computationally effective means for minimizing the effects of delusional bias, while scaling to practical problems. Our main contributions are as follows. First, we define novel augmentations of Q-regression to increase the degree of policy consistency across training batches. Since testing exact consistency is expensive, we introduce an efficient *soft-consistency penalty* that promotes consistency of labels with earlier policy commitments. Second, using information-set structure [@LuEtAl_Qdelusion:nips18], we define a search space over Q-regressors to explore multiple sets of policy commitments. Third, we propose heuristics to guide the search, critical given the combinatorial nature of information sets. Finally, experimental results on the Atari suite [@bellemare:jair2013] demonstrate that [<span style="font-variant:small-caps;">ConQUR</span>]{} can add (sometimes dramatic) improvements to Q-learning. These results further show that delusion does emerge in practical applications of Q-learning. We also show that straightforward consistency penalization on its own (i.e., without search) can improve both standard and double Q-learning. Background {#sec:preliminaries} ========== \[sec:background\] We assume a discounted, infinite horizon *Markov decision process (MDP)*, ${\mathbf{M}}= ({\mathcal{S}}, A, P, p_0, R, \gamma)$. The state space ${\mathcal{S}}$ can reflect both discrete and continuous features, but we take the action space $A$ to be finite (and practically enumerable). We consider *Q-learning* with a function approximator $Q_\theta$ to learn an (approximately) optimal Q-function [@watkins:1989; @sutton:rlbook_2ed], drawn from some approximation class parameterized by $\Theta$ (e.g., the weights of a neural network). When the approximator is a deep network, we generically refer to this as *DQN*, the method at the heart of many RL successes [@dqn-atari; @alphaGo:nature2016]. For online Q-learning, at a transition $s,a,r,s'$, the Q-update is given by: $$\theta \leftarrow \theta + \alpha \Big(r + \gamma\max_{a'\in A} Q_\theta(s', a') - Q_\theta(s, a)\Big) \nabla_\theta Q_\theta(s, a). \label{eq:online_q_update}$$ Batch versions of Q-learning are similar, but fit a regressor repeatedly to batches of training examples [@ernst_batchQ:jmlr2005; @riedmiller:ecml2005], and are usually more data efficient and stable than online Q-learning. Batch methods use a sequence of (possibly randomized) data batches $D_1,\ldots, D_T$ to produce a sequence of regressors $Q_{\theta_1},\ldots, Q_{\theta_T} = Q_\theta$, estimating the Q-function.[^1] For each $(s,a,r,s')\in D_k$, we use a prior estimator $Q_{\theta_{k-1}}$ to bootstrap the *Q-label* $q = r+\gamma\max_{a'} Q_{\theta_{k-1}}(s',a')$. We then fit $Q_{\theta_{k}}$ to this data using a regression procedure with a suitable loss function. Once trained, the (implicit) induced policy $\pi_\theta$ is the *greedy policy* w.r.t. $Q_\theta$, i.e., $\pi_\theta(s) = \operatorname*{arg\!max}_{a\in A} Q_\theta(s,a)$. Let ${\mathcal{F}}(\Theta)$ (resp., $G(\Theta)$) be the class of expressible Q-functions (resp., greedy policies). Intuitively, *delusional bias* occurs whenever a backed-up value estimate is derived from action choices that are not (jointly) realizable in $G(\Theta)$ [@LuEtAl_Qdelusion:nips18]. Standard Q-updates back up values for each $(s,a)$ pair by *independently* choosing maximizing actions at the corresponding next states $s'$. However, such updates may be “inconsistent” under approximation: if no policy in $G(\Theta)$ can jointly express all past action choices, *backed up values may not be realizable by any expressible policy*. @LuEtAl_Qdelusion:nips18 show that delusion can manifest itself with several undesirable consequences (e.g., divergence). Most critically, it can prevent Q-learning from learning the optimal representable policy in $G(\Theta)$. To address this, they propose a non-delusional *policy consistent* Q-learning (PCQL) algorithm that provably eliminates delusion. We refer to the original paper for details, but review the main concepts.[^2] The first key concept is that of *policy consistency*. For any $S\subseteq {\mathcal{S}}$, an *action assignment* $\sigma_S: S\rightarrow A$ associates an action $\sigma(s)$ with each $s\in S$. We say $\sigma$ is *policy consistent* if there is a greedy policy $\pi\in G(\Theta)$ s.t. $\pi(s) = \sigma(s)$ for all $s\in S$. We sometimes equate a set ${\mathit{SA}}$ of state-action pairs with the implied assignment $\pi(s)=a$ for all $(s,a) \in{\mathit{SA}}$. If ${\mathit{SA}}$ contains multiple pairs with the same state $s$, but different actions $a$, it is a *multi-assignment* (we use the term “assignment” when there is no risk of confusion). In (batch) Q-learning, each new regressor uses training labels generated by assuming maximizing actions (under the prior regressor) are taken at its successor states. Let $\sigma_k$ be the collection of states and corresponding maximizing actions used to generate labels for regressor $Q_{\theta_k}$ (assume it is policy consistent). Suppose we train $Q_{\theta_{k}}$ by bootstrapping on $Q_{\theta_{k-1}}$. Now consider a training sample $(s,a,r,s')$. Q-learning generates label $r + \gamma\max_{a'} Q_{\theta_{k-1}}(s',a')$ for input $(s,a)$. Notice, however, that taking action $a^*=\operatorname*{arg\!max}_{a'} Q_{\theta_k}(s',a')$ at $s'$ may not be *policy consistent* with $\sigma_k$. Thus Q-learning will estimate a value for $(s,a)$ assuming execution of a policy that cannot be realized given the approximator. PCQL prevents this by ensuring that any assignment used to generate labels is consistent with earlier assignments. This means Q-labels will often *not* be generated using maximizing actions w.r.t. the prior regressor. The second key concept is that of *information sets*. One will generally not be able to use maximizing actions to generate labels, so tradeoffs can be made when deciding which actions to assign to different states. Indeed, even if it is feasible to assign a maximizing action $a$ to state $s$ early in training, say at batch $k$, since it may prevent assigning a maximizing $a'$ to $s'$ later, say batch $k+\ell$, we may want to use a different assignment to $s$ to give more flexibility to maximize at other states later. PCQL does not anticipate the tradeoffs—rather it maintains *multiple information sets*, each corresponding to a different assignment to the states seen in the training data this far. Each gives rise to a *different Q-function estimate*, resulting in multiple hypotheses. At the end of training, the best hypothesis is that with maximum expected value w.r.t. an initial state distribution. PCQL provides strong convergence guarantees, but it is a tabular algorithm: the function approximator *restricts* the policy class, but is not used to generalize Q-values. Furthermore, its theoretical guarantees come at a cost: it uses *exact* policy consistency tests—tractable for linear approximators, but impractical for large problems and DQN; and it maintains *all* consistent assignments. As a result, PCQL cannot be used for large RL problems of the type tackled by DQN. The [<span style="font-variant:small-caps;">ConQUR</span>]{} Framework {#sec:conqur} ====================================================================== We develop the [<span style="font-variant:small-caps;">ConQUR</span>]{} framework to provide a practical approach to reducing delusion in Q-learning, specifically addressing the limitations of PCQL identified above. [<span style="font-variant:small-caps;">ConQUR</span>]{} consists of three main components: a practical soft-constraint penalty that promotes policy consistency; a search space to structure the search over multiple regressors (information sets, action assignments); and heuristic search schemes (expansion, scoring) to find good Q-regressors. Preliminaries {#sec:conqurprelim} ------------- We assume a set of training data consisting of quadruples $(s,a,r,s')$, divided into (possibly non-disjoint) *batches* $D_1, \ldots, D_T$ for training. This perspective is quite general: online RL corresponds to $|D_i| = 1$; offline batch training (with sufficiently exploratory data) corresponds to a single batch (i.e., $T=1$); and online or batch methods with replay are realized when the $D_i$ are generated by sampling some data source with replacement. For any batch $D$, let $\chi(D) = \{s' : (s,a,r,s') \in D\}$ be the set of *successor states* of $D$. An *action assignment* $\sigma_D$ for $D$ is an assignment (or multi-assignment) from $\chi(D)$ to $A$, dictating which action $\sigma_D(s')$ is considered “maximum” when generating a Q-label for pair $(s,a)$; i.e., $(s,a)$ is assigned training label $r + \gamma Q(s',\sigma(s'))$ rather than $r + \gamma \max_{a'\in A} Q(s',a')$. The set of all such assignments $\Sigma(D) = A^{\chi(D)}$ grows exponentially with $|D|$. Given a Q-function parameterization $\Theta$, we say $\sigma_D$ is *$\Theta$-consistent (w.r.t. $D$)* if there is some $\theta\in\Theta$ s.t.  $\pi_\theta(s') = \sigma(s')$ for all $s'\in\chi(D)$.[^3] This is simple policy consistency, but with notation that emphasizes the policy class. Let $\Sigma_\Theta(D)$ denote the set of all $\Theta$-consistent assignments over $D$. The union $\sigma_1 \cup \sigma_2$ of two assignments (over $D_1, D_2$, resp.) is defined in the usual way. Consistency Penalization {#sec:consistency} ------------------------ Enforcing strict $\Theta$-consistency as regressors $\theta_1, \theta_2, \ldots, \theta_T$ are generated is computationally challenging. Suppose the assignments $\sigma_1,\ldots,\sigma_{k-1}$, used to generate labels for $D_1, \ldots D_{k-1}$, are jointly $\Theta$-consistent (let $\sigma_{\leq k-1}$ denote their multi-set union). Maintaining $\Theta$-consistency when generating $\theta_{k}$ imposes two requirements. First, one must generate an assignment $\sigma_k$ over $D_{k}$ s.t. $\sigma_{\leq k-1} \cup\sigma_k$ is consistent. Even testing assignment consistency can be problematic: for linear approximators this is a linear feasibility program [@LuEtAl_Qdelusion:nips18] whose constraint set grows linearly with $|D_1 \cup \ldots\cup D_{k}|$. For DNNs, this is a complex, more expensive polynomial program. Second, the regressor $\theta_k$ should itself be consistent with $\sigma_{\leq k-1} \cup\sigma_k$. This too imposes a severe burden on regression optimization: in the linear case, it is a constrained least-squares problem (solvable, e.g., as a quadratic program); while with DNNs, it can be solved, say, using a more involved projected SGD. However, the sheer number of constraints makes this impractical. Rather than enforcing consistency, we propose a simple, computationally tractable scheme that “encourages” it: a penalty term that can be incorporated into the regression itself. Specifically, we add a penalty function to the usual squared loss to encourage updates of the Q-regressors to be consistent with the underlying information set, i.e., the prior action assignments used to generate its labels. When constructing $\theta_{k}$, let $D_{\leq k} = \cup \{D_j: j \leq k\}$, and $\sigma\in\Sigma_\Theta({D_{\leq k}})$ be the collective assignment used to generate labels for all prior regressors (including $\theta_k$ itself). The multiset of pairs $B = \{(s', \sigma(s')) | s'\in \chi(D_{\le k})\}$, is called a *consistency buffer*. The assignment need not be consistent (as we elaborate below), nor does regressor $\theta_k$ need to be consistent with $\sigma$. Instead, we use the following *soft consistency penalty* when constructing $\theta_{k}$: $$\begin{aligned} C_\theta(s', a) &= \sum_{a'\in A} [Q_\theta(s', a') - Q_\theta(s', a)]_+ ,\\ C_\theta(B) &= \sum_{(s', \sigma(s'))\in B} C_\theta(s', \sigma(s')), \label{eq:loss_fucntion_Q_delusion}\end{aligned}$$ where $[x]_+ = \max(0, x)$. This penalizes Q-values of actions at state $s$ that are larger than that of action $\sigma(s)$. Notice $\sigma$ is $\Theta$-consistent iff $\min_{\theta\in\Theta} C_\theta(B) = 0$. We add this penalty into our regression loss for batch $D_k$: $$\begin{gathered} L_\theta(D_k, B) = \sum_{(s,a,r,s')\in D_k} \Big[r + \gamma Q_{\theta_{k-1}}(s', \sigma(s')) - \\ Q_\theta(s, a)\Big]^2 + \lambda C_\theta(B). \label{eq:total_loss} $$ Here $Q_{\theta_{k}}$ is the prior estimator on which labels are bootstrapped (other regressors may be used). The penalty effectively acts as a “regularizer” on the squared Bellman error, where $\lambda$ controls the degree of penalization, allowing a tradeoff between Bellman error and consistency with the assignment used to generate labels. It thus *promotes* consistency without incurring the expense of *enforcing* strict consistency. It is straightforward to replace the classic Q-learning update with one using our consistency penalty: $$\begin{gathered} \theta_{k} \leftarrow \theta_{k-1} + \sum_{(s, a, r, s')\in D_k} \alpha \Big[r + \gamma Q_{\theta_{k-1}}(s', \sigma(s')) \\ - Q_\theta(s, a)\Big] \nabla_\theta Q_\theta(s, a) +\alpha\lambda\nabla_\theta C_\theta(B) \Bigr\rvert_{\theta=\theta_{k-1}}. \label{eq:penalized_q_update}\end{gathered}$$ This scheme is quite general. First, it is agnostic as to how the prior action assignments are made (e.g., standard maximization w.r.t. the prior regressor as in DQN, Double DQN (DDQN) [@double-dqn], or other variants). It can also be used in conjunction with a search through alternate assignments (see below). Second, the consistency buffer $B$ may be populated in a variety of ways. Including all max-action choices from all past training batches promotes full consistency. However, this may be too constraining since action choices early in training are generally informed by inaccurate value estimates. $B$ may be implemented to focus only on more recent data (e.g., with a sliding recency window, weight decay, or subsampling); and the degree of recency bias may adapt during training (e.g., becoming more inclusive as training proceeds and the Q-function converges). Reducing the size of $B$ also has computational benefits. We discuss other ways of promoting consistency in Sec. \[sec:conclusion\]. The proposed consistency penalty resembles the temporal-consistency loss of @observe_look_further, but our aims are very different. Their temporal consistency notion penalizes changes in a next state’s Q-estimate over all actions, whereas we discourage inconsistencies in the greedy policy induced by the Q-estimator, regardless of the actual estimated values. The Search Space {#sec:searchspace} ---------------- ![A generic search tree.\[fig:searchtree\]](SearchTree.png){width="0.40\linewidth"} Ensuring optimality requires that PCQL track *all $\Theta$-consistent assignments*. While the set of such assignments has polynomial size [@LuEtAl_Qdelusion:nips18], it is impractical to track in realistic problems. As such, in [<span style="font-variant:small-caps;">ConQUR</span>]{} we recast information set tracking as a *search problem* and propose several strategies for managing the search process. We begin by defining the search space and discussing its properties. We discuss search procedures in Sec. \[sec:heuristics\]. As above, assume training data is divided into batches $D_1, \ldots, D_T$ and we have some initial Q-function estimate $\theta_0$ (for bootstrapping $D_1$’s labels). The regressor $\theta_{k}$ for $D_k$ can, in principle, be trained with labels generated by *any assignment* $\sigma \in \Sigma_\Theta(D_k)$ of actions to its successor states $\chi(D_k)$, not necessarily maximizing actions w.r.t. $\theta_{k-1}$. Each $\sigma$ gives rise to a different updated Q-estimator $\theta_{k}$. There are several restrictions we can place on “reasonable” $\sigma$-candidates: (i) $\sigma$ is $\Theta$-consistent; (ii) $\sigma$ is jointly $\Theta$-consistent with all $\sigma_j$, for $j<k$, used to construct the prior regressors on which we bootstrap $\theta_{k-1}$; (iii) $\sigma$ is not *dominated* by any $\sigma' \in \Sigma_\Theta(D_k)$, where we say $\sigma'$ dominates $\sigma$ if $Q_{\theta_{k-1}}(s',\sigma'(s')) \geq Q_{\theta_{k-1}}(s',\sigma(s'))$ for all $s'\in\chi(D)$, and this inequality is strict for at least one $s'$. Conditions (i) and (ii) are the strict consistency requirements of PCQL. We relax these below as discussed in Sec. \[sec:consistency\]. Condition (iii) is inappropriate in general, since we may add additional assignments (e.g., to new data) that render all non-dominated assignments inconsistent, requiring that we revert to some dominated assignment. This gives us a generic *search space* for finding policy-consistent, delusion-free Q-function (see Fig. \[fig:searchtree\]). Each node $n^i_k$ at depth $k$ in the search tree is associated with a regressor $\theta^i_k$ defining $Q_{\theta^i_k}$ and assignment $\sigma^i_k$ that justifies the labels used to train $\theta^i_k$ ($\sigma^i_k$ can be viewed as an information set). The root $n_0$ is based on an initial $\theta_0$, and has an empty assignment $\sigma_0$. Nodes at level $k$ of the tree are defined as follows. For each node $n_{k-1}^i$ at level $k-1$—with regressor $\theta_{k-1}^i$ and $\Theta$-consistent assignment $\sigma_{k-1}^i$—we have one child $n_k^j$ for each $\sigma_k^j\in\Sigma_\Theta(D_k)$ such that $\sigma_{k-1}^i \cup \sigma_k^j$ is $\Theta$-consistent. Node $n_k^j$’s assignment is $\sigma_{k-1}^i \cup \sigma_k^j$, and its regressor $\theta_k^i$ is trained using the data set: $$\{(s,a) \mapsto r + \gamma Q_{\theta_{k-1}^i} (s',\sigma_k^j(s')) \,:\, (s,a,r,s')\in D_k\}.$$ The entire search space constructed in this fashion to a maximum depth of $T$. See Appendix \[app:algos\], Algorithm \[alg:cellsearch\] for pseudocode of a simple depth-first recursive specification. The exponential branching factor in this search tree would appear to make complete search intractable; however, since we only allow $\Theta$-consistent “collective” assignments we can bound the size of the tree—it is *polynomial* in the VC-dimension of the approximator. \[thm:treesize\] The number of nodes in the search tree is no more than $O(nm\cdot [\binom{m}{2}n]^{\operatorname*{\mathsf{VCDim}}({\mathcal{G}})})$ where $\operatorname*{\mathsf{VCDim}}(\cdot)$ is the VC-dimension [@Vapnik1998] of a set of boolean-valued functions, and ${\mathcal{G}}$ is the set of boolean functions defining all feasible greedy policies under $\Theta$: ${\mathcal{G}}= \{ g_\theta(s, a, a') := {{\mathbf 1}}[f_\theta(s, a) - f_\theta(s, a') > 0], \forall s,a\ne a' ~|~ \theta \in \Theta\}.$ A linear approximator with a fixed set of $d$ features induces a policy-indicator function class ${\mathcal{G}}$ with VC-dimension $d$, making the search tree polynomial in the size of the MDP. Similarly, a fixed ReLU DNN architecture with $W$ weights and $L$ layers has VC-dimension of size $O(W L \log W)$ again rendering the tree polynomially sized. Even with this bound, navigating the search space exhaustively is generally impractical. Instead, various search methods can be used to explore the space, with the aim of reaching a “high quality” regressor at some leaf of the tree. Search Heuristics {#sec:heuristics} ----------------- Even with the bound in Thm. \[thm:treesize\], traversing the search space exhaustively is generally impractical. Moreover, as discussed above, enforcing consistency when generating the children of a node, and their regressors, may be intractable. Instead, various search methods can be used to explore the space, with the aim of reaching a “high quality” regressor at some (depth $T$) leaf of the tree. We outline three primary considerations in the search process: child generation, node evaluation or scoring, and the search procedure. [**Generating children.**]{} Given node $n^i_{k-1}$, there are, in principle, exponentially many action assignments, or children, $\Sigma_\Theta(D_k)$ (though Thm. \[thm:treesize\] limits this if we enforce consistency). Thus, we develop heuristics for generating a small set of children, driven by three primary factors. The first factor is a preference for generating *high-value assignments*. To accurately reflect the intent of (sampled) Bellman backups, we prefer to assign actions to state $s'\in\chi(D_k)$ with larger predicted Q-values i.e., a preference for $a$ over $a'$ if ${Q}_{\theta_{k-1}^j}(s',a)>{Q}_{\theta_{k-1}^j}(s',a')$. However, since the maximizing assignment may be $\Theta$-inconsistent (in isolation, jointly with the parent information set, or with future assignments), candidate children should merely have *higher probability* of a high-value assignment. Second, we need to ensure *diversity* of assignments among the children. Policy commitments at stage $k$ constrain the assignments at subsequent stages. In many search procedures (e.g., beam search), we avoid backtracking, so we want the stage-$k$ commitments to offer flexibility in later stages. The third factor is the degree to which we enforce consistency. There are several ways to generate high-value assignments. We focus on one natural technique: sampling action assignments using a Boltzmann distribution. Let $\sigma$ be the assignment of some node (parent) at level $k-1$ in the tree. We generate an assignment $\sigma_k$ for $D_k$ as follows. Assume some permutation $s_1', \ldots, s'_{|D_k|}$ of $\chi(D_k)$. For each $s'_i$ in turn, we sample $a_i$ with probability proportional to $e^{Q_{\theta_{k-1}}(s'_i,a_i)/\tau}$. This can be done *without regard to consistency*, in which case we use the consistency penalty when constructing the regressor $\theta_k$ for this child to “encourage” consistency rather than enforce it. If we want strict consistency, we can use rejection sampling without replacement to ensure $a_i$ is consistent with $\sigma^j_{k-1} \cup \sigma_{\leq i-1}$ (we can also use a subset of $\sigma^j_{k-1}$ as a less restrictive consistency buffer).[^4] The temperature parameter $\tau$ controls the degree to which we focus on maximizing assignments versus diverse, random assignments. While sampling gives some diversity, this procedure biases selection of high-value actions to states that occur early in the permutation. To ensure further diversity, we use a new random permutation for each child. [**Scoring children.**]{} Once the children of some expanded node are generated, we must assess the quality of each child to decide which new nodes to expand. One possiblity is to use the average Q-label (overall, or weighted using an initial distribution), Bellman error, or loss incurred by the regressor. However, care must be taken when comparing nodes at different depths of the tree. Since deeper nodes have a greater chance to accrue rewards or costs, simple calibration methods can be used. Alternatively, when a simulator is available, rollouts of the induced greedy policy can be used evaluate the node quality. However, rollouts incur considerable computational expense during training relative to the more direct scoring methods. [**Search Procedure.**]{} Given a method for generating and scoring children, different search procedures can be applied: best-first search, beam search, local search, etc. all fit very naturally within the [<span style="font-variant:small-caps;">ConQUR</span>]{} framework. Moreover, hybrid strategies are possible—one we develop below is a variant of beam search in which we generate multiple children only at certain levels of the tree, then do “deep dives” using consistency-penalized Q-regression at the intervening levels. This reduces the size of the search tree considerably and, when managed properly, adds only a constant-factor (proportional to beam size) slowdown to methods like DQN. An Instantiation of the [<span style="font-variant:small-caps;">ConQUR</span>]{} Framework {#sec:conqurdetails} ------------------------------------------------------------------------------------------ We now outline a specific instantiation of the [<span style="font-variant:small-caps;">ConQUR</span>]{} framework that effectively navigates the large search spaces that arise in practical RL settings. We describe a heuristic, modified beam-search strategy with backtracking and priority scoring. We outline only key features (see details in Algorithm \[alg:beam\], Appendix \[app:algos\]). Our search process alternates between two phases. In an *expansion phase*, parent nodes are expanded, generating one or more child nodes with assignments sampled from the Boltzmann distribution. For each child, we create target Q-labels, then optimize its regressor using consistency-penalized Bellman error Eq. \[eq:total\_loss\], foregoing strict policy consistency. In a *dive phase*, each parent generates *one* child, whose action assignment is given by standard max-action selection w.r.t. the parent’s regressor. No diversity is considered but we continue to use consistency-penalized regression. From the root, the search begins with an expansion phase to create $c$ children—$c$ is the *splitting factor*. Each child inherits its parent’s consistency buffer to which we add the new assignments used for that child’s Q-labels. To limit the tree size, we track a subset of the children (the *frontier*), selected using some scoring function. We select the top $\ell$-nodes for expansion, proceed to a dive phase and iterate. We consider backtracking strategies that return to unexpanded nodes at shallower depths of the tree below. Related Work {#sec:related} ------------ Other work has considered multiple hypothesis tracking in RL. One direct approach uses ensembling, with multiple Q-approximators updated in parallel [@fausser2015neural; @osband2016deep; @ADQN-2018] and combined to reduce instability and variance. Population-based methods, inspired by evolutionary search, are also used. @ns-2018 combine novelty search and quality diversity to improve hypothesis diversity and quality. @es-2018 augment an off-policy RL method with diversified population information derived from an evolutionary algorithm. These techniques do not target a specific weaknesses of Q-learning, such as delusion. Empirical Results {#sec:empirical} ================= We assess the performance of [<span style="font-variant:small-caps;">ConQUR</span>]{} using the Atari test suite [@bellemare:jair2013]. Since [<span style="font-variant:small-caps;">ConQUR</span>]{} directly tackles delusion, any performance improvement over Q-learning baselines strongly suggests the presence of delusional bias in the baselines in these domains. We first assess the impact of our consistency penalty in isolation (without search), treating it as a “regularizer” that promotes consistency with both DQN and DDQN. We then test our modified beam search to assess the full power of [<span style="font-variant:small-caps;">ConQUR</span>]{}. We do not directly compare [<span style="font-variant:small-caps;">ConQUR</span>]{} to policy gradient or actor-critic methods—which for some Atari games offer state-of-the-art performance [@model_atari:2019; @r2d2_atari:iclr2020]—because our aim with [<span style="font-variant:small-caps;">ConQUR</span>]{} is to improve the performance of (widely used) Q-learning-based algorithms. Consistency Penalization {#subsec:exp_reg} ------------------------ We first study the effects of augmenting both DQN and DDQN with soft-policy consistency in isolation. We train models using an open-source implementation of DQN and DDQN, using default hyperparameters [@tf_agents] . We refer to the consistency-augmented algorithms as ${\mathsf{DQN}{(\lambda)}}$ and ${\mathsf{DDQN}{(\lambda)}}$, respectively, where $\lambda$ is the penalty weight (see Eq. \[eq:total\_loss\]). When $\lambda = 0$, these correspond to DQN and DDQN themselves. This policy-consistency augmentation is lightweight and can be applied readily to any regression-based Q-learning method. Since we do not use search (i.e., do not track multiple hypotheses), these experiments use a small consistency buffer drawn only from the *current* data batch by sampling from the replay buffer—this prevents getting “trapped” by premature policy commitments. No diversity is used to generate action assignments—standard action maximization is used. We evaluate ${\mathsf{DQN}{(\lambda)}}$ and ${\mathsf{DDQN}{(\lambda)}}$ for $\lambda \in \{0.25, 0.5 , 1, 1.5, 2 \}$ on 19 Atari games.[^5] In training, $\lambda$ is initialized at $0$ and annealed to the desired value to avoid premature commitment to poor assignments.[^6] Unsurprisingly, the best $\lambda$ tends to differ across games depending on the extent of delusional bias. Despite this, $\lambda = 0.5$ works well across all games tested. Fig. \[fig:e1\] illustrates the effect of increasing $\lambda$ on two games. In Gravitar, it results in better performance in both ${\mathsf{DQN}{}}$ and ${\mathsf{DDQN}{}}$, while in SpaceInvaders, $\lambda = 0.5$ improves both baselines, but relative performance degrades at $\lambda=2$. We also compare performance on each game for each $\lambda$ value, as well as using the best $\lambda_{\textrm{best}}$ (see Fig. \[fig:all\_penalty\], Table \[tab:best\_reg\] in Appendix \[app:consis\_detailed\]). ${\mathsf{DQN}{(\lambda_{\text{best}})}}$ and ${\mathsf{DDQN}{(\lambda_{\text{best}})}}$ outperform their “potentially delusional” counterparts in all but 3 and 2 games, respectively. In 9 games, both ${\mathsf{DQN}{(\lambda_{\text{best}})}}$ and ${\mathsf{DDQN}{(\lambda_{\text{best}})}}$ beat *both* baselines. With a fixed $\lambda = 0.5$, ${\mathsf{DQN}{(\lambda)}}$ and ${\mathsf{DDQN}{(\lambda)}}$ each beat their respective baseline in 11 games. These results suggest that consistency penalization—independent of the general [<span style="font-variant:small-caps;">ConQUR</span>]{} model—can improve the performance of DQN and DDQN by addressing delusional bias. Moreover, promoting policy consistency appears to have a different effect on learning than double Q-learning, which addresses maximization bias. Indeed, consistency penalization, when applied to ${\mathsf{DQN}{}}$, achieves greater gains than ${\mathsf{DDQN}{}}$ in 15 games. Finally, in 9 games ${\mathsf{DDQN}{(\lambda)}}$ improves unaugmented ${\mathsf{DQN}{(\lambda)}}$. Further experiment details and results can be found in Appendix \[app:ablation\_penalty\]. ![Varying penalization $\lambda$ (no search procedure).[]{data-label="fig:e1"}](plots/reg_gravitar_spaceinv2.pdf){width="\columnwidth"} Full [<span style="font-variant:small-caps;">ConQUR</span>]{} {#subsec:exp_conqur} ------------------------------------------------------------- {width="\columnwidth"} \[fig:cells\_bar\] {width="\columnwidth" height="3.5cm"} \[fig:cells\] {width="\columnwidth"} \[fig:conqur\_reg\] ![Improvements of ConQUR($\lambda=10$) over multi-DQN baseline on all 59 games. A frontier $F=16$ nodes was used.[]{data-label="fig:winning_margin"}](plots/baseline_margin.pdf){width="0.98\columnwidth"} We test the full [<span style="font-variant:small-caps;">ConQUR</span>]{} framework using our modified beam search (Sec. \[sec:conqurdetails\]) on the full suite of 59 Atari games. Rather than training a full Q-network using [<span style="font-variant:small-caps;">ConQUR</span>]{}, we leverage pre-trained networks from the Dopamine package [@castro18dopamine],[^7] and use [<span style="font-variant:small-caps;">ConQUR</span>]{} to learn final layer weights, i.e., a new “linear approximator” w.r.t. the learned feature representation. We do this for two reasons. First, this allows us to test whether delusional bias occurs in practice. By freezing the learned representation, any improvements offered by [<span style="font-variant:small-caps;">ConQUR</span>]{} when learning a linear Q-function over those same features provides direct evidence that (a) delusion is present in the original trained baselines, and (b) [<span style="font-variant:small-caps;">ConQUR</span>]{} does in fact mitigate its impact (without relying on novel feature discovery). Second, from a practical point of view, this “linear tuning” approach offers a relatively inexpensive way to apply our methodology in practice. By bootstrapping a model trained in standard fashion and extracting performance gains with a relatively small amount of additional training (e.g., linear tuning requires many fewer training samples, as our results show), we can offset the cost of the [<span style="font-variant:small-caps;">ConQUR</span>]{} search process itself. We use DQN-networks with the same architecture as in @dqn-atari, trained on 200M frames as our baseline. We use [<span style="font-variant:small-caps;">ConQUR</span>]{} to retrain only the last (fully connected) layer (freezing other layers), which can be viewed as a linear Q-approximator over the features learned by the CNN. We train Q-regressors in [<span style="font-variant:small-caps;">ConQUR</span>]{} using *only 4M additional frames*.[^8] We use a splitting factor of $c=4$ and frontier size 16. The dive phase is always of length nine (i.e., nine batches of data), giving an expansion phase every ten iterations. Regressors are trained using soft-policy consistency (Eq. \[eq:total\_loss\]), with the consistency buffer comprising *all* prior action assignments. We run [<span style="font-variant:small-caps;">ConQUR</span>]{} with $\lambda \in \{1,10\}$ and select the best performing policy. We use larger $\lambda$ values than in Sec. \[subsec:exp\_reg\] since full [<span style="font-variant:small-caps;">ConQUR</span>]{} maintains multiple Q-regressors and can “discard” poor performers. This allows more aggressive consistency enforcement—in the extreme, with exhaustive search and $\lambda\to \infty$, [<span style="font-variant:small-caps;">ConQUR</span>]{} behaves like PCQL, finding a near-optimal greedy policy. See Appendix \[app:full\_atari\] for further details (e.g., hyperparameters) and results. We first test two approaches to scoring nodes: (i) policy evaluation using rollouts; and (ii) scoring using the loss function (Bellman error with soft consistency). Results on a small selection of games are shown in Table \[tab:scoringfn\]. While rollouts, unsurprisingly, tend to induce better-performing policies, consistent-Bellman scoring is competitive. Since the latter much less computationally intense, and does not require a simulator (or otherwise sampling the environment), we use it throughout our remaining experiments. We next compare [<span style="font-variant:small-caps;">ConQUR</span>]{} with the value of the pre-trained DQN. We also evaluate a “multi-DQN” baseline that trains multiple DQNs independently, warm-starting from the same pre-trained DQN. It uses the same number of frontier nodes as [<span style="font-variant:small-caps;">ConQUR</span>]{}, and is trained identically to [<span style="font-variant:small-caps;">ConQUR</span>]{}, but uses direct Bellman error (no consistency penalty). This gives DQN the same advantage of multiple-hypothesis tracking as [<span style="font-variant:small-caps;">ConQUR</span>]{} (without its policy consistency). **Rollouts** **Bellman + Consistency Penalty** -------------- -------------- ----------------------------------- BattleZone 33796.30 32618.18 BeamRider 9914.00 10341.20 Boxing 83.34 83.03 Breakout 379.21 393.00 MsPacman 5947.78 5365.06 Seaquest 2848.04 3000.78 SpaceInvader 3442.31 3632.25 StarGunner 55800.00 56695.35 Zaxxon 11064.00 10473.08 : Results of [<span style="font-variant:small-caps;">ConQUR</span>]{} with 8 nodes (split 2) on 9 games: comparing loss-based node scoring with scoring using rollouts.[]{data-label="tab:scoringfn"} We test on 59 games. [<span style="font-variant:small-caps;">ConQUR</span>]{} with frontier size 16 and expansion factor 4 and splitting factor 4 (16-4-4) with backtracking (as described in the Appendix \[app:full\_atari\]) results in significant improvements over the pre-trained DQN, with an average score improvement of 189%. The only games without improvement are Montezuma’s Revenge, Tennis, Freeway, Pong, PrivateEye and BankHeist. This demonstrates that, *even when simply retraining the last layer of a highly tuned DQN network*, removing delusional bias frequently improves policy performance significantly. [<span style="font-variant:small-caps;">ConQUR</span>]{} exploits the reduced parameterization to obtain these gains with only 4M frames of training data. A half-dozen games have outsized improvements over pre-trained DQN, including Venture (35 times greater value), ElevatorAction (23 times), Tutankham (5 times) and Solaris (5 times).[^9] We found that $\lambda=10$ provided the best performance across all games. Fig. \[fig:winning\_margin\] shows the percentage improvement of [<span style="font-variant:small-caps;">ConQUR</span>]{}[($\lambda=10$)]{} over the multi-DQN baseline for all 59 games. The improvement is defined as $(s_C - s_B)/|s_B|$ where $s_C$ and $s_B$ are the average scores (over 5 runs) of the policy generated by [<span style="font-variant:small-caps;">ConQUR</span>]{} and that by the multi-DQN baseline (16 nodes), respectively. Compared to this stronger baseline, [<span style="font-variant:small-caps;">ConQUR</span>]{} wins by a margin of at least 10% in 16 games, while 19 games see improvements of 1–10%, 16 games show little effect ($\pm 1$%) and 8 games show a decline of greater than 1%. Tables of complete results and figures of training curves (all games) appears in Appendix \[subapp:conqur\_results\], Table \[tab:allnoop\] and Fig. \[fig:allgames\_allreg\]. Figs. \[fig:cells\_bar\] and \[fig:cells\] (smoothed, best frontier node) show node policy values and training curves, respectively, for Solaris. When examining nodes ranked by their policy value (Fig. \[fig:cells\_bar\]), we see that nodes of any given rank generated by [<span style="font-variant:small-caps;">ConQUR</span>]{} dominate their by multi-DQN (baseline) counterparts: the three highest-ranked nodes exceed their baseline counterparts by 18%, 13% and 15%, respectively, while the remaining nodes show improvements of roughly 11–12%. Fig. \[fig:conqur\_reg\] (smoothed, best frontier node) shows the effect of varying $\lambda$. In Alien, increasing $\lambda$ from 1 to 10 improves performance, but performance starts to decline for higher $\lambda$ (we tested both 100 and 1000). This is similar to patterns observed in Sec. \[subsec:exp\_reg\] and represents a trade-off between emphasizing consistency and not over-committing to action assignments. In Atlantis, stronger penalization tends to degrade performance. In fact, the stronger the penalization, the worse the performance. Concluding Remarks {#sec:conclusion} ================== We have introduced [<span style="font-variant:small-caps;">ConQUR</span>]{}, a framework for mitigating delusional bias in various forms of Q-learning that relaxes some of the strict assumptions of exact delusion-free algorithms like PCQL to ensure scalability. Its main components are a search procedure used to maintain diverse, promising Q-regressors (and corresponding information sets); and a consistency penalty that encourages “maximizing” actions to be consistent with the approximator class. [<span style="font-variant:small-caps;">ConQUR</span>]{} embodies elements of both value-based and policy-based RL: it can be viewed as using partial policy constraints to bias the Q- value estimator, and as a means of using candidate value functions to bias the search through policy space. Empirically, we find that [<span style="font-variant:small-caps;">ConQUR</span>]{} can improve the quality of existing approximators by removing delusional bias. Moreover, the consistency penalty applied on its own, in either DQN or DDQN, can improve policy quality. There are many directions for future research. Other methods for nudging regressors to be policy-consistent include exact consistency (i.e., constrained regression), other regularization schemes that push the regressor to fall within the information set, etc. Further exploration of search, child-generation, and node-scoring strategies should be examined within [<span style="font-variant:small-caps;">ConQUR</span>]{}. Our (full) experiments should also be extended beyond those that warm-start from a DQN model. We believe our methods can be extended to both continuous actions and soft max-action policies. We are also interested in the potential connection between maintaining multiple “hypotheses” (i.e., Q-regressors) and notions in distributional RL [@bellemare:icml17]. An Example of Delusional Bias {#app:delusionexample} ============================= We describe an example, taken directly from [@LuEtAl_Qdelusion:nips18], to show concretely how delusional bias causes problems for Q-learning with function approximation. The MDP in Fig. \[fig:4chain\] illustrates the phenomenon: @LuEtAl_Qdelusion:nips18 use a linear approximator over a specific set of features in this MDP to show that: (a) No $\pi\in G(\Theta)$ can express the optimal (unconstrained) policy (which requires taking $a_2$ at each state); (b) The optimal *feasible* policy in $G(\Theta)$ takes $a_1$ at $s_1$ and $a_2$ at $s_4$ (achieving a value of $0.5$). (c) Online Q-learning (Eq. \[eq:online\_q\_update\]) with data generated using an $\varepsilon$-greedy behavior policy must converge to a fixed point (under a range of rewards and discounts) corresponding to a “compromise” admissible policy which takes $a_1$ at both $s_1$ and $s_4$ (value of $0.3$). Q-learning fails to find a reasonable fixed-point because of delusion. Consider the backups at $(s_2, a_2)$ and $(s_3, a_2)$. Suppose $\hat{\theta}$ assigns a “high” value to $(s_3,a_2)$, so that $Q_{\hat{\theta}}(s_3,a_2) > Q_{\hat{\theta}}(s_3,a_1)$ as required by $\pi_{\theta^{\ast}}$. They show that any such $\hat{\theta}$ also accords a “high” value to $(s_2, a_2)$. But $Q_{\hat{\theta}}(s_2,a_2) > Q_{\hat{\theta}}(s_2,a_1)$ is inconsistent the first requirement. As such, any update that makes the Q-value of $(s_2,a_2)$ higher *undercuts the justification* for it to be higher (i.e., makes the “max” value of its successor state $(s_3,a_2)$ lower). This occurs not due to approximation error, but the inability of Q-learning to find the value of the optimal *representable* policy. ![A simple MDP [@LuEtAl_Qdelusion:nips18]. []{data-label="fig:4chain"}](4chain_example.pdf){width="0.9\columnwidth"} Algorithms {#app:algos} ========== The pseudocode of (depth-first) version of the [<span style="font-variant:small-caps;">ConQUR</span>]{} search framework is listed in Algorithm \[alg:cellsearch\]. Data sets $D_k, D_{k+1}, \ldots D_{T}$; regressor ${\hat Q}_{k-1}$; and assignment $\sigma$ over $D_{\leq k-1}= \cup_{1\leq j \leq k-1} D_{j}$ reflecting prior data; policy class $\Theta$.\ Let $\Sigma_{\Theta,\sigma} = \{\sigma_k\in\Sigma_\Theta(D_j) : \sigma_k \cup \sigma \textrm{ is consistent}\}$ \[line:consistentsigma\] Training set $S\leftarrow \{\}$ $q \leftarrow r + \gamma {\hat Q}_{k-1}(s',\sigma_k^j(s'))$ $S\leftarrow S\cup\{((s,a),q)\}$ Train ${\hat Q}^j_{k}$ using training set $S$ \[line:training\] Return ${\hat Q}^j_{k}$ // terminate \[line:terminate\] Return <span style="font-variant:small-caps;">Search</span>($D_{k+1}, \ldots D_{T}; \, {\hat Q}^j_{k} ; \, \sigma_k^j \cup \sigma ; \, \Theta$) // recurse \[line:recurse\] As discussed in Sec. \[sec:conqurdetails\], a more specific instantiation of the [<span style="font-variant:small-caps;">ConQUR</span>]{} algorithm is listed in Algorithm \[alg:beam\]. Search control parameters: $m$, $\ell$, $c$, $d$, $T$ list of data batches $D_1,...,D_k$, initialized empty candidate pool $P$ of at most $m$ nodes, initialized $P=\{n_0\}$ frontier list $F$ of $\ell^c$ nodes for each node $n^i_k$ a regressor $\theta^i_k$ and an ancestor assignment $\sigma^i_k$ 1ex Find top scoring node $n^1\in P$ Use $\varepsilon$-greedy policy extracted from $Q_{\theta^1}$ to collect next data batch $D_k$ 1ex Select top $\ell$ scoring nodes $n^1,...,n^\ell\in P$ Generate $c$ children $n^{i,1},...,n^{i,c}$ using Boltzmann sampling on $D_k$ with $Q_{\theta^i}$ Let assignment history $\sigma^{i,j}$ be $\sigma^i\cup\{\textit{new assignment}\}$ Determine regressor $\theta^{i,j}$ by applying update from $\theta^i$ Score and add child nodes to the candidate pool $P$ Assign frontier nodes to set of child nodes, $F=\{n^{i,j}\}$ evict bottom scoring nodes, keeping top $m$ in $P$ 1ex Update regressor $\theta^{i,j}$ by applying update to $\theta^{i,j}$ 1ex Run $d$ “dive” levels after each expansion level 1ex Additional Detail: Effects of Consistency Penalization {#app:ablation_penalty} ====================================================== Delusional bias in DQN and DDQN ------------------------------- Both DQN and DDQN uses a delayed version of the $Q$-network $Q_{\theta^-}(s', a')$ for label generation, but in a different way. In DQN, $Q_{\theta^-}(s', a')$ is used for both value estimate and action assignment $\sigma_{\text{DQN}}(s') = \operatorname*{arg\!max}_{a'} Q_{\theta_{k}}(s', a')$, whereas in DDQN, $Q_{\theta^-}(s', a')$ is used only for value estimate and the action assignment is computed from the current network $\sigma_{\text{DDQN}}(s') = \operatorname*{arg\!max}_{a'} Q_{\theta_{k}}(s', a')$. With respect to delusional bias, action assignment of DQN is consistent for all batches after the latest network weight transfer, as $\sigma_{\text{DQN}}(s')$ is computed from the same $Q_{\theta^-}(s', a')$ network. DDQN, on the other hand, could have very inconsistent assignments, since the action is computed from the current network that is being updated at every step. Training Methodology and Hyperparameters ---------------------------------------- We implement consistency penalty on top of the DQN and DDQN algorithm by modifying the open-source TF-Agents library [@tf_agents]. In particular, we modify existing `DqnAgent` and `DdqnAgent` by adding a consistency penalty term to the original TD loss. We use TF-Agents implementation of DQN training on Atari with the default hyperparameters, which are mostly the same as that used in the original DQN paper [@dqn-atari]. For conveniece to the reader, some important hyperparameters are listed in Table \[tab:hps\_cp\]. The reward is clipped between $[-1, 1]$ following the original DQN. Hyper-parameter Value ------------------------------------------------ -------------------------------------------- Mini-batch size 32 Replay buffer capacity 1 million transitions Discount factor $\gamma$ 0.99 Optimizer RMSProp Learning rate 0.00025 Convolution channel $32, 64, 64$ Convolution filter size $(8 \times 8), (4 \times 4), (3 \times 3)$ Convolution stride 4, 2, 1 Fully-connected hidden units 512 Train exploration $\varepsilon_{\text{train}}$ 0.01 Eval exploration $\varepsilon_{\text{eval}}$ 0.001 Evaluation Methodology ---------------------- We empirically evaluate our modified DQN and DDQN agents trained with consistency penalty on 15 Atari games. Evaluation is run using the training and evaluation framework for Atari provided in TF-Agents without any modifications. Detailed Results {#app:consis_detailed} ---------------- Fig. \[fig:all\_penalty\] shows the effects of varying $\lambda$ on both DQN and DDQN. Table \[tab:best\_reg\] summarizes the best penalties for each game and their corresponding scores. Fig. \[fig:best\_penalty\] shows the training curves of the best penalization constants. Finally, Fig. \[fig:0.5\_penalty\] shows the training curves for a fixed penalization of $\lambda=0.5$. The datapoints in each plot of the aforementioned figures are obtained by averaging over window size of 30 steps, and within each window, we take the largest policy value (and over $\approx$2–5 multiple runs). This is done to reduce visual clutter. {width="1.4\columnwidth"} ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ${\mathsf{DQN}{}}$ $\lambda_{\text{best}}$ ${\mathsf{DQN}{(\lambda_{\text{best}})}}$ ${\mathsf{DDQN}{}}$ $\lambda'_{\text{best}}$ ${\mathsf{DDQN}{(\lambda'_{\text{best}})}}$ --------- -------------------- ------------------------- ----------------------------------------------------------------------------------- --------------------- -------------------------- --------------------------------------------- Assault 2546.56 1.5 **[3451.07]{} & 2770.26 & 1 & 2985.74\ Atlantis & 995460.00 & 0.5 & **[1003600.00]{} & 940080.00 & 1.5 & 999680.00\ BattleZone & **[67500.00]{} & 2 & 55257.14 & 47025.00 & 2 & 48947.37\ BeamRider & 7124.90 & 0.5 & **[7216.14]{} & 5926.59 & 0.5 & 6784.97\ Boxing & 86.76 & 0.5 & 90.01 & 82.80 & 0.5 & **[91.29]{}\ Breakout & 220.00 & 0.5 & 219.15 & 214.25 & 0.5 & **[242.73]{}\ Enduro & 1206.22 & 0.5 & **[1430.38]{} & 1160.44 & 1 & 1287.50\ Gravitar & 475.00 & 1.5 & **[685.76]{} & 462.94 & 1.5 & 679.33\ JourneyEscape & -1020.59 & 0.25 & -696.47 & -794.71 & 1 & **[-692.35]{}\ MsPacman & **[4104.59]{} & 2 & 4072.12 & 3859.64 & 0.5 & 4008.91\ NameThisGame & 7230.71 & 1 & 9013.48 & 9618.18 & 0.5 & **[10210.00]{}\ Qbert & 13270.64 & 0.5 & **[14111.11]{} & 13388.92 & 1 & 12884.74\ Seaquest & 5849.80 & 1 & 6123.72 & **[12062.50]{} & 1 & 7969.77\ SpaceInvaders & 2389.22 & 0.5 & 2707.83 & 3007.72 & 0.5 & **[4080.57]{}\ StarGunner & 40393.75 & 0.5 & 55931.71 & 55957.89 & 0.5 & **[60035.90]{}\ TimePilot & 4205.83 & 2 & 7612.50 & 6654.44 & 2 & **[7964.10]{}\ Tutankham & 222.76 & 1 & **[265.86]{} & 243.20 & 0.25 & 247.17\ VideoPinball & **[569502.19]{} & 0.25 & 552456.00 & 509373.50 & 0.25 & 562961.50\ Zaxxon & 5533.33 & 1 & **[10520.00]{} & 7786.00 & 0.5 & 10333.33\ ************************************** ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- {width="1.4\columnwidth"} {width="1.4\columnwidth"} {width="1.4\columnwidth"} Additional Detail: [<span style="font-variant:small-caps;">ConQUR</span>]{} Results {#app:full_atari} =================================================================================== Our results use a frontier queue of size ($F$) 16 (these are the top scoring leaf nodes which receive gradient updates and rollout evaluations during training). To generate training batches, we select the best node’s regressor according to our scoring function, from which we generate training samples (transitions) using $\varepsilon$-greedy. Results are reported in Table \[tab:allnoop\], and training curves in Fig. \[fig:allgames\_allreg\]. We used Bellman error plus consistency penalty as our scoring function. During the training process, we also calibrated the scoring to account for the depth difference between the leaf nodes at the frontier versus the leaf nodes in the candidate pool. We calibrated by taking the mean of the difference between scores of the current nodes in the frontier with their parents. We scaled this difference by multiplying with a constant of 2.5. In our implementation, we initialized our Q-network with a pre-trained DQN. We start with the expansion phase. During this phase, each parent node splits into $l$ children nodes and the Q-labels are generated using action assignments from the Boltzmann sampling procedure, in order to create high quality and diversified children. We start the dive phase until the number of children generated is at least $F$. In particular, with $F = 16$ configuration, we performed the expansion phase at the zero-th and first iterations, and then at every tenth iteration starting at iteration 10, then at 20, and so on until ending at iteration 90. All other iterations execute the “dive” phase. For every fifth iteration, Q-labels are generated from action assignments sampled according to the Boltzmann distribution. For all other iterations, Q-labels are generated in the same fashion as the standard Q-learning (taking the max Q-value). The generated Q-labels along with the consistency penalty are then converted into gradient updates that applies to one or more generated children nodes. Training Methodology and Hyperparameters ---------------------------------------- Each iteration consists of 10k transitions sampled from the environment. Our entire training process has 100 iterations which consumes 1M transitions or 4M frames. We used RMSProp as the optimizer with a learning rate of $2.5\times 10^{-6}$. One training iteration has 2.5k gradient updates and we used a batch size of 32. We replace the target network with the online network every fifth iteration and reward is clipped between $[-1, 1]$. We use a discount value of $\gamma =0.99$ and $\varepsilon$-greedy with $\varepsilon=0.01$ for exploration. Details of hyper-parameter settings can be found in Table \[tab:hps\], \[tab:hps\_nodes\]. Evaluation Methodology ---------------------- We empirically evaluate our algorithms on 59 Atari games [@bellemare:jair2013], and followed the evaluation procedure as in @double-dqn. We evaluate our agents on every 10-th iteration (and also the initial and first iteration) by suspending our training process. We evaluate on 500k frames, and we cap the length of the episodes for 108k frames. We used $\varepsilon$-greedy as the evaluation policy with $\varepsilon =0.001$. We evaluated our algorithm under the *no-op starts* regime—in this setting, we insert a random number of “do-nothing” (or *no-op*) actions (up to 30) at the beginning of each episode. Detailed Results {#subapp:conqur_results} ---------------- Fig. \[fig:allgames\_allreg\] shows training curves of [<span style="font-variant:small-caps;">ConQUR</span>]{} with 16 nodes under different penalization strengths $\lambda\in \{1, 10\}$. While each game has its own optimal $\lambda$, in general, we found that $\lambda=10$ gave the best performance for most games. Each plotted step of each training curve (including the baseline) shows the best performing node’s policy value as evaluated with full rollouts. Table \[tab:allnoop\] shows the summary of the highest policy values achieved for all 59 games for [<span style="font-variant:small-caps;">ConQUR</span>]{} and the baseline under 16 nodes. Both the baseline and [<span style="font-variant:small-caps;">ConQUR</span>]{} improve overall, but [<span style="font-variant:small-caps;">ConQUR</span>]{}’s advantage over the baseline is amplified. These results all use a splitting factor of $c=4$. (We show results with 8 nodes and a splitting factor of 2 below.) {width="90.00000%"} {width="70.00000%"} Additional Results: [<span style="font-variant:small-caps;">ConQUR</span>]{} with 8 Nodes ----------------------------------------------------------------------------------------- As an additional study of [<span style="font-variant:small-caps;">ConQUR</span>]{}, we present results of the running our method using 8 nodes (rather than the 16 used above), and compare it to a multi-DQN baseline that also uses 8 “nodes” (i.e., 8 separate DQN runs). We use a splitting factor $c=2$ for ${\textsc{ConQUR}}{}.$ Table \[tab:allnoop\_8\] shows the average scores for each game using [<span style="font-variant:small-caps;">ConQUR</span>]{} and the baseline with 8 nodes. Unsurprisingly, [<span style="font-variant:small-caps;">ConQUR</span>]{} with 8 nodes does not perform as well as [<span style="font-variant:small-caps;">ConQUR</span>]{} with 16 nodes; but as in the 16-node case, [<span style="font-variant:small-caps;">ConQUR</span>]{} outperforms the baseline when each uses 8 nodes. More importantly, the average improvement of $24.5\%$ for [<span style="font-variant:small-caps;">ConQUR</span>]{} with 16 nodes over the corresponding baseline *exceeds* the $19.6\%$ improvement of [<span style="font-variant:small-caps;">ConQUR</span>]{} in the 8-node case. This is a strong indication that increasing the number of nodes increases the performance gap *relative* to the corresponding multi-DQN baseline; this implies that a good search heuristic is critical to effectively navigate the search space (as compared to randomly selected nodes) with a greater number of candidate hypotheses.[^10] [^1]: We describe our approach using batch Q-learning, but it can accommodate many variants, e.g., where the estimators generating max-actions and value estimates are different, as in double Q-learning [@hasselt:2010; @double-dqn]; indeed, we experiment with such variants. [^2]: While delusion may not arise in other RL approaches (e.g., policy iteration, policy gradient), our contribution focuses on mitigating delusion to derive maximum performance from widely used Q-learning methods. [^3]: We suppress mention of $D$ when clear from context. [^4]: Notice that at least one action for state $s'_i$ must be consistent with any previous (consistent) information set. [^5]: These 19 games were selected arbitrarily simply to test soft-consistency in isolation. See Appendix \[app:ablation\_penalty\] for details. [^6]: The annealing schedule is $\lambda_t = \lambda_{\textrm{final}} t / (t+2\times 10^6)$. Without annealing, the model tends anchor on poorly informed assignments during early training, adversely impacting performance. [^7]: See [^8]: This reduces computational/memory footprint of our experiments, and suffices since we re-train a simpler approximator. Nothing in the framework *requires* this reduced training data. [^9]: This may be in part, but not fully, due to the sticky-action training of the pre-trained model. [^10]: Average score improvements exclude games where the baseline score is zero.

--- bibliography: - 'icrc0180.bib' title: The Northern Site of the Pierre Auger Observatory --- Introduction ============ This paper describes the current design of the Northern Auger Observatory in the context of the science of ultra-high energy (UHE) cosmic rays. The design takes into consideration both the initial science results from the Southern Auger Observatory and our experience with the technologies and methods used. The need for two observatories, one in each hemisphere, for complete sky coverage at the highest energies was clear from the inception of the Auger Project. The Southern observatory site will be completed in 2007[@icrc299]. The Southern Observatory with its 1.5 km triangular spacing and an area of 3,000 $ \mathrm{km}^2 $ will be able to measure accurately the spectrum and composition from below $10^{18}$ eV to about $10^{20}$ eV[@icrc318; @icrc594; @icrc596; @icrc602]. The statistics above $10^{19}$ eV are sufficient to identify the GZK feature[@greisen; @zk], but marginal for definitive studies of the source distribution by looking for strong anisotropies in the distribution of arrival directions[@icrc074; @icrc075; @icrc076]. However, the data indicate that the bending power of extragalactic magnetic fields is small enough to do charged particle astronomy above $10^{19}$ eV and to therefore be able to observe the sources of ultra-high energy cosmic rays, given sufficient aperture. This is the main goal of the planned Northern Auger Observatory. ![Exposures above $10^{19}$ eV of Auger North and Auger South as a function of time. Also indicated are the expected exposures of the Telescope Array and the final exposures of the HiRes (monocular) and AGASA experiments.[]{data-label="fig:exposure-time"}](icrc0180_fig01.eps){width="0.90\columnwidth"} Auger North will retain the basic functionality and features of Auger South. This is important for seamless data integration, e.g. for an anisotropy analysis on the whole sky. The Northern hemisphere is chosen to be at roughly the same latitude and elevation as the Southern site. An important site feature is the usable area both for initial deployment and possible future expansion. The chosen site in Southeast Colorado has an initial area of 4,000 square miles (10,370 $\mathrm{km}^2 $), 3.3 times larger than Auger South. Deployment of the *Surface Detectors* (SD) is greatly facilitated when they are placed at the corners of a *square-mile grid*, corresponding to the grid of roads that exists in Southeast Colorado. Fluorescence Detectors (FD) will again be used for calibration of the SD, as well as hybrid analysis with accurate composition information and superior angular resolution on a subset of events. ![Average number of 5$\sigma$ source candidates over the lifetime of the full Auger Observatory for events above $10^{20}$ eV. Black triangles show fake sources from statistical fluctuations, whereas red circular points show the expected number of source candidates.[]{data-label="fig:Nsources"}](icrc0180_fig02.eps){width="0.90\columnwidth"} Science ======= The spectrum and composition of UHECRs below $10^{19}$ eV is most likely the same in both hemispheres as extragalactic particles below this energy can reach Earth from the entire universe and galactic ones are isotropized by magnetic fields. Spectral and composition differences may occur once isotropy is broken. As data accumulate above $10^{19}$ eV, departure from isotropy is expected both from the limited horizon in particle propagation and the weakening of the effects of cosmic magnetic fields. Fig. \[fig:exposure-time\] shows the expected accumulated exposure above $10^{19}$ eV of Auger South, Auger North, and Auger South+North as a function of time, assuming the construction of Auger North begins in 2009 and is completed in 2012. Fig. \[fig:Nsources\] shows the expected number of candidate “point” sources detected for Auger South alone by 2014, for both Auger North and South by 2014, and for Auger North and South combined by 2030. The number of source candidates was found by generating maps for each exposure for energies above $10^{20}$ eV and for source densities of $10^{-5}$ Mpc$^{-3}$, $10^{-4}$ Mpc$^{-3}$, and $10^{-3}$ Mpc$^{-3}$. The average intensity of each source is adjusted to match the observed spectrum of cosmic rays. Isotropic maps were used to estimate the number of fake sources. The large exposure and full sky coverage provided by Auger North will reward us with the detection of 15 to 40 sources by 2030. In recent years, the great potential for discoveries in UHE neutrino detections has triggered several experiments, which cover energies from $10^{14}$ eV up to $10^{26}$ eV. Given the expected shape of the cosmogenic neutrino flux, which peaks around $10^{18}$ eV, the combination of both Auger sites provides the best chance to detect cosmogenic neutrinos[@icrc607]. Implementation ============== The layout of the planned Auger North Observatory is indicated in Fig. \[fig:fd\_layout\]. Surface detectors are situated on a square-mile grid covering a 84x48 mile area in the Southeast corner of Colorado. Three FD eyes overlook the area to provide hybrid coverage. The square-mile grid layout of the Surface Detector will slightly decrease the acceptance for small hadron showers yielding an increase of the threshold energy. The efficiency is $>90\%$ for hadron showers with 5 triggered detectors for energies above $10^{19}$ eV, while in Auger South it is $3\times10^{18}$ eV. ![Topographic map of the Auger North site with the fields of view of the 3 fluorescence detector eyes indicated.[]{data-label="fig:fd_layout"}](icrc0180_fig03.eps){width="0.55\columnwidth"} Surface Detectors ----------------- The surface detector electronics planned for Auger North is a natural evolution of that which is used in Auger South. The philosophy of real-time station control software, triggering based upon processing flash ADC traces, and GPS based time-stamping, which work so well in Auger South, will remain the same. However, parts obsolescence requires a redesign of the electronics. One of the improvements to the electronics is increased dynamic range. Conversely, as a cost saving measure, the number of PMTs per tank is reduced from three to one. Studies in Auger South indicate that this does not significantly degrade either the triggering or the reconstruction of the highest energy events. The integration of the electronics will be increased in order to reduce cabling and improve reliability. Increasing the FADC sampling rate from 40MHz to 100MHz compensates for the reduction in PMTs. The station controller operating system will be changed to a variant of real-time Linux. Unlike the Auger South tanks, Auger North tanks will require thermal insulation. One technique being developed is rotationally molded polyethylene foam insulation on the interior of the tanks. This technique is commonly used to increase the stiffness of the walls of parts being roto-molded. The Auger North tank design has the main access port in the center for the single main PMT. Fluorescence Detector ===================== The Auger North FD will be split into 3 half eyes, in order to maximize the number of hybrid events. The design of the FD eyes is similar to that of the South. The HEAT enhancement telescopes[@icrc065] serve as a prototype for the North. Communications Network ---------------------- Design of the SD communications system for the North takes advantage of advances during the past decade in wireless network communications. The southern tanks each communicate independently with local collectors situated on towers at the FD buildings. Point-to-point microwave links to the campus complete the system. This scheme works well at the southern site, where the FDs and the towers are situated substantially higher than the remarkably flat intervening terrain. The topography of the Southeast Colorado makes this architecture less suitable for the North. Fig. \[fig:unreachable\] shows the results of a study, using digital elevation maps (DEM) of the site, to determine how many of the 4,000 stations would not have a clear line of site to a collector. Three different scenarios were considered: 1) each tank communicates to a tower-mounted base station as in Auger South; 2) Mini-clusters, where each station communicates with a local tower, which are then networked together; 3) A peer-to-peer network where each station communicates with one or more of its nearest neighbors. The peer-to-peer network has many fewer problematic links, and we are thus pursuing that option for Auger North. ![Number of stations (out of 4000) without a line of site communications link.[]{data-label="fig:unreachable"}](icrc0180_fig04.eps){width="0.90\columnwidth"} A network of fiber optic cables crossing the site will be used to make the trunk connections to the central campus facility (instead of the microwave links used in the South). Data Acquisition ---------------- For Auger South, a comprehensive Central Data Acquisition System (CDAS) was developed. CDAS includes both the hardware and software required to collect incoming data packets from both FD and SD systems, form and relay triggers, and to save and organize experiment data online. Minimal changes will be required to adapt the existing Auger South CDAS system for Auger North. Summary ======= By pioneering charged particle astronomy, Auger North will address some of the most compelling questions in science today: - Where do the highest energy particles that reach the Earth originate? - What process in nature can reach such extremely high energies? - What clues to these particles and their interactions offer about the universe and its fundamental laws? Answering these questions will transform our view of the most energetic sites in the present Universe. Acknowledgments =============== The presenter gratefully acknowledges the support of the US Department of Energy and thanks the Michigan Tech Vice-President for Research for providing additional funding.

[SLAC–PUB–8965\ October 2001]{} [[**$B$ Physics and $CP$ Violation**]{}[^1]]{} [**Abstract** ]{} These lectures provide a basic overview of topics related to the study of $CP$ Violation in $B$ decays. In the first lecture, I review the basics of discrete symmetries in field theories, the quantum mechanics of neutral but flavor-non-trivial mesons, and the classification of three types of $CP$ violation [@texts]. The actual second lecture which I gave will be separately published as it is my Dirac award lecture and is focussed on the separate topic of strong $CP$ Violation. In Lecture 2 here, I cover the Standard Model predictions for neutral $B$ decays, and in particular discuss some channels of interest for $CP$ Violation studies. Lecture 3 reviews the various tools and techniques used to deal with the hadronic physics effects. In Lecture 4, I briefly review the present and planned experiments that can study $B$ decays. I cannot teach all the details of this subject in this short course, so my approach is instead to try to give students a grasp of the relevant concepts and an overview of the available tools. The level of these lectures is introductory. I will provide some references to more detailed treatments and current literature, but this is not a review article so I do not attempt to give complete references to all related literature. By now there are some excellent textbooks that cover this subject in great detail [@texts]. I refer students to these for more details and for more complete references to the original literature. [*Lectures given at the\ Particle Physics School\ International Centre for theoretical Physics\ Trieste, Italy\ 2–6 July 2001* ]{}\ Lecture 1: Preliminaries: Symmetries, Hermiticity, Rephasing Invariance ======================================================================= We begin with the basics of symmetries in Lagrangian Field Theory. Physicists use the term symmetry to denote an invariance of the Lagrangian, and thus of the associated equations of motion, under some change of variables. Such changes can be local, that is coordinate dependent, or global; and they can be a continuous set or a discrete set of changes. The value of such symmetries lies in the simplification they achieve by limiting possible terms in the Lagrangian and by their relationship to conservation laws and the conserved quantum numbers that then characterize physical states. The invariance may be with respect to coordinate redefinitions, as in the case of Lorentz Invariance, or field redefinitions, as in the case of gauge invariance. The particular invariances of interest to us in these lectures are the global discrete invariances known as $C$, $P$, and $T$. These are charge conjugation or $C$ (replacement of a field by its particle-antiparticle conjugate), parity or $P$ (sign reversal of all spatial coordinates), and time reversal or $T$ (sign reversal of the time coordinate, which reverses the role of in and out states). Table \[tab:1\] shows the effect of these operations on a Dirac spinor field $\psi$, and Table \[tab:2\] summarizes the effect of the particular combination $CP$ on some quantities that appear in a gauge theory Lagrangian. In Table \[tab:2\], the symbol $(-1)^\mu$ denotes a factor +1 for $\mu=0$ and -1 for $\mu=1,2,3$. ------------------------- -------------------------------------------------- $P\psi(t,x)P\!$ $=\gamma^0\psi(t,-x)$ , \[1ex\] $T\psi(t,x)T\!$ $=-\gamma^1\gamma^3\psi(-t,x)$ , \[1ex\] $C\psi(t,x)C\!$ $=-i({\overline{\psi}}(t,x)\gamma^0\gamma^2)^T $ ------------------------- -------------------------------------------------- : The operation of $P$,$C$, and $T$ on a Dirac spinor field \[tab:1\] ------------------------------- ------------------------------ --------------------------------------- ------------------------------------------------ -------------------------------------------------------- term ${\overline{\psi}}_i\psi_j$ $i{\overline{\psi}}_i\gamma^5\psi_j$ ${\overline{\psi}}_i\gamma^\mu\psi_j$ ${\overline{\psi}}_i\gamma^\mu\gamma^5\psi_j$ \[1ex\] $CP$-transformed term ${\overline{\psi}}_j\psi_i $ $-i{\overline{\psi}}_j\gamma^5\psi_i$ $-(-1)^\mu{\overline{\psi}}_j\gamma^\mu\psi_i$ $-(-1)^\mu{\overline{\psi}}_j\gamma^\mu\gamma^5\psi_i$ \[1ex\] term $H$ $A$ $W^{\pm\mu}$ $\partial_\mu$ \[1ex\] $CP$-transformed term $H$ $-A$ $-(-1)^\mu W^{\mp\mu}$ $(-1)^\mu\partial_\mu$ ------------------------------- ------------------------------ --------------------------------------- ------------------------------------------------ -------------------------------------------------------- : The effect of a $CP$ transformation on various quantities \[tab:2\] When constructing a field theory we always require locality, the symmetries of Lorentz Invariance, and hermiticity of $\cal L$. That is sufficient to make any field theory automatically also invariant under the product of operations $CPT$. In many theories, for example for QED with fermion masses included, the combination $CP$, and thus also $T$ are also separately automatic. This is the reason why the experimental discovery that $CP$ is not an exact symmetry of nature caused such a stir. All the field theories that had been studied up to that time had automatic $CP$ conservation. So we need to examine how $CP$ non-conservation manifests itself, and then ask what theories will give such effects. $CP$ non-conservation shows up, for example, as a rate difference between two processes that are the $CP$ conjugates of one-another. How can such a rate difference appear? Consider a particle decay for which two different terms in the Lagrangian (two different Feynman diagrams) give possible contributions. The amplitude for such a process can be written as $$A= A(A \rightarrow B) = g_1 r_1 e^{i\phi_1} + g_2 r_2 e^{i\phi_2} \ .$$ Here $g_1$ and $g_2$ are two different, possibly complex, coupling constants in the theory. The transition amplitudes corresponding to each coupling are written as $re^{i\phi}$ to emphasize that they too can have both a real part or magnitude and a phase or absorptive part. The physical source of this phase is that there may be multiple real intermediate states which can contribute to the process in question via rescattering effects. In the jargon of the field the phases $\phi$ are called strong phases because the rescattering effects among the various coupled channels are dominated by strong interactions. These phases are the same for a process and its $CP$ conjugate because the $CP$-related sets of intermediate states must contribute the same absorptive part to the two processes. The phases of the coupling constants are often called weak phases because, in the Standard Model, the relevant complex couplings are in the weak interaction sector of the theory. When we look at the amplitude for the $CP$ conjugate process we find $${\overline{A}} = A({\overline{A}}\rightarrow {\overline{B}}) = g_1^* r_1 e^{i\phi_1} + g_2^* r_2 e^{i\phi_2} \ .$$ Note that the phases of the coupling constants change sign between any process and its $CP$ conjugate process, while the strong phases, which arise from absorptive parts in the amplitudes, do not. So now let us calculate the $CP$-violating difference in rates for these two processes. With a little algebra we find $$|A|^2 -|{\overline{A}}|^2 = 2r_1 r_2 {\rm Im} g_1 g_2^* \sin (\phi_1 -\phi_2) \ . \label{ratediff}$$ This shows that the effect will vanish if the two coupling constants can be made relatively real. In addition it depends on the difference of strong phases in the two amplitude contributions, and vanishes if this quantity is zero. Such a $CP$ violation in the comparison of two $CP$-related decay rates is often called direct $CP$ violation. I prefer the more descriptive term $CP$ violation in the decay amplitudes. Whatever you choose to call it, this effect is characterized by the condition $|{\overline{A}}/A|\neq 1$. It is obvious that in any process where there is only a single contributing term in the decay amplitude the phase of the coupling constant is irrelevant and $|{\overline{A}}/A|=1$ is automatic. You need two different couplings contributing, with non-zero relative phase of the two couplings to see any $CP$ violation. This statement applies for al types of $CP$ violation. The phase of any single complex coupling in a Lagrangian is not a physically meaningful quantity. In general it can be redefined, and even made to vanish by simply redefining some field or set of fields by appropriate phase factors. But such rephasing of fields can never change the relative phase between two couplings (or products of couplings) that contribute to the same process. Both contributing terms must involve the same nett set of fields, and hence both change in the same way under any rephasings of those fields. These rephasing-invariant quantities are the physically meaningful phases in any Lagrangian, the existence of such a quantity signals the possibility of $CP$ violation. The second feature we note is that the $CP$-violating rate difference in Eq. (\[ratediff\]) also depends on a difference of strong phases. Typically, this makes it difficult to calculate. Strong phases are, in general, long-range strong interaction physics effects, not amenable to perturbative calculation. One of the things that makes the decays of neutral but flavored mesons particularly interesting is that there we find other types of $CP$-violation effects where the role played here by the strong phases is replaced by other coupling constant phases, those relevant to the processes that mix the meson with its $CP$ (and thus also flavor) conjugate meson. In such a case we may be able to relate a measured $CP$ violation directly to phase-differences in the Lagrangian couplings, with no need to calculate any strong-interaction quantities. Only in the case of neutral but flavor non-trivial mesons can such mixing-dependent effects occur. We have seen that only a theory with two coupling constants that are not relatively real can give $CP$ violation. Thus we only can have $CP$ violation in a theory where there is some set of couplings for which rephasing of all fields cannot remove all phases. $CP$ conservation is automatic for any theory for which the most general form of the Lagrangian allows all complex phases to be removed by rephasing of some set of fields. Let us examine a few of the terms that occur in the QED Lagrangian to see why $CP$ conservation is automatic in that theory. For the gauge coupling terms we have, after requiring hermiticity $$g A^\mu {\overline{\psi}} \gamma_\mu \psi +g^* A_\mu {\overline{\psi}} \gamma^\mu \psi \ .$$ Thus hermiticity clearly makes the QED gauge coupling real, ($g +g^*$), because the term it multiplies is itself a hermitian quantity. After imposing hermiticity you will find that the fermion mass term must take the form $$Re( m){\overline{\psi}} \psi +i{\rm Im} (m){\overline{\psi}} \gamma_5\psi$$ for any complex $m$. Hermiticity alone does not require that the fermion mass be real, but it does require that the imaginary part multiplies a factor of $\gamma_5$. But a chiral rephasing of the fermion field $\psi \rightarrow e^{i\phi \gamma_5}\psi $ can be made. This does not change the kinetic or gauge coupling terms at all. In QED, one can always choose the angle $\phi$ in this rotation in such a way that it makes m a real quantity. This tells us that, in such a theory, the phase of m is not a physically meaningful quantity. Hence the theory is indeed automatically $CP$ conserving for any choice of $m$. (It is merely for convenience that we always choose to write QED with real particle masses; it is unnecessary to include additional parameters that you know are irrelevant to complicate your calculations.) Tomorrow we will see that this same rephasing is not so innocuous in $QCD$, and how this leads to the strong $CP$ problem [@dirac]. Given these examples you may be beginning to wonder how we ever get a $CP$ violating coupling into a Lagrangian field theory. That is the question that puzzled everyone in 1964. The trick is to have a sufficient number of different terms in the Lagrangian involving the same set of fields. For example imagine a theory with multiple flavors of fermions and multiple scalar fields. In such a theory there can be Yukawa couplings of the form $Y_{ijk} \phi_k{\overline{\psi}}_i \psi_j$. Hermiticity then requires only that we also have a term $Y^*_{ijk} \phi^*_k {\overline{\psi}}_j\psi_i$ in the Lagrangian. Note that this is a different product of fields from the original term, so hermiticity does not disallow phases for the various $Y_{ijk}$ in such a theory. But we still must ask whether we can make every such coupling real, by systematically redefining the phases of the various fields. That depends on the details of the theory. As we add more fields of a given type, either fermions or scalars, the number of possible coupling terms grows more rapidly than total number of fields. With enough fields of the each type there will be more couplings that there are possible phase redefinitions, and then not all couplings can be made real by rephasing the fields. We can always make all couplings real by imposing $CP$ invariance as a postulate, but it no longer an automatic feature of the theory. The Standard Model with only one Higgs doublet and only two fermion generations has automatic $CP$ invariance; all possible couplings can be made simultaneously real (ignoring for now the issue of strong $CP$-violation via a QCD-theta parameter). Adding one more generation of fermions or adding an additional Higgs doublet with no further symmetries imposed opens up the possibility of $CP$ violating couplings [@CPVpossibilities]. The three generation Standard Model with a single Higgs doublet has only one $CP$-violating parameter, that is only one independent phase difference survives after as many couplings as possible are made real by field rephasing. This means that all $CP$-violating effects in this theory are related. That is what makes it so interesting to test the pattern of $CP$ violation in $B$ decays. Here there are many different channels in which possible $CP$-violating effects may be observed. In the Standard Model there are predicted relationships between these effects, and between $CP$ violating effects and the values of other $CP$-conserving Standard Model parameters. Thus the patterns of the $B$ decays, as well as their relationships to the observed $CP$ violation in $K$-decays, provide ways to test for the effects of physics beyond the Standard Model. Such effects can disrupt the predicted Standard Model relationships between the different measurements. Quantum Mechanics of Neutral Mesons ----------------------------------- We now we turn to a general discussion of the physics of flavored neutral mesons, those made from different quark and antiquark types of the same charge. These are the $K$, $D$, $B_d$ and $B_s$ mesons, which we denote generically by $M^0$. (I use the notation $B_d$ as a reminder of the quark content, even though the official name of this particle is simply $B^0$.) There is a beautiful quantum mechanical story here. In each case there are two $CP$-conjugate flavor eigenstates, $M^0 = {\overline{q}} q^\prime$ and ${\overline{M}}^0 ={\overline{q}}^\prime q$. In general $CP M^0 = e^{i\xi} {\overline{M}}^0 $. The phase $\xi$ is convention dependent and can be altered by redefining one or other of the quark fields by a phase. In much of the literature on this subject the convention $\xi=0$ is chosen without comment, but elsewhere $\xi=\pi$ is used. Physical results are convention independent, but only as long as you consistently use the same convention. You can get into trouble if you combine formulae taken from two different sources without first checking that both are using the same convention. From this point on I will use the convention $\xi=0$; if you want to see the equations with arbitrary phase factors explicitly displayed, go to the textbooks [@texts]. Let us for the moment assume that $CP$ is a symmetry of our theory. What does this tell us about the neutral mesons? It says that the physical propagation-eigenstates of the system, that is the particles which propagate with a distinct mass (and lifetime), must be eigenstates of $CP$. These are the combinations $(M^0 \pm {\overline{M}}^0)/ \sqrt{2}$. Particles produced by the strong interactions are produced as flavor eigenstates. This means initially one always has a coherent superposition of the two $CP$ eigenstates. Then as time goes on, because of the difference in masses of these two states, their relative phases change. Thus, if both states are long-lived enough, the flavor composition oscillates. However there is also a difference in lifetime of the two $CP$ eigenstates. If this is large then eventually the shorter-lived eigenstate decays away. Once one of the two mass eigenstates has decayed the other combination dominates, terminating the flavor oscillation and giving essentially a fixed admixture from that time on (in vacuum). For the kaon system the difference in lifetime is large compared to the difference in mass, so one does not talk about kaon oscillation, but rather about long-lived and short-lived states. Conversely for $B_d$ the mass difference is large compared to the width difference, and one can discuss either oscillating flavor states, or, discuss the same phenomena in the language of mass eigenstates, $B_{H={\rm heavy}}$ and $B_{L={\rm light}}$. For the $B_s$ both the mass and lifetime differences must be both be considered in analyzing the evolution of states. For the $D$ mesons, in contrast, the mass and width differences are both small in the Standard model. Thus both mass eigenstates decay before any significant oscillation occurs. These particles are thus typically described in terms of flavor eigenstates. Experimental searches for evidence of mixing (mass or width differences) for the $D^0$ states are another way to seek non-Standard Model physics effects, since the effect as predicted in the standard Model is small [@dmixing]. Notice that the peculiar phenomenon of oscillating particles, here and in the neutrino case as well, occurs only if you insist on describing the process in terms of flavor eigenstates. The more physical description is to use the mass eigenstates as the things you call particles (as we do for the quarks themselves). Then all that changes with time is the proportion of the two eigenstates that are present, because of their different half-lives, and the relative phase of the two states, because of their different masses. Now let us review the story of $CP$ for neutral $K$ mesons. The flavor quantum number strangeness is conserved in strong interactions. Strangeness-changing weak decays are suppressed by the Cabibbo factors tan($\theta_{\rm Cabibbo}$) compared to strangeness conserving $u <-> d$ transitions. This first fact means strange mesons are typically pair produced, the second that they are relatively long lived. The assumption of $CP$-conservation in neutral Kaon decays “explains” the observation of the two very different half-lives for neutral kaons. If $CP$ were exact, then only the $CP$-even state, $K_{\rm even} =(K^0 +{\overline{K}}^0)/\sqrt{2}$, can decay to two pions, since a spin zero neutral state of two pions can only be $CP$-even. (By Bose statistics, it can have no I=1 part.) Three-pion final states can be either $CP$-even or $CP$-odd. But the phase space for the three pion decay of a neutral kaon is quite small compared to that for two pions. This predicts two very different half-lives for the two $CP$-eigenstates. They are different, in fact, by more than a factor of ten. This successful picture was challenged in 1964 by the discovery by Christensen, Cronin, Fitch and Turlay [@ccft], that the long-lived (and hence putatively $CP$-odd) kaon state did indeed sometimes decay into the $CP$-even two pion state. This result immediately shows that $CP$-invariance is violated. Comparison of the rates for charged and neutral pions further showed that the violation is principally in the fact that the mass eigenstate does not have a unique $CP$. This result was initially very puzzling. Until then almost any field theory that had been considered as a realistic physical theory had automatic $CP$ conservation once the other desired symmetries of were imposed. Now, however, we know that the three generation Standard Model in its most general form includes one $CP$-violating parameter in the matrix of weak couplings, which is called the CKM matrix (for Cabibbo, Kobayashi and Maskawa). Thus $CP$ violation [*per se*]{} is no longer a puzzle, but rather a natural part of the Standard Model. What we do not yet know is whether the Standard Model correctly describes the $CP$-violation found in nature. Exploration of that question is a major goal of the B-physics program. Any theory for physics beyond the Standard Model will have, in general, possible additional $CP$-violating parameters. Any further fields, such as any additional Higgs fields, can introduce further $CP$-violating couplings. Such effects may then enter into $B$ decay physics. For example, in many models additional Higgs particles lead to additional contributions to $B^0$-${\overline{B}}^0$ mixing. This in turn gives possible deviations from the patterns predicted by the Standard Model for $CP$-violation in $B$ decays. One of the motivations to search for such effects is that it is not possible to fit the observed matter-antimatter imbalance (or rather the consequent matter to radiation balance) of the Universe with the $CP$-violation in the quark mixing matrix as the only such effect [@baryofailure]. (This failure suggests that there must be additional sources of $CP$-violation beyond those in the quark coupling matrix of the Standard Model, but does not require that any such effects will be apparent in $B$ decays.) Even with no other new particles, an extension of the Standard Model to include neutrino masses now appears to be needed. Then the weak couplings of the neutrino mass eigenstates are given by a CKM-like matrix. This introduces the possibility of further $CP$-violating parameters. Indeed if the neutrinos have Majorana type masses there are more $CP$-violating parameters in this matrix than in the quark case [@numass]. These parameters will be very difficult to determine and they play essentially no role in $B$ physics. However they may have played an important role in the early universe, giving the matter-antimatter imbalance via leptogensis [@leptogenesis]. I will not discuss neurtrino masses further in these lectures. As I will discuss tomorrow [@dirac], once there is any $CP$ violation in the Standard Model theory it becomes a problem to understand how it happens that $CP$ is conserved in the strong interaction sector of the theory. Experiment tells us this is so to very high accuracy, chiefly via the upper limit on the electric dipole moment of the neutron. This result tells us that, far as the $CP$-violating effects that we want to explore in $B$ decays go, we can ignore strong $CP$ violation. So apart from tomorrow’s Dirac lecture, I will not discuss it further in this series of talks. General Formalism for Neutral Mesons with $CP$ Violation -------------------------------------------------------- Once we know that $CP$ is not a symmetry of our theory we must allow a more general form for the two mass eigenstates of neutral but flavored mesons. In the following I use the convention that these two states are defined to be $M_H$ and $M_L$ where the $H$ and $L$ stand for heavy and light, which really means heavier and less heavy, since the mass difference may indeed be quite tiny. I define the two eigenstates to be $$M_{H } = p M^0 + q {\overline{M}}^0 \qquad M_{L} = p M^0 - q {\overline{M}}^0 , \label{masseigen}$$ where $|p|^2 + |q|^2 = 1$. Note that this equation is again convention dependent, I have not specified a sign or phase for $q$, but I have defined the more massive state to be the one with a plus sign before $q$. In combination with my convention that $CP M^0 = {\overline{M}}^0$ this makes the phase of $q$ a meaningful quantity. (Be aware however that, once again, other conventions are also used in the literature.) The quantity q/p is determined from the mass and mixing matrix for the two-meson system, ${\cal M} = M +i \Gamma$. This matrix is written in the basis of the two flavor eigenstates. Note that both M and $\Gamma$ are complex 2$\times$2 matrices, $M$ is hermitian and $\Gamma$ is anti-hermitian. The off-diagonal (or mixing) elements are calculated from Feynman Diagrams that can convert one flavor eigenstate to the other. In the Standard Model these are dominated by the one loop box diagrams, shown in Fig. \[fig:1\]. Actual calculation of such quantities will be discussed in later lectures, for now we simply note that they exist. Then $$q/p = \frac{\Delta M -i/2\Delta\Gamma}{2(M_{12} -i/2 \Gamma_{12})} = \frac{M_{12} -i/2 \Gamma_{12}}{2(\Delta M -i/2\Delta\Gamma)}\ . \label{qoverp}$$ Notice that the two mass eigenstates of this mixed system do not have to be orthogonal, in fact in general they will not be so, unless $|q/p|=1$. The Three Types of $CP$ Violation --------------------------------- In the above discussion we have already mentioned two possible ways that $CP$ violation can occur. The first was $CP$ violation in the decay, or direct $CP$ violation, which requires that two $CP$-conjugate processes to have differing absolute values for their amplitudes. A second possibility, seen for example in $K$ decays, occurs if $|q/p| \neq 1$. It is very clear in this case that no choice of phase conventions can make the two mass eigenstates be $CP$ eigenstates. This is generally called $CP$-violation in the mixing. As we will see later, in decays of the neutral mesons to a $CP$-eigenstate $f$, there is a third possibility. This can occur even when both the ratio of amplitudes and the quantity $q/p$ have absolute value 1. The $CP$ violation effects in such decays will be shown to depend only on the deviations from unity of the parameter $\lambda_f =(q/p) A({\overline{B}}^0 \rightarrow f)/A(B^0 \rightarrow f)$ . The third option is $CP$ violation in the interference between decays to $f$ with and without mixing. This effect is proportional to the imaginary part of $\lambda_f$ and thus can be non-zero even when the absolute value satisfies $|\lambda_f|=1$. Decays where this latter condition is true are particularly interesting. In such cases one can interpret any observed asymmetry as a direct measurement of some difference of phases of CKM matrix elements, with no theoretical uncertainties. We will see this in more detail in the next lecture. Lecture 2: Standard Model Predictions for $CP$ Violations in $B$ Decays ======================================================================= CKM Unitarity ------------- The CKM matrix of quark weak couplings has been discussed in some detail in previous lecture series in this school. It can be written, in the Wolfenstein parameterization [@wolfensteinparam], as $$\begin{aligned} V &=&\pmatrix{V_{ud}&V_{us}&V_{ub}\cr V_{cd}&V_{cs}&V_{cb}\cr V_{td}&V_{ts}&V_{tb} }\nonumber \\[1ex] &\simeq& \pmatrix{1-\lambda^2/2&\lambda &A\lambda^3(\rho-i\eta)\cr -\lambda &1-\lambda^2/2&A\lambda^2\cr A\lambda^3(1-\rho-i\eta)&-A\lambda^2&1} + O(\lambda^4)~. \label{ckm}\end{aligned}$$ In the previous lecture I talked about the ability to remove, or move, a complex phase of a coupling by redefining the phase of any field involved. This parameterization corresponds to a particular choice of phase convention which eliminates as many phases as possible and puts the one remaining, possibly large, complex phase in the matrix elements $V_{ub}$ and $V_{td}$. In this convention the upper right off-diagonal elements define the parameters. The parameterization is a convenient way to make the unitarity of the matrix explicit, up to higher order corrections in powers of $\lambda \equiv V_{us}$. (The higher order terms may also have phases, as required by the unitarity relationships, but bring in no new independent phase parameters.) The quantity $\lambda$ is essentially the sine of the Cabibbo angle. It is a small number, of order 0.2. Wolfenstein’s parameterization uses powers of $\lambda$ is a convenient way to keep track of the relative sizes of the terms in the matrix. The other independent magnitude parameters $A$ and $\rho^2 +\eta^2$ are known to be roughly of order unity. There is no theory behind which powers of $\lambda$ enter each term. The Wolfenstein parameterization simply summarizes the observations in a neat way. The fact that $V_{cb}$ and $V_{ub}$ are both small (of order $\lambda^2$ and $\lambda^3$ respectively in Wolfenstein’s parameterization) is responsible for the relatively long lifetimes of $B$-mesons (and $b$-containing baryons too). This is a fortunate property; it is essential to the feasibility of most $B$-physics experiments because it allows us to identify $B$ decays by the spatial separation of the decay vertex from the production point. It is an observational fact, not a theoretical prediction. Independent of the parameterization used, in the three generation Standard Model the CKM matrix must be unitary. This leads to a number of relationships among its elements of the form \[(row)\*x(column)\]=0. Examples are $$\begin{aligned} V_{ud}V^*_{us} +V_{cd}V^*_{cs}+ V_{td}V^*_{ts} &=&0 \hspace{.5in} {a}\nonumber\\ V_{us}V^*_{ub} +V_{cs}V^*_{cb}+ V_{ts}V^*_{tb} &=&0\hspace{.5in} {b}\\ V_{ub}V^*_{ud} +V_{cb}V^*_{cd}+ V_{tb}V^*_{td} &=&0\hspace{.5in} {c} \ .\nonumber \label{unitarity}\end{aligned}$$ In the Wolfenstein parameterization the relationship that arises from unitarity can be used to express the diagonal and lower left hand elements of the matrix in terms of the upper right elements, to any desired order in $\lambda$. The form given above drops terms of order $\lambda^4$ and above. It is a trivial fact that any relationship of the form of a sum of three complex numbers equal to zero can be drawn as a closed triangle in the complex plane. Hence these, and the other similar relationships, are referred to as the Unitarity Triangle relationships. The fact that there is only one independent $CP$-violating quantity in the CKM matrix can be expressed in phase-convention-invariant form by defining the quantity $J$, called the Jarlskog invariant for Cecilia Jarlskog who first pointed out this form [@jarlskog], $${\rm Im} V_{ij}V_{kl}V_{il}^*V_{kj}^* = J \Sigma_{m,n =1}^3 \epsilon_{ikm}\epsilon_{jln} \label{Jarlskog}$$ where $i,j,k,l$ run over the values $1,2,3$ and $\epsilon_{ijk} $ takes the value +1 if the three indices are all different and in cyclic order, and -1 if they are all different and in anti-cyclic order, but is zero if any two are the same. All the unitarity triangles have the same area, $J/2$. This area shrinks to zero if the $CP$-violating phase differences in the matrix vanish. Notice however that, while the triangles have the same area, the three examples given above are triangles of very different shapes. Triangle ${a}$ has two sides of order $\lambda$ and one of order $\lambda^5$. It would be very difficult to measure the area using such a triangle. Triangle ${b}$ is a little better, but still a has one small angle, its larger sides are of order $\lambda^2$ while its small side is of order $\lambda^4$ giving an angle of order $\lambda^2$. Finally triangle ${c}$ is the most interesting, because it has all three sides of order $\lambda^3$ so all three angles are [*a priori*]{} of comparable and large magnitude. The price one pays is that all the sides are small, but this is not as serious as the problem of measuring an asymmetry proportional to a very small angle. This triangle is the one most often discussed in relation to $B$-meson decays. Since these angles are large one expects some channels in both $B_d$ and $B_s$ decays with order 1 $CP$-violating asymmetries . Fixing the Parameters --------------------- The triangle is conventionally drawn by dividing all sides by $V_{cb}V_{cd}^*$, which gives a triangle with base of unit length whose apex is the point ($\rho,\eta$) in the complex plane. Prior to considering the asymmetry measurements we can try to determine the shape of this triangle from measurements of $CP$-conserving quantities which fix the sides, plus the measured $CP$ violation in $K$-decays. Notice that this information is already sufficient (in principle) to over constrain the set of parameters. The quantity $V_{cb}$ is determined from $B$ decays to charmed final states, $V_{ub}$ from final states with no charm, while measurements of the $B_d$ and $B_s$ mass differences constrain $V_{td}$. The $CP$ violation in $K \rightarrow \pi \pi$ gives an allowed band for the apex of the triangle. In each case there is both an experimental uncertainty in the measurement and a theoretical uncertainty in the relationship between the measured quantity and the theoretical parameter(s). The theoretical uncertainties dominate. They are typically not statistical in nature, but rather have to do with the part of the calculation which involves models or approximations needed to allow for strong interaction physics effects. There is a large literature by now on the topic of how best to combine the various measurement and deal with both statistical and theoretical uncertainties [@traingleconstraints]. New measurements from Belle and BaBar on a $CP$ asymmetry in $B$-decays constraining the angle at the lower left of the triangle have recently been announced [@newtwobeta]. This is one measurement where the theoretical uncertainties are very small, so the constraint will improve as the statistics of the measurement improve for some time to come. So far all the various results give a consistent picture; the Standard Model fits the data. This means that, within the ranges of the various theoretical uncertainties, there is a region of possible choices for the Lagrangian parameters that are consistent with all data. One hope of many physicists involved in the large effort in $B$ physics is that at some point some measurements will give discrepant answers for some Standard Model parameters or predictions. This would be evidence for physics beyond the Standard Model, and cause for much excitement in the physics community. If results for some set of measurements should begin to look discrepant, then the question of the statistical significance of the discrepancy will be much debated, as different treatments of theoretical uncertainties will give different conclusions on this point. Let us examine one of these quantities in a little more detail to see how the theoretical uncertainties arise. In each case there is a mix of weak interaction and short-distance strong-interaction physics, which both are perturbatively calculable and long range strong-interaction physics which is not perturbatively calculable. Tomorrow’s lecture will introduce some of the methods that are used to deal with (or avoid) possible long-range strong interaction effects. Here I simply want to show how such effects can enter. Consider the question of the mass difference between the two mass eigenstates for $B_d$. The two one-loop diagrams given in Fig. \[fig:1\] are the dominant contribution to this effect. Each loop-diagram can have either a $t$-, $c$-, or $u$-quark for each of the two internal quark lines. Calculation of the matrix element of these diagrams between a $B^0$ and a ${\overline{B}}^0$ meson would give $M_{12} +i\Gamma_{12}/2$. The diagrams can be written as a local four-quark operator multiplied by a calculable coefficient which includes CKM factors. I will write the quark-propagator and coupling dependent part of this coefficient schematically as $$Q =|V_{td}V_{tb}^*D_t + V_{tcd}V_{cb}^*D_c +V_{ud}V_{ub}^*D_u|^2$$ where the $D_q$ factors are the quark propagators. This expression is schematic because in writing it as a perfect square I ignored the differences in the momenta of the two quark lines in the diagram (which are typically small, ${\cal{O}}(m_b/m_W)$, compared to the loop momentum itself). Notice that if all the quarks had equal mass then $D_t=D_c=D_u$ and the unitarity condition Eq. (\[unitarity\]c) would say that this factor $Q$ vanishes. Indeed we can use this condition to rewrite the expression as $$Q = |V_{td}V_{tb}^*(D_t-D_u) + V_{cd}V_{cb}^*(D_c-D_u)|^2. \label{Q2}$$ Because of the two $W$-propagators the loop integral is dominated by momenta of order $M_W$, which is large compared to either the $c$ or $u$ quark masses. Thus the two quark propagators in the second term of Eq. (\[Q2\]) above essentially cancel one-another, so the term is suppressed by a factor of order ${(M_c^2 -M-u^2)}/{m_W^2}$. Thus the mass difference is effectively proportional to the square of the coefficient of the remaining term, which $|V_{td}|^2$ (since $V_{tb}$ is 1 up to order $|\lambda|^4$). (Note that this argument also shows why the mixing matrix is small in the $D$-meson case. There the three propagators are the down-type quarks, all three of which have masses that are small compared to $M_W$, so the Unitarity cancellations suppress the entire effect. Furthermore the contribution of the most-massive quark in this case, the $b$-quark, is Cabibbo-suppressed, further reducing the effect. ) To find the value of this $V_{td}$ by measuring the $B$ meson mass differences we need to know the matrix element of the four quark operator between the $B^0$ and ${\overline{B}}^0$ meson states. This is where the long-distance hadronic physics sneaks into the problem, this matrix element depends on the form of the $B$ wavefunction, including all effects of soft gluons. The best available method to determine it is to use lattice QCD calculation [@latticevtd]. A measurement of the mass difference of the two $B_d$ mass eigenstates thus gives a measurement of $V_{td}$ with a theoretical uncertainty that is dominated by the theoretical uncertainty in the lattice determination of the relevant four-quark matrix element. The result is usually written as some “known” factors times $B_B f_B^2$. (The “known” factors include quark masses, which are actually not so well-known and must be carefully defined.) Here the factor $f_b^2$ is the vacuum to one meson matrix element of the axial current which arises in the naive approximation to the matrix element obtained by splitting the four-quark operator into two-quark terms and inserting the vacuum state between them. This is known as the vacuum-insertion approximation. The quantity $B_B$ is simply the correction factor between that approximate answer and the true answer. It can be estimated in various model calculations. The lattice calculation does not need to make this subdivision, it directly calculates the full matrix element. However the result is often quoted in terms of the $B_B$ and $f_B$ parameters. Lattice methods can also directly calculate the latter. Eventually $f_b$ will be measured and that will provide a separate test of the lattice calculation. Once there is a good measurement of the $B_s$ mass difference the ratio $\Delta m_b /\Delta m_s$ will provide a better determination of $V_{td}$ via the ratio $V_{td}/V_{ts}$. This mass ratio is relatively free of theoretical uncertainties, as most of these cancel in the ratio of matrix elements. The matrix elements for the $B_d$ and the $B_s$ mesons are similar. Only a small correction due to the difference of the $s$ and $d$ quark masses remains. The uncertainty in this correction gives a relatively small theoretical uncertainty in $V_{td}$. At present only a lower limit for the $B_s$ mass difference is known; even this gives an important constraint (upper limit) on the range of $V_{td}$. Time Evolution of the $B$ States and Time-Dependent Measurements ---------------------------------------------------------------- Now I turn to the topic of decays of neutral $B$ mesons. What can we measure and what does it tell us? To discuss this we need to understand the time evolution of state which at time $t=0$ is known to be a pure $B^0$ meson. This means that at t=0 we have $$B(t=0) = (B_H + B_L)/2p \ .$$ Since the two mass states evolve with different time-dependent exponential prefactors we find $$B(t) =g_+(t)B^0 +(q/p)g_-(t){\overline{B}}^0$$ where the functions $g_\pm$ are just the sums and differences of the exponential mass and lifetime factors $$\begin{aligned} g_\pm &=& [e^{(-iM_Ht-\Gamma_Ht/2)} \pm e^{(-im_Lt -\Gamma_Lt/2)}]/2\nonumber \\ &=& e^{-iMt-\Gamma t/2}[e^{(-i\Delta M -\Delta \Gamma/2)/2} \pm e^{(i\Delta M +\Delta \Gamma/2)/2}]/2 \ . \label{gplusminus}\end{aligned}$$ Here we introduce the notation $M$ and $\Gamma $ for the average mass and width and $\Delta M$ and $\Delta \Gamma$ for the differences between the two sets of eigenvalues. In the case of $B_d$ the width difference is small compared to the mass difference (and to the width itself) so to a good approximation we can neglect $\Delta \Gamma$. Then the expressions for the $g_\pm$ simplify in an obvious way. For $B_s$ it is likely that the width difference is comparable to the mass difference and the full expressions must be used. The time-dependent state that is a pure ${\overline{B}}^0$ at $t=0$ can likewise be written in terms of these same functions $${\overline{B}}(t) = (p/q) g_-(t)B^0 +g_+(t){\overline{B}}^0.$$ It is now straightforward to derive the time-dependent rate to reach a particular $CP$ eigenstate final state $f$ with $CP$ quantum number $\eta_f$. It is given by $$|A(B(t)\rightarrow f|^2 = |A(B^0\rightarrow f)|^2[|g_+(t)|^2 + |\lambda_f g_-(t)|^2 + 2Re[g^*_+(t)g_-(t)\lambda_f]]$$ where the quantity $$\lambda_f = (q/p) {A({\overline{B}} \rightarrow f)\over A(B\rightarrow f)} =\eta_f (q/p) {A({\overline{B}} \rightarrow {\overline{f}})\over A(B\rightarrow f)}.$$ In the second equality here we have used the fact that f is a $CP$ eigenstate, $ CP f ={\overline{f}} = \eta_f f$ where $\eta_f = \pm 1$, to write the ratio of amplitudes in a form that shows explicitly that one amplitude is simply the $CP$ conjugate of the other. The $CP$-violating asymmetry between the rates is defined to be $$a(t) ={|A({\overline{B}}(t)\rightarrow {\overline{f}})|^2 -|A( B(t)\rightarrow f)|^2\over |A({\overline{B}}(t)\rightarrow {\overline{f}})|^2+|A( B(t)\rightarrow f)|^2}\ . \label{asymdef}$$ (Note once again you must beware of conventions, some of the literature defines the asymmetry with the opposite sign.) If $\Delta \Gamma/\Gamma$ can be neglected, which is a very good approximation for $B_d$ decays, then $|q/p|=1$ and the asymmetry takes the form $$a(t) = -[(1-|\lambda_f|^2)\cos(\Delta M t)+ 2Im\lambda_f \sin(\Delta Mt)]/(1+|\Lambda_f|^2) \ . \label{asym}$$ As promised previously, this relationship shows that the $CP$-violating effects measure properties of $\lambda_f$, in particular its magnitude and imaginary part. (In the more general case the expressions are somewhat more complicated and depend also on the width difference.) In particular, if only the third type of $CP$ violation is present, namely if in addition to $|q/p|=1$ we have $|{\overline{A}}/A|=1 $ so that $|\lambda_f|=1$, then this expression simplifies to $$a(t) = -Im \lambda_f \sin(\Delta Mt)] \ .$$ The argument of $\lambda$ depends simply on weak phases, so that $$Im\lambda_f = \eta_f \sin(2\phi_{\rm mixing} -2\phi_{\rm decay})\ . \label{imlambda}$$ Here $2\phi_{\rm mixing}$ is the phase of $q/p$ and $2\phi_{\rm decay}$ is the phase of $A({\overline{B}} \rightarrow {\overline{f}})/A(B \rightarrow f)$ while $\eta_f$ is the $CP$ quantum number of the state $f$. These phases are each given by some combination of $CKM$ matrix-element phases. While each of them separately can be changed by changes in phase convention (rephasing of quark fields) the difference is convention independent, as must be so for any physically measurable quantity. Thus the asymmetry directly measures the phase differences between particular CKM matrix elements with no uncertainties introduced by our inability to calculate strong interaction physics effects such as the magnitude or strong phase of an amplitude. These strong interaction effects all cancel exactly when $|\lambda_f|$ is 1. CP Eigenstate Channels for $b\rightarrow c {\overline{c}} s$ ------------------------------------------------------------ There are many possible channels to investigate. The interest lies not just in one measurement but in whether the pattern of $CP$-violating asymmetries fits the predictions of the Standard Model. What channels should we study? We need a final state of definite $CP$. In general for a multibody final state even when the particle content is $CP$-self conjugate there will be an admixture of $CP$-even and $CP$-odd contributions because of different possible orbital angular momenta among the particles. The simplest way to get a definite $CP$ final state is to require that the $B$ decay to a two-body or quasi-two body final state with only one allowed orbital angular momentum. (Quasi-two-body here simply means a two-body state with one or two unstable particles, such as a $\rho \pi$ or $\rho \rho$. The actual observed final state is then three or four pions.) Given that the $B$ has spin zero, the final state has a unique orbital angular momentum between the pair of particles if (and only if) at least one of the two particles has spin zero. For quasi-two body states where both particles have non-zero spin but at least one of them is unstable one can possibly separate out the $CP$-even and $CP$-odd final state contributions using an angular analysis of the distribution of secondary decay products [@anganal]. The price is that, in general, a larger data sample is needed to achieve the same accuracy on the $CP$ asymmetry measurement. Note that the Feynman diagram structure is the same for all channels with the same quark content. Results from multiple channels can sometimes be combined to improve statistical accuracy. For example for the quark decay $b\rightarrow c{\overline{c}} s$ the $B^0$ decay channels $J/\psi K_S, \psi^\prime K_S,\eta_c K_S$ $J/\psi K_L, \psi^\prime K_L,\xi_c K_L$ ([*etc.*]{}) all depend on the same set of quark diagrams. For the $b\rightarrow u{\overline{u}} d$ (and $d{\overline{d}} d$)quark content there are likewise many channels: $\pi\pi,\rho\pi, \rho\rho$, [*etc.*]{} (The last of these needs angular analysis.) Let us then examine what the predicted $CP$ asymmetry is in each of these two cases. We begin with the modes such as $B\rightarrow J/\psi K_s$. These have been called the golden modes for analyzing $CP$ violation in $B$ decay. For once we have a situation where the mode for which the theoretical analysis is straightforward is also one with good experimental accessibility. One still needs a large sample of $B$ decays because the branching fraction to these channels is not large. (In $B$ decays there are so many open channels that branching fractions are small and smaller: the “large” modes occur at the few percent level; $J/\psi K_S$ and similar modes are about a tenth of a percent; a “rare” mode in this game has a branching fraction a few times $10^{-5}$.) First we need a little terminology. We use the term spectator quark for the quark other than the $b$-type quark (or antiquark) that is present in the initial $B$ meson, since it is generally not involved in the $b$-decay diagram. There are two topologies of weak decay Feynman diagram that can contribute to $B$ decays to leading order in the weak interactions. These are called “tree” and “penguin” diagrams and are shown in Fig. \[fig:2.1\]. A tree diagram is one where the $W$-boson creates or connects to a different quark line from the line that starts out as the $b$-quark. I thus also include any annihilation diagram or any diagram where the $W$-boson connects to the spectator quark as part of what I call the tree amplitude. Whenever such a diagram is allowed it will enter with the same CKM factors as the other tree diagram processes. A penguin diagram is a loop-diagram where the $W$ reconnects to the quark line from which it was emitted. Then a hard gluon is emitted from the quark line in the loop, and either makes a pair or is absorbed by the spectator quark. When higher order strong interaction rescattering effects are included the distinction between tree and penguin diagrams becomes blurred. However, it is useful (and standard) to start out by describing processes in this language as it allows us to identify all the relevant CKM factors, and the operators which they multiply. As we will shortly see, that is the essence of the story. Eventually we will group terms not by the diagrams, but by the CKM factors. That grouping is not blurred by any subsequent strong interactions. The language tree and penguin persists, but the “tree contribution", in my terminology will be taken to include not only the tree diagrams (including those that involve the spectator in the weak vertex), but also that part of the contribution from the penguin diagrams that has the same CKM factor as the tree diagrams. Obviously, if one wants to try to calculate the size of the contribution to the amplitude one must keep track of each diagram separately, but if we are only concerned with whether there is more than one CKM structure in the significant contributions we can lump together all the terms with a given CKM factor. The cleanest cases theoretically are those where we can make a prediction without knowing anything about the sizes of the amplitudes because we are looking at a ratio of rates where these cancel to a good approximation. The $CP$-violating asymmetry in channels arising from quark transition $b\rightarrow c{\overline{c}} s$ in a $B_d$ meson is just this type. The tree diagram has a CKM factor $V_{cb}^*V_{cs}$. Any time that penguin diagrams contribute to an amplitude there are three terms, corresponding to the three different up-type quarks that inside the loop. Thus we can write the $b$ to $s$ penguin amplitude $P$ in the form $$\begin{aligned} P&=&V_{tb}^*V_{ts} f(m_t)+V_{cb}^*V_{cs} f(m_c)+V_{ub}^*V_{us} f(m_u)\nonumber \\ &=&V_{cb}^*V_{cs}[ f(m_c)-f(m_t)]+V_{ub}^*V_{us} [f(m_u)-f(m_t)] \label{cspenguin}\end{aligned}$$ where the $f(m_q)$ is some function of the quark mass. In the second expression I have once again used the Unitarity relationship Eq. (\[unitarity\]c) to rewrite the three terms in $P$ in terms of two independent CKM factors. Notice that the first of these is the same as that for the tree term, so for this discussion we call that contribution part of the “tree amplitude”. The remaining term is CKM suppressed by an additional factor of $\lambda^2$. The two differences of quark-mass-dependent factors are expected to be comparable in magnitude. Furthermore, ignoring CKM factors, the penguin graph contribution is expected to be suppressed by about 0.3 compared to the tree graph, because it is a loop graph and has an additional hard gluon. This means the suppressed second term in Eq. (\[cspenguin\]) is negligible (a few percent) compared to the “tree amplitude” which here is the sum of the tree term and the dominant penguin term. Thus we have an amplitude that effectively has only a single CKM coefficient and hence one overall weak phase. This then ensures $|{\overline{A}}/A|=1$, which means there is no decay-type (direct) $CP$ violation. (You will recall we needed two terms with different weak phases to get such an effect. ) Remember too that for $B_d$ we expect $|q/p|=1$ to a good approximation. Thus we have a case where $|\lambda_f|=1$ and the measured asymmetry arises purely from the interference of decay before and after mixing. We find $$a_{J/\psi K_S} = -Im (\lambda_{J/\psi K_S}) \sin (\Delta M t )= \sin(2\beta)\sin (\Delta M t ) \ .$$ Here the quantity $\beta$ is the lower left-hand angle in the standard $B$ physics Unitarity triangle (also sometimes called $\phi_1$). (The minus sign disappears because $\eta_f=-1$ for $f=J/\psi K_S$.) Thus this asymmetry directly measure the phase of a rephasing-invariant combination of CKM elements. Furthermore all the channels in the $c{\overline{c}} s$ list above measure the same asymmetry, up to an overall sign, the $\eta_f$ factor of the channel in question. For example $K_S$ and $K_L$ are states of opposite $CP$, as are the $\psi$ and $\eta_c$. Care must be taken to include the correct $\eta_f$ factor for each state in combining the results. One can also include a state such as $J/\psi K^*$ provided the $K^*$ decays to a flavor-blind combination such as $K_S\pi^0$, and angular analysis is used to separate $CP$-even and $CP$-odd contributions. One can apply this same diagrammatic analysis to the decays $b\rightarrow c{\overline{c}}s$ in a $B_s$ meson. This gives a prediction for channels such as $J/\psi\phi$ that the $CP$ asymmetry is zero in the Standard Model, as the $B_s$ mixing term is dominated by CKM factors with the same weak phase as this decay. Thus, in the Standard Model, only the CKM suppressed penguin terms which we neglected above can give $CP$ violating asymmetries here, so at most a few percent asymmetry is expected. Such predictions of small or vanishing asymmetries give another way to examine the patterns of the Standard Model. Any theory of new physics effects which give additional mixing contributions could destroy the cancellation of mixing phase and decay phase which makes this asymmetry small in the Standard Model. However to interpret such a result one indeed needs some calculation of decay amplitudes, in order to quantify more precisely how big the “few percent” Standard Model asymmetry could be. The trick of rewriting the sum of three penguin terms as two terms using the Unitarity relationships is a generally useful tool. In any channel one then has at most two CKM factors to consider. The next step is to get a rough estimate of the relative size of the two terms. This becomes important when $|{\overline{A}}/A|\neq 1$. Some further $B$ Physics Jargon ------------------------------- The $B$ physics jargon distinguishes contributions by three attributes, because these three things give a first estimate of how big the contribution is. The first size factor is whether the diagram is tree or penguin. The penguin is suppressed relative to the tree because it is a loop diagram and because it involves a factor of $\alpha_{\rm strong}$ at a scale of order $m_b$ due to the hard gluon, together this makes for a suppression factor of order about 0.3, all else being equal. The next size factor is the powers of the Wolfenstein parameter $\lambda$ in the associated CKM factors. All $B$-decay amplitudes have at least two powers of $\lambda$. Amplitudes with higher powers are called CKM-suppressed. The third size factor is the color flow pattern that forms the particular final state of interest. Diagrams where a quark-antiquark pair produced by a W finish up in the same meson are called color-allowed, because this pair is produced in the requisite color-singlet combination. In terms of color-flow diagrams there are two independent color-flow loops as shown in Fig. \[fig:2.2\](a). When the quark and antiquark produced by the $W$ end up in different final mesons the diagram is called color-suppressed (Fig. \[fig:2.2\](b)). There is then only a single color-flow loop so that diagram is expected to be of the order of $1/N_c$ smaller than the corresponding color-allowed diagram. For penguin diagrams color suppression, if it works at all, works the other way around. Diagrams where the quark and antiquark from the gluon end up in two different mesons, Fig. \[fig:2.3\](a), are color allowed, and indeed can be seen to have two-color-flow loops just as do the tree color-allowed contributions. Diagrams where the flavor-structure says the quark and antiquark produced by the hard gluon must be in the same meson are called color suppressed. In Fig. \[fig:2.3\](b) there is only one color loop. However in this diagram the gluon makes a color singlet object. But a gluon is a color-octet state. Taken literally, the diagram vanishes. A second gluon must be exchanged here. If we were to count the extra gluon as a hard gluon, there would be an additional suppression factor of $\alpha_{\rm strong}$, but no $1/N_C$, because we would again see two color loops, Fig. \[fig:2.3\](c). However the second gluon is not necessarily hard, so the relevant scale for the $\alpha_{\rm strong}$ is not large. In some estimates these contributions are treated as $1/N_C$ suppressed terms, but there is no good argument that justifies this counting. As you can see from these arguments, the naive color-counting is not a very reliable measure of the relative strengths of the two types of penguin contributions. QCD-improved operator-product expansion calculations at leading order in $\Lambda/m_b$ [@qcda; @qcdb; @qcdc] can be made. These treat the color factors correctly. We will return to this approach at later, in Lecture 3. However there is a large literature of estimates that use the language of color-allowed and color-suppressed contributions, so it is important to know how these terms arose and how they are used. All these size-counting factors are generally used to give first estimates of the order of magnitude of the various contributions. Clearly a more serious calculation can significantly change the relative sizes. The kinematics of the different diagrams are different. The matrix elements of the various operators are different. Indeed there is an interplay between the wave function of the mesons and the counting factors discussed above which in the end determines the size of an amplitude. Powers of $\Lambda_{QCD}/m_b$ can arise from the wavefunction for particular kinematic configurations relative to others. Higher-order hard QCD effects can be systematically included, but the soft hadronization part of the calculation needs some additional input, either from a model or from some other measurement. Another Sample Channel ---------------------- Now let us look at one more set of channels to see what happens when this size counting says two CKM factors can occur with comparable coefficients. The case I choose to examine is the decay $B_d \rightarrow \pi^+\pi^-$. At the quark level this process is governed by decays $b \rightarrow u{\overline{u}} d$. You can readily find from the diagrams of Fig. \[fig:2.1\] that there are both tree and penguin contributions for this quark content. The tree diagrams have a CKM factor $V_{ub}^*V_{ud}$. For the penguin contributions we can again use unitarity to rewrite the three different intermediate quark contributions as a sum of two terms. In this case all three CKM coefficients are of the same magnitude. I choose to eliminate $V_{cb}^*V_{cd}$ because then the second penguin term (the one that does not have the same weak phase as the tree term) has the same weak phase as the mixing term in the Standard Model. Then only one difference of CKM phases will enter my eventual formulae for the asymmetry. However we cannot ignore the second penguin term. The only thing that makes it small compared to the “tree amplitude” (which includes the first penguin term as well as the contribution from the tree diagram) is the fact it is a penguin loop. That is not sufficient to completely discard it. So here we have a situation where there can be $|{\overline{A}}/A|\neq 1$ effects. We must use Eq. (\[asym\]) to interpret the the measured asymmetry. One would like to extract from the measurement the CKM phase difference between mixing and tree decay contribution (which in this case is $\alpha \equiv \pi -\beta - \gamma$). One can measure two quantities, $|\lambda_f|$ from the coefficient of cos($\Delta M t$), and Im$\lambda_f$ from the coefficient of sin($\Delta M t$). However three unknown quantities enter in the expressions for $\lambda_f$ in such a case. These are the relative weak phase of mixing and the tree decay amplitude $\alpha$, and both the absolute value ratio, r, and the relative strong phase, $\delta$ of the penguin and tree terms. We can write $$\lambda_f = e^{-2i\alpha}{1 +re^{i(\delta +\alpha)} \over 1+r^{i(\delta -\alpha)}} \ .$$ Here the phase $\alpha = \pi - \gamma -\beta$ is the angle at the top vertex of the standard $B$-physics unitarity triangle; it is the difference between the weak phases of the mixing and that of the tree contribution to the decay. Obviously, knowledge of both the real and imaginary parts of $\lambda_f$ is not enough to fix all three quantities. So we cannot extract a value of $\alpha$ from this asymmetry measurement alone. (Note, however that for very small r the expression simplifies so that the measurement of Im$\lambda$ determines sin$2\alpha$.) We must use further theory or measurement inputs (or both) to determine $\alpha$ if r is not small. (A note of warning here, one often sees the statement that one tests the Standard Model by testing the relationship $\alpha = \pi -\beta -\gamma$ between the angles in the triangle. The relationship is a definition. The tests of the Standard Model are tests of whether one finds the same result for the two independent angles, usually chosen to be $\beta$ and $\gamma$, using a variety of independent ways to measure them.) Note also that the ratio, $re^{i\delta}$, of the tree to the penguin amplitudes will be different for the different channels with the same quark content. The kinematics of the tree and penguin diagrams are different, and so are the wave functions for forming a $\pi$ or a $\rho$, for example. Thus, unlike the $c{\overline{c}} s$ decays, we cannot simply combine channels to improve statistical accuracy. Instead we must devise methods to remove the dependence on the additional parameters; these methods are different for each set of final state particles. For the $\pi \pi$ case there are two ways to proceed. One is to rely on isospin symmetry and isospin-related channels to give the needed additional information. The second is to develop methods to calculate these various amplitudes more reliably. This may also involve using relationships to other channels where the tree and penguin amplitudes enter with different relative strengths because of different CKM structure. For example by using measurements on $K\pi$ channels as well with those from $\pi \pi$ channels one can gain some information on the size of the penguin amplitude which dominates the decay in the former case. One can then use SU(3) symmetry to relate that to the size of the penguin in the $\pi\pi$ case. Eventually such methods can much reduce the theoretical uncertainty in the extraction of the CKM parameter $\gamma$, or equivalently $\alpha = \pi -\beta -\gamma$. Tomorrow I will discuss both of these approaches in a little more detail. The set of all possible $B$ decays can be summarized by reviewing all possible $b$-quark decays and the channels to which they can contribute. A little care must be applied to this logic, as strong rescattering can turn one quark-antiquark combination into another, one must include this possibility in a full treatment. For example in any channel involving a $\pi^0$ or $\rho^0$ meson the penguin diagrams for $b\rightarrow d{\overline{d}} d$ must be added to the diagrams for $b\rightarrow u{\overline{u}} d$. I refer you to the table in the Particle Data Book review on this topic [@pdgreview] that summarizes the quark decays and gives the CKM factors that enter for each (after using the Unitarity trick to get two terms only.) Any time you start thinking about a specific process you will find you want this information. You can rederive it readily by drawing the allowed quark diagrams and investigating their CKM factors. Lecture 3. Theorist’s Tools for $B$-physics =========================================== Today’s lecture will briefly introduce a number of theoretical tools for calculating $B$ decay processes. There are only a few examples of measurements for which we do not need to know the relative magnitude of various contributions to the decay amplitudes in order to relate the measurement to some parameters in the theory. We would like to go further and interpret the multitude of other measurements that are possible because of the many different $B$-decay channels. To do this we must devise methods to calculate or relate amplitudes. The available calculational methods all involve some mix of systematic expansion in powers of one or more small parameters, lattice calculation of matrix elements of operators, relationships based on symmetries of the strong interactions such as isospin and SU(3) flavor symmetry, and some input for transition matrix elements and or quark distribution functions. These last can be calculated reliably only in certain limits and in general require models and approximations. Alternately one can measure some of these quantities in one set of processes and use the measured values as input in the interpretation of other measurements. This lecture will give a general picture of the toolkit of approaches, what each tool is, and how it can be used. There will not be time here to teach the details of any of the methods. This lecture summarizes a large body of theoretical work. I will not attempt to reference all the relevant papers, but will include references to some current papers as examples of the type of work now underway. I apologize in advance to the many whose papers I do not mention. There are two small parameters in this game, namely $\Lambda_{QCD}/m_b$ and $\alpha_{\rm strong}(m_b)$. Here $m_b$ is the mass of the $b$-quark and $\Lambda_{QCD}$ is the scale that defines the running of the strong interaction coupling. The detailed definition of each of these quantities is fraught with technical problems, but there is a clear physical meaning for the rough size of these parameters. $\Lambda_{QCD}$ is related to the inverse size of a typical hadron while the $b$-quark mass can be characterized as roughly the same scale as the mass of a $B$ meson (up to corrections of order $\Lambda_{QCD}/m_b$). The strong coupling $\alpha_s(m_b)$ scales as a logarithm of $\Lambda_{QCD}/m_b$; we treat it as a separate small parameter because we can count powers of this parameter separately from the powers of $\Lambda_{QCD}/m_b$; they arise in different ways. The fact that $\Lambda_{QCD}/m_b$ is indeed quite small leads to a simple intuitive picture of a $B$ meson at rest. It is an essentially static $b$ quark with the light quark forming a cloud around it. The light-quark distribution is sometimes called the brown muck, because we cannot reliably calculate the details of it. However we do know that certain properties are rigorously true in the limit $m_b\rightarrow \infty$. For example in that limit the wavefunction does not depend on the spin orientation of the $b$-quark and hence is the same for a spin 0 $B$ meson and a spin 1 $B^*$. A second way in which the large mass of the $b$-quark simplifies the problem is that any gluon that carries off a significant fraction of the $b$-quark mass is a hard gluon that can be treated perturbatively; it introduces the small parameter $\alpha_{\rm strong}(m_b)$. In addition to these expansions there is another part of the picture that is true because $m_b/M_W$ is small. This means that weak decays of the $b$-quark are essentially local four-quark effects. Thus the $B$ meson decay can, to a reasonable approximation, be thought of as proceeding in two stages: a $b$-quark decays and then the remnants hadronize to give the final state under study. It is this second stage, the hadronization, that introduces all the uncertainties into the calculations. We have good methods for applying QCD to things like jet-formation for well-separated high momentum quarks, but a $B$ decay does not give us large enough quark momenta to use this formalism reliably. Further, we want to know amplitudes for specific few-body (quasi-two-body) final states (states of definite $CP$). Most likely these arise when the four quarks that are present after the $b$ decay are not well-separated (so even if the $B$ mass were much larger a jet calculation would not provide the answer). We cannot calculate these amplitudes completely from first principles. So my purpose in this lecture is to review the tools that we do have and how they can be used to minimize the theoretical uncertainty on the extraction of the desired quantities, such as CKM parameters, from experiment. Operator Product Expansion -------------------------- The operator product expansion is a way to formalize the separation of hard or short-distance physics from soft or long-distance physics. It begins by rewriting the Feynman diagrams into the form of local operators, defined at a given scale, with calculable, scale-dependent coefficients. First we look at all the tree and penguin Feynman diagrams for the weak decay of the $b$-quark. Each can be written as a sum of four quark operators with definite coefficients at the scale $M_W$. This is the leading order operator product expansion. There are actually two types of penguin diagrams, those I mentioned earlier that involve a gluon, and a second set called electroweak penguins that involve a photon or a $Z$ particle emitted from the loop. These last give an additional set of four-quark operators. At first glance one might guess that the electroweak penguin contributions are very small, with $\alpha_{QED}$ replacing the $\alpha_{\rm strong}$ of the gluon case. However it turns out there is a part of the $Z$-penguin contribution which is enhanced by a factor $M_t^2/M_W^2$ and so there are cases where these terms can be important too. Each class of diagrams corresponds to a distinct set of four quark operators at leading order. When hard QCD corrections are included, one must introduce a new scale into the problem, which is the hard-soft separation scale $\mu$ that defines which gluons are absorbed into the new scale-dependent operator coefficients and which are defined to be included in the scale-dependent matrix elements of operators. In addition, these corrections can mix the operators, and thereby blur the distinction between tree and penguin contributions. Thus the labels of each operator as being tree or penguin type is a leading order distinction only. However they are usually listed in that way as it is a useful way to keep track of which operator arises with which CKM coefficients. In addition, if a hard gluon connects the weak decay vertex to the spectator quark this can also introduce additional local operators that involve six quark fields, again with calculable coefficients that begin at order $\alpha_s(m_b)$. One must choose the $\mu$-scale that separates hard and soft physics. In principle no physics depends on this choice. In practice if one makes approximations for the matrix elements one does not usually get the correct scale-dependence in their values. So results do to some extent depend on the choice of scale. This dependence is minimized by doing higher order QCD calculations, but in general is not fully removed even with that laborious step. Each four-quark operator takes the form $${\cal O}_n = {\overline{b}} \Gamma_{n1} q^i {\overline{q}}^j\Gamma_{n1}q^k$$ where each $\Gamma_{ni}$ denote a specific combination of gamma matrices and QCD color structure and the $q^i$ denote the relevant quark flavor (and color) content. The details of the color and flavor flow in the diagram can be read off once these operators are written. I do not include here the detailed list nor any discussion of the coefficients. That is available many places [@texts]; my point here is not to discuss this well-developed technical subject, but rather to talk about the additional steps between writing down an operator and its coefficient and calculating an amplitude for any particular channel. The matrix elements of the operators between the initial $B$ state and the final set of mesons are where hadronic physics enters the game. Our methods for calculating that physics are limited. We can however use information that we do have about symmetries of the strong interactions, for example, to tell us about the ratios of matrix elements that occur in different decays. The Factorization Approximation ------------------------------- The simplest approach to the problem, for example for calculation of a color-allowed tree diagram, is to approximate the matrix element in a two-hadron decay as the product of the transition matrix element of a two-quark weak current between the $B$ meson and one final state meson (that can be measured in a semileptonic decay), times the matrix element for the $W$ to create the second meson, which is also measured elsewhere. This approach is called factorization, (or sometimes “naive factorization”) because it factorizes the four-quark hadronic operator matrix element into a product of two two-quark matrix elements. This idea can be generalized to divide any four-quark operator into two two-quark operators, which can either be extracted from experiment or estimated using models for the quark distribution functions of the mesons. The approximation neglects any effect of interactions between the two mesons in the final state, effects known as final state interactions. Now we know that two mesons (for a concrete example think of two pions) colliding at the energy corresponding to a $B$-mass certainly do interact. So at first glance you may think this approximation has no reason to be accurate. It is certainly not rigorously true, except in a few special cases. However it is motivated by a reasonable physical picture, usually attributed to Bjorken [@bjfactorization] (although in this reference he says the argument is common knowledge). The idea is that the weak decay is a very local process which converts one quark to three. Only for the kinematic configuration where two of these quarks (or rather one quark and one antiquark) go off essentially together, with the third one recoiling in the opposite direction, is there any significant probability that the system will hadronize as a two-body final state. (All other configurations are assumed to make multi-body final states, for example by fragmentation of the four final-state quarks.) In the special case that gives two-body states the quark and anti-quark that travel together start out much closer together in the transverse direction than the size of a typical hadron. They get quite far from the region containing the other quark and the “brown muck" of the spectator quark before they evolve into the hadronic-sized meson that is observed. They must start out in a color-singlet state to form such a meson. In a local color-singlet configuration (small compared to a meson) the strong interactions must cancel. So initially there are no strong interactions because the pair is in a local color-singlet configuration. Later there is no strong interaction because the two mesons are well-separated and strong interactions are a short-range phenomenon. The justification of the factorization approximation, as described above, applies for a tree diagram with no direct involvement of the other valence quark of the $B$ meson quark in the weak decay vertex. More generally one can try to factorize any four quark operator (possibly after making a Fierz rearrangement to group the relevant quark fields as flavor-flow dictates they must be grouped to form the mesons of interest). One then uses other measurements, or possibly lattice calculations, to fix the two two-quark matrix elements. In the case of a color-suppressed contribution, or one arising from a penguin diagram the flavor-flow does not automatically match two color-singlet quark pairings. However, if a color-singlet meson is to be formed then there must be a color-singlet piece of the amplitude, and for this piece the factorization argument applies. In some processes the flavor content of the final state allows a contribution either from annihilation (in the case of a charged $B$ meson) or from exchange of a $W$ between the two initial state valence quarks (for neutral $B$’s). Both processes are suppressed in the heavy quark limit by the quark-mass dependence of the wave-function at the origin (the $B$ to vacuum transition matrix element of a local two-quark current). These contributions are typically neglected in rough estimates of two-hadron decay rates. Despite all the caveats, the factorization approximation is generally used to make first guess estimates of the sizes of various partial rates. To determine the reliability of this calculation one must look more carefully at what is being done here. I mentioned previously that the operator coefficients can be calculated with hard QCD corrections taken into account. This introduces a scale dependence into their definition, the scale of the separation between hard and soft corrections in QCD. This is not a physical scale, but an arbitrarily chosen one, so the true answer cannot depend on it. Any scale-dependence in the coefficients must be compensated by cancelling scale-dependence in the matrix elements. But when we use measurement of a semi-leptonic process to determine the matrix element there is no reference to any hard-soft division scale; the measured quantity is scale independent. So we clearly have a problem, even in the best cases, factorization cannot be quite correct. The naive way to deal with this problem is to say it is reasonable to pick a scale somewhere between $m_b/2$ and $2m_b$ since the mass of the $b$-quark sets the typical momentum scale for the quarks arising from its decay. One then asks how the quantity in question varies as one changes the scale within this range and uses this variation to assign a central value and a theoretical uncertainty to the result. While this seems quite a plausible approach there is no way to be sure it is right. The problem is alleviated somewhat, though not completely removed, when higher order QCD calculations of the operator coefficients are used. It can only be dealt with correctly when a consistent treatment of higher order matrix elements is used, along with the higher order coefficients. Any finite order calculation, however, will typically have some residual scale-dependence problems. The issue of determining the theoretical uncertainty, that is the reasonable range of values of a theoretical estimate, is one to which we will return again and again in this lecture. Our ability to test the Standard Model by comparing its predictions with experiment depends on our ability to determine how big the uncertainties in our theoretical calculation are. A clean result is one where we know that these uncertainties are very small, or at least where we know very well how big they can be. But more often than not we find a part of the calculation is not so clean. The methods of determining the possible range of the predictions of the Standard Model are all too often subjective and ill-defined. Theorists continue to work to remove such ambiguities, and to find those measurements, or sets of measurements, for which they are minimal. This is an important task. Heavy Quark Limit Relationships between $B$ and $D$ Mesons ---------------------------------------------------------- One powerful technique for dealing with $B$ decays is use the fact that the $b$-quark mass is large compared to the QCD scale and to calculate quantities in terms of a power series expansion in that ratio. If one also treats the charm quark as heavy compared to the QCD scale then one has an even more powerful set of relationships. Then to leading order in $\Lambda_{QCD}/m_q$ the distribution of the light quark in a heavy-light meson is independent of the spin orientation or the mass of the heavy quark. This means it is the same for a $B$ or a $B^*$ or a $D$ or a $D^*$ meson. This is a very important statement because it gives us at least one limit in which we know the transition matrix element between a $B$ and a $D$ or $D^*$ meson. Consider for example the semi-leptonic decay $B^0\rightarrow D^*\ell\nu$. In the kinematic limit where the $D^*$ is at rest in the $B$ rest frame the wave-function overlap is 1. There is a small but calculable QCD correction to the unit wave-function overlap. Then there are the corrections to the heavy-quark limit relationships, which in this case turn out to be quadratic in $\Lambda_{QCD}/m_q$. This is reasonably small even for the charm quark. This means that we can, in principle, use a measurement of this quantity to extract the CKM matrix element $V_{cb}$ with very little theoretical uncertainty. The only problem is that the configuration where this relationship holds is, as I said, a kinematic limit. That means that the rate vanishes at that point! One must measure the rate as a function of $q^2$, and use an extrapolation to extract the quantity of interest. The extrapolation requires some knowledge about the behavior of the form factor as one goes away from the perfect-overlap situation, and that introduces some theoretical uncertainty into the answer for $V_{cb}$. However as more data is collected one can measure the rate ever closer to the end point, thereby reducing the sensitivity to the extrapolation. There are some other technical issues that appear in this problem. One interesting one that crops up here, and in other problems too, is the choice of the definition of the quark mass $m_b$ (or $m_c$). If you remember from muon decay, the semileptonic decay rate for a fermion (here the $b$-quark) goes like the fifth power of the mass of the decaying particle. Thus any uncertainty in the definition of the quark mass translates into a huge uncertainty in the predicted rate. But it is even worse than this. If you try to define the quark mass as the mass at the pole of the quark propagator this definition is scale dependent and even diverges as the scale is reduced (known as the renormalon problem). Clearly this is an unphysical effect, because you chose an unphysical definition of the quark mass. The problem is to find a definition that avoids this problem and leads to a well-controlled result. This can indeed be done. The full discussion of how one does it is beyond the scope of this lecture. I merely warn you that you can get into trouble by blithely assuming you know what someone means when they write $m_b$. This quantity cannot be directly measured. It is dependent on definition convention and on renormalization scale. As you compare results of different calculations you must always be aware of the conventions and definitions that have been used. Otherwise you will not be able to interpret and apply the results correctly. QCD-Improved Factorization -------------------------- The word picture explanation of factorization is to some extent confirmed by explicit calculation of QCD corrections up to order $\alpha_S$ and at leading order in $\Lambda/m_q$. It is found that the color-singlet nature of the meson leads to cancellation of the soft-gluon exchange between the two final-state mesons. In general, particularly for processes dominated by penguin or color-suppressed diagrams, there are found to be additional contributions which cannot be described by the simple factorization of a four-quark operator, but rather add to the picture a local six-quark operator. They arise because of a hard-gluon exchange between the so-called spectator quark (now no longer just a spectator) and another quark within the same meson. The matrix elements of this operator can be approximated as the a product of three valence-quark-distribution functions, one for each meson (one initial and two final) times the hard coefficient which begins in order $\alpha_s(m_b)$. Uncertainties arise from limitations on our knowledge of the quark distribution functions. One has to be careful here when matching the calculated hard-quark coefficient with measured transition matrix elements and form factors. The scale-dependence matching must be done correctly. One must also ensure that one is not double counting contributions of hard quarks that are effectively inside one of the measured quantities. But these are technical problems that can be dealt with correctly. This treatment is known as qcd-improved factorization [@qcda]. Here the term factorization is used for the factorization of the hard and soft physics. This form of factorization has been demonstrated to work for the leading order in $\Lambda/m_b$ and one order in $\alpha_s({m_b})$ corrections to the leading diagrams. The actual $\Lambda/M_b$ power counting is dependent on the assumptions about quark distribution functions; it assumes they vanish as a power of x at their end-point. As the calculation includes all gluon energy scales it is argued that all final state interactions are included in the formalism. The question remains as to whether this argument applies to all orders. It has been proven true to all orders in $\alpha_s$ and leading order in $\Lambda/m_q$ for the special case of a $D\pi$ final state with flavor such that the spectator quark in the $B$ ends up in the $D$ and the charm quark is treated as a heavy quark in the $\Lambda/m_q$ power counting [@dpifac]. It turns out that the numerical results depend quite sensitively on the details of input assumptions on the quark distribution functions [@qcdb; @qcdc]. A variant of the approach making quite different, and indeed additional, assumptions about the quark distribution function end-point behavior gets numerically very different results [@qcdc]. The second approach is called perturbative QCD by its proponents. It is claimed in this approach that the entire result is perturbatively calculable. While these claims are open to question [@sachrajda], one can simply regard the results of this work as the output of a set of ansaetze for the distribution functions. The results raise issues that have contributed important points to the discussion. One is the question of exactly how small some of the $(\Lambda/m_b)$-suppressed contributions are in actuality. The annihilation-graph contribution, for example, is found to be significant, even though formally suppressed. The sensitivity of results to inputs is unfortunate. It means that even these more sophisticated calculations leave us with some significant theoretical uncertainties. The best one can do to quantifying these uncertainties is to see how much the results change when one varies over some reasonable set of assumptions for the various inputs such as quark distribution functions and transition matrix elements. But how do you decide what is a reasonable range? As the existing debates show, in many cases this comes down to some subjective choices, not all rigorously decidable! (Some choices are, however, quite clearly unreasonable and should be excluded from discussion, for example a calculation that sets the scale of transverse momenta in a hadron at $k_\perp^2 =\Lambda m_b$, or a form-factor model that does not fit a rigorous theoretical limit relationship.) As data and calculations for multiple channels are obtained it is likely that we will develop a better understanding of such issues, and a more consistent view of what range of assumptions are reasonable will emerge. Meanwhile it is very important that any calculation reported should include an honest estimate of its uncertainties, and a clear explanation of the assumptions made and the ranges of input variables that were included in obtaining this estimate. Isospin ------- Another useful tool for extracting clean results for strong decay amplitudes is the symmetries of the strong interactions. The best of these, in that it most close to a true symmetry of the hadronic decays, is Isospin symmetry. I find I must explain this symmetry from scratch for current students. It is a piece of old fashioned physics knowledge which is not always taught in modern courses. Isospin is a symmetry under interchange of $u$ and $d$ quark flavors. It is called “iso”, because atoms which differ by such an interchange (originally by replacing a neutron by a proton or vice versa) are called isomers because they have nearly equal mass, and “spin” because the two quarks form an SU(2) doublet and the mathematics of SU(2) is the familiar mathematics of spin doublets. Isospin has nothing to do with any angular momentum. Notice also that I do not here mean the weak isospin (so called because it is yet another SU(2)); the isospin doublet is truly $u$ with $d$, not with some admixture of $d$,$s$, and $b$. Isopin is, quite obviously, broken by electromagnetic effects since these distinguish quark charges, and it is also broken by quark masses. Now the up and down quark mass are nowhere near the same, the ratio $({m_u-m_d})/({m_u+m_d})$ is not a small number. So why is Isospin ever a good symmetry? The answer is that in many cases, (including most but not all hadron decays) the relevant scale with which to compare the quark mass difference is not the quark mass sum but the hadron mass scale. That scale is set either by $\Lambda_{QCD}$ or by some heavy quark mass. Then the corrections to isopin-based predictions are small. One must be careful, however, to look out for the cases where the effect is one that is “chirally enhanced” that is where the sum of up and down masses does appear in the denominator. (A similar issue may also arise when making a heavy-quark expansion; terms that behave like ${\Lambda_{QCD}^2}/{m_b(m_u+m_d)}$, though formally suppressed in the large $m_b$ limit, are not always numerically negligible.) How does isospin help clarify $B$ decay processes? Its chief value is that it allows us to make an experimental separation of some tree and QCD-penguin type contributions. In some processes these have different isospin structure, as well as having different CKM structure. Let us take the example of $B$ decaying to two pions. First let us look at the final states, two pions in a spin zero state. A pion has isopin 1. Naively there are three possible isospins for the two-pion states, 0, 1 and 2. However Bose statistics says the overall state must be even under pion interchange. Since the spin zero spatial state is even, the isopin state must be even too. This eliminates the $I=1$ possibility. Now let us examine the quark decays. The tree $b\rightarrow u{\overline{u}} d$ contribution contains both $\Delta I = 1/2$ and $\Delta I = 3/2$ contributions. These combine with the spectator quark to contribute to the $I=0$ and $I=2$ final states respectively. But a gluon is an isosinglet particle—it has no isospin. Hence the $b \rightarrow d$ QCD penguin graph is purely $\Delta I=1/2$ and contributes only to the $I=0$ final state. (In quark language the gluon makes $u{\overline{u}} + d {\overline{d}}$. ) We can use measurements of several isospin-related channels (Here $B^0 \rightarrow \pi^+\pi^-$, $B^0 \rightarrow \pi^0\pi^0$ and $B^+ \rightarrow \pi^+\pi^0$ and their CP conjugates) to isolate the $I=2$ contribution [@gronaulondon]. Then we have found a pure tree process, which thus depends on only one weak phase (up to small corrections from electroweak penguin effects.) Thus the isospin analysis gives us a way to separate out the dependence on $\alpha$, the difference of the weak phase of the mixing and the weak phase of the tree diagram, without having to calculate the relative strength of the penguin and tree contributions. The theoretical uncertainty that we found in the previous lecture in trying to extract the CKM parameter $\alpha$ from the asymmetry in $B \rightarrow \pi^+\pi^-$ decays can then be much reduced. If, in addition to measuring that time-dependent asymmetry in that channel, one also measures the rates for the isospin related channels, one has, in principle, enough information to determine sin(2$\alpha$). Unfortunately, the $\pi^0\pi^0$ rate is expected to be small, so that it may be some time before the experimental uncertainties of this approach are small enough that the result is actually improved by it. However even an upper bound on the neutral pion rate can provide useful constraints [@bounds]. Electroweak penguin effects can also be considered in an isospin analysis, by writing the isospin structure of the $Z$-boson decay. However, since this decay has isospin 1 as well as isospin 0 parts, there is a $\Delta I = 3/2, I_{\rm final}=2$ contribution, and this cannot be separated from the tree term via any multichannel analysis. This results in some residual theoretical uncertainty in the extraction of $\alpha$, but it is significantly smaller than that from the gluonic penguin contribution without isospin analysis. A similar situation makes isospin analysis useless in separating tree and penguin parts for $b\rightarrow c{\overline{c}} d$ channels such as $D^+D^-$. Here both the tree and penguin contributions are pure $\Delta I = 1/2$, so there is no way to distinguish them via their isospin structure. SU(3) Symmetry -------------- One can get further relationships between different processes if one extends the idea of isospin to the full flavor SU(3), which treats the three lightest quarks as a degenerate triplet. In particular the subgroup of SU(3) known as U-spin under which the down and strange quarks are a doublet gives lots of interesting relationships between amplitudes [@gronauetc]. As with any approximate method, the challenge here is to estimate the size of possible corrections from symmetry breaking effects, that is to estimate the theoretical uncertainty in the predictions. One can distinguish three different types of SU(3) breaking effects. First there are kinematic factors that occur because of the different quark (and hence different meson) masses give different phase space factors. These may be large but can be well-estimated and lead to small theoretical uncertainties for any given set of channels. Second there are the factors of $F_\pi$ (or $f_\pi$) versus the similar factors for the kaon. These are measured numbers so, where a vector or pseudoscalar meson is directly produced by a $W$, they again lead to no significant uncertainties. However when the local operator that produces the light meson is not an axial current then the corresponding ratio is not so well determined. Calculations often use the known ratio of $F$ (or $f$) factors to estimate the SU(3) breaking in such cases also, but now the uncertainty is not so well-controlled. Finally there are cases where the prediction depends also on assuming an SU(3) relationship between the phases of decay amplitudes. Results sensitive to this assumption may have a larger theoretical uncertainty. The application of SU(3)symmetry can allow one to use measured penguin-dominated amplitudes such as $B\rightarrow K\pi$ to constrain the penguin contribution to a tree-dominated amplitude such as $B\rightarrow \pi\pi$. This provides a collection of additional approaches to fix the CKM parameter $\gamma$ from the combined $\pi\pi$ and $K \pi$ data [@gamma]. Another value of both Isospin and SU(3) relationships is that they provide a window to search for effects of physics beyond the Standard Model. There are a number of cases where possible new physics effects do not respect the relationships predicted by these symmetries [@trojanpenguins]. Tests of these relationships may then provide a window for new physics. Lattice Calculations -------------------- Perhaps the best way to include hadronic physics and QCD effects in a calculation of the matrix element of any operator is to use lattice QCD methods. Methods to treat heavy-light mesons on the lattice have been developed and are steadily improving. There are a number of cases where this method will eventually yield theoretical predictions with well controlled errors. Lattice calculation is particularly useful for quantities such as the $B$-mixing matrix element which is a one-particle to one-particle transition, or $f_B$, which is a one-particle to vacuum transition. For one particle to multiparticle transitions (where multi here means two or more) the problem of including final state interactions is not solved by lattice calculations. These calculations are performed in Euclidean space-time and require analytic continuation to give the actual physical result. The uncertainties introduced by this step are difficult to quantify and can be large. There are basically four sources of uncertainties in lattice of calculations of the one-particle to one-particle (or one to zero-particle) matrix elements. The first is the statistical reliability of the Monte-Carlo treatment. This is simply a matter of doing enough calculation, and is very well understood. Second there are the extrapolations and scale-matching to match the finite-volume, finite-lattice-spacing parameters and results with the infinite-volume continuum quantities. Again the process is highly developed and for the most part in good control. Third are the methods of handling the heavy quark on the lattice, which are also now quite well-developed. The critical last ingredient in this progression is for the lattice calculation to be “unquenched”. This means that the lattice allows the development of virtual light quark-antiquark loops. Such calculations require significantly more computer time than the corresponding “quenched calculation” which suppresses quark-loop effects. Unquenched calculations are beginning to appear, for example for the matrix element that is relevant to the mixing between $B$ and ${\overline{B}}$ mesons. There then remains some extrapolation in the light quark masses and in the number and degeneracies of the light quarks. The prospect is that all sources of uncertainty can be investigated, and that, at least for some of the critical quantities, the lattice will eventually provide the most accurate and well-controlled estimates of the matrix elements. Well-controlled here means that the uncertainty in the estimate can be reliably constrained. Quark-Hadron Duality -------------------- Even with all these methods we are again and again confronted with data that cannot be interpreted without further input. We are reduced to using models, or to making further assumptions. One commonly used assumption goes under the name of “quark-hadron duality”. This is the assumption that if I can calculate a quantity, such as an inclusive rate, at the quark level then that calculation must also give the correct answer at the hadronic level. In a situation where we can average over a range of energies one can indeed prove that this must be true for certain averages, for example the energy-averaged total cross-section for electron-positron collisions to produce hadrons. On the other hand it is clear that if we look in detail at any process the quark result, calculated at low order in QCD, can not reproduce all the details of the hadronic spectrum correctly. In particular, thresholds or end-points of spectra are different for quarks and for mesons. Perturbative quark calculations know nothing about resonance masses, at least not in any fixed-order calculation. In a $B$ decay we cannot average over energies, the energy of the decay is set by the $B$ mass. Even so it is popularly believed that inclusive $B$ decays can be well-described using the assumption of quark hadron duality. At the quark level we can calculate the $b$-quark decay. Now we assume that gives the inclusive meson decay correctly, because, if the quark has decayed it must hadronize to something. The level of assurance with which one can make an estimate for the corrections to this approximation varies with the process. For inclusive semi-leptonic decays integrating over lepton momenta provides integration over a range of hadron invariant mass. This can be expected to reduce the corrections. It has thus been argued that these are very small in the inclusive semileptonic case [@bigiuraltsev]. The demands of realistic measurements can also dilute the power of quark-hadron duality. Consider for example inclusive semi-leptonic decays of $B$ mesons to hadrons that contain no charm. In principle the measurement of this total rate can be used to extract a value for the CKM parameter $V_{ub}$, if we can calculate the expected rate. We assume quark-hadron duality gives an accurate result for the full inclusive rate, by the arguments given above. However in any experimental measurement, we must make some kinematic restriction in order to exclude backgrounds coming from the much larger rate of decays to hadrons containing charm quarks. This introduces dependence on details of the spectrum, rather than just a particular integral of it. There is more than one way to choose the kinematic cut: one can for example restrict the electron momentum to be large enough that charm production is excluded; or one can restrict the hadronic invariant mass to be small enough to exclude charm. Because of the unseen neutrino these restrictions are not identical. Each keeps some fraction of the total rate. To extract $V_{ub}$ we must know what that fraction is. But to calculate that fraction we are looking at details of the spectrum for which the use of a quark-level calculation may not be so safe. Recent work has suggested using some combination of cuts on hadron mass and on lepton invariant mass (which requires neutrino reconstruction). A carefully chosen combination can minimize sensitivity to the spectrum end-point details. One can also make some tests as to the stability of the result as the cut prescription is varied [@bauer; @rothstein]. Models and Other Approximations ------------------------------- In many other channels, even once one uses QCD-improved factorization calculations one needs to know a meson-meson transition matrix and/or quark distribution functions for both initial and final state particles to calculate a rate. Lattice calculation, or measurement in a semi-leptonic decay, can be used to fix the transition matrix element. In certain cases one obtains self-consistent quark distribution functions using light-cone QCD arguments. Or one can parameterize these distributions, for example by their moments, and use some set of measurements to fix the set of parameters that dominate an effect (making sure that such parameters are indeed carefully and consistently defined in both processes). Finally one can simply resort to making models for the unknown quantities. One can using rigorous limits obtained from QCD sum rules [@sumrules] and from the heavy quark limit to constrain the models and reduce the number of independent inputs needed. However this is not sufficient to remove all model dependence of the results. There are often still large (and not well-constrained) uncertainties that arise in this stage of the calculation. Summary ------- For two-body hadronic decays even QCD-improved calculations require some input of transition matrix elements and quark distribution functions for the mesons in question in order to calculate amplitudes. These input quantities can sometimes be constrained by symmetries. Rigorous limits for some can be derived for example from the heavy quark limit and from QCD ([*e.g.*]{} the QCD sum rule methods). Some of the quantities of interest can eventually be accurately calculated on the lattice. Some can be measured in semileptonic processes. Data on a great variety of decays will help refine our understanding. This process has already begun. Data from CLEO and from the two asymmetric $B$ factories gives us much to study, and will continue to do so. Our ability to see whether different measurements yield consistent or inconsistent values for the Standard Model parameters is only as good as our ability to constrain the theoretical uncertainties in a reliable fashion. As one applies any method to a multitude of channels one can learn from experience what accuracy is obtained and refine the method on the basis of that experience. Because there are indeed many possible quasi-two-body $B$ decays this process will eventually improve our ability to constrain the theoretical uncertainty of a given calculational method. To achieve this ability it is important for theorists to be as precise and as honest as possible about the sensitivity of any results to input assumptions or models, and to explore this sensitivity in some detail. Only in this way can we find those sets of measurements which truly give us sensitive tests of the Standard Model. Lecture 4. Experiments to Measure $B$ Decays ============================================ In this last lecture I will review how one goes about studying these questions experimentally. Even though you (in this audience) are mostly theory students, it is important that you have some idea of how the measurements are made. The aim of the game is to make multiple measurements that can check Standard Model predictions in a redundant fashion. There are a number of ways that physics from beyond the Standard Model could show up. One could find inconsistent results for a particular Standard Model parameter (or set of parameters) when determining the same parameters by multiple independent methods. One could find a large $CP$-violating asymmetry in a mode for which the Standard model predicts a small or vanishing effect. One could find decay modes that are predicted to be rare present at a rate different from that expected or with a pattern of isospin or SU(3) symmetry violations that cannot be accommodated within the theoretical uncertainty of Standard Model predictions. Each of these possibilities requires ongoing work on both the theory front, to reduce theoretical uncertainties, and the experimental one, to make all the suggested measurements. I will focus on $B$ decay experiments, but rare $K$-decay results also contribute to the picture, as do the existing results on $CP$-violation in $K$ decays. Tagging $B$ Flavor ------------------ Up until now we have talked about various decays of an individual $B$ meson as if we knew what meson we had at time $t=0$. The flavor conservation of strong and electromagnetic interactions means that one produces a $b$-quark and an anti-$b$-quark in the same event. In general one has no [*a priori*]{} knowledge of which type of neutral $B$ meson was formed at production. One must use other properties of the total event in order to determine whether one had a $B^0$ or ${\overline{B}}^0$ meson at production (or at some other known time). This process is called tagging. For example one can tag a $B$ meson when another $B$ meson in the same event decays in such a way that its $b$-flavor is identifiable. An example of a tag is a semileptonic decay; the charge of the lepton then identifies whether it came from the weak decay of a $b$ or a ${\overline{b}}$ quark. The tagging possibilities and efficiencies are quite different in $e^+e^-$ collisions and in hadronic collisions, but the requirement for tagging is common to both types of experiments. In principle almost every event has some tagging information. Often this information is not precise. For example consider the lepton-charge tag suggested above. If the $b$-quark decays hadronically to a $c$-quark which then decays semileptonically then the detected lepton comes from the decay of the $c$ instead of that of the $b$. Assuming it came from the $b$ will give a wrong sign tag. The spectrum of such secondary-decay leptons is different from that of the primary ones. One can use such additional information to improve the correctness of the tag. However the two spectra overlap, so there will still be cases where there is an ambiguity. Only a probability for each tag-type can be determined. Each type of tag event thus has two properties that must be understood, its efficiency, $\epsilon$, and the wrong tag fraction, $w$ associated with it. Some methods have very high purity but low efficiency, others with much higher efficiency may have lower purity. The measure of tagging quality that eventually determines how well we can measure a $CP$-violating asymmetry is the product $\epsilon(1-2w)^2$. We will see below how this comes about. Both the efficiency and the wrong tag fraction are determined by a combination of Monte Carlo modelling of events and measurements, for example from samples of doubly tagged events. A significant systematic uncertainty in the result for any asymmetry arises from the uncertainty in determining the wrong tag fraction. Since that determination is at least in part data driven, this uncertainty will decrease as data samples increase. $e^+e^-$ Collisions ------------------- In an electron-positron collider the most efficient way to produce $B^0$ mesons is to tune the energy to the $\Upsilon_{4s}$, since that large resonant peak in event rate is just above threshold to decay into either a $B^+ $ and a $B^-$ or into a $B^0$ and a ${\overline{B}}^0$. Hence the $\Upsilon_{4s}$ decays essentially $50\%$ to each of these states. Furthermore, the two neutral mesons are produced in a coherent state which, even though both particles are oscillating as described previously, remains exactly one $B^0$ and one ${\overline{B}}^0$ until such time as one of the particles decays. For studies of $CP$-violation this turns out to be either a disaster or a very useful property depending on the design of your collider. To observe $CP$ violation we must look for decays where one of the two neutral $B$’s decays in a way that identifies its flavor, so that it gives a good tag, and the other decays to the $CP$ eigenstate of interest for the study. Then we examine the decay rate as a function of the time, $t$, between the tagging decay (defined to occur at $t=0$) and the $CP$-eigenstate decay. When the tag is a ${\overline{B}}^0$ this means that the particle which decayed to the $CP$ eigenstate is known to have been a $B^0$ at time $t=0$ (or, for $t<0$, to be that combination which would have evolved to be a $B^0$ at time $t=0$). We denote this state as $B^0(t)$. Its decay rate as a function of time is given by $$R(B^0(t)\rightarrow f) = |A(B^0\rightarrow f)|^2e^{-\Gamma |t|} [1 +(1-|\lambda_f|^2)\cos(\Delta m t) +Im\lambda_f \sin(\Delta m t)]$$ where once again $\lambda_f= (q/p)[A({\overline{B}}^0\rightarrow f)/A(B^0\rightarrow f)]$. In this equation and all following discussion of $B_d$ decays we neglect $\Delta \Gamma$, and, equivalently, assume $|q/p|=1$. (The corresponding formulae for $B_s$ decays are a little more complicated as this approximation cannot be used in that case, you can find them in the textbooks [@texts]. ) Likewise, the rate when the tagging decay is a $B^0$ is $$R({\overline{B}}^0(t)\rightarrow f) = |A(B^0\rightarrow f)|^2 e^{-\Gamma |t|} [|\lambda_f|^2 +(|\lambda_f|^2-1)\cos(\Delta m t) -Im\lambda_f\sin(\Delta m t)] \ .$$ Notice that if we were to integrate over all times, $-\infty \le 0\le \infty$ the term proportional to sin$(\Delta M t)$ would integrate to zero. This would destroy our sensitivity to the $CP$-violating quantity Im$\lambda_f$. We must measure the asymmetry between $B$ tags and ${\overline{B}}$ tags as a function of time to avoid this cancellation. For a symmetric electron positron collider running at the $\Upsilon_{4s}$ this is essentially impossible. (This is the disaster referred to above.) The two $B$ mesons are produced with small momenta. Even with the best detectors one cannot accurately measure the difference in distance from the collision point of the two decays. Indeed the size of the beam-beam interaction region is typically sufficient to destroy any possibility of resolving this difference. Hence cannot measure the time-difference between the decays. Pier Oddone suggested an idea that allowed $B$ factories to be built to tackle $CP$ violation [@oddone]. The idea was to build two storage rings with different energies and collide the electrons and positrons so that the $\Upsilon_{4s}$, and likewise the pair of $B$’s to which it decays, are produced moving, with a significant relativistic gamma-factor. Then the physical separation of the decay vertices of the two $B$’s is increased via the time dilation of the decay half-life. (A decay vertex is the point from which the tracks of the particles produced in the decay diverge.) In this case one can indeed, using a precision tracking device known as a vertex detector, resolve the two decay vertices and measure their separation with a resolution that is small compared to the average separation. Furthermore, since any transverse motion of the $B$ mesons is small compared to the overall center-of-mass momentum, the distance between the decays (in the higher-energy beam direction) gives a good measure of the time between them. The uncertainty in the production point due to beam size is irrelevant for this measurement, as we are not concerned with time from production, but only the time between the two decays. Thus the initial coherent state gives a beautiful prediction for a measurable time-dependent asymmetry. The experiment has many internal cross checks that can be made to confirm that the effect is seen as predicted. For a detailed discussion of the physics capabilities of such a facility see for example the BaBar Physics Book, which is available via the web [@babarbook]. To see how the tagging efficiency affects the result consider how the measured asymmetry is related to the actual asymmetry. The total number of events that we count as $B$-tagged events is $\epsilon (N_B (1-w) + N_{{\overline{B}}}w) $ where $N_B$ and $N_{{\overline{B}}}$ are the actual numbers of $B$ and ${\overline{B}}$ events produced. Likewise the total count of ${\overline{B}}$ events is $\epsilon (N_B w + N_{{\overline{B}}}(1-w)) $. Thus the measured asymmetry is $$a_{\rm meas} =(1-2w){(N_B - N_{{\overline{B}}})\over N_B + N_{{\overline{B}}})} = (1-2w)a_{\rm true}$$ where $a_{\rm true}$ is the true asymmetry. In addition the total number of events included in the result scales with $\epsilon$, the tagging efficiency, since only tagged events can be used. Since statistical accuracy grows like the square root of the number of events, the accuracy of the measurement is proportional to the square root of epsilon. Combining these two facts gives you an understanding of the earlier statement that the quality measure for tagging is $\epsilon(1-2w)^2$. This is sometimes called the effective tagging efficiency. Both asymmetric $B$ factory projects, one at SLAC [@pep2] and the other at KEK [@KEKB]), have succeeded spectacularly in building and operating a two-storage-ring facility together with a detector and computer system capable of detecting and recording all the relevant details of millions of $B{\overline{B}}$ events. Interesting data from these facilities is now beginning to be reported and will continue over the next several years to yield new insights. See the websites of the BaBar [@babar] and Belle [@belle] experiments for details. In addition to measuring $CP$-violating asymmetries these facilities are also compiling and analyzing large data samples for a variety of $B_d$ decays. Together with measurements from the symmetric $B$ factory at Cornell [@cesr] and its detector CLEO [@cleo], this data will considerably refine our ability to measure the $CP$-conserving parameters and to test theoretical calculations. I have talked in previous lectures about the uncertainties that plague many theoretical calculation methods, and in particular about the difficulty in quantifying these uncertainties. As data on multiple modes accumulates we can refine our understanding of the accuracy of various approaches by comparison with this data. Proton Colliders ---------------- Because the $B$-factory machine’s are optimized to run at the $\Upsilon_{4s}$ they are below the threshold to produce any $B_s$ mesons. In principle they could do so by running at the $\Upsilon_{5s}$. The smaller peak height of this resonance, together with the fact that it has many possible decay channels combine to make the production rate for $B_s {\overline{B}}_s$ pairs significantly lower than that for $B_d$ at the $\Upsilon_{4s}$. The machines would have to be be re-optimized to run at this higher energy, which itself is not a simple change. All these factors combine to make it unlikely that this will be attempted any time soon, while there is still so much to learn about the $B_d$ decays. So for measurements of $B_s$ decays, and also for those of baryons containing $b$-quarks, we need to look elsewhere, to hadron colliders. For the time being that means the Fermilab TeVatron [@tevatron], eventually it will also mean LHC [@lhc] at CERN. At a hadron collider the $b$ and ${\overline{b}}$ quarks hadronize independently and each $B$ meson is part of a large jet of many particles. Many more $B$’s are produced in high energy hadron-hadron collisions than in an electron-positron $B$ factory. Hadronic collisions also produce many other types of events, with yet higher cross-sections. Thus, for these experiments, it is critical to devise ways to identify $B$-events fast enough to trigger the system to record the event. The trigger is typically two charged tracks emerging from a $B$-decay vertex that is separated from the beam-beam collision region. The design of the trigger and its efficiency is a very important and challenging feature of these experiments. The triggering requirements restrict the decay channels that can be studied in a hadronic environment. The methods and efficiencies for tagging the flavor of the produced $B$ are also quite different in the hadronic case than in the electron-positron $B$ factory environment. The tagging particle may be a charged $B$ or a baryon, or it may be deduced from properties of the leading particles in the jet containing the neutral $B$. Furthermore, since the two $b$-quark (or antiquark) containing particles are not in a coherent state, the time evolution of the $CP$-study particle (and also the tagging particle if it is a neutral $B$-meson) starts at production time. There are a number of interesting quantities that can only be studied in a hadron facility, others where the two types of machines are competitive, and some where the electron-positron machines have unique capabilities. Both approaches are needed to gather all the information we would like to have. An example of a quantity where hadron collider results will be important is the determination of the side $V_{td}$ of the unitarity triangle. Currently this quantity is determined by measuring the $B_d$ mass difference. However there is a significant theoretical uncertainty that arises when relating the measurement to the parameter $V_{td}$. Much of this uncertainty would be removed by a measurement of the $B_s$ mass difference as well as that for $B_d$. The ratio of the two mass differences gives $V_{td}/V_{ts}$ with relatively controlled theoretical uncertainties. If the value predicted by the Standard Model is correct this measurement can be done at Fermilab in the CDF experiment, probably within the next couple of years. There has been a detailed study of the opportunities for $B$ physics in Run II at Fermilab [@TevatronB]. The CDF [@cdf] and D-Zero [@D0] detectors have just completed upgrades and are beginning to take data, including some $B$-physics-triggered data. In addition a new experiment,known as BTeV, with a detector optimized for $B$-physics capability, is planned [@btev]. At CERN there is also such an experiment planned, known as LHCB [@lhcb]. These detectors will give expanded $B$ physics capability and perhaps allow some rare modes to be studied, with branching fractions that are too small to measure in the current experiments. (After my talk I was told there is also a study underway of a possible future $B$ experiment at HERA, a follow-up to the HERA-$B$ experiment [@herab] using a wire target in the proton beam of that $e$-$p$ collider.) Another future option is an intense $Z$-production facility at a linear collider, where study of $Z\rightarrow b {\overline{b}}$ decays can yield useful additional possibilities.) All in all, the problem has many aspects. The complementarity of the different experiments will allow a rich program of measurements. Eventually we will have a clear picture of whether the pattern of results matches the Standard Model or requires some physics beyond the Standard Model to describe the data. Some Final Remarks ------------------ As theorists search for ways to extract interesting information from $B$ decays they will often describe desired measurements that are beyond present capabilities. This is not new. When Bigi and Sanda [@bigisanda] first talked about $CP$-violation in $B$ decays we did not know the $B$ lifetime, so the measurements that they proposed seemed out of reach. Sometimes nature is kind and the numbers work out better than present knowledge suggests. Sometimes clever technical ideas, such as the asymmetric $e^+e^-$ collider, extend our experimental reach. Improvements in the technology of particle tracking and particle identification have been essential in the $B$ factory experiments and will continue to be so for BTeV and LHCB. The history of discovery in science continues because measurements deemed impossible in one era become feasible with new developments. Likewise new developments on the theory side, such as new techniques for unquenched lattice calculations are important, as they allow more measurements to be interpreted with good control of theoretical uncertainties. To conclude this lecture series I would like to remind you that the aim of the game in studying $CP$ is to examine this least-explored corner of the Standard Model in two ways. The first is to pin down the value of the remaining Standard Model parameters. The second is to test whether multiple measurements give consistent answers, both for the parameters and for other Standard Model predictions. The hope is that any discrepancy will be a clue to the nature of physics beyond the Standard Model, physics that can, for example, change the relative phase of a mixing amplitude compared to a decay amplitude. Indirect searches for new physics, such as these $B$ physics probes, are a blunt instrument. Many extensions of the Standard Model may predict similar effects, for example additional contributions to the mixing. The challenge to theorists is to reduce theoretical uncertainties to the point that we sharpen that instrument enough to see the effects if they are there, rather than losing them in the ranges of possible answers given by our poor control of hadronic physics effects. This work is well begun, but there is more to do. I hope some of the students here will make interesting contributions to it in the near future. [99]{} $CP$ VIOLATION. By Gustavo Castelo Branco, Luis Lavoura, Joao Paulo Silva. Oxford Univ. Press, 1999. 511p. (The International Series of Monographs on Physics, Vol. 103) QCD161:B721:1999. $CP$ VIOLATION. By I. I. Bigi and A. I. Sanda. Cambridge Univ. Press, 2000. 382p. (Cambridge Monographs on Particle Physics, Nuclear Physics, and Cosmology, Vol. 9)QCD161:B54:2000. The Babar Physics Book SLAC-Report-504 H. Quinn, 2000 Dirac Medal Lecture, ICTP Trieste July 3, 2001. M. Kobayashi and T. Maskawa, Prog. Theor. Phys.  [**49**]{}, 652 (1973). S. Weinberg, Phys. Rev. Lett.  [**37**]{}, 657 (1976). See, for example, H. Georgi, Phys. Lett. B [**297**]{}, 353 (1992) \[arXiv:hep-ph/9209291\]. J. H. Christenson, J. W. Cronin, V. L. Fitch and R. Turlay, Phys. Rev. Lett.  [**13**]{}, 138 (1964). See for example P. Huet and E. Sather, Phys. Rev. D [**51**]{}, 379 (1995) \[arXiv:hep-ph/9404302\]. R. Svoboda \[Super-Kamiokande Collaboration\], Nucl. Phys. Proc. Suppl.  [**98**]{}, 165 (2001). Q. R. Ahmad [*et al.*]{} \[SNO Collaboration\], solar neutrinos at the Sudbury Neutrino Observatory,” nucl-ex/0106015. Y. Fukuda [*et al.*]{} \[Super-Kamiokande Collaboration\], Phys. Rev. Lett.  [**81**]{}, 1562 (1998) \[hep-ex/9807003\]. See, for example, W. Buchmuller, arXiv:hep-ph/0107153. L. Wolfenstein, Phys. Rev. Lett.  [**51**]{}, 1945 (1983). C. Jarlskog, Phys. Rev. Lett.  [**55**]{}, 1039 (1985). M. Ciuchini [*et al.*]{}, JHEP [**0107**]{}, 013 (2001) \[hep-ph/0012308\]. A. Hocker, H. Lacker, S. Laplace and F. Le Diberder, LAL-01-14, see alsohttp://www.slac.stanford.edu/ laplace/ckmfitter.html B. Aubert [*et al.*]{} \[BaBar Collaboration\], Phys. Rev. Lett.  [**87**]{}, 091801 (2001) \[hep-ex/0107013\]. K. Abe [*et al.*]{} \[Belle Collaboration\], Phys. Rev. Lett.  [**87**]{}, 091802 (2001) \[hep-ex/0107061\]. See, for example, N. Yamada and S. Hashimoto \[JLQCD collaboration\] \[arXiv:hep-ph/0104136\] and references contained therein. I. Dunietz, H. R. Quinn, A. Snyder, W. Toki and H. J. Lipkin, Phys. Rev. D [**43**]{}, 2193 (1991). M. Beneke, G. Buchalla, M. Neubert and C. T. Sachrajda, Phys. Rev. Lett.  [**83**]{}, 1914 (1999) \[hep-ph/9905312\]. M. Beneke, G. Buchalla, M. Neubert and C. T. Sachrajda, Nucl. Phys. B [**606**]{}, 245 (2001) \[hep-ph/0104110\]. Y. Y. Keum, H. Li and A. I. Sanda, Phys. Rev. D [**63**]{}, 054008 (2001) \[hep-ph/0004173\]. H. Quinn and A. I. Sanda, Eur. Phys. J. C [**15**]{} (2000) 626 Web version at pdg.lbl.gov/ J. D. Bjorken, Nucl. Phys. Proc. Suppl.  [**11**]{}, 325 (1989). C. W. Bauer, D. Pirjol and I. W. Stewart, hep-ph/0107002. S. Descotes-Genon and C. T. Sachrajda, arXiv:hep-ph/0109260. M. Gronau and D. London, Phys. Rev. Lett.  [**65**]{}, 3381 (1990). Y. Grossman and H. R. Quinn, Phys. Rev. D [**58**]{}, 017504 (1998) \[arXiv:hep-ph/9712306\]. M. Gronau, D. London, N. Sinha and R. Sinha, Phys. Lett. B [**514**]{}, 315 (2001) \[arXiv:hep-ph/0105308\]. See, for example, M. Gronau, O. F. Hernandez, D. London and J. L. Rosner, Phys. Rev. D [**52**]{}, 6356 (1995) \[hep-ph/9504326\]. A. J. Buras and R. Fleischer, Eur. Phys. J. C [**11**]{}, 93 (1999) \[hep-ph/9810260\]. See, for example, Y. Grossman, M. Neubert and A. L. Kagan, JHEP [**9910**]{}, 029 (1999) \[hep-ph/9909297\]. I. I. Bigi and N. Uraltsev, hep-ph/0106346. C. W. Bauer, Z. Ligeti and M. Luke, hep-ph/0107074. A. K. Leibovich, I. Low and I. Z. Rothstein, Phys. Lett. B [**513**]{}, 83 (2001) \[hep-ph/0105066\]. See, for example, M. Shifman TASI Lectures 1995 in “QCD and Beyond” World Scientific 1995. P. Oddone in Proceedings of the UCLA Workshop: Linear Collider $B{\overline{B}}$ Factory Conceptual Design, D. Stork ed. p243 (1987) SLAC Report 504 (1998) www.slac.stanford.edu/pubs/slacreports/slac-r-504.html www.slac.stanford.edu/accel/pepii/home.html www-acc.kek.jp/WWW-ACC-exp/KEKB/KEKB-home.html www.slac.stanford.edu/BFROOT/www/Public/index.html http://bsunsrv1.kek.jp/ w4.lns.cornell.edu/public/CESR/ w4.lns.cornell.edu/public/CLEO/ adcon.fnal.gov/userb/www/tevatron/ lhc.web.cern.ch/lhc/ Talks and a preliminary draft of the report can be found atwww-theory.fnal.gov/people/ligeti/Brun2/ www-cdf.fnal.gov/ www-d0.fnal.gov/ www-btev.fnal.gov/btev.html lhcb.cern.ch/ www-hera-b.desy.de I. I. Bigi and A. I. Sanda, Nucl. Phys. B [**193**]{}, 85 (1981). A. B. Carter and A. I. Sanda, Phys. Rev. D [**23**]{}, 1567 (1981). [^1]: Work supported by Department of Energy contract DE–AC03–76SF00515.

--- abstract: 'We present a method to characterize non-Abelian anyons that is based only on static measurements and that does not rely on any form of interference. For geometries where the anyonic statistics can be revealed by rigid rotations of the anyons, we link this property to the angular momentum of the initial state. We test our method on the paradigmatic example of the Moore-Read state, that is known to support excitations with non-Abelian statistics of Ising type. As an example, we reveal the presence of different fusion channels for two such excitations, a defining feature of non-Abelian anyons. This is obtained by measuring density-profile properties, like the mean square radius of the system or the depletion generated by the anyons. Our study paves the way to novel methods for characterizing non-Abelian anyons, both in the experimental and theoretical domains.' author: - 'E. Macaluso' - 'T. Comparin' - 'L. Mazza' - 'I. Carusotto' bibliography: - 'bibliography.bib' title: 'Fusion channels of non-Abelian anyons from angular-momentum and density-profile measurements' --- [*Introduction.—*]{} The standard classification of particles into bosons and fermions breaks down in two spatial dimensions, where exotic objects known as *anyons* can exist [@Wilczek_PRL.49.957; @Halperin_PRL.52.1583; @Arovas_Wilczek_PRL.53.722; @Leinaas_Myrheim_Nuovo_Cimento; @Wu_PRL.52.2103; @Stern_AoP]. The key concepts for defining the statistics of anyons are the adiabatic motion of one anyon around another, hereafter the *braiding*, and the adiabatic *exchange* of the anyons positions [@Tong_notes]. Anyons can be characterized by merging two of them, and the properties of the new composite object depend on the fusion rules of the original anyons. When there is the possibility of fusing in more than one way, anyons can be non-Abelian [@Moore_NPB.360.2.362; @ReadGreen_PRB.61.10267; @Ivanov_PRL.86.268; @Read_Rezayi_PRB.59.8084]: they are the heart of topological quantum computation [@Nayak_RMP.80.1083], and their experimental realization is thus highly desired. Several existing platforms are expected to host them as emergent quasi-particles, but the unambiguous experimental demonstration of their properties is still the matter of an intense debate [@Willett_PNAS.106.22; @Mourik_Science.336.6084]. In the last twenty years, several works addressed the problem of extracting the properties of the anyons hosted by the ground states of a given Hamiltonian. The simplest approach relies on explicitly following the ground-state evolution when anyons are exchanged [@ParedesZoller_PRL.87.010402; @Tserkovnyak_PRL.90.016802; @Baraban_Simon_numerical_analysis_MR_qh_wfs; @Wu_PRL.113.116801; @Nielsen_PRB.91.041106]. Within other approaches, the analytical study of paradigmatic wave functions has also clarified important issues about the statistics of excitations [@NayakWilczek_NuclPhysB479.3_1996; @Nayak_plasma_Ising-type_FQH; @Ivanov_PRL.86.268]. On the experimental side, interferometric schemes have been proposed to compare the state before and after the adiabatic time evolution [@Halperin_Rosenow_PRB.83.155440; @Campagnano_Gefen_PRL.109.106802; @ParedesZoller_PRL.87.010402; @DasSarma_Nayak_PRL.94.166802; @Stern_Halperin_PRL.96.016802; @Bonderson_Shtengel_PRL.96.016803], but none of them has produced unambiguous results [@Camino_Goldman_PRB.72.075342; @Rosenow_Halperin_PRL.98.106801]. We propose a method to characterize non-Abelian anyons: By considering geometries where the anyonic statistics can be revealed through rigid rotations of the anyons (see Fig. \[fig:braiding\_rigid\_rotations\]), we relate their statistical phase to the angular momentum and to the density profile of the system. This protocol allows one to identify the existence of different fusion channels, a defining property of non-Abelian anyons, with remarkable experimental simplicity in the context of ultracold atoms [@Pitaevskii_Stringari_2016; @Cooper_etal_TopoBands_RMP] and photons [@Carusotto_Ciuti_RMP.85.299; @Ozawa_etal_TopoPhoto_RMP]. Moreover, our study represents a powerful theoretical tool to inspect excitations with unknown statistics, going beyond the observation of multiple fusion channels. As a showcase study, we discuss our method for the case of the Moore-Read (MR) state [@Moore_NPB.360.2.362], and outline an experimental procedure for computing the statistical phases of its quasiholes. ![**a**-**c**, Rigid rotations of two anyons (panels [**a**]{} and [**b**]{}) or four anyons (panels [**c**]{}). Rotation angles are such that the set of anyonic coordinates (red dots) remains the same.[]{data-label="fig:braiding_rigid_rotations"}](fig_rigid_rotations_bis2.pdf){width="48.00000%"} [*Rigid rotations of the anyons.—*]{} We consider a two-dimensional (2D) system of $N$ particles (bosons or fermions) supporting anyonic excitations. The Hamiltonian $\hat H_1$ is a function of particle positions and momenta, as well as of time. We use the complex coordinate notation $z_j = x_j + i y_j$ for the position of the $j$-th particle. The time dependence of $\hat H_1 (\partial_{z_j}, \partial_{\bar z_j},z_j, \bar z_j; t)$ is only due to a set of parameters $\eta_\mu(t)$ defining the centers of some external local potentials $V_\mathrm{ext}(z, \eta_\mu(t))$. These potentials typically couple with the particle density, creating and pinning the anyons at positions $\eta_\mu(t)$ [@ParedesZoller_PRL.87.010402; @MacalusoCarusotto_PRA.98.013605; @Wan_Rezayi_PRL.97.256804; @Wan_Yang_PRB.77.165316; @Prodan_Haldane_PRB.80.115121]. To reveal the anyonic statistics, one option is to braid the anyons through rigid rotations of the pinning-potential coordinates (see Fig. \[fig:braiding\_rigid\_rotations\]). These transformations are defined as $$\eta_{\mu} (t) = \eta_{\mu} (0) e^{i \theta(t)}, \qquad \theta(t) = \frac {t}{T} \theta_f , \label{Eq:Anyon:Rotation}$$ where $\theta_f$ is the final rotation angle and $T$ is the time duration of the process. Since we consider rigid rotations, we can study the problem in the reference frame $R_2$ co-rotating with the anyons, rather than using the laboratory reference frame $R_1$. We assume that $V_\mathrm{ext}(z,\eta_\mu(t))$ is a function of the distance $|z - \eta_\mu(t)|$ between particles and anyons, and that the remaining terms in $\hat{H}_1$ are rotationally invariant. Under these assumptions, the generator of the time evolution in $R_2$ in the time span $[0,T]$ reads [@Pitaevskii_Stringari_2016]: $$\hat H_2 (\partial_{z_j}, \partial_{\bar z_j},z_j, \bar z_j; t) = \hat H_1 (\partial_{z_j}, \partial_{\bar z_j},z_j, \bar z_j; t=0) - \frac{\theta_f}{T} \hat L_z,$$ which is manifestly time-independent. The first term on the right-hand side is the initial Hamiltonian in $R_1$, while the second one describes the effect of the rotation. Being interested in an adiabatic process, we consider $T \to \infty$. The rotation term is then a small contribution and can be treated perturbatively. To describe the dynamics in $R_2$, we consider an initial state $\ket{\Psi_0}$ belonging to the $m$-fold degenerate ground-state manifold $ \mathcal{H}_{E_0}$, spanned by the basis $\lbrace\ket{\psi_{\alpha}}\rbrace_{\alpha=1,\dots,m}$, with $\hat{H}_1(t=0)\ket{\psi_{\alpha}}=E_0\ket{\psi_{\alpha}}$ and $\braket{\psi_{\alpha}|\psi_{\beta}} = \delta_{\alpha\beta}$. If the dynamics is slow enough, we can use the adiabatic theorem to state that the dynamics is restricted to $\mathcal{H}_{E_0}$ (an explicit proof is in [@SuppMat]), and make the following ansatz: $$\ket{\Psi_2(t)} = e^{- i E_0 t / \hbar} \sum_{\alpha = 1}^m \gamma_{\alpha}(t) \ket{\psi_{\alpha}}, \quad \gamma_\alpha(0) = \braket{\psi_{\alpha} | \Psi_0}. \label{Eq:State:2:Ansatz}$$ By applying the Schrödinger equation, we recover the time-evolution equation of the $\gamma_\alpha$’s: $$i \hbar \frac{\mathrm d \gamma_{\alpha}(t)}{\mathrm d t} = -\frac{\theta_f}{T}\sum_{\beta=1}^m \mathcal{L}_{\alpha\beta}\, \gamma_{\beta} (t), \label{Eq:Diff:gamma}$$ where $\mathcal{L}_{\alpha\beta} = \braket{\psi_{\alpha} | \hat{L}_z | \psi_{\beta}}$ is the angular momentum restricted to $\mathcal{H}_{E_{0}}$. The solution reads $$\ket{\Psi_2(T)} = e^{- i \hat{H}_2 T / \hbar} \ket{\Psi_0} = e^{- i E_0 T / \hbar}\, e^{i \theta_f \mathcal L / \hbar} \ket{\Psi_0}, \label{Eq:Psi2}$$ in terms of the matrix exponential $\exp \left[i \theta_f \mathcal L / \hbar \right]$. To find the state $\ket{\Psi_1(T)}$ in the laboratory frame, we need to rotate $\ket{\Psi_2(T)}$ by an angle $\theta_f$: $$\begin{aligned} \ket{\Psi_1(T)} &= e^{- i \theta_f \hat{ L}_z / \hbar} \ket{\Psi_2(T)} \\ &= e^{- i E_0 T / \hbar}\, e^{- i \theta_f \hat{ L}_z / \hbar}\, e^{i \theta_f \mathcal L / \hbar} \ket{\Psi_0}. \end{aligned} \label{Eq:Psi1}$$ The state in equation  is the exact result for an adiabatic braiding process performed through a rigid rotation of all anyons by an angle $\theta_f$. We recognize a dynamical phase proportional to $T$, that is unessential to the discussion of non-Abelian statistics and therefore neglected from now on. The remaining geometric contribution is the product of two unitary matrices: $\mathcal{B}$, with matrix elements $\mathcal{B}_{\alpha\beta} = \bra{\psi_{\alpha}} e^{- i \theta_f \hat L_z /\hbar} \ket{\psi_{\beta}}$, and $\mathcal{U}_\mathrm{B} \equiv e^{i \theta_f \mathcal L / \hbar}$, which is the Berry matrix of the adiabatic process under study, once one makes a suitable choice of the basis states for each angle $\theta(t)$ [@SuppMat]. To guarantee that the ground-state manifold is $\mathcal{H}_{E_0}$ at both times [@Nayak_RMP.80.1083], the angle $\theta_f$ must be such that $\hat{H}_1(t)$ is the same at times $t=0$ and $T$. Depending on the anyon positions, this constraint can be satisfied even for rotation angles which are not multiple of $2\pi$ \[see Fig. \[fig:braiding\_rigid\_rotations\] **b**-**c**\]. When $\theta_f = 2 \pi k$, with $k$ integer, $\mathcal{B}$ is trivially the identity matrix. In this case, $\mathcal{U}_\mathrm{B}$ encodes the full geometrical contribution to the time evolution, made up of both topological and non-topological parts. We stress that $\mathcal{U}_\mathrm{B}$ only depends on measurable properties of the ground-state manifold at the initial time, namely the angular-momentum matrix elements. Therefore no actual time evolution is needed to measure it, which constitutes an undeniable experimental advantage. The case of $\theta_f \neq 2 \pi k$ is relevant in the theoretical context, where –in contrast with experimental studies– nothing precludes the extraction of $\mathcal{B}$ \[see example in Ref. [@SuppMat]\]. A comprehensive analysis of this case is left for a future work. {width="99.00000%"} [*Moore-Read state and its quasihole excitations.—*]{} We now consider the MR state, which is described by the wave function [@Moore_NPB.360.2.362] $$\Psi (\{z_j\}) = \text{Pf}(W) \prod_{i<j} \left(z_{i} - z_{j} \right)^{M} e^{- \sum_{i} |z_{i}|^{2} / 4 l^{2}_{B} }, \label{eq:MR_wf}$$ where $l_{\text{B}}$ is the magnetic length. $\text{Pf}(W)$ denotes the Pfaffian of the $N\times N$ anti-symmetric matrix $W$, with $W_{ij} = 1 / (z_i - z_j)$ for $i\neq j$. For even (odd) values of the positive integer $M$, this wave function represents a fermionic (bosonic) FQH state at filling $\nu = 1/M$, which belongs to the lowest Landau level (LLL) [@Tong_notes]. This state is the ground state for 2D charged particles, in the presence of a transverse magnetic field and of a specific three-body repulsion [@Greiter_PRL.66.3205], and it is believed to be in the same universality class of the FQH state observed at filling $\nu=5/2$ [@Greiter_NPB.374.567; @Morf_PRL.80.1505; @Rezayi_Haldane_PRL.84.4685]. In the presence of properly designed external potentials, the ground state may also host a specific number of localized anyonic excitations [@Wan_Rezayi_PRL.97.256804; @Wan_Yang_PRB.77.165316; @Prodan_Haldane_PRB.80.115121]. The quasihole (QH) excitations of the MR state obey non-Abelian statistics [@Moore_NPB.360.2.362; @NayakWilczek_NuclPhysB479.3_1996; @Nayak_RMP.80.1083]. In particular, they are Ising anyons with an additional Abelian contribution to their statistical phase, and they can fuse in two different ways. For a given set of coordinates $(\eta_{1}, \dots, \eta_{2n})$ of $2n$ such QHs, there is a $2^{n-1}$-fold degenerate set of states [@NayakWilczek_NuclPhysB479.3_1996]. In the following, we will consider the case $2n = 2$, for which the system is not degenerate. In this case, the MR wave function $\Psi^{2\text{QH}}$ has the same form as in Eq. ; yet the anti-symmetric matrix $W$ depends on the even/odd parity $P_{N} = 0,1$ of the particle number $N$. For $P_N = 0$, it is $N \times N$ and reads $$W_{ij} = \dfrac{(\eta_1 - z_i)(\eta_2 - z_j) + (i \leftrightarrow j)}{z_i - z_j} \quad \forall \, i \neq j. \label{eq:MR_2QH}$$ For $P_N = 1$, on the other hand, $W$ is a $(N+1)\times(N+1)$ matrix. The $N \times N$ upper-left block is defined as in Eq. , while the entries of the $(N+1)$-th row (column) are equal to $+1$ ($-1$) [@SuppMat]. The fusion channel of the two QHs depends on $P_N$ [@Nayak_plasma_Ising-type_FQH]. As a consequence, the braiding of two MR QHs induces a phase $\varphi_{\rm br}$ that depends on $P_N$: $$\dfrac{\varphi_{\text{br}}}{2\pi} = \dfrac{1}{4M} - \dfrac{1}{8} + \dfrac{P_{N}}{2} . \label{eq:phi_br_MR_QHs}$$ The dependence of the braiding phase $\varphi_\text{br}$ on $P_N$ is thus a direct indication of the non-Abelian statistics of QHs, because it indicates that the two QHs are in different fusion channels when $N$ is even or odd [@Nayak_RMP.80.1083]. [*$\varphi_{\text{br}}$ from the mean square radius.—*]{} As previously mentioned, for a $2\pi$-rotation of the QHs, $\mathcal{B}$ is the identity matrix. For the non-degenerate MR state with two QHs, the unitary transformation $\mathcal{U}(T)$ associated with this process reduces to the phase factor $\mathcal{U}_{\text{B}} = e^{i \varphi_{\text{B}}}$, where $\varphi_{\text{B}} = 2\pi \mathcal{L}/\hbar$ is the Berry phase. In this case, $\mathcal{L}$ is the expectation value of the angular-momentum operator over the initial state, $\langle \hat{L}_{z} \rangle$. The Berry phase $\varphi_{\text{B}}$ has a non-topological contribution, which can be interpreted as an Aharonov-Bohm phase [@SuppMat]. Although this phase factor contains information on the QH fractional charge, we have to remove it to isolate the QH braiding phase $\varphi_{\text{br}}$. To this purpose we consider the difference between the Berry phases for two particular states \[see Fig. \[fig:phi\_br\_depletions\] **a** and **b**\]: $$\dfrac{\varphi_{\text{br}}}{2\pi} = \dfrac{1}{\hbar} \left[\langle \hat{L}_{z} \rangle_{|\eta_{1}|=|\eta_{2}|} - \langle \hat{L}_{z} \rangle_{\eta_{1} = \eta_{2}} \right] . \label{eq:phi_br_Lz}$$ The expectation value $\langle \hat{L}_{z} \rangle_{|\eta_{1}|=|\eta_{2}|}$ is taken on a state with QHs sufficiently far from each other, at positions $\eta_{1}$ and $\eta_{2}$ such that $|\eta_1| = |\eta_2|$. On the other hand, $\langle \hat{L}_{z} \rangle_{\eta_{1}=\eta_{2}}$ is measured on the state with the two QHs on top of each other at $\eta_{1} = \eta_{2}$ \[for details, see Ref. [@SuppMat]\]. The mean angular momentum of a state in the LLL is related to its mean square radius: $\langle \hat{L}_{z} \rangle / \hbar + N = N \langle r^{2} \rangle / 2 l^{2}_{B}$ [@Ho_Mueller_PRL.89.050401; @Umucalilar_PRL.120.230403]. This simplifies equation  which reads $$\dfrac{\varphi_{\text{br}}}{2\pi} = \dfrac{N}{2 l^{2}_{B}} \left[\langle r^{2} \rangle_{|\eta_{1}|=|\eta_{2}|} - \langle r^{2} \rangle_{\eta_{1} = \eta_{2}} \right] . \label{eq:phi_br_msr}$$ Moreover, within the LLL approximation, the mean square radius of the cloud, and so $\varphi_{\text{br}}$, can be measured after time-of-flight expansion [@Read_Cooper_PRA.68.035601; @Umucalilar_PRL.120.230403]. To validate equation , we compute $\langle r^{2} \rangle$ through the Monte Carlo technique [@SuppMat]. Numerical results –reported in Table \[tab:phi\_br\_2qhs\] for both $M=2$ (fermionic case) and $M=1$ (bosonic case) and for different parities $P_N$ of the particle number $N$– are fully compatible with equation . This demonstrates that the existence of multiple fusion channels for the MR QHs can be experimentally probed without braiding them. $M$ $P_{N}$ $\varphi_{\text{br}}^\mathrm{MC}$ $[2 \pi]$ $\varphi_{\text{br}}$ $[2 \pi]$ ----- --------- --------------------------------------------- --------------------------------- $0$ $0.05\pm0.06$ $0$ $1$ $0.49\pm0.07$ $0.5$ $0$ $0.13\pm0.04$ $0.125$ $1$ $0.59\pm0.04$ $0.625$ : Quasihole braiding phase $\varphi_{\text{br}}^\mathrm{MC}$ obtained numerically via equation  (third column, with the Monte Carlo statistical uncertainty) and its prediction $\varphi_{\text{br}}$ in equation  (fourth column), for $M=2, 1$ and for different parities $P_{N}$ of the particle number $N=150$ and $N=149$. For the $|\eta_1| = |\eta_2|$ term in equation , we set $\eta_1 = - \eta_2$, which is the optimal configuration for a finite-size system. For $M=2$ ($M=1$) case, $|\eta_1|/l_B$ is equal to 7.5 (6.5). []{data-label="tab:phi_br_2qhs"} [*$\varphi_{\text{br}}$ from the quasihole density depletions.—*]{} Although the protocol suggested in equation  is already close to the current experimental capabilities, it requires the ability to pin QHs with high precision and the knowledge of the particle number. Moreover, $\varphi_\text{br}$ is difficult to compute for large systems, since it is a $\mathcal{O}(1)$ number obtained as the difference between two $\mathcal{O}(N^2)$ quantities. However, equation  can be recast in a form which does not depend neither on $N$ nor on the precise QH positions, as we prove in the following. Due to the incompressibility of the FQH states [@Tong_notes], the densities of the configurations under study only differ in the regions $A_1$ and $A_2$ surrounding the QHs \[see red circles in Fig. \[fig:phi\_br\_depletions\] **a** and **b**\]. Therefore, the integrals in equation  can be restricted to $A_{1}$ and $A_{2}$: $$\dfrac{\varphi_{\text{br}}}{2\pi} = \dfrac{1}{2 l^{2}_{B}} \int_{A_{1}, A_{2}} r^{2} \left[ n_{|\eta_{1}|=|\eta_{2}|}(\vec{r}) -n_{\eta_{1}=\eta_{2}}(\vec{r}) \right] d\vec{r} . \label{eq:phi_br:integral}$$ In these regions, the densities in Eq.  can be expressed in terms of the density depletions $d_{1\text{QH}}$ and $d_{2\text{QH}}$ caused by a single QH and two overlapping QHs [@SuppMat]. This allows us to write the braiding phase as $$\dfrac{\varphi_{\text{br}}}{2\pi} = \dfrac{1}{2 l^{2}_{B}} \int d\vec{\rho} \, \rho^{2} \left[ d_{2\text{QH}} (\vec{\rho}) - 2 d_{1\text{QH}} (\vec{\rho}) \right] , \label{eq:phi_br_from_depletions}$$ in which $\vec{\rho}$ is the distance from a QH position, $d_{1\text{QH}}(\vec{\rho}) = n_{\text{b}} - n_{|\eta_{1}|=|\eta_{2}|}(\vec{\rho} + \eta_{i})$ and $d_{2\text{QH}}(\vec{\rho}) = n_{\text{b}} - n_{\eta_{1}=\eta_{2}}(\vec{\rho} + \eta_{i})$ are the aforementioned QH density depletions, w.r.t. the bulk density $n_{\text{b}} = 1/2 \pi M l^{2}_{\text{B}}$ \[see Fig. \[fig:phi\_br\_depletions\] **c** and **e**\]. The integration region must be large enough to ensure an appropriate decay of the density oscillations induced by the QHs. At the same time, a cutoff $\rho < R_\mathrm{max}$ is needed to avoid spurious contributions coming from the density deformations generated at the cloud boundaries. The numerical validation of equation  is shown in Fig. \[fig:phi\_br\_depletions\] **d**, **f** for the different parities $P_{N}$, and for $M=2, 1$. Residual deviations from the expected $\varphi_\text{br}$ are due to finite-size effects. Equation  constitutes an operative way to measure $\varphi_\text{br}$, which depends only on local properties in the bulk region. As such, it is robust against edge modes, which are the typical low-energy excitations due to finite-temperature effects [@MacalusoCarusotto_PRA.96.043607; @MacalusoCarusotto_PRA.98.013605]. Moreover, since $d_{1\text{QH}}(\rho)$ does not depend on $P_{N}$ \[see Fig. \[fig:phi\_br\_depletions\] **c** and **e**\], all the information on the fusion channels is encoded in $d_{2\text{QH}}(\rho)$, which is completely different for even and odd values of $N$. Although this dependence on $P_N$ was already known [@Prodan_Haldane_PRB.80.115121; @Baraban_thesis], the key result of our work is that the depletion profiles also contain quantitative information on the braiding phase. Note that this result holds for the QH excitations of any state in the LLL. [*Experimental procedure.—*]{} While $d_{1\text{QH}}(\rho)$ can be indifferently measured in the ground state with either a single QH or two well-separated ones [@Wan_Rezayi_PRL.97.256804; @Wan_Yang_PRB.77.165316], the characterization of two overlapping QHs involves more subtleties: First, the state in Eq.  with overlapping QHs may not be the ground state in the presence of a given external potential. For instance, for odd parity $P_{N}$, having two QHs close to each other might cost more energy than just exciting a low-energy fermionic excitation at the boundary [@Wen_AdvPhys.44.5.405; @Milovanovic_Read_PRB.53.13559; @Wan_Rezayi_PRL.97.256804]. Furthermore, the presence of these fermionic edge modes may modify the relation between the QHs fusion channel and the particle number parity $P_{N}$ \[see footnote \[33\] in Ref. [@Wan_Yang_PRB.77.165316]\]. We thus propose to proceed as follows for the measurement of $d_{2\text{QH}}(\rho)$: two QHs are created far apart, by cooling the system in the presence of pinning potentials. The two QHs are then slowly brought closer and fused [@SuppMat]. According to the general theory of topological quantum computation [@Kitaev_AOP.303; @Nayak_RMP.80.1083], the fusion channel cannot change during this process, so the system is adiabatically transported into the (possibly metastable) desired state, where the depletion profile $d_{2\text{QH}}(\rho)$ is measured. Note that unless special care is taken, we can argue that in an actual experiment the QH fusion channel will be randomly chosen at each repetition [@SuppMat]. Nonetheless, the non-Abelian statistics of the QHs will be still visible in the bi-peaked probability for $\varphi_{\text{br}}$. A rigorous proof of this statement requires numerical experiments based on a model Hamiltonian and a particular cooling mechanism; we leave it for a future study. [*Conclusions and Outlook.—*]{} In this work, we presented a scheme to assess the statistical properties of anyonic excitations which does not rely on any kind of interference. Our protocol is based on a mathematical link between statistics and angular-momentum measurements, derived by considering rigid rotations of the anyons. This relation further simplifies for states in the LLL, where anyonic statistics is encoded in the density profile. Having access to the anyonic statistics without performing any interference scheme is remarkable in itself; moreover, relating statistics to density measurements makes our protocol readily applicable to state-of-the-art experiments with ultracold atoms [@Cooper_etal_TopoBands_RMP] and photons [@Ozawa_etal_TopoPhoto_RMP]. Beyond the identification of the Moore-Read fusion channels, on which our scheme has been validated, the study of the two-anyons case opens several other perspectives. For example, our method can be employed to distinguish the Moore-Read and anti-Pfaffian states, whose quasiholes have different Abelian contributions to the braiding phase [@Levin_Rosenow_PRL.99.236806; @Lee_Fisher_PRL.99.236807; @Son_PRX.5.031027; @Simon_PRB.97.121406; @Feldman_PRB.98.167401; @Simon_PRB.98.167402]. Moreover, it gives access to a key property in topological quantum computation [@Kitaev_AOP.303; @Nayak_RMP.80.1083], namely the dependence of the braiding phase on the distance between the anyons [@Baraban_Simon_numerical_analysis_MR_qh_wfs]. Our method can also be useful for theoretical studies of states supporting anyons of unknown type. When one can compute the matrix elements of the angular-momentum and rotation operators in the ground-state manifold, our scheme gives access to all contributions to the time-evolution operator, for any rigid rotation of the anyons. We stress that in the case of non-Abelian anyons rigid rotations are sufficient to induce non-trivial mixing of the ground states [@Nayak_RMP.80.1083], although only a subset of the possible anyonic exchanges is accessible in this way [@SuppMat]. Therefore, we envision the possibility of a more precise theoretical characterization of the anyons, beyond the present identification of fusion channels. Natural extensions of our analysis include other states in the LLL –like the Read-Rezayi state [@Read_Rezayi_PRB.59.8084]– or the p-wave superconductor, closely related to the Moore-Read state [@ReadGreen_PRB.61.10267]. An exciting question is whether the link between the anyonic statistics and the system density remains valid also for lattice systems [@Hafezi_Lukin_PRA.76.023613; @Mazza_Cirac_PRA.82.043629; @Regnault_Bernevig_PRX.1.021014; @Wu_PRL.113.116801; @Hafezi_PRB.90.060503]; this is the subject of ongoing study [@EM_etal_lattice_QHs]. This work was supported by the EU-FET Proactive grant AQuS, Project No. 640800, and by the Autonomous Province of Trento, partially through the project “On silicon chip quantum optics for quantum computing and secure communications" (“SiQuro"). Stimulating discussions with P. Bonderson, M. Fremling, M. O. Goerbig, V. Gurarie, C. Nayak, N. Regnault, M. Rizzi, S. H. Simon, J. K. Slingerland, and R. O. Umucal[i]{}lar are warmly acknowledged.

--- abstract: 'In this letter, we investigate the resource allocation for downlink multi-cell coordinated OFDMA wireless networks, in which power allocation and subcarrier scheduling are jointly optimized. Aiming at maximizing the weighted sum of the minimal user rates (WSMR) of coordinated cells under individual power constraints at each base station, an effective distributed resource allocation algorithm using a modified decomposition method is proposed, which is suitable by practical implementation due to its low complexity and fast convergence speed. Simulation results demonstrate that the proposed decentralized algorithm provides substantial throughput gains with lower computational cost compared to existing schemes.' author: - | Zesong Fei$^{*}$ , Shuo Li, Chengwen Xing, Yiqing Zhou and Jingming Kuang\ [^1] title: 'Distributed Resource Allocation Algorithm Design for Multi-Cell Networks Based on Advanced Decomposition Theory' --- Resource allocation, multi-cell network, optimization theory Introduction ============ OFDMA has been widely accepted as a promising multiple access technique for future mobile communication systems. Recently, OFDMA-based multi-cell coordination becomes a hot research topic, which could provide superior performance over the traditional single-cell processing network through joint signal processing among the involved based stations [@Wang2009]. Moreover, the performance of the multi-cell coordination network can be further improved by employing multi-cell resource allocation. However, the traditional centralized multi-cell resource allocation approach is too complicated to be practical because of the large number of parameters and constraints [@KTH]. Therefore, various distributed algorithms have been proposed to reduce the computation complexity and increase the scalability, such as the distributed algorithm for the resource allocation using game theory in [@game] and the one based on the Lagrange duality method in [@Zhangrui]. In this paper, a novel distributed resource allocation algorithm is proposed for multi-cell OFDMA networks. Different from existing works in which weighted sum of user rates (WSR) are usually used, the weighted sum of the minimal user rates (WSMR) of coordinated cells is taken as the performance criterion, which should be maximized subject to individual power and subcarrier allocation constraint at each BS. An iterative algorithm is proposed to solve the problem, which optimizes the subcarrier scheduling and the power allocation alternatively. For the power allocation subproblem, motivated by a natural decomposition of the optimality conditions of the original problem, we propose a novel distributed algorithm. Unlike traditional Lagrangian based algorithms, the proposed technique does not need to solve subproblems, which results in computational savings and fast convergence. Furthermore, in the proposed procedure the central agent only distributes information and checks the convergence condition without the need to update information as [@Zhangrui], which makes the scheme simpler. System Model And Problem Formulation {#sect:system} ==================================== System Model ------------ In this paper, we investigate the distributed resource allocation in the downlink of a cellular OFDMA network consisting of $M$ cooperative cells. In the $m^{\rm{th}}$ cell, BS $m$ serves $K_{m}$ users and the active user set is denoted as $U_m$. The channel power coefficient of subcarrier $n$ from BS $l$ to user $u$ in the $m^{\rm{th}}$ cell is denoted by $g^{l}_{u,m,n}$ . If $l \neq m$, $g^{l}_{u,m,n}$ is related to the interfering channel power from BS $l$ , and $g^{l}_{u,m,n}$ denotes the desired channel power coefficient from BS $m$ if $l=m$. In addition, $\sigma^{2}_{u,m,n}$ is the power of the additive white Gaussian noise at user $u$ in cell $m$ on subcarrier $n$. The number of OFDMA subcarriers is $N$ and each subcarrier is allocated to only one user exclusively in each cell. It is also assumed that all channel state information (CSI) is perfectly known at each BS. The maximum transmit power of BS $m$ is denoted as $P_{m,max}$, and $P_{m,n}$ represents the power allocated to subcarrier $n$ by BS $m$. For convenience, $\{P_{m,n}\}^{M}_{m=1}$ is stacked into a $M \times 1$ vector $\mathbf{p}_{n}=[ P_{1,n},\ldots,P_{M,n}]^{T}$, and then $\{\mathbf{p}_{n}\}^{N}_{n=1}$ is stacked into a $M \times N$ matrix $\mathbf{P}$ referring to power allocation in following sections. As for the subcarrier allocation notations, we define a binary variable $A_{u,m,n}$, which indicates that subcarrier $n$ is allocated to user $u$ in cell $m$ if $A_{u,m,n}=1$. Stack $\{A_{u,m,n}\}^{N}_{n=1}$ into a $N \times 1$ vector $\mathbf{A}_{u,m}=[A_{u,m,1},\ldots,A_{u,m,N}]^{T}$, and merge $\mathbf{A}_{u,m}$ of all users in all cells to a $N \times K_{m} \times M$ subcarrier allocation matrix $\mathbf{A}$ which indicates how subcarriers are assigned among all users. Then the information rate of user $u$ in cell $m$ can be expressed as a function of $\mathbf{A}_{u,m}$ and $\mathbf{P}$ $$\label{user rate with A} \begin{split} R_{u,m}(\mathbf{A}_{u,m},\mathbf{P}) &= {\sum}_{n=1}^N A_{u,m,n} \cdot R_{u,m,n}(\mathbf{p}_n)\\ % &= \sum_{n:A_{u,m,n=1}}R_{u,m,n}(\mathbf{p}_n) \end{split}$$ where $R_{u,m,n}$ is the achievable rate of user $u$ on subcarrier $n$ in cell $m$, and can be calculated as $$\label{user rate without A} R_{u,m,n}(\mathbf{p}_n) = \textrm{ln} \left(1+ \frac{P_{m,n}g^{m}_{u,m,n}}{(\sigma^{2}_{u,m,n}+\sum_{l=1,l\neq m}^M P_{l,n}g^{l}_{u,m,n})\Gamma}\right)$$ in the unit of nats per OFDM symbol, where $\gamma_{u,m,n}(\mathbf{p}_n)$ is the signal to interference plus noise ratio (SINR), and $\Gamma$ represents the signal to noise ratio gap between the adopted modulation and coding scheme and the one achieving capacity. Problem Formulation ------------------- For resource allocation, sum capacity is the most widely used performance criterion. However, since OFDMA multi-cell networks with multiple users in each cell, maximizing the sum capacity of a multi-cell network may result in a serious performance imbalance. It is because that the users and cells with better channel conditions will be allocated with much more resources and those experiencing worse channels may be sacrificed. In order to overcome this kind of problems fairness among the users and cells must be taken into account. Therefore, an objective function called weighted sum minimal user rate (WSMR) is used. As shown in [@KIM], maximizing the minimal user rate in a single cell could provide the maximum fairness among the users in this cell. Therefore, in cooperative multi-cell systems, by introducing different priorities to the involved cells, WSMR can be employed to provide fairness to multiple users as well as cells. The objective function in this problem is given by $$\label{object function} f(\mathbf{A,P}) = {\sum}_{m=1}^M \omega_m \cdot \min_{u \in U_m} R_{u,m}(\mathbf{A}_{u,m},\mathbf{P})$$ where $\omega_m \geq 0$ represents the weight assigned to cell $m$’s minimal user rate. In particular, increasing $\omega_m$ leads to a higher resource allocation priority assigned to the users in cell $m$. Based on (\[object function\]), the resource allocation problem is formulated as $$\begin{aligned} \label{original optimization problem} \max_{\mathbf{A,P}}&\quad f(\mathbf{A,P}) = {\sum}_{m=1}^M \omega_m \cdot \min_{u \in U_m} R_{u,m}(\mathbf{A}_{u,m},\mathbf{P}) \nonumber \\ \textrm{s.t}& \quad {\sum}_{n=1}^N P_{m,n} \leq P_{m,max} \ \ \ \forall{m}, \nonumber \\ & \quad {\sum}_{u \in U_m}A_{u,m,n} \leq 1 \ \ \ \forall{m,n},\nonumber \\ & \quad A_{u,m,n} \in \{0,1\} \quad\quad\forall{u, m, n}.\end{aligned}$$ Note that the minimization operation in (\[original optimization problem\]) prohibits the objective function from being differentiable. Thus, an auxiliary optimization variable $R_m$ is introduced, and the optimization problem (\[original optimization problem\]) is reformulated as follows $$\begin{aligned} \label{equivalent optimization problem} \max_{\mathbf{A,P},R_m}&\ \ \ {\sum}_{m=1}^M \omega_m \cdot R_m \nonumber \\ \textrm{s.t}&\quad {\sum}_{n=1}^N A_{u,m,n} \cdot R_{u,m,n}(\mathbf{p}_n) \geq R_m \ \ \forall{m,u \in U_m}\nonumber \\ &\quad {\sum}_{n=1}^N P_{m,n} \leq P_{m,max} \nonumber \\ & \quad {\sum}_{u \in U_m}A_{u,m,n} \leq 1 \ \ \forall{m,n}\nonumber \\ &\quad A_{u,m,n} \in \{0,1\} \ \ \forall{m,n,u \in U_m}.\end{aligned}$$ Obviously, the optimization problem (\[equivalent optimization problem\]) is a constrained nonlinear optimization program with both integer and continuous variables. Furthermore, the problem is not convex and thus generally speaking it is difficult to directly solve the problem. Following a similar logic as those in [@Wong1999; @Rhee2000], an effective algorithm is proposed in the following, which optimizes subcarrier scheduling and power allocation alternatively. The Proposed Algorithm {#sect:Algorithm } ====================== Proposed Decomposition Power Allocation Algorithm ------------------------------------------------- Without loss of generality, the power allocation optimization is carried out first at each iteration. The subcarrier allocation variables **A** can be directly removed from (\[equivalent optimization problem\]) as they are assumed to be already computed. For notational simplicity, all superscript of variables are omitted in following description. Thus, the power allocation optimization subproblem can be expressed as $$\begin{aligned} \label{power allocation optimization problem} \max_{\mathbf{P},R_m} \ & \ \ \ \ {\sum}_{m=1}^M \omega_m \cdot R_m \nonumber \\ \textrm{s.t} & \ \ \ \ {\sum}_{{\tiny{n:A_{:,:,n=1}}}}{R_{u,m,n}(\mathbf{p}_n)} \ge R_m \nonumber \\ & \ \ \ \ {\sum}_{n=1}^N P_{m,n} \leq P_{m,max}.\end{aligned}$$ Based on the definition of $R_{u,m,n}(\mathbf{p}_n)$ in (\[user rate without A\]), it is obvious that (\[power allocation optimization problem\]) is a nonconvex optimization problem. For multi-cell OFDMA networks, the optimization problem (\[power allocation optimization problem\]) usually has high dimensions. Although centralized optimization methods can be used to solve the problem, they are impractical due to the high complexity. Therefore, instead of centralized optimization algorithms, we focus on distributed power allocation schemes using decomposition techniques. Various decomposition algorithms have been developed. The frequently used decomposition algorithm in wireless communications is Lagrangian relaxation procedure [@Bazaraa1993; @Kelley1960], and its variant relaxation techniques based on augmented Lagrangian functions [@Carpentier1996; @Cohen1978]. However, Lagrangian procedures may present drawbacks in some cases, such as difficulties to converge to an optimal solution for the global system (in the absence of convexity assumptions), uncontrollable convergence rates that depend on the correct choice of the values for several parameters which are difficult to update, and the requirement of the intervention of a central agent to update this complicated information. To overcome these drawbacks, a novel decomposition algorithm which is based on the decomposition of the optimality conditions of the global problem (\[power allocation optimization problem\]) is proposed in this paper [@Bazaraa1993]. The proposed decomposition algorithm improves both the computational efficiency and implementation simplification compared to previously mentioned algorithms. In the following, the decomposition methodology is presented in detail. First of all, the optimization problem (\[power allocation optimization problem\]) can be written in a compact form as follows for convenience $$\begin{aligned} \label{compact form of power subproblem} \max_{\boldsymbol{x}_m}&\quad {\sum}_{m=1}^M f_m(\boldsymbol{x}_m) \nonumber \\ \textrm{s.t}&\quad \boldsymbol{h}(\boldsymbol{x}_1,\dots,\boldsymbol{x}_M) \leq \boldsymbol{0} \quad \boldsymbol{g}_m(\boldsymbol{x}_m) \leq \boldsymbol{0}\end{aligned}$$ where $\boldsymbol{x}_m$ is a vector for cell $m$, that contains power allocation vector $\mathbf{p}_m$ and the minimum user rates $R_m$ within cell $m$. The first constraint in (\[compact form of power subproblem\]) is known as complicating constraint, which represents the rate constraint in (6). These equations contain variables and parameters from different cells and prevent each system from operating independently to each other. On the contrary, the second constraint is the power constraint for each BS and thus only related to one single cell. Therefore, in order to decompose the global optimization problem (\[compact form of power subproblem\]) into local subproblems for each cell, those equations of rate constraints are removed from (\[compact form of power subproblem\]). Thus, the problem is equivalent to $$\begin{aligned} \label{compact form of power subproblem-decomposed} \max_{\boldsymbol{x}_m}&\quad {\sum}_{m=1}^M f_m(\boldsymbol{x}_m) + {\sum}_{l\neq m} \boldsymbol{\lambda}_m^T \boldsymbol{h}_m(\boldsymbol{x}_1,\dots,\boldsymbol{x}_M) \nonumber \\ \textrm{s.t}& \quad \boldsymbol{h}_m(\boldsymbol{x}_1,\dots,\boldsymbol{x}_M) \leq \boldsymbol{0} \quad \boldsymbol{g}_m(\boldsymbol{x}_m) \leq \boldsymbol{0}\end{aligned}$$ where the first constraint has been separated into different cells compared with that in (\[compact form of power subproblem\]). The dual variable vector corresponding to the first constraint is denoted by $\boldsymbol{\lambda}_m$. Fixing the values of all variables and multipliers (indicated by over line) except those in cell m, (\[compact form of power subproblem-decomposed\]) reduces to $$\begin{aligned} \label{reduced form of power subproblem-decomposed} \max_{\boldsymbol{x}_m}&\quad k + f_m(\boldsymbol{x}_m)+{\sum}_{l\neq m}\bar{\boldsymbol{\lambda}}_l^{\rm{T}} \boldsymbol{h}_l({\bf{\bar x}}^m)\nonumber \\ \textrm{s.t}& \quad {\bf{\bar x}}^m =[\bar{\boldsymbol{x}}_1,\dots,\bar{\boldsymbol{x}}_{m-1}, \boldsymbol{x}_m,\bar{\boldsymbol{x}}_{m+1},\dots,\bar{\boldsymbol{x}}_{M}] \nonumber \\ & \quad \boldsymbol{h}_m({\bf{\bar x}}^m) \leq \boldsymbol{0} \quad \boldsymbol{g}_m(\boldsymbol{x}_m) \leq \boldsymbol{0}\end{aligned}$$ where $k= \sum_{l=1,l\neq m}^M f_l(\bar{\boldsymbol{x}}_l)$ is a constant. Similarly, the reduced problem (\[reduced form of power subproblem-decomposed\]) can be reduced for every cell $(m=1,\dots,M)$. To prove that the proposed decomposition method is based on the solutions of these reduced cell subproblems, we revisit the first-order optimality conditions of the problem (\[compact form of power subproblem\]) which is written as $$\begin{aligned} \label{first-order optimality conditions} &\nabla_{\boldsymbol{x}_m} f_m(\boldsymbol{x}_m^*)+ {\sum}_{m=1}^M \nabla_{\boldsymbol{x}_m}^{\rm{T}} \boldsymbol{h}_m(\boldsymbol{x}_1^* ,\dots,\boldsymbol{x}_M^*)\boldsymbol{\lambda}_m^* \nonumber \\ & +\nabla_{\boldsymbol{x}_m}^{\rm{T}}\boldsymbol{g}_m(\boldsymbol{x}_m^*)\boldsymbol{\mu}_m^* = \mathbf{0} \nonumber \\ & \boldsymbol{h}_m(\boldsymbol{x}_1^*,\dots,\boldsymbol{x}_M^*) \leq \mathbf{0},\quad \quad \boldsymbol{h}_m(\boldsymbol{x}_m^*)^{\rm{T}} \boldsymbol{\lambda}_m^* = 0\nonumber \\ & \boldsymbol{\lambda}_m^* \geq \mathbf{0}, \quad \quad \boldsymbol{g}_m(\boldsymbol{x}_m^*) \leq \mathbf{0},\quad \boldsymbol{g}_m(\boldsymbol{x}_m^*)^{\rm{T}} \boldsymbol{\mu}_m^* = 0\nonumber \\ & \boldsymbol{\mu}_m^* \geq \mathbf{0} \quad \quad m=1,\dots,M\end{aligned}$$ These conditions have been constructed using the optimal values $\boldsymbol{x}_m^*$, $\boldsymbol{\lambda}_m^*$ and $\boldsymbol{\mu}_m^*$ which are assumed to be known. The values $\boldsymbol{\lambda}_m^*$ and the values $\boldsymbol{\mu}_m^*$ are the optimal Lagrange multipliers associated with the two constraints in (\[compact form of power subproblem\]), respectively. If the first-order optimality conditions of every cell reduced subproblem (\[reduced form of power subproblem-decomposed\]) $(m=1,\dots,M)$ are stuck together, it can be observed that they are identical to the first-order optimality conditions (\[first-order optimality conditions\]) of the global problem (\[compact form of power subproblem\]). Thus, it is obvious that the decomposition mythology holds. Based on the previously proposed decomposition methodology, removing constants and substituting the objection function and constraints of (\[power allocation optimization problem\]) into (\[reduced form of power subproblem-decomposed\]), the original power allocation problem can be decomposed into $M$ subproblems, which are computed in each cell independently. The $m^{\rm{th}}$ subproblem is formulated as $$\begin{aligned} \label{power allocation subproblem} \max_{\mathbf{P}_m,R_m}&\quad \omega_m R_m+{\sum}_{l\neq m} \bar{\lambda}_{u,l}(\bar{R}_l-{\sum}_{n:A_{:,:,n}=1}R_{u,l,n}(\mathbf{\bar p}^m)) \nonumber \\ \textrm{s.t} \ \ & \ \ \ {\sum}_{n:A_{:,:,n}=1}R_{u,m,n}(\mathbf{\bar p}^m) \geq R_m \ \ \forall{u \in U_m} \nonumber \\ &\quad {\sum}_{n=1}^N P_{m,n} \leq P_{m,max} \quad\quad \forall{m}.\end{aligned}$$ where $\mathbf{p}^m = (\bar{\mathbf{p}}_1,\cdots,\bar{\mathbf{p}}_{m-1},\mathbf{p}_m,\bar{\mathbf{p}}_{m+1},\cdots,\bar{\mathbf{p}}_{M})$, and $\lambda_{u,m}$ is the optimal Lagrange multiplier which guarantees that user rate is larger than minimal value $R_m$ in each cell. A summary of the proposed decomposition algorithm for power allocation optimization is as follows: **Return** $\mathbf{P}^{t}$ The search directions, $\Delta \mathbf{p}_m, \Delta R_m, \Delta \lambda_{u,m}$, for subproblem (\[power allocation subproblem\]) can be computed independently of each other, allowing a parallel and distributed implementation. This step requires a central agent to coordinate the process, which receives certain information ($\bar{\mathbf{p}}_m,\bar{R}_m, \bar{\lambda}_{u,m}$ after each iteration) from all cells and returns it to the appropriate cells. It can be noted that the information exchanged between the areas and the central agent is little. Unlike other decomposition algorithm, in the proposed algorithm the central agent only distributes information and checks the convergence condition. It does not need to update any information, because this information is updated by the areas of the system, implying a simpler process. The main difference between the Lagrangian relaxation algorithm and the proposed decomposition one is that Lagrangian relaxation adds all the complicating constraints into the objective function. Therefore it needs auxiliary procedures to update the Lagrange multipliers. On the contrary, the proposed technique does not need any procedure to update the multipliers because this updating is automatic and results directly from the foreign optimization problem. For example, the Lagrangian multipliers $\lambda_l$ $(l\neq m)$ included in the objective of cell $m$ are obtained from cell $l$ ’s optimization by keeping its own complicating constraints. What’s more, the proposed approach has the advantage that convergence properties do not require an optimal solution of the subproblems at each iteration of the algorithm. It is enough to perform a single iteration for each subproblem, and then to update variable values. As a consequence, computation times can be significantly reduced with respect to other methods that require the computation of the optimum for the subproblems in order to attain convergence. Subcarrier Allocation Optimization ---------------------------------- When the power allocation is fixed, the remaining optimization problem is to find the optimal subcarrier allocation. It can be directly decomposed into $M$ subproblem, each of which corresponds to problem of maximizing the minimal user rate in cell $m$. The $m^{th}$ is a Mixed Integer Linear Problem (MILP) problem involving both integer and continuous variables, it can be solved by various algorithms such as exhaustive search, implicit enumeration method and branch-and-bound algorithm[@Stephen; @Boyed]. Simulation Results {#sec:simulation} ================== In this section, the performance of the proposed distributed resource allocation algorithm is investigated. The downlink of a cellular OFDMA system is considered with three coordinated cells and 32 subcarriers. The radius of each cell is 40m and there are two users randomly located in each cell. To focus on the performance of proposed algorithm, all weighting coefficients assigned to each cell are assumed to be identical and keep constant. In addition, we set $\Gamma = 0$dB, $\sigma^{2}_{u,m,n}= -60$ dBw and $\Psi_1=\Psi_2=\Psi_3=10^{-1}$. 500 independent channel realizations are carried out to obtain the final results. Fig.\[WSMR\] compares the average WSMR performance of the proposed algorithm and the one using Lagrangian relaxation method (LR) as a benchmark power allocation algorithm. The average WSMR at initialization which use uniform power allocation and even subcarrier allocation is also shown as a reference. It can be seen that the performance of the proposed distributed resource allocation algorithm is better than that of the algorithm using traditional Lagrangian decomposition method. Fig.\[convergence\] demonstrates the convergence performance of different algorithm for power allocation. Unlike the slow and oscillating behavior of LR procedure, proposed algorithm converges rapidly and stably. It can be concluded that the proposed decomposition algorithm are more effective than the traditional Lagrangian based algorithms. Conclusions {#sec:conclusion} =========== In this letter, we proposed a distributed algorithm for joint resource allocation in a coordinated multi-cell OFDMA network. The performance criterion termed as WSMR has been introduced to guarantee the fairness among multiple users and cells. Targeting at maximizing WSMR, an iterative algorithm was proposed to solve the joint optimization of power allocation and subcarrier scheduling. At each iteration, the power allocation is updated by applying a modified decomposition methodology which has low complexity and fast convergence speed. On the other hand, the subcarrier allocation is updated by solving a MILP. It is shown that the proposed distributed resource allocation algorithm achieve a better WSMR performance compared compared with the traditional decomposition algorithms. At the same time the proposed power allocation scheme also provides a better convergence performance than the Lagrangian algorithm. [99]{} L. Venturino, N. Prasad, and X. Wang, “Coordinated scheduling and power allocation in downlink multicell OFDMA networks,” *IEEE Trans. Vehicular Technology.*, vol. 58, no. 6, pp. 2835-2848, Jul. 2009. P. Soldati and M. Johansson, “Dynamic resource allocation in OFDMA multi-cellular systems,” *Technical Report TRITA-EE 2008:048.*, School of Electrical Engineering (KTH), Aug. 2008. Z. Han, Z. Ji, and K. J. R. Liu, “Power minimization for multi-cell OFDM networks using distributed noncooperative game approach,” in *Proc. IEEE Globecom*, vol. 6, pp. 3742-3747, 2004. B. Da and R. Zhang, “Cooperative Interference Control for Spectrum Sharing in OFDMA Cellular Systems,” in *Proc. IEEE Int. Conf. Communications*, pp. 1-5, Jun. 2011. I. Kim, I.-S. Park, and Y. Lee, “Use of linear programming for dynamic subcarrier and bit allocation in multiuser OFDM,” *IEEE Trans. Veh. Technol.*, vol. 55, no. 4, pp. 1195-1207, Jul. 2006. C. Y. Wong, R. S. Chen, K. Ben Letaief, and R. D. Murch, “Multiuser OFDM with adaptive subcarrier, bit, and power allocation,” *IEEE J.Select. Areas Commun.*, vol. 17, no. 10, pp. 1747-1758, Oct. 1999. W. Rhee and J. M. Cioffi, “Increase in capacity of multiuser OFDM sys-tem using dynamic subchannel allocation,” in *Proc. Vehicular Technology Conference*, 2000, vol. 2, pp. 1085-1089. M.S. Bazaraa, H.O. Sherali and C.M. Shetty, *Nonlinear Programming: Theory and Algorithms,2nd ed.* Wiley, New York, 1993. J.E. Kelley, “The cutting-plane method for solving convex programs,” *J. Soc. Industr. Appl. Math.*, vol. 8, pp. 703-712, 1960. P. Carpentier, G. Cohen, J.C. Culioli and A. Renaud, “Stochastic optimization of unit commitment: A new decomposition framework,” *IEEE Trans. Power Systems.*, vol. 11, no. 2 pp. 1067-1073, 1996. G. Cohen, “Optimization by decomposition and coordination: A unified approach,” *IEEE Trans. Automat.Control.*, vol. 23, no. 2, pp. 222-232, 1978. S. Boyd and L. Vandenberghe, *Convex Optimization.* Cambridge,U.K.: Cambridge Univ. Press, 2004. ![Average WSMR of proposed algorithm and the algorithm whose power allocation use Lagrangian relaxation method.[]{data-label="WSMR"}](WSMR.eps){width=".3\textwidth"} ![Convergence performances of proposed decomposition algorithm and the Lagrangian relaxation method for power allocation.[]{data-label="convergence"}](convergence.eps){width=".3\textwidth"} [^1]: Z. Fei, S. Li, J. Kuang and C. Xing are with RCDCT, Modern Communication Lab, Dept. of E. E., Beijing Institute of Technology, Beijing, China (email: {feizesong, surelee, jmkuang}@bit.edu.cn; chengwenxing@ieee.org) Y. Zhou is with Wireless Communication Research Center, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China (email: zhouyiqing@ict.ac.cn)

--- abstract: 'We apply a method to filter relevant information from the correlation coefficient matrix by extracting a network of relevant interactions. This method succeeds to generate networks with the same hierarchical structure of the Minimum Spanning Tree but containing a larger amount of links resulting in a richer network topology allowing loops and cliques. In Tumminello et al. [@TumminielloPNAS05], we have shown that this method, applied to a financial portfolio of $100$ stocks in the USA equity markets, is pretty efficient in filtering relevant information about the clustering of the system and its hierarchical structure both on the whole system and within each cluster. In particular, we have found that triangular loops and 4 element cliques have important and significant relations with the market structure and properties. Here we apply this filtering procedure to the analysis of correlation in two different kind of interest rate time series (16 Eurodollars and 34 US interest rates).' author: - 'T. Aste, T. Di Matteo Department of Applied Mathematics, Research School of Physical Sciences, The Australian National University, 0200 Canberra, ACT, Australia. M. Tumminello, R. N. Mantegna INFM Unità di Palermo and Dipartimento di Fisica e Tecnologie Relative, Università di Palermo, Viale delle Scienze, Palermo, I-90128, Italy.\' title: Correlation filtering in financial time series --- Introduction ============ The collective behavior of a system comprised of many elements is well described by the matrix of correlation coefficient among the elements. The main difficulty in the analysis of such correlations arises from the fact that the information in the whole correlation matrix is often huge, containing correlation coefficients for all the pairs of elements in the system. The challenge is to extract the smallest sub-set of such correlations which is able to describe accurately the collective behavior of the whole system. From a topological perspective, the correlation matrix can be represented by the complete graph (each node connected with all the other nodes) with edges weighted by the value of the correlation coefficient. The problem of extracting a system of meaningful interactions is now translated into the problem of reducing the complete graph into a sub-graph which keeps only the edges which best describe the system of interactions. Such a sub-graph must contain the maximum amount of information about the system’s collective behavior while keeping the simplest possible structure. It was shown in Ref. [@Mantegna99] that a method to investigate correlations in financial systems consists in extracting a minimal set of relevant interactions associated with the strongest correlations belonging to the Minimum Spanning Tree (MST). The MST maintains only the minimum number of links necessary to connect the network ($n-1$ links for a network with $n$ nodes) and, by construction, it keeps the links associated with the strongest correlations. This method turns out to be very effective in revealing the hierarchical structure contained in the correlation matrix. However, the reduction to a minimal skeleton of links is necessarily very drastic in filtering correlation based networks loosing therefore valuable information. The necessity of a less drastic filtering procedure has already been raised in the literature. For example, an extension from trees to more general graphs generated by selecting the most correlated links, up to a given threshold, was proposed by Onnela et al. [@Onnela03]. However, this method depends on the choice of the threshold and, when the number of kept links is constrained to be of the same order of the number of nodes (as for the MST), this method tends to generate disconnected networks made of several isolated, highly connected, sub graphs. On the contrary, we seek for a connected network with no isolated nodes as in the case of the MST. Since the MST method has been proved to be very effective in extracting hierarchies from the correlation matrix, its extension to a richer graph must be constructed in a way to maintain such a hierarchical skeleton while including a larger amount of links. A solution to this problem was recently proposed by introducing a new technique which enables to produce networks with tunable information content [@Noi; @TumminielloPNAS05]. This method is based on the idea of connecting iteratively the most correlated nodes while constraining the resulting network to be embedded on a given surface. We here apply this filtering procedure to the correlation matrix of *16 Eurodollar Interest Rates* [@DiMatteo] and *$34$ different kinds of interest rates in money and capital markets* [@DiMatteoMant; @DiMatteoAless04]. We investigate the data clustering and differentiations seeking for basic sub-structures and exploring their hierarchical gathering and growth. The paper is organized as following: in Section \[Embedding\], we present the general idea regarding the embedding of the complete graph on hyperbolic surfaces. In Section \[Corrnet\], we discuss the case of planar embeddings by considering the Planar Maximally Filtered Graph (PMFG) introduced in Ref. [@Noi; @TumminielloPNAS05] and we compare it with the MST. Section \[filtering\] shows results for different sets of interest rates and in Section \[clique\] the emergent clique structures are analyzed and discussed. Finally the conclusions are given in Section \[conclusion\]. Embedding networks on hyperbolic surfaces {#Embedding} ========================================= Any orientable surface can be topologically classified in terms of its [*genus*]{} which is the largest number of non-intersecting simple closed cuts that can be made on the surface without disconnecting a portion. The genus $g$ is a good measure of complexity for a surface: under such a classification, the sphere ($g=0$) is the simplest surface; the torus is the second-simpler ($g=1$); etc. To a given network a genus can always be assigned. It is defined to be equal to the minimum number of handles that must be added to the plane to *embed* the graph without edge-crossings. Consider a system of $n$ interacting elements which are collectively fluctuating in a stochastic way. In first place, these mutually-related stochastic variables can be considered all connected to each other in an $n$-th order complete graph ($K_n$). A *weight* can associated to each link and a natural choice for the weight of a link between node $i$ and node $j$ is the correlation coefficient $c_{i,j}$. The challenge is to locally simplify the network by keeping only the most significant interactions (largest correlations) and simultaneously extracting global information about the hierarchical organization of the whole system. Hereafter we show that a very powerful way to proceed is to map the complete graph into a significant sub network on a 2-dimensional (2D) hyperbolic surface. This approach has several attractive features: 1) it provides new measures to characterize complexity; 2) it gives a locally-planar representation; 3) it provides a hierarchical ensemble classification; 4) it allows the application of topologically invariant elementary moves [@p16; @p17; @AsteSherr; @Noi]. In addition, let us stress that *any* network can be embedded on a surface, therefore this approach is completely general and the genus of the embedding surface acts as a constraint on the complexity of the network. Indeed, Ringel and Youngs have shown that an embedding of $K_n$ is always possible in an orientable surface $S_g$ of genus [@Ringel1968] $$\label{g*} g \ge g^* = \lceil{ \frac{(n-3)(n-4)}{12} }\rceil$$ (for $n \ge 3$ and with $\lceil x \rceil$ the ceiling function which returns the smallest integer number $\ge x$). Therefore, providing a sufficiently high genus $g \ge g^*$, we can always generate embeddings that contain the whole information present in the complete graph. Such a reduction to 2D cannot be in general implemented simply on the plane. Indeed, the complexity of the surface will increase with the complexity of the graph and multi-handled hyperbolic surfaces must be used (for $n > 7$ one has $g^* >1$, and hyperbolic surfaces are needed to embed the complete graph). The embedding of the complete graph on $S_g$ is locally planar but this local simplification has been achieved at the expenses of an high complexity in the global surface ($g^*$ scales with $n^2$). A reduction to simpler surfaces with lower genus is therefore necessary. This can be done only by removing some of the ${n(n-1)}/{2}$ links of $K_n$ transforming the complete graph into a less connected network. For instance, the opposite extreme to $K_n$ is the spanning tree which is the connected graph with minimum number of links ($n-1$) and it can be embedded on a surface of genus $g=0$ (the sphere). In the next section we discuss an algorithm [@TumminielloPNAS05] which allows to generate networks embedded on surfaces with arbitrary genus and we discuss in details the case $g=0$. From correlations to networks {#Corrnet} ============================= Starting from a correlation coefficient matrix, we construct a graph G embedded on a surface $S_g$ of genus $g$ by means of the following procedure: from a set of disconnected elements, by following an ordered list of pair of elements sorted in decreasing order of the correlation coefficient between two elements, we connect two elements if and only if the resulting network can still be embedded on $S_g$. Elsewhere the two elements are left disconnected and the subsequent pair in the list is considered. It is important to stress out that this construction is analogous to the procedure which generates the MST. Indeed, the only difference between the two procedures is that in the case of MST a link is inserted if and only if the network after such connection is a forest or a tree. It has been proved in Tumminello et al. [@TumminielloPNAS05] that *at any step of construction of the MST and graph G of genus $g$ if two elements are connected via at least one path in one of the considered graphs then they are connected also in the other one*. This fact implies that the MST is always contained in G. Moreover, this also implies that the formation of connected clusters of nodes during the construction of a graph G coincides with the formation of the same clusters in the MST at the same stage of the construction. In other words the hierarchical structure associated to G coincides with the one of the MST *at any stage of the construction*. The graph G of genus $g$ is the most connected graph for a given embedding $S_g$ and number of nodes $n$. The graph G is a triangulation of $S_g$. It has $3(n-2+2g)$ (for $g \leq g^*$) links and the addition of a further link is impossible without edge-crossings or the increase of the surface genus. A special case is the PMFG [@TumminielloPNAS05] which corresponds to $g=0$, i.e. to a planar embedding [@Planar]. This is the simplest embedding and coincides with the one for the MST. In this respect, the MST and the PMFG are the two extreme cases having the same topological complexity but having the minimum ($n-1$) and the maximum ($3n -6$) number of links respectively. The main structural difference between the PMFG and the MST is that the PMFG allows the existence of loops and cliques. A clique of $r$ elements ($r$-clique) is a complete subgraph that links all $r$ elements. Only cliques of 3 and 4 elements are allowed in the PMFG. Indeed, topological constraints expressed by the Kuratowski’s theorem [@Planar] do not allow cliques with a number of elements larger than 4 in a planar graph. Larger cliques can only be present in graphs with genus $g>0$ and the larger the value of $g$ the larger is the number of elements $r$ of the maximal allowed clique (specifically $r \le \frac{7 + \sqrt{1+48 g}}{2}$) [@Ringel]. Correlation filtering in financial data {#filtering} ======================================= In the previous section we have introduced a *general* method for constructing a network of genus $g$ from the correlation matrix. Here we constraint ourselves to the case $g=0$, i.e. to the PMFG and we present two examples of PMFG graphs obtained by means of such correlation based procedure. The first example concerns a set of 16 Eurodollars Interest rates [@DiMatteo]. The second example concerns a set of 34 US interest rates [@DiMatteoMant]. PMFG for $16$ Eurodollar interest rates US stocks ------------------------------------------------- Let us here start with the analysis of the Eurodollar interest rates. For several economic reasons, interest rates have very similar statistical behaviors and follow similar trends in time. This makes the subject very challenging since one is no more dealing with the statistics of single objects but with the notion of a whole complex set of interacting elements which collectively fluctuate [@Pagan; @Rebonato; @Bouchaud; @DiMatteo; @DiMatteoMant; @DiMatteoScalas; @Nuyts]. Here we apply our filtering procedure to daily values in the time period $1990-1996$ for $16$ Eurodollars interest rates with maturity date between $3$ to $48$ months [@DiMatteo; @DiMatteoAless04]. These interest rates are very highly correlated with correlation coefficients values between $0.46$ and $0.98$ with $76\%$ of the coefficients larger than $0.8$. Fig. \[f.MSTEurod\] shows the MST associated with this set of data. The labels represent the maturity dates (in units of $3$ months) and the distances between the nodes are set to the corresponding metric distances $d_{i,j}=\sqrt{2(1-c_{i,j})}$. Fig. \[f.3DEurodollar\] shows a three dimensional representation of the PMFG network where the distance between nodes is $d_{i,j}$. Fig. \[f.3DEurodollar1\] reports a drawing of the PMFG network on the Euclidean plane. In this last figure, the thicker lines indicate links belonging to both the MST and the PMFG. This figure gives an immediate graphical proof that the PMFG network is planar ($g=0$): indeed it can be drawn on the plane without edge-crossings. In this PMFG we count $38$ cliques of 3 elements (triangles). $28$ of such triangles lie on the planar surface whereas the other $10$ are ‘collar rings’. The total number of possible combinations of $n$ elements in groups of three (i.e. the total number of cliques of $3$ elements in the complete graph) is $\binom{n}{3}=560$. This number is much larger than the number of $3$-cliques in PMFG ($38$). Similarly, the number of cliques of $4$ elements (tetrahedra) is $10$, a number which is much smaller than the number of cliques of $4$ elements present in the fully connected graph $\binom{n}{4} = 1820$. Indeed, the planar embedding reduces the interconnectivity of the network simplifying the resulting graph. In Table \[t.1\] all the cliques of $4$ elements are reported. In this table are also reported the average correlation coefficients $\left<c_{i,j}\right>$ inside each clique, the difference $\Delta$ between the maximum and minimum correlation coefficient and the standard deviation $\sigma$. PMFG for $34$ US interest rates ------------------------------- The second example regards the analysis of correlation coefficients among $34$ different weekly interest rates recorded during a time period of $16$ years between $1982$ and $1997$ and stored in the Statistical Release database of the Federal Reserve [@DiMatteoMant; @data]. For such a data-set we construct both the MST (Fig. \[f.MSTutTas\]) and the PMFG (Figs. \[f.3DTuttiTassi\] and \[f.3DTuttiTassi1\]). In this case we observe $94$ cliques of $3$ elements and $31$ cliques of $4$ elements. Note that, also in this case, such numbers are much smaller than the number of all possible cliques of $3$- and $4$-elements in $K_{34}$ which are respectively $\binom{n}{3} = 5984$ and $\binom{n}{4} = 46376$. The complete list of $4$ cliques together with $\left<c_{i,j}\right>$, $\Delta$ and $\sigma$ are listed in Table \[t.2\]. Labels $\left< c_{i,j} \right>$ $\Delta$ $\sigma$ ------------ ----------- ----------- ------------ -------------------------- ---------- ---------- [**12**]{} 14 15 [**16**]{} 0.977 0.017 0.006 9 10 11 13 0.972 0.016 0.006 [**8**]{} 9 10 11 0.969 0.021 0.008 7 [**8**]{} 9 11 0.965 0.021 0.007 6 7 [**8**]{} 9 0.964 0.026 0.009 5 6 7 9 0.959 0.029 0.012 [**4**]{} 5 6 7 0.958 0.044 0.016 3 [**4**]{} 5 7 0.94 0.06 0.02 2 3 [**4**]{} 5 0.91 0.13 0.05 1 2 3 [**4**]{} 0.83 0.30 0.11 : \[t.1\] The entire set of $4$-cliques in the Eurodollar interest rates. In bold are reported the labels corresponding to maturity dates of $1$, $2$, $3$ and $4$ years (labels $4$, $8$, $12$, and $16$ respectively.) Labels $\left< c_{i,j} \right>$ $\Delta$ $\sigma$ ---- ---- -------- ---- -------------------------- ---------- ---------- 22 28 21 20 0.967 0.045 0.016 21 28 20 19 0.97 0.06 0.02 20 21 19 18 0.97 0.05 0.02 9 12 5 4 0.962 0.032 0.013 18 27 17 16 0.96 0.07 0.03 19 20 18 17 0.96 0.06 0.02 26 27 16 15 0.96 0.07 0.03 10 13 9 5 0.957 0.040 0.018 12 13 9 5 0.955 0.040 0.017 30 31 13 5 0.951 0.044 0.017 19 27 18 17 0.95 0.09 0.03 13 30 12 5 0.949 0.044 0.016 25 26 15 14 0.93 0.10 0.05 13 16 10 5 0.93 0.11 0.05 26 27 16 10 0.93 0.11 0.04 17 27 16 13 0.93 0.13 0.05 16 27 13 10 0.93 0.11 0.05 11 12 9 4 0.93 0.10 0.05 7 12 5 4 0.92 0.10 0.04 11 29 4 3 0.92 0.07 0.03 28 32 22 21 0.92 0.11 0.06 6 11 4 3 0.92 0.09 0.03 25 26 16 15 0.92 0.17 0.06 12 29 11 4 0.91 0.10 0.04 7 8 5 4 0.90 0.10 0.04 24 31 13 5 0.87 0.21 0.10 32 33 28 21 0.87 0.17 0.06 21 28 19 2 0.84 0.26 0.13 23 24 13 5 0.80 0.30 0.12 33 34 32 21 0.76 0.23 0.10 4 6 3 1 0.7 0.5 0.2 : \[t.2\] The entire set of $4$-cliques in the $34$ US Interest rates. Clique structure {#clique} ================ The construction algorithm and the topological constraint of the PMFG force each element to participate to at least a clique of $3$ elements. The PMFG is a topological triangulation of the sphere. Therefore the [*triangular rings*]{} ($3$-cliques) are the elementary building blocks of this network and all the possible PMFG networks associated to the correlations between a set of $n$ elements can be generated by exploring the ensemble of planar triangulations with $n$ nodes. In a triangulation of the topological sphere the number of triangles on the surface is $2n -4$. However, in the two systems considered (Eurodollars and US interest rates) as well as in the $100$ stock portfolio examined in Ref. [@TumminielloPNAS05], the number of $3$-cliques is systematically larger than the number of triangles on the surface indicating therefore that there are several triangular rings which are collar rings and do not lie on the surface. By observing in details Figs. \[f.3DEurodollar\] and \[f.3DTuttiTassi\], we note that these internal rings belong to tetrahedra which pack together some 3-clique by sharing a triangular ring which then becomes a collar ring. The basic structures in these graphs are the $4$-cliques, which during the formation of the PMFG clusterize together locally at similar correlation values and then connect to each other by following the MST as skeleton structure. If such $4$ elements cliques are the ‘emerging building blocks’ of the PMFG, then there must be strong relations between their properties and the ones of the system from which they have been generated. One can verify that indeed the network for the $34$ US interest rates is entirely made by a set of tetrahedra packed together by sharing a triangular face. Such a packing leads to a structure made of $3 n-8$ cliques of $3$ elements and $n -3$ cliques of $4$ elements. This is exactly what we find in the case of the $34$ US interest rates. Interestingly, the planar network associated with the $16$ Eurodollar interest rates does not follow the above scheme having $38$ cliques of $3$ elements (instead of $3 n -8 = 40$) and $10$ cliques of $4$ elements (instead of $n-3 = 13$). By analyzing the network in detail one can see that such a discrepancy comes from the region where interest rates with large maturity dates of $30$-$48$ months gather together. However, it is clear from Fig. \[f.3DEurodollar\] that, apart from this region, the basic structure is also in this case identifiable with a tetrahedral packing. A further analysis of the clique structure for the case of the Eurodollar interest rates shows that these cliques gather together interest rates forming clusters which are characterized by restricted ranges of maturity dates. From Table \[t.1\], one can note that the most correlated cliques are the ones with largest maturity dates and the spread of the correlation coefficients inside each clique increases as the correlation decreases. A similar differentiation in term of maturity dates arises also from the analysis of the $34$ US interest rates. In ref. [@DiMatteoMant], it was shown that these data gather together in $6$ main clusters and three isolated elements. By comparing the structure of $4$-cliques in Table \[t.2\] with this cluster gathering we observe that there are $13$ cliques composed by elements all belonging to the same cluster, $11$ cliques with $3$ elements belonging to the same cluster, $5$ cliques which mix two elements for one cluster and two from another and finally $2$ cliques with elements belonging to three different clusters. Fig.\[f.US\_CLU\] shows schematically such a differentiation into different clusters.\ Conclusions {#conclusion} =========== The correlation coefficient matrix has been studied in different financial data by analyzing the structure of a network obtained by linking the most correlated elements while constraining the genus of the resulting graph. We find that in the case $g=0$ corresponding to the PMFG, the basic structure of such network is formed by packing together cliques of $4$ elements which share one or more triangular ring. The study of such $4$-cliques in the case of $16$ Eurodollar interest rates and $34$ interest rates reveals that the network hierarchy spontaneously generates clusters grouping together interest rates with similar maturity dates. This result confirms previous findings [@TumminielloPNAS05] which show such hierarchical gathering and differentiation in the case of $100$ US stocks. [99]{} M. Tumminello, T. Aste, T. Di Matteo and R. N. Mantegna, “A new tool for filtering information in complex systems, ” Submitted 2005 (cond-mat/0501335). R. N. Mantegna, “Hierarchical structure in financial markets,” [*Eur. Phys. J. B*]{} [**11**]{}, pp. 193-197, 1999. J.-P. Onnela, A. Chakraborti, K. Kaski, J. & Kertész, “Dynamic asset trees and portfolio analysis,” [*Eur. Phys. J. B*]{} [**30**]{}, pp. 285-288, 2002.\ J.-P. Onnela, A. Chakraborti, K. Kaski, J. Kertész, A. & Kanto, “Dynamics of market correlations: Taxonomy and portfolio analysis,” [*Phys. Rev. E*]{} [**68**]{}, pp. 056110-056122, 2003.\ J.-P. Onnela, A. Chakraborti, K. Kaski, J. & Kertész, “Dynamic asset trees and Black Monday,” [*Physica A*]{} [**324**]{}, pp. 247-252, 2003. T. Aste, T. Di Matteo, S. T. Hyde, “Complex networks on hyperbolic surfaces,” [*Physica A*]{} [**346**]{}, pp. 20-26, 2005. T. Di Matteo and T. Aste, “How does the Eurodollars interest rate behave?,” [*J. Theoret. Appl. Finance* ]{} [**5**]{}, pp. 122-127, 2002, (cond-mat/0101009). T. Di Matteo, T. Aste, R. N. Mantegna, “An interest rate cluster analysis,” [*Physica A* ]{} [**339**]{}, pp. 181-188, 2004, (cond-mat/0401443). T. Di Matteo, T. Aste, S. T. Hyde and S. Ramsden, “Interest rates hierarchical structure,” [*Physica A* ]{} to appear, 2005. H. M. Ohlenbusch, T. Aste, B. Dubertret and N. Rivier, “The topological structure of 2D disordered cellular systems,” [*Eur. Phys. J. B*]{} [**2**]{}, pp. 211-220, 1998. B. Dubertret, T. Aste, H. M. Ohlenbusch and N. Rivier, “Two-dimensional froths and the dynamics of biological tissues,” [*Phys. Rev. E*]{} [**58**]{}, pp. 6368-6378, 1998. T. Aste and D. Sherrington, “Glass transition in self organizing cellular patterns,” [*J. Phys. A*]{} [**32**]{}, pp. 7049-7056, 1999. G. Ringel and J. W. T. Youngs, “Solution of the Heawood Map-Coloring Problem,” [*Proc. Nat. Acad. Sci. USA*]{} [**60**]{}, pp. 438-445, 1968. D. B. West, in *An Introduction to Graph Theory*, Prentice Hall, pp. 247-281, 2001. G. Ringel, in *Map Color Theorem*, Springer, Berlin, Chapter 4, 1974. A. R. Pagan, A. D. Hall, and V. Martin, “Modeling the Term Structure,” in [*HandBook of Statistics*]{} [**14**]{}, G. S. Maddala and C. R. Rao, ed., pp.91-118, 1997. R. Rebonato, [*Interest-rate option models*]{}, John Wiley & Sons, New York, 1998. J. P. Bouchaud, N. Sagna, R. Cont, N. EL-Karoui and M. Potters, “Phenomenology of the interest rate curve,” [*Applied Mathematical Finance*]{} [**6**]{}, pp. 209-232, 1999 (cond-mat/9712164). T. Di Matteo, M. Airoldi, E. Scalas, “On pricing interest rate derivatives,” [*Physica A* ]{} [**339**]{}, pp. 189-196, 2004 (cond-mat/0401445). T. Alderweireld, J. Nuyts, “Detailed empirical study of the term structure of interest rates. Emergence [*Physica A*]{} [**331**]{}, pp. 602-616, 2004. http://www.federalreserve.gov/releases/h15/data.htm

--- abstract: 'We propose a factor-graph-based approach to joint channel-estimation-and-decoding (JCED) of bit-interleaved coded orthogonal frequency division multiplexing (BICM-OFDM). In contrast to existing designs, ours is capable of exploiting not only sparsity in sampled channel taps but also clustering among the large taps, behaviors which are known to manifest at larger communication bandwidths. In order to exploit these channel-tap structures, we adopt a two-state Gaussian mixture prior in conjunction with a Markov model on the hidden state. For loopy belief propagation, we exploit a “generalized approximate message passing” (GAMP) algorithm recently developed in the context of compressed sensing, and show that it can be successfully coupled with soft-input soft-output decoding, as well as hidden Markov inference, through the standard sum-product framework. For $N$ subcarriers and any channel length $L<N$, the resulting JCED-GAMP scheme has a computational complexity of only ${\ensuremath{\mathcal{O}}}(N\log_2 N + N |{\mathbb{S}}|)$, where $|{\mathbb{S}}|$ is the constellation size. Numerical experiments using IEEE 802.15.4a channels show that our scheme yields BER performance within 1 dB of the known-channel bound and 3-4 dB better than soft equalization based on LMMSE and LASSO.' author: - 'Philip Schniter$^*$ [^1] [^2]' bibliography: - 'macros\_abbrev.bib' - 'books.bib' - 'misc.bib' - 'comm.bib' - 'multicarrier.bib' - 'sparse.bib' - 'stc.bib' - 'underwater.bib' title: 'A Message-Passing Receiver for BICM-OFDM over Unknown Clustered-Sparse Channels' --- Introduction {#sec:intro} ============ When designing a digital communications receiver, it is common to model the effects of multipath propagation in discrete time using a convolutive linear channel that, in the slow-fading scenario, can be characterized by a fixed impulse response $\{x_j\}_{j=0}^{L-1}$ over the duration of one codeword. When the communication bandwidth is sufficiently low, the “taps” $\{x_j\}_{j=0}^{L-1}$ are well modeled as independent complex Gaussian random variables, resulting in the “uncorrelated Rayleigh-fading” and “uncorrelated Rician-fading” models that have dominated the wireless communications literature for many decades [@Molisch:Book:05]. For receiver design, the Gaussian tap assumption is very convenient because the optimal estimation scheme is well known to be linear [@Poor:Book:94]. As the communication bandwidth increases, however, the channel taps are no longer well-modeled as Gaussian nor independent. Rather, they tend to be heavy-tailed or “sparse” in that only a few values in $\{x_j\}_{j=0}^{L-1}$ have significant amplitude [@Cramer:TAP:02; @Preisig:JAcSA:04; @Molisch:TVT:05; @Czink:TWC:07]. Moreover, groups of large taps are often clustered together in lag $j$. These behaviors are both a blessing and a curse: a blessing because, of all tap distributions, the independent Gaussian one is most detrimental to capacity [@Medard:TIT:00], but a curse because optimal channel estimation becomes non-linear and thus receiver design becomes more complicated. Recently, there have been many attempts to apply breakthrough non-linear estimation techniques from the field of “compressive sensing” [@Mar:SPM:08] (e.g., LASSO [@Tibshirani:JRSSb:96; @Chen:JSC:98]) to the wireless channel estimation problem. We refer to this approach as “compressed channel sensing” (CCS), after the recent comprehensive overview [@Bajwa:PROC:10]. The CCS literature generally takes a *decoupled* approach to the problem of channel estimation and data decoding, in that pilot-symbol knowledge is first exploited for sparse-channel estimation, after which the channel estimate is used for data decoding. However, this decoupled approach is known to be suboptimal [@Kannu:arXiv:10]. The considerations above motivate a *joint* approach to structured-sparse-channel-estimation and decoding that offers both near-optimal decoding performance and low implementation complexity. In this paper, we propose exactly such a scheme. In particular, we focus on orthogonal frequency-division multiplexing (OFDM) with bit-interleaved coded modulation (BICM), and propose a novel factor-graph-based receiver that leverages recent results in “generalized approximate message passing” (GAMP) [@Rangan:10b], soft-input/soft-output (SISO) decoding [@MacKay:Book:03], and structured-sparse estimation [@Schniter:CISS:10]. Our receiver assumes a clustered-sparse channel-tap prior constructed using a two-state Gaussian mixture with a Markov model on the hidden tap state. The scheme that we propose has only ${\ensuremath{\mathcal{O}}}(N\log_2 N\!+\!N|{\mathbb{S}}|)$ complexity, where $N$ denotes the number of subcarriers and $|{\mathbb{S}}|$ denotes the constellation size, facilitating large values of $N$ and channel length $L\!<\!N$ (e.g., we use $N\!=\!1024$ and $L\!=\!256$ for our numerical results). For rich non-line-of-sight (NLOS) channels generated according to the IEEE 802.15.4a standard [@Molisch:802.15.4a], our numerical experiments show bit error rate (${\textsf{BER}}$) performance within $1$ dB of the known-channel bound and $3$–$4$ dB better than soft equalization based on LMMSE and LASSO. We now place our work in the context of existing factor-graph designs. Factor-graph based joint channel-estimation and decoding (JCED) was proposed more than a decade ago (see, e.g., the early overview [@Worthen:TIT:01]). To calculate the messages passed among the nodes of the factor graph, first instincts suggest to apply the standard “sum-product algorithm” (SPA) [@Pearl:Book:88; @Loeliger:Proc:07; @Kschischang:TIT:01]. Exact SPA on the JCED factor graph is computationally infeasible, however, and so it must be approximated. For this, there are many options, since many well-known iterative inference algorithms can themselves be recognized as SPA approximations, e.g., the expectation-maximization (EM) algorithm [@Dauwels:ISIT:05], particle filtering [@Dauwels:ISIT:06], variational (or “mean-field”) techniques [@Dauwels:ISIT:07], and even steepest descent [@Dauwels:ITW:05]. Moreover, because the JCED factor graph is loopy, even non-approximate SPA is not guaranteed to yield the correct output distributions, because exact inference is NP hard [@Cooper:AI:90]. It is perhaps not surprising that, amidst this uncertainty about exact SPA and its “best” approximation, a number of different factor-graph approaches to JCED over frequency-selective channels have been proposed (e.g., [@Novak:ICASSP:09; @Liu:PIMRC:09; @Knievel:ICC:10; @Kirkelund:GLOBE:10]). Our approach differs from existing factor-graph JCED designs in that it uses 1) a sparse (i.e., non-Gaussian) channel-tap prior, 2) a clustered (i.e., non-independent) channel-tap prior, and 3) a state-of-the-art SPA approximation known as “generalized approximate message passing” (GAMP), which has been shown to admit rigorous analysis as $N,L\!\rightarrow\!\infty$ [@Rangan:10b]. In fact, we conjecture that the success of our method is due in large part to the principled approximations used within GAMP. We also note that, although we focus on the case of clustered-sparse channels, our approach could be applied to non-sparse (i.e., Gaussian) or non-clustered (i.e., independent) channel-taps or, e.g., non-sparse channels with unknown length $L$ [@Novak:ICASSP:09], with minor modifications of our assumed channel prior. Finally, we mention that this work is an evolution of our earlier work [@Schniter:ASIL:10; @Schniter:PHYCOM:11] that was limited to an exactly sparse channel, that did not exploit clustering, and that was based on the “relaxed belief propagation” (RBP) algorithm [@Rangan:10v2], which has higher implementation complexity than GAMP. For example, the JCED scheme from [@Schniter:ASIL:10; @Schniter:PHYCOM:11] has complexity ${\ensuremath{\mathcal{O}}}(NL\!+\!N|{\mathbb{S}}|)$, which grows with the channel length $L$. Our paper is organized as follows. In [Section \[sec:model\]]{} we detail our assumptions on the OFDM system and the channel prior, and provide an illustrative example of clustered-sparse behavior with the IEEE 802.15.4a channel model. In [Section \[sec:jced\]]{} we detail our GAMP-based JCED approach, in [Section \[sec:sims\]]{} we report the results of our simulation study, and in [Section \[sec:conc\]]{} we conclude. Throughout the paper, we use the following notation. ${{\mathbb{R}}}$ denotes the field of reals and ${{\mathbb{C}}}$ the complex field. $(\cdot)^*$ denotes conjugate and $\operatorname{Re}(\cdot)$ extracts the real part. Furthermore, $\delta(\tau)$ denotes the Dirac delta waveform while $\{\delta_n\}_{n=-\infty}^\infty$ denotes the Kronecker delta sequence. Also, $\langle j\rangle_N$ denotes $j$-modulo-$N$, ${\star}$ convolution, and $\propto$ denotes equality up to a scaling. We use boldface capital letters like ${\ensuremath{\boldsymbol{B}}}$ to denote matrices and boldface small letters like ${\ensuremath{\boldsymbol{b}}}$ to denote vectors. ${\ensuremath{\boldsymbol{I}}}$ denotes the identity matrix, ${\ensuremath{\boldsymbol{1}}}$ denotes the vector of ones, and $\operatorname{{{\ensuremath{\mathcal{D}}}}}({\ensuremath{\boldsymbol{b}}})$ constructs a diagonal matrix from the vector ${\ensuremath{\boldsymbol{b}}}$. For matrices and vectors, $(\cdot){^\textsf{T}}$ denotes transpose and $(\cdot){^\textsf{H}}$ denotes conjugate transpose. When $x_j$ is a realization of random variable $X_j$, we write $x_j\!\sim\! X_j$ and use $\operatorname{E}_{X_j}\{x_j\}$ to denote the mean, $\operatorname{var}_{X_j}\{x_j\}$ the variance, $p_{X_j}(x_j)$ the pdf, and $p_{X_j|D_j}(x_j{\,|\,}d_j)$ the pdf conditioned on the event $D_j\!=\!d_j$. Sometimes we omit the subscript when there is no danger of confusion, yielding, e.g., $\operatorname{E}\{x_j\}$, $\operatorname{var}\{x_j\}$, $p(x_j)$ and $p(x_j{\,|\,}d_j)$. ${\ensuremath{\mathcal{CN}}}(x;\hat{x},\nu^x)\!{\triangleq}\!(\pi\nu^x)^{-1}\exp(-|x-\hat{x}|^2/\nu^x)$ denotes the circular Gaussian pdf with mean $\hat{x}$ and variance $\nu^x$. In fact, we often use $(\hat{v}_j,\nu_j^v)$ when referring to the mean and variance of $V_j$. For a random vector ${\ensuremath{\boldsymbol{x}}}$, we use $\operatorname{Cov}({\ensuremath{\boldsymbol{x}}})$ to denote the covariance matrix. System Model {#sec:model} ============ The BICM-OFDM model {#sec:OFDM} ------------------- We consider an OFDM system with $N$ subcarriers, each modulated by a QAM symbol from a $2^M$-ary unit-energy constellation ${\mathbb{S}}$. Of the $N$ subcarriers, ${N_\textsf{p}}$ are dedicated as pilots,[^3] and the remaining ${N_\textsf{d}}\!{\triangleq}\! N\!-\!{N_\textsf{p}}$ are used to transmit a total of ${M_\textsf{t}}$ training bits and ${M_\textsf{d}}\! {\triangleq}\! {N_\textsf{d}}M \!-\! {M_\textsf{t}}$ coded/interleaved data bits. The data bits are generated by encoding ${M_\textsf{i}}$ information bits using a rate-$R$ coder, interleaving them, and partitioning the resulting ${M_\textsf{c}}\!{\triangleq}\!{M_\textsf{i}}/R$ bits among an integer number $Q\!{\triangleq}\!{M_\textsf{c}}/{M_\textsf{d}}$ of OFDM symbols. We note that the resulting scheme has a spectral efficiency of $\eta\!{\triangleq}\! {M_\textsf{d}}R/N$ information bits per channel use (bpcu). In the sequel, we use $s{^{(k)}}\!\in\!{\mathbb{S}}$ for $k\!\in\!\{1,\dots,2^M\}$ to denote the $k^{th}$ element of the QAM constellation, and ${\ensuremath{\boldsymbol{c}}}{^{(k)}}\!{\triangleq}\! [c_{1}{^{(k)}},\dots,c_{M}{^{(k)}}]{^\textsf{T}}$ to denote the corresponding bits as defined by the symbol mapping. Likewise, we use $s_i[q]\!\in\!{\mathbb{S}}$ for the QAM symbol transmitted on the $i^{th}$ subcarrier of the $q^{th}$ OFDM symbol and ${\ensuremath{\boldsymbol{c}}}_i[q] \!{\triangleq}\! [c_{i,1}[q],\dots,c_{i,M}[q]]{^\textsf{T}}$ for the coded/interleaved bits corresponding to that symbol. We use ${\ensuremath{\boldsymbol{c}}}[q] \!{\triangleq}\! [{\ensuremath{\boldsymbol{c}}}_0[q],\dots,{\ensuremath{\boldsymbol{c}}}_{N-1}[q]]{^\textsf{T}}$ to denote the coded/interleaved bits in the $q^{th}$ OFDM symbol and ${\ensuremath{\boldsymbol{c}}} \!{\triangleq}\! [{\ensuremath{\boldsymbol{c}}}[1],\dots,{\ensuremath{\boldsymbol{c}}}[Q]]{^\textsf{T}}$ to denote the entire (interleaved) codeword. The elements of ${\ensuremath{\boldsymbol{c}}}$ that are apriori known as pilot or training bits will be referred to as ${\ensuremath{\boldsymbol{c}}}{_\textsf{pt}}$. The remainder of ${\ensuremath{\boldsymbol{c}}}$ is determined from the information bits ${\ensuremath{\boldsymbol{b}}}\!{\triangleq}\![b_1,\dots,b_{{M_\textsf{i}}}]{^\textsf{T}}$ by coding/interleaving. To modulate the $q^{th}$ OFDM symbol, an $N$-point inverse discrete Fourier transform (DFT) ${\ensuremath{\boldsymbol{\Phi}}}{^\textsf{H}}$ is applied to the QAM sequence ${\ensuremath{\boldsymbol{s}}}[q]\!=\![s_0[q],\dots,s_{N-1}[q]]{^\textsf{T}}$, yielding the time-domain sequence ${\ensuremath{\boldsymbol{\Phi}}}{^\textsf{H}}{\ensuremath{\boldsymbol{s}}}[q] \!=\! {\ensuremath{\boldsymbol{a}}}[q] \!=\! [a_0[q],\dots,a_{N-1}[q]]{^\textsf{T}}$. The OFDM waveform $a(t)$ is then constructed using $L$-cyclic-prefixed versions of $\{a_{j}[q]\}$ and the transmission pulse ${g_{\text{\sf t}}}(\tau)$: $$\begin{aligned} a(t) &=& \sum_{q=1}^Q \sum_{j=-L}^{N-1} a_{\langle j\rangle_N}\![q]\, {g_{\text{\sf t}}}\big(t-jT-q(N+L)T\big),\end{aligned}$$ with $T$ denoting the baud interval (in seconds) and $L<N$. The waveform $a(t)$ propagates through a noisy channel with an impulse response $h(\tau)$ that is supported on the interval $[\tau_{\min},\tau_{\max}]$, resulting in the receiver input waveform $$\begin{aligned} r(t) &=& w(t) + \int_{\tau_{\min}}^{\tau_{\max}} h(\tau) a(t-\tau) d\tau,\end{aligned}$$ where $w(t)$ is a Gaussian noise process with flat power spectral density $N_o$. We note that a time-invariant channel is assumed for simplicity. The receiver samples $r(t)$ through the reception pulse ${g_{\text{\sf r}}}(\tau)$, obtaining $$\begin{aligned} r_j[q] &=& \int r(t) \, {g_{\text{\sf r}}}\big(jT +q(N+L)T-t\big) dt, $$ and applies an $N$-DFT ${\ensuremath{\boldsymbol{\Phi}}}$ to each time-domain sequence ${\ensuremath{\boldsymbol{r}}}[q]\!=\![r_0[q],\dots,r_{N-1}[q]]{^\textsf{T}}$, yielding the frequency-domain sequences ${\ensuremath{\boldsymbol{\Phi}}}{\ensuremath{\boldsymbol{r}}}[q] \!=\! {\ensuremath{\boldsymbol{y}}}[q] \!=\! [y_0[q],\dots,y_{N-1}[q]]{^\textsf{T}}$ for $q=1\dots Q$. Defining the pulse-shaped channel response $x(\tau) \!{\triangleq}\! ({g_{\text{\sf r}}}{\star}h{\star}{g_{\text{\sf t}}})(\tau)$, it is well known (e.g., [@Cimini:TCOM:85]) that, when the support of $x(\tau)$ is contained within the interval $[0,LT)$, the frequency domain observation on the $i^{th}$ subcarrier can be written as $$\begin{aligned} y_i[q] &=& s_i[q] z_i[q] + w_i[q] , \label{eq:yi}\end{aligned}$$ where $z_i[q]\in{{\mathbb{C}}}$ is the $i^{th}$ subcarrier’s gain and $\{w_i[q]\}$ are Gaussian noise samples. Furthermore, defining the uniformly sampled channel “taps” $x_j[q]\!{\triangleq}\!x(jT\!+\!q(N\!+\!L)T)$, the subcarrier gains are related to these taps through the DFT: $$\begin{aligned} z_i[q] &=& \sum_{j=0}^{L-1} \Phi_{ij} x_j[q] . \label{eq:zi}\end{aligned}$$ In addition, when $({g_{\text{\sf r}}}{\star}{g_{\text{\sf t}}})(\tau)$ is a Nyquist pulse, $\{w_i[q]\}_{\forall i,q}$ are statistically independent with variance $\nu^w\!=\!N_o$. To simplify the development, we assume that $Q=1$ in the sequel (but not in the simulations), and drop the index $[q]$ for brevity. A clustered-sparse tap prior {#sec:GM2} ---------------------------- Empirical studies [@Cramer:TAP:02; @Preisig:JAcSA:04; @Molisch:TVT:05; @Czink:TWC:07] have suggested that, when the baud rate $T^{-1}$ is sufficiently large, the channel taps $\{x_j\}$ are “sparse” in that the tap distributions tend to be heavy tailed. The same empirical studies suggest that large taps tend to be clustered in the lag domain. Furthermore, both the sparsity and clustering behaviors can be lag-dependent, such as when the receiver’s timing-synchronization mechanism aligns the first strong multipath arrivals with a particular reference lag $j$. A concrete example of these behaviors will be given in [Section \[sec:IEEE\]]{}. Since our message-passing-based receiver design is inherently Bayesian, we seek a prior on the taps $\{x_j\}$ that is capable of representing this lag-dependent clustered sparsity. For this purpose, we assume a two-state Gaussian mixture (GM2) prior,[^4] $$\begin{aligned} p(x_j) &=& (1-\lambda_j) {\ensuremath{\mathcal{CN}}}(x_j;0,\nu^0_j) + \lambda_j {\ensuremath{\mathcal{CN}}}(x_j;0,\nu^1_j) , \label{eq:pxj}\end{aligned}$$ where $\nu_j^0\!\geq\! 0$ denotes the variance while in the “small” state, $\nu_j^1\!>\!\nu_j^0$ denotes the variance while in the “big” state, and $\lambda_j\!{\triangleq}\! \Pr\{d_j\!=\!1\}$ denotes the prior probability of $x_j$ being in the “big” state. Here, we use $d_j\!\in\!\{0,1\}$ to denote the hidden state, implying the state-conditional pdf $ p(x_j{\,|\,}d_j) = {\ensuremath{\mathcal{CN}}}(x_j;0,\nu^{d_j}_j). $ For example, if $x_j$ was presumed to be a “sparse” tap, then we would choose $\lambda_j\!\ll\! 1$ and $\nu_j^1 \!\gg\! \nu_j^0$ in [(\[eq:pxj\])]{}. If, on the other hand, $x_j$ is presumed to be (non-sparse) Rayleigh-fading, we would choose $\lambda_j\!=\!1$ and set $\nu_j^1$ equal to the tap variance, noting that $\nu_j^0$ becomes inconsequential. If $x_j$ is presumed to be Nakagami-fading or similar, we could fit the GM2 parameters $[\lambda_j,\nu_j^0,\nu_j^1]$ appropriately using the EM algorithm, as described in [@Bishop:Book:07 p. 435]. The GM2 prior has been used successfully in many other non-Gaussian inference problems (see, e.g., [@Ishwaran:AS:05]), and our premise here is that the GM2 model achieves a good balance between fidelity and tractability when modeling channel taps as well. To capture the big-tap clustering behavior, we employ a hidden Markov model (HMM). For this, we model the tap states $\{d_j\}_{j=0}^{L-1}$ as a Markov chain (MC) with switching probabilities $p_j^{01}\!{\triangleq}\!\Pr\{d_{j+1}\!=\!0{\,|\,}d_{j}\!=\!1\}$ and $p_j^{10}\!{\triangleq}\!\Pr\{d_{j+1}\!=\!1{\,|\,}d_{j}\!=\!0\}$. Here, $p_j^{01}<0.5$ implies that the neighbors of a big $x_j$ tend to be big, and $p_j^{10}<0.5$ implies that the neighbors of a small $x_j$ tend to be small. We note that $\{p_j^{01},p_j^{10}\}_{j=0}^{L-1}$ must be consistent with $\{\lambda_j\}_{j=0}^{L-1}$ in that the following must hold for all $j$: $$\begin{aligned} {\ensuremath{\begin{bmatrix}\lambda_{j+1}~&~1\!-\!\lambda_{j+1}\end{bmatrix}}} &=& {\ensuremath{\begin{bmatrix}\lambda_{j}~&~1\!-\!\lambda_{j}\end{bmatrix}}} {\ensuremath{\begin{bmatrix}1\!-\!p_j^{01} & p_j^{01}\\ p_j^{10} & 1\!-\!p_j^{10}\end{bmatrix}}}.\end{aligned}$$ Although we allow correlation among the tap states, we assume that the tap *amplitudes* are conditionally independent, i.e., $p(x_{j+1},x_j{\,|\,}d_{j+1},d_j) \!=\! p(x_j{\,|\,}d_j) p(x_{j+1}{\,|\,}d_{j+1})$. Our experiences with IEEE 802.15.4a channels (see below) suggest that this is a valid assumption. We emphasize that the model parameters $\{\lambda_j,p_j^{01},p_j^{01},\nu_j^1,\nu_j^0\}$ are allowed to vary with lag $j$, facilitating the exploitation of apriori known lag-dependencies in sparsity and/or clustering. An illustrative example: IEEE 802.15.4a channels {#sec:IEEE} ------------------------------------------------ As an illustrative example of the clustered-sparse tap behavior described above, we generated realizations of the tap vector ${\ensuremath{\boldsymbol{x}}}{\triangleq}[x_0,\dots,x_{L-1}]{^\textsf{T}}$ from channel impulse responses $h(\tau)$ generated according to the method specified in the IEEE 802.15.4a “ultra-wideband” standard [@Molisch:802.15.4a], which uses the Saleh-Valenzuela model [@Saleh:JSAC:87] $$\begin{aligned} h(\tau) &=& \sum_{c=0}^C\sum_{k=0}^K h_{k,c} e^{j\phi_{k,c}} \delta(\tau-T_c-\tau_{k,c}),\end{aligned}$$ where $C$ denotes the number of clusters, $T_c$ the delay of the $c^{th}$ cluster, $K$ the number of components per cluster, $\{\tau_{k,c}\}$ the relative component delays, $\{h_{k,c}\}$ the component amplitudes, and $\{\phi_{k,c}\}$ the component phases. In particular, the 802.15.4a standard specifies the following. - The cluster arrival times are a Poisson process with rate $\Lambda$, i.e., $p(T_c{\,|\,}T_{c-1})\!=\! \Lambda \exp(-\Lambda(T_c-T_{c-1}))$. The initial cluster delay $T_0\!\geq\! \tau_{\min}$, as seen by the receiver, is a function of the timing synchronization algorithm. - The component arrivals are a mixture of two Poisson processes: $p(\tau_{k,c} |\tau_{k-1,c}) \!=\! \beta\lambda_1 \exp(-\lambda_1(\tau_{k,c}-\tau_{k-1,c})) \!+\! (1-\beta)\lambda_2 \exp(-\lambda_2(\tau_{k,c}-\tau_{k-1,c}))$ with $\tau_{0,c}=0$. - The component energies obey $$\begin{aligned} \operatorname{E}\{|h_{k,c}|^2\} &=& \frac{\exp(-T_l/\Gamma-\tau_{k,l}/\gamma)} {\gamma [(1-\beta)\lambda_1 + \beta\lambda_2 + 1]} ,\end{aligned}$$ where $\Gamma$ is the cluster decay time constant and $\gamma$ is the intra-cluster decay time constant. - The amplitudes $\{h_{k,c}\}$ are i.i.d Nakagami with $m$-factors randomly generated via i.i.d $m\sim{\ensuremath{\mathcal{N}}}(m_0,\hat{m}^2_0)$. - The phases $\{\phi_{k,c}\}$ are i.i.d uniform on $[0,2\pi)$. - The number of clusters, $C$, is Poisson distributed with mean $\bar{C}$, i.e., $p(C) \!=\! (\bar{C})^{C} \exp(-\bar{C}) / (C!)$. - The number of components per cluster, $K$, is set large enough to yield a desired modeling accuracy. Beyond the above specifications, we assume the following. - The parameters $\{\Lambda,\lambda_1,\lambda_2,\beta,\Gamma,\gamma,m_0,\hat{m}_0,\bar{C}\}$ are set according to the 802.15.4a “outdoor NLOS” scenario [@Molisch:802.15.4a]. - $K\!=\!100$ components per cluster are used. - The pulses ${g_{\text{\sf t}}}(\tau)$ and ${g_{\text{\sf r}}}(\tau)$ are square-root raised cosine (SRRC) designs with parameter $0.5$. - The system bandwidth equals $T^{-1}=256$ MHz. - The number of taps (and CP length) was set at $L=256$ (implying a maximal delay spread of $1\,\nu$sec) in order to capture all significant energy in $h(\tau)$. - The initial delay was generated via $T_0\!=\! {L_{\text{\sf pre}}}T+ \tilde{T}_0$, where ${L_{\text{\sf pre}}}\!=\!20$ and where $\tilde{T}_0$ is exponentially distributed with mean $T$, i.e., $p(\tilde{T}_0) \!=\! \Lambda_0 \exp(-\Lambda_0 \tilde{T}_0)$ for $\Lambda_0 \!=\! 1/T$. Here, ${L_{\text{\sf pre}}}$ was chosen so that $\{x_j\}_{j=0}^{{L_{\text{\sf pre}}}}$ captures the “pre-cursor” energy contributed by the pulse shape, while $\Lambda_0$ models a positive synchronization uncertainty. We now show results from an experiment conducted using $U=10000$ realizations of the tap vector ${\ensuremath{\boldsymbol{x}}}$. In [Fig. \[fig:hist\]]{}, we show histograms of $\operatorname{Re}(x_j)$ for lags $j\!\in\!\{5,23,128,230\}$. There it can be seen that the empirical distribution of $\operatorname{Re}(x_j)$ changes significantly with lag $j$: for pre-cursor lags $j\!<\!{L_{\text{\sf pre}}}$, it is approximately Gaussian; for near-cursor lags $j\!\approx\! {L_{\text{\sf pre}}}$, it is approximately Laplacian; and, for post-cursor lags $j\!\gg\! {L_{\text{\sf pre}}}$, it is extremely heavy-tailed. In [Fig. \[fig:realization\]]{}, we show a typical realization of ${\ensuremath{\boldsymbol{x}}}$ and notice clustering among the big taps. For comparison, we also plot an empirical estimate of the power-delay profile (PDP) ${\ensuremath{\boldsymbol{\rho}}}{\triangleq}[\rho_0,\dots,\rho_{L-1}]{^\textsf{T}}$ in [Fig. \[fig:realization\]]{}, where $\rho_j \!{\triangleq}\! \operatorname{E}\{|x_j|^2\}$. Next, we fit the GM2 parameters $\{\lambda_j,\nu^0_j,\nu^1_j\}_{j=0}^{L-1}$ from the realizations $\{{\ensuremath{\boldsymbol{x}}}_u\}_{u=1}^U$ using the EM algorithm [@Bishop:Book:07 p. 435], which iterates the steps [(\[eq:EM\_post\])]{}-[(\[eq:EM\_lam\])]{} until convergence: $$\begin{aligned} \omega_{j,u} &=& \textstyle \frac{\lambda_j {\ensuremath{\mathcal{CN}}}(x_{j,u}; 0,\nu_j^1)} {(1-\lambda_j) {\ensuremath{\mathcal{CN}}}(x_{j,u}; 0,\nu_j^0) + \lambda_j {\ensuremath{\mathcal{CN}}}(x_{j,u}; 0,\nu_j^1)} ~\forall j,u \label{eq:EM_post}\\ \nu^1_j &=& \textstyle \sum_{u=1}^U \omega_{j,u} |x_{j,u}|^2/ \sum_{u=1}^U \omega_{j,u} ~\forall j\\ \nu^0_j &=& \textstyle \sum_{u=1}^U (1-\omega_{j,u}) |x_{j,u}|^2/ \sum_{u=1}^U (1-\omega_{j,u}) ~\forall j \quad\\ \lambda_j &=& \textstyle \frac{1}{U} \sum_{u=1}^U \omega_{j,u} ~\forall j. \label{eq:EM_lam}\end{aligned}$$ Above, $\omega_{j,u}$ is the posterior on the state $d_{j,u}$ of tap $x_{j,u}$, i.e., $\omega_{j,u}=\Pr\{d_{j,u}\!=\!1{\,|\,}x_{j,u}; \lambda_j,\nu^0_j,\nu^1_j\}$. The EM-estimated big-variance profile ${\ensuremath{\boldsymbol{\nu}}}^1{\triangleq}[\nu^1_0,\dots,\nu^1_{L-1}]{^\textsf{T}}$ and small-variance profile ${\ensuremath{\boldsymbol{\nu}}}^0$ are shown in [Fig. \[fig:realization\]]{}, while the sparsity profile ${\ensuremath{\boldsymbol{\lambda}}}{\triangleq}[\lambda_0,\dots,\lambda_{L-1}]{^\textsf{T}}$ is shown in [Fig. \[fig:gm2\]]{}. Not surprisingly, the best-fit GM2 parameters also change significantly with lag $j$. In particular, as $j$ becomes larger, the variance ratio $\nu^1_j/\nu^0_j$ increases while the big-tap-probability $\lambda_j$ decreases, corresponding to an increase in sparsity. Meanwhile, there exists a peak in $\lambda_j$ near $j\!=\!{L_{\text{\sf pre}}}$ that results from synchronization. Next, we empirically estimated the switching probabilities ${\ensuremath{\boldsymbol{p}}}^{01}{\triangleq}[p^{01}_0,\dots,p^{01}_{L-1}]{^\textsf{T}}$ and ${\ensuremath{\boldsymbol{p}}}^{10}$ using maximum a-posteriori (MAP) state estimates, i.e., $\hat{d}_{j,u} = \lfloor \omega_{j,u} + 0.5\rfloor$. In particular, $$\begin{aligned} p_j^{01} &=& \textstyle \sum_{u=1}^U 1_{\{\hat{d}_{j+1,u}=0 \,\&\, \hat{d}_{j,u}=1\}} / \sum_{u=1}^U 1_{\{\hat{d}_{j,u}=1\}} \\ p_j^{10} &=& \textstyle \sum_{u=1}^U 1_{\{\hat{d}_{j+1,u}=1 \,\&\, \hat{d}_{j,u}=0\}} / \sum_{u=1}^U 1_{\{\hat{d}_{j,u}=0\}} ,\end{aligned}$$ where $1_{\{A\}}$ denotes the indicator function for event $A$. From the plots in [Fig. \[fig:gm2\]]{}, we see that the estimated switching probabilities are lag-dependent as well. Finally, using the MAP state estimates $\{\hat{d}_{j,u}\}$, we empirically estimated the normalized conditional correlation $$\textstyle \frac{\sum_{u=1}^U 1_{\{\hat{d}_{j+1,u}=1,\hat{d}_{j,u}=1\}} x_{j+1,u} x_{j,u}^* } {\sqrt{\sum_{u=1}^U 1_{\{\hat{d}_{j+1,u}=1,\hat{d}_{j,u}=1\}} |x_{j+1,u}|^2 \sum_{u=1}^U 1_{\{\hat{d}_{j+1,u}=1,\hat{d}_{j,u}=1\}} |x_{j,u}|^2}}$$ and found that the magnitudes were $<\!0.1$, validating our assumption of conditionally independent tap amplitudes. In summary, we see that IEEE 802.15.4a channels do indeed yield taps with the lag-dependent clustered sparsity described in [Section \[sec:GM2\]]{}. Moreover, we have shown how the GM2-HMM parameters can be estimated from realizations of ${\ensuremath{\boldsymbol{x}}}$. Next, we propose an efficient factor-graph based approach to joint channel-estimation and decoding (JCED) for BICM-OFDM using the GM2-HMM prior proposed in [Section \[sec:GM2\]]{}. Joint Channel Estimation and Decoding {#sec:jced} ===================================== Our goal is to infer the information bits ${\ensuremath{\boldsymbol{b}}}$ from the OFDM observations ${\ensuremath{\boldsymbol{y}}}$ and the pilot/training bits ${\ensuremath{\boldsymbol{c}}}{_\textsf{pt}}$, without knowing the channel state ${\ensuremath{\boldsymbol{x}}}$. In particular, we aim to maximize the posterior pmf $p(b_m{\,|\,}{\ensuremath{\boldsymbol{y}}},{\ensuremath{\boldsymbol{c}}}{_\textsf{pt}})$ of each info bit. To exploit prior knowledge that ${\ensuremath{\boldsymbol{x}}}$ is clustered-sparse, we employ the GM2-HMM prior described in [Section \[sec:GM2\]]{}. As a result, the info-bit posterior can be decomposed into the following product of factors: $$\begin{aligned} \lefteqn{ p(b_m{\,|\,}{\ensuremath{\boldsymbol{y}}},{\ensuremath{\boldsymbol{c}}}{_\textsf{pt}}) \,=\, \sum_{{\ensuremath{\boldsymbol{b}}}_{-m}} p({\ensuremath{\boldsymbol{b}}}{\,|\,}{\ensuremath{\boldsymbol{y}}},{\ensuremath{\boldsymbol{c}}}{_\textsf{pt}}) \,\propto\, \sum_{{\ensuremath{\boldsymbol{b}}}_{-m}} p({\ensuremath{\boldsymbol{y}}}{\,|\,}{\ensuremath{\boldsymbol{b}}},{\ensuremath{\boldsymbol{c}}}{_\textsf{pt}}) p({\ensuremath{\boldsymbol{b}}}) } \quad \label{eq:propto} \\ &=& \int_{{\ensuremath{\boldsymbol{x}}}} \sum_{{\ensuremath{\boldsymbol{s}}},{\ensuremath{\boldsymbol{d}}},{\ensuremath{\boldsymbol{c}}},{\ensuremath{\boldsymbol{b}}}_{-m}} \hspace{-4mm} p({\ensuremath{\boldsymbol{y}}}{\,|\,}{\ensuremath{\boldsymbol{s}}}, {\ensuremath{\boldsymbol{x}}}) p({\ensuremath{\boldsymbol{x}}}{\,|\,}{\ensuremath{\boldsymbol{d}}}) p({\ensuremath{\boldsymbol{d}}}) p({\ensuremath{\boldsymbol{s}}}{\,|\,}{\ensuremath{\boldsymbol{c}}}) p({\ensuremath{\boldsymbol{c}}}{\,|\,}{\ensuremath{\boldsymbol{b}}},{\ensuremath{\boldsymbol{c}}}{_\textsf{pt}}) p({\ensuremath{\boldsymbol{b}}}) \nonumber\\ &=& \int_{{\ensuremath{\boldsymbol{x}}}} \sum_{{\ensuremath{\boldsymbol{d}}}} \prod_{j=0}^{L-1} p(x_j{\,|\,}d_j) p(d_j{\,|\,}d_{j-1}) \sum_{{\ensuremath{\boldsymbol{s}}}} \prod_{i=0}^{N-1} p(y_i{\,|\,}s_i, {\ensuremath{\boldsymbol{x}}}) \nonumber\\&&\mbox{}\times \sum_{{\ensuremath{\boldsymbol{c}}}} p(s_{i}{\,|\,}{\ensuremath{\boldsymbol{c}}}_{i}) \sum_{{\ensuremath{\boldsymbol{b}}}_{-m}} p({\ensuremath{\boldsymbol{c}}}{\,|\,}{\ensuremath{\boldsymbol{b}}},{\ensuremath{\boldsymbol{c}}}{_\textsf{pt}}) \prod_{m=1}^{{M_\textsf{i}}} p(b_{m}),\quad \label{eq:factored}\end{aligned}$$ where ${\ensuremath{\boldsymbol{b}}}_{-m}\!{\triangleq}\![b_1,\dots,b_{m-1},b_{m+1},\dots,b_{{M_\textsf{i}}}]{^\textsf{T}}$. This factorization is illustrated by the *factor graph* in [Fig. \[fig:factor\_graph\_noncoh\_clust\]]{}, where the round nodes represent random variables and the square nodes represent the factors of the posterior exposed in [(\[eq:factored\])]{}. Background on belief propagation {#sec:bp} -------------------------------- Although exact evaluation of the posteriors $\{p(b_m{\,|\,}{\ensuremath{\boldsymbol{y}}},{\ensuremath{\boldsymbol{c}}}{_\textsf{pt}})\}$ is computationally impractical for the problem sizes of interest, these posteriors can be approximately evaluated using *belief propagation* (BP) [@Frey:ANIPS:98] on the factor graph in [Fig. \[fig:factor\_graph\_noncoh\_clust\]]{}. In textbook BP, beliefs take the form of pdfs/pmfs that are propagated among nodes of the factor graph via the *sum/product algorithm* (SPA) [@Pearl:Book:88; @Loeliger:Proc:07; @Kschischang:TIT:01]: 1. Say the factor node $f$ is connected to the variable nodes $\{v_a\}_{a=1}^A$. The belief passed from $f$ to $v_b$ is $p_{f\rightarrow v_b}(v_b) \propto \int_{\{v_a\}_{a\neq b}} f(v_1,\dots,v_A) \prod_{a\neq b} p_{v_a \rightarrow f}(v_a)$, given the beliefs $\{p_{v_a\rightarrow f}(\cdot)\}_{a\neq b}$ recently passed to $f$. 2. Say the variable node $v$ is connected to the factor nodes $\{f_1,\dots,f_B\}$. The belief passed from $v$ to $f_a$ is $p_{v\rightarrow f_a}(v) \propto \prod_{b\neq a} p_{f_b\rightarrow v}(v)$, given the beliefs $\{p_{f_b\rightarrow v}(\cdot)\}_{b\neq a}$ recently passed to $v$. 3. Say the variable node $v$ is connected to the factor nodes $\{f_1,\dots,f_B\}$. The posterior on $v$ is the product of all recently arriving beliefs, i.e., $p(v) \propto \prod_{b=1}^B p_{f_b\rightarrow v}(v)$. When the factor graph contains no loops, SPA-BP yields exact posteriors after two rounds of message passing (i.e., forward and backward). But, in the presence of loops, convergence to the exact posteriors is not guaranteed [@Cooper:AI:90]. That said, there exist many problems to which loopy BP [@Frey:ANIPS:98] has been successfully applied, including inference on Markov random fields [@Freeman:IJCV:00], LDPC decoding [@MacKay:Book:03], and compressed sensing [@Baron:TSP:10; @Donoho:PNAS:09; @Bayati:10; @Rangan:10b; @Schniter:CISS:10; @Rangan:10v2]. Our work not only leverages these past successes, but unites them. Background on GAMP {#sec:gamp} ------------------ An important sub-problem within our larger bit-inference problem is the estimation of a vector of independent possibly-non-Gaussian variables ${\ensuremath{\boldsymbol{x}}}$ that are linearly mixed via ${\ensuremath{\boldsymbol{\Phi}}}\in{{\mathbb{C}}}^{N\times L}$ to form ${\ensuremath{\boldsymbol{z}}}\!=\!{\ensuremath{\boldsymbol{\Phi x}}}\!=\![z_0,\dots,z_{N-1}]{^\textsf{T}}$, and subsequently observed as noisy measurements ${\ensuremath{\boldsymbol{y}}}$ through the possibly non-Gaussian pdfs $\{p_{Y_i|Z_i}(.{\,|\,}.)\}_{i=0}^{N-1}$. In our case, [(\[eq:pxj\])]{} specifies a GM2 prior on $x_j$ and [(\[eq:yi\])]{}—given the finite-alphabet uncertainty in $s_i$—yields the non-Gaussian measurement pdf $p_{Y_i|Z_i}$. This “linear mixing” sub-problem is described by the factor graph shown within the middle dashed box in [Fig. \[fig:factor\_graph\_noncoh\_clust\]]{}, where each node “$y_i$” represents the measurement pdf $p_{Y_i|Z_i}$ and the node rightward of each node “$x_j$” represents the GM2 prior on $x_j$. Building on recent work on multiuser detection by Guo and Wang [@Guo:ISIT:07], as well as recent work on message passing algorithms for compressed sensing by Donoho, Maleki, Montanari, and Bayati [@Donoho:PNAS:09; @Bayati:10], Rangan proposed a so-called *generalized approximate message passing* (GAMP) scheme that, for the sub-problem described above, admits rigorous analysis[^5] as $N,L\!\rightarrow\!\infty$ [@Rangan:10b]. The main ideas behind GAMP are the following. First, although the beliefs flowing leftward from the nodes $\{x_j\}$ are clearly non-Gaussian, the corresponding belief about $z_i = \sum_{j=0}^{L-1}\Phi_{ij} x_j$ can be accurately approximated as Gaussian, when $L$ is large, using the central limit theorem. Moreover, to calculate the parameters of this distribution (i.e., its mean and variance), only the mean and variance of each $x_j$ are needed. Thus, it suffices to pass only means and variances leftward from each $x_j$ node. It is similarly desirable to pass only means and variances rightward from each measurement node. Although the exact rightward flowing beliefs would be non-Gaussian (due to the non-Gaussian assumption on the measurement channels $p_{Y_i|Z_i}$), GAMP approximates them as Gaussian using a 2nd-order Taylor series, and passes only the resulting means and variances. A further simplification employed by GAMP is to approximate the *differences* among the outgoing means/variances of each left node, and the incoming means/variances of each right node, using Taylor series. The GAMP algorithm[^6] is summarized in [Table \[tab:gamp\]]{}. Joint estimation and decoding using GAMP {#sec:jced_gamp} ---------------------------------------- We now detail our application of GAMP to joint channel-estimation and decoding (JCED) under the GM2-HMM tap prior, frequently referring to the factor graph in [Fig. \[fig:factor\_graph\_noncoh\_clust\]]{}. Because our factor graph is loopy, there exists considerable freedom in the message passing schedule. Roughly speaking, we choose to pass messages from the left to the right of [Fig. \[fig:factor\_graph\_noncoh\_clust\]]{} and back again, several times, stopping as soon as the messages converge. Each of these full cycles of message passing will be referred to as a “turbo iteration.” However, during a single turbo iteration, there may be multiple iterations of message passing *between* the GAMP and MC sub-graphs, which will be referred to as “equalizer” iterations. Furthermore, during a single equalizer iteration, there may be multiple iterations of message passing *within* the GAMP sub-graph, while there is at most one forward-backward iteration *within* the MC sub-graph. Finally, the SISO decoding block may itself be implemented using message passing, in which case it may also use several internal iterations. The message passing details are discussed below. At the start of the first turbo iteration, there is total uncertainty about the information bits, so that $\Pr\{b_m\!=\!1\}\!=\!\frac{1}{2}~\forall m$. Thus, the initial bit beliefs flowing rightward out of the coding/interleaving block are uniformly distributed. Meanwhile, the pilot/training bits are known with certainty. Coded-bit beliefs are then propagated rightward into the symbol mapping nodes. Since the symbol mapping is deterministic, the corresponding pdf factors take the form $p(s{^{(k)}}{\,|\,}{\ensuremath{\boldsymbol{c}}}{^{(l)}}) = \delta_{k-l}$. The SPA dictates that the message passed rightward from symbol mapping node “${\ensuremath{\mathcal{M}}}_i$” takes the form $$\begin{aligned} p_{{\ensuremath{\mathcal{M}}}_i\rightarrow s_i}(s{^{(k)}}) &\propto& \sum_{{\ensuremath{\boldsymbol{c}}}\in\{0,1\}^M} p(s{^{(k)}}|{\ensuremath{\boldsymbol{c}}}) \prod_{m=1}^M p_{c_{i,m}\rightarrow {\ensuremath{\mathcal{M}}}_i}(c_m) \quad \\ &=& \prod_{m=1}^M p_{c_{i,m}\rightarrow {\ensuremath{\mathcal{M}}}_i}(c_m{^{(k)}}) ,\end{aligned}$$ which is then copied forward as the message passed rightward from node $s_i$ (i.e., $p_{{\ensuremath{\mathcal{M}}}_i\rightarrow s_i}(s{^{(k)}})=p_{s_i\rightarrow y_i}(s{^{(k)}})$). Recall, from [Section \[sec:gamp\]]{}, that the symbol-belief passed rightward into the measurement node “$y_i$” determines the pdf $p_{Y_i|Z_i}$ used in GAMP. Writing this symbol belief as ${\ensuremath{\boldsymbol{\beta}}}_i{\triangleq}[\beta_i{^{(1)}},\dots,\beta_i{^{(|{\mathbb{S}}|)}}]{^\textsf{T}}$ for $\beta_i{^{(k)}}{\triangleq}p_{s_i\rightarrow y_i}(s{^{(k)}})$, equation [(\[eq:yi\])]{} implies the measurement pdf $$\begin{aligned} p_{Y_i|Z_i}(y|z) &=& \sum_{k=1}^{|{\mathbb{S}}|} \beta_i{^{(k)}} \,{\ensuremath{\mathcal{CN}}}(y;s{^{(k)}}z;\nu^w) . \label{eq:pY|Z}\end{aligned}$$ From [(\[eq:pY|Z\])]{}, it is shown in [Appendix \[app:out\]]{} that the quantities in (D2)-(D3) of [Table \[tab:gamp\]]{} become $$\begin{aligned} g{_{\textsf{out},i}}(y,\hat{z},\nu^z) &=& \frac{1}{\nu^z} \hat{e}_i(y,\hat{z},\nu^z) \label{eq:gout}\\ g'{_{\textsf{out},i}}(y,\hat{z},\nu^z) &=& \frac{1}{\nu^z}\Big(\frac{\nu^e_i(y,\hat{z},\nu^z)}{\nu^z}-1\Big) \label{eq:g'out}\end{aligned}$$ for $$\begin{aligned} \xi_i{^{(k)}}(y,\hat{z},\nu^z) &{\triangleq}& \frac{\beta_i{^{(k)}} {\ensuremath{\mathcal{CN}}}(y;s{^{(k)}}\hat{z},|s{^{(k)}}|^2\nu^z\!+\!\nu^w)} {\sum_{k'} \beta_i{^{(k')}} {\ensuremath{\mathcal{CN}}}(y;s{^{(k')}}\hat{z},|s{^{(k')}}|^2\nu^z\!+\!\nu^w)} \quad\label{eq:xi}\\ \zeta{^{(k)}}(\nu^z) &{\triangleq}& \frac{|s{^{(k)}}|^2\nu^z}{|s{^{(k)}}|^2\nu^z+\nu^w} \label{eq:zeta}\\ \hat{e}{^{(k)}}(y,\hat{z},\nu^z) &{\triangleq}& \Big(\frac{y}{s{^{(k)}}}-\hat{z}\Big) \zeta{^{(k)}}(\nu^z) \\ \hat{e}_i(y,\hat{z},\nu^z) &{\triangleq}& \sum_{k=1}^{|{\mathbb{S}}|} \xi_i{^{(k)}}\!(y,\hat{z},\nu^z) \,\hat{e}{^{(k)}}(y,\hat{z},\nu^z) \label{eq:ei} \\ \nu^{e}_i(y,\hat{z},\nu^z) &{\triangleq}& \sum_{k=1}^{|{\mathbb{S}}|} \xi_i{^{(k)}}\!(y,\hat{z},\nu^z) \,\bigg( |\hat{e}{^{(k)}}(y,\hat{z},\nu^z) - \hat{e}_i(y,\hat{z},\nu^z)|^2 + \frac{\nu^w \zeta{^{(k)}}(\nu^z)}{s{^{(k)}}} \bigg) \label{eq:muei} \end{aligned}$$ where ${\ensuremath{\boldsymbol{\xi}}}_i\!{\triangleq}\![\xi_i{^{(1)}},\dots,\xi_i{^{(|{\mathbb{S}}|)}}]{^\textsf{T}}$ characterizes the posterior pmf on $s_i$ under the channel model $z_i\sim{\ensuremath{\mathcal{CN}}}(\hat{z},\nu^z)$. Likewise, from [(\[eq:pxj\])]{}, it is shown in [Appendix \[app:in\]]{} that the quantities (D5)-(D6) take the form $$\begin{aligned} g{_{\textsf{in},j}}(\hat{r},\nu^r) &=& \Big( \alpha_j \,\gamma^1_j + \big(1-\alpha_j\big) \,\gamma^0_j \Big)\hat{r} \label{eq:gin}\\ g'{_{\textsf{in},j}}(\hat{r},\nu^r) &=& \alpha_j(1-\alpha_j)(\gamma^1_j-\gamma^0_j)^2 \,|\hat{r}|^2/\nu^r + \alpha_j\gamma^1_j + (1-\alpha_j)\gamma^0_j , \label{eq:g'in}\end{aligned}$$ for $\gamma^0_j(\nu^r) {\triangleq}(1+\nu^r/\nu_j^0)^{-1}$ and $\gamma^1_j(\nu^r) {\triangleq}(1+\nu^r/\nu_j^1)^{-1}$ and $$\begin{aligned} \alpha_j(\hat{r},\nu^r) &{\triangleq}& \frac{1}{1+\Bigg( \underbrace{ \frac{\lambda_j}{1-\lambda_j} }_{ \displaystyle {\ensuremath{\mathcal{L}}}{^\text{\sf apri}}_j } \underbrace{ \frac{{\ensuremath{\mathcal{CN}}}(\hat{r};0,\nu_j^1+\nu^r)} {{\ensuremath{\mathcal{CN}}}(\hat{r};0,\nu_j^0+\nu^r)} }_{ \displaystyle {\ensuremath{\mathcal{L}}}{^\text{\sf ext}}_j(\hat{r},\nu^r) } \Bigg)^{-1}} . \label{eq:alfj}\end{aligned}$$ Above, ${\ensuremath{\mathcal{L}}}_j{^\text{\sf apri}}$ is the apriori likelihood ratio $\frac{\Pr\{d_j=1\}}{\Pr\{d_j=0\}}$ on the hidden state, ${\ensuremath{\mathcal{L}}}_j{^\text{\sf ext}}(\hat{r},\nu^r)$ is GAMP’s extrinsic likelihood ratio, and $\alpha_j(\hat{r},\nu^r)$ is the corresponding posterior probability that $d_j=1$. Using [(\[eq:gout\])]{}-[(\[eq:alfj\])]{}, the GAMP algorithm in [Table \[tab:gamp\]]{} is iterated until it converges.[^7] In doing so, GAMP generates (a close approximation to) both the conditional means ${\ensuremath{\Hat{\boldsymbol{x}}}}$ and variances ${\ensuremath{\boldsymbol{\nu}}}^x\!{\triangleq}\![\nu_0^x,\dots,\nu_{L-1}^x]{^\textsf{T}}$ given the observations ${\ensuremath{\boldsymbol{y}}}$, the soft symbol priors ${\ensuremath{\boldsymbol{\beta}}}\!{\triangleq}\![{\ensuremath{\boldsymbol{\beta}}}_{0},\dots,{\ensuremath{\boldsymbol{\beta}}}_{L-1}]{^\textsf{T}}$ and the sparsity prior ${\ensuremath{\boldsymbol{\lambda}}}$. Conveniently, GAMP also returns (close approximations to) both the conditional means ${\ensuremath{\Hat{\boldsymbol{z}}}}$ and variances ${\ensuremath{\boldsymbol{\nu}}}^z$ of the subchannel gains ${\ensuremath{\boldsymbol{z}}}$, as well as posteriors ${\ensuremath{\boldsymbol{\xi}}}\!{\triangleq}\![{\ensuremath{\boldsymbol{\xi}}}_{0},\dots,{\ensuremath{\boldsymbol{\xi}}}_{L-1}]{^\textsf{T}}$ on the symbols ${\ensuremath{\boldsymbol{s}}}$. Before continuing, we discuss some GAMP details that are specific to our OFDM-JCED application. First, we notice that, to guarantee that the variance $\nu_i^u(n)$ in (R5) is positive, we must have $\nu^e_i\!<\!\nu_z$ in [(\[eq:g’out\])]{}. Since this is not necessarily the case during the first few GAMP iterations, we clip $\nu^e_i$ at the value $0.99\nu^z$, where $0.99$ was chosen heuristically. Second, due to unit-modulus property of the DFT elements $\Phi_{ij}$, step (R2) in [Table \[tab:gamp\]]{} simplifies to $\nu^z_i(n) \!=\! \sum_j \nu_j^x(n)$ and (R6) simplifies to $\nu_j^r(n) \!=\! \big(\sum_i \nu_i^u(n)\big)^{-1}$. With these simplifications, the complexity of GAMP is dominated by either the matrix-vector products $\sum_j \Phi_{ij}\hat{x}_{j}(n)$ in (R1) and $\sum_i \Phi_{ij}^*\hat{u}_{i}(n)$ in (R7), which can be implemented using a $N\log_2 N$-multiply FFT when $N$ is a power-of-two, or by the calculation of $\{\hat{e}_i,\nu_i^e\}_{i=0}^{N-1}$ in [(\[eq:ei\])]{}-[(\[eq:muei\])]{}, which requires ${\ensuremath{\mathcal{O}}}(N|{\mathbb{S}}|)$ multiplies. Thus, GAMP requires only ${\ensuremath{\mathcal{O}}}(N\log_2 N +N|{\mathbb{S}}|)$ multiplies per iteration. After the messages within the GAMP sub-graph have converged, tap-state beliefs are passed rightward to the MC sub-graph. In particular, the SPA dictates that GAMP passes tap-state likelihoods or, equivalently, the extrinsic likelihood ratios ${\ensuremath{\mathcal{L}}}_j{^\text{\sf ext}}$. Since the MC sub-graph is non-loopy, only one iteration of forward-backward message passing is performed,[^8] after which the resulting tap-state likelihoods are passed leftward back to GAMP, where they are treated as tap-state priors ${\ensuremath{\boldsymbol{\lambda}}}$ in the next equalizer iteration. This interaction between the GAMP and MC sub-blocks can be recognized as an incarnation of the structured-sparse reconstruction scheme recently proposed by the authors in [@Schniter:CISS:10]. When the tap-state likelihoods passed between GAMP and MC have converged,[^9] the equalizer iterations are terminated and messages are passed leftward from the GAMP block. For this, SPA dictates that a symbol-belief propagates leftward from the $y_i$ node with the form $$\begin{aligned} p_{s_i\leftarrow y_i}(s) &\propto& \int_{z} {\ensuremath{\mathcal{CN}}}(y_i;s z,\nu^w) \,{\ensuremath{\mathcal{CN}}}(z;\hat{z}_i,\nu_i^z) \\ &=& {\ensuremath{\mathcal{CN}}}(y_i;s \hat{z}_i,|s|^2\nu_i^z+\nu^w) , \label{eq:p_y_to_s}\end{aligned}$$ where $(\hat{z}_i,\nu_i^z)$ play the role of soft channel estimates. The SPA then implies that $p_{{\ensuremath{\mathcal{M}}}_i\leftarrow s_i}(s)=p_{s_i\leftarrow y_i}(s)$. Next, beliefs are passed leftward from each symbol-mapping node ${\ensuremath{\mathcal{M}}}_i$ to the corresponding bit nodes $c_{i,m}$. From the SPA, they take the form $$\begin{aligned} \lefteqn{ p_{c_{i,m}\leftarrow{\ensuremath{\mathcal{M}}}_i}(c) }\nonumber\\ &\propto& \sum_{k=1}^{|{\mathbb{S}}|} \sum_{{\ensuremath{\boldsymbol{c}}}:c_m=c} p(s{^{(k)}}{\,|\,}{\ensuremath{\boldsymbol{c}}}) ~ p_{{\ensuremath{\mathcal{M}}}_i\leftarrow s_i}(s{^{(k)}}) \prod_{m'\neq m} p_{c_{i,m'}\rightarrow{\ensuremath{\mathcal{M}}}_i}(c_{m'}) \nonumber\\ &=& \sum_{k: c_m{^{(k)}}=c} p_{{\ensuremath{\mathcal{M}}}_i\leftarrow s_i}(s{^{(k)}}) \frac{\prod_{m'=1}^M p_{c_{i,m'}\rightarrow{\ensuremath{\mathcal{M}}}_i}(c_{m'}{^{(k)}})} {p_{c_{i,m}\rightarrow{\ensuremath{\mathcal{M}}}_i}(c)} \\ &=& \frac{1}{p_{c_{i,m}\rightarrow{\ensuremath{\mathcal{M}}}_i}(c)} \sum_{k: c_m{^{(k)}}=c} p_{{\ensuremath{\mathcal{M}}}_i\leftarrow s_i}(s{^{(k)}}) p_{{\ensuremath{\mathcal{M}}}_i\rightarrow s_i}(s{^{(k)}}) \end{aligned}$$ for pairs $(i,m)$ that do not correspond to pilot/training bits. (Since the pilot/training bits are known with certainty, there is no need to update their pmfs.) Finally, messages are passed leftward into the coding/interleaving block. Doing so is equivalent to feeding extrinsic soft bit estimates to a soft-input/soft-output (SISO) decoder/deinterleaver, which treats them as priors. Since SISO decoding is a well-studied topic [@MacKay:Book:03; @Richardson:Book:09] and high-performance implementations are readily available (e.g., [@Kozintsev:SW]), we will not elaborate on the details here. It suffices to say that, once the extrinsic outputs of the SISO decoder have been computed, they are re-interleaved and passed rightward from the coding/interleaving block to begin another turbo iteration. These turbo iterations continue until either the decoder detects no bit errors, the soft bit estimates have converged, or a maximum number of iterations has elapsed. Numerical Results {#sec:sims} ================= In this section, we present numerical results that compare JCED using our GAMP-based scheme to that using soft-input soft-output (SISO) equalizers based on linear MMSE (LMMSE) and LASSO [@Tibshirani:JRSSb:96], as well as to performance bounds based on perfect channel state information (CSI). Setup ----- For all results, we used irregular LDPC codes with codeword length $\sim\! 10000$ and average column weight $3$, generated (and decoded) using the publicly available software [@Kozintsev:SW], with random interleaving. We focus on the case of $N\!=\!1024$ subcarrier OFDM with $16$-QAM (i.e., $M\!=\!4$) operating at a spectral efficiency of $\eta\!=\!2$ bpcu. For bit-to-symbol mapping, we used multilevel Gray-mapping [@deJong:TCOM:05], noting recent work [@Samuel:ASIL:09] that conjectures the optimality of Gray-mapping when BICM is used with a strong code. In some simulations, we used ${N_\textsf{p}}\!>\!0$ pilot-only subcarriers and ${M_\textsf{t}}\!=\!0$ interspersed training bits, whereas in others we used ${N_\textsf{p}}\!=\!0$ and ${M_\textsf{t}}\!>\!0$. When ${N_\textsf{p}}\!>\!0$, the pilot subcarriers were placed randomly and modulated with (known) QAM symbols chosen uniformly at random. When ${M_\textsf{t}}\!>\!0$, the training bits were placed at the most significant bits (MSBs) of uniformly spaced data-subcarriers and modulated with the bit value $1$. Realizations of the tap vector ${\ensuremath{\boldsymbol{x}}}[q]$ were generated from IEEE 802.15.4a outdoor-NLOS impulse responses and SRRC pulses, as described in [Section \[sec:IEEE\]]{}, and *not* from the GM2-HMM model. The tap vectors generated for our simulations are thus as realistic as one can hope to obtain in software. All reported results are averaged over $5000$ channel realizations (i.e., $10^7$ info bits). The GM2-HMM parameters ${\ensuremath{\boldsymbol{\nu}}}^0,{\ensuremath{\boldsymbol{\nu}}}^1,{\ensuremath{\boldsymbol{p}}}^{01},{\ensuremath{\boldsymbol{p}}}^{10}$ were fit from $10000$ realizations of the tap-vector ${\ensuremath{\boldsymbol{x}}}$ using the procedure described in [Section \[sec:IEEE\]]{}. In doing so, we implicitly assumed[^10] that the receiver is designed for the outdoor scenario, and we leverage the prior information made available by the extensive measurement campaign conducted for the IEEE 802.15.4a standard [@Molisch:802.15.4a]. In all cases, we used a *maximum* of $10$ turbo iterations, $5$ equalizer iterations, $15$ GAMP iterations, and $25$ LDPC decoder iterations, although in most cases the iterations converged early (as described in [Section \[sec:jced\_gamp\]]{}). Comparison with other schemes ----------------------------- The proposed GAMP-based equalizer was compared with soft-input soft-output (SISO) equalizers based on LMMSE and LASSO [@Tibshirani:JRSSb:96], whose constructions are now detailed. All SISO equalizers are provided with the soft inputs ${\ensuremath{\Hat{\boldsymbol{s}}}}[q]$ and ${\ensuremath{\boldsymbol{\nu}}}^s[q]$, i.e., the means and variances, respectively, of the symbols ${\ensuremath{\boldsymbol{s}}}[q]\in{\mathbb{S}}^N$. (Note that, if certain elements in ${\ensuremath{\boldsymbol{s}}}[q]$ are known perfectly as pilots, then the corresponding elements in ${\ensuremath{\boldsymbol{\nu}}}^s[q]$ will be zero-valued.) Then, writing ${\ensuremath{\boldsymbol{s}}}[q]={\ensuremath{\Hat{\boldsymbol{s}}}}[q]+{\ensuremath{\Tilde{\boldsymbol{s}}}}[q]$, where ${\ensuremath{\Tilde{\boldsymbol{s}}}}[q]$ an unknown zero-mean deviation, the subcarrier observations ${\ensuremath{\boldsymbol{y}}}[q]=\operatorname{{{\ensuremath{\mathcal{D}}}}}({\ensuremath{\boldsymbol{s}}}[q]){\ensuremath{\boldsymbol{\Phi}}}{\ensuremath{\boldsymbol{x}}}[q]+{\ensuremath{\boldsymbol{w}}}[q]$ can be written as $$\begin{aligned} {\ensuremath{\boldsymbol{y}}}[q] &=& \operatorname{{{\ensuremath{\mathcal{D}}}}}({\ensuremath{\boldsymbol{s}}}[q]){\ensuremath{\boldsymbol{\Phi}}}{\ensuremath{\boldsymbol{x}}}[q]+{\ensuremath{\boldsymbol{v}}}[q], \label{eq:yq}\end{aligned}$$ where ${\ensuremath{\boldsymbol{v}}}[q]{\triangleq}\operatorname{{{\ensuremath{\mathcal{D}}}}}({\ensuremath{\Tilde{\boldsymbol{s}}}}[q]){\ensuremath{\boldsymbol{\Phi}}}{\ensuremath{\boldsymbol{x}}}[q]+{\ensuremath{\boldsymbol{w}}}[q]$ is a zero-mean noise. Treating the elements within ${\ensuremath{\Tilde{\boldsymbol{s}}}}[q]$ as uncorrelated and doing the same with ${\ensuremath{\boldsymbol{x}}}[q]$, and leveraging the fact that ${\ensuremath{\boldsymbol{\Phi}}}$ is a truncated DFT matrix, it is straightforward to show that $\operatorname{Cov}({\ensuremath{\boldsymbol{v}}}[q]) = \operatorname{{{\ensuremath{\mathcal{D}}}}}({\ensuremath{\boldsymbol{\nu}}}^v[q])$ with ${\ensuremath{\boldsymbol{\nu}}}^v[q] = \nu^w{\ensuremath{\boldsymbol{1}}} + ({\ensuremath{\boldsymbol{1}}}{^\textsf{T}}{\ensuremath{\boldsymbol{\rho}}}){\ensuremath{\boldsymbol{\nu}}}^s[q]$, where ${\ensuremath{\boldsymbol{\rho}}}$ denotes the channel’s PDP. Without loss of generality, [(\[eq:yq\])]{} can then be converted to the equivalent white-noise model $$\begin{aligned} {\ensuremath{\boldsymbol{u}}}[q] &{\triangleq}& \operatorname{{{\ensuremath{\mathcal{D}}}}}({\ensuremath{\boldsymbol{\nu}}}^v[q])^{-\frac{1}{2}} {\ensuremath{\boldsymbol{y}}}[q] = {\ensuremath{\boldsymbol{A}}}{\ensuremath{\boldsymbol{x}}}[q]+{\ensuremath{\boldsymbol{n}}}[q], \label{eq:u}\end{aligned}$$ where $\operatorname{Cov}({\ensuremath{\boldsymbol{n}}}[q])={\ensuremath{\boldsymbol{I}}}$ and ${\ensuremath{\boldsymbol{A}}}[q]{\triangleq}\operatorname{{{\ensuremath{\mathcal{D}}}}}({\ensuremath{\boldsymbol{\nu}}}^v[q]^{-\frac{1}{2}}{\ensuremath{\boldsymbol{s}}}[q]){\ensuremath{\boldsymbol{\Phi}}}$ is a known matrix. In summary, [(\[eq:u\])]{} provides a mechanism to handle soft inputs for both LASSO and LMMSE. For LMMSE equalization, we first used [(\[eq:u\])]{} to compute $$\begin{aligned} {\ensuremath{\Hat{\boldsymbol{x}}}}{_\text{\sf lmmse}}[q] &=& \operatorname{{{\ensuremath{\mathcal{D}}}}}({\ensuremath{\boldsymbol{\rho}}}){\ensuremath{\boldsymbol{A}}}{^\textsf{H}}[q] \big({\ensuremath{\boldsymbol{A}}}[q]\operatorname{{{\ensuremath{\mathcal{D}}}}}({\ensuremath{\boldsymbol{\rho}}}){\ensuremath{\boldsymbol{A}}}{^\textsf{H}}[q] + {\ensuremath{\boldsymbol{I}}}\big)^{-1} {\ensuremath{\boldsymbol{u}}}[q] \quad\end{aligned}$$ from which we obtain the subcarrier gain estimate ${\ensuremath{\Hat{\boldsymbol{z}}}}{_\text{\sf lmmse}}[q]={\ensuremath{\boldsymbol{\Phi}}}{\ensuremath{\Hat{\boldsymbol{x}}}}{_\text{\sf lmmse}}[q]$. The covariance matrix of ${\ensuremath{\Hat{\boldsymbol{z}}}}{_\text{\sf lmmse}}[q]$ is [@Poor:Book:94] $${\ensuremath{\boldsymbol{\Phi}}} \big(\! \operatorname{{{\ensuremath{\mathcal{D}}}}}({\ensuremath{\boldsymbol{\rho}}}) - \operatorname{{{\ensuremath{\mathcal{D}}}}}({\ensuremath{\boldsymbol{\rho}}}){\ensuremath{\boldsymbol{A}}}{^\textsf{H}}[q] \big({\ensuremath{\boldsymbol{A}}}[q]\operatorname{{{\ensuremath{\mathcal{D}}}}}({\ensuremath{\boldsymbol{\rho}}}){\ensuremath{\boldsymbol{A}}}{^\textsf{H}}[q]+{\ensuremath{\boldsymbol{I}}}\big)^{-1} {\ensuremath{\boldsymbol{A}}}[q]\operatorname{{{\ensuremath{\mathcal{D}}}}}({\ensuremath{\boldsymbol{\rho}}}) \!\big) {\ensuremath{\boldsymbol{\Phi}}}{^\textsf{H}}$$ whose diagonal elements ${\ensuremath{\boldsymbol{\nu}}}^z{_\text{\sf lmmse}}[q]$ are variances on the gain estimates ${\ensuremath{\Hat{\boldsymbol{z}}}}{_\text{\sf lmmse}}[q]$. Finally, we obtain soft symbol estimates from the soft gain estimates $({\ensuremath{\Hat{\boldsymbol{z}}}}{_\text{\sf lmmse}}[q],{\ensuremath{\boldsymbol{\nu}}}^z{_\text{\sf lmmse}}[q])$ via [(\[eq:p\_y\_to\_s\])]{}. For LASSO,[^11] we first computed the tap estimate ${\ensuremath{\Hat{\boldsymbol{x}}}}{_\text{\sf lasso}}[q]$ from [(\[eq:u\])]{} using the celebrated SPGL1 algorithm [@vandenBerg:JSC:08]. In doing so, we needed to specify the target residual variance, i.e., $\nu^u{_\text{\sf lasso}}\!{\triangleq}\! \frac{1}{N}{\ensuremath{\| {\ensuremath{\boldsymbol{u}}}[q]-{\ensuremath{\boldsymbol{A}}}[q]{\ensuremath{\Hat{\boldsymbol{x}}}}{_\text{\sf lasso}}[q] \|}}_2^2$. Because $\operatorname{Cov}({\ensuremath{\boldsymbol{n}}}[q])={\ensuremath{\boldsymbol{I}}}$, we expect the optimal value of $\nu^u{_\text{\sf lasso}}$ to be near $1$ and, after extensive experimentation, we found that the value $\nu^u{_\text{\sf lasso}}\!=\!0.9$ works well at high SNR and that the value $\nu^u{_\text{\sf lasso}}\!=\!1.5$ works well at low SNR. Thus, for each ${\ensuremath{\boldsymbol{u}}}[q]$, we computed SPGL1 estimates using each of these two[^12] targets, and kept the one that minimized the squared error $\nu^x{_\text{\sf lasso}}[q]{\triangleq}\frac{1}{L}{\ensuremath{\| {\ensuremath{\boldsymbol{x}}}[q]-{\ensuremath{\Hat{\boldsymbol{x}}}}{_\text{\sf lasso}}[q] \|}}_2^2$, which we assume a genie is able to provide. For the soft outputs, we set ${\ensuremath{\Hat{\boldsymbol{z}}}}{_\text{\sf lasso}}[q]={\ensuremath{\boldsymbol{\Phi}}}{\ensuremath{\Hat{\boldsymbol{x}}}}{_\text{\sf lasso}}[q]$ and take ${\ensuremath{\boldsymbol{\nu}}}^z{_\text{\sf lasso}}[q]$ to be the diagonal elements of ${\ensuremath{\boldsymbol{\Phi}}}\operatorname{Cov}({\ensuremath{\Hat{\boldsymbol{x}}}}{_\text{\sf lasso}}[q]){\ensuremath{\boldsymbol{\Phi}}}{^\textsf{H}}$. Assuming $\operatorname{Cov}({\ensuremath{\Hat{\boldsymbol{x}}}}{_\text{\sf lasso}}[q])=\nu^x{_\text{\sf lasso}}[q]{\ensuremath{\boldsymbol{I}}}$ and leveraging the fact that ${\ensuremath{\boldsymbol{\Phi}}}$ is a truncated DFT matrix, we find ${\ensuremath{\boldsymbol{\nu}}}^z{_\text{\sf lasso}}[q]= L\nu^x{_\text{\sf lasso}}[q] {\ensuremath{\boldsymbol{1}}}$. Finally, using [(\[eq:p\_y\_to\_s\])]{}, we obtain soft symbol estimates from the soft gain estimates $({\ensuremath{\Hat{\boldsymbol{z}}}}{_\text{\sf lasso}}[q],{\ensuremath{\boldsymbol{\nu}}}^z{_\text{\sf lasso}}[q])$. Due to the genie-aided steps, the performance attained by our LASSO implementation is better than what could be obtained in practice. These LMMSE- and LASSO-based SISO equalizers were then embedded in the overall factor graph in the same manner as GAMP, with the following exceptions: 1) The LMMSE and LASSO algorithms could not be connected to the MC sub-block, since they are not based on a two-state mixture model; 2) For LASSO, if the genie-aided MSE $\nu^x{_\text{\sf lasso}}[q]$ did not improve during a given turbo iteration, then the corresponding outputs $({\ensuremath{\Hat{\boldsymbol{z}}}}{_\text{\sf lasso}}[q],{\ensuremath{\boldsymbol{\nu}}}^z{_\text{\sf lasso}}[q])$ were not updated. This rule was employed to prevent turbo-LASSO from occasionally diverging at low SNR; 3) For LASSO, if ${N_\textsf{p}}\!>\!0$ and ${M_\textsf{t}}\!=\!0$, then the LASSO estimates computed during the first turbo iteration use only pilot subcarriers. This makes the performance of SISO-LASSO after the first turbo iteration equal to the performance of the standard pilot-aided LASSO. ${\textsf{BER}}$ versus the number of pilot subcarriers ${N_\textsf{p}}$ ------------------------------------------------------------------------ [Figure \[fig:ber\_vs\_Np\]]{} shows bit error rate (${\textsf{BER}}$) versus the number of pilot subcarriers ${N_\textsf{p}}$ at $E_b/N_o\!=\!11$ dB and a fixed spectral efficiency of $\eta\!=\!2$ bpcu. In this and other figures, “” refers to algorithm with \# turbo iterations (and “” after turbo convergence; see [Fig. \[fig:time\_vs\_snr\_Np\]]{}) with the MC block disconnected (i.e., there was no attempt to exploit tap clustering). Meanwhile “” refers to GAMP+MC after \# turbo iterations, each containing $5$ equalizer iterations. Finally, refers to MAP equalization under perfect CSI, which yields a bound on the ${\textsf{BER}}$ performance of any equalizer. The curves in [Fig. \[fig:ber\_vs\_Np\]]{} exhibit a “U” shape because, as ${N_\textsf{p}}$ increases, the code rate $R$ must decrease to maintain the fixed spectral efficiency $\eta=2$ bpcu. While an increase in ${N_\textsf{p}}$ generally makes channel estimation easier, the reduction in $R$ makes data decoding more difficult. For all schemes under comparison, [Fig. \[fig:ber\_vs\_Np\]]{} suggests that the choice ${N_\textsf{p}}\!\approx\!224$ is optimal under the operating conditions. Overall, we see GAMP significantly outperforming both LMMSE and LASSO. Moreover, we see a small but definite gain from the MC block. ${\textsf{BER}}$ versus the number of interspersed training bits ${M_\textsf{t}}$ --------------------------------------------------------------------------------- Although ${N_\textsf{p}}>0$ pilot subcarriers are required for decoupled channel estimation and decoding, JCED can function with ${N_\textsf{p}}\!=\!0$ as long as a sufficient number ${M_\textsf{t}}$ of training bits are interspersed among the coded bits used to construct each QAM symbol. To examine this latter case, [Fig. \[fig:ber\_vs\_Mt\]]{} shows ${\textsf{BER}}$ versus ${M_\textsf{t}}$ at $E_b/N_o\!=\!10$ dB, a fixed spectral efficiency of $\eta\!=\!2$ bpcu, and ${N_\textsf{p}}\!=\!0$. Again we see the “U” shape, but with GAMP working very well for a relatively wide range of ${M_\textsf{t}}$, and again we see a small but noticeable BER improvement when the MC block is used. SISO-LASSO seems to work to some degree with ${N_\textsf{p}}=0$, but SISO-LASSO does not. ${\textsf{BER}}$ versus $E_b/N_o$ --------------------------------- [Figure \[fig:ber\_vs\_snr\_Np\]]{} shows ${\textsf{BER}}$ versus $E_b/N_o$ using ${N_\textsf{p}}\!=\!224$ pilot subcarriers (as suggested by [Fig. \[fig:ber\_vs\_Np\]]{}) and ${M_\textsf{t}}\!=\!0$ training bits. Relative to the perfect-CSI bound, we see SISO-LASSO performing within $5$ dB during the first turbo iteration and within $4.5$ dB after convergence. Meanwhile, we see SISO-LMMSE performing very poorly during the first turbo iteration, but eventually surpassing SISO-LASSO and coming within $4$ dB from the perfect-CSI bound. Remarkably, we see GAMP+MC performing within $0.6$ dB of the perfect-CSI bound (and within $1$ dB after only $2$ turbo iterations). This excellent performance confirms that the proposed GM2-HMM channel model and equalizer design together do an excellent job of capturing and exploiting the lag-dependent clustered-sparse characteristics of the 802.15.4a channel taps. Comparing the GAMP traces to the GAMP+MC traces, we see that the MC block yields a small but noticeable benefit. [Figure \[fig:ber\_vs\_snr\_Mt\]]{} shows ${\textsf{BER}}$ versus $E_b/N_o$ using ${M_\textsf{t}}\!=\!448$ interspersed training bits (as suggested by [Fig. \[fig:ber\_vs\_Mt\]]{}) and ${N_\textsf{p}}\!=\!0$ pilot subcarriers. There we see that SISO-LASSO does not perform well at all. SISO-LMMSE works to some degree after several turbo iterations, although not as well as in the ${N_\textsf{p}}>0$ case. Meanwhile, we see GAMP+MC performing within $1$ dB of the perfect-CSI case, and GAMP alone performing within $1.5$ dB. Comparing [Fig. \[fig:ber\_vs\_snr\_Mt\]]{} to [Fig. \[fig:ber\_vs\_snr\_Np\]]{}, we see GAMP with training bits performing about $1$ dB better than GAMP with dedicated pilot subcarriers. The perfect-CSI bound likewise improves because, with $16$-QAM, ${M_\textsf{t}}\!=\!448$ training bits constitutes half the overhead of ${N_\textsf{p}}\!=\!224$ pilot subcarriers, allowing [Fig. \[fig:ber\_vs\_snr\_Mt\]]{} the use of a stronger code at $\eta\!=\!2$ bpcu. Channel-tap ${\textsf{NMSE}}$ versus $E_b/N_o$ ---------------------------------------------- [Figure \[fig:nmse\_vs\_snr\_Np\]]{} shows the channel estimates’ normalized mean-squared error (${\textsf{NMSE}}$) $\operatorname{E}\{\|{\ensuremath{\boldsymbol{x}}}[q]\!-\!{\ensuremath{\Hat{\boldsymbol{x}}}}[q]\|_2^2/\|{\ensuremath{\boldsymbol{x}}}[q]\|_2^2\}$ versus $E_b/N_o$, at the point that the turbo iterations were terminated, using ${N_\textsf{p}}\!=\!224$ pilot subcarriers and ${M_\textsf{t}}\!=\!0$ training bits. (For comparison, [Fig. \[fig:ber\_vs\_snr\_Np\]]{} shows ${\textsf{BER}}$ for this configuration.) We also show the ${\textsf{NMSE}}$ attained by the “bit and support genie” (BSG), which calculates MMSE channel estimates using perfect knowledge of both the coded bits and the hidden channel states $\{d_j\}$, and which provides a lower bound for any channel estimator. In the figure, we see that the ${\textsf{NMSE}}$s of LMMSE and LASSO channel estimates are within $8$-to-$12$ dB of the BSG, whereas those of GAMP are within $2$-to-$4$ dB. Meanwhile, we see that GAMP+MC has a small but noticeable advantage over GAMP alone. We reason that the LMMSE estimates are worse than the GAMP estimates because they do not exploit the non-Gaussianity of the channel taps $x_j$, and the LASSO estimates are worse than the GAMP estimates because they do not exploit the known priors on the channel taps (i.e., the lag-dependent sparsity ${\ensuremath{\boldsymbol{\lambda}}}$ and PDP ${\ensuremath{\boldsymbol{\rho}}}$). Computational complexity versus $E_b/N_o$ ----------------------------------------- [Figure \[fig:time\_vs\_snr\_Np\]]{} shows the average time per turbo iteration (in Matlab seconds on a $2.6$GHz CPU), the average number of turbo iterations, and the average total time (to turbo convergence), as a function of $E_b/N_o$, using ${N_\textsf{p}}\!=\!224$ pilot subcarriers and ${M_\textsf{t}}\!=\!0$ training bits. (For comparison, [Fig. \[fig:ber\_vs\_snr\_Np\]]{} shows ${\textsf{BER}}$ for this configuration and [Fig. \[fig:nmse\_vs\_snr\_Np\]]{} shows ${\textsf{NMSE}}$.) Regarding the average time per turbo iteration, we see GAMP$\pm$MC taking $\approx 1.5$ sec at low $E_b/N_o$ and $\approx 0.5$ sec at high $E_b/N_o$. GAMP+MC takes only slightly longer than GAMP alone due to the efficiency of the message computations within the MC block, and the fact that both the GAMP iterations and equalizer iterations are terminated as soon as the messages converge. In comparison, SISO-LMMSE takes $\approx 4.5$ sec per turbo iteration, and SISO-LASSO takes between $1$ and $7$ sec, depending on $E_b/N_o$. Regarding the number of average number of turbo iterations until convergence, we see that—at low $E_b/N_o$—GAMP+MC takes about $5$ turbo iterations, GAMP alone takes about $7$, SISO-LMMSE takes about $5$, and SISO-LASSO takes about $3$, while—at high $E_b/N_o$—all algorithms converge after only $1$ turbo iteration. Regarding the total time for equalization, GAMP+MC and GAMP are about the same at low $E_b/N_o$, whereas GAMP alone takes about $30\%$ less time at high $E_b/N_o$. Meanwhile, SISO-LASSO and SISO-LMMSE are uniformly slower than GAMP and GAMP+MC over the entire $E_b/N_o$ range, in some cases by a factor of $10$. Conclusion {#sec:conc} ========== In this paper, we presented a factor-graph approach to joint channel-estimation and decoding (JCED) for BICM-OFDM that merges recent advances in approximate message passing algorithms [@Rangan:10b] with those in structured-sparse signal reconstruction [@Schniter:CISS:10] and SISO decoding [@MacKay:Book:03]. Different from existing factor-graph approaches to JCED, ours is able to exploit not only sparse channel taps, but also clustered sparsity patterns that typify large-bandwidth communication channels, such as those that result from pulse-shaped communication over IEEE 802.15.4a modeled channels. For this purpose, we proposed the use of a two-state Gaussian mixture prior with a Markov model on the hidden tap states. The implementation complexity of our JCED scheme is dominated by ${\ensuremath{\mathcal{O}}}(N\log_2 N \!+\! N|{\mathbb{S}}|)$ multiplies per GAMP iteration, facilitating the application to systems with many subcarriers $N$ and many channel taps $L<N$. Experiments with IEEE 802.15.4a modeled channels showed ${\textsf{BER}}$ performance within $1$ dB of the known-channel bound, and $3$–$4$ dB better than LMMSE- and LASSO-based soft equalizers. These experiments also suggested that, with our proposed approach, the use of interspersed training bits is more efficient than the use of dedicated pilot subcarriers. For very large constellations (e.g., $|{\mathbb{S}}|\!=\!1024$), future work is motivated to reduce the linear complexity dependence on $|{\mathbb{S}}|$. Derivation of GAMP Functions [$g{_{\textsf{out},i}}$]{.nodecor} and [$g'{_{\textsf{out},i}}$]{.nodecor} {#app:out} ======================================================================================================= In this appendix, we derive the GAMP quantities $g{_{\textsf{out},i}}(y,\hat{z},\nu^z)$ and $g'{_{\textsf{out},i}}(y,\hat{z},\nu^z)$ given in [(\[eq:gout\])]{}-[(\[eq:ei\])]{}. From (D1), we have that $$\begin{aligned} \operatorname{E}_{Z_i|Y_i}\{z {\,|\,}y;\hat{z},\nu^z\} &=& \frac{1}{p_{Y_i}(y)} \int_z z \, p_{Y_i|Z_i}(y|z) \, {\ensuremath{\mathcal{CN}}}(z;\hat{z},\nu^z) , \quad \label{eq:gout2}\end{aligned}$$ where $p_{Y_i}(y) {\triangleq}\int_z p_{Y_i|Z_i}(y|z) {\ensuremath{\mathcal{CN}}}(z;\hat{z},\nu^z)$. From [(\[eq:pY|Z\])]{}, we rewrite $p_{Y_i|Z_i}(y|z)$ as $$\begin{aligned} p_{Y_i|Z_i}(y|z) &=& \sum_{k=1}^{2^M} \frac{\beta_i{^{(k)}}}{s{^{(k)}}} \,{\ensuremath{\mathcal{CN}}}\Big(z;\frac{y}{s{^{(k)}}},\frac{\nu^w}{|s{^{(k)}}|^2}\Big) ,\end{aligned}$$ so that $$\begin{aligned} \int_z z\, p_{Y_i|Z_i}(y|z) {\ensuremath{\mathcal{CN}}}(z;\hat{z},\nu^z) &=& \sum_{k=1}^{2^M} \frac{\beta_i{^{(k)}}}{s{^{(k)}}} \int_z z \, {\ensuremath{\mathcal{CN}}}\Big(z;\frac{y}{s{^{(k)}}},\frac{\nu^w}{|s{^{(k)}}|^2}\Big) {\ensuremath{\mathcal{CN}}}(z;\hat{z},\nu^z) \quad \\ p_{Y_i}(y) &=& \sum_{k=1}^{2^M} \frac{\beta_i{^{(k)}}}{s{^{(k)}}} \int_z {\ensuremath{\mathcal{CN}}}\Big(z;\frac{y}{s{^{(k)}}},\frac{\nu^w}{|s{^{(k)}}|^2}\Big) {\ensuremath{\mathcal{CN}}}(z;\hat{z},\nu^z).\end{aligned}$$ Using the property that $$\begin{aligned} {\ensuremath{\mathcal{CN}}}(x;\hat{\theta},\nu^\theta){\ensuremath{\mathcal{CN}}}(x;\hat{\phi},\nu^\phi) &=& {\ensuremath{\mathcal{CN}}}\Big(x;\frac{\hat{\theta}/\nu^\theta+\hat{\phi}/\nu^\phi} {1/\nu^\theta+1/\nu^\phi},\frac{1}{1/\nu^\theta+1/\nu^\phi}\Big) {\ensuremath{\mathcal{CN}}}(0;\hat{\theta}-\hat{\phi},\nu^\theta+\nu^\phi) , \label{eq:pogr}\end{aligned}$$ we can rewrite $$\begin{aligned} \lefteqn{ \int_z z \, p_{Y_i|Z_i}(y|z) \, {\ensuremath{\mathcal{CN}}}(z;\hat{z},\nu^z) }\nonumber\\ &=& \sum_{k=1}^{2^M} \frac{\beta_i{^{(k)}}}{s{^{(k)}}} {\ensuremath{\mathcal{CN}}}\Big(0;\frac{y_i}{s}-\hat{z},\frac{\nu^w}{|s{^{(k)}}|^2}+\nu^z\Big) \int_z z \, {\ensuremath{\mathcal{CN}}}\bigg(z;\frac{ \frac{y}{s{^{(k)}}}\frac{|s{^{(k)}}|^2}{\nu^w}+\frac{\hat{z}}{\nu^z} }{ \frac{|s{^{(k)}}|^2}{\nu^w}+\frac{1}{\nu^z} },\frac{1}{\frac{|s{^{(k)}}|^2}{\nu^w}+\frac{1}{\nu^z}}\bigg) \label{eq:prod}\\ &=& \sum_{k=1}^{2^M} \frac{\beta_i{^{(k)}}}{s{^{(k)}}} {\ensuremath{\mathcal{CN}}}\Big(\frac{y_i}{s};\hat{z},\frac{\nu^w}{|s{^{(k)}}|^2}+\nu^z\Big) \frac{ \frac{y}{s{^{(k)}}}\frac{|s{^{(k)}}|^2}{\nu^w}+\frac{\hat{z}}{\nu^z} }{ \frac{|s{^{(k)}}|^2}{\nu^w}+\frac{1}{\nu^z} } \quad \\ &=& \sum_{k=1}^{2^M} \beta_i{^{(k)}} {\ensuremath{\mathcal{CN}}}\big(y_i;s{^{(k)}}\hat{z},|s{^{(k)}}|^2\nu^z+\nu^w\big) \bigg( \underbrace{ \Big(\frac{y}{s{^{(k)}}}-\hat{z}\Big) \frac{|s{^{(k)}}|^2 \nu^z}{|s{^{(k)}}|^2 \nu^z + \nu^w} }_{\displaystyle {\triangleq}\hat{e}{^{(k)}}(y,\hat{z},\nu^z)} + \hat{z} \bigg) \label{eq:intz}\end{aligned}$$ and, using the same procedure, we get $$\begin{aligned} p_{Y_i}(y) &=& \sum_{k=1}^{2^M} \beta_i{^{(k)}} {\ensuremath{\mathcal{CN}}}\big(y_i;s{^{(k)}}\hat{z},|s{^{(k)}}|^2\nu^z+\nu^w\big) . \label{eq:pY}\end{aligned}$$ With $\xi_i{^{(k)}}(y,\hat{z},\nu^z)$ defined in [(\[eq:xi\])]{}, equations [(\[eq:gout2\])]{} and [(\[eq:intz\])]{} and [(\[eq:pY\])]{} combine to give $$\begin{aligned} \operatorname{E}_{Z_i|Y_i}\{z {\,|\,}y;\hat{z},\nu^z\} &=& \sum_{k=1}^{2^M} \xi_i{^{(k)}}(y,\hat{z},\nu^z) \big( \hat{e}{^{(k)}}(y,\hat{z},\nu^z) + \hat{z} \big) . \quad \label{eq:EZY}\end{aligned}$$ Finally, from [(\[eq:EZY\])]{} and the definition of $g{_{\textsf{out},i}}(y,\hat{z},\nu^z)$ in (D2), equation [(\[eq:gout\])]{} follows immediately. From (D1), we have that $$\begin{aligned} \operatorname{var}_{Z_i|Y_i}\{z {\,|\,}y;\hat{z},\nu^z\} &=& \frac{1}{p_{Y_i}(y)} \int_z |z-\operatorname{E}_{Z_i|Y_i}\{z {\,|\,}y;\hat{z},\nu^z\}|^2 \, p_{Y_i|Z_i}(y|z) \, {\ensuremath{\mathcal{CN}}}(z;\hat{z},\nu^z) . \label{eq:varZY2}\end{aligned}$$ Similar to [(\[eq:prod\])]{}, we can write $$\begin{aligned} \lefteqn{ \int_z |z-\operatorname{E}_{Z_i|Y_i}\{z {\,|\,}y;\hat{z},\nu^z\}|^2 \, p_{Y_i|Z_i}(y|z) \, {\ensuremath{\mathcal{CN}}}(z;\hat{z},\nu^z) }\nonumber\\ &=& \sum_{k=1}^{2^M} \frac{\beta_i{^{(k)}}}{s{^{(k)}}} {\ensuremath{\mathcal{CN}}}\Big(0;\frac{y_i}{s}-\hat{z},\frac{\nu^w}{|s{^{(k)}}|^2}+\nu^z\Big) \nonumber\\&&\mbox{}\times \int_z |z-\operatorname{E}_{Z_i|Y_i}\{z {\,|\,}y;\hat{z},\nu^z\}|^2 \, {\ensuremath{\mathcal{CN}}}\bigg(z;\frac{ \frac{y}{s{^{(k)}}}\frac{|s{^{(k)}}|^2}{\nu^w}+\frac{\hat{z}}{\nu^z} }{ \frac{|s{^{(k)}}|^2}{\nu^w}+\frac{1}{\nu^z} },\frac{1}{\frac{|s{^{(k)}}|^2}{\nu^w}+\frac{1}{\nu^z}}\bigg) .\end{aligned}$$ Then, using the change-of-variable $\tilde{z}{\triangleq}z-\operatorname{E}_{Z_i|Y_i}\{z {\,|\,}y;\hat{z},\nu^z\}$, and absorbing the $s{^{(k)}}$ terms as done in [(\[eq:intz\])]{}, we get $$\begin{aligned} \lefteqn{ \int_z |z-\operatorname{E}_{Z_i|Y_i}\{z {\,|\,}y;\hat{z},\nu^z\}|^2 \, p_{Y_i|Z_i}(y|z) \, {\ensuremath{\mathcal{CN}}}(z;\hat{z},\nu^z) }\nonumber\\ &=& \sum_{k=1}^{2^M} \beta_i{^{(k)}} {\ensuremath{\mathcal{CN}}}\big(y_i;s{^{(k)}}\hat{z},|s{^{(k)}}|^2\nu^z+\nu^w\big) \nonumber\\&&\mbox{}\times \int_{\tilde{z}} |\tilde{z}|^2 \, {\ensuremath{\mathcal{CN}}}\Big(\tilde{z}; \hat{e}{^{(k)}}+ \underbrace{\hat{z}-\operatorname{E}_{Z_i|Y_i}\{z {\,|\,}y;\hat{z},\nu^z\}}_{\displaystyle = -\hat{e}_i}, \frac{\nu^w\nu^z}{|s{^{(k)}}|^2\nu^z+\nu^w}\Big) \\ &=& \sum_{k=1}^{2^M} \beta_i{^{(k)}} {\ensuremath{\mathcal{CN}}}\big(y_i;s{^{(k)}}\hat{z},|s{^{(k)}}|^2\nu^z+\nu^w\big) \Big( |\hat{e}{^{(k)}}-\hat{e}_i|^2 + \frac{\nu^w\nu^z}{|s{^{(k)}}|^2\nu^z+\nu^w} \Big).\label{eq:intz2}\end{aligned}$$ Using $\xi_i{^{(k)}}(y,\hat{z},\nu^z)$ defined in [(\[eq:xi\])]{} and $\zeta{^{(k)}}(y,\hat{z},\nu^z)$ defined in [(\[eq:zeta\])]{}, equations [(\[eq:pY\])]{} and [(\[eq:varZY2\])]{} and [(\[eq:intz2\])]{} combine to give $$\begin{aligned} \operatorname{var}_{Z_i|Y_i}\{z {\,|\,}y;\hat{z},\nu^z\} &=& \sum_{k=1}^{2^M} \xi_i{^{(k)}}(y,\hat{z},\nu^z) \Big( \frac{\nu^w \zeta{^{(k)}}(y,\hat{z},\nu^z)}{|s{^{(k)}}|^2} + \big|\hat{e}_i(y,\hat{z},\nu^z)-\hat{e}{^{(k)}}(y,\hat{z},\nu^z)\big|^2 \Big). \label{eq:varZY}\end{aligned}$$ which is rewritten as $\nu^e_i(y,\hat{z},\nu^z) {\triangleq}\operatorname{var}_{Z_i|Y_i}\{z {\,|\,}y;\hat{z},\nu^z\}$ in [(\[eq:muei\])]{}. Finally, plugging $\nu^e_i(y,\hat{z},\nu^z)$ into the definition of $g'{_{\textsf{out},i}}(y,\hat{z},\nu^z)$ in (D3), we immediately obtain [(\[eq:g’out\])]{}. Derivation of GAMP Functions [$g{_{\textsf{in},j}}$]{.nodecor} and [$g'{_{\textsf{in},j}}$]{.nodecor} {#app:in} ===================================================================================================== In this appendix, we derive the GAMP quantities $g{_{\textsf{in},j}}(\hat{r},\nu^r)$ and $g'{_{\textsf{in},j}}(\hat{r},\nu^r)$ given in [(\[eq:gin\])]{}-[(\[eq:alfj\])]{}. From (D4)-(D6), we note that $g{_{\textsf{in},j}}(\hat{r},\nu^r)$ and $\nu^r g'{_{\textsf{in},j}}(\hat{r},\nu^r)$ are the mean and variance, respectively, of the pdf $$\begin{aligned} \frac{1}{Z_j} p_{X_j}\!(r) \,{\ensuremath{\mathcal{CN}}}(r;\hat{r},\nu^r) ,\label{eq:pdfin}\end{aligned}$$ where $Z_j{\triangleq}\int_r p_{X_j}\!(r) \,{\ensuremath{\mathcal{CN}}}(r;\hat{r},\nu^r)$. Using [(\[eq:pogr\])]{} together with the definition of $p_{X_j}\!(.)$ from [(\[eq:pxj\])]{}, we find $$\begin{aligned} \lefteqn{ p_{X_j}\!(r) \,{\ensuremath{\mathcal{CN}}}(r;\hat{r},\nu^r) }\nonumber\\ &=& \lambda_j {\ensuremath{\mathcal{CN}}}(r;0,\nu_j^1) \,{\ensuremath{\mathcal{CN}}}(r;\hat{r},\nu^r) + (1-\lambda_j) {\ensuremath{\mathcal{CN}}}(r;0,\nu_j^0) \,{\ensuremath{\mathcal{CN}}}(r;\hat{r},\nu^r) \\ &=& \textstyle \lambda_j {\ensuremath{\mathcal{CN}}}(\hat{r};0,\nu^1_j+\nu^r) \, {\ensuremath{\mathcal{CN}}}\big(r;\hat{r}\gamma^1_j(\nu^r),\nu^r\gamma^1_j(\nu^r)\big) \nonumber\\&&\mbox{} \textstyle + (1-\lambda_j) {\ensuremath{\mathcal{CN}}}(\hat{r};0,\nu^0_j+\nu^r) \, {\ensuremath{\mathcal{CN}}}\big(r;\hat{r}\gamma^0_j(\nu^r),\nu^r\gamma^0_j(\nu^r)\big) \end{aligned}$$ for $\gamma^0_j(\nu^r) {\triangleq}(1+\nu^r/\nu_j^0)^{-1}$ and $\gamma^1_j(\nu^r) {\triangleq}(1+\nu^r/\nu_j^1)^{-1}$. This implies that $$\begin{aligned} Z_j &=& \lambda_j {\ensuremath{\mathcal{CN}}}(\hat{r};0,\nu^1_j+\nu^r) + (1-\lambda_j) {\ensuremath{\mathcal{CN}}}(\hat{r};0,\nu^0+\nu^r) .\end{aligned}$$ Thus, the mean obeys $$\begin{aligned} g{_{\textsf{in},j}}(\hat{r},\nu^r) &=& \frac{1}{Z_j}\int_r r \, p_{X_j}\!(r) \,{\ensuremath{\mathcal{CN}}}(r;\hat{r},\nu^r) \\ &=& \underbrace{ \frac{\lambda_j{\ensuremath{\mathcal{CN}}}(\hat{r};0,\nu_j^1+\nu^r)}{Z_j} }_{\displaystyle = \alpha_j(\hat{r},\nu^r)} \gamma_j^1(\nu^r) \,\hat{r} + \underbrace{ \frac{(1-\lambda_j){\ensuremath{\mathcal{CN}}}(\hat{r};0,\nu_j^0+\nu^r)}{Z_j} }_{\displaystyle = 1-\alpha_j(\hat{r},\nu^r)} \gamma_j^0(\nu^r) \,\hat{r} , \label{eq:gin2} $$ yielding [(\[eq:gin\])]{}, where a straightforward manipulation relates the expression for $\alpha_j(\hat{r},\nu^r)$ above with its definition in [(\[eq:alfj\])]{}. Since, for the pdf in [(\[eq:pdfin\])]{}, $g{_{\textsf{in},j}}$ is the mean and $\nu^r g'{_{\textsf{in},j}}$ is the variance, we can write $$\begin{aligned} \nu^r g'{_{\textsf{in},j}}(\hat{r},\nu^r) &=& \frac{1}{Z_j}\int_r |r|^2 \, p_{X_j}\!(r) \,{\ensuremath{\mathcal{CN}}}(r;\hat{r},\nu^r) -|g{_{\textsf{in},j}}|^2 \\ &=& \alpha_j \big( |\hat{r}\gamma^1_j|^2 + \nu^r\gamma^1_j \big) +(1-\alpha_j) \big( |\hat{r}\gamma^0_j|^2 + \nu^r\gamma^0_j \big) -\big|\alpha_j\gamma^1_j\hat{r} + (1-\alpha_j)\gamma^0_j\hat{r}\big|^2 , \label{eq:g'in2}\end{aligned}$$ which can be simplified to yield [(\[eq:g’in\])]{}. [^1]: Please direct all correspondence to Prof. Philip Schniter, Dept. ECE, The Ohio State University, 2015 Neil Ave., Columbus OH 43210, e-mail: schniter@ece.osu.edu, phone 614.247.6488, fax 614.292.7596. [^2]: This work has been supported in part by NSF grant CCF-1018368 and DARPA/ONR grant N66001-10-1-4090, and an allocation of computing time from the Ohio Supercomputer Center. [^3]: For our GAMP decoder, we recommend ${N_\textsf{p}}\!=\!0$; see [Section \[sec:sims\]]{}. [^4]: The message passing algorithm described in [Section \[sec:gamp\]]{} can also handle non-Gaussian mixtures and/or mixtures with more than two terms. [^5]: Since it is difficult to give a concise yet accurate account of GAMP’s technical properties, we refer the interested reader to [@Rangan:10b]. [^6]: To be precise, the GAMP algorithm in [Table \[tab:gamp\]]{} is an extension of that proposed in [@Rangan:10b]. [Table \[tab:gamp\]]{} handles circular *complex-valued* distributions and *non-identically* distributed signals and measurements. [^7]: More precisely, GAMP is iterated until the mean-square tap-estimate difference $\frac{1}{L}\sum_{j=0}^{L-1}|\hat{x}_j(n)-\hat{x}_j(n-1)|^2$ falls below a threshold or a maximum number of GAMP iterations has elapsed. [^8]: Message passing on the MC factor graph is a standard procedure. For details, we refer the reader to [@MacKay:Book:03; @Bishop:Book:07]. [^9]: More precisely, the equalizer iterations are terminated when the mean-square difference in tap-state log-likelihoods falls below a threshold or a maximum number of equalizer iterations has elapsed. [^10]: If, instead, we knew that the receiver would be used in a different operating scenario, then we could generate representative realizations of ${\ensuremath{\boldsymbol{x}}}$ for that scenario and fit the GM2-HMM parameters accordingly. Furthermore, one could optimize the receiver for any desired balance between “typical” and “worst-case” operating conditions by simply choosing appropriate training realizations ${\ensuremath{\boldsymbol{x}}}$. [^11]: The criterion employed by LASSO [@Tibshirani:JRSSb:96] is equivalent to the one employed in “basis pursuit denoising” [@Chen:JSC:98]. [^12]: We also tried running SPGL1 for a dense grid of $\nu^u{_\text{\sf lasso}}$ values, but often it would get “stuck” at one of them and eventually return an error.

--- abstract: 'We study the subgap spin and charge transport in normal metal-ferromagnet-superconductor trilayers induced by bias voltage and/or magnetization precession. Transport properties are discussed in terms of time-dependent scattering theory. We assume the superconducting gap is small on the energy scales set by the Fermi energy and the ferromagnetic exchange splitting, and compute the non-equilibrium charge and spin current response to first order in precession frequency, in the presence of a finite applied voltage. We find that the voltage-induced instantaneous charge current and longitudinal spin current are unaffected by the precessing magnetization, while the pumped transverse spin current is determined by spin-dependent conductances and details of the electron-hole scattering matrix. A simplified expression for the transverse spin current is derived for structures where the ferromagnet is longer than the transverse spin coherence length.' author: - Hans Joakim Skadsem and Arne Brataas - Jan Martinek - Yaroslav Tserkovnyak title: 'FMR and voltage induced transport in normal metal$-$ferromagnet$-$superconductor trilayers' --- Introduction {#sec:introduction} ============ Experimental and theoretical studies of spin polarized transport in hybrid magnetic nanostructures is a frontier in mesoscopic physics. The most prominent example of conceptual, technological, and commercial impact is the giant magnetoresistance effect utilized in magnetic information storage devices. In order to gain a deeper understanding of spin and charge transport, and to enhance circuit functionality and efficiency, more complex structures are fabricated and studied. In recent years, hybrid nanoscale circuits containing normal conductors, ferromagnets, and superconductors have been realized. These structures allow observation and understanding of competing mechanisms of electron-electron interactions. The simultaneous existence of ferromagnetism and superconductivity is rare. In ferromagnets, the exchange interaction lifts the spin-degeneracy and induces an itinerant spin polarization. In *s*-wave superconductors, on the other hand, electrons with anti-parallel spins form Cooper pairs. In homogenous conventional ferromagnets (Fe, Ni, Co, and alloys thereof), the large exchange interaction efficiently dephases electron-hole pairs, and eliminates singlet superconducting correlations over distances larger than the ferromagnetic coherence length. This would suggest a short-range superconducting proximity effect in transition metal ferromagnets.[kawaguchi92:\_magnet\_fe\_nb,PhysRevB.55.15174]{} Such a simple picture cannot explain recent measurements on Co and Ni ferromagnets coupled to Al superconductors, however, where a substantial resistance drop was observed at the onset of superconductivity.[giroud98:\_super\_proxim\_effec\_in\_mesos\_ferrom\_wire,petrashov99:\_giant\_mutual\_proxim\_effec\_in]{} The simple picture also fails to explain the long-range superconducting proximity effect recently observed via the Josephson supercurrent through a half-metallic ferromagnet.[keizer06:\_spin\_tripl\_super\_throug\_half,anwar10]{} Subsequent theoretical work show that induced triplet superconducting correlations give rise to long ranged proximity effect in transition metal ferromagnets.[bergeret01:\_long\_range\_proxim\_effec\_super\_ferrom\_struc,kadigrobov01:\_quant]{} Triplet superconducting correlations are insensitive to the pair-breaking exchange interaction and exhibit a longer coherence length, similar to that of superconducting correlations in normal metals. It is now established that spin-flip processes in a ferromagnet can convert singlet into triplet pair correlations. A spatially inhomogeneous magnetization texture [bergeret05:\_odd\_tripl\_super\_and\_relat]{} or magnons [tkachov01:\_subgap\_trans\_in\_ferrom\_super,takahashi07:\_super\_pumpin\_in\_josep\_junct,houzet08:\_ferrom\_josep\_junct\_with\_preces\_magnet]{} are examples of spin-flip sources that are able to induce long ranged triplet correlations. In this report, we focus our attention on the influence of magnons on the transport properties in normal metal-ferromagnet-superconductor systems. Even normal metal-ferromagnet systems without superconductors exhibit intriguing physics, and especially the interaction between spin and charge currents and the magnetic order parameter in such structures have attracted tremendous interest. For instance, a non-collinear spin flow towards a ferromagnet exerts a torque on the magnetization, a spin transfer torque, that can excite the magnetization and even induce steady state, precessional motion of the ferromagnetic order parameter.[spintransfer2,spintransfer1]{} The inverse effect is also of significant interest: A precessing ferromagnet in electrochemical equilibrium with its environment, acts as a “spin battery” by emitting (or “pumping”) pure spin currents into neighboring materials.[@spinpumping] When emitted spins are dissipated in adjacent materials, spin pumping enhances magnetic dissipation in the precessing ferromagnet, and thus increases observed linewidths in FMR experiments.[@mizukami01:_ferrom_nm_nm_nm_cu] Some ideas from spin transfer physics in normal metal-ferromagnet structures were recently used to study superconductor-ferromagnet systems. A FMR experiment[@bell08:_spin_dynam_in_super_ferrom_proxim_system] and the following theoretical analysis[@morten08:_proxim] have shown how spin pumping can be used to visualize proximity effects and spin relaxation processes inside the superconductor. In essence, in metallic contacts, ferromagnetic correlations reduce the superconducting order parameter close to the layer interface, enabling pumped sub-gap electrons to enter and deposit spin in the superconductor. This is a prime example of how the inverse proximity effect affects the FMR linewidth broadening when typical spin-flip lengths are comparable to the superconducting coherence length.[@sillanpaeae01:_inver] We direct our attention to a different aspect of the interplay between magnetization and carrier dynamics in ferromagnet-superconductor structures. In contrast to the works mentioned above, where the magnetization dynamics have been the primary concern, we will consider how a precessing magnetization and an applied voltage bias induce spin and charge currents in a normal metal-ferromagnet-superconductor (N|F|S) trilayer. The computed charge currents can be measured directly, whereas spin currents can possibly be measured by its dissipative effect on the precessing ferromagnet, its spin transfer torque effect on a second ferromagnet, or via spin-filtering as a charge buildup on another ferromagnet.[@spinpumping] Related to our work, sub-gap transport properties have recently been studied in a normal metal-ferromagnetic superconductor structure.[brataas04:\_spin\_and\_charg\_pumpin\_by]{} In ferromagnetic superconductors, magnetic and electron-hole correlations coexist which can result in novel transport and dynamical magnetic phenomena. It was shown how superconducting correlations, namely Andreev reflections at the layer interface, add features to the results of spin and charge pumping in normal metal-ferromagnet systems. In this report, we also consider how pumping in the N|F|S trilayer is related to pumping in the normal metal-ferromagnetic superconductor system as studied in Ref. . Diffusive transport in hybrid superconductor-normal metal systems is usually formulated within a quasiclassical description.[@rammersmith] Although this description give qualitative insight into transport properties of superconductor-ferromagnet systems,[bergeret05:\_odd\_tripl\_super\_and\_relat,houzet08:\_ferrom\_josep\_junct\_with\_preces\_magnet]{} the formalism is limited to ferromagnets with exchange interactions much smaller than the Fermi energy. Thus, a quasiclassical description cannot be used to quantitatively study transport in transition metal ferromagnets Fe, Ni and Co used in experiments. This is one of the reaons why we adopt the scattering theory to transport.[@landauerbuttiker] Another reasone is that scattering theory captures adiabatic slow time-dependent variations of the magnetization direction well. Scattering theory has proven most useful in the study of stationary charge and spin currents in magnetoelectronic structures,[brataas06:\_non\_collin\_magnet]{} and the time-dependent generalization has successfully been applied to describe parametric pumping of charge[buettiker94:\_curren,brouwer98:\_scatt,buttiker06:\_scatt\_theor\_of\_dynam\_elect\_trans]{} and spin currents.[@spinpumping] For the N|F|S structure under consideration, we derive charge- and spin currents in the normal metal conductor in response to a slowly precessing ferromagnetic exchange field and applied bias voltage. We focus on sub-gap energies, and how Andreev scattering contributes to the conductivites of the currents. In electro-chemical equilibrium, we make contact with the results for pumping in normal metal-ferromagnetic superconductor structures.[brataas04:\_spin\_and\_charg\_pumpin\_by]{} We proceed by detailing how time- and energy gradients of the total scattering matrix contribute to non-equilibrium pumped currents, and find that both charge and longitudinal spin currents are unaffected by the precessing magnetization. Finally, we consider non-equilibrium charge and spin currents for trilayers where the ferromagnetic region is longer than the transverse spin coherence length. This paper is organized in the following way: The N|F|S system is described in Sec. \[sec:model-description\]. In Sec. \[sec:time-depend-scatt\], we use time-dependent scattering theory to derive general expressions for charge and spin currents to first order in pumping frequency. The total scattering matrix for the system is then invoked in Sec. [sec:evaluation-currents]{} to obtain non-equilibrium pumped currents. Our conclusions are in Sec. \[sec:conclusion\]. Model description {#sec:model-description} ================= The system is sketched in Fig. \[fig:1\]. It consists of a superconductor (S) in series with a ferromagnet (F) and a normal metal lead (N$_1$). N$_1$ is ideally coupled to a normal metal reservoir (N$_{\text{res}}$). We assume N$_{\text{res}}$ and S to be in local thermal equilibrium, and denote a possible chemical potential difference between the normal and the superconducting side as $\mu_{N} - \mu_{S} = eV$. Spin-orbit interactions are disregarded, and the ferromagnetic order parameter is assumed to be homogeneous and with a fixed magnitude $\Delta_{xc}$ inside F. Its direction is along the time-dependent unit vector $\boldsymbol{m}(t) = (\sin \theta(t) \cos \Omega t, \sin \theta(t) \sin \Omega t, \cos \theta(t))$. The precessing magnetization serves as the pumping parameter in the system. ![A ferromagnetic scattering region (F) is connected to a superconductor (S) and a normal metal reservoir (N$_{\text{res}}$) via two normal metal leads (N$_1$ and N$_2$). Amplitudes of outgoing (incoming) carrier states are given by $b$ ($a$).[]{data-label="fig:1"}](model_v2 "fig:") (-140,80) (-112,75)[$\boldsymbol{m}(t)$]{} (-30,80) (-75,80) (-185,80) (-230,100) (-185,65)[$b$]{} (-185,51)[$a$]{} (-160,67)[(-1,0)[15]{}]{} (-175,53)[(1,0)[15]{}]{} (-125,45)[(1,4)[10]{}]{} We focus on sub-gap transport properties. Thus, possible scattering processes include Andreev reflections at the F|S interface[@andreev64] and spin-dependent normal scattering inside F. Following a standard procedure,[@beenakker92:_quant_trans_in_semic_super_microj] the scattering problem is greatly simplified by utilizing spatially separated regions where scattering processes occur. This is achieved by inserting a fictitious normal metal lead (N$_2$) between F and S. We assume that N$_2$ is longer than the Fermi wavelength, so that asymptotic, plane wave solutions are applicable in this region. The total scattering matrix is a concatenation of the scattering matrices for N$_1$|F|N$_2$ and for Andreev reflections at the N$_2$|S interface. Transport between F and S is mediated by the ballistic N$_2$ lead. The singlet superconductor is described by the BCS Hamiltonian $$\begin{gathered} \label{eq:7} \hat{\mathcal{H}} = \sum_{\sigma = \uparrow,\downarrow} \int\mathrm{d}% \boldsymbol{r} \: \hat{\Psi}^{\dagger}_{\sigma} (\boldsymbol{r}) H_0 (% \boldsymbol{r}) \hat{\Psi}_{\sigma}(\boldsymbol{r}) \\ + \int \mathrm{d}\boldsymbol{r} \: \left\{ \Delta (\boldsymbol{r}) \hat{\Psi}% ^{\dagger}_{\uparrow}(\boldsymbol{r}) \hat{\Psi}^{\dagger}_{\downarrow}(% \boldsymbol{r}) + \Delta^*(\boldsymbol{r}) \hat{\Psi}_{\downarrow}(% \boldsymbol{r}) \hat{\Psi}_{\uparrow}(\boldsymbol{r}) \right\},\end{gathered}$$ where $H_0$ is the normal state, single-particle Hamiltonian and $\Delta(% \boldsymbol{r})$ the superconducting gap. We model the gap by a step function, $\Delta(\boldsymbol{r}) = \Delta\mathrm{e}^{\mathrm{i}\phi} \Theta (x)$, where the phase $\phi$ is constant, and $x$ is the coordinate perpendicular to the N$_2$|S interface. We take the Fermi energy $E_F$ to be the largest energy scale, and focus on the low energy transport properties in regimes when the superconducting gap is much less than the exchange interaction in the ferromagnet $\Delta_{xc}$, $eV \le \Delta \ll \Delta_{xc},E_F$. The Hamiltonian (\[eq:7\]) is diagonalized by the following Bogoliubov transformation[@ketterson99:_super] $$\label{eq:8} \hat{\Psi}_{\sigma}(\boldsymbol{r}) = \sum_n \left\{ \hat{\gamma}_n u_n (% \boldsymbol{r},\sigma) + \hat{\gamma}^{\dagger}_n v^*_n (\boldsymbol{r}% ,\sigma) \right\},$$ where $\hat{\gamma}_n^{(\dagger)}$ are quasiparticle annihilation (creation) operators that satisfy the fermionic anti-commutation relation $$\label{eq:24} \{ \hat{\gamma}_m, \hat{\gamma}^{\dagger}_{n} \} = \delta_{m,n}.$$ The transformation (\[eq:8\]) results in a matrix equation for the quasiparticle eigenfunctions $u_n$ and $v_n$: $$\label{eq:9} \begin{pmatrix} H_0(\boldsymbol{r}) & \mathrm{i}\Delta(\boldsymbol{r}) \sigma^y \\ - \mathrm{i}\Delta^*(\boldsymbol{r}) \sigma^y & - H_0^*(\boldsymbol{r})% \end{pmatrix} \begin{pmatrix} u_n(\boldsymbol{r}) \\ v_n(\boldsymbol{r})% \end{pmatrix} = \varepsilon_n \begin{pmatrix} u_n(\boldsymbol{r}) \\ v_n(\boldsymbol{r})% \end{pmatrix}% .$$ The quasiparticle excitation energy $\varepsilon_n$ is measured with respect to the chemical potential of the superconductor, which is set to zero. $% \sigma^y$ is a Pauli matrix operating in spin space. The Bogoliubov-de Gennes Hamiltonian (\[eq:9\]) is the starting point when we in Sec. [sec:scatt-matr-nfs]{} derive the appropriate reflection amplitudes for quasiparticles impinging on the superconductor interface. Time-dependent scattering theory {#sec:time-depend-scatt} ================================ We now focus on the time-dependent scattering theory for the N|F|S structure in Fig. \[fig:1\], apply the general framework established in Refs. , and make use of the scattering theory for hybrid superconductor-normal metal structures discussed in Refs. . We find it most convenient to study a slowly precessing magnetization by a scattering matrix expressed in the Wigner representation,[@rammersmith] making the derivation of pumped currents similar to that carried out for normal systems in Refs. . In order to describe a scattering potential of arbitrary time-dependence, we start by considering the two-time scattering matrix $\mathcal{S}(t,t^{\prime })$, that relates annihilation operators between states outgoing and incoming from the scattering region: $$\hat{b}_{\alpha }(t)=\sum_{\beta }\int \mathrm{d}t^{\prime } \mathcal{S}% _{\alpha \beta }(t,t^{\prime })\hat{a}_{\beta }(t^{\prime }). \label{eq:28}$$As indicated in Fig. \[fig:1\], $\hat{b}_{\alpha }:(\hat{a}_{\alpha })$ annihilates the outgoing (incoming) state $\alpha $. $\alpha $ labels electron-hole Nambu space index, spin and transverse wave-guide number. We assume that the reservoirs connected to the scattering region are in local thermal equilibrium, and that incoming carriers from the normal metal reservoir fulfill $$\langle \hat{a}_{\alpha }^{\dagger }(\varepsilon )\hat{a}_{\alpha ^{\prime }}(\varepsilon ^{\prime })\rangle =\delta _{\alpha ,\alpha ^{\prime }}\delta (\varepsilon -\varepsilon ^{\prime })f_{\alpha }(\varepsilon ), \label{eq:29}$$where the brackets indicate a quantum and statistical average, and $$\begin{gathered} f_{e(h)}(\varepsilon )=f_{0}(\varepsilon -\sigma ^{e(h)}eV) \label{eq:5} \\ =\left[ 1+\mathrm{e}^{(\varepsilon -\sigma ^{e(h)}eV)/k_{B}T_{el}}\right] ^{-1},\end{gathered}$$where $\sigma ^{e(h)}=+(-)1$, and $f_{e(h)}(\varepsilon )$ is the Fermi-Dirac distribution of incoming electrons (holes) at a charge bias $eV$ and electron temperature $T_{el}$. We will eventually consider electron temperature to be lower than the superconducting gap. the We will now proceed by computing charge and spin currents in the system. Matrix current {#sec:matrix-current} -------------- We seek the right-going charge and spin currents in normal metal lead 1, and start by introducing the matrix current[tserkovnyak01:\_shot\_noise\_in\_ferrom\_normal\_metal\_system]{} $$\label{eq:20} \hat{I}_{1,\alpha\beta}(t) = 2\pi e \tau^z_{\alpha\beta} \left( \hat{a}% ^{\dagger}_{\beta}(t) \hat{a}_{\alpha}(t) - \hat{b}^{\dagger}_{\beta}(t) \hat{b}_{\alpha}(t) \right),$$ where $e$ is the electronic charge, and $\tau^z$ is a Pauli matrix in electron-hole space: $$\label{eq:30} \tau^z = \begin{pmatrix} 1 & 0 \\ 0 & -1% \end{pmatrix}% .$$ Charge and spin currents are obtained from the matrix current (\[eq:20\]) as follows: $$\label{eq:21} I_{c} (t) = \sum_{\alpha} \langle \hat{I}_{1,\alpha\alpha}(t) \rangle,$$ and $$\label{eq:22} \boldsymbol{I}_s (t) = \frac{1}{2e} \sum_{\alpha,\beta} \boldsymbol{\rho}% _{\alpha\beta} \langle \hat{I}_{1,\beta\alpha} (t) \rangle,$$ respectively. Summations run over electron-hole, spin and mode space, and $% \boldsymbol{\rho}$ is a matrix with diagonal structure in electron-hole space: $$\label{eq:31} \boldsymbol{\rho}_{\alpha\beta} \equiv \begin{pmatrix} \boldsymbol{\sigma}_{\alpha\beta} & 0 \\ 0 & \boldsymbol{\sigma}^*_{\alpha\beta}% \end{pmatrix}% ,$$ and with a vector of the Pauli matrices and their complex conjugates, as the diagonal elements. For a slowly oscillating scatterer, it is convenient to express the scattering matrix in the Wigner representation[rammersmith,wang02:\_heat\_curren\_in\_param\_quant\_pump,wang02:\_optim\_quant\_pump\_in\_presen]{} $$\mathcal{S}_{\alpha \beta }(t,t^{\prime })=\int_{0}^{\infty }\frac{\mathrm{d}% \varepsilon }{2\pi }\mathrm{e}^{-\mathrm{i}\varepsilon (t-t^{\prime })}% \mathcal{S}_{\alpha \beta }\left( \varepsilon ;\frac{t+t^{\prime }}{2}% \right) . \label{eq:32}$$In this representation, the matrix current is: $$\begin{gathered} \label{eq:81} \langle \hat{I}_{1,\alpha\beta}(t) \rangle = \frac{e}{2\pi} \tau^z_{\alpha\beta} \biggr\{ \delta_{\alpha,\beta} \int_0^{\infty} \mathrm{d}\varepsilon \: f_{\alpha} (\varepsilon) - \sum_{\gamma} \int_{-\infty}^{\infty} \mathrm{d}\tau \: \mathrm{d}T \int_0^{\infty} \frac{\mathrm{d}\varepsilon_1 \: \mathrm{d}\varepsilon_2}{2\pi} f_{\gamma}(\tau) \\ \times \mathrm{e}^{-\mathrm{i}\varepsilon_1 (T - \tau/2)} \mathrm{e}^{\mathrm{i}\varepsilon_2(T + \tau/2)} \mathcal{S}_{\alpha\gamma}\left (\varepsilon_2; t + \frac{T + \tau/2}{2} \right) \mathcal{S}^*_{\beta\gamma} \left( \varepsilon_1; t + \frac{T - \tau/2}{2} \right) \biggr\}. \end{gathered}$$ The current is expressed in terms of the center and relative time coordinates $T=(t^{\prime }+t^{\prime \prime })/2$ and $\tau =t^{\prime \prime }-t^{\prime }$, and the Fourier transform of the distribution function $$f_{\gamma }(\tau )\equiv \int_{0}^{\infty }\frac{\mathrm{d}\varepsilon }{% 2\pi }\mathrm{e}^{-\mathrm{i}\varepsilon \tau }f_{\gamma }(\varepsilon ). \label{eq:33}$$ When the scattering matrix $\mathcal{S}(\varepsilon;t)$ is a concatenation of multiple time-dependent scattering elements, the Wigner representation of $\mathcal{S}$ will also be an infinite sum of time and energy gradients.[rammersmith]{} The magnetization dynamics is slow as compared to the time an electron spends in the scattering region. In the *adiabatic* approximation, we assume the scattering matrix evolves on a much longer timescale than the typical dwell times of particles inside the scattering region. In this regime, we formally expand $\mathcal{S}$ as[moskalets04:\_adiab\_quant\_pump\_in\_presen]{} $$\label{eq:50} \mathcal{S}(\varepsilon;t) = S_0(\varepsilon;t) + A(\varepsilon;t) + \mathcal{O}(\partial_t^2 S_0)$$ where $S_0$ is the “frozen” or instantaneous scattering matrix, and the matrix $A$ represents all first-order gradient corrections to $S_0$ resulting from the concatenation of time-dependent scattering elements that describe the device. Unitarity of $\mathcal{S}$ to all orders in time- and energy-gradients implies[@moskalets04:_adiab_quant_pump_in_presen] $$\label{eq:51} S_0 A^{\dagger} + A S_0^{\dagger} = \frac{\mathrm{i}}{2} \left( \partial_t S_0 \partial_{\varepsilon} S_0^{\dagger} - \partial_{\varepsilon} S_0 \partial_t S_0^{\dagger} \right) \equiv \frac{1}{2} P \left\{ S_0; S_0^{\dagger} \right\} ,$$ where a Poisson bracket $P\{.;.\}$ has been defined to ease the notation. In the following, scattering matrix arguments $(\varepsilon;t)$ are omitted in places where there is no risk of confusion. To obtain a local (in time) expression for the matrix current (\[eq:81\]), we Taylor expand $\mathcal{S}$ to first order in time derivatives, and obtain the matrix current $$\begin{gathered} \label{eq:82} \langle \hat{I}_{1,\alpha\beta}(t) \rangle = \frac{e}{2\pi} \tau^z_{\alpha\beta} \sum_{\gamma} \int_0^{\infty} \mathrm{d}\varepsilon \biggr\{ \left( f_{\alpha}(\varepsilon) - f_{\gamma}(\varepsilon) \right) \left( S_{0,\alpha\gamma} S_{0,\beta\gamma}^* + A_{\alpha\gamma} S^*_{0,\beta\gamma} + S_{0,\alpha\gamma} A^*_{\beta\gamma} - \frac{1}{2} P \left\{ S_{0,\alpha\gamma}; S^*_{0,\beta\gamma} \right\} \right) \\ + \frac{\mathrm{i}}{2} \left(- \partial_{\varepsilon} f_{\gamma} (\varepsilon)\right) \left( S_{0,\alpha\gamma} \partial_t S^*_{0,\beta\gamma} - \partial_t S_{0,\alpha\gamma} S^*_{0,\beta\gamma} \right) \biggr\} + \mathcal{O} \left( \partial_t^2 S_0 \right) \end{gathered}$$ where Eqs. (\[eq:50\]) and  have been used. The matrix current in Eq. (\[eq:82\]) is *exact* to first order in frequency of the pumping parameter. Finally, we observe that in the absence of a voltage bias, the gradient corrections to the frozen scattering matrix, represented by $A$, vanish from the matrix current. In electro-chemical equilibrium, when $V = 0$, $% f_e(\varepsilon) = f_h(\varepsilon)$, the first line of Eq. (\[eq:82\]) vanishes, and the pumped current is determined by the frozen scattering matrix. Naturally, the same is also true for the time-dependent theory based on Floquet scattering matrices.[buttiker06:\_scatt\_theor\_of\_dynam\_elect\_trans]{} Scattering matrix for a N|F|S structure {#sec:scatt-matr-nfs} --------------------------------------- In this section, the scattering matrix formalism derived for N|S structures[@beenakker92:_quant_trans_in_semic_super_microj] is applied to our N|F|S trilayer. As described in Sec. \[sec:model-description\], the scattering description of a N|F|S structure is greatly simplified by inserting a fictitious normal metal lead (N$_2$) between the two scattering regions, thereby spatially separating spin-dependent scattering in F and Andreev reflection at the N$_2$|S interface.[beenakker92:\_quant\_trans\_in\_semic\_super\_microj]{} The scattering matrix $S_F$, describing the disordered ferromagnetic region, is block-diagonal in electron-hole space. We write $S_F$ as $$\label{eq:34} S_F(\varepsilon;t) = \begin{pmatrix} s_F(\varepsilon;t) & 0 \\ 0 & s_F(-\varepsilon;t)^*% \end{pmatrix}% ,$$ where the diagonal elements are $$\label{eq:35} s_F = \begin{pmatrix} r_{11} & t_{12} \\ t_{21} & r_{22}% \end{pmatrix}% .$$ Here, $r_{ii}$ and $t_{ij}$ are matrices in spin-space that describe reflection of an incoming electron in lead $i$, and transmission of an electron from lead $j$ to lead $i$, respectively. Electrons and holes with opposite spins are coupled by Andreev reflection at the superconductor interface, where an incoming electron (hole) is reflected as a hole (electron) with reversed spin direction. The reflection amplitudes are derived by matching propagating wave functions in N$_2$ with evanescent wave functions in the superconductor. The resulting scattering matrix reads[beenakker92:\_quant\_trans\_in\_semic\_super\_microj,waintal02:\_magnet\_exchan\_inter\_induc\_by\_josep\_curren]{} $$\label{eq:36} r^A = \begin{pmatrix} 0 & r^A_{eh} \\ r^A_{he} & 0% \end{pmatrix} = \begin{pmatrix} 0 & \mathrm{i}\alpha\sigma^y \mathrm{e}^{\mathrm{i}\phi} \\ -\mathrm{i}\alpha \sigma^y \mathrm{e}^{-\mathrm{i}\phi} & 0% \end{pmatrix}% ,$$ where $\alpha = \mathrm{exp} \left[ -\mathrm{i}\arccos (\varepsilon / \Delta) \right]$. The total scattering matrix of the N|F|S structure is a concatenation of $S_F $ and $r^A$, and in terms of the frozen scattering matrices, we obtain the familiar results[beenakker92:\_quant\_trans\_in\_semic\_super\_microj,blaauboer02:\_charg\_pumpin\_in\_mesos\_system]{} \[eq:6\] $$\begin{aligned} \label{eq:16} S_0^{ee} (\varepsilon&;t) = r_{11}(\varepsilon) \notag \\ & + t_{12}(\varepsilon) r^A_{eh}(\varepsilon) r^*_{22}(-\varepsilon) M_e (\varepsilon) r^A_{he}(\varepsilon) t_{21}(\varepsilon), \\ S_0^{hh} (\varepsilon&;t) = r_{11}^*(-\varepsilon) \notag \\ & + t^*_{12}(-\varepsilon) r^A_{he}(\varepsilon) r_{22}(\varepsilon) M_h (\varepsilon) r^A_{eh}(\varepsilon) t^*_{21}(-\varepsilon), \\ S_0^{eh} (\varepsilon&;t) = t_{12}(\varepsilon) M_h(\varepsilon)r^A_{eh}(\varepsilon) t^*_{21}(-\varepsilon), \\ S_0^{he} (\varepsilon&;t) = t^*_{12}(-\varepsilon) M_{e}(\varepsilon) r^A_{he} (\varepsilon) t_{21}(\varepsilon),\end{aligned}$$ where time arguments are omitted on the right hand side of the equations for sake of notation. Multiple reflections between S and F, mediated by propagations through N$_2$, are described by $$\begin{aligned} \label{eq:38} M_e (\varepsilon) & = \left[1 - r^A_{he}(\varepsilon) r_{22}(\varepsilon) r^A_{eh}(\varepsilon) r^*_{22}(-\varepsilon) \right]^{-1}, \\ M_h(\varepsilon) & = \left[1 - r^A_{eh}(\varepsilon) r^*_{22}(-\varepsilon) r^A_{he}(\varepsilon) r_{22}(\varepsilon) \right]^{-1}.\end{aligned}$$ From Eqs. (\[eq:6\]), and using $r^A_{eh}(-\varepsilon)^* = r^A_{he}(\varepsilon)$, one obtains the following symmetry relations for the total scattering matrix: $$\label{eq:4} \mathcal{S}^{ee}(\varepsilon;t) = \left[ \mathcal{S}^{hh}(-\varepsilon;t) % \right]^*,$$ and $$\label{eq:10} \mathcal{S}^{eh}(\varepsilon;t) = \left[ \mathcal{S}^{he}(-\varepsilon;t) % \right]^*.$$ The frozen scattering matrices in Eqs. (\[eq:6\]) are all time-dependent due to the slowly varying magnetization in the ferromagnet. Arguably the easiest way to evaluate the matrix current, is to perform a spinor rotation that aligns the spin quantization axis with the instantaneous magnetization direction.[@spinpumping; @brataas04:_spin_and_charg_pumpin_by] The total scattering matrix $$\label{eq:40} S_0 (\varepsilon;t) = \begin{pmatrix} S_0^{ee} & S_0^{eh} \\ S_0^{he} & S_0^{hh}% \end{pmatrix}%$$ can be related to the total scattering matrix $\underline{S}$ in the rotating frame by the spinor rotations $$\label{eq:25} S_0 (\varepsilon;t) = {W}^{\dagger}(t) \underline{S} (\varepsilon) W(t),$$ where $W(t) = V(t) U(t)$, with $$\label{eq:26} U(t) = \begin{pmatrix} \mathcal{U}(t) & 0 \\ 0 & \mathcal{U}^{\dagger}(t)% \end{pmatrix} = \begin{pmatrix} \mathrm{exp} \left[\frac{\mathrm{i} \Omega t}{2} \sigma^z\right] & 0 \\% [0.3cm] 0 & \mathrm{exp}\left[- \frac{\mathrm{i} \Omega t}{2} \sigma^z \right]% \end{pmatrix}% ,$$ and $$\label{eq:27} V(t) = \begin{pmatrix} \mathcal{V}(t) & 0 \\ 0 & \mathcal{V}(t)% \end{pmatrix} = \begin{pmatrix} \mathrm{exp}\left[\frac{\mathrm{i}\theta(t)}{2} \sigma^y\right] & 0 \\% [0.3cm] 0 & \mathrm{exp} \left[\frac{\mathrm{i}\theta(t)}{2} \sigma^y \right]% \end{pmatrix}% .$$ In the rotating frame, $\underline{S}_0^{ee}$ and $\underline{S}_0^{hh}$ are both diagonal in spin space, while $\underline{S}_0^{eh}$ and $\underline{S}% _0^{he}$, which mix spin $\sigma$ electrons with spin $-\sigma$ holes, only have off-diagonal elements. Now that the matrix current and relevant scattering matrices are derived, we proceed to study pumped charge and spin currents for a voltage biased trilayer structure. Pumped currents out of equilibrium {#sec:evaluation-currents} ================================== A complication that arises when the system is driven out of equilibrium, is that time- and energy gradients of the frozen scattering matrix must be evaluated. Before presenting the detailed expressions for charge and spin currents in the normal metal lead, we derive the required gradient corrections. Due to electron-hole symmetry (\[symmetry\]), it is sufficient to consider only $A^{he}$ in the gradient correction. Gradient correction matrix {#sec:grad-corr-matr-1} -------------------------- In the following, we determine $A^{he}$ by a formal gradient expansion of the corresponding scattering matrix $\mathcal{S}^{he}$, whose full time and energy dependence of $\mathcal{S}^{he}$ is given by (see Eq. (\[eq:16\])): $$\label{eq:42} \mathcal{S}^{he}(\varepsilon;t) = \left( t^*_{12} \circ M_e \circ r^{A}_{he} \circ t_{21} \right)(\varepsilon;t).$$ Evaluating the convolutions in the Wigner representation can be done by systematically expanding the exponentials:[@rammersmith] $$\label{eq:43} (A \circ B)(\varepsilon;t) = \mathrm{e}^{\mathrm{i} \left( \partial^A_{\varepsilon} \partial^B_t - \partial^A_t \partial^B_{\varepsilon} \right)/2} A(\varepsilon;t) B(\varepsilon;t),$$ where the superscripts indicate which matrix the operator works on. A significant simplification of the final result is achieved when the superconducting gap is much less than the exchange energy, $\Delta \ll \Delta_{xc}, E_F$. The energy dependence is then only determined by the energy dependence of the Andreev reflection. Since we are evaluating the energy gradients close to the Fermi level, $\partial_{\varepsilon} s_F \ll \partial_{\varepsilon} r^A$, and we obtain the simplified expression for the gradient matrix $A^{he}$: $$\begin{gathered} \label{eq:46} A^{he}(\varepsilon;t) \approx - \frac{\mathrm{i}}{2} \partial_{\varepsilon} \partial_t S^{he}_0 + \mathrm{i} t^*_{12} \partial_{\varepsilon} (M_e r^A_{he}) \partial_t t_{21} \\ + \mathrm{i} t^*_{12} \partial_{\varepsilon} \partial_t M_e r^A_{he} t_{21} + \mathrm{i} t^*_{12} \partial_t M_e \partial_{\varepsilon} M_e^{-1} M_e r^A_{he} t_{21} \\ - \mathrm{i} t^*_{12} M_e r^A_{he} \partial_t r_{22} \partial_{\varepsilon} r^A_{eh} r^*_{22} M_{e} r^A_{he} t_{21} \\ \equiv - \frac{\mathrm{i}}{2} \partial_{\varepsilon} \partial_t S_0^{he} + \Gamma^{he}.\end{gathered}$$ Here, $S_0^{he}$ is the frozen scattering matrix from Eq. (\[eq:16\]), and Before evaluating the currents, we observe that $\Gamma^{he}$ in the rotating frame is diagonal in spin space. This fact, which is important when evaluating non-equilibrium pumped charge and spin currents, can be seen from $$\label{eq:19} \Gamma^{he} = \mathcal{U}\mathcal{V}^{\dagger} \underline{\Gamma}^{he} \mathcal{VU},$$ with $$\begin{gathered} \label{eq:60} \underline{\Gamma}^{he} = \frac{\mathrm{i}}{2} \underline{t}^*_{12} \partial_{\varepsilon} (\underline{M}_e r^A_{he}) \Lambda ( \underline{t}% _{21\uparrow} - \underline{t}_{21\downarrow}) \\ + \frac{\mathrm{i}}{2} \underline{t}^*_{12} (\underline{M}_{e\uparrow} - \underline{M}_{e\downarrow}) \partial_{\varepsilon}(r^A_{he} \Lambda \underline{r}_{22} r^A_{eh}) \underline{r}^*_{22} \underline{M}_e r^A_{he} \underline{t}_{21} \\ - \frac{\mathrm{i}}{2} \underline{t}^*_{12} \underline{M}_e r^A_{he} (% \underline{r}_{22\uparrow} - \underline{r}_{22\downarrow}) \Lambda \partial_{\varepsilon} r^A_{eh} \underline{r}^*_{22} \underline{M}_e r^A_{he} \underline{t}_{21} \\ - \frac{\mathrm{i}}{2} \underline{t}^*_{12} \partial_{\varepsilon} (% \underline{M}_{e\uparrow} - \underline{M}_{e\downarrow}) r^A_{he} \Lambda \underline{t}_{21},\end{gathered}$$ where $$\label{eq:75} \Lambda \equiv \mathcal{VU} \partial_t (\boldsymbol{m}\cdot\boldsymbol{\sigma% }) \mathcal{U}^{\dagger} \mathcal{V}^{\dagger} = \partial_t \theta \sigma^x + \sin\theta \Omega \sigma^y.$$ Multiplying $r^A_{he}$, which is $\sim \sigma^y$, with $\Lambda$, and using that the other components in the equation are all diagonal, brings us to the conclusion that $\underline{\Gamma}^{he}$ is diagonal in spin space. Finally, we note that $\Gamma^{he} \to 0$ for a vanishing ferromagnetic ordering parameter. Once the gradient corrections to the frozen scattering matrix are derived, one can obtain non-equilibrium pumped currents to first order in pumping frequency. Pumped charge current {#sec:pump-charge-curr} --------------------- According to Eq. (\[eq:21\]), the charge current is obtained by tracing the matrix current (\[eq:82\]) over electron-hole, spin and mode space. Making use of the electron-hole symmetries from Eqs. (\[eq:4\])(\[eq:10\]), and using that both $\mathrm{Tr} \left\{ \partial_t S_0^{ee} S_0^{ee\dagger} \right\} = 0$ and $\mathrm{Tr} \left\{ \partial_t S_0^{he} S_0^{he\dagger} \right\} = 0$, one finds that the pumped charge current is determined by $$\begin{gathered} \label{eq:1} I_c (t) = \frac{e}{2\pi} \int_{-\infty}^{\infty} \mathrm{d}\varepsilon \: % \biggr( \left[f_e (\varepsilon) - f_h (\varepsilon)\right] \mathrm{Tr} % \Big\{ S_{0}^{he} S_{0}^{he\dagger} \\ + A^{he} S_0^{he\dagger} + S_0^{he} A^{he\dagger} - \frac{1}{2} P \left\{ S_0^{he}; S_0^{he\dagger} \right\} \Big\} \biggr),\end{gathered}$$ to first order in pumping parameter frequency. Using that $A^{he} = - \frac{% \mathrm{i}}{2} \partial_{\varepsilon} \partial_t S_0^{he} + \Gamma^{he}$, the current  simplifies to $$\begin{gathered} \label{eq:2} I_c (t) = \frac{e}{2\pi} \int_{-\infty}^{\infty} \mathrm{d}\varepsilon \: % \biggr( \left[f_e (\varepsilon) - f_h (\varepsilon)\right] \\ \times \mathrm{Tr} \Big\{ S_{0}^{he} S_{0}^{he\dagger} + \Gamma^{he} S_0^{he\dagger} + S_0^{he} \Gamma^{he\dagger} \Big\} \biggr).\end{gathered}$$ Any non-equilibrium pumped contributions to the current are determined by the remainder $\Gamma^{he}$ from Eq. (\[eq:60\]). However, as pointed out at the end of Sec. \[sec:grad-corr-matr-1\], $\underline{\Gamma}^{he}$ is a diagonal matrix in spin space. From Eq. (\[eq:16\]), we know that $% \underline{S}_0^{he}$ is strictly off-diagonal in spin space. This implies that $\mathrm{Tr} \left\{ \Gamma^{he} S_0^{he\dagger} \right\} = 0$, and the charge current is reduced to the stationary result: $$\label{eq:3} I_c = \frac{e}{2\pi} \int_{-\infty}^{\infty} \mathrm{d}\varepsilon \: \left[% f_e (\varepsilon) - f_h (\varepsilon)\right] \tilde{g} (\varepsilon),$$ where the total conductance is defined as $$\label{eq:47} \tilde{g} \equiv \sum_{m,n} \left\{ \left|\underline{S}^{he}_{\downarrow% \uparrow,mn} \right|^2 + \left|\underline{S}^{he}_{\uparrow\downarrow,mn}% \right|^2 \right\}.$$ The result in Eq. (\[eq:3\]) shows that there is no pumped charge current in N|F|S structures, even when there is an additional bias voltage driving the system, *e.g.* there are no bilinear contributions proportional to the bias voltage and the FMR frequency. The stationary result is similar to that obtained in FS|N structures,[brataas04:\_spin\_and\_charg\_pumpin\_by]{} a result that indicates that the total scattering matrix for a disordered region coupled to a ferromagnetic superconductor, is structurally equivalent to that of a disordered ferromagnetic region coupled to a superconductor. The two structures have different scattering matrices, however, and therefore the expressions for the conductances differ. Pumped spin current {#sec:pumped-spin-current} ------------------- We proceed by evaluating the pumped spin current to first order in pumping parameter frequency. Utilizing the electron-hole symmetry relations for the total scattering matrix, we obtain $$\begin{gathered} \label{eq:17} \boldsymbol{I}_s(t) = \frac{1}{4\pi} \int_{-\infty}^{\infty} \mathrm{d}\varepsilon\: \left(f_e(\varepsilon) - f_h(\varepsilon)\right) \left[\mathrm{Tr} \left\{ \boldsymbol{\sigma}^* \left( S_0^{he} S_0^{he\dagger} + \Gamma^{he} S_0^{he\dagger} + S_0^{he} \Gamma^{he\dagger} \right) \right\} + \partial_{\varepsilon} \mathrm{Im} \mathrm{Tr} \left\{ \boldsymbol{\sigma}^* \partial_t S_0^{he} S_0^{he\dagger} \right\} \right] \\ + \frac{1}{4\pi} \int_{-\infty}^{\infty} \mathrm{d}\varepsilon \: (-\partial_{\varepsilon} f_e (\varepsilon)) \left[ \mathrm{Im} \mathrm{Tr} \left\{ \boldsymbol{\sigma} \partial_t S_0^{ee} S_0^{ee\dagger} \right\} - \mathrm{Im} \mathrm{Tr} \left\{ \boldsymbol{\sigma}^* \partial_t S_0^{he} S_0^{he\dagger}\right\} \right]. \end{gathered}$$ Introducing the conductance polarization $$\label{eq:48} \tilde{p} \equiv \frac{1}{\tilde{g}} \sum_{m,n} \left\{ \left|\underline{S}% ^{he}_{\downarrow\uparrow,mn} \right|^2 - \left|\underline{S}% ^{he}_{\uparrow\downarrow,mn}\right|^2 \right\},$$ and the generalized mixing conductance[brataas04:\_spin\_and\_charg\_pumpin\_by]{} $$\label{eq:18} \tilde{g}^{\uparrow\downarrow} \equiv \sum_{m,n} \left\{ \delta_{m,n} - \underline{S}^{ee}_{\uparrow,mn} \underline{S}^{ee*}_{\downarrow,mn} + \underline{S}^{he}_{\downarrow\uparrow,mn} \underline{S}^{he*}_{\uparrow% \downarrow,mn} \right\}.$$ we find the following expression for the spin current: $$\begin{gathered} \label{eq:62} \boldsymbol{I}_s(t) = - \frac{1}{4\pi} \int_{-\infty}^{\infty} \mathrm{d}\varepsilon \: (f_e (\varepsilon) - f_h(\varepsilon)) \left( \tilde{p}\tilde{g} \boldsymbol{m}(t) - \mathrm{Tr} \left\{ \boldsymbol{\sigma}^*( \Gamma^{he} S_0^{he\dagger} + S_0^{he} \Gamma^{he\dagger}) \right\} \right) \\ + \frac{1}{8\pi} \int_{-\infty}^{\infty} \mathrm{d}\varepsilon \: (f_e (\varepsilon) - f_h(\varepsilon)) \partial_{\varepsilon} \biggr( \boldsymbol{m} \times \partial_t \boldsymbol{m} \Big(\tilde{g} + 2 \mathrm{Re} \sum_{m,n} \underline{S}^{he}_{\downarrow\uparrow,mn} \underline{S}^{he*}_{\uparrow\downarrow,mn} \Big) + 2 \partial_t \boldsymbol{m} \mathrm{Im} \sum_{m,n} \underline{S}^{he}_{\downarrow\uparrow,mn} \underline{S}^{he*}_{\uparrow\downarrow,mn} \biggr) \\ + \frac{1}{4\pi} \int_{-\infty}^{\infty} \mathrm{d}\varepsilon \: \partial_{\varepsilon} f_e(\varepsilon) \left( \boldsymbol{m} \times \partial_t \boldsymbol{m} \mathrm{Re} \tilde{g}^{\uparrow\downarrow} + \partial_t \boldsymbol{m} \mathrm{Im} \tilde{g}^{\uparrow\downarrow} \right). \end{gathered}$$ The term $\sim \tilde{p}\tilde{g}\boldsymbol{m}(t)$ on the right hand side of Eq.  corresponds to the non-equilibrium bias voltage spin current observed also in the absence of a precessing magnetization vector. Terms in the final line are similar to those derived previously within electro-chemical equilibrium pumping theory for F|N[@spinpumping], and FS|N structures[@brataas04:_spin_and_charg_pumpin_by]. However, we ask the reader to note that the generalized mixing conductance in Eq. (3) in Ref.  is valid for triplet superconductors only; the correct mixing conductance for a singlet superconductor is given by Eq.  above. The remaining terms on the right hand side of Eq.  are non-equilibrium, pumped contributions to the spin current. They depend on pumping parameter frequency via $\partial_t \boldsymbol{m}$ and the $\Lambda$ term from Eq. (\[eq:75\]), which is contained in the gradient remainder $\Gamma^{he}$. Finally, we would like to point out that there are no pumped contributions to the *longitudinal* spin current $I^{||}_{s} \equiv \boldsymbol{m} \cdot \boldsymbol{I}_s$. The terms in the second and third line of Eq. ([eq:62]{}) are transverse with respect to the magnetization $\boldsymbol{m}$, so this leaves only a possible gradient remainder contribution coming from $% \Gamma^{he}$. However, due to the particular matrix structure of $\Gamma^{he} $ mentioned in Sec. \[sec:grad-corr-matr-1\], $\boldsymbol{m} \cdot \mathrm{Tr}\{ \boldsymbol{\sigma}^* \Gamma^{he} S_0^{he\dagger} \}$ vanishes. This observation implies that the longitudinal spin current is stationary and unaffected by the precessing magnetization. Thus, to first order in precession frequency: $$\label{eq:41} I_s^{||} = \boldsymbol{m}(t) \cdot \boldsymbol{I}_s (t) = - \frac{1}{4\pi} \int_{-\infty}^{\infty} \mathrm{d} \varepsilon \: (f_e (\varepsilon) - f_h (\varepsilon)) \tilde{p}\tilde{g}.$$ In the following, we will investigate pumped charge and spin currents when the ferromagnetic region is longer than the typical transverse spin coherence length. Long ferromagnet limit {#sec:limit-1:-long-2} ---------------------- When the length $L_{F}$ of the ferromagnet is longer than the transverse spin coherence length, $$L_{F}>\lambda _{F}\equiv \frac{\pi }{k_{F\uparrow }-k_{F\downarrow }}, \label{eq:39}$$where $k_{F\sigma }$ is the Fermi wave vector of a spin $\sigma $ electron, we expect to find a mixing conductance that is determined by the properties of the N-F subsystem, characterized by the spin-dependent conductances[@spinpumping] $$g^{\sigma \sigma ^{\prime }}=\sum_{m,n}\left( \delta _{m,n}-\underline{r}% _{\sigma ,mn}\underline{r}_{\sigma ^{\prime },mn}^{\ast }\right) . \label{eq:49}$$Indeed, in the limit (\[eq:39\]), one can disregard mixing transmission terms, $\sum_{m,n}\underline{t}% _{\sigma ,mn}\underline{t}_{-\sigma ,mn}^{\ast }\rightarrow 0$, so that $% \sum_{m,n}\underline{S}_{\downarrow \uparrow ,mn}^{he}\underline{S}% _{\uparrow \downarrow ,mn}^{he\ast }\rightarrow 0$. Disregarding interference terms between reflected and transmitted electronic wave functions, one obtains $$\sum_{m,n}\underline{S}_{\uparrow ,mn}^{ee}\underline{S}_{\downarrow ,mn}^{ee\ast }\rightarrow \sum_{m,n}\underline{r}_{11\uparrow ,mn}\underline{% r}_{11\downarrow ,mn}^{\ast }, \label{eq:45}$$for a long ferromagnet. This implies that $\tilde{g}^{\uparrow \downarrow }\rightarrow g^{\uparrow \downarrow }$, while the total conductance $\tilde{g% }$ and the conductance polarization $\tilde{p}$ remain unchanged. Since the mixing conductance is now determined by properties of the N-F structure, energy gradients of the mixing conductance should be disregarded in the limit $\Delta \ll \Delta _{xc},E_{F}$, as described in Sec. \[sec:grad-corr-matr-1\]. Finally, by an explicit calculation, one can show that $\mathrm{Tr}\{\Gamma ^{he}S_{0}^{he\dagger }\boldsymbol{\sigma }^{\ast }\}\sim \underline{t}_{\sigma }\underline{t}_{-\sigma }^{\ast }$, which vanishes when Eq. (\[eq:39\]) holds. To summarize, when the ferromagnet is longer than the transverse spin coherence length, the charge current and longitudinal spin current are still given by $$I_{c}=\frac{e}{2\pi }\int_{-\infty }^{\infty }\mathrm{d}\varepsilon \left( f_{e}(\varepsilon )-f_{h}(\varepsilon )\right) \tilde{g}, \label{eq:11}$$and $$I_{s}^{||}=-\frac{1}{4\pi }\int_{-\infty }^{\infty }\mathrm{d}\varepsilon (f_{e}(\varepsilon )-f_{h}(\varepsilon ))\tilde{p}\tilde{g}, \label{eq:55}$$while the transverse spin current is simplified to $$\begin{gathered} \boldsymbol{I}_{s}^{\perp }(t)=-\frac{1}{8\pi }\int_{-\infty }^{\infty }% \mathrm{d}\varepsilon (\partial _{\varepsilon }f_{e}(\varepsilon )-\partial _{\varepsilon }f_{h}(\varepsilon ))\tilde{g}\boldsymbol{m}\times \partial _{t}\boldsymbol{m} \label{eq:12} \\ +\frac{1}{4\pi }\int_{-\infty }^{\infty }\mathrm{d}\varepsilon \partial _{\varepsilon }f_{e}(\varepsilon )\left( \mathrm{Re}{g}^{\uparrow \downarrow }\boldsymbol{m}\times \partial _{t}\boldsymbol{m}+\mathrm{Im}{g}^{\uparrow \downarrow }\partial _{t}\boldsymbol{m}\right) .\end{gathered}$$With no applied bias voltage, the pumped spin current in Eq. (\[eq:12\]) is identical to that found in N-F systems[@spinpumping], as should be expected. In this situation, emission of spins from the ferromagnet into the normal metal are unaffected by the superconductor. \ To compare the exact result (\[eq:62\]) with the long ferromagnet approximation of Eq. (\[eq:12\]), we plot in Fig. \[fig:plot1\] the spin current along $\partial_t\boldsymbol{m}$ for a ballistic N|F|S trilayer, as a function of the ratio between the ferromagnet length ($L_F$) and the transverse spin coherence length ($\lambda_F$) defined in Eq. (\[eq:39\]). When $L_F \le \lambda_F$, non-negligible “mixing transmission” terms combine with energy gradients of the scattering matrix and produce large deviations between the two equations. As $L_F$ exceeds $\lambda_F$, the fit improves and the exact result oscillates towards the spin current obtained by the approximate Eq. . Conclusion {#sec:conclusion} ========== In conclusion, we have derived non-equilibrium pumped charge and spin currents to first order in pump frequency, using time-dependent scattering theory. Magnetization precession induces transverse spin currents, but neither charge nor longitudinal spin currents, which are both given by their stationary values. The currents are expressed in terms of generalized, spin dependent conductances, that include spin-dependent scattering in the ferromagnet and Andreev reflection at the F|S interface. Finally, we consider trilayers where the ferromagnetic region is longer than the transverse spin coherence length, and derive an approximate expression for the transverse spin current. Numerical calculation of the spin current in a ballistic trilayer shows good agreement between exact and approximate spin currents for ferromagnets whose layer thicknesses exceed the transverse spin coherence length. [Sillanpää et al.(2001)Sillanpää, Heikkilä, Lindell, and Hakonen]{} natexlab bibnamefont bibfnamefont citenamefont url urlprefix \[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} and , ****, (). , , and , ****, (). , , , , and , ****, (). [petrashov99:\_giant\_mutual\_proxim\_effec\_in]{} , , , , and , ****, (). [keizer06:\_spin\_tripl\_super\_throug\_half]{} , , , , and , ****, (). , , , and , . [bergeret01:\_long\_range\_proxim\_effec\_super\_ferrom\_struc]{} , , and , ****, (). [kadigrobov01:\_quant]{} , , and , ****, (). [bergeret05:\_odd\_tripl\_super\_and\_relat]{} , , and , ****, (). [tkachov01:\_subgap\_trans\_in\_ferrom\_super]{} , , and , ****, (). , , , , and , ****, (). , ****, (). , ****, (). , ****, (). [spinpumping]{} , , , and , ****, (). [mizukami01:\_ferrom\_nm\_nm\_nm\_cu]{} , , and , ****, (). [bell08:\_spin\_dynam\_in\_super\_ferrom\_proxim\_system]{} , , , and , ****, (). [morten08:\_proxim]{} , , , , and , ****, (). , , , and , ****, (). and , ****, (). and , ****, (). , ****, (). [brataas06:\_non\_collin\_magnet]{} , , and , ****, (). [buettiker94:\_curren]{} , , and , ****, (). , ****, (). [buttiker06:\_scatt\_theor\_of\_dynam\_elect\_trans]{} and , ** (, ), vol. of **, p. . , ****, (), . , ****, (). and , ** (, ). [vavilov01:\_charg\_pumpin\_and\_photov\_effec]{} , , and , ****, (). , ****, (). and , ****, (). and , ****, (). [tserkovnyak01:\_shot\_noise\_in\_ferrom\_normal\_metal\_system]{} and , ****, (). [moskalets04:\_adiab\_quant\_pump\_in\_presen]{} and , ****, (). [waintal02:\_magnet\_exchan\_inter\_induc\_by\_josep\_curren]{} and , ****, ().

--- abstract: 'With the nearing completion of the first-generation experiments at asymmetric $e^+ e^-$ colliders running at the $\Upsilon(4S)$ resonance (“B-Factories”) a new era of high luminosity machines is at the horizon. We report here on the plans at KEK in Japan to upgrade the KEKB machine (“SuperKEKB”) with the goal of achieving an instantaneous luminosity exceeding $8 \times 10^{35}$ cm$^{-2}$ s$^{-1}$, which is almost two orders of magnitude higher than KEKB. Together with the machine, the Belle detector will be upgraded as well (“Belle-II”), with significant improvements to increase its background tolerance as well as improving its physics performance. The new generation of experiments is scheduled to take first data in the year 2013.' author: - Christian Kiesling title: CP Violation and the Future of Flavor Physics --- [ address=[Max-Planck-Institute for Physics, Munich, Germany]{} ]{} Introduction ============ The Belle Collaboration, together with BaBar, has made essential contributions to establish the theory of Kobayashi and Maskawa, who explain all known CP violation phenomena within the framework of the Standard Model (SM) by a single irreducible phase appearing in the quark mixing matrix. For their outstanding achievement Kobayashi and Maskawa were awarded the Nobel Prize of 2008. Although the SM has been extremely successful in describing virtually all data, most importantly the CP violation phenomena of the $K$- and the $B$-systems (for a comprehensive overview see, e.g. [@buras]), there are a number of arguments why the SM cannot be the regarded as a complete theory. In fact, there is clear evidence for physics beyond the SM, as suggested by the non-vanishing mass of the neutrinos, the matter-antimatter asymmetry observed in the universe, and the apparent necessity for dark matter. Most likely, the effects mentioned have to do with CP violation of a yet undiscovered source. The “New Physics” (NP) generating these sources is expected to appear at large, so far unreached (multi-TeV) energy scales. While the discovery and exploration of New Physics is the central motivation for the LHC program, flavor physics is expected to play a key role in unraveling possible NP at this scale and to solve the puzzle of CP violation. Colliding $e^+$ and $e^-$ beams with different (“asymmetric”) energies to produce the $\Upsilon(4S)$ resonance, with just enough energy for the creation of a pair of $B$-mesons (or background with other quark flavors), is an alternative approach to the high-energy frontier experiments at the LHC. With high luminosity, and consequently large statistics, as achieved in the next generation of flavor factories (“Super Flavor Factory”, SFF) [@superkekb; @superb], very large energy scales can be reached, when quantum loop corrections to the SM are considered (see fig. \[fig:penguin\]). Depending on the flavor changing couplings of the NP particle spectrum, the sensitivity to large mass scales in a SFF may be from many hundred GeV up to tens of TeV. In this respect a SFF is truely complementary to the LHC. A recent review of the physics potential of a future high luminosity B factory can be found, e.g., in [@newphys]. One should note here that the discovery potential of a future SFF is indeed extraordinary and might reach even beyond the LHC. There are several distinct approaches to look for NP at SFF’s: Many of them concentrate on the precise measurement of the angles and sides of the unitarity triangle for $B$-meson couplings, derived from the Cabibbo-Kobayashi-Maskawa (CKM) matrix which connects the mass eigenstates to the flavor eigenstates of the down-type quarks. Within the SM the CKM matrix is unitary and their elements describe the coupling strengths of the flavor changing currents at the quark level, such as $b {\rightarrow}c$ or $b {\rightarrow}u$. The unitarity of the CKM matrix gives rise to a total of six so-called unitarity triangles, one of them involving the $b$-quark couplings (“$B$-triangle”). All triangles have the same area, but the $B$-triangle has all three sides of the same order, corresponding to large angles and thus giving rise to large CP violating effects. Within the SM the $B$-triangle is highly over-constrained (5 observables for only 2 independent quantities), so a precise measurement of all the three angles and the two sides is a crucial check of the validity of the SM: If the $B$-triangle “does not close”, New Physics must be the reason. One should mention that the all present measurements of the CKM unitarity are in agreement with the SM, although a few “tensions” have become noticeable (for details see, e.g.  [@CKMfitter]). Another way of gaining sensitivity at the SFF to NP is the study of rare decays of $B$ mesons and $\tau$ leptons. Some examples should illustrate this point, such as $B \rightarrow X_d \nu \bar{\nu}$ or $\tau \rightarrow \mu \gamma$. These decays involve (several) neutral particles in the final state and can therefore only be measured at SFF’s. Such decays are highly suppressed (in the $B$ case) or even completely forbidden (the $\tau$ case). With a SFF, branching fractions down to several $10^{-9}$ can be probed. More details on rare decays and their potential to search for New Physics can be found in the proposals for the SFFs [@superkekb; @superb]. ![\[fig:penguin\]Example of a SM process at the quantum loop level (“penguin diagram”, left) with “New Physics” contributing (right) to $B$-meson decay amplitudes.](Loop_Diagram){height=".17\textheight"} One of the flagship measurements at the SFF is the precise determination of the time-dependent CP violating asymmetries. These asymmetries are measured by observing the decay rate $\Gamma$, as function of time, of a ${B^0}$ meson decaying into a specific CP eigenstate $f_{CP}$, as compared to the same final state coming from the ${{\bar{B}}^0}$. The CP violating time-dependent asymmetry is defined as: $${\cal A}(f_{CP},\Delta t) =\frac{\Gamma({{\bar{B}}^0}{\rightarrow}f_{CP};\Delta t)-\Gamma({B^0}{\rightarrow}f_{CP};\Delta t)}{\Gamma({{\bar{B}}^0}{\rightarrow}f_{CP};\Delta t)+\Gamma({B^0}{\rightarrow}f_{CP};\Delta t)},$$ where $\Gamma$ is the decay rate of the ${B^0}({{\bar{B}}^0})$ into the CP eigenstate $f_{CP}$ (“CP side”) within some time interval $\Delta t$, to be explained below. In order to determine which of the two flavors (${B^0}$ or ${{\bar{B}}^0}$) has decayed into the common final state $f_{CP}$, the other $B$-decay (“tag side”) is analyzed for a specific flavor, using, e.g., semi-leptonic decays (${B^0}{\rightarrow}X l^- \bar{\nu}, {{\bar{B}}^0}{\rightarrow}X l^+ \nu$). The charge of the lepton uniquely identifies (“tags”) the flavor of the $B$ meson, where a positive (negative) lepton signals a ${{\bar{B}}^0}$(${B^0}$). Since the two $B$ mesons are produced in an entangled state by virtue of the quantum numbers of the $\Upsilon (4S)$, the tag side uniquely[^1] determines the flavor of the $B$-meson that decayed into the CP eigenstate. The time difference $\Delta t$ is given by the difference in decay times between the tag side and the CP side. Note that $\Delta t$ can be positive or negative. Depending on the CP eigenstate chosen, any of the three angles $\phi_1 (\beta), \phi_2(\alpha)$ or $\phi_3 (\gamma)$ of the $B$-triangle can be measured. Chosing, e.g., the final state ${J/\psi}K^0$, the angle $\phi_1$ (or $\beta$) is determined. A recent measurement [@jpsiks] of the time-dependent CP asymmetry for the ${J/\psi}K^0$ channel is shown in fig. \[fig:jpsiksacp\]. Here, the asymmetry is given for both odd (${J/\psi}{K_S}$) and even (${J/\psi}{K_L}$) CP final states. Since the $\Delta t$ distributions for the CP eigenstates are different, CP violation is established. Recent summaries on the measurements of the angles $\phi_1, \phi_2, \phi_3$ ($\alpha, \beta, \gamma$) have been presented at this conference [@angle_summ]. ![Recent measurements of the time-dependent CP asymmetry using the final states ${J/\psi}{K_S}$ and ${J/\psi}{K_L}$ (see [@jpsiks]).[]{data-label="fig:jpsiksacp"}](jpsiksacp){height=".3\textheight"} Machine Upgrade: SuperKEKB ========================== While no significant deviations from the SM predictions have been observed so far, there are some tantalizing hints for possible New Physics in $B$ decays (see, e.g., [@hints]). Clarification can only come with a new generation of SFF’s, which should aim at integrated luminosities in excess of 50 /ab (the present world record KEKB accelerator is about to accumulate 1 /ab). Such large integrated luminosities require an equally large increase of the instantaneous luminosity ${\cal L}$ which, in its simplified form, is given by $${\cal L}=\frac{N_1N_2f}{4\pi\sigma_x\sigma_y}.$$ Here, $N_i$ are the numbers of particles in each of the two colliding bunches, $f$ is the bunch collision frequency, and $\sigma_{x,y}$ are the transverse dimensions of the colliding bunches. At KEK, an extremely strong accelerator research program is focussing on an asymmetric $e^+e^-$ collider with instantaneous luminosities in excess of 8 $\times 10^{35}$ cm$^{-2}$ s$^{-1}$ (which is about 40 times the present world record luminosity of 2.11 $\times 10^{34}$ cm$^{-2}$ s$^{-1}$, reached in May 2009 with KEKB). The new machine, called “SuperKEKB”, is an upgrade of the present KEKB machine and should start producing luminosity by the year 2013. According to the current plan the KEKB accelerator should stop running by the end of 2009, so that the construction work for SuperKEKB can start in 2010. The expected luminosity development of the SuperKEKB machine is shown in fig. \[fig:lumidev\]. ![Expected development of the luminosity for the SuperKEKB machine.[]{data-label="fig:lumidev"}](Luminosity){height=".35\textheight"} Two options for the SuperKEKB machine have been discussed: The “high-current” (HC) option and the “nano-beam” (NB) option (see Table \[tab:machine\]). Initially, the HC option was favored. It was characterized by a mild decrease of the beta function in the low energy ring (LER), and a dramatic increase of the beam current in the LER and the high energy ring (HER). In addition, an ingenious scheme was developed to counteract the luminosity loss due to the finite angle under which the two colliding bunches cross each other, i.e. the crab crossing scheme. In this scheme special crab cavities before and after the interaction region rotate the bunches, so that they collide head-on, instead of at an angle. The crab crossing scheme was verified with the running KEKB accelerator, increasing the instantaneous luminosity by about 20 percent to a new world record. However, it became apparent that a new effect would make the HC option difficult: For the crab crossing to work efficiently the bunch length must be smaller than the $\beta$ function at the interaction point (IP), i.e. $\sigma_z < \beta^*_y$. This condition, imposed to avoid the so-called hour-glass effect, creates no problems at low beam currents. However, at the large beam currents of the HC option (see Table \[tab:machine\]) coherent synchrotron radiation becomes an issue which has the effect of lengthening the bunch, thus running into the hour-glass effect. The unavoidable bunch lengthening leads to a decrease of the obtainable maximum luminosity to slightly over 50 $\times 10^{34}$. Failing the goal of ${\cal L} \ge 80 \times 10^{34}$, this option is disfavored now. Following the ideas of the final focus system envisaged for a future linear collider, transferred to a circular machine as laid down in [@superb], the NB option has become the baseline [@machine_par]. Here, the key parameters are a factor of two increase in the beam currents with respect to the present KEKB, but strongly reduced $\beta$ functions at IP (factor of about 20 smaller than in the HC option). In addition, low emittance beams are necessary to achieve the desired small (less than 100 nano meters) transverse bunch sizes. Such a beam size has been achieved at the ATF damping ring facility at KEK. While a low emittance electron beam can be prepared by virtue of a carefully designed injection system, the transverse phase-space of the positrons needs to be cooled in a special new damping ring. ------------------------------------------------------------------ --------- -------------------------------------------- ------------------- --------------- $\beta^*_y$ (mm) (LER/HER) 10 / 10 $\begin{array}{c} 6.5 / 5.9 \\ 3 / 6 0.26 / 0.26 (5.9 / 5.9) \end{array}$ $\epsilon_x$ (nm) 18 / 18 18 / 24 24 / 18 2.8 / 2.0 $\sigma_y$ ($\mu$m) 1.9 1.9 (0.94) 0.85 / 0.73 0.073 / 0.097 $\xi_y$ 0.052 $\begin{array}{c}0.108 / 0.057 \\ (0.129 / 0.3 / 0.51 0.079 / 0.079 0.090) \end{array}$ $\sigma_z$ ($\mu$m) 4  7 5 (LER) / 3 (HER) 5 $I_{\rm beam}$(A) 2.6/1.1 $\begin{array}{c}1.66 / 1.34 \\ (1.64 / 9.1 / 4.1 3.84 / 2.21 1.19) \end{array}$ $N_{\rm bunches}$ 5000 1388 (1585) 5000 2252 $\begin{array}{l}{\rm Luminosity} \\ (10^{34} {\rm cm}^{-2} {\rm 1 1.76 (2.11) 53 80 s}^{-1}) \end{array} $ ------------------------------------------------------------------ --------- -------------------------------------------- ------------------- --------------- : Comparison of parameters for the two high luminosity options of the SuperKEKB machine, compared to the presently running KEKB factory. The high current option for SuperKEKB has been discarded recently (see text) and the nano-beam option is now the baseline.[]{data-label="tab:machine"} Table \[tab:machine\] gives an overview of the machine parameters for the presently running KEKB accelerator, and for both the HC and NB options (status of July 2009). Note that the presently achieved luminosity of KEKB is a factor of two larger than the original design value. Part of it is due to the new crab crossing scheme, which is being successfully tested with the running machine. It should be noted, however, that the luminosity increase predicted in simulation for the crab crossing scheme has not yet been reached in the real machine. The discrepancy is attributed to so-called “machine errors” which may result from a finite precision in aligning the focussing quadrupoles along the ring. Small (unknown) deviations from the ideal position can lead to a coupling of the horizontal and vertical betatron oscillations and therefore to an increase in the vertical beam size at the IP, reducing the luminosity. During the last winter shutdown some errors of this kind have been diagnosed and corrected for by installing a set of so-called “skew sextupoles” around the ring. As a consequence, the luminosity could be immediately increased by more than 15 percent, leading to new world records for the instantaneous luminosity (see table \[tab:machine\]). Preliminary parameters of both collision schemes for the SuperKEKB machine are given in the table [@machine_par]. Intense optimization studies for the new machine are going on, concentrating now mainly on the NB option. One should note, however, that the parameters given in the table have not yet been consolidated. In addition, the so-called crab waist scheme, as proposed in [@superb], has not yet been included in the present design. While the crab waist scheme is expected to improve the instantaneous luminosity only by a few percent, it is very effective in reducing the coupling between vertical and horizontal betatron oscillations and is therefore quite useful to stabilize the machine operation. Some of the virtues of the NB option are quite evident: There is no need to have short bunches (the condition $\sigma_z < \beta^*_y$ vital for the HC option to avoid the hour glass effect does not apply here), and the beam-beam parameter $\xi_y$ can be relaxed substantially (it should be noted that the large $\xi_y$ parameter required for the HC option has never been reached so far). Also the synchrotron radiation is less of an issue due to the smaller beam currents. This is certainly good news for the detectors close to the beampipe (e.g. the silicon detectors). On the other hand, a new source of background may arise from the Touschek effect [@touschek], a kind of intra-beam Coulomb scattering, which couples betatron and synchrotron oscillations. Due to this effect particles may be sent to off-momentum orbits, causing particle loss and a severe decrease of the beam lifetime. To counteract the shorter lifetimes (order of a few minutes) an elaborate scheme is envisaged for SuperKEKB which keeps the bunch currents almost constant by permanent re-injection at a rate of about 50 Hertz. With the NB option another important physics parameter may be chosen more favorably, i.e. the radius of the beampipe, which may be as small as roughly 1 cm, whereas in the HC option 1.5 cm had to be chosen. It is expected that the final design for the SuperKEKB machine in its nano-beam version will be completed in the fall of 2009. Detector Upgrade: BELLE-II ========================== Parallel to the machine studies, a Letter of Intent [@belleloi] for the Belle upgrade was issued in the year 2004, with a recent update from 2008 [@sbelle]. The main (re)design goals are to cope with the much higher physics rates and the much larger backgrounds to be expected, as well as improving the overall physics performance. A comparison of Belle and its upgraded version (“Belle-II”) is shown in a sideview in fig. \[belleii\]. We will first give a short overview of the different detector components and then pick out two specific systems which will be very important for the precise measurement of the CP quantities. As a general side condition, the performance of the new Belle-II detector should be as good or better than Belle. This is a non-trivial requirement in view of the anticipated very high background at the SuperKEKB machine, which is estimated to be roughly an order of magnitude larger as compared to the present KEKB machine (for details, see [@sbelle]). ![Upgrade scenario from the Belle detector (lower half) to the Belle-II detector (upper half).[]{data-label="belleii"}](Belle_BelleII){height="0.47\textheight"} The tracking system in the “old” Belle detector consisted of 4 layer of Si strip vertex detectors (SVD), followed by a Central Drift Chamber (CDC). Due to the largely increased background, strip detectors are no longer an option for the innermost Si layers. Instead a two-layer pixel detector (PXD) for the innermost Si layers is planned (see below). The SVD will be replaced entirely as well as the CDC: due to the harsh backgrounds the inner radius of the CDC has to be moved out and the two outer layers of the new SVD will cover the gap. The momentum resolution of charged particles will be improved by extending the CDC to a larger radius. Since the magnet and the barrel part of the electromagnetic calorimeter (ECL) will not be changed, the particle identification system has to be replaced by a thinner, low material budget detector (see below). In the endcap parts of the ECL the present CsI(Tl) crystals will be replaced by pure CsI that provide faster signals, and the forward part of the KLM (${K_L}$ and muon detector) in the iron flux return yoke will be instrumented by scintillator strips with SiPM readout, replacing the present RPCs. To extend the physics reach for Belle-II, the $K/\pi$ separation ability will be improved by a new particle identification (PID) system. Two types of detectors are being proposed, one is a time-of-propagation (TOP) counter for the barrel region (Barrel PID), and the other a proximity-focussing Cherenkov ring imaging counter with aerogel radiators (ARICH) for the endcap region (Endcap PID). ![Comparison of the PID system performance of Belle (left) and Belle-II (right) for the decay $B^0 {\rightarrow}\rho^0 \gamma$. The abscissa is $\Delta E= {E_{\rm beam}}-{E_{B^0}}$ (in GeV) in the $\Upsilon(4S)$ center-of-mass system (see text).[]{data-label="fig:pid"}](PID){width="\textwidth"} The present time-of-flight and aerogel Cherenkov counters in the barrel region of Belle are replaced by a TOP counter made from quartz radiator bars (thickness 2 cm), in which the time of propagation of Cherenkov photons is measured, which are internally reflected and focussed onto micro-channel plate (MCP) PMTs at the end surfaces of the quartz bars. The MCP-PMTs have an excellent time resolution of about 50 ps, so that the difference in arrival times of the Cherenkov photons, radiated by pions or kaons, can be determined. This time difference results from the different Cherenkov emission angles (for equal particle momenta) and consequently different path lengths for the photons which undergo multiple internal reflections on their way to the ends of the quartz bars. The ARICH counters are located in front of the endcap ECL, where the space is quite limited. For this reason a proximity focussing scheme is envisaged for the Cherenkov photons, with an expansion thickness of only 20 cm. In the present design three layers of silica aerogel, each 10 mm thick, are foreseen, with refractive indices varying between 1.045 and 1.055, so that the photons emitted from the three regions produce overlapping images on the photon detector surface. Since the photon detectors have to work in a strong magnetic field of 1.5 Tesla, candidates are a hybrid avalanche photon detector (HAPD) or a MCP-PMT. An example of the expected performance of the new PID system relative to the present Belle ACC system is shown in fig. \[fig:pid\]. Here, the decay $B^0 {\rightarrow}\rho^0 \gamma$ is studied at the $\Upsilon(4S)$ resonance and the distribution of the quantity $\Delta E={E_{\rm beam}}-{E_{B^0}}$ is chosen for reference, where ${E_{\rm beam}}$ is the beam energy in the $\Upsilon(4S)$ center-of-mass system and ${E_{B^0}}$ is the energy of the $B$ meson, reconstructed from the two pions forming the $\rho$ and the photon. An overwhelming background from $B {\rightarrow}K^* \gamma$ is expected due to the CKM couplings (about a factor 40). While for the case of Belle (left side) the background from misidentified kaons is very large, the improved PID in Belle-II (right) largely reduces the background and shows a much clearer signal. ![Principle of operation of the DEPFET pixel sensor.[]{data-label="fig:depfet"}](depfet_schematics){width="27.00000%"} A second example for the improved instrumentation at Belle-II is the new two-layer Si pixel detector, located closest to the beampipe, for precise vertexing. As mentioned above, the expected strong increase of background relative to the present KEKB machine, proportional to $1/r^2$ where $r$ is the radial distance from the interaction region, creates large occupancies within the strip detectors, making an efficient reconstruction of tracks and vertices impossible. The solution is to use a pixel detector which intrinsically provides three-dimensional space points. However, pixel segmentation is not the only requirement to be fulfilled at Belle-II. Due to the small (transverse) momenta of the $B$ decay products, the momentum and vertex resolutions are dominated by multiple scattering, so a pixel sensor should have a very low material budget. Furthermore, the harsh radiation environment at the SuperKEKB factory requires low-noise and radiation-hard technology, and the limited space between the beampipe and the strip detector requires a low power consumer. Finally, a pixel detector should be operational from the start of the Belle-II running, which is envisaged for the year 2013. It turns out that there is only one mature technology at present fulfilling all these requirements: The DEPFET pixel sensor [@depfet], invented and developed at the Max-Planck-Institute for Physics (MPI) in Munich, Germany. The name DEPFET stands for $DEP$leted $F$ield $E$ffect $T$ransistor and combines detection and amplification. The DEPFET principle of operation is shown in fig. \[fig:depfet\]. A MOS field effect transistor is integrated onto a fully depleted silicon substrate forming the detector. By means of an additional $n$-implant underneath the transistor channel a potential minimum for electrons is created. This potential well can be considered as an internal gate of the transistor. A particle entering the detector creates electron-hole pairs in the fully depleted silicon substrate. The electrons (“signal”) are collected and stored in the internal gate. The electron charges change the potential of the internal gate, resulting in a modulation of the transistor channel current. After readout of the channel current, the signal electrons are removed by a positive voltage at the clear contact. Another read cycle then establishes the baseline which is subtracted from the current readout before the clear. The whole readout cycle (read-clear-read) takes about 80 ns. Because of the small capacitance of the internal gate, the noise in the DEPFET is very low, about 100-200 electrons are expected when operating at Belle-II. For a real detector a matrix of DEPFET pixels must be constructed. Such matrices, using the SOI technique, where the sensor wafer is bonded to a “handling” wafer, have already been built and subjected to extensive beam tests. Matrices up to 512 $\times$ 512 pixels have been obtained, with pixel sizes of about 17 $\times$ 13 $\mu$m$^2$. For the DEPFET pixel detector (“PXD”) at Belle-II typical pixel sizes would be around 50 $\times$ 50 $\mu$m$^2$. An additional asset of the DEPFET technology is the fact that the detector can be thinned down to about 50 $\mu$m thickness. The thinning procedure of the handling wafer has already been demonstrated at the MPI semiconductor laboratory. With such a thin sensor still a signal to noise ratio of about 40:1 can be reached. Finally, the radiation hardness of the DEPFET matrix has been tested and found to be satisfactory for a few years of running at SuperKEKB with full luminosity. Further details of the DEPFET project at Belle-II can be found elsewhere [@depfetpage]. Conclusions =========== Complementary to the LHC, a new generation of Super Flavor Factories is being planned, probing the Standard Model at energy scales beyond tens of TeV. Flavor physics continues to be the key to the puzzle of the observed matter-antimatter asymmetry in the universe, intimately connected to CP violation. The SuperKEKB project, together with a substantially upgraded detector, Belle-II, will contribute to this fascinating chapter of particle physics, taking first data in the year 2013, with the prospect of accumulating 50 times the presently available data by the year 2020. The author would like to acknowledge the extremely stimulating and pleasant atmosphere created by the organizers, especially Marvin Marshak, of the CIPANP 2009 Conference. He also wishes to thank Tom Browder, Yoshihiro Funakoshi, Katsunobu Oide, and Yoshi Sakai for their careful reading of the manuscript and their valuable advice. [^1]: strictly speaking, the quantum entanglement is broken when the first of the $B$-mesons decays. From then on the other $B$-meson is freely oscillating between ${B^0}$ and ${{\bar{B}}^0}$, but with a known time dependence.

--- abstract: 'We present a novel predictive model architecture based on the principles of predictive coding that enables open loop prediction of future observations over extended horizons. There are two key innovations. First, whereas current methods typically learn to make long-horizon open-loop predictions using a multi-step cost function, we instead run the model open loop in the forward pass during training. Second, current predictive coding models initialize the representation layer’s hidden state to a constant value at the start of an episode, and consequently typically require multiple steps of interaction with the environment before the model begins to produce accurate predictions. Instead, we learn a mapping from the first observation in an episode to the hidden state, allowing the trained model to immediately produce accurate predictions. We compare the performance of our architecture to a standard predictive coding model, and demonstrate the ability of the model to make accurate long horizon open loop predictions of simulated Doppler radar altimeter readings during a six degree of freedom Mars landing. Finally, we demonstrate a 2X reduction in sample complexity by using the model to implement a Dyna style algorithm to accelerate policy learning with proximal policy optimization.' bibliography: - 'PCM.bib' --- Introduction ============ A model capable of accurate multi-step prediction prediction over long horizons has the potential to reduce the sample complexity of reinforcement learning as compared to model free methods. For example, such a model would allow an agent to plan by estimating the effect of some sequence of actions. Alternately, the model can be used in a Dyna [@sutton1990integrated] style algorithm to generate rollouts starting from observations sampled from a replay buffer. A fundamental trade-off exists when learning a model suitable for model-based reinforcement learning. Complex dynamics necessitate a complex model, which is prone to overfitting without large amounts of data. However, for model-based RL, we want the model to learn a good representation of the dynamics quickly to reduce sample complexity. If learning the model is as difficult as learning a model-free policy, we have gained nothing (unless the model can be re-used for other tasks). In this work we develop a predictive model using the principles of predictive coding [@rao1999predictive], which was originally developed to explain endstopping in receptive fields of the visual cortex. An action conditional predictive coding model maps from a history of prediction errors and actions to the predicted observation $f: \mathbf{e}_{0:t}, \mathbf{a}_{0:t} \mapsto \hat{\mathbf{o}}_{t+1}$, where $\mathbf{e}_{t+1}=\hat{\mathbf{o}}_{t+1} - \mathbf{o}_{t+1}$, and $\mathbf{e}_0=\mathbf{0}$. When the trained model is run open loop to make predictions, the error feedback signal $\mathbf{e}$ is set to zero, i.e., the model implements the function $f: \mathbf{0}, \mathbf{a}_{0:t} \mapsto \hat{\mathbf{\hat{o}}}_{t+1}$, which is equivalent to using the predicted observation in place of the measured observation to compute the error feedback. In contrast, most action conditional models reported in the literature implement a mapping from a history of observations and actions to a predicted observation $f: \mathbf{o}_{0:t}, \mathbf{a}_{0:t} \mapsto \hat{\mathbf{o}}_{t+1}$. The model can then be used to predict a trajectory open loop given some initial observation and a sequence of actions by feeding back the predicted observation into the model $f: \mathbf{\hat{o}}_{0:t}, \mathbf{a}_{0:t} \mapsto \hat{\mathbf{o}}_{t+1}$, where in model-based RL, the action would be output by some policy $\pi: \mathbf{o} \mapsto \mathbf{a}$. Our model borrows from the PredNet architecture [@lotter2016deep], which applies modern deep learning techniques to predictive coding, but is modified to make the model action conditional, and implements two key innovations. To understand the first improvement, first consider that a predictive coding model (PCM) does not have access to the observations, but only to the prediction error. Consequently, it typically takes several steps of closed loop interaction with the environment for a trained PCM to begin making accurate open loop predictions, which can be a problem in some applications. Our solution is to learn a mapping from the initial observation in an episode to the model’s recurrent network hidden state $f: \mathbf{o_0} \mapsto \mathbf{h_0}$. This effectively bootstraps the hidden state, allowing the model to immediately begin making accurate predictions from the onset of an episode. Concretely, the model implements a mapping $f: \mathbf{o}_0, \mathbf{e}_{0:t}, \mathbf{a}_{0:t} \mapsto \hat{\mathbf{o}}_{t+1}$. The second improvement concerns the forward pass through the model’s network during training. If the network is unrolled for T steps to implement back propagation through time, we input the prediction error $\mathbf{e}$ only on the first step, and set $\mathbf{e}=\mathbf{0}$ for steps 2 through T. This forces the network to learn a representation that is good for multi-step prediction, even when using a conventional single step cost function. Open loop prediction of trajectories is accomplished by initializing the recurrent network’s hidden state using the learned mapping from the initial observation at the start of the episode, and setting the prediction error $\mathbf{e}$ to zero, i.e., the trajectory is generated using the mapping $f: \mathbf{o}_0, \mathbf{0}, \mathbf{a}_{0:t} \mapsto \hat{\mathbf{o}}_{t+1}$. We demonstrate through a series of experiments that even without our enhancements, action conditional predictive coding allows accurate open-loop trajectory prediction of high dimensional trajectories [^1] over long horizons, and that our refinements further enhance this accuracy. We attribute the high performance of predictive coding to the fact that a predictive coding model must make predictions using only a history of prediction errors and actions (as opposed to observations and actions). Since the model must integrate a history of prediction errors, minimizing the loss requires the model to learn important temporal dependencies. In these experiments we use two versions of a Mars powered descent phase environment. The first version is a simple 3 degree of freedom (3-DOF) environment that is used to compare a standard PCM to a model with our enhancements. Although only 3-DOF, the environment is realistic with respect to the lander capability, dynamics, and deployment ellipse. The model is then tested in a 6-DOF version of the Mars landing environment. In one experiment we demonstrate the model’s ability to predict the lander’s $\mathbb{R}^{13}$ trajectory open loop over an entire episode, whereas in the second experiment the model must predict simulated Doppler radar altimeter readings open loop during the descent, with the simulated readings generated using a digital terrain map of the Uzboi Vallis region on Mars. We then couple the model with an agent implementing proximal policy optimization (PPO) [@schulman2017proximal] and demonstrate a reduction in sample complexity in the 3-DOF landing task using a Dyna style algorithm. To our knowledge, this is the first demonstration of a sample efficiency improvement of PPO using concurrent training of policy and model, as previous work has focused on accelerating Q learning. Although [@nagabandi2018neural] used TRPO, the training had two distinct phases of model-based and model free learning, with model predictive control used in the 1st phase. In addition, most previous work in model-based acceleration has assumed that a ground-truth reward function is known, e.g., see [@nagabandi2018neural], [@gu2016continuous], and [@feinberg2018model]. In contrast, our model learns a reward function using a separate reward head in the model network. Preliminaries ============= Our goal is for the model to learn a representation that is effective for making long-term open loop predictions of high dimensional trajectories. In this work the model will learn to predict future observations in a reinforcement learning setting where an agent interacts with an environment and learns to accomplish some task, with the learning driven by a scalar reward signal. The agent will instantiate a policy $\pi: \mathbf{o} \mapsto \mathbf{a}$ that maps observations to actions, and the model learns a representation by observing the sequence of observations and actions induced by the episodic interaction between the agent and environment. Ideally, after a limited amount of experience observing this interaction, the trained model will have the ability to accurately predict future observations from some sequence of on-policy actions, operating open loop. This ability to “imagine” the long-term consequences of some sequence of actions will then be a powerful tool for planning in model-based reinforcement learning, or alternatively, could be used to accelerate model-free algorithms by augmenting the rollouts collected via interaction with the environment with simulated rollouts from the model. We will evaluate our model using two criteria. The first is the ability of the model to be able to generate accurate extended horizon trajectory predictions given some sequence of actions, with the policy that is generating actions having access to the ground truth observation. The second criterion is whether the model can significantly accelerate proximal policy optimization (PPO), a model-free policy gradient with baseline algorithm. Related Work ============ Recent work in developing predictive models include [@finn2017deep], where a model learns to predict future video frames by observing sequences of observation and actions. The model is then used to generate robotic trajectories using model predictive control (MPC), choosing the trajectory that ends with an image that best matches a user specified goal image. The action conditional architecture of [@oh2015action] has proven successful in open loop prediction of long sequences of rendered frames from simulated Atari games. Predictive coding [@lotter2016deep] has been been applied to predicting images from objects that are sequentially rotated, and has been used to predict steering angles from video frames captured from a car mounted camera. [@wang2017predrnn] and [@wang2017predrnn] extend the PredNet architecture described in [@lotter2016deep]. As for using predictive models to accelerate reinforcement learning, in [@nagabandi2018neural], the authors use an action conditional model to predict the future states of agents operating in various openAI gym environments with high dimensional state spaces, and use the model in a model predictive control algorithm that quickly learns the tasks at a relatively low level of performance. The model is then used to create a dataset of trajectories to pre-train a TRPO policy [@schulman2015trust], which then achieves a high level of performance on the task through continued model-free policy optimization. The authors report a 3 to 5 times reduction in sample efficiency using the combined model free and model based algorithm. The approach assumes that a ground truth reward function is known. In [@gu2016continuous], the authors develop a normalized advantage function that gives rise to a Q-learning architecture suitable for continuous high dimensional action spaces. They then improve the algorithm’s sample efficiency by adding “imagination rollouts” to the replay buffer, which are created using a time varying linear model. Importantly, the authors state that their architecture had no success using neural network based models to improve sample efficiency, and assume a ground-truth reward function is available. [@feinberg2018model] accelerates learning a state-action value function by using a neural network based model to generate synthetic rollouts up to a horizon of H steps. The rewards beyond H steps are replaced by the current estimate of the value of the penultimate observation in the synthetic rollouts, as in an H-step temporal difference estimate. The accelerated Q function learning is then used to improve the sample efficiency of the DDPG algorithm [@lillicrap2015continuous]. Again, the algorithm assumes that the ground-truth reward function is available. Learning to predict over long horizons ====================================== Although a model can easily learn a representation mapping observations and actions at step $t$ to a predicted next observation at step $t+1$ (one-step predictions), these representations typically do not capture the underlying dynamics of the system. The equations of motion encoding the laws of physics typically constrain states in configuration space that are close to each other in time to be numerically close in value, and the observations that are functions of these states (such as a sequence of video frames) will by similar for two consecutive frames. Consequently, the model’s network will tend to learn a trivial mapping that does not capture the underlying dynamics. When such a network is run forward open loop to make multi-step predictions (as required when using such a model for planning), the prediction accuracy falls of rapidly as a function of the number of steps. Using Differences in Observations as Training Targets ----------------------------------------------------- Some environments (such as many of the open AI Mujoco environments) use dynamical systems where the system dynamics are dominated by rigid body rotations. For example, the half-cheetah environment has an observation space $S \in \mathbb{R}^{17}$, but the translational motion is constrained to $\mathbb{R}^2$ (position and velocity along a line). In these dynamical systems, having an action conditional model’s network use the difference between consecutive observations (deltas) as a training target can lead to a representation with improved multi-step prediction performance (see for example [@nagabandi2018neural]). Concretely, the training target is given as $o_{t+1}-o_{t}$ for all $t$ in the rollouts. In our work applying RL to aerospace problems, we have found that using deltas as targets fails for environments where translational motion plays a more important roll in the system dynamics. We postulate that the reason using deltas as targets works well for rotational dynamics is that the mapping between torques and rotational velocities using Euler’s rotational equations of motion has sufficient complexity to insure that the model does not learn an identity mapping between components of the current observation and components of the deltas. On the other hand, the differential equations governing position, velocity, and commanded thrust are such that using deltas actually makes matters worse, as the delta velocity component of the training target is approximately proportional to the force (the action input) and the delta position component of the training target is proportional to the velocity input to the model. k-step prediction loss functions -------------------------------- Another common approach to improving multi-step prediction is to use a K-step prediction loss function [@oh2015action]. Consider training data accumulated over $M$ episodes, with $T$ steps in a given episode. Then the K-step loss function is as shown below in Equation \[eq:kstep\_loss\], with the more commonly used 1-step loss given in \[eq:1step\_loss\]. Minimizing the K-step loss function requires the model to learn a representation that is good at multi-step predictions. However, this approach also increases the size of the data used for a model update by a factor of K, as each sample from the rollouts is used to produce a K-step open loop prediction. Moreover, [@oh2015action] required curriculum learning [@bengio2009curriculum], with the value of K was increased once a given value of K converged. $$\begin{aligned} L(\theta)=\frac{1}{2MTK}\sum_i^M\sum_t^T\sum_{k=1}^K\|\hat{\mathbf{o}}_i^{(t+k)}-\mathbf{o}_i^{(t+k)}\|\label{eq:kstep_loss}\\ L(\theta)=\frac{1}{2MT}\sum_i^M\sum_t^T\|\hat{\mathbf{o}}^{(i)}-\mathbf{o}^{(i)}\|\label{eq:1step_loss} \end{aligned}$$ Open Loop Forward Pass ---------------------- Recently, we have found a new method to improve multi-step prediction performance in predictive models that use one or more recurrent layers. First, we will review a common approach to implementing the forward pass in a network with one or more recurrent layers that allows parallel computation over a sequence where we want to preserve the temporal dependencies. Prior to the recurrent layer(s), the network unrolls the output of the previous layer, reshaping the data from $\mathbb{R}^{m\times n}$ to $\mathbb{R}^{T\times (m/T)\times n}$, where $m$ is the batch size, $n$ the feature dimension, and $T$ the number of steps we unroll the network. We then implement a loop from step 1 to step $T$, where at step one, we input the hidden state from the rollouts to the recurrent layer, but for all subsequent steps up to $T$, we let the state evolve according to the current parameterization of the recurrent layer. After the recurrent layer (or layers), the recurrent layer output is then reshaped back to $\mathbb{R}^{m\times n}$. Now consider a modification to this approach where we unroll the entire network from the network’s inputs to the final recurrent layer, as opposed to just unrolling the recurrent layers. For this example, we will assume a predictive coding architecture. At the first step of the loop over the T steps the network is unrolled, we input the prediction error $\mathbf{e}$ and action $\mathbf{a}$ from the rollouts to the network’s first layer. However, on subsequent timesteps, we set $\mathbf{e}$ to zero, making the remainder of the unrolling open loop, in that there is no feedback from the prediction error captured in the rollouts. Consequently, to minimize the 1-step cost function, the network must learn a representation that is useful for multi-step prediction. This approach is also applicable to action conditional models. Here we need to unroll the entire network so that we can feed back the predicted observation as the observation for steps 2 through T. This approach has the advantage over a k-step prediction loss function in that it does not increase the training set size by a factor of K. Moreover, in contrast to [@oh2015action], this method does not require curriculum learning. Predictive Coding Model used in Experiments {#section:pcm_exp} =========================================== Our model’s operation over a single episode is listed in Algorithm \[alg:pcm\], where $\pi: \mathbf{o} \mapsto \mathbf{a}$ is a policy, $\mu: \mathbf{e},\mathbf{a} \mapsto \mathbf{\mathbf{\hat{o}}}, v$ is our model, with both operating in environment $env$. The model consists of four major components. A representation network maps the previous prediction error $\mathbf{e}$ and current action $\mathbf{a}$ to a representation $\mathbf{r}$ (Eq. \[eq:pcm\_r\]), and consists of a fully connected layer followed by a recurrent layer implemented as a gated recurrent unit (GRU)[@chung2015gated] with hidden state $\mathbf{h}$. A prediction network consists of two fully connected layers, and maps the representation $\mathbf{r}$ to a predicted observation $\mathbf{\hat{o}}$ (Eq. \[eq:pcm\_p\]). In an application using high dimensional observations, the prediction network would contain several deconvolutional layers mapping from the representation to a predicted image $\mathbf{\hat{o}}$. The recurrent state initialization network learns a mapping from an observation at the start of an episode $\mathbf{o}_0$ to an initialization for hidden state $\mathbf{h}$ in the representation network $\mathbf{h}$ (Eq. \[eq:pcm\_i\]). This allows the trained model to immediately produce accurate predictions. Finally the value prediction network maps the representation $\mathbf{r}$ to $v$, an estimate of the sum of discounted rewards that would be received by starting in representation $\mathbf{r}$ and following policy $\pi$ (Eq. \[eq:pcm\_v\]). Since in theory a value function could be implemented in an external network using representation $\mathbf{r}$ as input, the primary purpose of the value network is to solve the “vanishing bullet” problem when the model is used with visual observations. The vanishing bullet problem, coined for the Atari space invaders environment, occurs when a very important visual feature consists of relatively few pixels. In this case, a model using an L2 loss between predicted and actual observations might not include the “bullets” fired by the space invaders in its predictions, leading to rather poor performance if the model were to be used for planning. Note that the value prediction head can be replaced by a reward prediction head by simply changing the target from the sum of discounted rewards to rewards. In section \[section:Dyna\], we use the reward prediction head in our experiment where we use the model to accelerate learning. In applications using images as observations, predicting rewards should have a similar effect on mitigating the “vanishing bullet” problem, except possibly in sparse reward settings. The network layers are detailed in Figure \[tab:network\], where $d_o$ is the observation dimension and $d_{o+a}$ is the sum of the observation and action dimensions. Our model is trained end-to-end using two L2 loss functions, one for the difference between the predicted and actual observation and another for the difference between the predicted value and the empirical sum of discounted rewards. Similar to the model used in [@lotter2016deep], this model can be treated as a layer and stacked, providing higher levels of abstraction for higher layers. This stacking can enhance performance when observations consist of images, but we found it unnecessary in this work. $$\begin{aligned} \mathcal{R}&: \mathbf{e}, \mathbf{a} \mapsto \mathbf{r} \label{eq:pcm_r}\\ \mathcal{P}&: \mathbf{r} \mapsto \mathbf{\hat{o}}\label{eq:pcm_p}\\ \mathcal{I}&: \mathbf{o}_0 \mapsto \mathbf{h}\label{eq:pcm_i}\\ \mathcal{V}&: \mathbf{r} \mapsto \hat{v}\label{eq:pcm_v}\end{aligned}$$ 0.15in ----------------- --------------------- --------------------- ------ ------ Layer in dim out dim type act $\mathcal{R}_1$ $d_o $ $10\times d_{o+a} $ FC tanh $\mathcal{R}_2$ $10\times d_{o+a}$ $10\times d_{o+a} $ GRU None $\mathcal{P}_1$ $10\times d_{o+a} $ $10\times d_{o+a} $ FC tanh $\mathcal{P}_2$ $10\times d_{o+a} $ $d_o $ FC None $\mathcal{I}_1$ $d_{o} $ $10\times d_{o+a} $ FC tanh $\mathcal{I}_2$ $10\times d_{o+a} $ $d_h $ FC tanh $\mathcal{V}_1$ $10\times d_{o+a} $ $10\times d_{o+a} $ FC tanh $\mathcal{V}_2$ $10\times d_{o+a} $ 1 FC None ----------------- --------------------- --------------------- ------ ------ : Network Layers[]{data-label="tab:network"} -0.1in Initialize $\mathbf{e}_0=\mathbf{0}$, $t=0$, $\mu.h=\mu(\mathbf{o}_0)$,$\mathbf{o}_0$ = env.reset(). $\mathbf{a}_t=\pi(\mathbf{o}_t)$ $\mathbf{o}_{t+1},r=\mathrm{env.step}(\mathbf{a}_t)$ $\mathbf{\hat{o}}_{t+1}, \hat{v}_{t+1} = \mu(\mathbf{e}_t,\mathbf{a}_t)$ $\mathbf{e}_{t+1}=\mathbf{\hat{o}}_{t+1}-\mathbf{o}_{t+1}$ $t=t+1$ Experiments =========== Comparison of standard and enhanced PCM --------------------------------------- Here we compare the performance of a standard predictive coding model and the model described in Section \[section:pcm\_exp\] using a simple 3-DOF Mars landing environment, with initial lander conditions in the target-centered reference frame as shown in Table \[tab:IC\]. The agent’s goal is to achieve a soft pinpoint landing with a terminal glideslope of greater than 5; a complete description of the environment can be found in [@gaudet2018deep]. For this comparison we use a policy implementing the DR/DV guidance law [@d1997optimal], which maps position and velocity to a commanded thrust vector. This is similar to using a pre-trained policy as good trajectories are generated from the onset. The model is updated using 30 episode rollouts and trained for 30,000 episodes. During the forward pass, the networks are unrolled 60 steps. After training, each model is evaluated on a 3600 episode test where the model is run forward open loop for 1, 10, 30, and 60 steps. This is the same process used to collect rollouts for model and policy training, but we reshape the rollouts to allow running multi-step predictions in parallel on the full set of rollouts, and collect prediction accuracy statistics at each step. Note that in most of these cases, the model state will have evolved since the start of an episode, which is why the average accuracy for the model without hidden state initialization is not that bad. The prediction accuracy is measured as the absolute value of the prediction error (position and velocity) and the value estimate accuracy is measured using explained variance. Note that model predictions are scaled such that each element of the prediction has zero mean and is divided by three times the standard deviation, and the error is calculated over the samples and features (which have equal scales). So a prediction error of 0.01 is 1% of the range of values encountered over all elements of the predicted position and velocity vectors; if the maximum altitude was say 2400m, then a mean absolute value of prediction error of 1% would correspond to 24m. To insure a fair comparison, the code is identical in each case, except that the appropriate network is instantiated in the model. ------------ ----- ----- ------- ------ min max min max Downrange -70 -10 0 2000 Crossrange -30 30 -1000 1000 Elevation -90 -70 2300 2400 ------------ ----- ----- ------- ------ : Mars Lander Initial Conditions[]{data-label="tab:IC"} Tables \[tab:model-pred\_perf\] and \[tab:model-ev\_perf\] shows prediction error performance for three model architectures. An explained variance of ’-’ indicates less than zero. PCM is a standard predictive coding model, i.e., without the mapping between initial observations and the hidden state and without zeroing the error feedback in the forward pass. The maximum prediction error for this model is high (close to the maximum range of each state variable), and occurs when the multi-step error checking begins at the start of an episode, so the model never gets any error feedback. When the error checking begins mid-episode, performance improves, and consequently the mean error does not look too bad. PCM-I adds the network layers that learn a mapping between an episodes initial observation and the recurrent layer’s hidden state. Here we see that the mean prediction error improves, and although not shown, the maximum prediction error is much reduced. PCM-I-OLT is our PCM model as described in Section \[section:pcm\_exp\], where the error feedback is set to zero during the forward pass in training. We see that running the forward pass open loop has a significant impact on multi-step prediction performance. We also include two standard action conditional models as comparison baselines. ACM is a standard action conditional model and ACM-OLT is modified to feed back the predicted observation as the observation during the training forward pass. Note that even without the open loop forward pass, the predictive coding model excels at multi-step prediction. We hypothesize that the performance is due to the model taking a prediction error rather than an observation as an input. This requires the model to integrate multiple steps of previous errors in order to make good predictions, thereby forcing the model to learn a representation capturing temporal dependencies. The model learns reasonably quickly, with the model’s performance as a function of training steps given in Table \[tab:model-ep\_perf\]. These statistics are captured during training, and importantly, are measured on the current rollouts before the model trains on these rollouts, and are consequently a good measure of generalization. 0.15in ----------- -------- -------- -------- -------- -- K 1 10 30 60 ACM 0.0150 0.3892 0.3605 0.3152 ACM-OLT 0.0165 0.1615 0.1954 0.2174 PCM 0.0360 0.0392 0.0439 0.0483 PCM-I 0.0019 0.0088 0.0153 0.0216 PCM-I-OLT 0.0020 0.0024 0.0033 0.0044 ----------- -------- -------- -------- -------- -- : K Step Mean Abs Prediction Error[]{data-label="tab:model-pred_perf"} -0.1in 0.15in ----------- -------- -------- -------- -------- -- K 1 10 30 60 ACM 0.9380 - - - ACM-OLT 0.9759 0.6626 0.3816 - PCM 0.9808 0.9789 0.9736 0.9604 PCM-I 0.9851 0.9723 0.9426 0.9167 PCM-I-OLT 0.9859 0.9877 0.9847 0.9798 ----------- -------- -------- -------- -------- -- : K Step Mean Explained Variance[]{data-label="tab:model-ev_perf"} -0.1in 0.15in ------------------- -------- -------- -------- -------- -- Steps $\times1e4$ 8 16 24 78 Pred. Err. 0.1155 0.0411 0.0176 0.0065 Exp. Var. 0.7214 0.8536 0.8626 0.9374 ------------------- -------- -------- -------- -------- -- : PCM 20-step Prediction Performance as function of Training Steps[]{data-label="tab:model-ep_perf"} -0.1in 6-DOF Mars Landing ------------------ Here we test the model’s ability to predict a lander’s 6-DOF state during a simulated powered descent phase. To make this more interesting, we let the model and agent learn concurrently. The agent uses PPO to learn a policy to generate a thrust command for each of the lander’s four thrusters, which are pointed downwards in the body frame. Rather than use a separate value function baseline, the policy uses the model’s value estimate. This is a difficult task, as to achieve a given inertial frame thrust vector, the policy needs to figure out how to properly rotate the lander so that the thrusters are properly oriented, but this also affects the lander’s translational motion. The goal is for the lander to achieve a soft pinpoint landing with the velocity vector directed predominately downward, an upright attitude, and negligible rotational velocity. The initial conditions are similar to that given in Table \[tab:IC\], except that the lander’s attitude and rotational velocity are perturbed from nominal. A full description of the environment can be found in [@gaudet2018deep]. In the first experiment, the policy and model get access to the ground truth state (position, velocity, quaternion attitude representation, and rotational velocity). In a variation on this experiment, the model only gets access to simulated Doppler radar altimeter returns, using a digital terrain map (DTM) of the Mars surface in the vicinity of Uzboi Valis. We assume a configuration similar to the Mars Science Laboratory lander, where there are four altimeters with fixed orientation in the body frame, each pointing downwards and outwards at a 22 degree angle from the body frame vertical axis. Since the DTM spans 40 square km, we had to cut some corners simulating the altimeter readings in order to reduce computation time enough to be practical for the large number of episodes (300,000) required to solve this problem using RL. This results in low accuracy, particularly at lower elevations, further complicating this task. Also, due to the relatively high pitch and roll limits we allow the lander, the altimeter beams occasionally completely miss the DTM, and return a max range reading. Training updates use 120 episode rollouts. Learning a good policy in 6-DOF typically takes around 300,000 episodes, which is over 80M steps of interaction with the environment. Similarly, we find that model takes longer to learn a good representation. There are two factors that can contribute to the increased model learning time. First, the dynamics are more complex, and the model’s network is larger. Second, since it takes a long time for the policy to converge, exploration causes the model to be presented with a wide range of actions for each observation during each training epoch. Table \[tab:6-DOF\_conv\] shows the model’s convergence during training as a function of training steps. We look at this for both a pre-trained policy and concurrent learning to establish the primary factor behind the slower model convergence. Although model convergence is slower in the case of concurrent learning, the difference is not extreme; consequently we attribute the slower convergence in the 6-DOF case primarily to the more complicated dynamics. It may be possible to speed model convergence by either increasing the number of model training epochs, training on a larger set of rollouts using a replay buffer, or both. Table \[tab:6-DOF\_perf\] tabulates the model’s performance for both the case where the model has access to the ground truth lander state and the case where the model only has access to simulated Doppler radar returns. For the latter case, the observation consists of four simulated altitudes and closing velocities, one for each altimeter. These observations do not satisfy the Markov property in that there are many lander states in configuration space that can give identical readings. However, the recurrent network layer allows the model to make reasonably accurate predictions. Figure \[fig:PCM\_traj\_6dof\] illustrates an entire episode of PCM open loop predictions. Here the policy gets access to the ground truth observation on all steps, but the model only gets access to the ground truth observation on the 1st step of the episode, and for remaining steps, the PCM’s prediction error input is set to zero, removing any feedback from actual observations. The “Value” plot shows the estimated value (sum of discounted rewards) of the observation at the current step as predicted by the model’s value head. The prediction error is small enough that is is barely discernible from the plot. We use a quaternion representation for attitude. Note that the model’s network was unrolled only 60 steps during training, but this sufficed to allow good prediction out to the end of a 300 step episode. 0.15in ------------------- -------- -------- -------- -------- -- Steps $\times1e6$ 0.37 0.74 1.11 3.69 $\mu$ TP 0.0883 0.0411 0.0290 0.0155 $\sigma$ TP 0.1166 0.0649 0.0464 0.0241 $\mu$ LP 0.0964 0.0575 0.0490 0.0341 $\sigma$ LP 0.0899 0.0593 0.0549 0.0368 ------------------- -------- -------- -------- -------- -- : 6-DOF 30 step Mean Abs Prediction Error as function of steps for pre-trained policy (TP) and concurrent learning (LP)[]{data-label="tab:6-DOF_conv"} -0.1in 0.15in ---------- -------- -------- -------- -------- -- K 1 10 30 60 GT Pred 0.0101 0.0089 0.0090 0.0085 GT EV 0.9845 0.9739 0.9562 0.8633 DTM Pred 0.0094 0.0129 0.0120 0.0093 DTM EV 0.9225 0.9442 0.8633 0.5988 ---------- -------- -------- -------- -------- -- : 6-DOF K Step Mean Abs Prediction Error and Explained Variance[]{data-label="tab:6-DOF_perf"} -0.1in 0.2in -0.2in Accelerated RL using a Predictive Model {#section:Dyna} --------------------------------------- In this experiment we demonstrate an improvement in sample effiency obtained using PPO in conjunction with our model. The model $\mu$, policy $\pi$, and value function $\nu$ learn concurrently in the 3-DOF Mars environment. After each update of $\mu$, $\pi$, $\nu$ from rollouts $\mathcal{R}$, $s$ observations are sampled from $\mathcal{R}$. These are fed through $\pi$ and $\mu$ in parallel for $k$ steps to create a set $\mathcal{S}$ of simulated rollouts as shown in Algorithm \[alg:dyna\]. Note that these segments do not result in a full episode, and consequently, when we update the policy using the simulated rollouts, we need to take care discounting the simulated rewards to generate the empirical return for updating the policy. Concretely, we need to use the correct $n$-step temporal difference return, and use this to compute the advantages $\mathcal{A}$; this is implemented in lines 12 through 17 of the algorithm. Figure \[fig:Dyna\_compare\] plots the norm of the lander terminal position and velocity as a function of training episode for a policy optimized using PPO and PPO with Dyna over the first 5000 episodes. The simulated rollouts were generated using 20,000 observations sampled from the previous rollouts with k=10, resulting in simulated rollouts containing 200,000 tuples of observations, actions, and advantages. We find that PPO enhanced with Dyna converges roughly twice as fast. Initialize Model, Policy, Value Function $\mathcal{M}, \mathcal{P}, \mathcal{V} $ Collect rollouts $\mathcal{R}$ from running E episodes Update $\mu,\pi,\nu$ using $\mathcal{R}$ Randomly select the set $\mathcal{O}$ of observations from $\mathcal{R}$ $\mathcal{S} = \{\}$ $\mathbf{A}_{t}=\pi(\mathbf{O_t})$ $\mathbf{O_{t+1},R_{t+1}}=\mu(\mathbf{A_{t}},\mathbf{0})$ append $\mathbf{O}_t,\mathbf{A}_t,\mathbf{R}_{t+1}$ to $\mathcal{S}$ $\mathbf{R}_f= \nu(\mathbf{O}_{t+1})$ $\mathbf{E}_{t}=\sum_{\tau=t}^{k}\gamma^{\tau-t}\mathbf{R}_{\tau}$ $\mathbf{E}_t=\mathbf{E}_t+\gamma^{K-t}\mathbf{R}_f$ Add $\mathcal{A}$ = $\mathbf{E}_t - \nu(\mathbf{O}_t)$ to $\mathcal{S}$ Update $\pi$ using $\mathcal{S}$ 0.2in ![Dyna versus Standard PPO Performance[]{data-label="fig:Dyna_compare"}](Figures/Dyna_compare.png){width="\columnwidth"} -0.2in Discussion ========== We have presented a novel predictive coding model architecture capable of generating accurate trajectory predictions over long horizons. Through a series of experiments, we have shown that our enhancements to predictive coding outperforms a standard implementation for long-horizon predictions, and performs well predicting both the agent’s ground truth state and simulated altimeter readings open loop over an entire episode during a 6-DOF Mars powered descent phase. The ability to generate long horizon open loop trajectory predictions is extremely useful for both model based reinforcement learning and model predictive control. We demonstrated the ability of the model to reduce the sample complexity of proximal policy optimization for a Mars 3-DOF powered descent phase task. Future work will extend the model to predicting observations in higher dimensional spaces consistent with flash LIDAR and electro-optical sensors. [^1]: Note that here we are referring to the dimension of the configuration space that the agent is operating in, not the dimension of the measurement space. For example, an agent operating in a simulated Atari environment has a low dimensional configuration space but a high dimension measurement space

--- abstract: 'We investigate the critical properties of the Ising model in two dimensions on [*directed*]{} small-world lattice with quenched connectivity disorder. The disordered system is simulated by applying the Monte Carlo update heat bath algorithm. We calculate the critical temperature, as well as the critical exponents $\gamma/\nu$, $\beta/\nu$, and $1/\nu$ for several values of the rewiring probability $p$. We find that this disorder system does not belong to the same universality class as the regular two-dimensional ferromagnetic model. The Ising model on [*directed*]{} small-world lattices presents in fact a second-order phase transition with new critical exponents which do not dependent of $p$, but are identical to the exponents of the Ising model and the spin-$1$ Blume-Capel model on [*directed*]{} small-world network.' author: - 'Ediones M. Sousa and F.W.S. Lima' title: 'Ising model on directed small-world Voronoi Delaunay random lattices' --- Introduction ============ Experimental studies of real magnetic materials show that their critical behavior can suffer the influence of either impurities or inhomogeneities [@bel]. A theoretical understanding of such impurities can be realized, with a very good approximation, in the case of quenched disorder. In this case, the criterion due to Harris [@harri] is an important theoretical tool to interpret the importance of the effect of quenched random disorder on the critical behavior of a physical system. The randomness can be classified solely by the specific heat exponent of the pure system, $\alpha_{pure}$. This criterion asserts that for $\alpha_{pure}>0$ the quenched random disorder is a relevant perturbation, leading to a different critical behavior than in the pure case (as for the three-dimensional Ising model). In particular, one expects [@fisher] in the disordered system that $\nu\geq 2/D$, where $\nu$ is the correlation length exponent and $D$ is the dimension of the system. Assuming hyper-scaling to be valid, this implies $\alpha=2-D\nu\leq 0$. On the other hand, for $\alpha_{pure}<0$ disorder is irrelevant (as, for instance, in the three-dimensional Heisenberg model) and, in the marginal case $\alpha_{pure}=0$ (like the $D=2$ Ising model), no prediction can be made. For the case of (non-critical) first-order phase transitions it is known that the influence of quenched random disorder can lead to a softening of the transition [@Imry]. Recently, the predicted softening effect at first-order phase transitions has been confirmed for D=3 q-state Potts models with $q\geq3 $ using Monte Carlo [@balestro; @chatelain1; @chatelain2] and high temperature series expansion [@M] techniques. The overall picture is even better in two dimensions ($D=2$) where several models with $\alpha_{pure}>0$ [@D; @Ma; @Gj; @SW] and the marginal ($\alpha_{pure}=0$) [@25; @28; @aarao; @puli; @puli2] have been investigated. In this paper we study a type of different quenched disorder, namely the effect of [*directed*]{} bound case with rewiring probability $p$ [@sanches]. Specifically, we consider $D=2$ small-world Voronoi Delaunay random lattices (SWVD) type, and performed an extensive computer simulation study of Ising model. We concentrated on the close vicinity of the transition point and applied finite-size scaling (FSS) techniques to extract the exponents and the “renormalized charges" $U_{4}^{*}$. Monte Carlo simulations of the disorder system was realized using the spin-flip heat bath algorithm to update the spins. Previous studies of connectivity disorder on $D=2$ lattices have been realized by Monte Carlo simulations of $q$-state Potts models on quenched random lattices of Voronoi Delaunay type for $q=2$ [@janke; @janke1; @FWSL], $q=3$ [@FWSL1] and $q=8$ [@janke2; @FWSL2]. In particular, it has been shown that for $q=2$ [@janke; @janke1; @FWSL] and $q=3$ [@FWSL1] the critical exponents are the same as those for the model on a regular $D=2$ lattice. This is indeed a surprising result since the relevant criterion of the Delaunay triangulations reduces to the well known Harris criterion such that disorder of this type should be relevant for any model with positive specific heat exponent [@JankeWeigel]. This means that for $q=3$, where $\alpha_{pure}>0$, one would expect a different universality class. For the spin-1 Ising model, where $\alpha_{pure}=0$, Fernandes et al. [@fernandes] showed that the exponents do no change in the [*undirected*]{} SWVD lattice, but for [*directed*]{} SWVD random lattice the situation is quite different. There is a second-order phase transition for $p<p_{c}$ and a first-order phase transition for $p>p_{c}$, where $p_{c}\approx 0.35$ is the rewiring probability where the system change the order phase transition. In addition, the calculated critical exponents for $p<p_{c}$ do not belong to the same universality class as the regular two-dimensional ferromagnetic model. Therefore both [*undirected*]{} and [*directed*]{} cases agree with Harris criterion for $\alpha_{pure}=0$. In the present spin-$1/2$ Ising model on [*directed*]{} SWVD lattice we show that the critical behavior is quite similar to that observed by Fernandes et al. in the spin-$1$ case[@fernandes]. However, now one has only a second-order phase transition for all $p$ values studied. The critical exponents do not belong to the same universality class as the regular two-dimensional ferromagnetic model, but they agree with the critical exponents of Blume-Capel model for $p<p_{c}$ [@fernandes]. In the next section we present the model and the simulation background. The results and conclusions are discussed in the last section. Model and Simulation ==================== We consider the ferromagnetic spin-$1/2$ Ising model, on [*directed*]{} SWVD random lattice by a set of spins variables ${S_i}$ taking the values $\pm 1$ situated on every site $i$ of a [*directed*]{} SWVD random lattice with $N=L\times L$ sites, were $L$ is the side of square cluster. In this random lattice, similar to Sánchez et al. [@sanches], we start from a two-dimensional SWVD random lattice consisting of sites linked to their $c$ ( where $3<c<20$ and different for each site of network) nearest neighbors by both outgoing and incoming links. Then, with probability $p$, we reconnect nearest-neighbor outgoing links to a different site chosen at random. After repeating this process for every link, we are left with a network with a density $p$ of SWVD [*directed*]{} links. Therefore, with this procedure every site will have $c$ outgoing links and varying (random) number of incoming links. The evolution in time of these systems is given by a single spin-flip like dynamics with a probability $P_{i}$ given by $$%\begin{center} P_{i}= 1/[1+\exp(2E_{i}/k_BT)], %\end{center}$$ where $T$ is the temperature, $k_B$ is the Boltzmann constant, and $E_i$ is the energy of the configuration obtained from the Hamiltonian $$H=-J\sum_{<i,j>}S_{i}S_{j},$$ where the sum runs over all neighbor pairs of sites (including the nearest-neighbor and the long ranged ones determined by the probability $p$) and the spin-$1/2$ variables $S_{i}$ assume values $\pm 1$. In the above equation $J$ is the exchange coupling. The spin-$1/2$ case is well known in the literature [@a3; @onsager]. The simulations have been performed on different SWVD random lattice sizes comprising a number $N=5000$, $10000$, $20000$, $40000$, $60000$ and $80000$ of sites. For simplicity, the length of the system is defined here in terms of the size of a regular lattice $L=N^{1/2}$. For each system size quenched averages over the connectivity disorder are approximated by averaging over $R=20$ independent realizations. For each simulation we have started with a uniform configuration of spins. We ran $4\times10^{5}$ Monte Carlo steps (MCS) per spin with $2\times10^{5}$ configurations discarded for thermalization using the “perfect" random-number generator [@nu]. We do not see any significant change by increasing the number of R and MCS. So, for the sake of saving computer time, the present values seem to give reasonable results for our simulation. In both cases we have employed the heat bath algorithm and for every MCS, the energy per spin, $e=E/N$, and the magnetization per spin, $m=\sum_{i}S_{i}/N$, were measured. From the energy measurements we can compute the average energy, specific heat and the fourth-order Binder cumulant of the energy, given respectively by $$u(T)=[<E>]_{av}/N,$$ $$C(T)=K^{2}N[<e^{2}>-<e>^{2}]_{av},$$ $$B(T)=1-[\frac{<e^{4}>}{3<e^{2}>^{2}}]_{av}.$$ In the above equations $<...>$ stands for thermodynamic averages and $[...]_{av}$ for averages over different realizations. Similarly, we can derive from the magnetization measurements the average magnetization, the susceptibility, and the fourth-order magnetic cumulant, $$m(T)=[<|m|>]_{av},$$ $$\chi(T)=KN[<m^{2}>-<|m|>^{2}]_{av},$$ $$U_{4}(T)=1-[\frac{<m^{4}>}{3<|m|>^{2}}]_{av}.$$ In order to calculate the exponents of this model, we apply finite-size scaling (FSS) theory. We then expect, for large system sizes, an asymptotic FSS behavior of the form $$C=C_{reg}+L^{\alpha/\nu}f_{C}(x)[1+...],$$ $$[<|m|>]_{av}=L^{-\beta/\nu}f_{m}(x)[1+...],$$ $$\chi=L^{\gamma/\nu}f_{\chi}(x)[1+...],$$ where $C_{reg}$ is a regular background term, $\nu$, $\alpha$, $\beta$, and $\gamma$ are the usual critical exponents, and $f_{i}(x)$ are FSS functions with $$x=(T-T_{c})L^{1/\nu}$$ being the scaling variable. The dots in the brackets $[1+...]$ indicate corrections-to-scaling terms. We calculated the error bars from the fluctuations among the different realizations. Note that these errors contain both, the average thermodynamic error for a given realization and the theoretical variance for infinitely accurate thermodynamic averages which are caused by the variation of the quenched, random geometry of the lattices. Results and conclusion ====================== By applying the standard heat bath algorithm to each of the $R$ energy data we determine the temperature dependence of $C_{i}(T)$, $\chi_{i}(T)$,..., $i=1$,...,$R$. Once the temperature dependence is known for each realization, we can easily compute the disorder average, e.g., $C(T)=\sum^{T}_{i=1}C_{i}(T)/R$, and then determine the maxima of the averaged quantities, e.g., $C_{max}(T_{max})=max_{T}C(T)$. The variable $R$ ($=20$) represents the number of replicas in our simulations. In Figure \[cum0\] we show the behavior of the magnetization versus temperature for several different lattice sizes and rewiring probability $p=0.5$. Figure \[cum1\] displays the behavior of the suscepbility versus temperature for the same parameters used in Figure \[cum0\]. From here on we set $J$ and $k_B$ to unity. One can see a typical behavior of a second-order phase transition. In order to estimate the critical temperature we calculate the fourth-order Binder cumulant given by eq. (8). It is well known that these quantities are independent of the system size and should intercept at the critical temperature [@binder]. In Figure \[cum2\] the fourth-order Binder cumulant is shown as a function of the $T$ for several lattice sizes for the rewiring probability $p=0.5$. Taking the largest lattices we have $T_{c}= 5.118(4)$. To estimate $U^{*}_{4}$ we note that it varies little at $T_{c}$, so we have $U^{*}_{4}= 0.283(4)$. One can see that $U^{*}_{4}$ is different from the universal value $U^{*}_{4}\sim 0.61$ for the Ising model on the regular $D=2$ lattice, and also for the Ising model on Voronoi-Delaunay random lattice in two-dimensions [@janke; @janke1; @FWSL]. By following this same procedure one can get the corresponding results for other values of $p$. ![\[cum0\] (color online) Magnetization as a function of $T$ for various lattice sizes with $N=5000$, $N=10000$, $N=20000$, $N=40000$, $N=60000$, and $N=80000$ and rewiring probability $p=0.5$.](fig1.eps){width="8.0cm"} ![\[cum1\] (color online) The same as Figure \[cum0\] for the susceptibility versus temperature $T$.](fig2.eps){width="8.0cm"} ![\[cum2\] (color online) The same as Figure \[cum0\] for the fourth-order Binder cumulant as a function of $T$.](fig3.eps){width="8.0cm"} The correlation length exponent can be estimated from $ T_{c}(L)=T_{c}+bL^{-1/\nu}$, where $T_c(L)$ is the pseudo-critical temperature for the lattice size $L$, $T_c$ is the critical temperature in the thermodynamic limit, and $b$ is a non-universal constant. In Figure \[expnu\] it is shown a plot of $\ln\left[T_c(L)-T_c\right]$ as a function of $\ln L$ for several values of $p$. One can clearly see that the exponent is, within the errors, independent of $p$, in agreement with universality ideas. The actual values of $1/\nu$ are displayed in Table I.  $ \ln\left[T_{c}(L)-T_{c}\right]$ as a function of $L$ for several values of $p$. The solid lines are the best linear fits. ](fig4_0.eps){width="8.0cm"} In order to go further in the present analysis we have also computed the modulus of the magnetization at the inflection point and the magnetic susceptibility at $T_c$. The logarithm of these quantities as a function of the logarithm of $L$ are presented in Figures \[mag\] and \[sus\], respectively. A linear fit of these data gives $\beta/\nu$ from the magnetization and $\gamma/\nu$ from the susceptibility. In addition, we plotted in Figure \[mag1\] the logarithm of the maximum value of the susceptibility $\chi_{max}$ as a function of $\ln L$ for several values of $p$. One can also see that the exponents $\beta/\nu$ and $\gamma/\nu$ are also independent of $p$, as expected. They are different from $\beta/\nu=0.125$ and $\gamma/\nu=1.75$ obtained for a regular $D=2$ lattice, but obey hiper-scaling relation (into the error bar) $$2\frac{\beta}{\nu}+ \frac{\gamma}{\nu}=D,$$ where $D=2$. The numerical values of the ratio $\beta/\nu$ and $\gamma/\nu$ are also shown in Table I. ![\[mag\] (color online) Plot of the logarithm of the modulus of the magnetization at the inflection point as a function of the logarithm of $L$. The solid lines are the best linear fit. ](fig4.eps){width="8.0cm"} ![\[sus\] (color online) Log-log plot of the susceptibility $\chi$ at $T_{c}$ as a function of the logarithm of $L$. The solid lines are the best linear fit. ](fig5.eps){width="7.9cm"} ![\[mag1\] (color online) Log-log plot of the susceptibility maxima $\chi_{max}$ as a function of the logarithm of $L$. The solid lines are the best linear fit.](fig6.eps){width="8.0cm"} In Figures \[cola1\] and \[cola2\] we display the data colapse for the magnetisation and the susceptibility for $p=0.5$. In these cases, we see that the estimative of the critical exponents ratio $\beta/\nu$ and $\gamma/\nu$ are in good agreement for all lattice sizes. The same qualitative results are obtained for other values of $p$. ![\[cola1\] (color online) Data colapse of magnetisation for various values of $L$ and $p=0.5$.](fig7.eps){width="8.0cm"} ![\[cola2\] The same as in Figure \[cola1\] for the susceptibility.](fig8.eps){width="8.0cm"} $p$ $1/\nu$ $\beta/\nu$ $\gamma/\nu$ $\gamma/\nu^{max}$ ------- ------------- ------------- -------------- -------------------- -- -- $0.1$ $1.036(49)$ $0.489(8)$ $1.003(11)$ $1.001(13)$ $0.2$ $1.098(82)$ $0.538(68)$ $1.016(11)$ $1.016(5)$ $0.3$ $1.009(49)$ $0.463 (4)$ $0.924(98)$ $1.012(3)$ $0.4$ $0.886(8)$ $0.491 (9)$ $1.017(14)$ $1.012(8)$ $0.5$ $0.987(64)$ $0.494(10)$ $0.998(18)$ $1.005(66)$ $0.6$ $0.927(92)$ $0.486(10)$ $1.042(13)$ $1.004(7)$ $0.7$ $1.107(60)$ $0.486(10)$ $1.016(13)$ $1.003(10)$ $0.8$ $0.972(57)$ $0.493(16)$ $1.018(23)$ $1.021(7)$ $0.9$ $1.032(66)$ $0.471(12)$ $1.038(16)$ $0.991(69)$ : The critical exponents, for spin-$1/2$ on [*directed*]{} SWVD random lattice with probability $p$. $\gamma/\nu^{max}$ are the results from the maximum of the magnetic susceptibility. Error bars are statistical only. In summary, from the above results, there is a strong indication that the spin-$1/2$ Ising model model on a [*directed*]{} SWVD random lattice is in a different universality class than the model on a regular two-dimensional lattice. The exponents here obtained are independent of $p$ and different from the Ising model on regular $D=2$ lattice, but they are equivalent to the exponents of the Ising model and the spin-$1$ Blume-Capel model on [*directed*]{} small-world network [@newpla]. One possible explanation for this change in universality can be ascribed to the influence of long range interactions that occur with the presence of $p$ directed bounds. However, our results agree with the Harris-Luck criterion for [*directed*]{} SWVD random lattice. [20]{} D. P. Belanger, Braz. J. Phys. [**30**]{}, 682 (2000). A. B. Harris, J. Phys. C [**7**]{}, 1671, (1974). J. Chayes, L. Chayes, D. S. Fisher, and T. Soencer, Phys. Rev. Lett. [**57**]{}, 299 (1986). Y. Imry, M. Wortis, Phys. Rev. B [**19**]{}, 3581 (1979). H. G. Ballestros, L. A. Fernandez, V. Martin-Mayor, A. Munoz Sudupe, G. Parisi, and J. J. Ruiz-Lorenzo, Phys. Rev. B [**61**]{}, 3215 (2000). C. Chatelain, B. Berche, W. Janke, and P. E. Berche, Phys. Rev. [**E 64**]{}, 036120(2001). C. Chatelain, P. E. Berche, B. Berche, and W. Janke, Nucl. Phys. B (Proc. Suppl.) [**106-107**]{}, 899 (2002). M. Hellmund and W. Janke, Nucl. Phys. B (Proc. Supl.) [**106-1-7**]{}, 923 (2002). D. Matthews-Morgan, D. P. landau, R. H. Swendsen, Phys. Rev. lett. [**53**]{}, 679 (1984). M. A. Novotony and D. P. Landau, J. Mag. magn. Mater. [**15-18**]{}, 247 (1980). G. Jug and B. N. Shalaev, Phys. Rev. B [**54**]{}, 3442 (1996). S. Wiseman and E. Domany, Phys. Rev. E [**51**]{}, 3074 (1995). Vik. S. Dotsenko and VI. S. Dotsenko, So. Phys. JEPT Lett. [**33**]{}, 37 (1981). A. Roder, J. Adler, and W. Janke, Phys. Rev. Lett. [**80**]{}, 4697 (1988). F. D. A. Aarão Reis, S. L. A. de Quaeiroz, and R. R. dos Santos, Phys. Rev. B [**56**]{} 6013 (1997). P. H. L. Martins and J. A. Plascak, Phys. Rev. E [**76**]{} (2007) 012102. P. H. L. Martins and J. A. Plascak, Braz. J. Phys. [**34**]{}, 433 (2004). Alejandro D. Sánchez, Juan M. Lopes, and Miguel A. Rodriguez, Phys. Rev. Lett. [**88**]{}, 048701 (2002). W. Janke, Phys. Rev. B 47, 14757 (1993). W. Janke, R. Villanova, Phys. Lett. A 209, 179 (1995). D. Spriu, M. Gross, P. E. L. Rakow, and J. F. Wheater, Nucl. Phys. B [**265**]{} \[FS15\], 92 (1986). W. Janke, M. katoot, and R. Villanova, Phys. Lett B [**315**]{}, 412 (1993). F. W. S. Lima, J. E. Moreira, J. S. Andrade Jr., and U.M.S. Costa, Phys. A [**283**]{}, 100 (2000). F. W. S. Lima, U.M.S. Costa, M. P. Almeida, J. S. Andrade Jr., Eur. Phys. J. B [**17**]{}, 111 (2000). W. Janke and R. Villanova, Phys. Lett. A [**209**]{}, 179 (1995). F. W. S. Lima, J. E. Moreira, J. S. Andrade Jr., and U.M.S. Costa, Eur. Phys. J. B [**13**]{}, 107 (2000). W. Janke and M. Weigel, Phys. Rev. E [**69**]{}, 144208 (2004). F. P. Fernandes, F. W. S. Lima, and J. A. Plascak. Computer Physics Communications, [**181**]{}, 1218 (2010). L. Onsager, Phys. Rev. 65, 117 (1944); B. Kaufmann, Phys. Rev. [**76**]{}, 1232 (1949). R. J. Baxter, [*Exactly solved models in statistical mechanics*]{}, London, Academic Press (1982). P. L’Ecuyer, Commun. ACM [**31**]{}, 742 (1988). K. Binder, Z. Phys. B [**43**]{}, 119 (1981). F. W. S. Lima ; J. A. Plascak . Critical behavior of the Ising and Blume-Capel models on directed two-dimensional small-world networks. Appear in The European Physical Journal. B, Condensed Matter Physics (Print), 2013.

--- author: - '[A. MAMONTOV]{}' title: 'On periodic groups isospectral to $A_7$' --- [^1] [^2] [^3] Introduction ============ Let $\mathfrak{M}$ be a set of periodic groups and $G \in \mathfrak{M}$. The [*spectrum*]{} of $G$ is the set $\omega (G)$ of its elements orders. Two groups with the same spectrum are called [*isospectral*]{}. We say that $G$ is [*recognizable by spectrum in*]{} $\mathfrak{M}$, if any group from $\mathfrak{M}$ is isomorphic to $G$, whenever it is isospectral to $G$. Many finite simple groups are known to be recognizable by spectrum in the class of finite groups [@fin_review], in particular, $A_7$ [@OCn]. It is clear that a group $G$ is recognizable by spectrum in the class of periodic groups, if it is recognizable among finite groups, and any group isospectral to $G$ is locally finite. Verification of the last condition is related to the Burnside problem [@lmreviewe]. In the paper we prove Assume $\omega(G)=\omega(A_7)$ and $G$ has a subgroup $H$ isomorphic to $A_4$, whose involutions are squares of elements of order $4$. Then either $O_2(H) \subseteq O_2(G)$ or $G$ has a finite nonabelian simple subgroup. Existence of such simple subgroup $K \leq G$ allows to investigate the structure of $G$ using the inclusion $C_K(i) \leq C_G(i)$ (where $i$ is an involution from $K$) and Shunkov’s theorem that a periodic group with a finite centralizer of involution is locally finite [@shu1972e]. We conjecture that $G \simeq A_7$. Notations and preliminary results ================================= Speaking of computations we refer to computations in GAP [@gap] using coset enumeration algorithm. The following notations are used: $\Gamma_n=\Gamma_n(G)$ is the set of elements of order $n$ from $G$; $\Delta=\Delta(G)=\{x^2\mid x \in \Gamma_4 \}$, $O_2(G)$ is the largest normal 2-subgroup of $G$. Write $A:B$ for an extension of $A$ by $B$, and $C_n$ for a cyclic group of order $n$. The center of $G$ is denoted as $Z(G)$. Writing $a \sim b$ for $a,b \in G$ we mean that orders of $a$ and $b$ are equal. Groups in the work are mostly presented by generators and relations. Any word identity in a group also holds in its homomorphic images. So it is convenient to keep same notations for generators of both groups. For visualization some relations are drawn in the form of labeled graph. Vertex label represents some group element, and the corresponding number is its order. Edge label shows relations between the corresponding vertices (number stands for the order of the vertex element product), or subgroup, which they generate. An edge between commuting vertices is drawn as dashed line. In this section let $G$ be a group whose element orders are not greater than $7$. The goal of this section is to describe small subgroups, which may be in $G$. For further discussion, we distinguish some elements, which are written as words of generators, and establish some relations between them. In the following lemma we describe $(2,3)$-generated subgroups of $G$. [l]{}\[340pt\][20mm]{} {width="20mm"} [**Lemma 2.1**]{} *Consider an involution $t$ and an element $x$ of order $3$ in $G$. Let $K=\langle t,x \rangle$. Then one of the following holds:* - $(xt)^{2}=1$ and $K \simeq S_3$; - $(xt)^{3}=1$ and $K \simeq A_{4}$; - $(xt)^{4}=1$ and $K \simeq S_{4}$; - $(xt)^{5}=1$ and $K \simeq A_{5}$; - $(xt)^{6}=1$, and $K$ is a homomorphic image of $(C_k \times C_k) \leftthreetimes C_6$, where $k$ is the order of $[x,y]$; - $(xt)^{7}=[x,t]^4=1$ and $K \simeq L_2(7)$. Items (a)-(d) are well-known. \(e) Let $(xt)^6=1$ and $k$ be the order of $[x,t]$. Set $a=[x,t]=x^{-1}txt$. Then $a^t=tx^{-1}txt\cdot t=tx^{-1}tx=a^{-1}$. Further, $aa^xa^{x^2}=(ax^{-1})^3=x^{-1}(tx)^6x=1$ and $a^{xt}=a^{txa}=(a^{-1})^{xa}$, therefore $\langle a,a^x \rangle \unlhd \langle x,t \rangle$. Direct computations show that $[a,a^x]=(tx)^6=1$. \(f) Let $(xt)^7=1$. Then $K$ is a homomorphic image of $K(j)=\langle x,t | 1=x^3=t^2=(xt)^7=[x,t]^j \rangle$, where $j \in \{4,5,6,7 \}$. Computations show that $K(5) \simeq 1$; $K(6) \simeq K(7) \simeq L_2(13)$ and therefore has an element of order $13$, which is not possible. Hence, $K \simeq K(4) \simeq L_2(7)$. The lemma is proved. Further $F_n$ will denote the Frobenius group $\langle x,t| 1=(xt)^6=[x,t]^k \rangle$ from item (e), where $k \in \{ 4,5,6,7 \}$; and $n=6k^2$ is the order of the group. In the following Lemma 2.3 we list some properties of $F_n$ that will be used later. [l]{}\[340pt\][20mm]{} {width="20mm"} [**Lemma 2.2**]{} [ *Let $x,y,t \in \Gamma_2$, $[x,y]=1$ and $\langle x,t \rangle \simeq S_3$. Then the group $\langle x,y,t \rangle$ is finite and $yt \not \in \Gamma_7$. Moreover, if $yt \in \Gamma_5$, then $\langle x,y,t \rangle \simeq A_5$.*]{} Note that $\langle x,y,t \rangle$ is a homomorphic image of $\langle x,y,t \mid 1=x^2=y^2=t^2=(xy)^2=(xt)^3=(yt)^i=(xyt)^j \rangle$, where $i,j \in \{4,5,6,7 \}$ are selected in correspondence with Lemma 2.1. Now the statement is easily verified by computations. The lemma is proved. [**Lemma 2.3**]{} [*Let $z=t^{xt}, b=z^xz, u=z^{x^2z}=t^{[x^{-1},t][x,t]}$ be the corresponding elements of $F_{294}$. Then $b^x=b^{4}$, $F_{42} = \langle x,z | 1=x^3=z^2=(xz)^6=b^7, b^x=b^4 \rangle$ is the Frobenius group of order $42$, and $[u,x]=1$.*]{} Let $a=t^xt$ as in the proof of Lemma 2.1. Then $z=ta$ and $$b=x^{-1}taxta=aa^{xt}a=a(a^{-1})^{xa}a=(a^{-1})^xa^2,$$ $$b^4=(a^3)^xa=x^{-1} \cdot (x^{-1}txt)^3 \cdot x \cdot x^{-1}txt =$$ $$=xtxtx^{-1}txtx^{-1}tx^{-1}t,$$ $$b^x=(a^2(a^{-1})^x)^x=x^{-1} \cdot x^{-1}txtx^{-1}txt \cdot x^{-1} \cdot tx^{-1}tx \cdot x^2=xtxtx^{-1}txtx^{-1}tx^{-1}t.$$ Note that a map $$x \rightarrow (1,2,3)(4,5,6)(7,8,9)(10,11,12)(13,14,15)(16,17,18)(19,20,21)$$ $$t \rightarrow (2,4)(3,5)(6,7)(8,10)(9,11)(12,13)(14,16)(15,17)(18,19)$$ can be extended to an isomorphism between $\langle x,t \rangle $ and the corresponding permutation group. Now the rest of the proof is straightforward. This notation $F_{42}$ for the corresponding Frobenius group will be used further. An important property of subgroups of $G$, which are isomorphic to $F_{42}$, is described in the following section. Subgroups isomorphic to $D_{10}$ and $F_{42}$ ============================================= The goal of the section is to prove two statements. Writing $F_{42}=\langle z,x \rangle$ we assume that elements $z$ and $x$ satisfy the relations, which were used to define $F_{42}$ in Lemma 2.3. [l]{}\[340pt\][20mm]{} {width="20mm"} [**Statement 1**]{} [*Assume that $G$ has no elements of order greater than $7$. Let $F_{42}= \langle z,x \rangle$ be a subgroup of $G$, $u=z^{x^2z}$, and $v$ be an involution in $G$ such that $[x,v]=1$. Then either $v=u$ or $G$ has a subgroup isomorphic to $L_2(7)$.*]{} The proof is preceded by two lemmas with the same assumptions. Denote by $B$ the following set of relations, which are assumed, and also shown on the graph: $$B=\{ v^2, z^2, x^3, (xz)^6, b^7, b^xb^{-4}, [x,v] \}, \text{where } b=z^xz.$$ [l]{}\[340pt\][20mm]{} {width="20mm"} [**Lemma 3.1**]{} [*If $[u,v]=1$, then $v=u$.*]{} Note that the subgroup $\langle v,z,x \rangle$ is a homomorphic image of $G(i_1,i_2,i_3,i_4) = \langle v,z,x| B \cup \{ [u,v], (zv)^{i_1}, (v^{{x^z}}v)^{i_2}, (vx^z)^{i_3}, (xzv)^{i_4} \} \rangle$, where $i_1,\ldots,i_4 \in \{ 4,5,6,7 \}$. Computations show that either the order of $G(i_1,i_2,i_3,i_4)$ divides $42$, or $v$ centralizes $F_{42}$, which is not possible. The lemma is proved. [l]{}\[340pt\][20mm]{} {width="20mm"} [**Lemma 3.2**]{} [*If $(uv)^3=1$, then either $v=u$ or $G$ has a subgroup isomorphic to $L_2(7)$.*]{} Assume the contrary. Then by Lemma 2.1.e the order of $v \cdot x^z$ and $v \cdot xx^z$ is not 7. Therefore $\langle x,z,v \rangle $ is a homomorphic image of $G(j_1,j_2,i_1,i_2,i_3) = \langle x,z,v| B \cup \{ (uv)^3, (vx^z)^{j_1}, (xx^zv)^{j_2}, (vz)^{i_1}, (vzx)^{i_2}, (xx^zxv)^{i_3} \} \rangle$, where $j_1, j_2 \in \{4,5,6\}$ and $i_1,i_2, i_3 \in \{4,5,6,7\}$. Computations show that indexes $|G(j_1,j_2,i_1,i_2,i_3): \langle x \rangle|$ are not greater than $14$, which means that $v \in \langle x,z \rangle$. A contradiction. The lemma is proved. Assume the contrary. The dihedral subgroup $\langle u,v \rangle$ centralizes $x$ and so does not contain elements of orders $4$, $5$ and $7$. If the order of $uv$ is even, then let $i$ be the involution from the center of $\langle u,v \rangle$. By Lemma 3.1 $i=u$ and hence $v=u$, a contradiction. If the order of $uv$ is 3, then obtain a contradiction by Lemma 3.2. The statement is proved. We will also need to following corollary. [l]{}\[340pt\][20mm]{} {width="20mm"} [**Lemma 3.3**]{} [*Let $F_{42}= \langle z,x \rangle$ be a subgroup of $G$, and let $y$ be an element of order $3$ from $C_G(x)$. Then $[y,u]=1$, where $u=z^{x^2z}$.*]{} There are no elements of order $12$ in $G$, and so there are no elements of order $4$ in $\langle u,y \rangle$. By Statement 1 there are no involutions in $\langle u,y \rangle$ other than $u$. By Lemma 2.1 we obtain $[u,y]=1$. The lemma is proved. [**Statement 2**]{} [*Assume that $4,5 \in \omega(G)$ and $G$ has no elements of order greater than $7$. Let $a \in \Delta$. If $a$ is inverting an element of order $5$ and $C_G(a)$ contains an involution $t \not = a$, then $G$ has a nonabelian finite simple subgroup.* ]{} Let $a\in \Delta$. Take $b \in \Gamma_4$ such that $b^2=a$. By assumption there is an involution $c$ such that $ac \in \Gamma_5$. First we use relations, which define $\langle b,c \rangle$, to prove that $G$ has a Frobenius subgroup $K \simeq F_{20}$, which contains $a$. Then we prove that $\langle K,t \rangle$ is finite and satisfies the statement conclusion. A group $\langle b,c \rangle$ is a homomorphic image of order $\geq 20$ of a group $G(i_1,i_2,i_3,i_4,i_5)=\langle b,c | 1=b^4=c^2=x^5=(xb)^{i_1}=[x,b]^{i_2}=(a^xb)^{i_3}=(bxb^3x^2)^{i_4}=((cb^2cb)^2cb^{-1})^{i_5} \rangle $, where $a=b^2$, $x=ac$; and $i_1,\ldots,i_5 \in \{4,5,6,7\}$. Computations show that $G(5,4,5,6,6)\simeq A_6$, $G(7,5,7,4,5)\simeq L_3(4)$, which is not possible; $H=G(4,5,4,5,4) \simeq (C_5 \times C_5):C_4$. For other values of the parameters the order of the corresponding group is not greater than $20$. In both suitable cases: when $\langle b,c \rangle$ is isomorphic to $H$ and when $\langle b,c \rangle$ is its homomorphic image $F_{20}=C_5:C_4$ we may choose $d$ such that $d^4=1, b=dad^2$ and $\langle a,d \rangle$ is a subgroup of $G$ isomorphic to $F_{20}$. Indeed, in $H$ consider $d=b^{cbcaca}$. In further proof we use the assumption that $C_G(a)$ has an involution $t \not =a$. By [@lmmj2014e Lemma 9] such involution $t$ may be selected in the normalizer of $b$. Therefore, $\langle d,a,t \rangle$ is a homomorphic image of $G(w,i_1,i_2,i_3,i_4) = \langle d,a,t|1=d^4=a^2=t^2=[a,t]=w=(td)^{i_1}=(dat)^{i_2}=(d^2t)^{i_3}=(d^2at)^{i_4}, a=b^2 \rangle$, where $b=dad^2$, $w \in \{ b^{-1}b^t, bb^t \}$, $i_1,i_2,i_3,i_4 \in \{ 4,5,6,7 \}$. Computations show that either the orders of these groups are not greater than 20, or they are isomorphic to $S_5$ or $S_6$, which is not possible. For example, $G(b^{-1}b^t,5,5,6,6) \simeq S_6$. The statement is proved. Subgroups isomorphic to $A_4$ ============================= The main goal of this section is to prove the theorem. So we assume that $G$ satisfies the assumptions of Theorem, and $H$ is the corresponding subgroup, which is isomorphic to $A_4$. Throughout the section, $a\in \Delta$ and $x$ denote generators of $H$, which satisfy the following relations: $x^3=a^2=(ax)^3$. Then $O_2(H)=\langle a,a^x \rangle$. Also in this section we assume that $G$ is a counterexample to the theorem, and in particular it does not contain finite nonabelian simple subgroups. As a consequence, by Lemma 2.1, for every involution $i \in G$ and element $y \in G$ of order $3$ the order of $iy$ divides $4$ or $6$. Proof of Theorem is based on the following Baer-Suzuki argument for $p=2$, that was obtained in [@bs2014e]. [**Lemma 4.1**]{} [ *Let $G$ be a group of period $n=4k$, where $k$ is odd, containing an involution $i \in G$. If any two elements in $i^G$ generate a $2-$group, then $\langle i^G \rangle$ is a normal $2-$subgroup in $G$.* ]{} Therefore to prove Theorem, it is sufficient to show that if $t$ is an involution, then either $(at)^4=1$ or $\langle t,a,x \rangle$ contains a finite nonabelian simple subgroup. This strategy is implemented in the following lemmas, where we consequently show that in corresponding situations our involution $t$ cannot be a counterexample. Possibilities for the subgroup $\langle t,x \rangle$ are restricted by Lemma 2.1. First of all we consider the case, when $\langle t,x \rangle$ is a homomorphic image of $F_{294}$, in particular, when $[t,x]=1$. [l]{}\[340pt\][20mm]{} {width="20mm"} [**Lemma 4.2**]{} [ *If $t$ is an involution and $[t,x]=1$, then $(at)^4=1$.* ]{} Note that $\langle t,a,x \rangle$ is a homomorphic image of $G(i_1,i_2,...,i_9)=\langle a,t,x| 1=a^2=t^2=x^3=(ax)^3=[t,x]=(ta)^{i_1}=(ta^x)^{i_2}=(ata^x)^{i_3}=(a^x(at)^2)^{i_4}=[a^x,(at)^2]^{i_5}=(a^x(at)^3)^{i_6}=(a(ta^x)^3)^{i_7}=(atx)^{i_8}=((at)^2x)^{i_9} \rangle$, where $i_1,...,i_9 \in \{ 4,5,6,7 \}$. Computations show that this group is finite and is either isomorphic to $S_5$, or is a $\{2,3 \}$-group, and in the last case $O_2(\langle a,x \rangle) \subseteq O_2( \langle a,t,x \rangle)$. The lemma is proved. Note that relations that we use are taken from [@lmmj2014e Lemma 7.A]. [**Lemma 4.3**]{} [*For any involution $t \in G$ the order of $tt^x$ is not $7$.*]{} Assume the contrary. By Lemmas 2.1 and 2.3 there is an involution $z$ such that $\langle z,x \rangle$=$F_{42}$. By Statement 1 $C_G(x)$ has only one involution $u=z^{x^2z}$. By Lemma 4.2 we have $(ua)^4=(ua^x)^4=1$. Therefore $\langle u,a,x \rangle$ is a homomorphic image of $L_0=\langle u,a,x | 1=u^2=a^2=x^3=(ax)^3=[u,x]=(ua)^4=(ua^x)^4 \rangle$. Computations show that this group is finite: $|L_0|=192$ and $z_0=(x^2uaxua)^2$ is the central involution of $L_0$. In particular, $z_0 \in C_G(x)$. If $z_0=u$, then looking at $\overline{L_0}=L_0/ \langle uz_0 \rangle \simeq A_4$ we see that $\overline{u}=\overline{1}$, which is not possible. Therefore by Lemma 3.1 we may further assume that $z_0=1$. Hence $\langle u,a,x \rangle$ is a homomorphic image of $L_1=L_0/ \langle z_0=1 \rangle$. In $L_1$ let $i=(ua)^2$. Then $i^2=(ix)^3=[u,i]=1$. First assume that $i \not = 1$. Then $\langle i,x \rangle \simeq A_4$, and we substitute $a$ by $i$ and consider the group $\langle i,z,x \rangle$ to simplify the relations. Observe that the following relations hold: $R= \{ i^2, z^2, x^3, (xz)^6, b^7, b^xb^{-4}, (ix)^3, (ui)^2 \}$, where $b=z^xz$, $u=z^{x^2z}$. Let $y=z^i$. By Lemma 2.1 for $\langle y ,x \rangle$ one of the following possibilities holds: - $(yx)^6=(y^xy)^7=1$. Let $v=y^{[x^{-1},y][x,y]}$. By Lemma 2.3 $[v,x]=1$ and by Statement 1 $u=v$. Computations show that the order of $G(k_1)=\langle i,z,x | R \cup \{ (iz)^{k_1}, (yx)^6, (y^xy)^7, uv^{-1} \} \rangle$ for $k_1 \in \{5,7 \}$ divides $3$. Note that $G(4)=G(6)=G(2) \simeq C_7:(C_2 \times A_4)$ has an element of order $14$, and the corresponding homomorphic image without it is just $\langle x,z \rangle$. A contradiction. - $(yx)^6=(y^xy)^5=1$. Let $v=y^{xy(x^{-1}y)^2}$. Then $[v,x]=1$ and by Statement 1 $u=v$. Computation show that the order of $G(k_1)=\langle i,z,x | R \cup \{ (iz)^{k_1}, (yx)^6, (y^xy)^5, uv^{-1} \} \rangle $ for $k_1 \in \{4,5,6,7 \}$ divides $24$. A contradiction. - $(yx)^6=(y^xy)^6=1$. Elements $h=(y^xy)^2$ and $k=[h,x]$ of the group $\langle y,x | y^2=x^3=(yx)^6=(y^xy)^6=1 \rangle$ satisfy the following conditions: $\langle h,x \rangle \simeq 3^{1+2}$ is the extraspecial group of exponent 3 and order $27$, and $k$ is the element, generating its center. Moreover, the involution $y$ inverts both $h$ and $k$. Note that $uy=uizi\sim iuiz=uz \in \Gamma_7$. By Lemma 3.3 $[u,k]=1$. Therefore the element $(uy)^2$ of order $7$ centralizes $k$. It follows that $k=1$. Changing $k$ to $h$, in an analogous way we obtain that $h=1$. Therefore this case is a particular case of the following one. - $(yx)^6=(y^xy)^4=1$. Let $v=yx^{-1}yxyxyx^{-1}y$. Then $[v,x]=1$ and by Statement 1 either $u=v$, or $v=1$. Computations show that the order of $$G(k_1,\epsilon)=\langle i,z,x | R \cup \{ (iz)^{k_1}, (yx)^6, (y^xy)^4, u^{\epsilon}v \} \rangle$$ for $k_1 \in \{4,5,6,7 \}, \epsilon \in \{0,1 \}$ divides $24$. A contradiction. - $(yx)^4=1$. Let $v=y^{xy}$. Then $x^v=x^{-1}$. Therefore $vuv \in C_G(x)$ and by Statement 1 $u=vuv$. Computations show that the order of $$G(k_1,k_2)=\langle i,z,x | R \cup \{ (iz)^{k_1}, (izx)^{k_2}, (yx)^4, (uv)^2 \} \rangle$$ divides $2$ for $k_1, k_2 \in \{4,5,6,7 \}$. A contradiction. In this way we obtain that $i=1$. Then $(ax)^3=(ua)^2=1$, changing $i$ to $a$ and repeating arguments above we obtain that $a=1$. A contradiction. The lemma is proved. [l]{}\[340pt\][20mm]{} {width="20mm"} [**Lemma 4.4**]{} [*Let $x,y,z$ be elements of order $3$ such that $(xy)^2=(yz)^2=1$ and $H=\langle x,y,z \rangle$. Then $H$ is an extension of a $2$-group by a group isomorphic to $A_4$ and the following relations hold: $R=\{(xz)^6, (x^{-1}z)^4, (xyz)^6, (xyzxz)^4, (xzy)^6, (x^{-1}z^y)^4 \}$ and $(y^{-1}zx)^4=1$.*]{} Note that $x \cdot zy$ and $xzy= xy \cdot z^y$ are products of an involution and an element of order $3$, hence their orders divide $4$ or $6$. Let’s use the following notations for sets of relations in the alphabet $ \{x,y,z \}$. $$B= \{ x^3, y^3, z^3, (xy)^2, (yz)^2 \},$$ $$R(k_1,\dots,k_6)= \{ (xz)^{k_1}, (x^{-1}z)^{k_2}, (xyz)^{k_3}, (xyzxz)^{k_4}, (xzy)^{k_5}, (x^{-1}z^y)^{k_6} \}.$$ Therefore, $H$ is a homomorphic image of $G(k_1,\dots,k_6)= \langle x,y,z | B \cup R(k_1,\dots,k_6) \rangle $, where $k_1,k_2,k_4, k_6 \in \{ 4,5,6,7 \}$, $k_3,k_5 \in \{4,6\}$. Note that $G(6,7,6,7,6,7)$ is a homomorphic image of $K(k_7)= \langle x,y,z | B \cup R(6,7,6,7,6,7) \cup (y \cdot (xz)^3)^{k_7} \rangle $, where $k_7 \in \{4,6\}$. Computations show that the index in $K(6)$ of subgroup $ \langle x, yz \rangle$, generated by an involution and an element of order $3$, is trivial. Which is not possible by Lemma 2.1. Similarly, $|K(4): \langle y,z \rangle|=1$. The group $G(6,5,6,6,6,5)$ is isomorphic to $M_{12} \times C_3$, which is not possible. Computation show that the index in $G(6,7,6,6,6,7)$ of $ \langle x, yz \rangle$ is trivial. The group $H=G(6,4,6,4,6,4)$ has a normal $2$-subgroup $V$ of order $2^{11}$ and nilpotency class $2$ with $Z(V)$ of order $2^3$, and $H/V \simeq A_4$. The coset action of $H$ on $\langle x,yz \rangle$ provides an exact permutation representation of $H$ on $2^8$ points. Now it is straightforward to check that $H$ and its appropriate homomorphic images satisfy the conclusion of the lemma. The group $G(7,7,6,6,6,7)$ is isomorphic to $L_2(13)$, which is not possible. Computations show that the orders of other groups $G(k_1,\dots,k_6)$ divide $96$ and they are homomorphic images of $H$. The lemma is proved. [l]{}\[340pt\][20mm]{} {width="20mm"} [**Lemma 4.5**]{} [*Let $t \in G$ be an involution such that $\langle t,x \rangle \simeq A_4$. Then $(at)^4=1$.*]{} Let $x_1=tx^{-1}$, $y_1=x$, $z_1=x^{-1}a$. Then $(x_1y_1)^2=(y_1z_1)^2=1$ and the statement follows from Lemma 4.4. Indeed, $at=y_1z_1x_1y_1 \sim y_1^{-1}z_1x_1$. The lemma is proved. [l]{}\[340pt\][20mm]{} {width="20mm"} [**Lemma 4.6**]{} [*Let $x,y,w \in G$ be elements of order $3$ satisfying the following relations: $(xy)^2=[y,w]=1$. Then $\langle x,y,w \rangle$ is finite and $((xy)(xy)^w)^4=1$.*]{} Let $z=x^w$. Then $xy \sim x^wy^w \sim x^wy \sim yz$, and hence $(yz)^2=1$. Тherefore conditions of Lemma 4.5 are true for $\langle x,y,z \rangle$. Denote by $R$ the relations from the conclusion of Lemma 4.5, and by $B$ the relations from the proof of Lemma 4.5. Then note that $\langle x,y,w \rangle$ is a homomorphic image of $G(k_1,k_2,k_3)=\langle x,y,w | R \cup B \cup \{ w^3, [y,w], (xw)^{k_1}, (x^{-1}w)^{k_2}, (xyw)^{k_3} \} \rangle$, where $k_1,k_2,k_3 \in \{4,5,6,7 \}$. Note that the group $H=G(6,6,6)$ has a normal $2$-subgroup $V$ of order $2^{10}$ and nilpotency class $2$ with $Z(V)$ of order $2^4$, and $H/V \simeq C_3 \times C_3$. Computations show that orders of $H/ \langle ((xy)(xy)^w)^4 \rangle$ and $H$ coincide, meaning that the required identity holds in $H$ and its homomorphic images. Computations show that $G(7,7,7) \simeq L_3(4)$, which is not possible, and orders of $G(k_1,k_2,k_3)$ divide $12$ in other cases. The lemma is proved. [**Lemma 4.7**]{} [*Let $t\in G$ be an involution such that $(at)^3=(tx)^6=1$. Then $(tt^x)^6 \not = 1$.* ]{} Assume the contrary. Note that elements $h=(t^xt)^2$ and $k=[h,x]$ of $\langle t,x | 1=t^2=x^3=(tx)^6=(t^xt)^6 \rangle \simeq F_{216}$ satisfy the following conditions: $\langle h,x \rangle$ is an extraspecial group of period $3$ and order $27$, $k$ is its central element. Moreover, $h^t=h^{-1}$ and $k^t=k^{-1}$. Elements $ax^{-1}, x, k$ satisfy the conditions of Lemma 4.6, hence $(a^ka)^4=1$. Let $c=tk$. Note that $\langle a,t,c \rangle$ is a homomorphic image of $G(k_1,k_2,k_3,k_4) = \langle 1=a^2=t^2=c^2=(at)^3=(ct)^3=(a^{tc}a)^4=(ac)^{k_1}=(at \cdot c)^{k_2}=(ta^c)^{k_3}=(at \cdot a^c)^{k_4} \rangle$, where $k_1, k_3 \in \{4,5,6,7\}$, $k_2, k_4 \in \{4,6\}$. Computations show that orders of these groups divide $6$. This way we obtain that $k=1$. It follows that $[h,x]=1$. Again using Lemma 4.6 $(a^ha)^4=1$. In an analogous way, considering $\langle a,t,h \rangle$, we obtain that $h=1$. So we may assume that $(tt^x)^2=1$. Note that $u=tt^xtt^{x^{-1}}t$ is the central involution from $\langle t,x | 1=t^2=x^3=(tx)^6=[t,x]^2 \rangle \simeq C_2 \times A_4$. By Lemma 4.2 in $G$ the following relations hold $(au)^4=(a^xu)^4=1$. Applying Lemma 2.2 to $\langle t,a,a^x \rangle$ we obtain that the order of $ta^x$ divides $4$ or $6$. Therefore the subgroup $\langle a,t,x \rangle$ is a homomorphic image of $G(k_1,k_2,k_3)=\langle a,t,x | 1=a^2=t^2=x^3=(at)^3=(ax)^3=(tx)^6=(t^xt)^2=(au)^4=(a^xu)^4= (ta^x)^{k_1}=(t \cdot xa)^{k_2}=(ax \cdot t)^{k_3} \rangle$, where $k_1,k_2,k_3 \in \{4,6\}$. Computations show that the order of $G(k_1,k_2,k_3)$ divides $12$. Which implies $t=1$. A contradiction. [l]{}\[340pt\][20mm]{} {width="20mm"} [**Lemma 4.8**]{} [*Let $t \in G$ be an involution. Then $\langle a,t \rangle \not \simeq S_3$.*]{} Assume the contrary. Suppose that $bt$ and $abt$ are elements of order $6$, where $b=a^x$. Then $w=t^{bta}$ satisfies the following relations $(aw)^3=(bw)^2=(abw)^6=1$ by [@mamontov2013e Лемма 5], and $\langle a,b,w \rangle \simeq D_{12}$. By Lemma 2.2 changing if necessary $t$ to $w$, and $a$ to $b$ or $ab$, we may assume that $(tb)^6=(tab)^4$. Suppose that $(tx)^4=1$. Then $\langle a,t,x \rangle$ is a homomorphic image of $G(k)=\langle a,t,x |1=a^2=t^2=x^3=(tx)^4=(ax)^3=(at)^3=(tb)^6=(tab)^4=(ax \cdot t)^k \rangle$, where $k \in \{4,6\}$. Computations show that $G(6) \simeq V \rtimes L_2(7)$ and $G(4) \simeq A_6$, where $V$ is elementary abelian of order $2^3$. Therefore we may assume that the order of $tx$ is not equal to $4$. Let $t_1=a^t$. Then $t_1$ and $t$ are conjugated with $a$ and so sit in $\Delta$ and centralize more than one involution. By Statement 2, Lemmas 4.3, 4.7, 2.1 and arguments above $(tx)^6=(t^xt)^4=1$. Similarly, $(t_1x)^6=(t_1^xt_1)^4=1$. The involution $u=tt^xtt^{x^{-1}}t$ of the group $\langle t,x | (tx)^6=(t^xt)^4=1 \rangle$ centralizes $x$. By Lemma 4.2 $(au)^4=1$. Note that $\langle a,t,x \rangle$ is a homomorphic image of $G(k_1,k_2) = \langle a,t,x | 1=a^2=t^2=x^3=(ax)^3=(at)^3=(au)^4=(tb)^6=(tab)^4=(tx)^6=(t^xt)^4=(t_1x)^6=({t_1}^xt_1)^4=(ax \cdot t)^{k_1}=(xa \cdot t)^{k_2} \rangle$, where $k_1,k_2 \in \{4,6\}$. Computations show that the order of $G(k_1,k_2)$ divides $12$. Hence $t=1$ and this case is not possible. The lemma is proved. Note that we obtained a different proof of [@mamontov2013e Statement 1]. [l]{}\[340pt\][20mm]{} {width="20mm"} [**Lemma 4.9**]{} [*Assume that involutions $a,t,f$ satisfy the following relations: $(at)^7=(af)^4=1$ and $[t,f]=1$. Then $(atf)^7=(t(af)^2))^4=1$ and $\langle a,t,f \rangle$ is double frobenius $V:C_7:C_2$, where $V$ is elementary abelian of order $2^6$.*]{} To prove that the group is finite note that $\langle a,t,f \rangle$ is a homomorphic image of $G(i_1,i_2)=\langle a,t,f |1=a^2=t^2=f^2=(at)^7=(af)^4=(tf)^2=(atf)^{i_1}=(t(af)^2)^{i_2} \rangle$, where $i_1, i_2 \in \{4,5,6,7\}$. Computations show that $G(7,4)$ is finite and satisfy conclusion of the lemma, while the order of $G(i_1,i_2)$ is not greater than $14$ for other values of the parameters, which is not possible. The lemma is proved. By Statement 2 and Lemma 4.8 we may assume that $(at)^7=1$. First suppose that $(tx)^6=(t^xt)^2=1$. Then we may assume that $\langle t^xt ,x \rangle \simeq A_4$ and that $u=tt^xtt^{x^{-1}}t$ is an involution, which centralizes $x$. By Lemma 4.5 $(a \cdot t^xt)^4=(a^x \cdot t^xt)^4=1$. By Lemma 4.2 $(au)^4=1$. Using Lemma 4.8 we may assume that $3$ does not divide the order of $a^xt$. Consequently, $\langle a,t,x \rangle$ is a homomorphic image of $G(k)= \langle a,t,x | 1=a^2=t^2=x^3=(ax)^3=(tx)^6=(tt^x)^2=(au)^4=(att^x)^4=(a^xtt^x)^4=(at)^{7}=(a^xt)^{k} \rangle$. Computations show that the order of $G(k)$ divides $12$. Now suppose that $tx \in \Gamma_6$ and $t^xt \in \Gamma_4$. Denote $y=(xt)^2$ and $v=(xt)^3$. Obviously, $[y,v]=1$. In $\langle x,t \rangle$ element $y$ of order $3$ sits inside the subgroup $\langle (t^xt)^2,y \rangle$ isomorphic to $A_4$. By Lemma 4.3 the order of $aa^y$ is not 7. Therefore, by Statement 2 and Lemma 4.8 $(aa^y)^4=1$. In $\langle x,t \rangle$ we also have $(v^xv)^2=1$. By arguments above $(av)^4=1$. Let’s prove that $(aa^{xt})^4=1$. - First assume that $ay \in \Gamma_6$. Let $u=aa^yayay^{-1}a$. Then $u^2=1$ and $[u,y]=1$. Having $u,v \in C_G(y)$, the order of $uv$ divides $6$. If $\langle u,v \rangle$ has an element of order $3$, then let $w$ be an involution such that the order of $uw$ is 3. Note that if $a^g$ is an arbitrary conjugate of $a$, then the order of $a^gv$ is even. In other case $a$ is conjugated with $v$ and hence inverts an element of order $3$, which is in contradiction with Lemma 4.8. Therefore $\langle a,y,w \rangle$ is a homomorphic image of $G(i)=\langle a,y,w | 1=a^2=y^3=w^2=(ay)^6=(a^ya)^4=w^yw=(aw)^4=(uw)^3=(au^w)^4=(a^{ya}w)^4=(ayw)^{i} \rangle$. Computations show that $G(5) \simeq S_6$ and for other $i \in \{4,5,6,7\}$ the order of $G(i)$ is not greater than $| \langle a,y \rangle | =96$. So this case is not possible. Now suppose that $(uv)^2=1$. Then $\langle a,y,v \rangle$ is a homomorphic image of $G(i_1,i_2)=\langle a,y,v | 1=a^2=y^3=v^2=(ay)^6=[a,y]^4=[y,v]=(av)^4=(uv)^2=(ayv)^{i_1}=((av)^2y)^{i_2} \rangle$, where $i_1,i_2 \in \{4,5,6,7\}$; $u=aa^yayay^{-1}a$. Computation show that such group is always finite and contains an element of order $3$ only if $i_1=i_2=6$. In the last case $G(6,6)$ is $\{2,3 \}$-group of order $2^{18}3$, where by Lemma 4.8 we have the required equality $(aa^{xt})^4=1$. - Let now $ay \in \Gamma_4$. Then $\langle a,y,v \rangle$ is a homomorphic image of $G(i_1,i_2) = \langle a,y,v | 1=a^2=y^3=v^2=(ay)^4=[y,v]=(av)^4=(ayv)^{i_1}=((av)^2y)^{i_2} \rangle$, where $i_1,i_2 \in \{4,5,6,7\}$. In $G(4,6)$ we have $|aa^{yv}|=6$, which is not possible by Lemma 4.8. Further note that $G(6,4) \simeq C_2 \times S_6$, which is no possible by Statement 2. Computation show that $G(i_1,i_2)$ does not have elements of order 3 for other values of the parameters $i_1$ and $i_2$. Therefore, this case is not possible. So further we assume that $(aa^{xt})^4=1$. Then $(a^xa^t)^4=1$ and by Lemma 4.9 we obtain that $\langle a,a^t,a^x \rangle =\langle a,t,a^x \rangle$ is the double frobenius group. In particular, $(ta^x)^4=1$ and $(taa^x)^7=(a^taa^x)^7=1$. Note that in $\langle t,x \rangle$ we have $(x^{-1}t)^6=1$, $\langle (x^{-1}t)^3,x \rangle \simeq C_2 \times A_4$. Therefore repeating arguments above we obtain that $(aa^{x^{-1}t})^4=1$. Consequently, $(a^ta^{x^{-1}})^4=(a^taa^x)^4=1$. A contradiction. The theorem is proved. [1]{} The GAP Group, , http://www.gap-system.org. D. Lytkina, V. Mazurov, , J. Sib. Fed. Univ. Math. Phys., 7(2):191–203, 2014. D. V. Lytkina, V. D. Mazurov, A. S. Mamontov, E. Jabara, , Algebra and Logic, 53(5):365–376, 2014. A. S. Mamontov, , Siberian Math. J., 54(1):114–118, 2013. A. S. Mamontov, , Algebra and Logic, 53(5):422–424, 2014. R.Brandl, W.Shi, , J. Algebra, 143:388–400, 1991. V. P. Shunkov, , Algebra and Logic, 11(4):260–272, 1972. A.V. Vasil’ev, , J. Algebra, 423, 2015. [^1]: Mamontov, A., On periodic groups isospectral to $A_7$ [^2]: © 2018 Mamontov A [^3]: This research is supported by RSF (project 14-21-00065)

--- abstract: 'Let $\Gamma$ be a lattice in $G=\hbox{\rm SL}(n,\mathbb{R})$ and $X=G/S$ a homogeneous space of $G$, where $S$ is a closed subgroup of $G$ which contains a real algebraic subgroup $H$ such that $G/H$ is compact. We establish uniform distribution of orbits of $\Gamma$ in $X$ analogous to the classical equidistribution on torus. To obtain this result, we first prove an ergodic theorem along balls in the connected component of Borel subgroup of $G$.' address: 'Department of Mathematics, University of Michigan, Ann Arbor, MI 48109' author: - Alexander Gorodnik title: 'Lattice action on the boundary of $\hbox{\rm SL}(n,\mathbb{R})$' --- [^1] Introduction ============ Let $G=\hbox{\rm SL}(n,\mathbb{R})$ and $\Gamma$ a lattice in $G$; that is, $\Gamma$ is a discrete subgroup of $G$ with finite covolume. We study the action of $\Gamma$ on a compact homogeneous space $X$ of algebraic origin. Namely, $X=G/S$ where $S$ is a closed subgroup of $G$ which contains the connected component of a real algebraic subgroup $H$ of $G$ such that $G/H$ is compact. An important example is provided by the Furstenberg boundary of $G$ [@f63]. In this case, $X=G/B$ where $B$ is the subgroup of upper triangular matrices in $G$. It is possible to deduce from a result of Dani [@st Theorem 13.1] that every orbit of $\Gamma$ in $X$ is dense. We will prove a quantitative estimate for the distribution of orbits. Introduce a norm on $G$: $$\label{eq_norm0} \|g\|=\left(\hbox{Tr}({}^tgg)\right)^{1/2}=\left(\sum_{i,j} g_{ij}^2\right)^{1/2}\quad\hbox{for}\quad g=(g_{ij})\in G.$$ For $T>0$, $\Omega\subseteq X$, and $x_0\in X$, define a counting function $$\label{eq_NT} N_T(\Omega, x_0)=|\{\gamma\in \Gamma: \|\gamma\|<T, \gamma\cdot x_0\in \Omega\}|.$$ Let $m$ be a normalized $\hbox{SO}(n)$-invariant measure on $X$. It follows from the Iwasawa decomposition (see (\[eq\_iwasawa\]) below) that $X$ is a homogeneous space of $\hbox{SO}(n)$. Therefore, the measure $m$ is unique. The following theorem shows that orbits of $\Gamma$ in $X$ are uniformly distributed with respect to the measure $m$. \[th\_asympt\] For a Borel set $\Omega\subseteq X$ such that $m(\partial \Omega)=0$ and $x_0\in X$, $$\label{eq_asympt} N_T(\Omega, x_0)\sim \frac{\gamma_n}{\bar\mu(\Gamma\backslash G)} m(\Omega) T^{n(n-1)}\quad\hbox{as}\quad T\rightarrow\infty,$$ where $\gamma_n$ is a constant (defined in (\[eq\_gamma\]) below), and $\bar\mu$ is a finite $G$-invariant measure on $\Gamma\backslash G$ (defined in (\[eq\_mubar\])). It would be interesting to obtain an estimate for the error term in (\[eq\_asympt\]). This, however, would demand introducing different techniques than those employed here. Theorem \[th\_asympt\] is analogous to the result of Ledrappier [@l] (see also [@n]) who investigated the distribution of dense orbits of a lattice $\Gamma\subset\hbox{\rm SL}(2,\mathbb{R})$ acting on $\mathbb{R}^2$. Ledrappier used the equidistribution property of the horocycle flow. Similarly, we deduce Theorem \[th\_asympt\] from an equidistribution property of orbits of Borel subgroup. Denote by $B^o$ the connected component of the upper triangular subgroup of $G$. To prove Theorem \[th\_asympt\], we use the following ergodic theorem for the right action of $B^o$ on $\Gamma\backslash G$. \[th\_ergodic\] Let $\varrho$ be a right Haar measure on $B^o$, and $\nu$ the normalized $G$-invariant measure on $\Gamma\backslash G$. Then for any $\tilde{f}\in C_c(\Gamma\backslash G)$ and $y\in \Gamma\backslash G$, $$\frac{1}{\varrho(B^o_T)}\int_{B^o_T} \tilde{f}(yb^{-1})d\varrho(b)\rightarrow \int_{\Gamma\backslash G} \tilde{f}d\nu\quad\hbox{as}\quad T\rightarrow\infty,$$ where $B^o_T=\{b\in B^o:\|b\|<T\}$. 1. One can consider the analogous limit for a left Haar measure on $B^o$. In this case, it may happen that the limit is $0$ for some $y\in \Gamma\backslash G$ and all $\tilde{f}\in C_c(\Gamma\backslash G)$ (see Proposition \[l\_contr\]). 2. Since $B^o$ is solvable (hence, amenable), one might expect that convergence for a.e. $y\in\Gamma\backslash G$ follows from known ergodic theorems for amenable group actions. Moreover, since $\nu$ is the only normalized $B^o$-invariant measure on $\Gamma\backslash G$, one could expect that convergence holds for every $y$. However, this approach does not work because the sets $B^o_{T}$ do not form a Følner sequence, and the space $\Gamma\backslash G$ is not compact in general. 3. To prove Theorem \[th\_ergodic\], we use Ratner’s classification of ergodic measures for unipotent flows [@r1]. In fact, we don’t need the full strength of this result. Since the subgroup $U$ (defined in (\[eq\_U\]) below) is horospherical, it is enough to know classification of ergodic measures for horospherical subgroups. The situation is much easier in this special case (see [@st §13]). 4. We expect that analogs of Theorems \[th\_asympt\] and \[th\_ergodic\] hold for a noncompact semisimple Lie group and its irreducible lattice with balls $B_T^o$ defined by the Riemann metric. The main difficulty here is to show that the measure of $B_T^o$ is “concentrated” on the “cone” $B_T^C$ (cf. Lemma \[lem\_BTC2\]). 5. It was pointed out by P. Sarnak that it might be possible to prove the results of this article using harmonic analysis on $\Gamma\backslash G$. In particular, Corollary \[th\_distr\_sl2\] below can be deduced from the result of Good (Corollary on page 119 of [@g]). Note that his method gives an estimate on the error term. The paper is arranged as follows. In the next section, we give examples of applications of Theorem \[th\_asympt\]. In Section \[sec\_ba\] we set up notations and prove some basic lemmas. Theorem \[th\_asympt\] is deduced from Theorem \[th\_ergodic\] in Section \[sec\_th1\]. In Sections \[sec\_uni\] and \[sec\_rep\], we review results on the structure of unipotent flows and prove auxiliary facts about finite-dimensional representations of $\hbox{\rm SL}(n,\mathbb{R})$. Finally, Theorem \[th\_ergodic\] is proved in Section \[sec\_th2\]. Examples {#s_ex} ======== 1. Let $X=\mathbb{R}\cup\{\infty\}$, which is considered as the boundary of the hyperbolic upper half plane. The group $G=\hbox{\rm SL}(2,\mathbb{R})$ acts on $X$ by fractional linear transformations: $$\label{eq_frac} g\cdot x=\frac{ax+b}{cx+d}\quad \hbox{for}\;x\in X,\; g=\left( \begin{tabular}{cc} $a$ & $b$\\ $c$ & $d$ \end{tabular} \right)\in G.$$ Let $\Gamma$ be a lattice in $\hbox{\rm SL}(2,\mathbb{R})$. For $\Omega\subseteq X$ and $x_0\in X$, define the counting function $N_T(\Omega,x_0)$ as in (\[eq\_NT\]). Its asymptotics can be derived from Theorem \[th\_asympt\]. Note that the asymptotics of $N_T(X,x_0)$ as $T\rightarrow\infty$ provides a solution of the so-called hyperbolic circle problem (cf. [@ter1 p. 266] and references therein). \[th\_distr\_sl2\] [(of Theorem \[th\_asympt\])]{} For $x_0\in X$ and $-\infty\le a<b\le +\infty$, $$N_T((a,b),x_0)\sim c_{\Gamma}\left(\int_a^b\frac{dt}{1+t^2}\right) T^2\quad\hbox{as}\quad T\rightarrow\infty,$$ where $c_\Gamma=\frac{1}{2\bar\mu(\Gamma\backslash G)}$ ($\bar\mu$ is the $G$-invariant measure on $\Gamma\backslash G$ defined in (\[eq\_mubar\])). It is easy to see from (\[eq\_frac\]) that $G$ acts transitively on $X$, and the stabilizer of $\infty$ in $G$ is the group of upper triangular matrices $B$. Thus, Theorem \[th\_asympt\] is applicable to the space $X$. Note that $$ K=(2)={k\_=( -------------------- ------------------- $\cos 2\pi\theta$ $\sin 2\pi\theta$ $-\sin 2\pi\theta$ $\cos 2\pi\theta$ -------------------- ------------------- ):\[0,1)}, $$ and the normalized Haar measure on $K$ is given by $dk=d\theta$. The measure $m$ on $X$ can be defined as the image of $dk$ under the map $K\rightarrow X:k\rightarrow k\cdot \infty$. By (\[eq\_frac\]), $k_\theta\cdot\infty=-\hbox{ctan}\; 2\pi\theta$. Then $$m((a,b))=\int_{k\cdot\infty\in (a,b)}dk=\mathop{\int_{-\hbox{\small ctan}\; 2\pi\theta\in (a,b)}}_{\theta\in [0,1)} d\theta =\frac{1}{\pi}\int_a^b\frac{dt}{1+t^2}.$$ We have used the substitution $t=-\hbox{ctan}\; 2\pi\theta$. Finally, by Theorem \[th\_asympt\], $$N_T((a,b),x_0)\mathop{\sim}_{T\rightarrow\infty} \frac{\gamma_2 m((a,b))}{\bar\mu(\Gamma\backslash G)}T^2 =c_{\Gamma}\left(\int_a^b\frac{dt}{1+t^2}\right) T^2.$$ Note that $\gamma_2=\frac{\pi}{2}$ by (\[eq\_gamma\]) below. 2. Let $X=\mathbb{P}^{n-1}$ be the projective space (more generally $X=\mathcal{G}_{n,k}$, Grassmann variety, or $X=\mathcal{F}_n$, flag variety), and $m$ be the rotation invariant normalized measure on $X$. Then the asymptotic estimate (\[eq\_asympt\]) holds for the standard action of $G=\hbox{\rm SL}(n,\mathbb{R})$ on $X$. This is a special case of Theorem \[th\_asympt\] because $X$ can be identified with $G/S$ where $S$ is a closed subgroup of $G$ that contains $B$, the group of upper triangular matrices. Basic facts {#sec_ba} =========== For $s=(s_1,\ldots, s_n)\in\mathbb{R}^n$, define $$a(s)=\hbox{diag}(e^{s_1},\ldots, e^{s_n}).$$ For $t=(t_{ij}:1\le i<j\le n)$, $t_{ij}\in\mathbb{R}$, denote by $n(t)$ the unipotent upper triangular matrix with entries $t_{ij}$ above the main diagonal. We use the following notations: $$\begin{aligned} G&=&\hbox{\rm SL}(n,\mathbb{R}),\nonumber\\ K&=&\hbox{SO}(n),\nonumber\\ A^o&=&\{a(s)|\;s\in\mathbb{R}^n,\;\sum_i s_i=0\},\nonumber\\ N&=&\{n(t)|\; t_{ij}\in\mathbb{R},1\le i<j\le n\},\nonumber\\ B^o&=&A^oN=NA^o. \nonumber\end{aligned}$$ For $s\in\mathbb{R}^n$, $\sum_i s_i=0$, denote $\alpha_{ij}(s)=s_i-s_j$, where $i,j=1,\ldots, n$, $i<j$. These are the positive roots of the Lie algebra of $G$. Note that $$\label{eq_adj} \hbox{Ad}_{a(s)}n\left(\{t_{ij}\}\right)=a(s)n\left(\{t_{ij}\}\right)a(s)^{-1}=n(\{e^{\alpha_{ij}(s)}t_{ij}\}).$$ Let $$\label{eq_varrho} \delta(s)=\frac{1}{2}\sum_{i<j} \alpha_{ij}(s)=\sum_{1\le k\le n} (n-k)s_k.$$ For $C\in \mathbb{R}$, define $$A^C=\{a(s)\in A^o|\;s_i>C,i=1,\ldots, n-1\}.$$ Also put $B^C=A^CN$. Let $dk$ be the normalized Haar measure on $K$. A Haar measure on $N$ is given by $dn=dt=\prod_{i<j}dt_{ij}$. A Haar measure on $A^o$ is $da=ds=ds_1\ldots ds_{n-1}$. The product map $A^o\times N\rightarrow B^o$ is a diffeomorphism. The image of the product measure under this map is a left Haar measure on $B^o$. Denote this measure by $\lambda$. Then a right Haar measure $\varrho$ on $B^o$ can be defined by $$\label{eq_lambda} \varrho (f)=\int_{B^o} f(b^{-1}) \lambda(b)=\int_{A^o\times N} f(a(s)n(t))e^{2\delta(s)}dsdt,\quad f\in C_c(B^o).$$ The map corresponding to the Iwasawa decomposition $$\label{eq_iwasawa} (k,a,n)\mapsto kan: K\times A^o\times N\rightarrow G$$ is a diffeomorphism. One can define a Haar measure $\mu$ on $G$ in terms of this decomposition: $$\label{eq_iwasawa2} \int_G fd\mu=\int_{K\times A^o\times N}f(ka(s)n(t))e^{2\delta(s)}dkdsdt=\int_{K\times B^o} f(kb)dkd\varrho(b)$$ for $f\in C_c(G)$. For a lattice $\Gamma$ in $G$, there exists a finite measure $\bar \mu_\Gamma$ on $\Gamma\backslash G$ such that $$\label{eq_mubar} \int_G fd\mu=\int_{\Gamma\backslash G}\sum_{\gamma\in\Gamma} f(\gamma g)d\bar\mu_\Gamma(g),\quad f\in C_c(G).$$ Let $\beta$ be an automorphism of $G$. Then $\beta(\Gamma)$ is a lattice too. Moreover, the following lemma holds. \[l\_auto\] Define a map $$\bar\beta:\Gamma\backslash G\rightarrow\beta(\Gamma)\backslash G:g\Gamma\mapsto \beta(g)\beta(\Gamma)$$ Then $\bar\beta(\bar\mu_\Gamma)=\bar\mu_{\beta(\Gamma)}$. Since the automorphism group of $G$ is a finite extension of the group of inner automorphisms, and $G$ is unimodular, it follows that that the measure $\mu$ is $\beta$-invariant. Every $\tilde{f}\in C_c(\beta(\Gamma)\backslash G)$ can be represented as $\sum_{\gamma\in\beta(\Gamma)} f(\gamma g)$ for some $f\in C_c(G)$ (see [@rag Ch. 1]). Then $$\begin{aligned} \bar\beta(\bar\mu_\Gamma)(\tilde{f})&=& \int_{\Gamma\backslash G}\sum_{\gamma\in\beta(\Gamma)} f(\gamma \beta(g))d\bar\mu_\Gamma(g) =\int_G f(\beta(g))d\mu(g)\\ &=&\int_G f(g)d\mu(g)=\bar\mu_\Gamma(\tilde{f}).\end{aligned}$$ For a subset $D\subseteq G$ and $T>0$, put $$D_T=\{d\in D: \|d\|<T\}.$$ Note that $$\label{eq_BT} B_T^o=\left\{a(s)n(t): \sum_{1\le i\le n} e^{2s_i}+\sum_{1\le i<j\le n} e^{2s_i}t_{ij}^2<T^2 \right\}.$$ For $s\in\mathbb{R}^n$, define $$\label{eq_Ns} N(s)=\sum_i e^{2s_i}.$$ \[lem\_BTC\] For $C\in\mathbb{R}$, $$\varrho(B_T^C)=c_n\int_{A^C_T} \Big(T^2-N(s)\Big)^{\frac{n(n-1)}{4}}\hbox{\rm exp}\left(\sum_{k} (n-k)s_k\right)ds,$$ where $c_n=\pi^{n(n-1)/4}/\Gamma(1+n(n-1)/4)$. Use formulas (\[eq\_lambda\]), (\[eq\_varrho\]), (\[eq\_BT\]), and make change of variables $t_{ij}\rightarrow e^{-s_i}t_{ij}$. Then the above formula follows from the fact that the volume of the unit ball in $\mathbb{R}^m$ is $\pi^{m/2}/\Gamma(1+m/2)$. It follows from Lemma \[lem\_BTC\] that $\varrho (B_T^o)=O\left(T^{(n^2-n)}\right)$ as $T\rightarrow\infty$. In fact, more precise statement is true: $$\label{eq_BTasy} \varrho(B_T^o)\sim \gamma_n T^{(n^2-n)}\quad\hbox{as}\quad T\rightarrow\infty,$$ where $$\label{eq_gamma} \gamma_n=\frac{\pi^{n(n-1)/4}}{2^{n-1}\Gamma\left(\frac{n^2-n+2}{2}\right)}\prod_{k=1}^{n-1} \Gamma\left(\frac{n-k}{2}\right).$$ Since the norm (\[eq\_norm0\]) is $K$-invariant, $G_T=KB^o_T$. By (\[eq\_iwasawa2\]), $\mu (G_T)=\varrho(B^o_T)$. The asymptotics of $\mu (G_T)$ as $T\rightarrow\infty$ was computed in [@drs Appendix 1]. \[lem\_BTC2\] For any $C\in\mathbb{R}$, $\varrho(B_T^C)\sim\varrho(B_T^o)$ as $T\rightarrow\infty$. For $i_0=1,\ldots, n-1$, put $$A_T^{i_0}=\{a(s)\in A_T^o: s_{i_0}\le C\}\;\;\hbox{and}\;\;B_T^{i_0}=\{a(s)n(t)\in B_T^o: s_{i_0}\le C\}.$$ We claim that $\varrho (B_T^{i_0})=o(\varrho (B_T^o))$ as $T\rightarrow\infty$. It follows from (\[eq\_BT\]) that if $a(s)\in A_T^o$, then $s_i<\log T$ for every $i=1,\ldots, n$. Then by Lemma \[lem\_BTC\], $$\begin{aligned} \nonumber \varrho(B_T^{i_0})&\le& c_nT^{\frac{n(n-1)}{2}}\int_{A_T^{i_0}} \hbox{\rm exp}\left(\sum_{k} (n-k)s_k\right)ds \\ \nonumber &\ll& T^{\frac{n(n-1)}{2}} \prod_{{k<n},{k\ne i_0}} \int_{-\infty}^{\log T} e^{(n-k)s_k}ds_k\ll T^{n(n-1)-(n-i_0)}.\end{aligned}$$ (Here and later on $A\ll B$ means $A < c\cdot B$ for some absolute constant $c>0$.) Now the claim follows from (\[eq\_BTasy\]). Since $B_T^o-B_T^C=\cup_{i_0<n} B_T^{i_0}$, $\varrho (B_T^o-B_T^C)=o(\varrho (B_T^o))$ as $T\rightarrow\infty$. Therefore, $\varrho(B_T^C)\sim\varrho(B_T^o)$ as $T\rightarrow\infty$. Next, we show that Theorem \[th\_ergodic\] fails for a left Haar measure on $B^o$. \[l\_contr\] Let $\Gamma$ be a lattice in $G=\hbox{\rm SL}(2,\mathbb{R})$, and $y\in\Gamma\backslash G$ be such that the orbit $yN$ is periodic. Then for every $\tilde{f}\in C_c(\Gamma\backslash G)$, $$\frac{1}{\lambda(B^o_T)}\int_{B^o_T} \tilde{f}(yb^{-1})d\lambda(b)\rightarrow 0 \quad\hbox{as}\quad T\rightarrow\infty.$$ For $C\in \mathbb{R}$, put $\hat{A}^C=\{a(s)\in A^o:s_1<C\}$ and $\hat{B}^C=\hat{A}^C N$. As in Lemma \[lem\_BTC2\], one can show that $\lambda (B^o_T-\hat{B}^C_T)=o(\lambda (B^o_T))$ as $T\rightarrow\infty$. Therefore, for every $C\in\mathbb{R}$, $$\frac{1}{\lambda(B^o_T)}\int_{B^o_T-\hat{B}^C_T} \tilde{f}(yb^{-1})d\lambda(b)\rightarrow 0 \quad\hbox{as}\quad T\rightarrow\infty.$$ On the other hand, according to [@st Lemma 14.2], $yN^{-1}a(s)^{-1}\rightarrow\infty$ as $s_1\rightarrow -\infty$. Thus, there exists $C\in\mathbb{R}$ such that $y(\hat{B}^C_T)^{-1}\cap \hbox{supp} (\tilde{f})=\emptyset$. Then $$\int_{\hat{B}^C_T} \tilde{f}(yb^{-1})d\lambda(b)= 0.$$ This proves the proposition. Proof of Theorem \[th\_asympt\] {#sec_th1} =============================== [**Claim**]{}. [*It is enough to prove Theorem \[th\_asympt\] for $X=G/B^o$*]{}. Suppose that the theorem is proved for $X=G/B^o$. At first, we consider a special case: $$\label{eq_spec} X=G/(B^o)^{g_0}\quad\hbox{for}\;\;\hbox{some}\;\; g_0\in G,$$ where $(B^o)^{g_0}=g_0^{-1}B^og_0$. By the Iwasawa decomposition (\[eq\_iwasawa\]), $(B^o)^{g_0}=(B^o)^{k_0}$ for some $k_0\in K$. The normalized $K$-invariant measure $m$ on $G/(B^o)^{k_0}$ is defined as $$m(C)=\int_{k: k(B^o)^{k_0}\in C} dk\quad\hbox{for}\;\; \hbox{Borel}\;\hbox{set}\;\; C\subseteq G/(B^o)^{k_0}.$$ Similarly, one defines the normalized $K$-invariant measure $m^*$ on $G/B^o$. Consider a map $$\beta:G/B^o\rightarrow G/(B^o)^{k_0}: gB^o\mapsto g^{k_0} (B^o)^{k_0}.$$ Clearly, $\beta$ is a diffeomorphism. Using that $K$ is unimodular, one proves that $$\label{eq_m} m^*(\beta^{-1}(C))=m(C)\quad\hbox{for}\;\; \hbox{Borel}\;\hbox{set}\;\; C\subseteq G/(B^o)^{k_0}.$$ Take $$\Omega\subseteq G/(B^o)^{k_0}\quad\hbox{and}\quad x_0=h_0(B^o)^{k_0}\in G/(B^o)^{k_0}$$ such that $m(\partial\Omega)=0$. Set $${\Omega}^*=\beta^{-1}(\Omega)\subseteq G/B^o\quad\hbox{and}\quad {x}^*_0=h_0^{k_0^{-1}}B^o\in G/B^o.$$ By (\[eq\_m\]), $m^*(\partial{\Omega^*})=0$ too. For $\gamma\in\Gamma$, $\gamma\cdot x_0\in\Omega$ iff $\gamma^{k_0^{-1}}\cdot {x}^*_0\in {\Omega}^*$. Note also that $\|\gamma^{k_0^{-1}}\|=\|\gamma\|$. Therefore, $$N_T(\Omega,x_0)=|\{\gamma\in \Gamma^{k_0^{-1}}: \|\gamma\|<T, \gamma\cdot {x}^*_0\in {\Omega}^*\}|$$ Applying the assumption to the lattice $\Gamma^*=\Gamma^{k_0^{-1}}$, one gets $$N_T(\Omega, x_0)\sim \frac{\gamma_n}{\bar\mu^*(\Gamma^*\backslash G)} m^*(\Omega^*) T^{n(n-1)}\quad\hbox{as}\quad T\rightarrow\infty,$$ where $\bar\mu^*$ is the measure on $\Gamma^*\backslash G$ defined in (\[eq\_mubar\]). Now the special case (\[eq\_spec\]) follows from Lemma \[l\_auto\] and (\[eq\_m\]). Let us consider the general case. Let $S$ be a closed subgroup of $G$ such that $S\supseteq H^o$, where $H$ is a real algebraic subgroup of $G$, and $G/H$ is compact. Since $H$ has finitely many connected components, $G/H^o$ is compact too. Recall that the homogeneous space $G/H^o$ is compact iff $H_\mathbb{C}$ contains a maximal connected $\mathbb{R}$-split solvable $\mathbb{R}$-subgroup of $G_\mathbb{C}$ (see, for example, [@pr Theorem 3.1]). Since maximal connected $\mathbb{R}$-split solvable $\mathbb{R}$-subgroups of $G_\mathbb{C}$ are conjugate over $G_\mathbb{R}$ (see [@bt], or Theorem 15.2.5 and Exercise 15.4.8 in [@spr]), for some $g_0\in G$, $B^{g_0}\subseteq H$. Hence, $(B^o)^{g_0}\subseteq H^o\subseteq S$. Denote by $\pi$ the projection map $G/(B^o)^{g_0}\rightarrow G/S$. Take $$\Omega\subseteq G/S\quad\hbox{and}\quad x_0\in G/S$$ such that $m(\partial\Omega)=0$. Set $${\Omega}^*=\pi^{-1}(\Omega)\subseteq G/(B^o)^{g_0}\quad\hbox{and}\quad {x}^*_0\in \pi^{-1}(x_0).$$ Let $m^*$ be the $K$-invariant normalized measure on $G/(B^o)^{g_0}$. Then $m=\pi(m^*)$ is the $K$-invariant normalized measure on $G/S$. It is easy to check that $\pi(\partial\Omega^*)\subseteq \partial\Omega$. Hence, $m^*(\partial\Omega^*)\le m^*(\pi^{-1}(\partial\Omega))=m(\partial\Omega)=0$. Finally, $$N_T(\Omega,x_0)=N_T(\Omega^*,x_0^*)\mathop{\sim}_{T\rightarrow\infty} \frac{\gamma_n}{\bar\mu(\Gamma\backslash G)} m^*(\Omega^*) T^{n(n-1)} =\frac{\gamma_n}{\bar\mu(\Gamma\backslash G)} m(\Omega) T^{n(n-1)}.$$ We need the following proposition that follows easily from Theorem \[th\_ergodic\]. \[p\_ergodic\] Let $f$ be the characteristic function of a relatively compact Borel subset $Z\subseteq G$ such that $\mu (\partial Z)=0$. Then for any $y\in \Gamma\backslash G$, $$\label{eq_char_last} \frac{1}{\varrho(B^o_T)}\int_{B^o_T} \tilde{f}(yb^{-1})d\varrho(b) \longrightarrow \frac{1}{\bar \mu(\Gamma\backslash G)}\int_{G} fd\mu\quad\hbox{as}\;\;T\rightarrow\infty,$$ where $\tilde{f}(\Gamma g)=\sum_{\gamma\in \Gamma} f(\gamma g)\in C_c(\Gamma\backslash G)$. The argument is quite standard. One chooses functions $\phi_n,\psi_n\in C_c(G)$, $n\ge 1$, such that $\phi_n\le f\le \psi_n$ and $\int_G (\psi_n-\phi_n)d\mu<\frac{1}{n}$. By Theorem \[th\_ergodic\] and (\[eq\_mubar\]), $$\lim_{T\rightarrow\infty} \frac{1}{\varrho(B^o_T)}\int_{B^o_T} \tilde{\phi}_n(yb^{-1})d\varrho(b) =\frac{1}{\bar \mu(\Gamma\backslash G)}\int_{\Gamma\backslash G} \tilde{\phi}_n d\bar\mu= \frac{1}{\bar \mu(\Gamma\backslash G)}\int_{G} \phi_n d\mu,$$ and $$\lim_{T\rightarrow\infty} \frac{1}{\varrho(B^o_T)}\int_{B^o_T} \tilde{\psi}_n(yb^{-1})d\varrho(b) =\frac{1}{\bar \mu(\Gamma\backslash G)}\int_{G} \psi_n d\mu$$ for every $n\ge 1$. This implies (\[eq\_char\_last\]). The proof of Theorem \[th\_asympt\] should be compared with similar arguments in [@drs], [@em], [@ems], [@emm] where other counting problems were also reduced to asymptotics of certain integrals. Let $\alpha: K\rightarrow G/B^o$ be a map defined by $\alpha(k)=kB^o$. By (\[eq\_iwasawa\]), $\alpha$ is a diffeomorphism. The measure $m$ is given by $$\label{eq_mOm} m(C)=\int_{\alpha^{-1}(C)}dk\quad\hbox{for}\;\;\hbox{Borel}\;\;\hbox{set}\;\; C\subseteq G/B^o.$$ Since $\alpha$ is surjective, $x_0=k_0^{-1}B^o$ for some $k_0\in K$. It follows from the Iwasawa decomposition (\[eq\_iwasawa\]) that the product map $$K\times B^o:(k,b)\mapsto kk_0^{-1}bk_0\in G$$ is a diffeomorphism. For $g\in G$, define $k_g\in K$ and $b_g\in B^o$ such that $$g=k_gk_0^{-1}b_gk_0.$$ Since $G$ and $K$ are unimodular, it follows from (\[eq\_iwasawa2\]) that for $f\in C_c(G)$, $$\int_G fd\mu= \int_{K\times B^o} f(kk_0^{-1}bk_0)dk d\varrho(b).$$ Let $\phi$ be the characteristic function of $\tilde{\Omega}\stackrel{def}{=}\alpha^{-1}(\Omega)k_0\subseteq K$, and $\psi_{\mathcal{O}}$ the characteristic function of an open bounded symmetric neighborhood ${\mathcal{O}}$ of $1$ in $B^o$ with boundary of measure $0$ normalized so that $\int_{B^o} \psi_{\mathcal{O}} d\varrho=1$. Then $\int_{B^o} \psi_{\mathcal{O}} d\lambda=1$ too. Note that for $g\in G$, $gx_0\in\Omega$ iff $k_g\in\tilde{\Omega}$. Put $f_{\mathcal{O}}(g)=\phi(k_g)\psi_{\mathcal{O}}(b_g)$. Let $\tilde{f}_{\mathcal{O}}(\Gamma g)=\sum_{\gamma\in\Gamma} f_{\mathcal{O}}(\gamma g)$. Now Proposition \[p\_ergodic\] can be applied to $\tilde{f}_{\mathcal{O}}$: $$\begin{aligned} \nonumber \frac{1}{\varrho(B^o_T)}\int_{B^o_T} \tilde{f}_{\mathcal{O}} (\Gamma k_0^{-1} b^{-1} {k_0})d\varrho (b) \mathop{\longrightarrow}_{T\rightarrow\infty} \frac{1}{\bar\mu(\Gamma\backslash G)}\int_{G} {f}_{\mathcal{O}} (gk_0)d\mu(g)\\ =\frac{1}{\bar\mu(\Gamma\backslash G)}\int_K \phi dk\cdot\int_{B^o} \psi_{\mathcal{O}} d\varrho=\frac{1}{\bar\mu(\Gamma\backslash G)}\int_{\tilde{\Omega}} dk=\frac{m(\Omega)}{\bar\mu(\Gamma\backslash G)}. \label{eq_bound0}\end{aligned}$$ The last equality follows from (\[eq\_mOm\]). Take $r>1$. There exists a bounded open symmetric neighborhood ${\mathcal{O}}$ of identity in $B^o$ (with boundary of measure $0$) such that for any $b\in {\mathcal{O}}={\mathcal{O}}^{-1}$ and $x\in M(n,\mathbb{R})$, $$\label{eq_norm} r^{-1}\|x\|\le \|b^{-1}x\|\le r\|x\|.$$ Then for ${\mathcal{O}}$ as above, $$\begin{aligned} \nonumber \sum_{\gamma\in \Gamma} \int_{B^o_T} f_{\mathcal{O}} (\gamma k_0^{-1}b^{-1}{k_0})d\varrho(b) =\sum_{\gamma\in \Gamma} \int_{(B^o_T)^{-1}} f_{\mathcal{O}} (k_\gamma k_0^{-1}b_\gamma b{k_0})d\lambda(b)\\ \nonumber =\sum_{\gamma\in \Gamma} \int_{b_\gamma(B^o_T)^{-1}} f_{\mathcal{O}} (k_\gamma k_0^{-1}b{k_0})d\lambda(b)= \sum_{\gamma\in \Gamma} \int_{\|b^{-1}b_\gamma \|<T} \phi (k_\gamma)\psi_{\mathcal{O}}(b)d\lambda(b)\\ \nonumber =\sum_{\gamma: k_\gamma\in\tilde{\Omega}} \int_{\|b^{-1}b_\gamma \|<T} \psi_{\mathcal{O}}(b)d\lambda(b) =\sum_{\gamma: \gamma\cdot x_0\in \Omega} \int_{\|b^{-1}b_\gamma \|<T} \psi_{\mathcal{O}}(b)d\lambda(b).\end{aligned}$$ The integral $$I_\gamma\stackrel{def}{=}\int_{\|b^{-1}b_\gamma \|<T} \psi_{\mathcal{O}}(b)d\lambda(b)$$ is not greater than $1$. By (\[eq\_norm\]), $I_\gamma=0$ for $\gamma\in\Gamma$ such that $\|\gamma\|=\|b_\gamma\|\ge rT$. Therefore, $$\label{eq_bound1} N_{rT}(\Omega, x_0)\ge\sum_{\gamma\in \Gamma} \int_{B^o_T} f_{\mathcal{O}} (\gamma k_0^{-1}b^{-1}{k_0})d\varrho(b).$$ By (\[eq\_norm\]), $I_\gamma=1$ for $\gamma\in \Gamma$ such that $\|\gamma\|=\|b_\gamma\|< r^{-1}T$. Thus, $$\label{eq_bound2} N_{r^{-1}T}(\Omega,x_0)\le\sum_{\gamma\in \Gamma} \int_{B^o_T} f_{\mathcal{O}} (\gamma k_0^{-1}b^{-1}{k_0})d\varrho(b).$$ It follows from (\[eq\_BTasy\]) that $\varrho(B_{r^{-1}T}^o)\sim \gamma_n r^{n-n^2} T^{n^2-n}$ as $T\rightarrow\infty$. Then using (\[eq\_bound1\]) and (\[eq\_bound0\]), we get $$\begin{aligned} \nonumber \liminf_{T\rightarrow\infty} \frac{N_T(\Omega,x_0)}{T^{n^2-n}}&\ge& \liminf_{T\rightarrow\infty} \frac{\gamma_n}{r^{n^2-n}\varrho(B^o_{r^{-1}T})}\int_{B^o_{r^{-1}T}} \tilde{f}_{\mathcal{O}} (\gamma k_0^{-1}b^{-1}{k_0})d\varrho(b)\\ \nonumber &=&\frac{\gamma_n m(\Omega)}{r^{n^2-n}\bar\mu(\Gamma\backslash G)}.\end{aligned}$$ This inequality holds for any $r>1$. Hence, $$\liminf_{T\rightarrow\infty} \frac{N_T(\Omega,x_0)}{T^{n^2-n}}\ge \frac{\gamma_n m(\Omega)}{\bar\mu(\Gamma\backslash G)}.$$ Similarly, using (\[eq\_bound2\]) and (\[eq\_bound0\]), one can prove that $$\limsup_{T\rightarrow\infty} \frac{N_T(\Omega,x_0)}{T^{n^2-n}}\le \frac{\gamma_n m(\Omega)}{\bar\mu(\Gamma\backslash G)}.$$ This finishes the proof of Theorem \[th\_asympt\]. Behavior of unipotent flows {#sec_uni} =========================== In this section we review some deep results on equidistribution of unipotent flows, which are crucial for the proof of Theorem \[th\_ergodic\]. Note that there are two different approaches available: Ratner [@r2] and Dani, Margulis [@dm93]. Both of these methods rely on Ratner’s measure rigidity [@r1]. We follow the method of Dani and Margulis. The results below were proved in [@dm91; @dm93] for the case of one-dimensional flows and extended to higher dimensional flows and even polynomial trajectories in [@sh94; @ems97; @sh96]. See [@st §19] and [@kss01] for more detailed exposition. Appropriate adjustments are made for the right $G$-action on $\Gamma\backslash G$ instead of left $G$-action on $G/\Gamma$. [Notations:]{} Let $G$ be a connected semisimple Lie group without compact factors, and $\Gamma$ a lattice in $G$. Let $\mathfrak g$ be the Lie algebra of $G$. For positive integers $d$ and $n$, denote by ${\mathcal{P}}_{d,n}(G)$ the set of functions $q:\mathbb{R}^n\rightarrow G$ such that for any $a,b\in\mathbb{R}^n$, the map $$t\in\mathbb{R}\mapsto \hbox{Ad}(q(at+b))\in\mathfrak{g}$$ is a polynomial of degree at most $d$ with respect to some basis of $\mathfrak{g}$. Let $V_G=\oplus_{i=1}^{\dim \mathfrak{g}}\wedge^i\mathfrak{g}$. There is a natural action of $G$ on $V_G$ induced from the adjoint representation. Fix a norm on $V_G$. For a Lie subgroup $H$ of $G$ with Lie algebra $\mathfrak{h}$, take a unit vector $p_H\in \wedge^{\dim \mathfrak{h}}\mathfrak{h}$. The following theorem allows us to estimate divergence of polynomial trajectories. For its proof, see [@sh96 Theorems 2.1–2.2]. \[th\_help1\] There exist closed subgroups $U_i$ ($i=1,\ldots, r$) such that each $U_i$ is the unipotent radical of a maximal parabolic subgroup, $\Gamma U_i$ is compact in $\Gamma\backslash G$, and for any $d,n\in\mathbb{N}$, $\varepsilon,\delta>0$, there exists a compact set $C\subseteq \Gamma\backslash G$ such that for any $q\in {\mathcal{P}}_{d,n}(G)$ and a bounded open convex set $D\subseteq \mathbb{R}^n$, one of the following holds: 1. There exist $\gamma\in\Gamma$ and $i=1,\ldots, r$ such that $\sup_{t\in D} \|q(t)^{-1}\gamma\cdot p_{U_i}\|\le\delta.$ 2. $\ell(t\in D: \Gamma q(t)\notin C)< \varepsilon \ell(D)$, where $\ell$ is the Lebesgue measure on $\mathbb{R}^n$. Denote by ${\mathcal{H}}_\Gamma$ the family of all proper closed connected subgroups $H$ of $G$ such that $\Gamma\cap H$ is a lattice in $H$, and $\hbox{Ad}(H\cap \Gamma)$ is Zariski-dense in $\hbox{Ad}(H)$. \[th\_help2\] The set ${\mathcal{H}}_\Gamma$ is countable. For any $H\in {\mathcal{H}}_\Gamma$, $\Gamma\cdot p_H$ is discrete. For proofs, see [@r1 Theorem 1.1] and [@dm93 Theorems 2.1 and 3.4]. Fix a subgroup $U$ generated by $1$-parameter unipotent subgroups. For a closed subgroup $H$ of $G$, denote $$X(H,U)=\{g\in G: gU\subseteq Hg\}.$$ Define a singular set $$\label{eq_sing} Y=\bigcup_{H\in {\mathcal{H}}_\Gamma} \Gamma X(H,U)\subseteq \Gamma\backslash G.$$ It follows from Dani’s generalization of Borel density theorem and Ratner’s topological rigidity that $y\in Y$ iff $yU$ is not dense in $\Gamma\backslash G$ (see [@st Lemma 19.4]). One needs to estimate behavior of polynomial trajectories near the singular set $Y$. The following result can be deduced from [@sh94 Proposition 5.4]. It is formulated in [@sh96] and [@kss01]. Note that it is analogous to Theorem \[th\_help1\] with the point at infinity being replaced by the singular set. \[th\_help3\] Let $d,n\in\mathbb{N}$, $\varepsilon>0$, $H\in {\mathcal{H}}_\Gamma$. For any compact set $C\subseteq \Gamma X(H,U)$, there exists a compact set $F\subseteq V_G$ such that for any neighborhood $\Phi$ of $F$ in $V_G$, there exists a neighborhood $\Psi$ of $C$ in $\Gamma\backslash G$ such that for any $q\in {\mathcal{P}}_{d,n}(G)$ and a bounded open convex set $D\subseteq \mathbb{R}^n$, one of the following holds: 1. There exists $\gamma\in\Gamma$ such that $q(D)^{-1}\gamma\cdot p_{H}\subseteq \Phi$. 2. $\ell(t\in D: \Gamma q(t)\in \Psi)< \varepsilon \ell(D)$, where $\ell$ is the Lebesgue measure on $\mathbb{R}^n$. Representations of $\hbox{\rm SL}(n,\mathbb{R})$ {#sec_rep} ================================================ In order to be able to use the results from the previous section, we collect here some information about representations of $\hbox{\rm SL}(n,\mathbb{R})$. The next lemma is essentially Lemma 5.1 from [@sh96]. We present its proof because more precise information about dependence on $\beta$ in the inequality (\[eq\_shineq\]) is needed. \[lem\_shah\] Let $V$ be a finite dimensional vector space with a norm $\|\cdot\|$, $\mathfrak n$ be a nilpotent Lie subalgebra of $\hbox{\rm End}(V)$ with a basis $\{b_i:i=1,\ldots, m\}$, and $N=\exp(\mathfrak{n})\subseteq \hbox{\rm GL}(V)$ be the Lie group of $\mathfrak n$. Define a map $\Theta:\mathbb{R}^m\rightarrow N$: $$\Theta(t_1,\ldots, t_m)=\exp\left(\sum_{i=1}^m t_ib_i\right),\quad (t_1,\ldots, t_m)\in\mathbb{R}^m.$$ For $\beta>0$, put $D(\beta)=\Theta\left([0,\beta]\times\cdots \times[0,\beta]\right)$. Let $$W=\{v\in V: \mathfrak{n}\cdot v=0\}.$$ Denote by $\hbox{\rm pr}_W$ the orthogonal projection on W with respect to some scalar product on $V$. Then there exists a constant $c_0>0$ such that for any $\beta\in (0,1)$ and $v\in V$, $$\label{eq_shineq} \sup_{n\in D(\beta)} \|\hbox{\rm pr}_W(n\cdot v)\|\ge c_0\beta^d\|v\|,$$ where $d$ is the degree of the polynomial map $\Theta$. Let $${\mathcal{I}}=\left\{(i_1,\ldots, i_m)\in\mathbb{Z}^m: i_k\ge 0,\sum_k i_k\le d\right\}.$$ For $t\in\mathbb{R}^m$ and $I=(i_1,\ldots, i_m)\in {\mathcal{I}}$, denote $t^I=\prod_k t_k^{i_k}$, and $|I|=\sum_k i_k$. One can write $\Theta(t)=\sum_{I\in {\mathcal{I}}} t^I B_I$ for some $B_I\in \hbox{End}(V)$. Then $$\hbox{pr}_W(\Theta(t)v)=\sum_{I\in {\mathcal{I}}} t^I\hbox{pr}_W(B_Iv).$$ Consider a map $T:V\rightarrow\oplus_{I\in {\mathcal{I}}} W$ defined by $$Tv=\sum_{I\in {\mathcal{I}}} \hbox{pr}_W(B_Iv),$$ and a map $A_t:\oplus_{I\in {\mathcal{I}}} W\rightarrow W$ for $t\in\mathbb{R}^m$ defined by $$A_t\left(\bigoplus_{I\in {\mathcal{I}}} w_I\right)=\sum_{I\in {\mathcal{I}}} t^Iw_I.$$ For ${I\in {\mathcal{I}}}$, take fixed $s_I\in\mathbb{R}^m$ such that $0<s_{I,k}<1$ and $s_{I_1,k}\ne s_{I_2,k}$ for $I_1\ne I_2$, and put $t_I=\beta s_I$. Let $$A=\bigoplus_{I\in {\mathcal{I}}} A_{t_I}:\bigoplus_{I\in {\mathcal{I}}} W\rightarrow\bigoplus_{I\in {\mathcal{I}}} W.$$ The map $A$ has a matrix form $\left(t_I^J\right)_{I,J\in {\mathcal{I}}}$. This matrix is a Kronecker product of Vandermonde matrices which implies that $A$ is invertible. Using elementary row and column operations, one can write $$\label{eq_norm1} \left(t_I^J\right)_{I,J\in {\mathcal{I}}}=B\cdot\hbox{diag}\left(\beta^{|I|}:I\in {\mathcal{I}}\right)\cdot C$$ for some $B,C\in\hbox{GL}\left(\oplus_{I\in {\mathcal{I}}} W\right)$, which are independent of $\beta$. It is convenient to use a norm on $\oplus_{I\in {\mathcal{I}}} W$ defined by $$\left\|\bigoplus_{I\in {\mathcal{I}}} w_I\right\|=\max_{I\in {\mathcal{I}}} \|w_I\|,\quad w_I\in W.$$ Then by (\[eq\_norm1\]), for $w\in\oplus_{I\in {\mathcal{I}}} W$, $$\label{eq_norm2} \|Aw\|\ge \|B^{-1}\|^{-1}\cdot\beta^d\cdot\|C^{-1}\|^{-1}\cdot\|w\|.$$ It follows from Lie-Kolchin theorem that $T$ is injective (see [@sh96 Lemma 5.1]). Therefore, there exists a constant $c_1>0$ such that $\|Tv\|\ge c_1\|v\|$ for $v\in V$. Then using (\[eq\_norm2\]), we get $$\sup_{n\in D(\beta)} \|\hbox{\rm pr}_W(n\cdot v)\|=\sup_{t:0<t_i<\beta} \|A_tTv\|\ge \|ATv\|\ge c_0\beta^d\|v\|,$$ where $c_0=\|B^{-1}\|^{-1}\cdot\|C^{-1}\|^{-1}\cdot c_1$. We will need the following elementary observation: \[lem\_polynom\] Let $F:\mathbb{R}^m\rightarrow\mathbb{R}^m$ be a $C^1$ bijection such that $F(0)=0$, and $F^{-1}$ is $C^1$ too. Fix a norm on $\mathbb{R}^m$ and denote by $D(r)$ a ball of radius $r$ centered at the origin. Then there are $c_1,c_2>0$ such that $$D(c_1r)\subseteq F(D(r))\subseteq D(c_2r)$$ for every $r\in (0,1)$. Let $\mathfrak g$ be the Lie algebra of $G=\hbox{\rm SL}(n,\mathbb{R})$, and $\mathfrak{g}_\mathbb{C}=\mathfrak{g}\otimes \mathbb{C}$. Recall the root space decomposition of $\mathfrak{g}_\mathbb{C}$: $$\mathfrak{g}_\mathbb{C}=\mathfrak{g}_0\oplus \sum_{i\ne j} \mathfrak{g}_{ij},$$ where $\mathfrak{g}_0$ is the subalgebra of diagonal matrices of $\mathfrak{g}_\mathbb{C}$, and $\mathfrak{g}_{ij}=\mathbb{C}E_{ij}$ ($E_{ij}$ is the matrix with $1$ in position $(i,j)$ and $0$’s elsewhere). Introduce [*fundamental weights*]{} of $\mathfrak{g}_\mathbb{C}$: $$\label{eq_fw} \omega_i (s)=s_1+\cdots +s_i,\quad 1\le i\le n-1,$$ where $s\in \mathbb{C}^n$ and $\sum_i s_i=0$. [*Dominant weights*]{} are defined as linear combinations with non-negative integer coefficients of the fundamental weights $\omega_i$, $1\le i\le n-1$. A [*highest weight*]{} of a representation of $\mathfrak{g}_\mathbb{C}$ is a weight that is maximal with respect to the ordering on the dual space of $\mathfrak{g}_0$. Recall that irreducible representations of $\mathfrak{g}_\mathbb{C}$ are classified by their highest weights (see, for example, [@gw Ch. 5]). The highest weights are precisely the dominant weights defined above. \[lem\_diverge0\] Let $\pi$ be a representation of $G$ on a finite dimensional complex vector space $V$. Let $x\in V-\{0\}$ be such that $\pi(N)x=x$. Then $x$ is a sum of weight vectors with dominant weights. Moreover, if $V$ does not contain non-zero $G$-fixed vectors, every weight in this sum is not zero. Consider a representation $\tilde \pi$ of $\mathfrak{g}_\mathbb{C}$ on $V$ induced by the representation $\pi$. Since this representation is completely reducible, it is enough to consider the case when it is irreducible. We claim that in this case, $x$ is a vector with the highest weight. Write $x=\sum_k x_k$ where each $x_k\in V$ is a weight vector with a weight $\lambda_k$. We may assume that $\lambda_k\ne \lambda_l$ for $k\ne l$. Since $\pi(N)x=x$, $\tilde\pi(E_{ij})x=0$ for $i<j$. Thus, $\sum_k \tilde\pi(E_{ij})x_k=0$. Since $\tilde\pi(E_{ij})x_k$ is either $0$ or a weight vector with the weight $\lambda_k+\alpha_{ij}$, the non-zero terms in the sum are linearly independent. Hence, $\tilde\pi(E_{ij})x_k=0$ for $i<j$. Suppose that $\lambda_k$ is not the highest weight. Note that $\tilde\pi(\mathfrak{g}_0)x_k=\mathbb{C}x_k$, and $\tilde\pi(E_{ji})x_k$ has weight $\lambda_k-\alpha_{ij}<\lambda_k$ for $i<j$. By Poincaré-Birkhoff-Witt theorem, the universal enveloping algebra $\mathcal{U}(\mathfrak g_\mathbb{C})=\mathcal{U}(\mathfrak b^-)\oplus\mathcal{U}(\mathfrak g_\mathbb{C})\mathfrak n$, where $\mathfrak b^-$ is the space of lower triangular matrices, and $\mathfrak n$ is the Lie algebra of $N$. Therefore, the space $\tilde\pi(\mathcal{U}(\mathfrak{g}_\mathbb{C}))x_k=\tilde\pi(\mathcal{U}(\mathfrak{b}^-))x_k$ does not contain a vector with the highest weight. This contradicts irreducibility of $\tilde\pi$. Thus, each $x_k$ is of the highest weight, and $x$ is a highest weight vector. Since every highest weight is dominant, the lemma is proved. For a fixed $g_0\in G$, define $q_s(t)=g_0n(t)^{-1}a(s)^{-1}$. We are going to study $q_s$ using techniques from Section \[sec\_uni\]. The next lemma guarantees that the first possibility in Theorems \[th\_help1\] and \[th\_help3\] does not occur. For $\beta>0$, define $$\label{eq_drho} D(\beta)=\left\{n(t)\in N: \sum_{i<j} t_{ij}^2<\beta^2\right\}.$$ \[l\_diverge\] Let $\pi$ be a nontrivial representation of $G$ on a real vector space $V$ equipped with a norm $\|\cdot\|$. Let $V_0=\{v\in V: \pi(G)v=v\}$, and $V_1$ be a $G$-invariant complement for $V_0$. (Note that $V_1$ exists because $\pi$ is completely reducible.) Denote by $\Pi$ the projection on $V_1$. Then for any relatively compact set $K\subseteq V$ and $r>0$, there exist $\alpha \in (0,1)$ and $C_0>0$ such that for any $s$ with $a(s)\in A_T^{C_0}$ and $x\in V$ with $\|\Pi(x)\|>r$, $$\label{eq_div} \pi\left(q_s\left(D(e^{-\alpha s_1})\right)\right)^{-1}\cdot x\nsubseteq K.$$ It is convenient to extend $\pi$ to $V_\mathbb{C}=V\otimes \mathbb{C}$. $(V_0)_\mathbb{C}$ is the space of $G$-fixed vectors in $V_\mathbb{C}$. Thus, we may assume $V$ to be complex. Also dealing with projections on $V_1$, we may assume that $V$ has no $G$-fixed vectors. Since $\{g_0^{-1}\cdot x:\|x\|>r\}\subseteq \{x:\|x\|>r_1\}$ for some $r_1>0$, we may assume that $g_0=1$. For some $R>0$, $K\subseteq \{x\in V: \|x\|<R\}$. If (\[eq\_div\]) fails for some $s\in\mathbb{R}^n$ and $x\in V$, then $$\label{eq_g0} \sup_{n\in D(e^{-\alpha s_1})} \|\pi(a(s)n)x\|<R.$$ Let $W=\{x\in V:\pi(N)x=x\}$. Clearly, the statement of the lemma is independent of the norm used. It is convenient to use the max-norm with respect to a basis $\left\{ v_i\right\}$ of $V$ consisting of weight vectors, i.e. $$\left\|\sum_i u_i v_i\right\|=\max_i |u_i|,\quad u_i\in \mathbb{C},$$ and each $v_i$ is an eigenvector of $A$. Moreover we can choose the basis $\{v_i\}$ so that it contains a basis of $W$. Let $\hbox{pr}_W$ be the projection onto $W$ with respect to this basis. Then $\hbox{pr}_W$ commutes with $a(s)$. Note that there exists $C'>0$ such that for any $v\in V$, $$\label{eq_g1} \|v\|\ge C'\|\hbox{pr}_W(v)\|.$$ Let $\mathcal{K}\subseteq \mathbb{N}$ be such that $k\in \mathcal{K}$ iff $v_k$ has a non-zero dominant weight. Denote this weight by $\lambda_k$. By Lemma \[lem\_diverge0\], $W\subseteq \left<v_k:k\in \mathcal{K}\right>$. In particular, for $n\in N$, $$\label{eq_g2} \hbox{pr}_W(\pi(n)x)=\sum_{k\in {\mathcal{K}}} u_k(n)v_k\quad\hbox{for}\;\hbox{some}\; u_k(n)\in \mathbb{C}.$$ Therefore, $$\label{eq_g3} \|\hbox{pr}_W(\pi(n)x)\|=\max_{k\in {\mathcal{K}}} |u_k(n)|.$$ Using the fact that $\hbox{pr}_W$ and $\pi(a(s))$ commute, (\[eq\_g1\]), (\[eq\_g2\]), and (\[eq\_g3\]), we have that for any $n\in N$, $$\begin{aligned} \|\pi(a(s)n)x\|\ge C'\|\hbox{pr}_W(\pi(a(s))\pi(n)x)\|= C'\left\|\pi(a(s))\left(\sum_{k\in {\mathcal{K}}} u_k(n)v_k\right)\right\| \nonumber\\ =C'\max_{k\in {\mathcal{K}}} \left(|u_k(n)|e^{\lambda_k(s)}\right)\ge C'\exp\left(\min_{k\in {\mathcal{K}}} \lambda_k(s)\right)\|\hbox{pr}_W(\pi(n)x)\|. \label{eq_g4}\end{aligned}$$ Let $\mathfrak n$ be the Lie algebra of $N$. Denote by $\tilde\pi$ the representation of $\mathfrak g$ induced by $\pi$. Since $\mathfrak g$ is simple, $\tilde\pi$ is faithful. Thus, $\tilde\pi$ defines an isomorphism between $\mathfrak n$ and $\tilde\pi(\mathfrak n)$. Since the exponential map $\mathfrak n\to N$ is a polynomial isomorphism, the coordinates on $N$ used in Lemma \[lem\_shah\] and the coordinates $\{t_{ij}\}$ are connected by a polynomial isomorphism too. By Lemma \[lem\_polynom\], (\[eq\_shineq\]) holds for the set $D(\beta)$ defined in (\[eq\_drho\]). Therefore, $$\label{eq_g5} \sup_{n\in D(e^{-\alpha s_1})}\|\hbox{pr}_W(\pi(n)x)\|\ge c_0 (e^{-\alpha s_1})^d\|x\|\ge c_0re^{-\alpha ds_1}$$ for some positive integer $d$. It follows from (\[eq\_g4\]) and (\[eq\_g5\]) that if (\[eq\_g0\]) holds, then $$\exp\left(\min_{k\in {\mathcal{K}}} \lambda_k (s)-\alpha d s_1\right)\le \frac{R}{c_0C'r}.$$ Take $\alpha<d^{-1}$. Since each $\lambda_k$ is a non-zero dominant weight, it follows from (\[eq\_fw\]) that that $\lambda_k(s)-\alpha d s_1\rightarrow\infty$ as $C\rightarrow\infty$ for $s$ such that $a(s)\in A^C$. Hence, there exists $C_0>0$ such that (\[eq\_g0\]) does not hold for $s$ with $a(s)\in A_T^{C_0}$. Since (\[eq\_g0\]) fails, (\[eq\_div\]) holds. \[l\_diverge00\] Use notations from Lemma \[l\_diverge\]. Let $H$ be a subgroup of $G$, and $x\in V$ such that $\pi(H)x$ is discrete in $V$. Then $\Pi(\pi(H)x)$ is discrete in $V_1$. Denote by $x_0\in V_0$ and $x_1\in V_1$ the components of $x$ with respect to the decomposition $V=V_0\oplus V_1$. Then $\Pi(\pi(H)x)=\pi(H)x_1$. Suppose that for some $\{h_n\}\subseteq H$, $\pi(h_n)x_1\rightarrow y$ for some $y\in V_1$. Then $\pi(h_n)x$ converges to $x_0+y$. It follows that $\pi(h_n)x$ is constant for large $n$. Therefore, $\pi(h_n)x_1=\pi(h_n)x-x_0$ is constant for large $n$ too. Proof of Theorem \[th\_ergodic\] {#sec_th2} ================================ Let ${\mathcal{Z}}=(\Gamma\backslash G)\cup\{\infty\}$ be the $1$-point compactification of $\Gamma\backslash G$. For $T>0$, define a normalized measure on ${\mathcal{Z}}$ by $$\nu_T(\tilde{f})=\frac{1}{\varrho(B^o_T)}\int_{B^o_T} \tilde{f}(yb^{-1})d\varrho(b),\quad \tilde{f}\in C_c(\Gamma\backslash G).$$ To prove Theorem \[th\_asympt\], we need to show that $\nu_T\rightarrow \nu$ in weak-$*$ topology as $T\rightarrow\infty$. Since the space of normalized measures on ${\mathcal{Z}}$ is compact, it is enough to prove that if $\nu_{T_i}\rightarrow \eta$ as $T_i\rightarrow\infty$ for some normalized measure $\eta$ on ${\mathcal{Z}}$, then $\eta$ is $G$-invariant, and $\eta (\{\infty\})=0$. It follows from Lemma \[lem\_BTC2\] that for any $C\in\mathbb{R}$, $$\label{eq_eta} \eta(\tilde{f})=\lim_{T_i\rightarrow\infty} \frac{1}{\varrho(B^o_{T_i})}\int_{B^C_{T_i}} \tilde{f}(yb^{-1})d\varrho(b),\quad \tilde{f}\in C_c(\Gamma\backslash G).$$ Let $$\label{eq_U} U=\{n(t)\in N: t_{ij}=0\;\hbox{for}\; i<j<n\}.$$ \[lem\_eta\] The measure $\eta$ is $U$-invariant. For $\tilde{f}\in C_c(\Gamma\backslash G)$, and $g_0\in G$, define $\tilde{f}^{g_0}(\Gamma g)=\tilde{f}(\Gamma gg_0)\in C_c(\Gamma\backslash G)$. Let $\tilde{f}\in C_c(\Gamma\backslash G)$. Take $M>0$ such that $|\tilde{f}|<M$. For $T>0$ and $s\in\mathbb{R}^{n-1}$, define a set $$\label{eq_Dts} D_{s,T}=\{n\in N: \|a(s)n\|<T\}.$$ Denote by $\chi_{s,T}(n)$ the characteristic function of the set $D_{s,T}$. Then we can rewrite (\[eq\_eta\]) as $$\label{eq_eta1} \eta(\tilde{f})=\lim_{T_i\rightarrow\infty}\frac{1}{\varrho(B^o_{T_i})} \int_{A_{T_i}^C}\int_{N} \tilde{f}(yn^{-1}a(-s))\chi_{s,T_i}(n)e^{2\delta(s)}dnds.$$ Let $$\label{eq_apr} A_T^{'C}=\{a(s)\in A_T^C: T^2-N(s)> T\},$$ where $N(s)$ is defined in (\[eq\_Ns\]), and $B_T^{'C}=A_T^{'C}N\cap B_T^o$. We claim that the equality (\[eq\_eta1\]) holds when $A_{T_i}^C$ is replaced by $A_{T_i}^{'C}$. By Lemma \[lem\_BTC\], $$\begin{aligned} &&\int_{A_{T_i}^C-A_{T_i}^{'C}}\int_{N} \tilde{f}(yn^{-1}a(-s))\chi_{s,T_i}(n)e^{2\delta(s)}dnds\\ &\ll& \int_{A_{T_i}^C-A_{T_i}^{'C}}\left(T_i^2-N(s)\right)^{\frac{n(n-1)}{4}}\exp\left(\sum_k (n-k)s_k\right)ds\\ &=& O\left(T_i^{\frac{3n(n-1)}{4}}\right)\quad\hbox{as}\quad T_i\rightarrow\infty.\end{aligned}$$ Now the claim follows from (\[eq\_BTasy\]). Take $u\in U$. Let $u(s)=\hbox{Ad}_{a(-s)}(u)$. Then $$\begin{aligned} \nonumber |\eta(\tilde{f}^{u})-\eta(\tilde{f})|\le \limsup_{T_i\rightarrow\infty}\frac{1}{\varrho(B_{T_i}^o)}\int_{A_{T_i}^{'C}}\int_{N} |\tilde{f}(yn^{-1}u(s)a(-s))\chi_{s,T_i}(n)\\ - \tilde{f}(yn^{-1}a(-s))\chi_{s,T_i}(n)|dn e^{2\delta(s)}ds. \label{eq_mess1}\end{aligned}$$ We estimate the last integral: $$\begin{aligned} \nonumber \int_{N} \left|\tilde{f}(yn^{-1}u(s)a(-s))\chi_{s,T_i}(n)-\tilde{f}(yn^{-1}a(-s))\chi_{s,T_i}(n)\right|dn\\ \nonumber =\int_{N} \left|\tilde{f}(yn^{-1}a(-s))\right|\cdot \left|\chi_{s,T_i}(u(s)n)-\chi_{s,T_i}(n)\right|dn\\ \le M \int_{N} \left|\chi_{s,T_i}(u(s)n)-\chi_{s,T_i}(n)\right|dn. \label{eq_mess2}\end{aligned}$$ Recall that $\alpha_{i,n}(-s)=-s_i+s_n$. Therefore, by (\[eq\_adj\]), $u(s)_{in}= e^{-s_i+s_n} u_{in}$ for $i=1,\ldots, n-1$. It follows from the triangle inequality that $$\|a(s)u(s)n\|\le \|a(s)n\|+\sqrt{\sum_{i=1}^{n-1} e^{2s_i}u(s)_{in}^2}\le \|a(s)n\|+ e^{s_n}\|u\|,$$ and similarly, $$\|a(s)n\|\le \|a(s)u(s)n\|+ e^{s_n}\|u^{-1}\|=\|a(s)u(s)n\|+ e^{s_n}\|u\|.$$ Hence, $$\chi_{s,T_i-e^{s_n}\|u\|}(n)\le \chi_{s,T_i}(u(s)n)\le \chi_{s,T_i+e^{s_n}\|u\|}(n)$$ for $n\in N$. Therefore, $$\label{eq_mess3} \int_{N} \left|\chi_{s,T_i}(u(s)n)-\chi_{s,T_i}(n)\right|dn\le\int_{N}\left(\chi_{s,T_i+e^{s_n}\|u\|}- \chi_{s,T_i-e^{s_n}\|u\|}\right)dn.$$ Let $\varepsilon>0$. We claim that there exists $C_0>0$ such that for $C>C_0$ and $a(s)\in A^{'C}_{T_i}$, $$\label{eq_mess4} \int_{N}\left(\chi_{s,T_i+e^{s_n}\|u\|}-\chi_{s,T_i-e^{s_n}\|u\|}\right)dn \le \varepsilon \int_{N}\chi_{s,T_i}dn.$$ Similarly to Lemma \[lem\_BTC\], $$\int_{N}\chi_{s,T_i}dn=c_n\left(T_i^2-N(s)\right)^{\frac{n(n-1)}{4}}\exp\left(\sum_k (n-k)s_k\right).$$ Also $e^{s_n}\rightarrow 0$ for $a(s)\in A^{'C}_{T_i}$ as $C\rightarrow\infty$. Therefore, the equation (\[eq\_mess4\]) will follow from the following. [*Claim*]{}. There exists $d_0=d_0(\varepsilon)>0$ such that for any $d\in (0,d_0)$ and $a(s)\in A^{'C}_{T_i}$, $$\label{eq_LLL} \Big((T_i+d)^2-N(s)\Big)^{\frac{n(n-1)}{4}}-\Big((T_i-d)^2-N(s)\Big)^{\frac{n(n-1)}{4}} <\varepsilon \Big((T_i-d)^2-N(s)\Big)^{\frac{n(n-1)}{4}}.$$ Note that by (\[eq\_apr\]), $(T_i-d)^2-N(s)>0$ for $a(s)\in A^{'C}_{T_i}$ and $d<1/2$. The inequality (\[eq\_LLL\]) is equivalent to $$(T_i+d)^2-N(s)<(1+\bar\varepsilon)\Big((T_i-d)^2-N(s)\Big),$$ where $\bar\varepsilon=(1+\varepsilon)^\frac{4}{n(n-1)}-1$. By (\[eq\_apr\]), the last inequality follows from $$(T_i+d)^2 +\bar\varepsilon (T_i^2-T_i)<(1+\bar\varepsilon)(T_i-d)^2,$$ or equivalently, $$T_i(4d-\bar\varepsilon+2d\bar\varepsilon)<\bar\varepsilon d^2.$$ If one takes $d<d_0=\bar\varepsilon/(2\bar\varepsilon+4)$, the left hand side is negative. This proves the claim (\[eq\_LLL\]). Thus, (\[eq\_mess4\]) holds. Then by (\[eq\_mess1\]), (\[eq\_mess2\]), (\[eq\_mess3\]), and (\[eq\_mess4\]), $$\begin{aligned} |\eta(\tilde{f}^{u})-\eta(\tilde{f})|&\le& (M\varepsilon)\limsup_{T_i\rightarrow\infty}\frac{1}{\varrho(B_{T_i}^o)}\int_{A_{T_i}^{'C}}\int_{N} \chi_{s,T_i}(n)e^{2\delta(s)} dnds\\ &=&(M\varepsilon)\limsup_{T_i\rightarrow\infty}\frac{\varrho(B_{T_i}^{'C})}{\varrho(B_{T_i}^o)}\le M\varepsilon.\end{aligned}$$ This shows that $\eta(\tilde{f}^{u})=\eta(\tilde{f})$. \[lem\_tildeA\] Let $\alpha \in (0,1)$. Let $$\label{eq_tildeA} \tilde{A}^C_T=\left\{a(s)\in A_T^C|\; (T^2-N(s))^{1/2}>\exp\left(\mathop{\max}_{1\le j\le n-1}\{s_j\}-\alpha s_1\right)\right\},$$ where $N(s)$ is as in (\[eq\_Ns\]), and $\tilde{B}^C_T=\tilde{A}^C_TN\cap B_T^o$. Then for $C>0$, $$\eta(\tilde{f})=\lim_{T_i\rightarrow\infty} \frac{1}{\varrho(B^o_{T_i})}\int_{\tilde{B}^C_{T_i}} \tilde{f}(yb^{-1})d\varrho(b),\quad \tilde{f}\in C_c(\Gamma\backslash G).$$ By (\[eq\_eta\]), it is enough to show that $$\label{eq_tBTC} \frac{\varrho\left(B^C_{T_i}-\tilde{B}^C_{T_i}\right)}{\varrho\left(B^o_{T_i}\right)}\rightarrow 0\quad\hbox{as}\quad T_i\rightarrow \infty.$$ As in Lemma \[lem\_BTC\], $$\varrho\left(B_{T_i}^C-\tilde{B}^C_{T_i}\right)=c_n\mathop{\int}_{A_{T_i}^C-\tilde{A}^C_{T_i}} \Big(T_i^2-N(s)\Big)^{\frac{n(n-1)}{4}}\hbox{\rm exp}\left(\sum_{k} (n-k)s_k\right)ds.$$ Therefore, $$\begin{aligned} \varrho\left(B_{T_i}^C-\tilde{B}^C_{T_i}\right)\le c_n \int_{A_{T_i}^C}\exp\Big(\frac{n(n-1)}{2}\mathop{\max}_{1\le j\le n-1}\{s_j\}\\ -\frac{\alpha n(n-1)}{2} s_1+\sum_{k} (n-k)s_k\Big)ds \le c_n \mathop{\sum}_{1\le j\le n-1} \int_{A_{T_i}^C}\exp\Big(\frac{n(n-1)}{2}s_j\\ -\frac{\alpha n(n-1)}{2} s_1+\sum_{k} (n-k)s_k\Big)ds\end{aligned}$$ Then as in the proof of Lemma \[lem\_BTC2\], for $j\ne 1$ $$\begin{aligned} \int_{A_{T_i}^C}\exp\left(\frac{n(n-1)}{2}s_j-\frac{\alpha n(n-1)}{2} s_1+\sum_{k} (n-k)s_k\right)ds\\ \le \int_{-\infty}^{\log T_i} \exp\left(\left(-\frac{\alpha n(n-1)}{2}+n-1\right)s_1\right)ds_1\\ \cdot\int_{-\infty}^{\log T_i} \exp\left(\left(\frac{n(n-1)}{2}+n-j\right)s_j\right)ds_j\cdot \prod_{{k<n},{k\ne 1,j}} \int_{-\infty}^{\log T_i} e^{(n-k)s_k}ds_k\\ \ll T_i^{n(n-1)-\alpha n(n-1)/2}\end{aligned}$$ as $T_i\rightarrow\infty$. For $j=1$, the same estimate can be obtained by a similar calculation. Now (\[eq\_tBTC\]) follows from (\[eq\_BTasy\]). Let $y=\Gamma g_0$ for $g_0\in G$. Define $q_s(t)=g_0n(t)^{-1}a(s)^{-1}$. We apply the results of Section \[sec\_uni\] to the map $q_s$. \[lem\_inf1\] $\eta(\{\infty\})=0$. Let $\varepsilon,\delta >0$. Apply Theorem \[th\_help1\] to the map $q_s(t)$. By Theorem \[th\_help2\], the set $\Gamma\cdot p_{U_i}\subseteq V_G$ is discrete. Write $V_G=V_0\oplus V_1$, where $V_0$ is the space of vectors fixed by $G$, and $V_1$ is its $G$-invariant complement. Denote by $\Pi$ the projection on $V_1$. By Lemma \[l\_diverge00\], $\Pi(\Gamma\cdot p_{U_i})$ is discrete. Also $0\notin \Pi(\Gamma\cdot p_{U_i})$. Otherwise $p_{U_i}$ is fixed by $G$, and it would follow that $U_i$ is normal in $G$ which is a contradiction. Therefore, there exists $r>0$ such that $$\|\Pi(x)\|>r\quad\textrm{for}\quad x\in\bigcup_i \Gamma\cdot p_{U_i}.$$ Now we can apply Lemma \[l\_diverge\]. Let $$K=\{x\in V_G: \|x\|\le \delta\}.$$ By Lemma \[l\_diverge\], there exist $\alpha \in (0,1)$ and $C_0>0$ such that the first case of Theorem \[th\_help1\] fails for $q_s$ when $D$ is a bounded open convex set which contains $D(e^{-\alpha s_1})$ (it is defined in (\[eq\_drho\])), and $s$ is such that $a(s)\in A^{C_0}_{T_i}$. Therefore, for some compact set $C\subseteq \Gamma\backslash G$, $$\label{eq_mC} \omega \left(\left\{n(t)\in D: \Gamma q_s(t)\notin C\right\}\right)< \varepsilon \omega (D)$$ when $D\supseteq D(e^{-\alpha s_1})$ and $a(s)\in A^{C_0}_{T_i}$, where $\omega=dt$ denotes the Lebesgue measure on $N$. Let $D_{s,T_i}$ be as in (\[eq\_Dts\]). By Lemma \[lem\_tildeA\], $$\label{eq_tA} \eta(\tilde{f})=\lim_{T_i\rightarrow\infty} \frac{1}{\varrho(B^o_{T_i})}\int_{\tilde{A}^C_{T_i}}\int_{D_{s,T_i}}\tilde{f}(\Gamma q_s(t))e^{2\delta(s)}dtds,\quad \tilde{f}\in C_c(\Gamma\backslash G).$$ Note that $$D_{s,T_i}=\left\{n(t)\in N: \sum_{i<j} e^{2s_i}t_{ij}^2<T_i^2-N(s)\right\},$$ where $N(s)$ is defined in (\[eq\_Ns\]) (cf. (\[eq\_BT\])). It follows that $D_{s,T_i}$ contains $D(\beta)$ for $$\beta<(T_i^2-N(s))^{1/2}\exp\left(-\mathop{\max}_{1\le i\le n-1} \{s_i\}\right).$$ When $a(s)\in\tilde{A}^{C_0}_{T_i}$, the right hand side is greater then $e^{-\alpha s_1}$ (see (\[eq\_tildeA\])). Therefore, $D_{s,T_i}\supseteq D(e^{-\alpha s_1})$ when $a(s)\in\tilde{A}^{C_0}_{T_i}$. By (\[eq\_mC\]), $$\label{eq_mC1} \omega \left(\left\{n(t)\in D_{s,T_i}: \Gamma q_s(t)\notin C\right\}\right)< \varepsilon \omega (D_{s,T_i})\quad \hbox{for}\;\; a(s)\in\tilde{A}^{C_0}_{T_i}.$$ Let $\chi_C$ be the characteristic function of the set $C$. Take $\tilde{f}\in C_c(\Gamma\backslash G)$ such that $\chi_C\le \tilde{f}\le 1$. Then using (\[eq\_tA\]) and (\[eq\_mC1\]), we get $$\begin{aligned} \eta(\hbox{supp}(\tilde{f}))&\ge& \lim_{T_i\rightarrow\infty} \frac{1}{\varrho(B^o_{T_i})}\int_{\tilde{A}^{C_0}_{T_i}}\int_{D_{s,T_i}}\chi_C(\Gamma q_s(t))e^{2\delta(s)}dtds\\ &\ge&\lim_{T_i\rightarrow\infty} \frac{1}{\varrho(B^o_{T_i})}\int_{\tilde{A}^{C_0}_{T_i}}(1-\varepsilon)\omega(D_{s,T_i})e^{2\delta(s)}ds\\ &=&(1-\varepsilon)\lim_{T_i\rightarrow\infty} \frac{\varrho(\tilde{B}^{C_0}_{T_i})}{\varrho(B^o_{T_i})}=1-\varepsilon.\end{aligned}$$ Hence, $\eta(\{\infty\})\le\eta(\hbox{supp}(\tilde{f})^c)\le\varepsilon$ for every $\varepsilon>0$. Recall that the singular set $Y$ of $U$ was defined in (\[eq\_sing\]). \[lem\_Y1\] $\eta(Y)=0$. By (\[eq\_sing\]) and Theorem \[th\_help2\], it is enough to show that $\eta(\Gamma X(H,U))=0$ for any $H\in {\mathcal{H}}_\Gamma$. Moreover, since $\Gamma X(H,U)$ is $\sigma$-compact, we just need to show that $\eta(C)=0$ for any compact set $C\subseteq \Gamma X(H,U)$. Take $\varepsilon>0$. Apply Theorem \[th\_help3\] to the map $q_s(t)$. Let $F\subseteq V_G$ be as in Theorem \[th\_help3\]. Fix a relatively compact neighborhood $\Phi$ of $F$ in $V_G$. Take $\Psi\supseteq C$ as in Theorem \[th\_help3\]. By Theorem \[th\_help2\], the set $\Gamma\cdot p_H$ is discrete. Let $\Pi$ be as in the proof of Lemma \[lem\_inf1\]. By Lemma \[l\_diverge00\], $\Pi(\Gamma\cdot p_H)$ is discrete. If $0\in\Pi(\Gamma\cdot p_H)$, the vector $p_H$ is fixed by $G$, and $H$ is normal in $G$, which is impossible. Therefore, for some $r>0$, $\|\Pi(x)\|>r$ for every $x\in \Gamma\cdot p_H$. Applying Lemma \[l\_diverge\] with $K=\Phi$, one gets that there exist $\alpha\in (0,1)$ and $C_0>0$ such that the first case of Theorem \[th\_help1\] fails for $q_s$ when $D$ is a bounded open convex set containing $D(e^{-\alpha s_1})$, and $s$ is such that $a(s)\in A^{C_0}_{T_i}$. Therefore, the second case should hold: $$\label{eq_mC3} \omega \left(\left\{n(t)\in D: \Gamma q_s(t)\in \Psi\right\}\right)< \varepsilon \omega (D)$$ when $D\supseteq D(e^{-\alpha s_1})$ and $a(s)\in A^{C_0}_{T_i}$. Let $D_{s,T_i}$ be as in (\[eq\_Dts\]). It was shown in the proof of Lemma \[lem\_inf1\] that $D_{s,T_i}\supseteq D(e^{-\alpha s_1})$ when $a(s)\in\tilde{A}^{C_0}_{T_i}$. It follows from (\[eq\_mC3\]) that $$\label{eq_mC4} \omega \left(\left\{n(t)\in D_{s,T_i}: \Gamma q_s(t)\in \Psi\right\}\right)< \varepsilon \omega (D_{s,T_i})\quad \hbox{for}\;\; a(s)\in\tilde{A}^{C_0}_{T_i}.$$ Take a function $\tilde{f}\in C_c(\Gamma\backslash G)$ such that $\tilde{f}=1$ on $C$, $\hbox{supp}(\tilde{f})\subseteq \Psi$, and $0\le \tilde{f}\le 1$. Let $\chi_\Psi$ be the characteristic function of $\Psi$. Then using (\[eq\_tA\]) and (\[eq\_mC4\]), we get $$\begin{aligned} \eta(C)\le \lim_{T_i\rightarrow\infty} \frac{1}{\varrho(B^o_{T_i})}\int_{\tilde{A}^{C_0}_{T_i}}\int_{D_{s,T_i}}\chi_\Psi(\Gamma q_s(t))e^{2\delta(s)}dtds\\ \le \lim_{T_i\rightarrow\infty} \frac{1}{\varrho(B^o_{T_i})}\int_{\tilde{A}^{C_0}_{T_i}}\varepsilon\omega(D_{s,T_i})e^{2\delta(s)}ds =\varepsilon\lim_{T_i\rightarrow\infty} \frac{\varrho(\tilde{B}^{C_0}_{T_i})}{\varrho(B^o_{T_i})}=\varepsilon.\end{aligned}$$ This shows that $\eta(C)=0$. Hence, $\eta (Y)=0$. Recall that by Lemma \[lem\_eta\], $\eta$ is $U$-invariant. By Ratner’s measure classification [@r1], an ergodic component of $\eta$ is either $G$-invariant or supported on $Y\cup\{\infty\}$. By Lemmas \[lem\_inf1\] and \[lem\_Y1\], the set of ergodic components of the second type has measure $0$. Therefore, $\eta$ is $G$-invariant, and $\eta=\nu$. This finishes the proof of Theorem \[th\_ergodic\]. Acknowledgments {#acknowledgments .unnumbered} ================= The author wishes to thank his advisor V. Bergelson for constant support, encouragement, and help with this paper, N. Shah and Barak Weiss for insightful comments about the paper, and H. Furstenberg and G. Margulis for interesting discussions. [30]{} A. Borel and J. Tits. Groupes réductifs. [*Inst. Hautes Études Sci. Publ. Math.*]{} [**27**]{} (1965), 55–150. S. G. Dani and G. A. Margulis. Asymptotic behavior of trajectories of unipotent flows on homogeneous spaces. [*Proc. Indian Acad. Sci. Math. Sci.*]{} [**101**]{} (1991), 1–17. S. G. Dani and G. A. Margulis. Limit distributions of orbits of unipotent flows and values of quadratic forms. I. M. Gelfand Seminar, 91–137, [*Adv. Soviet Math.*]{} [**16**]{}, Part 1, AMS, Providence, RI, 1993. W. Duke, Z. Rudnick and P. Sarnak. Density of integer points on affine homogeneous varieties. [*Duke Math. J.*]{} [**71**]{} (1993), 143–179. A. Eskin and C. McMullen. Mixing, counting, and equidistribution in Lie groups. [*Duke Math. J.*]{} [**71**]{} (1993), 181–209. A. Eskin, G. Margulis and S. Mozes. Upper bounds and asymptotics in a quantitative version of the Oppenheim conjecture. [*Ann. Math.*]{} [**147**]{} (1998), 93–141. A. Eskin, S. Mozes and N. Shah. Nimish Unipotent flows and counting lattice points on homogeneous varieties. [*Ann. Math.*]{} [**143**]{} (1996), 253–299. A. Eskin, S. Mozes and N. Shah. Non-divergence of translates of certain algebraic measures.[*Geom. Funct. Anal.*]{} [**7**]{} (1997), 48–80. H. Furstenberg. A Poisson formula for semi-simple Lie groups. [*Ann. Math.*]{} [**77**]{} (1963), 335–386. A. Good, [*Local analysis of Selberg’s trace formula*]{}, Lecture Notes in Mathematics, 1040, Springer-Verlag, Berlin, 1983. R. Goodman and N. R. Wallach. [*Representations and invariants of the classical groups*]{}. Cambridge University Press, Cambridge, 1998. D. Kleinbock, N. Shah and A. Starkov. Dynamics of subgroup actions on homogeneous spaces of Lie groups and applications to number theory. in [*Handbook of dynamical systems*]{}, Vol. 1A, 813–930, North-Holland, Amsterdam, 2002. F. Ledrappier, Distribution des orbites des réseaux sur le plan réel. [*C. R. Acad. Sci. Paris Sér. I Math.*]{} [**329**]{} (1999), 61–64. A. Nogueira, Orbit distribution on $\mathbb{R}^2$ under the natural action of $\hbox{\rm SL}(2,\mathbb{Z})$, preprint, 2000. V. Platonov and A. Rapinchuk. [*Algebraic groups and number theory*]{}. Academic Press, Inc., Boston, MA, 1994. M. S. Raghunathan. [*Discrete Subgroups of Lie Groups*]{}. Springer-Verlag, New York, 1972. M. Ratner. On Raghunathan’s measure conjecture. [*Ann. Math.*]{} 134 (1991), 545–607. M. Ratner. Raghunathan’s topological conjecture and distributions of unipotent flows. [*Duke Math. J.*]{} [**63**]{} (1991), 235–280. N. A. Shah. Limit distributions of polynomial trajectories on homogeneous spaces. [*Duke Math. J.*]{} [**75**]{} (1994), 711–732. N. Shah. Limit distributions of expanding translates of certain orbits on homogeneous spaces. [*Proc. Indian Acad. Sci. Math. Sci.*]{} [**106**]{} (1996), 105–125. T. A. Springer. [*Linear algebraic groups*]{}. Birkhäuser, Boston, 1998. A. N. Starkov. [*Dynamical systems on homogeneous spaces*]{}. Translations of Mathematical Monographs 190, AMS, Providence, RI, 2000. A. Terras. [*Harmonic analysis on symmetric spaces and applications I*]{}. Springer-Verlag, New York, 1985. [^1]: This article is a part of author’s PhD thesis at Ohio State University done under supervision of Prof. Bergelson.

--- abstract: 'The main goal of this paper is to define and study new methods for the computation of effective coefficients in the homogenization of divergence-form operators with random coefficients. The methods introduced here are proved to have optimal computational complexity, and are shown numerically to display small constant prefactors. In the spirit of multiscale methods, the main idea is to rely on a progressive coarsening of the problem, which we implement via a generalization of the Green-Kubo formula. The technique can be applied more generally to compute the effective diffusivity of any additive functional of a Markov process. In this broader context, we also discuss the alternative possibility of using Monte-Carlo sampling, and show how a simple one-step extrapolation can considerably improve the performance of this alternative method.' address: 'Ecole normale supérieure de Lyon, CNRS, Lyon, France' author: - 'Jean-Christophe Mourrat' bibliography: - 'efficient.bib' title: Efficient methods for the estimation of homogenized coefficients --- Introduction ------------ Let $-\nabla \cdot {\mathbf{a}}(x) \nabla$ be a divergence-form operator with random coefficients. If the law of the matrix field $({\mathbf{a}}(x))$ is stationary and ergodic, then the operator homogenizes over large scales to an operator with constant, deterministic coefficients $-\nabla \cdot {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}\nabla$. The main goal of this paper is to explore numerical methods to estimate the homogenized matrix ${{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}$. We can formalize this task as follows. **Problem.** *Find an algorithm that, given ${\delta}> 0$ and a realization of the random coefficient field $({\mathbf{a}}(x))$, outputs a matrix ${\widehat}{\mathbf{a}}_{\delta}$ such that* $$\label{e.prob.algo} {\mathbb{E}}{\left}[|{{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}- {\widehat}{\mathbf{a}}_{\delta}|^2 {\right}]^\frac 1 2 {\leqslant}{\delta}.$$ In , we chose to measure the quality of the approximation ${\widehat}{\mathbf{a}}_{\delta}$ in an $L^2$ sense for definiteness, but other choices (e.g. convergence in probability, or higher moments) are equally valid, and in fact would only marginally alter the results presented below. The point is that ${{{\widehat}{\mathbf{a}}_{\delta}}}$ should typically be within ${\delta}$ or less from ${{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}$. Naturally, the real question is to find an *efficient* algorithm. We focus on a discrete-space setting, and assume throughout that the coefficients are uniformly elliptic, independent and identically distributed (i.i.d.). In order to measure computational effort, we use a slightly loosely defined notion of “elementary” operation: any memory access, boolean operation, floating point addition or multiplication counts as one operation. We first give a lower bound on the number of operations any algorithm must use. \[p.lower.bound\] No algorithm can output an approximation ${\widehat}{\mathbf{a}}_{\delta}$ of ${{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}$ in the sense of within $o({\delta}^{-2})$ operations. The main method we propose to compute an approximation of ${{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}$ is as follows. We fix $\xi \in {{\mathbb{R}^d}}$ of unit norm, set $$v_{-1}(x) := (\nabla \cdot {\mathbf{a}}\xi)(x) \qquad (x \in {\mathbb{Z}}^d),$$ and for each $k \in {\mathbb{N}}$, we let $v_k$ be the solution to $$\label{e.def.vk} (2^{-k} - \nabla \cdot {\mathbf{a}}\nabla)v_k = 2^{-k} v_{k-1} \qquad \text{on } {{\mathbb{Z}^d}}.$$ (We refer to Section \[s.defs\] for precise notation and assumptions.) Note that the functions $(v_k)_{k {\geqslant}-1}$ depend on the coefficient field ${\mathbf{a}}$, although we keep this dependence implicit. For each $r {\geqslant}0$, we write $$B_r := \{-\lfloor r \rfloor, \ldots, \lfloor r \rfloor\}^d.$$ \[t.main\] For each ${\varepsilon}\in (0, \frac{d-1}{2d})$, there exists a constant $C < \infty$ such that the following holds. Fix $n \in {\mathbb{N}}$, and denote $$\label{e.def.main.rk} r_k := 2^{n- {\left}( \frac 1 2 - {\varepsilon}{\right}) k} \qquad (k \in \{0,\ldots,n\}),$$ $$\label{e.def.final.hatsigma} {\widehat}\sigma^2_n := \sum_{k = 0}^n \frac{1}{|B_{r_k}|} \sum_{x \in B_{r_k}} 2^k {\left}( v_{k-1}(x) v_k(x) + v_k^2(x) {\right}).$$ We have $$\label{e.main} {\mathbb{E}}{\left}[ {\left}( \xi \cdot {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}\xi - {\mathbb{E}}[\xi \cdot {\mathbf{a}}\xi] + {\widehat}\sigma^2_n {\right}) ^2 {\right}] ^\frac 1 2 {\leqslant}C 2^{-\frac {dn}{2}}.$$ Moreover, the quantity ${\widehat}\sigma^2_n$ can be computed in at most $C n 2^{dn}$ operations. By the central limit theorem, the quantity ${\mathbb{E}}[\xi \cdot {\mathbf{a}}\xi]$ can be computed at precision $\delta > 0$ in $C {\delta}^{-2}$ operations. Hence, Theorem \[t.main\] gives us a method that requires ${\delta}^{-2} \log({\delta}^{-1})$ operations to compute an approximation of ${{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}$ at precision ${\delta}> 0$ in the sense of , essentially matching the lower bound given by Proposition \[p.lower.bound\]. For ${\varepsilon}= 0$, the statement of Theorem \[t.main\] remains valid provided that we replace the right side of by $C n 2^{-\frac{dn}{2}}$. I expect that the sharp upper bound in this case is $C n^\frac 1 2 2^{-\frac{dn}{2}}$. Whatever may be, this additional multiplicative factor of $n$ or $n^\frac 1 2$ does not degrade the bound much. For simplicity, it is the version with ${\varepsilon}= 0$ that was implemented numerically. The proof of Theorem \[t.main\] relies in particular on the assumption that the law of the coefficient field ${\mathbb{P}}$ is invariant under the action of ${\mathbb{Z}}^d$. If this assumption is weakened and the law is only assumed to be invariant under the action of a sublattice of ${\mathbb{Z}}^d$ which may be unknown to us a priori, then one can use masks in the spatial averages over $B_{r_k}$ appearing in in order to avoid boundary layer effects. When weakening the independence assumption on the coefficients, I expect Theorem \[t.main\] to remain unchanged provided that the field $x \mapsto {\mathbf{a}}(x)$ is sufficiently uncorrelated for partial sums over boxes to satisfy a central limit theorem. The power of the method presented in Theorem \[t.main\] comes from the fact that the domain of interest shrinks rapidly as $k$ increases. Moreover, when $k$ is of order one, the condition number of the elliptic problems we need to solve and the size of the associated boundary layers are also of order one. As $k$ increases, condition numbers and boundary layers become larger, but these adverse effects are more than compensated by the reduction of size of the domain of interest. In fact, in order to prove the complexity upper bound in Theorem \[t.main\], we do not need to use any preconditioner for the resolution of the elliptic problems: a direct application of the conjugate gradient method suffices. Moreover, the overhead caused by the presence of boundary layers only participates to a negligible fraction of the total computational time: this fraction of time is of the order of $2^{-\frac n 2}$ in dimension $d = 2$, and of the order of $2^{- n}$ in dimension $d {\geqslant}3$. It is straightforward to adapt the method of Theorem \[t.main\] to accommodate for parallel computing. Indeed, we will see that the correlation length of the random field $$\label{e.mixing.field} x \mapsto v_{k-1}(x) v_k(x) + v_k^2(x)$$ is of the order of $2^{\frac k 2}$. Computing spatial averages of this random field is the best strategy if all computations are done sequentially. On the other hand, if parallel computing is available, then we may proceed differently and compute separately $v_k^{(1)}, v_k^{(2)}, \ldots$, based on independent copies of the coefficient field but otherwise defined in the same way as $v_k$, and then average the results. The general rule is that we may replace any spatial average $$\frac 1 {|B_r|} \sum_{x \in B_r} {\left}( v_{k-1}(x) v_k(x) + v_k^2(x) {\right})$$ by $$\frac{1}{\lceil 2^{-\frac {kd} 2} r^d \rceil} \sum_{i = 1}^{\lceil 2^{-\frac {kd} 2} r^d \rceil} {\left}( v_{k-1}^{(i)}(0) v_k^{(i)}(0) + {\left}(v_k^{(i)}(0){\right})^2 {\right}) ,$$ or with any other intermediate combination of spatial averaging and independent sampling. Disregarding logarithmic factors, the previously known best procedure for computing ${{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}$ at precision ${\delta}> 0$ consists in averaging ${\delta}^{-1}$ samples of the energy of an approximate corrector calculated on a box of side length ${\delta}^{-\frac 1 d}$—see [@Gloria] and [@cemracs Section 3.3.4] for details. If no preconditioner is used, then the computational cost of this method is of the order of ${\delta}^{-{\left}(2 + \frac 1 d{\right})}$, up to logarithmic factors. While efficient preconditioners can bring this estimate back to being of the order of ${\delta}^{-2}$, up to logarithmic factors, the prefactors involved may be very large in practice. I did some modest tests with an incomplete Cholesky factorization as a preconditioner. While this did reduce the number of iterations involved in the conjugate gradient method, it was not sufficient to compensate for the induced overhead and resulted in degraded performance. Other preconditioners will certainly show better performance (using for instance the homogeneous Laplacian, which can be solved in almost linear time by fast Fourier transform), and I did not investigate this question further. In any case, the computational overhead caused by boundary layers for this older method in fact takes up an asymptotically full proportion of the computational time, since boundary layers occupy a volume of the order of ${\delta}^{-1} \log^d({\delta}^{-1})$ on each of the ${\delta}^{-1}$ domains. Naturally, this observation calls for an adjustment of the strategy: one should solve fewer problems, each of larger volume. But this runs counter to the fact that it is computationally best to solve smaller elliptic problems, and only suboptimal compromises can be found in this way. An additional advantage of the method described in Theorem \[t.main\] is that it is cumulative: all computations involved in the evaluation of ${\widehat}{\sigma}^2_n$ remain useful if the approximation needs to be refined. In contrast, in the older method, computations done for a coarse precision must be thrown away if greater precision is then desired. The theoretical arguments presented here are supplemented by numerical tests. The numerical results confirm the theoretical predictions, and display small constant prefactors. As was said above, we assume that the coefficient field $({\mathbf{a}}(x))$ is uniformly elliptic and i.i.d., and we describe algorithms that perform well under these assumptions. In practice, it seems more appropriate to implement more “agnostic” algorithms, that would evaluate on the fly the behavior of the mean and variance of the spatial averages appearing in as $r$ and $k$ vary, and adapt the numerical scheme dynamically. We will not discuss this possibility any further here. It would be interesting to investigate how the new method presented here adapts to the continuous-space setting. In this case, I expect that the error analysis remains unchanged. The complexity analysis would however also have to take into account the cost of the small-scale resolution of the equations, below the typical length scale of the correlations of the coefficient field. This problem is separate from the homogenization phenomenon, and is likely to come as the same multiplicative factor in the complexity analysis of any method. Another interesting line of further research concerns the behavior of the effective conductivity of percolation clusters. We refer to [@hughes; @kum-stflour] for very nice surveys on this problem. The method presented here should allow for a sharper numerical analysis of the effective conductivity as the percolation probability approaches criticality. Despite immense progress on the understanding of critical and near-critical percolation in two dimensions, it is still not known whether the effective conductivity behaves as a power law near criticality (let alone compute the exponent) in small dimensions. See however [@kesten-subdiff; @jarnac; @damronetal], as well as [@barandco; @koznac; @BACF1; @BACF2] for very fine results in high dimensions. It turns out that the underlying idea of the method can be recast in the more general framework of additive functionals of Markov processes. The paper is thus split into two parts. In the first part, we describe the general strategy in this broader context. We also discuss alternative methods based on Monte-Carlo sampling, and show that a simple one-step extrapolation can in general improve the efficiency of this approach dramatically. The second part of the paper is concerned with the computation of ${{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}$ per se. In particular, we prove Theorem \[t.main\] and report on numerical results there. Readers who are only interested in the problem of computing ${{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}$ are encouraged to go directly to the second part of this paper, where we explain how to extract relevant information from the first part without having to refer to Markov processes. We conclude this introduction with a brief review of related work. The core idea of the method presented in Theorem \[t.main\] is to decompose the problem into a series of scales. The approach is inspired by the renormalization scheme introduced in [@AKM2], which progressively coarsens the operator and relies on a “quadratic response” behavior around the homogenization limit—see also the discussion below . Other relevant recent references for quantitative homogenization include [@vardecay; @GO1; @GO2; @GNO; @MO; @AS; @GNO3; @AM; @AKM; @GO5; @AKMBook]. More generally, the approach presented here is similar in spirit to multiscale methods [@brandt; @eq-free; @hmm]. In this context, it was quickly realized that the resolution of cell problems or the computation of adapted finite-element base functions can display so-called “resonant errors” due to inappropriate boundary condition [@hw; @hwc; @yue-e; @eh-book]. A powerful approach was introduced in [@blb; @GO2; @Gloria; @GloriaH; @vardecay; @approx], based on the introduction of a massive term that smoothly makes the faraway boundary condition irrelevant. The methods presented here are faithful to this approach; see also [@malpet; @henpet; @time-dep]. A different line of research focused on developing more efficient methods for computing homogenized coefficients of small random perturbations of a simpler, e.g. homogeneous, medium, in the spirit of earlier fundamental insights [@maxwell; @rayleigh]. A typical example is that of spherical inclusions randomly placed at a small volume fraction $p >0$ in an otherwise constant medium. One looks for asymptotic formulas of the homogenized coefficients involving easily computable coefficients, in the regime $p \ll 1$. We refer to [@kozlov; @papa-survey; @bm1; @alb1; @alb2; @almog1; @almog2; @dl-diff; @DG] for a mathematical analysis of the problem. Note that the seemingly simpler task of determining the value of the volume fraction $p$ at precision ${\delta}$ from a sample of the coefficient field already requires of the order of ${\delta}^{-2}$ operations (the proof of the lower bound is a minor variant of the proof of Proposition \[p.lower.bound\]). Lastly, several techniques have been explored to reduce the size of the fluctuations of quantities of interest, typically by a constant multiplicative factor. We refer to [@lebris-survey] for a recent survey. General methods {#part.one} =============== Main results ------------ In this first part of the paper, we explore the problem of computing the effective diffusivity of general additive functionals of a Markov process. Let $(X(t))_{t {\geqslant}0}$ be a Feller process taking values in a metric space ${\mathcal}X$, and denote the associated family of probability laws by $({\mathbf{P}}_x)_{x \in {\mathcal}X}$, with corresponding expectations $({\mathbf{E}}_x)_{x \in {\mathcal}X}$. We assume that there exists a probability measure ${\mathbb{P}}$ on ${\mathcal}X$ (with corresponding expectation ${\mathbb{E}}$) such that the process $(X_t)$ is reversible and ergodic with respect to ${\mathbb{P}}$; we denote the law of $(X(t))$ started from the measure ${\mathbb{P}}$ by ${\mathbf{P}}_{\mathbb{P}}$ (with corresponding expectation ${\mathbf{E}}_{\mathbb{P}}$). The assumption of reversibility will be dropped shortly, and is only meant to facilitate the exposition here. Given $f \in L^2({\mathbb{P}})$ of mean zero, one can check using reversibility that the limit $$\label{e.def.H-1} \sigma^2(f) := \lim_{t \to + \infty} {\mathbf{E}}_{\mathbb{P}}{\left}[ {\left}( \frac 1 {\sqrt{t}}\int_0^t f(X(s)) \ {\mathrm{d}}s {\right})^2 {\right}] \in [0,+\infty]$$ always exists. Denote by ${{\mathcal L}}$ the (non-negative) infinitesimal generator of $(X(t))$, and $\langle f, g \rangle := {\mathbb{E}}[fg]$. An equivalent characterization of $\sigma^2(f)$ is given by $$\label{e.Poisson} \sigma^2(f) = 2 {\left\langle}f, {{\mathcal L}}^{-1} f {\right\rangle}:= 2 \lim_{\lambda \to 0} {\left\langle}f, (\lambda + {{\mathcal L}})^{-1} f {\right\rangle}.$$ Kipnis and Varadhan [@kipvar] showed that the condition $\sigma^2(f) < \infty$ implies that the additive functional $$\label{e.additive.func} \frac 1 {\sqrt{t}} \int_{0}^t f(X(s)) \, {\mathrm{d}}s$$ converges in law under ${\mathbf{P}}_{\mathbb{P}}$ to a centered Gaussian random variable of variance $\sigma^2(f)$ as $t$ tends to infinity. Our goal is to discuss practical ways to compute $\sigma^2(f)$. There are two standard methods to address this question (see for instance [@cemracs]). The first one uses the formula as the basis for a Markov chain Monte-Carlo (MCMC) algorithm. The second method dispenses with the simulation of the Markov chain, and uses the formula instead. The problem then becomes that of computing an approximation of ${{\mathcal L}}^{-1} f$, and computing the average. We study MCMC methods in Section \[s.mcmc\]. The main contribution of this section is to show that a one-step extrapolation of the naive MCMC algorithm yields significant improvement of the approximation in most cases. We then explore methods based on the formula . Similarly to multiscale methods, we aim to split the computational effort involved in the computation of ${{\mathcal L}}^{-1} f$ into a hierarchy of scales. A good starting point for witnessing this multiscale decomposition is the Green-Kubo formula, which in our present context is a direct consequence of and reads $$\label{e.gk1} \sigma^2(f) = 2 \int_0^{+\infty} {\left\langle}f, e^{-t{{\mathcal L}}} f {\right\rangle}\, {\mathrm{d}}t = 2 \int_0^{+\infty} {\mathbf{E}}_{\mathbb{P}}{\left}[ f(X(0)) f(X(t)) {\right}] \, {\mathrm{d}}t.$$ We think of the part of the integral with $t \lesssim 1$ as carrying the small-scale information of the process, while the part with $t \gg 1$ carries the larger-scale information. The first, very elementary but crucial, modification to we then bring forward is that, by reversibility, $$\label{e.reversible} {\left\langle}f, e^{-t {{\mathcal L}}} f {\right\rangle}= {\left\langle}e^{-\frac t 2 {{\mathcal L}}}f, e^{-\frac{t}{2}{{\mathcal L}}} f {\right\rangle}$$ so that, after a change of variables, $$\begin{aligned} \label{e.gk2} \sigma^2(f) & = 4 \int_0^{+\infty} {\left\langle}e^{-t {{\mathcal L}}}f, e^{-t{{\mathcal L}}} f {\right\rangle}\, {\mathrm{d}}t . $$ The point of this simple modification is that in order to compute ${\mathbb{E}}[(e^{-t{{\mathcal L}}} f)^2]$, we need to average over samples of $(e^{-t{{\mathcal L}}} f)^2$, and for large $t > 0$, each of these samples will be much smaller than realizations of $f e^{-t{{\mathcal L}}} f$ (by a square factor). For large $t > 0$, we thus need much fewer samples to compute the average on the right side of than we do to compute the average on the left side of . This is relevant for practical computations since it is usually more difficult to compute samples of $e^{-t{{\mathcal L}}} f$ when $t$ is large. In other words, we decomposed the problem of computing $\sigma^2(f)$ into a small-scale part ($t \lesssim 1$) which has a contribution of order one and is an average over many samples that are relatively easy to compute, and a large-scale part ($t \gg 1$) which refines the estimate, is computationally more demanding, but requires much fewer samples due to the own smallness of each of these samples. This is what will ultimately allow for the rapid shrinking of the domain of interest in as $k$ increases. The second modification to the Green-Kubo formula we consider is meant to replace the integral over time by a sum over a moderate number of terms. The starting point is the resolvent formula, which states that for every $\lambda, \mu > 0$, $$\label{e.resolvent} R_\lambda = R_\mu + (\mu - \lambda) R_\lambda R_\mu, \quad \text{where } R_\lambda := (\lambda + {{\mathcal L}})^{-1}.$$ For any given sequence $(\mu_k)_{k \in {\mathbb{N}}}$ of real numbers in $(0,1]$, we thus apply recursively and obtain, at least formally, $$\label{e.sum.resolv} R_0 = \sum_{k = 0}^{+\infty} \mu_0 \cdots \mu_{k-1} R_{\mu_0} \cdots R_{\mu_k}.$$ As will be explained in more details below (see Theorem \[t.pgk\]), defining recursively $$f_{-1} := f \quad \text{and} \quad \forall k \in {\mathbb{N}}, \quad f_k := \mu_k R_{\mu_{k}} f_{k-1},$$ we can use the formula and reversibility to get $$\label{e.induc.Poisson} {\left\langle}f, {{\mathcal L}}^{-1} f {\right\rangle}= \sum_{k = 0}^{+\infty} \mu_k^{-1} {\left}( {\left\langle}f_{k-1}, f_k {\right\rangle}+ {\left\langle}f_k, f_k {\right\rangle}{\right}).$$ This is the variant of the Green-Kubo formula that we will use. The choice of the parameters $\mu_k$ can be adjusted according to the problem at hand. Here we mostly have in mind dynamics that relax to equilibrium at a polynomial rate. In this case, it seems appropriate to choose $\mu_k$ to scale to $0$ geometrically, e.g. by fixing $\mu_k \equiv 2^{-k}$. While the series is infinite, we will show that the partial sums typically converge to their limit exponentially fast. In Section \[s.1notation\], we drop the reversibility assumption and define our notation precisely. We discuss Monte-Carlo methods in Section \[s.mcmc\]. We then prove and, under suitable assumptions, give estimates on the rate of convergence of the partial sums to their limit in Section \[s.variants.gk\]. We show how to adapt these considerations to discrete-time Markov chains in Section \[s.dynsys\]. Finally, in Section \[s.examples\], we discuss briefly and at a heuristic level the comparative advantages of the different methods for interacting particle systems and Langevin dynamics in random potential. Notation and assumptions {#s.1notation} ------------------------ In this section, we drop the reversibility assumption and clarify some of the notation used in the previous section. From now on, we only assume that the measure ${\mathbb{P}}$ is invariant for the Markov process $X$. We still denote by ${{\mathcal L}}$ the infinitesimal generator of the process, and choose the sign convention so that $(e^{-t{{\mathcal L}}})_{t {\geqslant}0}$ is the associated semigroup. That is, for every bounded and continuous function $f : {\mathcal}X \to {\mathbb{R}}$ and every $t {\geqslant}0$, we have $$(e^{-t {{\mathcal L}}} f)(x) = {\mathbf{E}}_x {\left}[ f(X_t) {\right}] .$$ By Jensen’s inequality, for each $t {\geqslant}0$ and $p \in [1,\infty]$, we have that $e^{-t{{\mathcal L}}}$ is a contraction from $L^p({\mathbb{P}})$ to itself. It thus follows that for each $\lambda > 0$ and $p \in [1,\infty]$, the resolvent $$\label{e.def.R} R_\lambda := \int_0^{+\infty} e^{-\lambda t} e^{-t {{\mathcal L}}} \, {\mathrm{d}}t$$ is a well-defined operator from $L^p({\mathbb{P}})$ to itself, and moreover, $\lambda R_\lambda$ is a contraction from $L^p({\mathbb{P}})$ to itself. One can then verify that $R_\lambda = (\lambda + {{\mathcal L}})^{-1}$ and that the resolvent formula holds (see e.g. [@lig1 Theorem 3.16 and (3.7)]). In particular, for every $\lambda > 0$ and $\mu > 0$, the resolvents $R_\lambda$ and $R_\mu$ commute. We denote by ${{\mathcal L}}^*$ the adjoint of ${{\mathcal L}}$ in $L^2({\mathbb{P}})$, and by $(e^{-t{{\mathcal L}}^*})_{t {\geqslant}0}$ and $(R_\lambda^*)_{\lambda > 0}$ the associated semigroup and resolvent. We fix a mean-zero function $f \in L^2({\mathbb{P}})$. In the non-reversible setting, it is more difficult to assert whether the additive functional $$\label{e.add.func2} \frac 1 {\sqrt{t}} \int_{0}^t f(X(s)) \, {\mathrm{d}}s$$ satisfies a central limit theorem. We refer to [@klo Chapter 2] for several results in this direction. We only record here that under relatively weak conditions, the limit $$\label{e.def.sigma2} \sigma^2(f) := 2 \lim_{\lambda \to 0} {\left\langle}f, (\lambda + {{\mathcal L}})^{-1} f {\right\rangle}$$ is well-defined, and as $t$ tends to infinity, the random variable in converges in law under ${\mathbf{P}}_{\mathbb{P}}$ to a centered Gaussian of variance $\sigma^2(f)$ (see in particular [@klo Theorem 2.7 and (2.14)]). Part of the difficulty with non-reversible dynamics is that quantities such as ${\left\langle}f, (\lambda + {{\mathcal L}})^{-1} f {\right\rangle}$ or ${\left\langle}f, e^{-t{{\mathcal L}}} f {\right\rangle}$ are no longer necessarily nonnegative. Since the emphasis of the present paper is on numerical methods for the computation of $\sigma^2(f)$ rather than on subtleties that may arise in bordeline cases, we make the simplifying assumption that $$\label{e.hyp.int} \int_0^{+\infty} \big| {\left\langle}f, e^{-t {{\mathcal L}}} f {\right\rangle}\big| \, {\mathrm{d}}t < \infty.$$ Under this assumption, the existence and finiteness of the limit in the definition of $\sigma^2(f)$, see , are easily shown using . The validity of the classical Green-Kubo formula  is obtained similarly. Whenever we want to go beyond the derivation of identities and want to discuss rates of convergence, we will assume that there exists ${\alpha}> 1$ such that for every $t {\geqslant}0$, $$\label{e.assump.decay} \big|{\left\langle}f, e^{-t{{\mathcal L}}} f {\right\rangle}\big| {\leqslant}(1+t)^{-\alpha}.$$ Note that a possible multiplicative constant in this inequality could be absorbed after a redefinition of $f$. We will indicate explicitly whenever or is assumed. Monte-Carlo methods {#s.mcmc} ------------------- In this section, we show that a single-step extrapolation of the naive Monte-Carlo estimator for $\sigma^2(f)$ allows for a significant improvement in computational complexity. In addition to the quantitative assumption , we also assume that $$\label{e.fourth} \sup_{t {\geqslant}1} {\mathbf{E}}_{\mathbb{P}}{\left}[ {\left}( \frac 1 {\sqrt{t}} \int_0^t f(X(s)) \, {\mathrm{d}}s {\right})^4 {\right}] {\leqslant}1.$$ The assumption of implies that the supremum on the left side of , with the fourth power replaced by the second power, is finite. The assumption of finiteness of the fourth moment in (which can then normalized to be of unit size, up to a redefinition of $f$) will be used to control the fluctuating part of the Monte-Carlo estimator. In order to describe MCMC methods, we assume that we are given a family $X^{(0)}, X^{(1)}, \ldots$ of processes that are independent and are each distributed according to the law ${\mathbf{P}}_{\mathbb{P}}$ (that is, the law of $X$ started from the measure ${\mathbb{P}}$). We denote their joint law by ${\mathbf{P}}_{\mathbb{P}}^{\otimes}$ (expectation ${\mathbf{E}}_{\mathbb{P}}^\otimes$). The next proposition quantifies the accuracy of the naive Monte-Carlo estimator. \[p.naive\] For each ${\alpha}> 1$, there exists a constant $C < \infty$ such that the following holds. Let $f \in L^2({\mathbb{P}})$ satisfy and , and for each $t {\geqslant}2$ and positive integer $N$, let $$\label{e.def.hat0} {\widehat}{\sigma}_0^2(N,t,f) := \frac 1 N \sum_{k = 1}^{N} {\left}(\frac 1 {\sqrt{t}}\int_0^t f(X^{(k)}(s)) \, {\mathrm{d}}s{\right})^2.$$ We have $$\label{e.naive} {\mathbf{E}}_{\mathbb{P}}^{\otimes} {\left}[ {\left}( {\widehat}{\sigma}_0^2(N,t,f) - \sigma^2(f) {\right}) ^2 {\right}] ^\frac 1 2 {\leqslant}C {\left}| \begin{array}{ll} N^{-\frac 1 2} + t^{-({\alpha}-1)} & \quad \text{if } {\alpha}< 2, \\ N^{-\frac 1 2} + t^{-1} \log(t) & \quad \text{if } {\alpha}= 2, \\ N^{-\frac 1 2} + t^{-1} & \quad \text{if } {\alpha}> 2. \end{array} {\right}.$$ We decompose the left side of into a variance and a bias terms: $$\begin{gathered} \label{e.varbias} {\mathbf{E}}_{\mathbb{P}}^{\otimes} {\left}[ {\left}( {\widehat}{\sigma}_0^2(N,t,f) - \sigma^2(f) {\right}) ^2 {\right}] \\ = {\mathbf{E}}_{\mathbb{P}}^{\otimes} {\left}[ {\left}( {\widehat}{\sigma}_0^2(N,t,f) - {\mathbf{E}}_{\mathbb{P}}^{\otimes} {\left}[{\widehat}{\sigma}_0^2(N,t,f) {\right}]{\right})^2 {\right}] + |{\mathbf{E}}_{\mathbb{P}}^{\otimes} {\left}[{\widehat}{\sigma}_0^2(N,t,f) {\right}] - \sigma^2(f)|^2.\end{gathered}$$ By , it is clear that the first term on the right side above is bounded by $N^{-1}$. By stationarity, the expectation appearing in the second term can be rewritten as $$\begin{aligned} \notag {\mathbf{E}}_{\mathbb{P}}{\left}[{\left}(\frac 1 {\sqrt t} \int_0^{t} f(X(s)) \, {\mathrm{d}}s{\right})^2 {\right}] & = \frac{2}{t} \int_{0 {\leqslant}u < v {\leqslant}t} {\mathbf{E}}_{\mathbb{P}}{\left}[ f(X(u)) f(X(v)) {\right}] \, {\mathrm{d}}u \, {\mathrm{d}}v \\ \notag & = \frac{2}{t}\int_{0 {\leqslant}u < v {\leqslant}t} {\mathbf{E}}_{\mathbb{P}}{\left}[ f(X(0)) f(X(v-u)) {\right}] \, {\mathrm{d}}u \, {\mathrm{d}}v \\ \label{e.naivebias} & = \frac{2}{t} \int_0^t (t-s) {\left\langle}f, e^{-s {{\mathcal L}}} f {\right\rangle}\, {\mathrm{d}}s.\end{aligned}$$ Recalling the formula for $\sigma^2(f)$ and using , we can bound the difference between the integral above and $\sigma^2(f)$ by $$\frac 2 t \int_0^t s (1+s)^{-{\alpha}} \, {\mathrm{d}}s + 2 \int_t^{+\infty} (1+s)^{-{\alpha}} \, {\mathrm{d}}s {\leqslant}C {\left}| \begin{array}{ll} t^{-({\alpha}-1)} & \quad \text{if } {\alpha}< 2, \\ t^{-1} \log(t) & \quad \text{if } {\alpha}= 2, \\ t^{-1} & \quad \text{if } {\alpha}> 2. \end{array} {\right}.$$ This completes the proof. The next proposition, which is the main result of this section, shows that a simple one-step extrapolation of the naive MCMC estimator yields an improved convergence rate in most cases. \[p.improved\] For each ${\alpha}> 1$, there exists a constant $C < \infty$ such that the following holds. Let $f \in L^2({\mathbb{P}})$ satisfy and , and for each $t {\geqslant}2$ and positive integer $N$, let $${\widehat}{\sigma}_1^2(N,t,f) := 2{\widehat}{\sigma}_0^2(N,2t,f) - {\widehat}{\sigma}_0^2(N,t,f),$$ where ${\widehat}{\sigma}_0^2(N,t,f)$ is defined in . We have $${\mathbf{E}}_{\mathbb{P}}^{\otimes} {\left}[ {\left}( {\widehat}{\sigma}_1^2(N,t,f) - \sigma^2(f) {\right}) ^2 {\right}] ^\frac 1 2 {\leqslant}C {\left}(N^{-\frac 1 2} + t^{-({\alpha}- 1)}{\right}).$$ We use the variance-bias decomposition (with ${\widehat}{\sigma}_0^2$ replaced by ${\widehat}{\sigma}_1^2$) and estimate the variance as in the proof of Proposition \[p.naive\]. There remains to estimate the bias term, which takes the form $${\left}|2{\mathbf{E}}_{\mathbb{P}}{\left}[{\left}(\frac 1 {\sqrt {2 t}} \int_0^{2t} f(X(s)) \, {\mathrm{d}}s{\right})^2 {\right}] - {\mathbf{E}}_{\mathbb{P}}{\left}[{\left}(\frac 1 {\sqrt t} \int_0^{t} f(X(s)) \, {\mathrm{d}}s{\right})^2 {\right}] - \sigma^2(f) {\right}|.$$ In view of and , we can rewrite this quantity as $$\begin{aligned} 2{\left}| \int_0^{+\infty} {\left}[2{\left}( 1 - \frac s {2t} {\right})_+ - {\left}( 1 - \frac s t {\right})_+ - 1 {\right}] {\left\langle}f, e^{-s{{\mathcal L}}} f {\right\rangle}\, {\mathrm{d}}s {\right}|,\end{aligned}$$ where we denote $x_+ = x \vee 0 = \max(x,0)$ for the positive part of $x$. We also write $x \wedge y = \min(x,y)$. Simplifying and using , we get that the expression above is $$\begin{aligned} \label{e.extrap} 2{\left}| \int_t^{+\infty} {\left}[{\left}(\frac s t-1 {\right}) \wedge 1{\right}] {\left\langle}f, e^{-s{{\mathcal L}}} f {\right\rangle}\, {\mathrm{d}}s{\right}| {\leqslant}2 \int_t^{+\infty} s^{-{\alpha}} \, {\mathrm{d}}s {\leqslant}C t^{-({\alpha}- 1)},\end{aligned}$$ as announced. The one-step extrapolated scheme described in Proposition \[p.improved\] is therefore more efficient than the naive scheme described in Proposition \[p.naive\] whenever ${\alpha}{\geqslant}2$. To make this point more precise, we may measure the computational complexity of each method in terms of its “total simulated time”: if an algorithm samples $n$ trajectories, each respectively up to time $t_1, \ldots, t_n$, then the total simulated time of the algorithm is $\sum_{i = 1}^n t_i$. For any fixed ${\alpha}> 2$, we compute the total simulated time required for each method to produce an output of the order of ${\delta}\in (0, 1]$ away from $\sigma^2(f)$. For the naive method, we need to choose $N \simeq {\delta}^{-2}$ and $t \simeq {\delta}^{-1}$, so that the total simulated time is of the order of ${\delta}^{-3}$. For the one-step extrapolated method, we choose instead $t \simeq {\delta}^{-\frac 1 {{\alpha}- 1}}$, still with $N \simeq {\delta}^{-2}$, so that the total simulated time is of the order of $ {\delta}^{-{\left}( 2 + \frac 1 {{\alpha}- 1} {\right}) } \ll {\delta}^{-3}$. If we replace the polynomial convergence to equilibrium in by an exponential one, then one can show that the one-step extrapolated method requires only of the order of ${\delta}^{-2} \log({\delta}^{-1})$ of simulated time, while the naive method still requires a total simulated time of the order of ${\delta}^{-3}$. It is striking that a single-step extrapolation suffices to get rid of “saturation” phenomena as seen in Proposition \[p.naive\] at all orders at once. Variants of the Green-Kubo formula {#s.variants.gk} ---------------------------------- In this section, we prove the non-reversible version of the variant of the Green-Kubo formula presented in . We then discuss the convergence rate of the series under the additional assumption of . \[t.pgk\] Let $f \in L^2({\mathbb{P}})$ be such that holds, let $(\mu_k)_{k \in {\mathbb{N}}}$ be a sequence of real numbers in $(0,1]$, and define recursively $$\label{e.def.fk} {\left}\{ \begin{aligned} & f_{-1} = f^*_{-1} := f, \\ & \forall k \in {\mathbb{N}}, \quad f_{k} := \mu_{k} (\mu_k + {{\mathcal L}})^{-1} f_{k-1} \quad \text{and} \quad f^*_{k} := \mu_{k} (\mu_k + {{\mathcal L}}^*)^{-1} f^*_{k-1} . \end{aligned} {\right}.$$ For every $n \in {\mathbb{N}}$, the quantity $$\label{e.def.rest} {\left\langle}f_n^*, {{\mathcal L}}^{-1} f_n {\right\rangle}:= \lim_{\lambda \to 0} {\left\langle}f_n^*, (\lambda + {{\mathcal L}})^{-1} f_n {\right\rangle}$$ is well-defined and finite, and $$\label{e.lim.rest} \lim_{n \to \infty} {\left\langle}f_n^*, {{\mathcal L}}^{-1} f_n {\right\rangle}= 0.$$ Moreover, for every $n \in {\mathbb{N}}$, we have $$\label{e.pgk} {\left\langle}f, {{\mathcal L}}^{-1} f {\right\rangle}= \sum_{k = 0}^{n} \mu_{k}^{-1} {\left}({\left\langle}f^*_{k-1}, f_k {\right\rangle}+ {\left\langle}f^*_k, f_k {\right\rangle}{\right}) + {\left\langle}f_n^*, {{\mathcal L}}^{-1} f_n {\right\rangle}.$$ Let $\lambda > 0$. From the formula , we see that if $T_\lambda$ is an exponential random variable of parameter $\lambda$, independent of any other quantity in the problem, then we can represent the resolvent operator $R_\lambda$ as $$\lambda R_\lambda g = E {\left}[ e^{-T_\lambda {{\mathcal L}}} g {\right}] ,$$ where $E$ denotes the expectation over the random variable $T_\lambda$. By induction, we see that if $S_n$ has the law of a sum of $(n+1)$ independent exponential random variables of parameter $\mu_0,\ldots, \mu_n$ respectively, then $$f_n = E {\left}[ e^{-S_n {{\mathcal L}}} f {\right}].$$ Here again the random variable $S_n$ is independent of any other quantity in the problem, and $E$ computes the average of this random variable. Denoting by $S_n'$ an independent copy of $S_n$, we have $$\begin{aligned} {\left\langle}f_n^*, (\lambda + {{\mathcal L}})^{-1} f_n {\right\rangle}& = \int_0^{+\infty} e^{-\lambda t} E {\left}[ {\left\langle}e^{-S_n' {{\mathcal L}}^*} f, e^{-(S_n + t) {{\mathcal L}}} f {\right\rangle}{\right}] \, {\mathrm{d}}t \\ & = \int_0^{+\infty} e^{-\lambda t} E {\left}[ {\left\langle}f, e^{-(S_n + S_n' + t){{\mathcal L}}} f {\right\rangle}{\right}] \, {\mathrm{d}}t.\end{aligned}$$ The assumption of ensures that for each fixed $s {\geqslant}0$, $$\int_0^{+\infty} {\left}| {\left\langle}f, e^{-(s+t){{\mathcal L}}} f {\right\rangle}{\right}| \, {\mathrm{d}}t {\leqslant}\int_0^{+\infty} {\left}| {\left\langle}f, e^{-t{{\mathcal L}}} f {\right\rangle}{\right}| \, {\mathrm{d}}t < +\infty.$$ Hence, the quantity $$\int_0^{+\infty} e^{-\lambda t} {\left\langle}f,e^{-(s+t){{\mathcal L}}} f {\right\rangle}\, {\mathrm{d}}t$$ is bounded uniformly over $s {\geqslant}0$ and $\lambda > 0$, and converges to $$\int_0^{+\infty} {\left\langle}f, e^{-(s+t) {{\mathcal L}}} f {\right\rangle}\, {\mathrm{d}}t$$ as $\lambda$ tends to $0$. By the dominated convergence theorem, this implies that $${\left\langle}f_n^*, (\lambda + {{\mathcal L}})^{-1} f_n {\right\rangle}$$ converges as $\lambda$ tends to infinity to $$\int_0^{+\infty} E {\left}[ {\left\langle}f, e^{-(S_n + S_n' + t){{\mathcal L}}} f {\right\rangle}{\right}] \, {\mathrm{d}}t \in {\mathbb{R}}.$$ We now verify that the integral above tends to $0$ as $n$ tends to infinity. It is clear that we can realize the sequences $S_n, S_n'$ such that $S_{n+1} {\geqslant}S_n$ and $S'_{n+1} {\geqslant}S_n'$, and moreover, since the sequence $(\mu_k)$ is bounded by $1$, we have that $S_n$ and $S_n'$ tend to infinity almost surely. For each $M {\geqslant}1$, we can bound the absolute value of the integral above by $$\label{e.easy.decomp} P {\left}[ S_n + S_n' {\leqslant}M {\right}] \int_0^{+\infty} {\left}| {\left\langle}f, e^{-t{{\mathcal L}}} f {\right\rangle}{\right}| \, {\mathrm{d}}t + \int_M^{+\infty} {\left}| {\left\langle}f, e^{-t{{\mathcal L}}} f {\right\rangle}{\right}| \, {\mathrm{d}}t.$$ By the assumption of , we may fix $M$ sufficiently large to make the second integral in as small as desired. Then taking $n$ sufficiently large ensures that the probability on the left of is arbitarily small. This completes the proof of . There remains to show . For each $\lambda > 0$, we introduce the shorthand notation $$\mu_k^\lambda := \mu_k-\lambda.$$ Applying the resolvent formula recursively, we see that $$\label{e.induc.resolv} R_\lambda = {\left}(\sum_{k = 0}^n \mu^\lambda_0 \cdots \mu^\lambda_{k-1} R_{\mu_0} \cdots R_{\mu_k}{\right}) + \mu^\lambda_0\cdots \mu^\lambda_n \, R_{\mu_0} \cdots R_{\mu_n} R_\lambda.$$ We define recursively $$f_{-1}^{\lambda} = f_{-1}^{\lambda, *} := f,$$ $$\forall k \in {\mathbb{N}}, \quad f^{\lambda}_{k} := \mu^\lambda_{k} R_{\mu_k} f^\lambda_{k-1} \quad \text{and} \quad f^{\lambda,*}_{k} := \mu^\lambda_{k} R_{\mu_k}^* f^{\lambda,*}_{k-1} .$$ Using the formula with the sequence $(\mu_k)_{k \in {\mathbb{N}}}$ replaced by $(\mu_0, \mu_0, \mu_1, \mu_1, \ldots)$, we get that $$\label{e.decomp.lambda} {\left\langle}f, R_\lambda f {\right\rangle}= \sum_{k = 0}^n \frac{1}{\mu^\lambda_k} {\left}( {\left\langle}f^{\lambda, *}_{k-1}, f^\lambda_k {\right\rangle}+ {\left\langle}f^{\lambda, *}_k, f^\lambda_k{\right\rangle}{\right}) + {\left\langle}f^{\lambda, *}_n, R_\lambda f^\lambda_n {\right\rangle},$$ where for simplicity, we may assume that $\lambda > 0$ is sufficiently small that $\mu_k^\lambda \neq 0$ for each $k {\leqslant}n$. Note that $f_n^\lambda$ differs from $f_n$ by a scalar factor only, and this factor tends to $1$ as $\lambda$ tends to $0$; and similarly for $f_n^{\lambda, *}$. It is therefore clear that $$\lim_{\lambda \to 0} {\left\langle}f^{\lambda, *}_n, R_\lambda f^\lambda_n {\right\rangle}= {\left\langle}f^*_n, {{\mathcal L}}^{-1} f_n {\right\rangle}.$$ Passing to the limit $\lambda \to 0$ in the sum on the right side of is similar (only simpler), and we obtain . We now give some simple criteria for evaluating the size of the remainder term in the decomposition \[e.pgk\]. This will depend on the sequence $(\mu_k)$ we choose. The simplest case is when we choose this sequence to be constant. \[p.constant.step\] For each ${\alpha}> 1$, there exists a constant $C < \infty$ such that the following holds. Let $f \in L^2({\mathbb{P}})$ satisfy , and let $(f_k)_{k \in {\mathbb{N}}}$, $(f^*_k)_{k \in {\mathbb{N}}}$ be defined according to with the choice $\mu_k \equiv 1$. For every $n \in {\mathbb{N}}$, we have $$\label{e.constant.step} |{\left\langle}f_n^*, {{\mathcal L}}^{-1} f_n {\right\rangle}| {\leqslant}C (1+n)^{-({\alpha}-1)}.$$ Denote by $S_n$ and $S_n'$ two idependent random variables, each having the law of a sum of $n$ independent exponential random variables of parameter $1$. We assume that these random variables are also independent of any other quantity discussed so far, and denote by $E$ the expectation with respect to these random variables. We have $$f_n = E {\left}[ e^{-S_n {{\mathcal L}}} f {\right}] \quad \text{and} \quad f_n^* = E {\left}[ e^{-S_n' {{\mathcal L}}^*} f {\right}] .$$ Moreover, $$\begin{aligned} {\left\langle}f_n^*, {{\mathcal L}}^{-1} f_n {\right\rangle}& = \int_0^{+\infty} {\left\langle}f^*_n, e^{-t{{\mathcal L}}} f_n {\right\rangle}\, {\mathrm{d}}t \\ & = \int_0^{+\infty} E {\left}[ {\left\langle}e^{-S_n'{{\mathcal L}}^*} f, e^{-(t+S_n) {{\mathcal L}}} f {\right\rangle}{\right}] \, {\mathrm{d}}t \\ & = \int_0^{+\infty} E {\left}[ {\left\langle}f, e^{-(t+ S_n + S_n'){{\mathcal L}}} f {\right\rangle}{\right}] \, {\mathrm{d}}t.\end{aligned}$$ By an elementary large deviation estimate, there exists a constant $C < \infty$ such that $$P {\left}[ {\left}|S_n + S_n' - 2n{\right}| {\leqslant}n {\right}] {\leqslant}C \exp {\left}( - n/C {\right}) .$$ Using the assumption , we thus get that $$\begin{aligned} \big| {\left\langle}f_n^*, {{\mathcal L}}^{-1} f_n {\right\rangle}\big| & {\leqslant}\int_0^{+\infty} E{\left}[(1+t+S_n + S_n')^{-{\alpha}}{\right}] \, {\mathrm{d}}t \\ & {\leqslant}C \exp(-n/C) + \int_0^{+\infty} (1+t+n)^{-{\alpha}} \, {\mathrm{d}}t \\ & {\leqslant}C (1+n)^{-({\alpha}- 1)},\end{aligned}$$ as announced. Under the assumption , it is also possible to use geometrically increasing step sizes with no degradation of the convergence rate, as the next proposition demonstrates. This can be interesting, since it reduces considerably the number of Poisson equations one needs to solve: for a desired precision of ${\delta}> 0$, using geometrically increasing step sizes requires only the resolution of a constant times $\log {\delta}^{-1}$ such equations, as opposed to ${\delta}^{-1/({\alpha}-1)}$ with the constant-step-size scheme of Proposition \[p.constant.step\]. (On the other hand, if the assumption was replaced by exponentially fast convergence, then the constant-step-size scheme should be preferred.) \[p.geom.step\] For each ${\alpha}> 1$, there exists a constant $C < \infty$ such that the following holds. Let $f \in L^2({\mathbb{P}})$ satisfy , and let $(f_k)_{k \in {\mathbb{N}}}$, $(f^*_k)_{k \in {\mathbb{N}}}$ be defined according to with the choice $\mu_k \equiv 2^{-k}$. For every $n \in {\mathbb{N}}$, we have $$\label{e.geom.step} |{\left\langle}f_n^*, {{\mathcal L}}^{-1} f_n {\right\rangle}| {\leqslant}C 2^{-n({\alpha}-1)}.$$ In both and , the right-hand side is a multiple of $(\mu_0^{-1} + \cdots + \mu_n^{-1})^{-({\alpha}-1)}$. Let $(T_k)_{k \in {\mathbb{N}}}$ be independent exponential random variables of respective parameter $(2^{-k})_{k \in {\mathbb{N}}}$, and for each $n \in {\mathbb{N}}$, let $$\label{e.def.Sn} S_n = \sum_{k = 0}^n T_k.$$ We assume that the random variables $(T_k)$ are independent of the other quantities of our problem, and denote by $E$ the expectation with respect to these random variables. We also give ourselves $S_n'$ an independent copy of $S_n$, and recall that $$f_n = E {\left}[ e^{-S_n {{\mathcal L}}} f {\right}] \quad \text{and} \quad f_n^* = E {\left}[ e^{-S_n{{\mathcal L}}^*} f {\right}] .$$ We decompose the rest of the proof into three steps. *Step 1.* In this first step, we show that for each $K < \infty$, there exists a constant $C < \infty$ such that for every $k {\leqslant}n$, $$\label{e.bound.Sn} P{\left}[S_n {\leqslant}2^k{\right}] {\leqslant}C \exp {\left}( -K(n-k) {\right}) .$$ By Chebyshev’s inequality, we have $$\begin{aligned} P{\left}[S_n {\leqslant}2^k{\right}] & {\leqslant}{\mathbb{E}}{\left}[ \exp {\left}(1 -2^{-k} {\left}( T_1 + \cdots + T_n {\right}) {\right}) {\right}] \\ & {\leqslant}\exp {\left}( 1 - \sum_{j = 0}^n \log {\left}( 1+2^{j-k} {\right}) {\right}).\end{aligned}$$ We select $k_0(K) < \infty$ such that $$\log(1+2^{k_0}) {\geqslant}K,$$ and deduce that $$P{\left}[S_n {\leqslant}2^k{\right}] {\leqslant}\exp {\left}( 1 - K (n-k-k_0)_+ {\right}) .$$ This shows . *Step 2.* We show that for every ${\beta}\in (0,\infty)$, there exists a constant $C < \infty$ such that for ever $n$, $$\label{e.Sn.moment} E {\left}[ (1+S_n)^{-\beta} {\right}] {\leqslant}C 2^{-n\beta }.$$ We rewrite the expectation on the left side above as $$E {\left}[ (1+S_n)^{-\beta} {\right}] = \beta \int_1^{+\infty} t^{-(\beta + 1)} P {\left}[ 1+S_n {\leqslant}t {\right}] \, {\mathrm{d}}t,$$ and decompose the integral into several parts. The easiest part is $$\beta \int_{2^n}^{\infty} t^{-(\beta + 1)} P {\left}[ 1+S_n {\leqslant}t {\right}] \, {\mathrm{d}}t {\leqslant}\beta \int_{2^n}^\infty t^{-(\beta + 1)} \, {\mathrm{d}}t {\leqslant}2^{-n\beta}.$$ We split the remaining part along a dyadic sequence: $$\begin{aligned} \int_1^{2^n} s^{-(\beta+1)} P[1+S_n {\leqslant}s] \, {\mathrm{d}}s & {\leqslant}\sum_{k = 0}^{n-1} \int_{2^k}^{2^{k+1}} s^{-(\beta+1)} P[1+S_n {\leqslant}s] \, {\mathrm{d}}s \\ & {\leqslant}\sum_{k = 0}^{n-1} 2^{-k\beta} P[S_n {\leqslant}2^{k+1}].\end{aligned}$$ Using the result of the first step with the choice of $K = (\beta+1) \log 2$, we obtain that the sum above is bounded by $$C \sum_{k = 0}^{n-1} 2^{-k\beta} \, 2^{-(n-k)(\beta+1)} {\leqslant}C 2^{-n\beta}.$$ This completes the proof of . *Step 3.* We complete the proof. As in the proofs of Theorem \[t.pgk\] and Proposition \[p.constant.step\], our starting point is to observe that $${\left\langle}f_n^*, {{\mathcal L}}^{-1} f_n {\right\rangle}= \int_0^{+\infty} E {\left}[ {\left\langle}f, e^{-(t+S_n + S_n') {{\mathcal L}}} f {\right\rangle}{\right}] \, {\mathrm{d}}t.$$ Using , Fubini’s theorem and a change of variables, we infer that $$\begin{aligned} \big|{\left\langle}f_n^*, {{\mathcal L}}^{-1} f_n {\right\rangle}\big| & {\leqslant}\int_0^{+\infty} E{\left}[(1+t+S_n + S_n')^{-{\alpha}}{\right}] \, {\mathrm{d}}t \\ & {\leqslant}\int_0^{+\infty} E{\left}[(1+t+S_n)^{-{\alpha}}{\right}] \, {\mathrm{d}}t \\ & {\leqslant}({\alpha}- 1)^{-1} E {\left}[(1+S_n)^{-({\alpha}- 1)} {\right}] . $$ The conclusion then follows from . Propositions \[p.constant.step\] and \[p.geom.step\] provide us with guidelines for the choice of the terminal index $n$ in the series expansion of ${\left\langle}f, {{\mathcal L}}^{-1} f {\right\rangle}$ given in . One central feature of this series expansion is the fact that as the index $k$ of the series increases and $f_k$ and $f_k^*$ become more difficult to compute, these quantities also become smaller, and we therefore need much fewer samples to evaluate the average ${\left\langle}f_k^*, f_k {\right\rangle}$ accurately. The smallness of $f_k$ is most clearly seen if we specialize the proposition below to the reversible case. \[p.size.fk\] There exists a constant $C < \infty$ such that the following holds. Let $f \in L^2({\mathbb{P}})$ satisfy , and let $(f_k)_{k \in {\mathbb{N}}}$, $(f_k^*)_{k \in {\mathbb{N}}}$) be defined according to with the choice $\mu_k \equiv 1$. For every $k \in {\mathbb{N}}$, we have $${\left}|{\left\langle}f_{k-1}^*, f_k {\right\rangle}{\right}| + {\left}|{\left\langle}f_k^*, f_k {\right\rangle}{\right}| {\leqslant}C (1+k)^{-{\alpha}}.$$ If instead, the sequences $(f_k)_{k \in {\mathbb{N}}}$ and $(f_k^*)_{k \in {\mathbb{N}}}$ are defined according to with the choice $\mu_k \equiv 2^{-k}$, then for every $k \in {\mathbb{N}}$, $${\left}|{\left\langle}f_{k-1}^*, f_k {\right\rangle}{\right}| + {\left}|{\left\langle}f_k^*, f_k {\right\rangle}{\right}| {\leqslant}C 2^{-k{\alpha}}.$$ The proof of this proposition is almost identical to the proofs of Propositions \[p.constant.step\] and \[p.geom.step\], so we omit it. The point we wish to stress here is that at least under the assumption of reversibility (that is, ${{\mathcal L}}^* = {{\mathcal L}}$ and thus $f_k^* = f_k$), the typical size of $f_k$ (as measured in $L^2({\mathbb{P}})$) is small: it is of the order of $k^{-{\alpha}/2}$ for constant step sizes, and of the order of $2^{-k{\alpha}/2}$ for geometrically increasing step sizes. In the non-reversible setting, the situation is more subtle. Indeed, the assumption of no longer informs us on the size of $e^{-t{{\mathcal L}}} f$ in $L^2({\mathbb{P}})$; similarly, we can no longer sharply estimate the size of $f_k$ (whatever the choice of $\mu_k$) from this assumption. In Proposition \[p.decay.discrete\] below, we consider dynamical systems (which in some sense is the “worst-case scenario” for these bounds) and show, under an assumption comparable to , that $f_k$ is of the order of $k^{-\frac 1 4}$ in the case of $\mu_k \equiv 1$, and that $f_k$ is of the order of $2^{-\frac k 2}$ in the case of geometrically increasing step sizes. This gives further motivation for using geometrically increasing step sizes. Discrete-time Markov chains {#s.dynsys} --------------------------- In this section, we explain how to adapt the arguments of the previous section to the setting of discrete-time Markov chains. We then discuss the rate of decay to zero of the remainder in the series expansion in the “extreme” case of dynamical systems. Let $(X_n)_{n \in {\mathbb{N}}}$ be a discrete-time Markov chain on ${\mathcal{X}}$, and let $P$ be its transition operator. That is, with hopefully transparent notation, $$Pf(x) := {\mathbf{E}}_x {\left}[ f(X_1) {\right}].$$ We define $$\label{e.newL} {{\mathcal L}}:= \mathrm{Id} - P.$$ We assume that the measure ${\mathbb{P}}$ is invariant for $(X_n)$, and continue denoting by ${\left\langle}\cdot, \cdot {\right\rangle}$ the scalar product in $L^2({\mathbb{P}})$. We fix $f \in L^2({\mathbb{P}})$ of mean zero. As in the continuous case, there is no simple and general theorem asserting whether the additive functional $$\label{e.additive.discrete} \frac 1 {\sqrt{n}} \sum_{k = 0}^n f(X_k)$$ converges in law to a Gaussian as $n$ tends to infinity. Note that at this level of generality, the question covers dynamical systems as a particular case; we refer to [@liver; @dolg; @melbourne] and references therein for results in this context. In analogy with the continuous case, we assume throughout that $$\label{e.hyp.int.dynsys} \sum_{k = 0}^{+\infty} \big| {\left\langle}f, P^k f {\right\rangle}\big| < \infty.$$ Under some additional assumptions (e.g. reversibility is sufficient [@kipvar]), the random variable in converges in law to a centered Gaussian of variance $$\begin{aligned} \notag \sigma^2(f) & := {\left\langle}f, f {\right\rangle}+ 2 \lim_{\lambda \to 0} {\left\langle}f, (\lambda + {{\mathcal L}})^{-1} f {\right\rangle}\\ \label{e.identity.resolv.discrete} & = - {\left\langle}f, f {\right\rangle}+ 2\sum_{k = 0}^{+\infty} {\left\langle}f, P^k f {\right\rangle}.\end{aligned}$$ The analogue of the formula , adapted to the non-reversible and discrete-time setting, reads $$\label{e.gk2.discrete} \sigma^2(f) = -{\left\langle}f, f {\right\rangle}+ 2 \sum_{k = 0}^{+\infty} \big( {\left\langle}(P^*)^k f, P^k f {\right\rangle}+ {\left\langle}(P^*)^k f, P^{k+1} f {\right\rangle}\big) ,$$ where $P^*$ is the adjoint of $P$ in $L^2({\mathbb{P}})$. This formula can be used directly to evaluate $\sigma^2(f)$ (this is essentially the formula used in Section \[s.discrete\] below). We also have that the exact statement of Theorem \[t.pgk\], with ${{\mathcal L}}$ now defined by , holds—and the proof given there applies with essentially no change besides a minor adaptation to the discrete-time setting. The minor adaptation boils down to replacing the formula for the resolvent in with $$\label{e.alt.R} R_\lambda = \sum_{k = 0}^{+\infty} (\lambda + 1)^{-(k+1)} \, P^k.$$ Strengthening , and in analogy with , we may assume that there exists an exponent ${\alpha}> 1$ such that for every $k \in {\mathbb{N}}$, $$\label{e.decay.disc} \big| {\left\langle}f, P^k f {\right\rangle}\big| {\leqslant}(1+k)^{-{\alpha}}.$$ This assumption implies the validity of Propositions \[p.constant.step\], \[p.geom.step\] and \[p.size.fk\]. The proofs are almost identical except for the use of in place of . With reversible dynamics, the assumption of directly translates into information on the size of $P^k f$ itself. This is not so in general. We can illustrate this phenomenon most clearly in the particular case of dynamical systems. Indeed, assume that there exists a measure-preserving mapping ${\mathsf T}: {\mathcal{X}}\to {\mathcal{X}}$ such that for every $n$, $X_n = {\mathsf T}^n(X_0)$, or equivalently, $$\label{e.def.dynsys} P f = f \circ {\mathsf T}.$$ Then the $L^2$ norm of $P^k f$ does not depend on $k$, since ${\mathsf T}$ is measure-preserving. In this case, we thus need many samples of ${\mathsf T}^k f$ and $({\mathsf T}^*)^k f$ if we wish to estimate the average ${\left\langle}({\mathsf T}^*)^k f, {\mathsf T}^k f {\right\rangle}$ using independent draws. On the other hand, we may expect the quantity $f_k$ appearing in Theorem \[t.pgk\] to display some cancellations that make it smaller. The next proposition quantifies this more precisely. \[p.decay.discrete\] There exists a constant $C < \infty$ such that the following holds. Let $P$ be the transition operator of a dynamical system, see , let $f \in L^2({\mathbb{P}})$ satisfy , and let $(f_k)_{k \in {\mathbb{N}}}$, $(f_k^*)_{k \in {\mathbb{N}}}$ be defined according to with the choice $\mu_k \equiv 1$. For every $k \in {\mathbb{N}}$, we have $$\label{e.disc1} {\left}|{\left\langle}f_{k}^*, f_k^* {\right\rangle}{\right}| + {\left}|{\left\langle}f_k, f_k {\right\rangle}{\right}| {\leqslant}C (1+k)^{-\frac 1 2}.$$ If instead, the sequences $(f_k)_{k \in {\mathbb{N}}}$ and $(f_k^*)_{k \in {\mathbb{N}}}$ are defined according to with the choice $\mu_k \equiv 2^{-k}$, then for every $k \in {\mathbb{N}}$, $$\label{e.disc2} {\left}|{\left\langle}f_{k}^*, f_k^* {\right\rangle}{\right}| + {\left}|{\left\langle}f_k, f_k {\right\rangle}{\right}| {\leqslant}C 2^{-k}.$$ We stress again that the exact statement of Proposition \[p.size.fk\] is true under the assumption of . Proposition \[p.decay.discrete\] is a statement about the respective sizes of $f_k^*$ and $f_k$, *not* about the size of the correlation between these quantities. Let $T_1, T_2, \ldots$ be independent random variables following the same law of geometric type: $$\label{e.def.geom} \forall k \in {\mathbb{N}}, \quad P {\left}[ T_i = k {\right}] = \mu_i (\mu_i + 1)^{-(k+1)} .$$ We assume that these random variables are independent of the other randomness of the problem, and denote by $E$ the expectation with respect to these variables. We also set, for each $k \in {\mathbb{N}}$, $$S_k := T_1 + \cdots + T_k,$$ and give ourselves an independent copy $S_k'$ of $S_k$. In view of the formula , we have the representation $$f_k = E {\left}[ P^{S_k} f {\right}] ,$$ and thus $${\left\langle}f_k, f_k {\right\rangle}= E {\left}[ {\left\langle}P^{S_k} f, P^{S_k'} f {\right\rangle}{\right}] .$$ Since the mapping ${\mathsf T}$ is measure-preserving, we deduce that $${\left\langle}f_k, f_k {\right\rangle}= E {\left}[ {\left\langle}f, P^{|S_k - S_k'|} f {\right\rangle}{\right}] ,$$ and by the assumption of and the triangle inequality, we get $$\begin{aligned} \big| {\left\langle}f_k, f_k {\right\rangle}\big| & {\leqslant}E {\left}[ (1+|S_k - S_k'|)^{-{\alpha}} {\right}] \\ & {\leqslant}C E {\left}[ (1+ |S_k - E[S_k]|)^{-{\alpha}} {\right}] .\end{aligned}$$ We now use the decomposition $$\label{e.fubini.friend} E {\left}[ (1+ |S_k - E[S_k]|)^{-{\alpha}} {\right}] = \int_1^\infty \frac{{\alpha}}{s^{{\alpha}+ 1}} \, P {\left}[ 1 + |S_k - E[S_k]| {\leqslant}s {\right}] \, {\mathrm{d}}s.$$ We first analyze the case of constant step size, that is $\mu_k \equiv 1$. In this case, the random variable $S_k$ is a sum of independent and identically distributed random variables. Denote by $\tau^2$ the variance of $T_1$. By the Berry-Esseen theorem, see [@petrov Theorem 5.5], there exists a constant $C < \infty$ such that for every $k \in {\mathbb{N}}$ and $s \in {\mathbb{R}}$, $${\left}|P{\left}[ {\left}| S_k - E[S_k] {\right}| {\leqslant}s\sqrt{k}{\right}] - \frac {2}{\tau\sqrt{2\pi}} \int_{0}^s \exp {\left}( -\frac{x^2}{2\tau^2} {\right}) \, {\mathrm{d}}x {\right}| {\leqslant}\frac{C}{\sqrt{k}}.$$ We thus obtain that $$\begin{aligned} E {\left}[ (1+ |S_k - E[S_k]|)^{-{\alpha}} {\right}] & {\leqslant}\frac{C}{\sqrt k} + C \int_1^\infty \frac{1}{s^{{\alpha}+ 1}} \int_{0}^{\frac s {\sqrt{k}}} \exp {\left}( -\frac{x^2}{2\tau^2} {\right}) \, {\mathrm{d}}x \, {\mathrm{d}}s \\ & {\leqslant}\frac{C}{\sqrt{k}} + C \int_1^{\infty} \frac 1 {s^{{\alpha}+ 1}} \, \frac{s}{\sqrt{k}} \, {\mathrm{d}}s \\ & {\leqslant}\frac{C}{\sqrt{k}},\end{aligned}$$ as announced. We now turn to the case of geometrically increasing step sizes, that is, $\mu_k \equiv 2^{-k}$. We first show that there exists a constant $C < \infty$ such that for every $k \in {\mathbb{N}}$ and $s {\geqslant}0$, $$\label{e.easy.decay} P{\left}[|S_k - E[S_k]| {\leqslant}s{\right}] {\leqslant}C 2^{-k} {\left}({s+1}{\right}).$$ In order to do so, it suffices to verify that there exists $C < \infty$ such that for every $k \in {\mathbb{N}}$ and $x \in {\mathbb{N}}$, $$P {\left}[ S_k = x {\right}] {\leqslant}C 2^{-k}.$$ It follows from the definition of $T_k$ that this property is satisfied if $S_k$ is replaced by $T_k$. By independence between $T_k$ and $S_{k-1}$, we deduce $$\begin{aligned} P {\left}[ S_k = x {\right}] & = E {\left}[ P {\left}[ T_k = x - S_{k-1} \ | \ S_{k-1} {\right}] {\right}] \\ & {\leqslant}E {\left}[ \sup_{z \in {\mathbb{Z}}} P{\left}[T_k = z \ | \ S_{k-1} {\right}] {\right}] \\ & {\leqslant}C 2^{-k}.\end{aligned}$$ This implies . Combining this with yields $$E {\left}[ (1+ |S_k - E[S_k]|)^{-{\alpha}} {\right}] {\leqslant}C 2^{-k} \int_1^\infty \frac{1}{s^{{\alpha}}} \, {\mathrm{d}}s {\leqslant}C 2^{-k},$$ completing the proof. One can verify that the estimates obtained in Proposition \[p.decay.discrete\] are sharp, in the sense that assuming the converse inequality to , one can show the converse inequalities to and , up to multiplicative constants. Examples {#s.examples} -------- In the second part of the paper, we explore thoroughly the example of a reversible random walk in a random environment, or equivalently the homogenization of discrete divergence-form operators with random coefficients. Under some assumptions, the relaxation property holds with ${\alpha}= \frac d 2 + 1$. We will explain how to implement numerical schemes derived from Theorem \[t.pgk\] by leveraging on the mixing properties of the functions $f_k$ under spatial translations. We will show that such numerical schemes achieve essentially optimal complexity, requiring $C {\delta}^{-2} \log ({\delta}^{-1})$ computations to compute the effective diffusivity at precision ${\delta}> 0$. An alternative route to solve the same problem would be to use Monte-Carlo methods. Computing the effective diffusivity at precision $\delta > 0$ using the method described in Proposition \[p.improved\] would call for the simulation of $\delta^{-2}$ random walk trajectories that run up to time $\delta^{-\frac 2 d}$, bringing the total cost of the method to a suboptimal $\delta^{-2-\frac 2 d}$. (See [@MCdisc Theorem 4.1] for the verification that a large deviation estimate much stronger than holds.) In this section, we discuss briefly other examples of interest where the methods described here can be applied. ### Tagged particle in exclusion process One example where the approach based on Theorem \[t.pgk\] seems very costly to implement is the problem of computing the effective diffusivity of a tagged particle in a system of particles evolving according to, say, exclusion dynamics on ${\mathbb{Z}}^d$. We refer to [@lig2] for the definition of this and other models of interacting particle systems. Indeed, the Poisson equations appearing in are posed in $\{0,1\}^{{\mathbb{Z}}^d}$. Even if we restrict the dynamics to a finite box $B {\subseteq}{\mathbb{Z}}^d$, the size of each ergodic component of the dynamics is exponentially large as a function of the volume of $B$, which seems to preclude any computationally efficient approach to the resolution of these equations. In contrast, the simulation of the dynamics on $B$ is feasible, and by exchangeability, the trajectory of each of the simulated particles can be used as a sample of the trajectory of a tagged particle. One would then need to assert precisely the independence properties of these trajectories, but it is clear that the diffusivity can be computed in polynomial time using an MCMC scheme, and possibly in about ${\delta}^{-2-\frac 2 d}$ operations. ### Langevin dynamics Hypoelliptic dynamics can also be considered. We explore briefly here the case of Langevin dynamics in a random potential. Let $\phi$ be a smooth stationary and ergodic function over ${{\mathbb{R}^d}}$. We consider the Markov process $(X_t, V_t)_{t {\geqslant}0}$, taking values in ${{\mathbb{R}^d}}\times {{\mathbb{R}^d}}$, whose infinitesimal generator is given by $$\label{e.def.LL} L f := \Delta_vf - v \cdot \nabla_vf - \nabla_x \phi \cdot \nabla_v f + v \cdot \nabla_x.$$ Note that in this case, the process $(X_t)$ itself is an additive functional, since $$X_t = \int_0^t V_s \, {\mathrm{d}}s,$$ while for random walks or tagged particles the connection between the dynamics and additive functionals is less immediate (but see [@kipvar]). The fact that $({\varepsilon}X_{{\varepsilon}^{-2} t})_{t {\geqslant}0}$ converges in law to Brownian motion as ${\varepsilon}$ tends to zero was proved in [@papvar]. Moreover, the invariant measure for the process of the environment as seen from $X$ is explicit. Under suitable assumptions on the mixing properties of $\phi$, we may expect, in analogy with the simpler setting explore in the second part of the paper, that the polynomial relaxation holds with ${\alpha}= \frac d 2 + 1$. It can be relatively costly (but not as prohibitive as for interacting particle systems) to compute approximate solutions to the Poisson equations appearing in . Compared with the situation encountered for random walks, the additional difficulty stems in part from the fact that the operator is hypoelliptic, but also simply from the fact that there are more variables to solve for (due to the addition of velocity coordinates). The superiority of this approach compared to that based on Monte-Carlo sampling is thus unclear, and probably depends on the situation. In one extreme case, if the dimension is small and $\phi$ is a deterministic periodic function (whose periods need not be known), then an approach similar to that exposed in the second part of the paper is most probably best (and it should be preferable to choose $\mu_k \equiv 1$ there, since the convergence to equilibrium is exponentially rapid in the periodic setting). At the other extreme, in the random case and in high dimension, the Monte-Carlo approach probably becomes superior. Stochastic homogenization {#part.two} ========================= Notation and basic properties {#s.defs} ----------------------------- In this second part of the paper, we study the problem of approximating the homogenized matrix of divergence-form operators with random coefficients, as stated in the introduction. In the present section, we set up notation and assumptions and recall basic facts of homogenization. We then proceed to prove the error estimate in Section \[s.error\]. In Section \[s.compl\], we complete the proof of Theorem \[t.main\] by estimating the computational complexity of the method, and prove the complexity lower bound given by Proposition \[p.lower.bound\]. We report on the results of numerical tests in Section \[s.numer\]. Finally, we explore a variant of the method based on discrete-time parabolic equations in Section \[s.discrete\]. We focus on the discrete space ${\mathbb{Z}}^d$, $d {\geqslant}2$. We say that $x, y \in {{\mathbb{Z}^d}}$ are neighbors, and write $x \sim y$, if $|x-y| = 1$. This turns ${\mathbb{Z}}^d$ into a graph, and we denote by ${\mathbb{B}}$ the corresponding set of undirected edges. For a fixed ellipticity parameter $\Lambda \in (1,\infty)$, we let $\Omega = [1,\Lambda]^{{\mathbb{B}}}$ denote the set of coefficient fields, equipped with the product $\sigma$-algebra. We denote by ${\mathbf{a}}= ({\mathbf{a}}_e)_{e \in {\mathbb{B}}}$ the canonical random variable on $\Omega$. We let ${\mathbb{P}}$ be a probability measure on $\Omega$ under which the random variables $({\mathbf{a}}_e)_{e \in {\mathbb{B}}}$ are independent and identically distributed (i.i.d.), and denote by ${\mathbb{E}}$ the associated expectation. Besides uniform ellipticity, this independence assumption is our main quantitative hypothesis. The group ${\mathbb{Z}}^d$ acts on $\Omega$ by translations via $${\left}(\theta_x {\mathbf{a}}{\right})_{e} = {\mathbf{a}}_{x + e},$$ where $x + e$ denotes the edge $e$ translated by the vector $x$. Let $({\mathbf{e}}_1,\ldots,{\mathbf{e}}_d)$ be the canonical basis of ${\mathbb{Z}}^d$. We may identify each vector ${\mathbf{e}}_i$ with the corresponding edge $\{0,{\mathbf{e}}_i\}$ when convenient. The discrete gradient $\nabla$ acts on functions $f : {{\mathbb{Z}^d}}\to {\mathbb{R}}$ according to $$\nabla f(x) := {\left}( f(x + {\mathbf{e}}_1) - f(x), \ldots, f(x+{\mathbf{e}}_d) - f(x){\right}) ,$$ while the discrete divergence acts on function $F = (F_1,\ldots,F_d) : \Omega \to {{\mathbb{R}^d}}$ via $$\nabla \cdot F(x) := \sum_{i = 1}^d {\left}( F_i(x) - F_i(x-{\mathbf{e}}_i) {\right}) .$$ One can check that up to a minus sign, the gradient and divergence operators are formal adjoints of one another with respect to the counting measure on ${{\mathbb{Z}^d}}$. The operator of interest to us is defined by $$L := -\nabla \cdot {\mathbf{a}}\nabla,$$ where here we interpret ${\mathbf{a}}(x)$ as the diagonal matrix $\operatorname{diag}({\mathbf{a}}_{x+{\mathbf{e}}_1}, \ldots,{\mathbf{a}}_{x+{\mathbf{e}}_d})$. More explicitly, for every $f : {{\mathbb{Z}^d}}\to {\mathbb{R}}$, $$Lf(x) = \sum_{y \sim x} {\mathbf{a}}_{xy} {\left}( f(x) - f(y) {\right}) .$$ It is useful to lift this operator to stationary functions on $\Omega$. We first define the corresponding discrete gradient $D$, which acts on functions $f : \Omega \to {\mathbb{R}}$ via $$D f({\mathbf{a}}) := {\left}( f(\theta_{{\mathbf{e}}_1}{\mathbf{a}}) - f({\mathbf{a}}), \ldots, f(\theta_{{\mathbf{e}}_d}{\mathbf{a}}) - f({\mathbf{a}}) {\right}),$$ and the discrete divergence, which acts on functions $F = (F_1, \ldots, F_d) : \Omega \to {{\mathbb{R}^d}}$ via $$D \cdot F({\mathbf{a}}) := \sum_{i =1}^d {\left}( F_i({\mathbf{a}}) - F_i(\theta_{-{\mathbf{e}}_i} {\mathbf{a}}) {\right}) .$$ One can verify that up to a minus sign, the gradient and divergence operators are formal adjoints of one another under the measure ${\mathbb{P}}$. The “lifted” operator is $$\label{e.def.mclL} {{\mathcal L}}:= -D \cdot {\mathbf{a}}D,$$ where we understand here that ${\mathbf{a}}= \operatorname{diag}({\mathbf{a}}_{{\mathbf{e}}_1}, \ldots,{\mathbf{a}}_{{\mathbf{e}}_d})$. In a more explicit form, for every $f : \Omega \to {\mathbb{R}}$, $${{\mathcal L}}f({\mathbf{a}}) = \sum_{z \sim 0} {\mathbf{a}}_{0z} {\left}( f({\mathbf{a}}) - f(\theta_z {\mathbf{a}}) {\right}) .$$ The operator ${{\mathcal L}}$ is self-adjoint on $L^2(\Omega)$. With each function $f : \Omega \to {\mathbb{R}}$, one can associate a stationary extension ${\widetilde}f : \Omega \times {{\mathbb{Z}^d}}\to {\mathbb{R}}$ defined by $${\widetilde}f({\mathbf{a}},x) := f(\theta_x {\mathbf{a}}).$$ Any function on $\Omega \times {{\mathbb{Z}^d}}$ that can be written in this way is said to be stationary. We let ${L^2_{\mathrm{pot}}}(\Omega)$ denote the closure in $L^2(\Omega,{{\mathbb{R}^d}})$ of the set of gradient fields $\{D f \ : \ f \in L^2(\Omega,{{\mathbb{R}^d}})\}$. Each element of ${L^2_{\mathrm{pot}}}(\Omega)$ can be realized as the gradient of a possibly non-stationary function $\Omega \times {{\mathbb{Z}^d}}\to {\mathbb{R}}$. Naturally, the gradient determines the function up to an additive constant. We denote by $D f$ any element of ${L^2_{\mathrm{pot}}}(\Omega)$, keeping in mind that $f$ may be a non-stationary function. We fix once and for all a unit vector $\xi \in {{\mathbb{R}^d}}$, and define the gradient of the corrector in the direction of $\xi$ as the unique $D\phi \in {L^2_{\mathrm{pot}}}(\Omega)$ solution to $$\label{e.def.corr} - D \cdot {\mathbf{a}}(\xi + D \phi) = 0.$$ The equation is interpreted in the weak sense of $$\label{e.weak.corr} \forall \, D f \in {L^2_{\mathrm{pot}}}(\Omega), \quad {\mathbb{E}}{\left}[ D f \cdot {\mathbf{a}}{\left}( \xi + D \phi {\right}) {\right}] = 0.$$ The existence of $D \phi$ can be showed by considering first the approximate correctors $(\phi_\mu)_{\mu > 0}$. For each $\mu > 0$, the approximate corrector $\phi_\mu \in L^2(\Omega)$ solves an equation with an additional massive term: $$\label{e.def.phimu} \mu \phi_\mu - D \cdot {\mathbf{a}}{\left}(\xi + D \phi_\mu{\right}) = 0,$$ and its gradient converges to $D\phi$: $$\label{e.conv.phimu} D \phi_\mu \xrightarrow[\mu \to 0]{L^2(\Omega)} D \phi.$$ Indeed, energy estimates show that $D\phi_\mu$ is bounded in $L^2(\Omega)$, and therefore converges weakly to some element of ${L^2_{\mathrm{pot}}}(\Omega)$ up to extraction. Using also that $\mu^\frac 1 2 \phi_\mu$ is bounded in $L^2(\Omega)$, we infer that the convergence in holds in the weak sense. In order to justify strong convergence in $L^2(\Omega)$, it then suffices to verify that $$\limsup_{\mu \to 0} {\mathbb{E}}{\left}[D \phi_\mu \cdot {\mathbf{a}}D \phi_\mu{\right}] {\leqslant}{\mathbb{E}}{\left}[D \phi \cdot {\mathbf{a}}D \phi{\right}].$$ This follows from $${\mathbb{E}}{\left}[ \mu \phi_\mu^2 + D \phi_\mu \cdot {\mathbf{a}}D \phi_\mu {\right}] = -{\mathbb{E}}[D \phi_\mu \cdot {\mathbf{a}}\xi]$$ and the fact that the right side converges to $$-{\mathbb{E}}[D \phi \cdot {\mathbf{a}}\xi] = {\mathbb{E}}[D \phi \cdot {\mathbf{a}}D \phi].$$ As is well-known, the discrete operator $-\nabla \cdot {\mathbf{a}}\nabla$ homogenizes to the continuous differential operator $-\nabla \cdot {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}\nabla$, where ${{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}$ is the constant-in-space, deterministic matrix characterized by the relation $$\label{e.charact.ahom} \xi \cdot {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}\xi = {\mathbb{E}}{\left}[ (\xi + D \phi) \cdot {\mathbf{a}}(\xi + D \phi) {\right}] .$$ Using the weak formulation , we can rewrite this formula as $$\label{e.basis} \xi \cdot {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}\xi = {\mathbb{E}}[\xi \cdot {\mathbf{a}}\xi] - {\mathbb{E}}{\left}[D \phi \cdot {\mathbf{a}}D \phi{\right}].$$ As was stated above, it is easy to compute the first term on the right side of . This term is the linear part of the mapping going from coefficient fields to homogenized matrices. One way to understand why formula may be better suited than for the development of numerical methods is by considering the case of small ellipticity contrast. If we assume, say, that $1 {\leqslant}{\mathbf{a}}{\leqslant}1 + \eta$ for some small constant $\eta > 0$, then the gradient $|D\phi|$ will be of the order of $\eta$, and thus enters as a perturbation of the order of $\eta$ in the computation of $(\xi + D \phi) \cdot {\mathbf{a}}(\xi + D \phi)$. It is only after we take expectations that we can discover that the difference $|{{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}- {\mathbb{E}}[{\mathbf{a}}]|$ is in fact of the order of $\eta^2$. In contrast, this property is directly apparent in . The method presented in Theorem \[t.main\] can be understood as a way to leverage on this phenomenon in general, through a coarsening of the problem which progressively inherits the properties of the small-ellipticity-contrast regime. Denoting by $$\label{e.def.f} f({\mathbf{a}}) := D \cdot {\mathbf{a}}\xi,$$ and by ${\left\langle}\cdot, \cdot {\right\rangle}$ the scalar product in $L^2({\mathbb{P}})$, we can rewrite the second term on the right side of as $$\label{e.Dphi.flf} {\mathbb{E}}{\left}[D\phi \cdot {\mathbf{a}}D \phi {\right}] = \lim_{\mu \to 0} {\left\langle}f, (\mu + {{\mathcal L}})^{-1} f {\right\rangle}= {\left\langle}f, {{\mathcal L}}^{-1} f {\right\rangle}.$$ Indeed, we have that $(\mu + {{\mathcal L}})^{-1} f = \phi_\mu$ and $${\left\langle}f, \phi_\mu {\right\rangle}= - {\mathbb{E}}{\left}[ D \phi_\mu \cdot {\mathbf{a}}\xi {\right}] \xrightarrow[\mu \to 0]{} -{\mathbb{E}}{\left}[ D \phi \cdot {\mathbf{a}}\xi {\right}] = {\mathbb{E}}{\left}[ D \phi \cdot {\mathbf{a}}D \phi {\right}] .$$ The problem of computing ${\left\langle}f, {{\mathcal L}}^{-1} f {\right\rangle}$ is precisely of the type discussed in the first part of the paper. Since the homogenization of the operator $-\nabla \cdot {\mathbf{a}}\nabla$ is essentially equivalent to the large-scale diffusive behavior of the corresponding random walk (which itself can be reduced to the study of an additive functional [@kipvar]), this is not surprising. We point out that no reference to random walks is necessary to prove the results of Section \[s.variants.gk\] there, which are the only results of the first part we will need here. Indeed, all the arguments in Section \[s.variants.gk\] only rely on the classical resolvent formula and on the property . That the latter is true follows from and the fact that, by reversibility, ${\left\langle}f, e^{-t{{\mathcal L}}} f {\right\rangle}{\geqslant}0$. Hence, the reader who may have skipped the first part of the paper can read Section \[s.variants.gk\] independently of the rest of the first part, with the understanding that ${{\mathcal L}}$ and $f$ are defined by and respectively. In order to compute ${\left\langle}f, {{\mathcal L}}^{-1} f {\right\rangle}$, we employ the stragegy described in Theorem \[t.pgk\] with $\mu_k \equiv 2^{-k}$. That is, we set $f_{-1} = f$ and, for each $k \in {\mathbb{N}}$, denote by $f_k \in L^2(\Omega)$ the unique solution to $$\label{e.def.fk.homog} (2^{-k} + {{\mathcal L}}) f_k = 2^{-k} f_{k-1}.$$ We thus infer from Theorem \[t.pgk\] that, for each $n \in {\mathbb{N}}$, $$\label{e.series.homog} {\left\langle}f, {{\mathcal L}}^{-1} f {\right\rangle}= \sum_{k = 0}^n 2^k {\left}( {\left\langle}f_{k-1}, f_k {\right\rangle}+ {\left\langle}f_k, f_k {\right\rangle}{\right}) + {\left\langle}f_n, {{\mathcal L}}^{-1} f_n {\right\rangle}.$$ For each $k {\geqslant}-1$, we denote by $(v_k(x), x \in {\mathbb{Z}}^d)$ the stationary extension of $f_k$, that is, $$v_k(x) := f_k(\theta_x \, {\mathbf{a}}).$$ The dependence of $v_k$ on the coefficient field ${\mathbf{a}}$ is left implicit in the notation. Using , one can check that $v_k$ solves . Error analysis {#s.error} -------------- The goal of this section is to prove the first part of Theorem \[t.main\], which gives an estimate of the error in the approximation provided by ${\widehat}{\sigma}^2_n$. Recall that $f$ is defined in and the identities and . \[t.eff\] For each ${\varepsilon}\in (0,\frac 1 2)$, there exists a constant $C < \infty$ such that the following holds. Fix $n \in {\mathbb{N}}$, and for each $k {\leqslant}n$, denote $$\label{e.def.rk} r_{k} := 2^{n-{\left}( \frac 1 2 - {\varepsilon}{\right}) k}.$$ We have $$\begin{gathered} \label{e.eff} {\mathbb{E}}{\left}[{\left}({\left\langle}f, {{\mathcal L}}^{-1} f {\right\rangle}- \sum_{k = 0}^n \frac{1}{|B_{r_k}|} \sum_{x \in B_{r_k}} 2^k {\left}( v_{k-1}(x) v_k(x) + v_k^2(x) {\right}){\right})^2 {\right}]^\frac 1 2 \\\ {\leqslant}C 2^{-\frac{nd}{2}} . \end{gathered}$$ \[e.sqrt.log\] The proof given below in fact gives the existence of a constant $C < \infty$ such that for any choice of $(r_k)_{0 {\leqslant}k {\leqslant}n}$, we have the inequality with the right side replaced by $$C 2^{-\frac {dn}{2}} + C \sum_{k = 0}^n 2^{-\frac {dk} 4} r_k^{-\frac d 2}.$$ Fixing $r_k$ according to for some ${\varepsilon}> 0$ indeed allows to bound the sum above by a constant times $2^{-\frac{dn}2}$, but other choices are possible, such as, say, $$r_k = 2^{n - \frac k 2} {\left}(1+ \frac{k^\frac 3 d}{100}{\right}).$$ For ${\varepsilon}= 0$, we obtain an upper bound of the form $C n 2^{-\frac{nd}{2}}$. I conjecture that the sharp upper bound in is $$\label{e.cond.rk} C 2^{-\frac {dn}{2}} + C 2^{-\frac{nd}{2}} {\left}(\sum_{k = 0}^n 2^{d {\left}(\frac k 2 - n{\right})} r_k^d {\right})^{\frac 1 2},$$ which for the choice of $r_k$ as in with ${\varepsilon}= 0$ gives the upper bound $C n^\frac 1 2 2^{-\frac{nd}{2}}$. As will be seen in the next section, for $r_k$ chosen as in , the asymptotic complexity of the method described in Theorem \[t.main\] does not depend on the choice of ${\varepsilon}\in [0,\frac{d-1}{2d})$. In order to prove Theorem \[t.main\], our first task is to estimate the size of the remainder in the series expansion displayed in . \[p.rest\] There exists a constant $C < \infty$ such that for every $n \in {\mathbb{N}}$, $$0 {\leqslant}{\left\langle}f_n, f_n {\right\rangle}{\leqslant}C 2^{-n {\left}( \frac d 2 + 1 {\right}) }$$ and $$0 {\leqslant}{\left\langle}f_n, {{\mathcal L}}^{-1} f_n {\right\rangle}{\leqslant}C 2^{-\frac{nd}2}.$$ Proposition \[p.rest\] is an immediate consequence of Propositions \[p.geom.step\] and \[p.size.fk\], together with the second moment bound proved in [@GNO Theorem 1] and recalled in the next proposition. \[p.moment.bound\] Recall that $f({\mathbf{a}}) = D \cdot {\mathbf{a}}\xi$. For each $p \in [1,\infty)$, there exists a constant $C < \infty$ such that for every $t {\geqslant}0$, $${\mathbb{E}}{\left}[ {\left}|e^{-t{{\mathcal L}}} f{\right}|^p {\right}] ^\frac 1 p {\leqslant}C (1 + t)^{- {\left}( \frac d 4 + \frac 1 2 {\right}) }.$$ We rely on the spatial mixing property of the random field in to compute the expected values in the series expansion in . This mixing property is quantified by the next proposition. For every random variable $X$, we denote by ${\mathbb{V}\mathrm{ar}}[X] := {\mathbb{E}}[(X-{\mathbb{E}}[X])^2]$ the variance of $X$. \[p.spat.mix\] There exists $C(d,\Lambda) < \infty$ such that for every $k \in {\mathbb{N}}$ and $r {\geqslant}0$, $$\label{e.spat.mix} {\mathbb{V}\mathrm{ar}}{\left}[ \frac{1}{|B_r|} \sum_{x \in B_r} {\left}(v_{k-1}(x) v_k(x) + v^2_k(x) {\right}){\right}] {\leqslant}C 2^{-2k {\left}( \frac d 2 + 1 {\right}) } {\left}( 1+2^{-\frac k 2} r {\right})^{-d}. $$ As will be seen shortly, we have $$\label{e.announce.fourth} {\mathbb{E}}[v_k^4(x)] = {\mathbb{E}}[f_k^4] {\leqslant}C 2^{-2k {\left}( \frac d 2 + 1 {\right}) },$$ which is the same scaling as the square of the second moment, see Proposition \[p.rest\]. Roughly speaking, Proposition \[p.spat.mix\] shows that the correlation length of the random field $x \mapsto v_{k-1}(x) v_k(x) + v_k^2(x)$ is of the order of $2^{\frac k 2}$. Indeed, the fact that the left side of is bounded by $C 2^{-2k {\left}( \frac d 2 + 1 {\right})}$ is a direct consequence of the bound . The additional term on the right side of witnesses CLT-type cancellations at length scale $2^{\frac k 2}$. In order to prove Proposition \[p.spat.mix\], we define, for each $t {\geqslant}0$ and $x \in {\mathbb{Z}}^d$, $$u(t,x) := (e^{-t{{\mathcal L}}} f)(\theta_x {\mathbf{a}}).$$ The function $u$ solves the parabolic equation $$\label{e.space.parab} {\left}\{ \renewcommand{\arraystretch}{1.5} \begin{array}{ll} \partial_t u = \nabla \cdot {\mathbf{a}}\nabla u \quad & \text{ on } {\mathbb{R}}_+ \times {{\mathbb{Z}^d}}, \\ u(0,\cdot) = \nabla \cdot {\mathbf{a}}\xi \quad & \text { on } {{\mathbb{Z}^d}}. \end{array} {\right}. $$ We also denote by $(G(t,x,y), t {\geqslant}0, x,y \in {{\mathbb{Z}^d}})$ the parabolic Green function associated with the operator $-\nabla \cdot {\mathbf{a}}\nabla$. We will make use of [@GNO Theorem 3], which we now recall. For each $e \in {\mathbb{B}}$, $x \in {{\mathbb{Z}^d}}$ and $t {\geqslant}0$, we introduce the weight function $$w_e(t,x) := 1 + \frac{|x-e|}{(1+t)^\frac 1 2},$$ where we use the notation $|x-e|$ to denote the distance between $x$ and the endpoint of the edge $e$ closest to $x$. \[p.gradient.green\] There exists an exponent $p_0(d,\Lambda) > 1$ and, for every ${\alpha}< \infty$, a constant $C(d,\Lambda,{\alpha}) < \infty$ such that for every $1 {\leqslant}p {\leqslant}p_0$, $t {\geqslant}0$ and $e \in {\mathbb{B}}$, $${\left}(\sum_{x \in {{\mathbb{Z}^d}}} {\left}(w_e^{\alpha}(t,x) |\nabla G(t,x,e)|^2 {\right})^p {\right})^\frac 1 {2p} {\leqslant}C (1 + t)^{- {\left}( \frac d 2 + \frac 1 2 {\right}) + \frac d 2 \frac 1 {2p}}.$$ We will also rely on the following observation. \[l.gradient\] There exists $C(d,\Lambda) < \infty$ such that for every $t {\geqslant}0$ and $e \in \mathbb{B}$, $${\mathbb{E}}{\left}[|\nabla u(t,e)|^2{\right}] {\leqslant}C (1 + t)^{- {\left}( \frac d 2 + 2 {\right})} .$$ We first observe that the mapping $t \mapsto {\mathbb{E}}[(\nabla u \cdot {\mathbf{a}}\nabla u)(t,x)]$ is nonincreasing. Indeed, by stationarity, and dropping the variables $(t,x)$ from the notation, $$\begin{aligned} \partial_t {\mathbb{E}}[\nabla u \cdot {\mathbf{a}}\nabla u] & = 2 {\mathbb{E}}{\left}[ \nabla {\left}( \nabla \cdot {\mathbf{a}}\nabla u {\right}) \cdot {\mathbf{a}}\nabla u {\right}] \\ & = - 2 {\mathbb{E}}{\left}[ {\left}( \nabla \cdot {\mathbf{a}}\nabla u {\right}) ^2 {\right}] {\leqslant}0.\end{aligned}$$ Moreover, $$\partial_t {\mathbb{E}}[u^2] = -2 {\mathbb{E}}{\left}[ \nabla u \cdot {\mathbf{a}}\nabla u {\right}].$$ Using also Proposition \[p.moment.bound\] with $p = 2$, we obtain $$\begin{aligned} {\mathbb{E}}[(\nabla u \cdot {\mathbf{a}}\nabla u)(t,x)] & {\leqslant}\frac 2 t \int_{\frac t 2}^t {\mathbb{E}}[(\nabla u \cdot {\mathbf{a}}\nabla u)(s,x)] \, {\mathrm{d}}s \\ & = t^{-1} {\left}( {\mathbb{E}}{\left}[ u^2{\left}(\tfrac t 2, x{\right}) {\right}] - {\mathbb{E}}{\left}[u^2(t,x){\right}] {\right}) \\ & {\leqslant}C (1+t)^{-{\left}( \frac d 2 + 2 {\right}) }.\end{aligned}$$ The announced result then follows by the assumption of uniform ellipticity of the coefficient field. For notational convenience, we focus on showing that $$\label{e.spat.mix2} {\mathbb{V}\mathrm{ar}}{\left}[ \frac 1 {|B_r|} \sum_{x \in B_r} v_k^2(x) {\right}] {\leqslant}C 2^{-2k {\left}( \frac d 2 + 1 {\right}) } {\left}( 1 + 2^{-\frac k 2} r {\right}) ^{-d}.$$ The proof that $${\mathbb{V}\mathrm{ar}}{\left}[ \frac 1 {|B_r|} \sum_{x \in B_r} v_{k-1}(x) v_k(x) {\right}] {\leqslant}C 2^{-2k {\left}( \frac d 2 + 1 {\right}) } {\left}( 1 + 2^{-\frac k 2} r {\right}) ^{-d}$$ is essentially identical, and we therefore omit it. We decompose the proof into four steps. *Step 1.* We show that there exists a constant $C < \infty$ such that for every $k \in {\mathbb{N}}$ and $x \in {{\mathbb{Z}^d}}$, $$\label{e.fourth.moment} {\mathbb{E}}{\left}[ v_k^4(x) {\right}] {\leqslant}C 2^{-k {\left}( d + 2 {\right}) }.$$ Denote by $S_k$ the random variable in . We assume that this random variable is independent of the coefficient field ${\mathbf{a}}$, and denote the expectation with respect to this random variable by $E$. We recall that we have the representation $$\label{e.recall.rep} v_k(x) = E {\left}[ u(S_k,x) {\right}] .$$ By Jensen’s inequality and Proposition \[p.moment.bound\] (with $p = 4$), we have $${\mathbb{E}}[v_k^4(x)] {\leqslant}E {\mathbb{E}}{\left}[ u^4(S_k,x) {\right}] {\leqslant}E {\left}[ (1+S_k)^{-(d+2)} {\right}] .$$ The conclusion then follows from . *Step 2.* By Jensen’s inequality and , we obtain the inequality provided that $r {\leqslant}2^{\frac k 2}$. From now on, we therefore assume that $r {\geqslant}2^{\frac k 2}$. We start by recalling the Efron-Stein (or spectral gap) inequality. Let $({\mathbf{a}}'_e)_{e \in {\mathbb{B}}}$ be an independent copy of the environment $({\mathbf{a}}_e)_{e \in {\mathbb{B}}}$, defined on the same probability space. For each edge $e \in {\mathbb{B}}$, we let ${\mathbf{a}}^e$ be the environment defined by $$\label{e.def.ae} {\mathbf{a}}^e_b := {\left}| \begin{array}{ll} {\mathbf{a}}_b & \text{if } b \neq e, \\ {\mathbf{a}}'_e & \text{if } b = e. \end{array} {\right}.$$ In other words, the environment ${\mathbf{a}}^e$ is obtained from the environment ${\mathbf{a}}$ by resampling the conductance at the edge $e$. By the independence assumption on the conductances $({\mathbf{a}}_e)$, every random variable $X$ satisfies the Efron-Stein inequality $$\begin{aligned} \label{SG} {\mathbb{V}\mathrm{ar}}[X] {\leqslant}\frac 1 2 \sum_{e \in {\mathbb{B}}} {\mathbb{E}}{\left}[ {\left}(\partial_e X {\right})^2 {\right}],\end{aligned}$$ where $\partial_e X$ denotes the Glauber derivative: $$\partial_eX(a):= X({\mathbf{a}}^e)-X({\mathbf{a}}).$$ Applying this inequality to our problem yields that $${\mathbb{V}\mathrm{ar}}{\left}[ \sum_{x \in B_r} v_k^2(x) {\right}] {\leqslant}\frac 1 2 \sum_{e \in {\mathbb{B}}} {\mathbb{E}}{\left}[ {\left}( \sum_{x \in B_r} \partial_{e} {\left}( v_k^2(x){\right}) {\right}) ^2 {\right}] .$$ For each edge $e$, we write $$\begin{aligned} v_{k}^{(e)}(x) & := |v_k({\mathbf{a}}^e,x)| + |v_k({\mathbf{a}},x)|. $$ where we stressed the dependency on the environment ${\mathbf{a}}$ or ${\mathbf{a}}^e$ in the notation. Note that the environment ${\mathbf{a}}^e$ has the same law as ${\mathbf{a}}$. Moreover, $${\left}|\partial_e {\left}( v_k^2(x) {\right}) {\right}|{\leqslant}v_{k}^{(e)}(x) \, {\left}|\partial_e v_k(x){\right}|.$$ We fix an exponent ${\alpha}> d$. By Hölder’s inequality, $$\begin{aligned} \sum_{x \in B_r} \partial_{e} {\left}( v_k^2(x){\right}) & {\leqslant}\sum_{x \in B_r} v_k^{(e)}(x) {\left}| \partial_e v_k(x) {\right}| \\ & {\leqslant}{\left}( \sum_{x \in B_r} w_e^{-{\alpha}}(2^k,x) |v_k^{(e)}(x)|^2 {\right})^\frac 1 2 {\left}( \sum_{x \in B_r} w_e^{\alpha}(2^k,x) |\partial_e v_k(x)|^2 {\right})^\frac 1 2.\end{aligned}$$ Therefore, $$\begin{gathered} \label{e.full.var} {\mathbb{V}\mathrm{ar}}{\left}[ \sum_{x \in B_r} v_k^2(x) {\right}] \\ {\leqslant}\sum_{e \in {\mathbb{B}}}{\mathbb{E}}{\left}[ {\left}(\sum_{x \in B_r} w_e^{-{\alpha}}(2^k,x) |v_k^{(e)}(x)|^2 {\right})^2{\right}]^\frac 1 2 {\mathbb{E}}{\left}[ {\left}(\sum_{x \in B_r} w_e^{\alpha}(2^k,x) |\partial_e v_k(x)|^2 {\right})^2{\right}]^\frac 1 2.\end{gathered}$$ By Hölder’s inequality, $${\left}(\sum_{x \in B_r} w_e^{-{\alpha}}(2^k,x) |v_k^{(e)}(x)|^2 {\right})^2{\leqslant}{\left}(\sum_{x \in B_r} w_e^{-{\alpha}}(2^k,x){\right}) {\left}( \sum_{x \in B_r} w_e^{-{\alpha}}(2^k,x) |v_k^{(e)}(x)|^4 {\right}).$$ Hence, by the definition of $v_k^{(e)}$ and , $$\label{e.bound2} {\mathbb{E}}{\left}[ {\left}(\sum_{x \in B_r} w_e^{-{\alpha}}(2^k,x) |v_k^{(e)}(x)|^2 {\right})^2{\right}]^\frac 1 2{\leqslant}C 2^{-k} {\left}( 1 + 2^{-\frac k 2} \operatorname{dist}(e,B_r) {\right}) ^{-{\alpha}}.$$ If we temporarily admit that there exists a constant $C(d,\Lambda,\alpha) < \infty$ such that for every $e \in {\mathbb{B}}$, $$\label{e.bound1} {\mathbb{E}}{\left}[ {\left}( \sum_{x \in B_r} w_e^{\alpha}(2^k,x) |\partial_e v_k(x)|^2 {\right}) ^2 {\right}]^\frac 1 2 {\leqslant}C 2^{-k {\left}( \frac d 2 + 1 {\right}) },$$ then we can combine the two bounds and into and thus complete the proof of the inequality . It therefore only remains to show that holds. We decompose the proof of this latter statement into two steps. *Step 3.* We show that there exists a constant $C < \infty$ such that for every $r,t {\geqslant}0$ and $e \in {\mathbb{B}}$, $$\label{e.bound1.ut} {\mathbb{E}}{\left}[ {\left}( \sum_{x \in B_r} w_e^{\alpha}(t,x) |\partial_e u(t,x)|^2 {\right}) ^2 {\right}]^\frac 1 2 {\leqslant}C (1 + t)^{- {\left}( \frac d 2 + 1 {\right}) }.$$ We recall from [@GNO Lemma 3] that there exists $C(d,\Lambda) < \infty$ such that $$\label{e.decomp.deu} |\partial_e u(t,x)| {\leqslant}C{\left}|\nabla G(t,x,e){\right}| + \int_0^t {\left}| \nabla G(t-s,x,e) g(s,e) {\right}| {\mathrm{d}}s ,$$ where $g(s,e)$ is a random variable that satisfies, for each $p < \infty$, $$\label{e.boundg} {\mathbb{E}}{\left}[ |g(s,e)|^p {\right}]^\frac 1 p {\leqslant}C {\mathbb{E}}{\left}[ |\nabla u(s,e)|^p {\right}]^\frac 1 p .$$ By Proposition \[p.gradient.green\] with $p = 1$, the first term on the right side of is handled without difficulty. It therefore suffices to show that $$\begin{gathered} \label{e.bound11} {\mathbb{E}}{\left}[ {\left}( \sum_{x \in {\mathbb{Z}}^d} w_e^{\alpha}(t,x) {\left}| \int_0^t {\left}| \nabla G(t-s,x,e) g(s,e) {\right}| \, {\mathrm{d}}s {\right}|^2 {\right})^2 {\right}]^\frac 1 2 \\ {\leqslant}C (1 + t)^{- {\left}( \frac d 2 + 1 {\right}) }.\end{gathered}$$ We first note for future reference that by Hölder’s inequality, , Lemma \[l.gradient\] and Proposition \[p.moment.bound\], there exists a constant $C(d,\Lambda) < \infty$ such that $$\label{e.boundg2} {\mathbb{E}}{\left}[ |g(s,e)|^4 {\right}]^\frac 1 4 {\leqslant}{\mathbb{E}}[|g(s,e)|^2]^\frac 1 2 \, {\mathbb{E}}[|g(s,e)|^6]^\frac 1 6 {\leqslant}C t^{- {\left}( \frac d 4 + 1 {\right}) }.$$ We now proceed to estimate the sum on the left side of . By the integral triangle inequality, $$\begin{gathered} {\left}(\sum_{x \in {\mathbb{Z}}^d} w_e^{\alpha}(t,x) {\left}| \int_0^t {\left}| \nabla G(t-s,x,e) g(s,e) {\right}| \, {\mathrm{d}}s {\right}|^2{\right})^\frac 1 2 \\ {\leqslant}\int_0^t {\left}( \sum_{x \in {\mathbb{Z}}^d} w_e^{\alpha}(t,x) {\left}| \nabla G(t-s,x,e) g(s,e) {\right}|^2 {\right})^\frac 1 2 \, {\mathrm{d}}s .\end{gathered}$$ Applying the triangle inequality again, we arrive at $$\begin{gathered} \label{e.bound.11.int} {\mathbb{E}}{\left}[ {\left}( \sum_{x \in {\mathbb{Z}}^d} w_e^{\alpha}(t,x) {\left}| \int_0^t {\left}| \nabla G(t-s,x,e) g(s,e) {\right}| \, {\mathrm{d}}s {\right}|^2 {\right})^2 {\right}]^\frac 1 {4} \\ {\leqslant}\int_0^t {\mathbb{E}}{\left}[ {\left}( \sum_{x \in {\mathbb{Z}}^d} w_e^{\alpha}(t,x) {\left}| \nabla G(t-s,x,e) g(s,e) {\right}|^2{\right})^{2} {\right}]^{\frac 1 {4}} \, {\mathrm{d}}s.\end{gathered}$$ We fix $p = p_0(d,\Lambda) > 1$ as given by Proposition \[p.gradient.green\], and denote by $q =\frac{p}{p-1}$ its conjugate exponent. We also give ourselves an exponent ${\beta}> 0$ to be specified later. By Hölder’s inequality, $$\begin{aligned} & \sum_{x \in {\mathbb{Z}}^d} w_e^{\alpha}(t,x) {\left}| \nabla G(t-s,x,e) g(s,e) {\right}|^2 \\& \quad {\leqslant}{\left}(\sum_{x \in {\mathbb{Z}}^d} w_e^{\alpha}(t,x) w_e^{{\beta}p}(t-s,x) {\left}| \nabla G(t-s,x,e) {\right}|^{2p} {\right})^\frac 1 {p} \\ & \qquad \times {\left}(\sum_{x \in {\mathbb{Z}}^d} w_e^{\alpha}(t,x) w_e^{-{\beta}q}(t-s,x) {\left}| g(s,e) {\right}|^{2q} {\right})^\frac 1 {q}\end{aligned}$$ By Proposition \[p.gradient.green\], the first term is deterministically bounded by a constant times $$(1 + |t-s|)^{- {\left}( d + 1 {\right}) + \frac d 2 \frac {1}{p}}.$$ Moreover, using the triangle inequality once more, we get $$\begin{aligned} & {\mathbb{E}}{\left}[ {\left}(\sum_{x \in {\mathbb{Z}}^d} w_e^{\alpha}(t,x) w_e^{-{\beta}q}(t-s,x) {\left}| g(s,e) {\right}|^{2q} {\right})^\frac 2 {q} {\right}]^\frac {q} 2 \\ & \quad {\leqslant}\sum_{x \in {\mathbb{Z}}^d} {\mathbb{E}}{\left}[ {\left}( w_e^{\alpha}(t,x) w_e^{-{\beta}q}(t-s,x) {\left}| g(s,e) {\right}|^{2q} {\right})^\frac 2 {q} {\right}]^\frac{q} 2 \\ & \quad {\leqslant}C {\left}(1 + |t-s|{\right})^{\frac d 2} \, {\left}( 1 + s {\right})^{-2q {\left}( \frac d 4 + 1 {\right})} ,\end{aligned}$$ where we chose ${\beta}< \infty$ sufficiently large and used and Proposition \[p.moment.bound\] in the last step. The quantity on the left of is therefore bounded by a constant times $$\begin{aligned} & \int_0^t (1 + |t-s|)^{- {\left}( \frac d 2 + \frac 1 2 {\right}) + \frac d 2 \frac {1}{2p}} \, (1 + |t-s|)^{\frac d 2 \frac 1 {2q}} \, {\left}( 1 + s {\right})^{-{\left}( \frac d 4 + 1 {\right})} \, {\mathrm{d}}s \\ & \qquad = \int_0^t (1 + |t-s|)^{- {\left}( \frac d 4 + \frac 1 2 {\right}) } \, (1 + s)^{- {\left}( \frac d 4 + 1 {\right})} \, {\mathrm{d}}s \\ & \qquad {\leqslant}C (1 + t)^{-{\left}(\frac d 4 + \frac 1 2{\right})}.\end{aligned}$$ This completes the proof of the inequality . *Step 4.* In order to complete the proof, there remains to show that implies . We use the representation  and Jensen’s and the triangle inequalities to bound $$\begin{aligned} & {\mathbb{E}}{\left}[ {\left}( \sum_{x \in B_r} w_e^{\alpha}(2^k,x) |\partial_e v_k(x)|^2 {\right}) ^2 {\right}]^\frac 1 2 \\ & \qquad {\leqslant}{\mathbb{E}}{\left}[ {\left}(E{\left}[\sum_{x \in B_r} w_e^{\alpha}(2^k,x) |\partial_e u(S_k,x)|^2 {\right}] {\right})^2 {\right}]^\frac 1 2 \\ & \qquad {\leqslant}E {\left}[ {\mathbb{E}}{\left}[ {\left}(\sum_{x \in B_r} w_e^{\alpha}(2^k,x) |\partial_e u(S_k,x)|^2 {\right})^2 {\right}]^\frac 1 2 {\right}] .\end{aligned}$$ By the result of the previous step and , this last term is bounded by a constant times $$ E {\left}[ (1+S_k)^{-{\left}( \frac d 2 + 1 {\right}) } {\right}] {\leqslant}C 2^{-k {\left}( \frac d 2 + 1 {\right}) }.$$ This completes the proof. We decompose the left side of into bias and variance. Using Theorem \[t.pgk\] and Proposition \[p.rest\], we can bound the bias by $$\begin{aligned} {\left}| {\left\langle}f, {{\mathcal L}}^{-1} f {\right\rangle}- \sum_{k = 0}^{n} \frac{1}{|B_{r_k}|} \sum_{x \in B_{r_k}} 2^k {\mathbb{E}}{\left}[ v_{k-1}(x) v_k(x) + v_k^2(x) {\right}] {\right}| & = {\left}| {\left\langle}f_n, {{\mathcal L}}^{-1} f_n {\right\rangle}{\right}| \\ & {\leqslant}C 2^{-\frac {dn}{2}}.\end{aligned}$$ By Proposition \[p.spat.mix\], we have $$\begin{aligned} {\mathbb{V}\mathrm{ar}}{\left}[ \frac{1}{|B_{r_k}|} \sum_{x \in B_{r_k}} 2^k {\left}( v_{k-1}(x) v_k(x) + v_k^2(x) {\right}) {\right}] & {\leqslant}C 2^{-\frac{dk}{2}} r_k^{-d} \\ & {\leqslant}C 2^{-dn-{\varepsilon}d k}.\end{aligned}$$ By the triangle inequality, we deduce that $$\begin{aligned} {\mathbb{V}\mathrm{ar}}{\left}[ \sum_{k = 0}^n \frac{1}{|B_{r_k}|} \sum_{x \in B_{r_k}} 2^k {\left}( v_{k-1}(x) v_k(x) + v_k^2(x) {\right}){\right}] & {\leqslant}C {\left}( \sum_{k = 0}^{n} 2^{-\frac 1 2 {\left}(dn + {\varepsilon}d k{\right})} {\right})^2 \\ & {\leqslant}C 2^{-dn}.\end{aligned}$$ This completes the proof. Complexity analysis {#s.compl} ------------------- In this section, we first complete the proof of Theorem \[t.main\] by evaluating the computational complexity of the method described there. We then prove the complexity lower bound announced in Proposition \[p.lower.bound\]. We recall that the method described in Theorem \[t.main\] is a family of approximations of the quantity of interest given, for each $n \in {\mathbb{N}}$, by $$\label{e.this.is.the.method} \sum_{k = 0}^n \frac{1}{|B_{r_k}|} \sum_{x \in B_{r_k}} 2^k {\left}( v_{k-1}(x) v_k(x) + v_k^2(x) {\right}),$$ where $v_k$ are functions on ${\mathbb{Z}}^d$ defined by , and $r_k$ is defined in and depends implicitly on $n$ and on an exponent ${\varepsilon}\in [0,\frac 1 2)$. (In view of Remark \[e.sqrt.log\], we will include the case ${\varepsilon}= 0$ in the complexity analysis.) We investigate the complexity of this method under the additional assumption $$\label{e.assumption.epsilon} {\varepsilon}< \frac{d-1}{2d}.$$ The question concerns the cost of computing the functions $(v_k, k \in \{0,\ldots n\})$ on $B_{r_k}$. These functions are solutions of elliptic equations. As we now argue, direct conjugate gradient methods, without any preconditioning, already yield the optimal scaling for the complexity of the method. Indeed, by standard Green function upper bounds (see e.g. [@dl-diff Proposition 3.6]), if we solve the equation for $v_k$ with Dirichlet boundary condition outside of $B_{r_k + C n 2^{\frac k 2}}$ for a sufficiently large constant $C < \infty$ (instead of the full space), then we can ensure that the solution thus obtained is $2^{-100dn}$ away from the true solution $v_k$. The problem of computing therefore boils down to the resolution of a series of elliptic problems, indexed by $k \in \{0,\ldots, n\}$, where for each $k$, we need to invert the operator $(2^{-k} + \nabla \cdot {\mathbf{a}}\nabla)$ on a domain of side length $r_k + C n 2^{\frac k 2}$. The condition number for this problem is of the order of $2^k$. Recall that using the conjugate gradient method, the computational cost of computing the solution to an elliptic problem with $N$ unknowns and condition number $\kappa$ at precision $\delta$ is of the order of $N \sqrt{\kappa} \log(\delta^{-1})$ (see e.g. [@quart Theorem 4.12]). Hence, for the problem of computing , the total complexity using conjugate gradient methods without preconditioning, and with a precision fixed at, say, $2^{-100dn}$, is bounded by a constant times $$\begin{aligned} n\sum_{k = 0}^n {\left}(r_k + C n 2^{\frac k 2}{\right})^d 2^{\frac k 2} & = n\sum_{k = 0}^n {\left}(2^{n-\frac k 2 + {\varepsilon}k} + C n 2^{\frac k 2}{\right})^d 2^{\frac k 2} \\ & {\leqslant}C n\sum_{k = 0}^n 2^{nd - {\left}( \frac {d-1} 2 - {\varepsilon}d {\right}) k} \\ & {\leqslant}C n2^{nd},\end{aligned}$$ where we used the assumption of in the last step. This completes the proof of Theorem \[t.main\]. One may argue that, under the specific i.i.d. assumption we make on the coefficient field, periodization would allow to simply get rid of any boundary layer. However, the idea of periodizing the medium is in general not suitable for solving the problem as stated at the onset of the introduction to this paper. Indeed, as soon as some short range of correlation is present in the medium, a periodization of the medium will create boundary defects that cause significant errors. This is discussed in more details in [@cemracs], see in particular Figure 9 there. In the method described in Theorem \[t.main\], the computational overhead caused by the necessity of boundary layers is an asymptotically small fraction of the total computational cost. More precisely, if we compare, say for ${\varepsilon}= 0$, the actual cost of the method with the fictitious cost one would face if boundary layers were not computed (i.e. where only values that enter into the formula are computed), we find that the difference is bounded by a constant times $$n^2 \sum_{k = 0}^n 2^{\frac k 2 + (d-1) {\left}( n - \frac k 2 {\right}) + \frac k 2} {\leqslant}C n^2 2^{(d-1)n} \times {\left}| \begin{array}{ll} 2^{\frac n 2} & \quad \text{if } d = 2, \\ n & \quad \text{if } d = 3, \\ 1 &\quad \text{if } d {\geqslant}4. \end{array} {\right}.$$ Hence, up to logarithmic corrections, the computational overhead of boundary layers is only of the order of $2^{(d-1)n}$ in dimension $d {\geqslant}3$. As we will argue shortly, no method can compute the homogenized coefficients at precision $2^{-\frac{nd}{2}}$ without evaluating the coefficients in a region of space of volume $2^{nd}$, and in this case a boundary layer of unit size already gives a computational overhead of $2^{(d-1)n}$, so this estimate is optimal. We now turn to proving the complexity lower bound announced in Proposition \[p.lower.bound\]. Since this proposition was stated relatively loosely, we first make it more precise. The number of operations of a given algorithm is a random variable; at precision ${\delta}> 0$, we call this random variable ${\mathcal}N_{\delta}$. The assumption we wish to contradict is that ${\delta}^2 {\mathcal}N_{\delta}$ tends to $0$ in probability, as ${\delta}$ tends to $0$. We will contradict it for a specific choice of the law of the random environment. However, it will be clear that the proof is very generic, and that the specific choice we make is only a matter of convenient notation and computation. Indeed, the only key point is to observe that “generically”, the Kullback-Leibler divergence (a.k.a. the relative entropy) responds quadratically to perturbations. For each $p \in [0,1]$, let ${\mathbb{P}}_p$ denote the product measure $${\mathbb{P}}_p = \bigotimes_{e \in {\mathbb{B}}}\, (p {\delta}_1 + (1-p) {\delta}_2),$$ where ${\delta}_x$ denotes the Dirac probability measure at $x \in {\mathbb{R}}$. We denote by ${\mathbb{E}}_p$ the associated expectation, and take $({\mathbf{a}}_e)_{e \in {\mathbb{B}}}$ to be the canonical random variable on ${\mathbb{R}}^{\mathbb{B}}$. Let ${{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(p)$ be the homogenized matrix associated with the law ${\mathbb{P}}_p$. By [@dl-diff], the mapping $p \mapsto {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(p)$ is $C^1$ on $[0,1]$, and it is not constant since ${{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(0) \neq {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(1)$. We will argue that in order for ${{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(p)$ to be determined within an error of ${\delta}$, the algorithm must query the values of at least $\simeq {\delta}^{-2}$ conductances, and therefore perform at least $\simeq {\delta}^{-2}$ operations. Since the law is of product form, we may assume without loss of generality that the algorithm queries ${\mathbf{a}}_{e_1}, {\mathbf{a}}_{e_2},\ldots$ in order, where $e_1,e_2,\ldots$ is an arbitrary sequence of distinct elements of ${\mathbb{B}}$. For every $p,q \in (0,1)$, every $n \in {\mathbb{N}}$ and every event $A \in \sigma {\left}( {\mathbf{a}}_{e_1},\ldots,{\mathbf{a}}_{e_n}, e \in {\mathcal}E {\right}) $, $$\begin{aligned} {\mathbb{P}}_q {\left}[ A {\right}] & = {\mathbb{E}}_p {\left}[ {\mathds{1}}_A \prod_{i = 1}^n {\left}(\frac q p {\mathds{1}}_{{\mathbf{a}}_{e_i} = 1} + \frac{1-q}{1-p} {\mathds{1}}_{{\mathbf{a}}_{e_i} = 2} {\right}){\right}] \notag\\ & = {\mathbb{E}}_p {\left}[ {\mathds{1}}_A \exp {\left}( {\widehat}{\mathrm{kl}}(q,p,n) {\right}) {\right}] , \label{e.hatkl}\end{aligned}$$ where we introduced the “empirical” Kullback-Leibler divergence $${\widehat}{\mathrm{kl}}(q,p,n) := \sum_{i = 1}^n \log{\left}(\frac q p {\mathds{1}}_{{\mathbf{a}}_{e_i} = 1} + \frac{1-q}{1-p} {\mathds{1}}_{{\mathbf{a}}_{e_i} = 2} {\right}).$$ The “true” Kullback-Leibler divergence between Bernoulli random variables with parameters $q$ and $p$ is the expectation of each of these summands, that is, $$\mathrm{kl}(q,p) := p \log \frac q p + (1-p) \log \frac{1-q}{1-p} {\geqslant}0.$$ Differentiating with respect to $q$, we verify that for each $\eta \in (0,1/2)$, there exists $C(\eta) < \infty$ such that $$\forall p,q \in (\eta,1-\eta), \quad \mathrm{kl}(q,p) {\leqslant}C(q-p)^2.$$ We also verify that $$\forall p,q \in (\eta,1-\eta), \quad \log{\left}(\frac q p {\mathds{1}}_{{\mathbf{a}}_{e_i} = 1} + \frac{1-q}{1-p} {\mathds{1}}_{{\mathbf{a}}_{e_i} = 2} {\right}) {\leqslant}C|q-p|.$$ In particular, uniformly over $p,q \in (\eta,1-\eta)$ and $n {\geqslant}1$, $$\label{e.control.hatkl} {\mathbb{E}}_p {\left}[ {\left}({\widehat}{\mathrm{kl}}(q,p,n) {\right})^2{\right}] {\leqslant}C n |q-p|^2 + C n^2 |q-p|^4.$$ If an algorithm can compute ${{{\widehat}{\mathbf{a}}_{\delta}}}$ in $o({\delta}^{-2})$ operations, we can define a deterministic $N_{\delta}= o({\delta}^{-2})$ such that the algorithm queries only the edges $e_1,\ldots,e_{N_{\delta}}$ with probability at least $1/2$. Denoting by ${\mathcal}E_{\delta}$ the event that only edges $e_1,\ldots,e_{N_{\delta}}$ are queried, we have $$\begin{aligned} {\mathbb{E}}_{q} {\left}[ ({{{\widehat}{\mathbf{a}}_{\delta}}}- {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(q))^2 {\right}] & {\geqslant}{\mathbb{E}}_{p} {\left}[ ({{{\widehat}{\mathbf{a}}_{\delta}}}- {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(q))^2 \, {\mathds{1}}_{{\mathcal}E_{\delta}} \, \exp {\left}( {\widehat}{\mathrm{kl}}(q,p,N_{\delta}){\right}) {\right}].\end{aligned}$$ We choose $q_{\delta}$ in such a way that $$\label{e.def.qde} {\delta}\ll q_{\delta}-p \ll N_{\delta}^{-1/2},$$ in the sense that the ratio of the right to the left side of both inequalities diverge to infinity as ${\delta}\to 0$. Let ${\mathcal}E'_{\delta}$ be the conjunction of ${\mathcal}E_{\delta}$ with the event that ${\widehat}{\mathrm{kl}}(q_{\delta},p,N_{\delta}) {\geqslant}-1$. Combining the three previous displays, we obtain, for ${\delta}> 0$ sufficiently small, $${\mathbb{E}}_{q_{\delta}} {\left}[ |{{{\widehat}{\mathbf{a}}_{\delta}}}- {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(q_{\delta})|^2 {\right}] {\geqslant}{\mathbb{E}}_p{\left}[ |{{{\widehat}{\mathbf{a}}_{\delta}}}- {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(q_{\delta})|^2 \, {\mathds{1}}_{{\mathcal}E'_{\delta}} {\right}] \, \exp(-1) , \qquad {\mathbb{P}}_p {\left}[ {\mathcal}E'_{\delta}{\right}] {\geqslant}1/4.$$ Since ${\mathbb{E}}_p {\left}[ {\left}| {{{\widehat}{\mathbf{a}}_{\delta}}}- {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(p) {\right}| ^2 {\right}] {\leqslant}{\delta}^2$, we deduce that $$\begin{aligned} |{{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(p) - {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(q_{\delta})| & {\leqslant}4 {\mathbb{E}}_p[|{{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(p) - {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(q_{\delta})| {\mathds{1}}_{{\mathcal}E'_{\delta}}] \\ & {\leqslant}4 {\mathbb{E}}_p {\left}[ |{{{\widehat}{\mathbf{a}}_{\delta}}}- {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(q_{\delta})|^2 {\mathds{1}}_{{\mathcal}E'_{\delta}}{\right}]^{1/2} + 4 {\mathbb{E}}_p {\left}[ |{{{\widehat}{\mathbf{a}}_{\delta}}}- {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(p)|^2 {\mathds{1}}_{{\mathcal}E'_{\delta}}{\right}]^{1/2} \\ & {\leqslant}8 {\mathbb{E}}_{q_{\delta}} {\left}[ |{{{\widehat}{\mathbf{a}}_{\delta}}}- {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(q_{\delta})|^2 {\right}]^{1/2} + 4{\delta}\\ & {\leqslant}12 {\delta}.\end{aligned}$$ Since our choice of $p \in (0,1)$ was arbitrary, and $p \mapsto {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}(p)$ is $C^1$ and non-constant over the interval $[0,1]$, this contradicts . Numerical tests {#s.numer} --------------- In this section, we report on numerical results for the method presented in Theorem \[t.main\]. For comparison, we also discuss the performance of the previously used method. The code was written in the Julia language, and can be downloaded as part of the source files of the arXiv posting of this paper. We focus on a two-dimensional example, with conductances taking the values $c_- = 1$ and $c_+ = 9$ with probability $1/2$ each. This is a rare example for which the homogenized matrix is know analytically: it is ${{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}= \sqrt{c_- c_+} \, \mathrm{Id} = 3 \, \mathrm{Id}$ (see e.g. [@Gloria Appendix A] or [@AKMBook Exercise 2.3]). In this case, we have ${\mathbb{E}}[{\mathbf{a}}] = 5 \, \mathrm{Id}$, and we thus expect that ${\widehat}{\sigma}_n^2$ converges to $2$. We fix $\xi = (1,0)$. The best numerical method should probably optimize the parameter $r_k$ in a dynamic way rather than rely on an a length scale obtained a priori from theoretical arguments. Here, I strived instead to avoid any fine-tuning and simply chose ${\varepsilon}= 0$ in the method described in Theorem \[t.eff\]. Whenever the operator $(\mu - \nabla \cdot {\mathbf{a}}\nabla)$ needs to be inverted, I chose the size of the boundary layer to be $$\label{e.boundary.size} 5 {\left}(1 \vee \mu^{-\frac 1 2}{\right}) {\left}(1 \vee \frac{\log {\left}(\mu^{-\frac 1 2}{\right})}{\log(2)}{\right}) ,$$ where we recall that $x \vee y := \max(x,y)$. The base-2 logarithm of the left side of (with ${\varepsilon}= 0$), as a function of $n$, is represented on Figure \[f.error.itresolv\]. The slope of the regression line there is $-1.0$, in full agreement with the theoretical prediction of Theorem \[t.main\]. Moreover, the intercept of the regression line is $1.0$, which indicates that the prefactor constant in can be chosen as $C \simeq 2.0$. ![ []{data-label="f.error.itresolv"}](l2_error_itresolv) The code was run on a laptop computer with $16$ Go of memory and a processor clocking at $2.40$ GHz. The base-2 logarithm of the time it takes to compute ${\widehat}{\sigma}^2_n$ on this machine, as a function of $n$, is reported on Figure \[f.cpu.itresolv\]. The slope of the regression line there is $1.9$, very closely agreeing with the prediction of Theorem \[t.main\]. The intercept is $-15.2$. It takes less than $3$ minutes to compute ${\widehat}{\sigma}^2_{12}$. ![ []{data-label="f.cpu.itresolv"}](cpu_time_itresolv) We compare these results with the following more classical approach. Recall the definition of the approximate corrector $\phi_\mu$ in . For each $n \in {\mathbb{N}}$, we denote $\mu_n := 2^{-n}$ and compute $2^n$ independent copies $\phi_{\mu_n}^{(1)}, \ldots, \phi_{\mu_n}^{(2^n)}$ of $\phi_{\mu_n}$ on a (two-dimensional) box of side length $2^\frac n 2$. (This involves the management of boundary layers of side length $C n 2^\frac n 2$.) The algorithm then outputs $${\widetilde}\sigma_n^2 := \frac 1 {k_n} \sum_{i = 1}^{k_n} \frac 1 {|B_{2^{\frac n 2}}|} \sum_{x \in B_{2^{\frac n 2}}} (\xi + \nabla \phi_{\mu_n}^{(i)}) \cdot {\mathbf{a}}(\xi + \nabla \phi_{\mu_n}^{(i)})) (x).$$ It was shown in [@GO1; @GO2] that there exists a constant $C < \infty$ such that, for every $n \in {\mathbb{N}}$, $$\label{e.l2.classical} {\mathbb{E}}{\left}[ {\left}(\xi \cdot {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}\xi - {\widetilde}\sigma_n^2 {\right})^2 {\right}] ^\frac 1 2 {\leqslant}C 2^{-n}.$$ (Recall that we assume $d = 2$ here.) We stick with the choice of for the size of the boundary layer. Figure \[f.error.classical\] displays the base-2 logarithm of the left side of , as a function of $n$. The slope of the regression line is $-0.9$, in very good agreement with the theoretical prediction. Figure \[f.cpu.classical\] reports on the cost of computing ${\widetilde}\sigma_n^2$, as a function of $n$. Two implementations were tried. The first uses no preconditioning, and in this case the slope of the regression line in Figure \[f.cpu.classical\] is about $2.7$, in close agreement with the theoretical prediction of $\frac 5 2$. I also implemented a version which uses an incomplete Cholesky factorization as a preconditioner. While this does reduce the number of iterations of the conjugate gradient method, this gain was not sufficient to offset the overhead caused by the preconditioning. The slope of the regression line on Figure \[f.cpu.classical\] in this case is $2.6$, but the intercept is $-12.1$, while the regression line for the non-preconditioned method has an intercept of $-15.0$. It takes more than $2$ minutes to compute ${\widetilde}\sigma_8^2$ using the non-preconditioned method, while computing the quantity ${\widehat}{\sigma}_8^2$ introduced in the present paper takes less than $1$ second and yields a better approximation. ![ []{data-label="f.error.classical"}](l2_error_classical) ![ []{data-label="f.cpu.classical"}](cpu_time_classical) The computational complexity reported here for the older method seems to be consistent with the numerical findings presented in [@Gloria]. Indeed, [@Gloria Figure 14] displays the computational cost of the method as a function of the error, and suggests that the computational time to reach precision $\delta$ scales like ${\delta}^{-2.6}$, in close agreement with the results reported on Figures \[f.error.classical\] and \[f.cpu.classical\][^1]. Discrete-time approach {#s.discrete} ---------------------- In this final section, we explore an alternative method based on a discrete-time dynamics asociated with the operator ${{\mathcal L}}$. Contrary to the approach presented in Theorem \[t.main\], the method explored here does not seem to generalize to the continuous-space setting. For discrete-space problems, it provides with an approach that does not involve the resolution of linear equations, and which I conjecture to have a computational cost for a precision of ${\delta}> 0$ of the order of $C {\delta}^{-2}$ (with no logarithmic correction) in dimension $d {\geqslant}3$. This would match exactly the complexity lower bound of Proposition \[p.lower.bound\]. In dimension $2$, I expect the number of operations to scale like ${\delta}^{-2} \log({\delta}^{-1})$ to yield a precision of ${\delta}> 0$. We set $$\pi({\mathbf{a}}) := \sum_{x \sim 0} {\mathbf{a}}_{0x},$$ and consider the operator $$\mathcal L_1 := -\pi^{-1} D \cdot {\mathbf{a}}D.$$ This operator is the generator of a discrete-time Markov chain on $\Omega$, and is self-adjoint on $L^2(\Omega, \pi({\mathbf{a}}) \, {\mathrm{d}}{\mathbb{P}}({\mathbf{a}}))$. We denote by ${\left\langle}\cdot, \cdot {\right\rangle}_\pi$ the scalar product in this $L^2$ space and set $$\label{e.def.g} g({\mathbf{a}}) := \pi^{-1} D \cdot {\mathbf{a}}\xi.$$ \[p.discrete.stuff\] We have $$\label{e.discrete.stuff} \xi \cdot {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}\xi = {\mathbb{E}}[\xi \cdot {\mathbf{a}}\xi] - {\left\langle}g, {{\mathcal L}}_1^{-1} g {\right\rangle}_\pi.$$ We recall that the rightmost term in is interpreted as $$\lim_{\mu \to 0} {\left\langle}g, (\mu + {{\mathcal L}}_1)^{-1} g {\right\rangle}_\pi,$$ and that $$ {\mathbb{E}}[\xi \cdot {\mathbf{a}}\xi] - \xi \cdot {{\overbracket[1pt][-1pt]{{\mathbf{a}}}}}\xi = {\mathbb{E}}[D\phi \cdot {\mathbf{a}}D \phi],$$ where $D \phi$ is the $L^2$ limit of $D \phi_\mu$, see . Let $\psi_\mu$ solve $${\left}( \mu + {{\mathcal L}}_1 {\right}) \psi_\mu = g,$$ that is, $${\left}( \mu - \pi^{-1} D \cdot {\mathbf{a}}D {\right}) \psi_\mu = \pi^{-1} D \cdot {\mathbf{a}}\xi.$$ By the same arguments allowing to justify , we can verify that $$D \psi_\mu \xrightarrow[\mu \to 0]{L^2(\Omega)} D \phi,$$ and that $\mu^\frac 1 2 \psi_\mu$ converges to $0$ in $L^2({\mathbb{P}})$. (This uses that $\pi$ is uniformly bounded from above and below.) In particular, $${\mathbb{E}}{\left}[ D \phi \cdot {\mathbf{a}}D \phi {\right}] = \lim_{\mu \to 0} {\mathbb{E}}{\left}[ D \psi_\mu \cdot {\mathbf{a}}D \psi_\mu {\right}] .$$ Moreover, by the weak formulation of the equation for $\psi_\mu$, we have $$\begin{aligned} {\left\langle}g, (\mu + {{\mathcal L}}_1)^{-1} g {\right\rangle}_\pi & = {\left\langle}g, \psi_\mu {\right\rangle}_\pi \\ & = -{\mathbb{E}}{\left}[ {\mathbf{a}}\xi \cdot D \psi_\mu {\right}] \\ & = {\mathbb{E}}{\left}[ D\psi_\mu \cdot {\mathbf{a}}D \psi_\mu {\right}] + \mu {\mathbb{E}}{\left}[ \pi \psi_\mu^2{\right}] \\ & \xrightarrow[\mu \to 0]{} {\mathbb{E}}[D \phi \cdot {\mathbf{a}}D \phi]. \qedhere\end{aligned}$$ ![ []{data-label="f.error.discrete"}](l2_error_discrete) ![ []{data-label="f.cpu.discrete"}](cpu_time_discrete) Note that the operator $\mathrm{Id} - {{\mathcal L}}_1$ is a contraction in $L^\infty(\Omega)$. It is also self-adjoint, and its spectrum is therefore contained in $[-1,1]$. The spectrum of ${{\mathcal L}}_1$ is thus contained in $[0,2]$. In order to avoid problems related to the periodicity of the underlying Markov chain, we modify slightly the formula  for computing ${\left\langle}g, {{\mathcal L}}_1^{-1} g {\right\rangle}_{\pi}$. We use instead that, for every $\lambda \in [0,2]$, $$\label{e.another.identity} \frac 1 {\lambda} = \frac 1 2\sum_{k = 0}^{+\infty}{\left}(1-\frac{\lambda}{2}{\right})^k ,$$ so that, setting $$\mathcal{P} := \mathrm{Id} - \frac 1 2 {{\mathcal L}}_1,$$ we have by spectral calculus that $$\begin{aligned} \notag {\left\langle}g, {{\mathcal L}}_1^{-1} g {\right\rangle}_\pi & = \frac 1 2 \sum_{k = 0}^{+\infty} {\left\langle}g, \mathcal P^k g {\right\rangle}_\pi \\ \label{e.self.mclT} & = \frac 1 2 \sum_{k = 0}^{+\infty} {\left}({\left\langle}\mathcal P^k g, \mathcal P^k g {\right\rangle}_\pi + {\left\langle}\mathcal P^k g, \mathcal P^{k+1} g {\right\rangle}_\pi{\right}).\end{aligned}$$ In view of the rigorous results in the continuous-time case, we expect that $$\label{e.decay.discrete} {\left}| {\left\langle}g, \mathcal P^k g {\right\rangle}_\pi {\right}| {\leqslant}C (1+k)^{-1 - \frac d 2}.$$ If we write $w(k,x) := ({\mathcal}P^k g)(\theta_x {\mathbf{a}})$, then the function $w$ solves the discrete-time parabolic equation $$\label{e.disc.parab} {\left}\{ \renewcommand{\arraystretch}{1.5} \begin{array}{rll} w(k+1,x) & = w(k,x) + \frac 1 {2\pi(x)} \nabla \cdot {\mathbf{a}}\nabla w(k,x) & \text{ for } (k,x) \in {\mathbb{N}}\times {{\mathbb{Z}^d}}, \\ w(0,x) & = \frac 1 {\pi(x)} (\nabla \cdot {\mathbf{a}}\xi)(x), & \text {for } x \in {{\mathbb{Z}^d}}, \end{array} {\right}.$$ where we write $\pi(x) := \pi(\theta_x {\mathbf{a}})$. In analogy with Theorem \[t.main\] and Remark \[e.sqrt.log\], we define, for each $L \in {\mathbb{N}}$, $$\label{e.def.Dl} D_L := \sum_{\ell = 0}^{L-1} \sum_{k = 2^\ell-1}^{2^{\ell+1}-2} \frac 1 {|B_{r(L,\ell)}|} \sum_{x \in B_{r(L,\ell)}} \pi(x) {\left}( w^2(k,x) + w(k,x) w(k+1,x) {\right}) ,$$ where $$r(L,\ell) := 2^{L - \frac \ell 2} ,$$ and we expect that $$\label{e.error.discrete} {\mathbb{E}}{\left}[ {\left}( {\left\langle}g, {{\mathcal L}}^{-1} g {\right\rangle}_\pi - D_L {\right})^2 {\right}]^\frac 1 2 {\leqslant}C L^\frac 1 2 2^{-\frac{dL}{2}}.$$ Moreover, the extraneous factor of $L^\frac 1 2$ is expected to be absent if we redefine $r(L,\ell)$ to be $2^{L - \frac \ell 2 + {\varepsilon}\ell}$, for ${\varepsilon}> 0$ arbitarily small. Finally, the computational cost of this scheme is given by $$\begin{aligned} \sum_{l = 0}^{L-1} \sum_{k = 2^{\ell} - 1}^{2^{\ell + 1} - 2} |B_{r(L,\ell)}| {\leqslant}C \sum_{l = 0}^{L-1} \sum_{k = 2^{\ell} - 1}^{2^{\ell + 1} - 2} 2^{d {\left}( L - \frac {\ell}{2} {\right}) } {\leqslant}C 2^{dL} \left| \begin{array}{ll} L & \qquad \text{if } d = 2, \\ 1 & \qquad \text{if } d {\geqslant}3. \end{array} {\right}.\end{aligned}$$ If we replace $r(L,\ell)$ by $2^{L - \frac \ell 2 + {\varepsilon}\ell}$, then the complexity of this method degrades in $d = 2$, no matter how small ${\varepsilon}> 0$ is chosen. This is in contrast with the method explored in the previous sections. On the other hand, in dimension $d {\geqslant}3$ and for ${\varepsilon}> 0$ sufficiently small, the complexity remains bounded by $C 2^{dL}$ and we that the error bound in holds with right-hand side replaced by $2^{-\frac {dL}{2}}$. We thus expect this scheme to match exactly the complexity lower bound given by Proposition \[p.lower.bound\] in dimensions $d {\geqslant}3$. We now report on numerical tests related to the quantity $D_L$ defined in . The code can be downloaded as part of the source files of the arXiv posting of this paper. We explore the same two-dimensional example as in Section \[s.discrete\], and use the same computer to run the code. Figure \[f.error.discrete\] displays the base-2 logarithm of the left side of , as a function of $L$, while Figure \[f.cpu.discrete\] displays the base-2 logarithm of the required CPU time to compute $D_L$. The numerical findings are very consistent with the heuristic predictions above. It takes less than one minute to compute $D_{12}$. The ratio of the slopes in Figures \[f.cpu.discrete\] and \[f.error.discrete\] is about $1.9$, in close agreement with the predicted theoretical complexity. The fact that we observe an exponent slightly below $2$ may be explained by the diminishing importance taken up by boundary layers as we move to larger values of $L$. ### Acknowledgments {#acknowledgments .unnumbered} I would like to thank Josselin Garnier for an inspiring talk which motivated me to revisit this problem, Tony Lelièvre for his helpful feedback, and Harmen Stoppels for his precious help with the Julia language. This work has been partially supported by the ANR Grant LSD (ANR-15-CE40-0020-03). [^1]: In order to estimate the slope of the regression line of the set of blue dots in [@Gloria Figure 14], I first drew the straight line going through the extremal blue dots, and observed that all the other dots are very close to this line. I then measured the coordinates of the extremal dots to be $(3.59, -0.62)$ and $(9.89,-3.06)$ respectively, which yields that the $dx/dy$ slope of the line is about $-2.58$. Moreover, the seven dots closest to the leftmost blue point are below this line, while the three dots closest to the rightmost blue point are above this line, and thus essentially any other reasonable choice of pair of of points to draw a line and measure the slope from would yield a larger absolute slope.

--- abstract: 'We present a state-of-the-art reanalysis of experimental results on Efimov resonances in the three-fermion system of $^6$Li. We discuss different definitions of the 3-body parameter (3BP) for Efimov states, and adopt a definition that excludes effects due to deviations from universal scaling for low-lying states. We develop a finite-temperature model for the case of three distinguishable fermions and apply it to the excited-state Efimov resonance to obtain the most accurate determination to date of the 3BP in an atomic three-body system. Our analysis of ground-state Efimov resonances in the same system yields values for the three-body parameter that are consistent with the excited-state result. Recent work has suggested that the reduced 3BP for atomic systems is a near-universal quantity, almost independent of the particular atom involved. However, the value of the 3BP obtained for $^6$Li is significantly ($\sim 20$%) different from that previously obtained from the excited-state resonance in Cs. The difference between these values poses a challenge for theory.' author: - 'Bo Huang(黄博)' - 'Kenneth M. O’Hara' - Rudolf Grimm - 'Jeremy M. Hutson' - 'Dmitry S. Petrov' title: 'The three-body parameter for Efimov states in lithium-6' --- [UTF8]{}[gbsn]{} Introduction ============ Ultracold atomic gases with resonant interactions provide experimental model systems to explore the universal physics of few-body quantum states [@Braaten2006uif; @Ferlaino2010fyo]. Efimov states, which are weakly bound three-body quantum states in systems of resonantly interacting particles, are a paradigm of this field. Efimov [@Efimov1970ela] showed that, when two bosons interact with an infinite scattering length, the corresponding three-particle system has an infinite number of three-body states just below threshold. For zero-range interactions, each successive Efimov state is larger than the previous one by a discrete length scaling factor, the ‘Efimov period’, which is 22.7 for a system of three identical bosons [@note:efactor] but can be widely different for other systems [@Dincao2006eto]. We refer to this universal scaling behavior as Efimov universality. The interactions between pairs of ultracold atoms may be varied by tuning an applied magnetic field in the vicinity of a zero-energy Feshbach resonance [@Chin2010fri]. The scattering length has a pole at resonance, corresponding to a 2-body bound state exactly at threshold. Signatures of Efimov states were first observed in an ultracold gas of cesium atoms [@Kraemer2006efe], and have since been found in many other systems, including other bosonic gases [@Zaccanti2009ooa; @Pollack2009uit; @Gross2009oou; @Gross2010nsi; @Wild2012mot; @Roy2013tot], three-component fermionic spin mixtures [@Ottenstein2008cso; @Huckans2009tbr; @Williams2009efa; @Nakajima2010nea], and mixtures of atomic species [@Barontini2009ooh; @Bloom2013tou; @Tung2014oog; @Pires2014ooe]. Moreover, extensions of the Efimov scenario to universal states of larger clusters [@Hammer2007upo; @vonStecher2009sou; @vonStecher2010wbc] have been demonstrated in experiments [@Ferlaino2009efu; @Pollack2009uit; @Zenesini2013rfb], highlighting the general nature of universal few-body physics. In addition to their discrete scaling property, Efimov states are characterized by a [*three-body parameter*]{} (3BP), which determines the position of the entire ladder of states. In the realm of nuclear systems, the 3BP is a non-universal quantity [@Braaten2006uif], determined by details of the short-range interaction. However, in atomic systems it has been found experimentally [@Berninger2011uot; @Roy2013tot] that the 3BP is nearly constant when expressed in terms of the van der Waals length $r_{\rm vdW}$ [@Chin2010fri], which quantifies the dispersion interaction between two neutral atoms. We refer to this feature of Efimov physics as van der Waals universality of the 3BP, and it has been the subject of a number of theoretical investigations [@Chin2011uso; @Wang2012oot; @Schmidt2012epb; @Sorensen2012epa; @Naidon2014moa; @Wang2014uvd]. Three-body recombination resonances occur when Efimov states cross the three-atom threshold as a function of magnetic field (and hence of scattering length) [@Esry1999rot; @Braaten2001tbr; @Ferlaino2011eri]. Recombination resonances due to Efimov ground states provide the most prominent observables in Efimov physics. Many experiments have focused on such features, including some that determined the 3BP [@Berninger2011uot; @Wild2012mot; @Roy2013tot]. In real atomic systems, however, finite-range corrections may significantly affect universal scaling, particularly for ratios involving the Efimov ground state [@Thogersen2008nbe; @Platter2009rct; @Naidon:2011; @Schmidt2012epb]. However, such corrections decrease substantially for higher Efimov states and are already very small for the first excited state. Excited-state resonances are therefore particularly interesting for precise measurements of the 3BP. Excited-state Efimov resonances occur at very large scattering lengths. They require extremely low temperatures for experimental observation, since the recombination peaks are less well defined when the de Broglie wavelengths are shorter than the scattering lengths [@Dincao2004lou; @Kraemer2006efe; @Rem2013lot]. Excited-state resonances have therefore been observed in only a very few experiments, carried out with $^6$Li [@Williams2009efa], with $^{133}$Cs [@Huang2014oot], and with mixtures of $^6$Li and $^{133}$Cs [@Tung2014oog; @Pires2014ooe]. Quantitative understanding of these resonances requires both very precise knowledge of the two-body scattering properties and an accurate theoretical description of finite-temperature effects. Ref. [@Huang2014oot] analyzed the excited-state Efimov resonance in cesium, using a highly accurate model of the two-body scattering [@Berninger2013frw] and a theoretical finite-temperature approach recently developed in Ref. [@Rem2013lot]. This study provided the most precise measurement of the Efimov period so far. In this Article, we re-analyze previous experimental results on the excited-state Efimov resonance in $^6$Li observed in Ref. [@Williams2009efa] and on the ground-state Efimov resonances observed in Refs. [@Ottenstein2008cso] and [@Huckans2009tbr]. In Sec. \[sec:3BP\], we discuss different definitions of the 3BP and how they are affected by deviations from ideal Efimov behavior. We adopt a definition that excludes effects due to deviations from universal scaling for low-lying states. In Sec. \[sec:threefermion\], we summarize the main properties of the three-fermion system. In Sec. \[sec:model\], we develop a new finite-temperature approach, which generalizes the theory introduced for the three-boson case in Ref. [@Rem2013lot] to the case of three distinguishable fermions. In Sec. \[sec:excited\], we present a refined analysis of the excited-state resonance observed in Ref. [@Williams2009efa]. This gives a high-precision value for the 3BP in $^6$Li, which deviates significantly from those found in other atomic systems. In Sec. \[sec:ground\], we re-analyze previous results on the ground-state Efimov resonance from Ref. [@Ottenstein2008cso] and investigate the possible influence of finite-range effects. In Sec. \[sec:discuss\], we discuss our findings in the context of other experiments in the field. The value of the 3BP found for $^6$Li is not well explained by current theories and presents a challenge for future theoretical work. Three-body parameter {#sec:3BP} ==================== For three identical bosons, ideal Efimov scaling leads to the simple relation $$\kappa^{(n+1)} = \kappa^{(n)} / 22.7 \, \label{eq:kappa_universal}$$ between the wavenumbers $\kappa^{(n)}$ that characterize the energies $E^{(n)}_{\rm res} = - (\hbar \kappa^{(n)})^2/m$ of successive Efimov states in the resonant limit $a \rightarrow \pm\infty$. Here $m$ is the atomic mass and $n$ is an integer quantum number. The corresponding relation between the the scattering lengths at the positions of successive recombination resonances is $$a_-^{(n+1)} = 22.7 \times a_-^{(n)} \, . \label{eq:a_universal}$$ The universal relation $$a_-^{(n)} = -1.508 / \kappa^{(n)} \, \label{eq:aminus}$$ connects a resonance position with the corresponding bound-state wavenumber. In the ideal case, knowledge of any of the above quantities $\kappa^{(n)}$ or $a_-^{(n)}$ fixes the infinite series and thus provides a proper representation of the 3BP. In a real system, where the interaction has a finite range, the Efimov spectrum is bounded from below. We refer to the lowest state as the Efimov ground state with $n=0$ and to the corresponding resonance at $a_-^{(0)}$ as the ground-state Efimov resonance. Eqs. (\[eq:kappa\_universal\]) and (\[eq:a\_universal\]) then represent approximations, subject to finite-range effects. One way to understand the Efimov effect is through a treatment in hyperspherical coordinates. Efimov states may be viewed as supported by an effective adiabatic potential that is a function of the hyperradius $R$. For a zero-range two-body potential with large scattering length $a$, this potential is attractive and proportional to $R^{-2}$ for $R\lesssim|a|$ [@Macek:1986] and supports an infinite number of bound states as $a\rightarrow\pm\infty$. For potential curves with long-range van der Waals tails, however, Wang [*et al.*]{} [@Wang2012oot] have shown that the effective adiabatic potential reaches a minimum and then rises to a wall or barrier near $R=2r_{\rm vdW}$. The position of the minimum and wall depend to some extent on the details of the two-body potential and the number of bound states it supports, but become near-universal as the number of 2-body bound states increases. The presence of the minimum and wall have two principal effects on the physics. First, the deviation of the effective potential from $R^{-2}$ behavior results in deviations from ideal Efimov scaling for the lowest-lying states. Secondly, the boundary condition provided by the wall defines the position of the entire ladder of Efimov states, and its nearly universal position is responsible for the near-universality of the 3BP. However, it should be noted that the wall itself is a product of physics around $2r_{\rm vdW}$, so that variations in the physics in this region can produce deviations from universality of the 3BP even in the limit $a\rightarrow\pm\infty$. Theoretical investigations [@Thogersen2008nbe; @Platter2009rct; @Wang2012oot; @Schmidt2012epb] have shown that the Efimov ground state may be subject to considerable modifications. For $n=0$ this may change the factor $22.7$ in Eqs. (\[eq:a\_universal\]) - (\[eq:aminus\]) by up to 25%. The relation (\[eq:aminus\]) is subject to similar modifications [@Schmidt2012epb; @Wang2012oot]. The recent experiment on the excited-state resonance in Cs [@Huang2014oot] and a related theoretical investigation [@Wang2014uvd] also hint at deviations from the ideal scaling. The deviations from universal scaling for low-lying Efimov states raise the question of the best representation of the 3BP. Definitions based on the limit $n \rightarrow \infty$ remove effects of this type from the 3BP. Accordingly, we adopt the definition [@Braaten2006uif] $$\kappa_* = \lim_{n \to \infty} \left( 22.7^n \kappa^{(n)} \right) \, , \label{eq:idealkappa}$$ and by analogy $$a_-^* = \lim_{n \to \infty} \frac{a_-^{(n)}}{22.7^n}\, . \label{eq:ideala}$$ The position of the ground-state Efimov resonance, $a_-^{(0)}$, is commonly used as a 3BP. However, it gives a somewhat crude approximation to $a_-^*$, and in some cases may deviate from it by as much as 25%. The quantities $a_-^{(1)}/22.7$ and $22.7\kappa^{(1)}$, obtained from the excited-state resonance, provide much better approximations to $a_-^*$ and $\kappa_*$, with corrections of only about 1% due to deviations from universal scaling [@Schmidt2012epb]; these corrections are comparable to the other uncertainties in current experiments. Efimov states are also characterized by a decay parameter $\eta_*$ [@Braaten2006uif], which describes their decay to lower-lying atom-dimer combinations. This parameter is usually considered to be a constant for a particular Efimov state, but may vary if the available product states change significantly. The resulting field dependence may be important when interpreting measurements that extend over wide ranges of field [@Wenz2009uti]. Efimov states in a three-component Fermi gas {#sec:threefermion} ============================================ Three-fermion system -------------------- Efimov states in a three-component gas of fermions [@Braaten2009tbr] exhibit the same discrete scaling behavior as in the three-boson case, provided that all three scattering lengths involved are large ($|a_{12}|, |a_{13}|, |a_{23}| \gg r_{\rm vdW}$). In particular, if the masses of the three components are equal, the Efimov period is given by the same discrete scaling factor of $22.7$ [@note:efactor]. The special case of three equal scattering lengths ($a_{12} = a_{13} = a_{23}$) is formally equivalent to the situation for three identical bosons. A gas of $^6$Li atoms prepared in a mixture of the lowest three spin states allows a realization of large scattering lengths by Feshbach tuning [@Chin2010fri]. However, the applied magnetic field offers only one degree of freedom for tuning, thus limiting the experimentally accessible combinations of scattering lengths. Arbitrary combinations and, in particular, the situation of three equal scattering lengths thus remain hypothetical cases, but universal theory allows them to be linked to the combinations that exist in real systems. In real experiments on a three-fermion system, Efimov resonances appear at certain combinations of large scattering lengths $a_{12}, a_{13}, a_{23}$, where typically $a_{12} \neq a_{13} \neq a_{23}$. A generalization of the Skorniakov–Ter-Martirosian (STM) equations [@Braaten2009tbr] can be employed to determine the 3BP from these generally unequal values. In the wavenumber representation, $\kappa_*$ then refers to the hypothetical case of three infinite scattering lengths, while $a_-^*$ refers to a hypothetical system with three equal scattering lengths. The STM approach is based on the zero-range approximation and therefore does not take account of finite-range corrections, which are significant at relatively small scattering lengths. It can thus be expected to provide an excellent approximation for excited Efimov states ($n \ge 1$), but it may be subject to significant corrections if applied to the Efimov ground state ($n=0$). Three-body recombination ------------------------ In a three-component Fermi gas, the dominant contribution to three-body losses results from triples of three non-identical particles. All other combinations involve pairs of identical fermions, which leads to a strong Pauli suppression of losses at ultralow temperatures [@Esry2001tlf]. Three-body losses can be modeled by the simple rate equation $$\frac{d}{dt} n_i = - L_3 n_1 n_2 n_3 \, ,$$ where the $n_i$ represent the number densities of the three different spin states. After a spatial integration of losses over the density profile of the trapped cloud, the loss rate coefficient $L_3$ can be experimentally determined by fitting the time-dependent decay of the total atom numbers [@Ottenstein2008cso; @Huckans2009tbr; @Williams2009efa]. Efimov states show up as distinct loss resonances [@Ferlaino2011eri] when they couple to the three-atom threshold. Lithium-6 --------- The situation of a three-component Fermi gas of $^6$Li is unique because of overlapping Feshbach resonances in all three combinations of the lowest three spin states together with large negative background scattering lengths. The two-body scattering properties are known to an extraordinarily high level of precision thanks to the characterization in Ref. [@Zurn2013pco], which significantly improved the conversion from magnetic field to scattering lengths compared to previous work [@Bartenstein2005pdo]. In the resonance region between 832 and 900G, all three scattering lengths are very large and negative, with absolute values of a few thousand times the Bohr radius $a_0$ that vastly exceed $r_{\rm vdW} = 31.26\,a_0$. In this extreme regime, an excited Efimov state exists [@Williams2009efa]. This trimer state crosses the three-atom threshold near 900G and leads to a strong enhancement of three-body recombination. The corresponding Efimov ground state exists over a much wider range of magnetic fields, but it does not cross threshold at currently accessible magnetic fields and thus does not lead to an observable recombination resonance. In the magnetic-field region below the zero crossings of the Feshbach resonances, the three scattering lengths are moderately large and negative, so that an Efimov ground state exists. This state crosses the three-atom threshold near 130G and near 500G [@Ottenstein2008cso; @Huckans2009tbr], leading to two observable Efimov resonances. In this low-field region, the scattering lengths never reach large enough values for an excited Efimov state to exist. The effective range ------------------- One way to quantify the finite (non-zero) range of an atomic interaction is through the effective range [@Bethe:1949; @Hinckelmann:1971], which characterizes the leading term in the energy-dependence of the scattering length. The effective range behaves very differently in the vicinity of Feshbach resonances of different types [@Blackley:eff-range:2014]. For a resonance that is strongly entrance-channel-dominated [@Chin2010fri], the effective range takes a small and fairly constant value close to $2.8r_{\rm vdW}$ at fields near the resonance pole [@Gao:QDT:1998]. By contrast, for resonances that are closed-channel-dominated, the effective range is much larger and varies very fast with magnetic field [@Blackley:eff-range:2014]. The Feshbach resonances used in the present work for $^6$Li are all strongly entrance-channel-dominated [@Chin2010fri], so that deviations from Efimov scaling due to finite-range effects are expected to be relatively small in comparison to some of the other atomic systems that have been studied. Finite-temperature theoretical approach {#sec:model} ======================================= A convenient way of modeling three-body losses in Efimovian systems is provided by the $S$-matrix formalism based on Efimov’s radial law [@Efimov1979lep], which is elaborated in Refs. [@Braaten2006uif; @Braaten2008tbr; @Rem2013lot] for the case of three identical bosons. Its generalization to three distinguishable atoms with different scattering lengths is straightforward and we will present only a brief derivation. This is a zero-range theory for which $\kappa_*$ and $\eta_*$ are external parameters. First, one introduces three-atom scattering channels describing the motion of free atoms at large distances. By contrast, all atom-dimer channels are substituted by the single Efimov channel defined in the scaling region $r_{\rm vdW}\ll R\ll {\rm min}\{1/k,|a_{12}|,|a_{23}|,|a_{13}|\}$, where $k=\sqrt{mE}/\hbar$, $E$ is the energy in the center of mass reference frame, $R$ is the hyperradius, and we consider the case of negative scattering lengths. The reason for this substitution is that when $r_{\rm vdW}\ll {\rm min}\{1/k,|a_{12}|,|a_{23}|,|a_{13}|\}$ this channel becomes essentially the only one that can conduct three atoms from large distances to the recombination region of size $\sim r_{\rm vdW}$. One can think of this short-distance channel and the long-distance three-atom channels as being fused together at intermediate distances where the transmission, reflection, and mixing of the channels takes place. We can then introduce a unitary matrix $s_{ij}$, which defines the amplitude of the outgoing wave in channel $j$ if the incoming wave is injected in channel $i$. The terms “incoming” and “outgoing” are defined with respect to the fusion region. In particular, the incoming Efimov wave $R^{-2+is_0}$ actually propagates towards larger distances and $R^{-2-is_0}$ describes the outgoing one. Here $s_0 \approx 1.00624$ is a constant and the ideal Efimov period of 22.7 is $e^{\pi/s_0}$. The simple fact that the matrix $s_{ij}$ is unitary turns out to be very useful in describing the scaling properties of Efimovian systems [@Braaten2006uif]. We point out that $s_{ij}$ does not depend on the 3BP $\kappa_*$ or the decay parameter $\eta_*$. These quantities come into play when one fixes the relative phase and amplitude of the incoming and outgoing Efimov waves, $$\label{ThreeBodyCond} R^2\Psi\,{\propto}\,(R/R_0)^{is_0}{-}e^{2\eta_*}(R/R_0)^{-is_0},$$ where $R_0$ is a three-body length related to $\kappa_*$ by $$(\kappa_* R_0/2)^{2is_0}=-\Gamma(is_0)/\Gamma(-is_0)$$ and $\Gamma$ is the gamma function. One can imagine that Efimov waves are reflected at small hyperradii by a lossy mirror with reflection/loss properties given by Eq. (\[ThreeBodyCond\]). The three-body problem is then analogous to a Fabry-Perot interferometer with the other mirror quantified by the matrix $s_{ij}$. This picture gives a convenient way of understanding and describing three-body loss peaks as resonances of the Fabry-Perot cavity. In particular, if we denote the Efimov channel by subscript 1, the loss probability for a given incoming channel $i\neq 1$ is [@Braaten2008tbr] $$P_i=\frac{(1-e^{-4\eta_*})|s_{i1}|^2}{|1+(kR_0)^{-2is_0}e^{-2\eta_*}s_{11}|^2},$$ where the denominator accounts for multiple reflections “inside” the resonator. The total loss rate constant for three distinguishable fermions is obtained by using unitarity ($\sum_{i=1}^\infty|s_{1i}|^2\,{=}\,1$) and averaging over the Boltzmann distribution, $$\begin{aligned} \label{L3result} L_3&=&\frac{24\sqrt{3}\pi^2\hbar (1-e^{-4\eta_*})}{mk_{\rm th}^{6}} \nonumber\\ &\times& \int_0^\infty\frac{(1-|s_{11}|^2)e^{-k^2/k_{\rm th}^2}}{|1+(kR_0)^{-2is_0}e^{-2\eta_*}s_{11}|^2}k\,dk,\end{aligned}$$ where $k_{\rm th}=\sqrt{mk_{\rm B}T}/\hbar$. Equation (\[L3result\]) differs from the bosonic result of Ref. [@Rem2013lot] only by the factor $1/3$, which is due to the bosonic bunching effect and different ways of counting triples in the two cases. A more profound change is hidden in the quantity $s_{11}$, which, in contrast to the case of identical bosons, now depends on three dimensionless numbers $ka_{12}$, $ka_{23}$, and $ka_{13}$. In order to determine $s_{11}$ we look for the three-body wave function that behaves as $A (kR)^{-2+is_0}+ B (kR)^{-2-is_0}$ in the scaling region and contains only outgoing waves at large distances. By definition, $s_{11}=B/A$. We solve this problem by using the STM equations in a very close analogy to the bosonic case (see Supplemental Material of [@Rem2013lot]). For distinguishable atoms with generally different scattering lengths we end up with three coupled STM equations (see Ref. [@PetrovLesHouches2010] for details of the method). In practice, we use the known dependence of $a_{ij}$ on $B$ [@Zurn2013pco] and tabulate $s_{11}$ as a function of $k$ and $B$. This then allows fast integration of Eq. (\[L3result\]) for any desired values of $T$, $\kappa_*$, and $\eta_*$. Excited-state Efimov resonance {#sec:excited} ============================== In Ref. [@Williams2009efa], the excited-state Efimov resonance was observed in the high-field region of $^6$Li. In Figure \[fig.2nd\] we show the experimental results for the three-body loss coefficient $L_3$ as a function of the magnetic field, measured for two different temperatures of about 30nK (set A) and 180nK (set B). In this section we reanalyze these results, taking account of finite-temperature effects using the theory described in Sec. \[sec:model\], in order to obtain a refined estimate of the 3BP for $^6$Li. ![(Color online). Finite-temperature fits to the excited-state Efimov resonance. The experimental results obtained for $L_3$ in Ref. [@Williams2009efa] for two different temperatures are plotted as filled blue squares (set A, 30nK) and filled red circles (set B, 180nK). The amplitude scaling parameter $\lambda$ is of order 1, see text. The corresponding solid lines are the fixed-temperature fits to both data sets, carried out on a linear scale (see first and fifth row in Table \[tab.2nd\]). The black dashed curve is calculated for the zero-temperature limit using the parameters from the fixed-temperature fit to the 30 nK set.[]{data-label="fig.2nd"}](2nd.eps){width="8.6cm"} The two free parameters in the temperature-dependent theory of Sec. \[sec:model\] are the 3BP $\kappa_*$ and the decay parameter $\eta_*$. In addition, experimental uncertainties in the number density calibration may considerably affect the amplitude of the observed losses. Such uncertainties may result from the atom number calibration, from the limited knowledge of the trap frequencies, and from errors in the temperature measurements. It is therefore useful to introduce an additional scaling parameter $\lambda$ for the amplitude of the observed losses [@Huang2014oot]. Under realistic experimental conditions, variations of up to a factor of two from the ideal value $\lambda = 1$ are plausible. To analyze the data we follow several different strategies, similar to those applied to the three-boson case of cesium [@Huang2014oot]. First, we fix the temperature $T$ to the measured values $T_{\rm meas} = 30$nK (set A) and 180nK (set B), and we perform a fit with $\kappa_*$, $\eta_*$, and $\lambda$ as the free parameters. Alternatively, we allow for a variable temperature $T$, and instead we fix $\lambda = (T/T_{\rm meas})^{-3}$ [@note:unstable] to take account of the resulting change in the volume of the harmonically trapped gas. Moreover, we fit the data sets A and B on either a linear or a logarithmic scale, which puts different weights on the different regions. In this way, we obtain four different fits for each data set. We note that the experimental results of Ref. [@Williams2009efa] indicated that the effect of heating during the decay of the trapped sample remained very small, so that this effect can safely be neglected in our fit analysis. Set $T$ (nK) $\kappa_* a_0$ $\eta_*$ $\lambda$ ------- ---------- ---------------- ----------- ------------ A 30$^a$ 0.006808(36) 0.032(5) 0.546(27) A log 30$^a$ 0.006744(91) 0.048(15) 0.498(107) A 35(5) 0.006774(39) 0.029(5) 0.644$^T$ A log 36(2) 0.006689(97) 0.042(14) 0.593$^T$ B 180$^a$ 0.006839(80) 0.088(15) 0.258(16) B log 180$^a$ 0.006665(130) 0.067(16) 0.270(49) B 237(5) 0.006736(84) 0.072(15) 0.438$^T$ B log 218(10) 0.006624(118) 0.034(8) 0.562$^T$ : Results of fits for the excited-state Efimov resonance, obtained from the two sets of measurements presented in Fig. \[fig.2nd\]. The fits using a logarithmic $L_3$ scale are indicated with ‘log’ in the column ‘Set’. The superscript $^a$ means that corresponding parameter is kept fixed. The superscript $^T$ indicates that the corresponding parameter is calculated from the fitted values for $T$.[]{data-label="tab.2nd"} ![(Color online). Fitted values for $\kappa_*$ corresponding to the third column in Table \[tab.2nd\]. The dashed line indicates the final result $\kappa_* = 0.00678(6)\,a_0^{-1}$, as obtained from a weighted average of the four data points of the low-temperature data set A (filled blue squares), and the gray-shaded region shows the corresponding uncertainty. The high-temperature data set B (filled red circles) is not used for deriving the final value, but within the uncertainties the values are fully consistent with the result from data set A.[]{data-label="fig.kappas"}](kappas.eps){width="8.6cm"} Table \[tab.2nd\] summarizes the results of our fits for both data sets, and Fig. \[fig.kappas\] shows the values obtained for the 3BP (from the third column of the Table). The comparison between the four different fits for each data set provides information on the robustness of the fits and possible systematic effects beyond simple statistical uncertainties. In our results from the low-temperature set A, the errors on $\kappa_*$ from individual fits range between 0.5% (for linear fits) and 1.5% (for logarithmic fits). Within the error bars no significant systematic deviations appear between the central values obtained from the different fits, which shows that the errors are consistent with purely statistical uncertainties. From the low-temperature data set A, by calculating a weighted average [@note:weights] over all four fitted values, we obtain the final value $\kappa_* = 0.00678(6)\,a_0^{-1}$, where the uncertainty includes both the weighted errors of the four individual fits and the standard deviation of the four slightly different values. The result for $\kappa_*$ and the error are shown by the dashed horizontal line and the gray-shaded region in Fig. \[fig.kappas\]. Note that all the statistical uncertainties specified in this work correspond to one standard deviation. The higher-temperature data set B yields similar results, but with somewhat larger uncertainties. Again, there are no systematic deviations between the four different fit strategies applied. Here the final result for the 3BP, $\kappa_* = 0.00674(13)\,a_0^{-1}$, is fully consistent with the result obtained at lower temperatures, with an uncertainty about two times larger than for set A. This confirms that temperature-induced shifts of the resonance are properly taken into account in our theoretical approach. The original data analysis in Ref. [@Williams2009efa] yielded $\kappa_* = 0.0069(2)\,a_0^{-1}$, remarkably close to the present result but with a quoted error about three times larger. However, the present work reveals two systematic shifts, which in the previous work partially canceled each other. The updated values of the scattering lengths [@Zurn2013pco] shift the value of $\kappa_*$ up by about 3%, while residual finite-temperature effects shift the value down by about 5% [@note:residualT]. A further contribution to our error budget comes from the uncertainty in the mapping from magnetic field to scattering length. The scattering lengths used here were obtained from the potential curves of Ref. [@Zurn2013pco], which were fitted to highly precise measurements of binding energies of $^6$Li$_2$ in the resonant region, together with measurements of collision properties. The fits have recently been extended to include binding energies for $^7$Li$_2$, with an explicit mass-dependence of the potential curves [@Julienne:Li67:2014]. In order to establish the uncertainties in the scattering lengths at the magnetic field of the excited-state resonance, we have repeated the fits of Ref. [@Zurn2013pco] and calculated explicit statistical uncertainties in the three scattering lengths $a_{12}$, $a_{13}$ and $a_{23}$ at 891G, using the procedure of Ref. [@LeRoy:1998]. The resulting contribution to the uncertainty in $\kappa_*$ is about 0.1%. We have also estimated the nonstatistical uncertainties in the scattering lengths by repeating the fits with the experimental binding energies set to the values at the upper and lower limits of their systematic errors, producing a further uncertainty of 0.07%. The uncertainty of 0.1G in the magnetic-field calibration of Ref. [@Williams2009efa] corresponds to a further error of 0.07%. All these error sources are thus negligibly small compared to the fitting errors described above. Based on the results of our fits for $\kappa_*$ and $\eta_*$, we can calculate the recombination rate coefficient $L_3$ in the zero-temperature limit. The resulting curve is shown as a dashed line in Fig. \[fig.2nd\]. The peak occurs at 891G, which marks the point where the Efimov state crosses the three-atom threshold. Here the three scattering lengths are $a_{12} = -8671(38)\,a_0$, $a_{13} = -2866(3)\,a_0$, and $a_{23} = -5728(16)\,a_0$. Ground-state Efimov resonances {#sec:ground} ============================== References [@Ottenstein2008cso; @Huckans2009tbr] reported the observation of the two ground-state Efimov resonances in the low-field region of $^6$Li near 130G and near 500G. The $L_3$ results of Ref. [@Ottenstein2008cso] have been further analyzed in Refs. [@Braaten2009tbr; @Wenz2009uti; @Naidon:2011], using different models within the zero-temperature approximation. Ref. [@Braaten2009tbr] treated the three different scattering lengths within the approach of the generalized STM equations, which is exact within the zero-range limit, while Ref. [@Wenz2009uti] introduced the approximation of an ‘effective scattering length’. As an important improvement, Ref. [@Wenz2009uti] introduced a magnetic-field dependence in the decay parameter $\eta_*$, determined by the binding energies of the different target molecular states. The latter has proved very useful to describe the different widths of the narrower Efimov resonance near 130G and the wider Efimov resonance near 500G. Ref. [@Naidon:2011] considered the effects of finite-range corrections and of a two-channel model of the atom-atom scattering. Our new analysis of the results of Ref. [@Ottenstein2008cso] is based on the generalized STM approach in combination with the magnetic-field-dependent decay parameter $\eta_*$. We also use the updated scattering length values from Ref. [@Zurn2013pco], instead of the ones from Ref. [@Bartenstein2005pdo], but this has negligible effect on the value resulting for the 3BP in the low-field region. All our fits assume a temperature of 215nK [@Ottenstein2008cso], but we find that including finite-temperature effects makes a negligible difference for the ground-state resonances, in contrast to the excited-state case. Our theoretical model to calculate $L_3$ from the three different scattering lengths relies on the zero-range approximation, and is applicable only for $|a_{12}|, |a_{13}|, |a_{23}| \gg r_{\rm vdW}$. However, at the resonance positions of 130G and 500G, the smallest of the three scattering lengths, $|a_{12}|$, exhibits rather small values of $\sim4\,r_{\rm vdW}$ and $\sim3\,r_{\rm vdW}$, respectively. This makes the analysis quite vulnerable to finite-range effects, and the extracted values for $\kappa_*$ can be expected to provide only an approximation to the limiting case of Eq. (\[eq:idealkappa\]). To deal with this nonideal situation, we introduce a lower cutoff scattering length $a_{\rm min}$ and restrict our fit to the region where $|a_{12}|, |a_{13}|, |a_{23}| > a_{\rm min}$. The dependence of the resulting values for $\kappa_*$ on $a_{\rm min}$ then gives an indication of the sensitivity to finite-range and model-dependent corrections. ![(Color online). Fits to the ground-state Efimov resonances. All three panels show the same experimental data on the loss rate coefficient $L_3$ from Ref. [@Ottenstein2008cso], where the filled squares, filled circles, and filled triangles refer to losses measured in the lowest three spin states. The theoretical curves represent our fits to the data on a linear scale. The solid lines indicate the region used for the fit with all three scattering lengths being larger than the cutoff value $a_{\rm min}$. The dashed lines extrapolate the theory to regions not used for the fit.[]{data-label="fig.1stfit"}](1stfit.eps){width="8cm"} Figure \[fig.1stfit\] shows three different fits to the same data points, differing in the cutoff scattering length, $a_{\rm min}/r_{\rm vdW} = 2, 4$, and $6$. The fits are applied globally to both resonances, appearing near 130G and near 500G. The three free parameters of the fit are $\kappa_*$, the amplitude scaling factor $\lambda$ (see Sec. \[sec:excited\]), and the parameter $A$ defined in Ref. [@Wenz2009uti], from which the magnetic-field-dependent $\eta_*$ can be calculated. The lines in Fig. \[fig.1stfit\] represent the behavior within the fit region (solid lines) and extrapolated beyond that region (dashed lines). We find that the fit with $a_{\rm min}/r_{\rm vdW} = 4$ captures both resonances and the overall behavior quite well. Here we obtain $\kappa_* = \, 0.00645(3) \, a_0^{-1}$ (linear scale) and $0.00641(3) \, a_0^{-1}$ (logarithmic scale). For the amplitude scaling factor the fits yield the plausible values $\lambda = 1.65(5)$ (linear) and $1.68(7)$ (logarithmic). From the corresponding values of $A$ we obtain the values $\eta_*=0.0814(3)$ (linear) and $0.0745(3)$ (logarithmic) for the decay parameter at the lower-field (sharper) resonance, which the fit locates at 132G. In contrast to the fit with $a_{\rm min}/r_{\rm vdW} = 4$, the two other fits in Fig. \[fig.1stfit\] are problematic. The fit for $a_{\rm min}/r_{\rm vdW} = 2$ puts some weight on regions where the applicability of zero-range theory is highly questionable, while the fit for $a_{\rm min}/r_{\rm vdW} = 6$ excludes the centers of the two resonances, which provide the most sensitive information on the Efimov resonance positions. ![(Color online). Dependence of the fitted values for $\kappa_*$ on the cutoff scattering length $a_{\rm min}$ for the ground-state Efimov resonance. The blue filled squares and red filled circles refer to fits performed with a linear and logarithmic $L_3$ scale. The error bars represent the $1\sigma$ uncertainties from the individual fits. The horizontal dashed line marks the value of $\kappa_*$ obtained from the excited-state Efimov resonance. The gray-shaded region marks the corresponding error range.[]{data-label="fig.1stBRange"}](1stBRange.eps){width="8.6cm"} Figure \[fig.1stBRange\] shows the values for $\kappa_*$ resulting from fits with different cutoff scattering lengths $a_{\rm min}$ in the range between $2$ and $6\,r_{\rm vdW}$. The filled blue squares represent the fit to the $L_3$ results on a linear scale. This fit puts most weight on the lower resonance, but as $a_{\rm min}$ increases it gives more weight to the region between the resonances, and the resulting value for $\kappa_*$ decreases by almost 10%. The fits to the $L_3$ data on a logarithmic scale (filled red circles) show a similar behavior with a trend towards smaller values of $\kappa_*$ at larger values of $a_{\rm min}$. The fits for $a_{\rm min}/r_{\rm vdW} \ge 5$ do not provide satisfactory results, mainly because of significant problems in reproducing the position of the resonance near 130G. The fits for $a_{\rm min}/r_{\rm vdW} \le 4$ (central panel in Fig. \[fig.1stfit\]) appear good, but for lower values of $a_{\rm min}$ the result may be subject to significant finite-range effects. We therefore consider $a_{\rm min}/r_{\rm vdW} = 4$ to be the best choice. It gives $\kappa_* = 0.00643(4) \, a_0^{-1}$, based on averaging the results of the linear and logarithmic fits. The error given here indicates only the statistical uncertainty, but the dependence of the results on $a_{\rm min}$ suggests additional systematic errors on the order of 10%. The dashed horizontal line and the gray-shaded region in Fig. \[fig.1stBRange\] indicate the value of $\kappa_*$ obtained from the excited-state resonance in Sec. \[sec:excited\], together with the corresponding error range. It may be seen that our results are consistent with discrete scaling as described by Eq. (\[eq:kappa\_universal\]) within the relatively large uncertainties due to finite-range effects in the low-field region. Conclusion {#sec:discuss} ========== We have reanalyzed experimental results for the Efimov recombination resonances in $^6$Li arising from the ground and excited Efimov states, using a very precise model of the two-body scattering [@Zurn2013pco] and a new model of temperature-dependent effects in three-body recombinantion of three nonidentical fermions. From the excited-state Efimov resonance [@Williams2009efa], we obtain the value for the 3BP in the wavenumber representation, $$\kappa_* = 0.00678(6)\,a_0^{-1} \, . \nonumber$$ This gives the reduced 3BP $$\kappa_* r_{\rm vdW} = 0.212(2)\, . \nonumber$$ According to Eq. (\[eq:aminus\]) this corresponds to a reduced 3BP in the scattering length representation, $$a_-^*/r_{\rm vdW} = -7.11(6) \, . \nonumber$$ This latter representation of the 3BP facilitates a direct comparison with three-boson systems, which are characterized by a single scattering length [@note:aeff]. Our analysis of the ground-state Efimov resonances [@Ottenstein2008cso; @Huckans2009tbr] yields values for the 3BP that are consistent with the above result within an estimated 10% uncertainty. Alternatively, they may be viewed as confirming that the lowest Efimov period in $^6$Li is within 10% of the universal value of 22.7. The uncertainties, which follow from systematic shifts that depend on the choice of the lower cutoff applied to the scattering lengths in the data analysis, place an upper bound on the magnitude of possible finite-range corrections to the lowest Efimov period. The rapid decrease of such shifts with increasing order of the Efimov state [@Thogersen2008upo; @Schmidt2012epb] gives us confidence that such corrections can be neglected for the 3BP if determined from the position of an excited-state resonance. It is very interesting to compare the present result with the recent measurement for cesium in Ref. [@Huang2014oot], which gave $a_-^{(1)}=-20190(1200)\,a_0$, implying a reduced 3BP $a_-^*/r_{\rm vdW}^{\rm Cs} = -8.8(4)$ with $r_{\rm vdW}^{\rm Cs} = 101.1\,a_0$ [@Berninger2013frw]. In both cases, the Feshbach resonances used for interaction tuning are strongly entrance-channel-dominated [@Chin2010fri]. The present result for the reduced 3BP in $^6$Li differs from that measured for Cs by a factor $0.81(4)$. This clearly demonstrates that the van der Waals length is not the only relevant quantity in determining the 3BP. Even for strongly entrance-channel-dominated Feshbach resonances, van der Waals universality of the 3BP is only approximate, and is subject to further influences. It remains a challenge for theory to understand fully the role of finite-range effects [@Naidon:2011], of the physics of particular Feshbach resonances [@Schmidt2012epb; @Wang2014uvd], of the role of genuine short-range three-body forces [@Axilrod1943iot; @Soldan2003tbn; @Dincao2009tsr], and of other species-dependent factors such as the number of bound states in the two-body potentials [@Wang2012oot]. It is also possible that light particles can tunnel through the barrier in the effective potential [@Wang2012oot] more effectively than heavy ones. It would be highly desirable to investigate other systems at the precision of the present work, by detecting excited-state Efimov resonances and thus accurately measuring the 3BP. The bosonic gas of $^7$Li [@Gross2009oou; @Pollack2009uit; @Rem2013lot] is a prime candidate for such experiments, because it provides another example of a light system with exceptionally well characterized two-body scattering properties [@Dyke2013frc; @Julienne:Li67:2014]. Atoms such as $^{39}$K [@Zaccanti2009ooa; @Roy2013tot; @Fletcher2013soa] and $^{85}$Rb [@Wild2012mot] also provide very interesting systems for future precision experiments: $^{85}$Rb offers access to another entrance-channel-dominated case, while $^7$Li and $^{39}$K offer Feshbach resonances of intermediate character [@Chin2010fri]. Acknowledgements {#acknowledgements .unnumbered} ================ We thank Selim Jochim for providing the experimental $L_3$ data for the ground-state Efimov resonances and Benno Rem for useful discussions. We acknowledge support by the Austrian Science Fund FWF within project P23106 and by EPSRC under grant no. EP/I012044/1. D.S.P. acknowledges support from the IFRAF Institute. The research leading to these results has received funding from the European Research Council (FR7/2007-2013 Grant Agreement no. 341197). K.M.O. acknowledges support from the NSF (Grant No. 1312430). [68]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , ****, (). , ****, (). , ****, (). , ****, (). , , , , ****, (). , , , , , , , , , , , ****, (). , , , , , , , , , ****, (). , , , ****, (). , , , , ****, (). , , , , ****, (). , , , , , ****, (). , , , , , , , , , ****, (). , , , , , ****, (). , , , , , ****, (). , , , , , , ****, (). , , , , , ****, (). , , , , , , , ****, (). , , , , ****, (). , , , , , (). , , , , , , , ****, (). , ****, (). , , , ****, (). , ****, (). , , , , , , , ****, (). , , , , , , , , , ****, (). , , , , , , , , , ****, (). , (). , , , , ****, (). , , , ****, (). , , , , ****, (). , , , ****, (). , ****, (). , , , ****, (). , ****, (). , , , , , , ****, (). , , , ****, (). , , , ****, (). , ****, (). , , , ****, (). , , , , , , , , , , , ****, (). , , , , ****, (). , , , , , , , , , ****, (). , ****, (). , , , , , , ****, (). , , , , ****, (). , , , ****, (). , , , , , , ****, (). , , , , , , , , , , , , ****, (). , ****, (). , ****, (). , , , ****, (). , ****, (). , ****, (). , , , , ****, (). . , ****, (). , ****, (). , , , ****, (). , ****, (). , , , ****, (). , , , ****, (). , , , ****, (). , , , , , ****, ().

--- abstract: 'Let $H$ be a connected semisimple linear algebraic group defined over $\mathbb C$ and $X$ a compact connected Riemann surface of genus at least three. Let ${\mathcal M}''_X(H)$ be the moduli space parametrising all topologically trivial stable principal $H$-bundles over $X$ whose automorphism group coincides with the centre of $H$. It is a Zariski open dense subset of the moduli space of stable principal $H$-bundles. We prove that there is a universal principal $H$-bundle over $X\times {\mathcal M}''_X(H)$ if and only if $H$ is an adjoint group (that is, the centre of $H$ is trivial).' address: - 'Chennai Mathematical Institute, 92, G.N. Chetty Road, Chennai 600017, India' - 'School of Mathematics, Tata Institute of Fundamental Research, Homi Bhabha Road, Mumbai 400005, India' - 'The Institute of Mathematical Sciences, CIT Campus, Taramani, Chennai 600113, India' - 'Department of Mathematical Sciences, The University of Liverpool, Peach Street, Liverpool, L69 7ZL, England' author: - 'V. Balaji' - 'I. Biswas' - 'D. S. Nagaraj' - 'P. E. Newstead' title: Universal families on moduli spaces of principal bundles on curves --- \[section\] \[theorem\][Lemma]{} \[theorem\][Proposition]{} \[theorem\][Corollary]{} \[theorem\][Definition]{} \[theorem\][Remark]{} [^1] Introduction ============ Let $X$ be a compact connected Riemann surface of genus at least three. Let ${\mathcal M}_X(n,d)$ denote the moduli space of all stable vector bundles over $X$ of rank $n$ and degree $d$, which is a smooth irreducible quasiprojective variety defined over $\mathbb C$. A vector bundle $\mathcal E$ over $ X\times {\mathcal M}_X(n,d)$ is called *universal* if for every point $m\in {\mathcal M}_X(n,d)$, the restriction of $\mathcal E$ to $X\times \{m\}$ is in the isomorphism class of holomorphic vector bundles over $X$ defined by $m$. A well-known theorem says that there is a universal vector bundle over $X\times {\mathcal M}_X(n,d)$ if and only if $d$ is coprime to $n$ (see [@Ty] for existence in the coprime case, [@Ra] for non-existence in the non-coprime case and [@Ne] for a topological version of non-existence in the case $d=0$). Let $H$ be a connected semisimple linear algebraic group defined over the field of complex numbers. Ramanathan extended the notion of (semi)stability to principal $H$-bundles and constructed moduli spaces for stable principal $H$-bundles over $X$ [@Ra1; @Ra2; @Ra3]. The construction works for any given topological type, yielding a moduli space which is an irreducible quasiprojective variety defined over $\mathbb C$. We are concerned here with the case of topologically trivial stable principal $H$-bundles. Let ${\mathcal M}_X(H)$ denote the moduli space of topologically trivial stable principal $H$-bundles over $X$. Let $Z(H)\, \subset\, H$ be the centre. For any $H$-bundle $E_H$, the group $Z(H)$ is contained in the automorphism group $\text{Aut}(E_H)$. Let $${\mathcal M}'_X(H)\, \subset \, {\mathcal M}_X(H)$$ be the subvariety consisting of all $H$-bundles $E_H$ over $X$ with the property $\text{Aut}(E_H)\, \cong\, Z(H)$. It is known that ${\mathcal M}'_X(H)$ is a dense Zariski-open subset contained in the smooth locus of ${\mathcal M}_X(H)$. A principal $H$-bundle $E$ over $X\times {\mathcal M}'_X(H)$ will be called a *universal bundle* if for every point $m\in {\mathcal M}_X(n,d)$, the restriction of $\mathcal E$ to $X\times \{m\}$ is in the isomorphism class of stable $H$-bundles over $X$ defined by the point $m$ of the moduli space. The following theorem is the main result proved here: \[theorem0\] There is a universal $H$-bundle over $X\times {\mathcal M}'_X(H)$ if and only if $Z(H) \, =\, e$. \[sl\] [Non-existence for $H={\rm SL}(n,{\mathbb C})$ was previously known [@Ra; @Ne], as also was existence for $H={\rm PGL}(n,{\mathbb C})$.]{} [*Acknowledgments*]{}. We thank A. King, A. Nair and P. Sankaran for useful discussions. Existence of universal bundle ============================= We begin this section by recalling very briefly certain facts from [@basa]. Unless otherwise specified, all bundles and sections considered will be algebraic. \[hb\] Let $E$ be a principal $G$-bundle over $X$, where $G$ is a reductive linear algebraic group defined over $\mathbb C$, and $H \subset G$ is a Zariski closed semisimple subgroup. For any variety $Y$ equipped with an action of $G$, the fibre bundle $(E\times Y)/G$ over $X$ associated to $E$ will be denoted by $E(Y)$. 1. There is a natural action of the group ${\mbox{{\rm Aut}$\,$}}_GE$, defined by all automorphisms of $E$ over the identity map of $X$ that commute with the action of $G$, on $\Gamma (X, E(G/H))$ (the space of all holomorphic sections of the fibre bundle $E(G/H) \,=\, E/H$ over $X$) and the orbits correspond to the equivalence classes of $H$-reductions of $E$ with two reductions being equivalent if the corresponding principal $H$-bundles are isomorphic. 2. Let $G = {\rm GL}(n,{{\mathbb C}})$, and let $\phi: H \hookrightarrow G$ be a faithful representation of the semisimple group $H$. Let $Q$ denote the open subset of semistable principal $G$-bundles (or equivalently, of trivial determinant semistable vector bundles of rank $n$) of the usual “Quot scheme”, and let $Q(\phi)$ be the “Quot scheme” which parametrises pairs of the form $(E',s)$, where $E'$ is a principal $G$-bundle and $s$ is a reduction of structure group of $E'$ to $H$. Then $Q(\phi)$ is in a sense a “relative Quot scheme”. As is clear from the definition and the notation, this scheme is dependent on the choice of the inclusion $\phi: H {\hookrightarrow}G$ (for details see [@basa; @Ra2; @Ra3]). One also has a [*tautological*]{} sheaf on $X \times Q$ which in fact is a vector bundle. We denote by ${{\mathcal E}}$ the [*associated tautological*]{} principal $G$-bundle on $X \times Q$. Recall that the moduli space of principal $G$-bundles ($G = {\rm GL}(n, {\mathbb C})$) is realised as a good quotient of $Q$ by the action of a reductive group ${{\mathcal G}}$. We may also assume that the group ${{\mathcal G}}$ is with trivial centre (see for example [@drezet]). It is immediate that the action of ${{\mathcal G}}$ on $Q$ lifts to an action on $Q(\phi)$, where ${{\mathcal G}}$, $Q$ and $Q(\phi)$ are defined in Remark \[hb\]. We have a morphism $$\label{d.psi} \psi \,:\, Q(\phi) \, {\longrightarrow}\, Q\, ,$$ which sends any $H$-bundle $E'$ to the ${\rm GL}(n,{{\mathbb C}})$-bundle obtained by extending the structure group of $E'$ using the homomorphism $\phi$. In fact, $\psi$ is a ${{\mathcal G}}$-equivariant [*affine morphism*]{} (see [@basa]). Continuing with the notation in the above two remarks, consider the ${{\mathcal G}}$-action on $Q(\phi)$ (defined in Remark \[hb\](2)) with the linearisation induced by the affine ${{\mathcal G}}$-morphism $\psi$ in . Since a good quotient of $Q$ by ${{\mathcal G}}$ exists and since $\psi$ is an affine ${{\mathcal G}}$-equivariant map, a good quotient $Q(\phi)// {{\mathcal G}}$ exists (see [@Ra3 Lemma 5.1]). Moreover by the universal property of categorical quotients, the canonical morphism $$\label{opsi} {\overline}{\psi} \,: \,Q(\phi)//{{\mathcal G}}\,{\longrightarrow}\, Q//{{\mathcal G}}$$ given by $\psi$ is also [*affine*]{}. \[moduli\] Let ${\mathcal M}_X(H)$ denote the scheme $Q(\phi)//{{\mathcal G}}$ (see ). Then this scheme is the coarse moduli scheme of semistable $H$-bundles. Further, the scheme ${\mathcal M}_X(H)$ is projective, and if $H \hookrightarrow {\rm GL}(V)$ is a faithful representation, then the canonical morphism $${\overline}{\psi}\, :\, {\mathcal M}_X(H) \,{\longrightarrow}\, {\mathcal M}_X({\rm GL}(V))\, =\, Q//{{\mathcal G}}$$ is finite. Let $Q(\phi)^{s}$ be the open subscheme of $ Q(\phi)$ consisting of stable $H$-bundles. \[free\] Let $Q_H' \subset Q(\phi)^s $ be the subset parametrising all stable $H$-bundles whose automorphism group is $Z(H)$. Then the action of ${{\mathcal G}}$ on the subset $Q_H'$ is free, and furthermore, the quotient morphism $Q_H' {\longrightarrow}Q_H'/{{{\mathcal G}}}$ is a principal ${{\mathcal G}}$-bundle. In fact, $Q_H'/{{{\mathcal G}}}$ is precisely the Zariski open subset ${\mathcal M}'_X(H)$ (the variety ${\mathcal M}'_X(H)$ is defined in the introduction). For any point $E_H \in Q(\phi)$, the isotropy subgroup of $E_H$ for the action of ${{\mathcal G}}$ coincides with ${\rm Aut}(E_H)/Z(H)$. This can be seen as follows: firstly, the point $E_H$ is a pair $(E,s)$, where $E \in Q$ and $s$ is a reduction of structure group to $H$ of the $G$-bundle $E$. It is well known for the action of ${{\mathcal G}}$ on $Q$ that the isotropy at $E$ is precisely the group ${\rm Aut}(E)/Z(G)$. From this it is easy to see that the isotropy of $E_H$ for the action of ${{\mathcal G}}$ on $Q(\phi)$ is the group ${\rm Aut}(E,s)/Z(H) = {\rm Aut}(E_H)/Z(H)$. Hence it follows that the action of ${{\mathcal G}}$ on the open subset $Q_H'$ is free, and the proof of the lemma is complete. We remark that, at least when $H$ is of adjoint type, $Q_H'$ is a non-empty open subset of $Q(\phi)^{s}$. Openness is easy and can be seen for example from [@faltings Theorem II.6 (ii)]. Non-emptiness follows from Proposition \[codim\] below. We prove a proposition on semisimple groups, possibly known to experts but which we could not locate in any standard text. \[faithrep\] Let $H$ be a semisimple algebraic group. Then $H$ has a faithful, irreducible representation $\phi:H {\longrightarrow}{\rm GL}(V)$ if and only if the centre of $H$ is cyclic. One way the implication is easy, namely suppose that a faithful irreducible representation exists, then the centre $Z(H)$ of $H$ is cyclic. To see this, first note that under the representation $\phi$, the centre $Z(H)$ maps to a subgroup which commutes with all elements of $\phi(H)$. Since $\phi$ is irreducible, this implies by Schur’s Lemma that $\phi(Z(H)) \subset Z({\rm GL}(V))$, where $Z({\rm GL}(V))$ is the centre of ${\rm GL}(V)$. Observe further that since $H$ is semisimple, we have $\phi(H) \subset {\rm SL}(V)$. Hence, $Z(H) \subset Z({\rm SL}(V))$ and is therefore cyclic. The contrapositive statement is harder to prove. We proceed as follows: Let $H'$ be the simply connected cover of $H$, and let $\overline H$ be the associated adjoint group, namely $H/Z(H)$. Let $\Lambda$ (respectively, $\Lambda_R$) be the weight lattice (respectively, root lattice). In other words, by the Borel-Weil theorem $\Lambda = {\mathcal X}(B)$ and $\Lambda_R = {\mathcal X}(\overline B)$, where $B$, $\overline B$ are fixed Borel subgroups of $H$, $\overline H$. Then, from the exact sequence $$e {\longrightarrow}Z(H) {\longrightarrow}H {\longrightarrow}\overline H {\longrightarrow}e$$ we see that $\Lambda/\Lambda_R \simeq {\mathcal X}(Z(H))$. In other words, the quotient group $\Lambda/\Lambda_R$ is cyclic of order $m$. Let $\overline \lambda$ be a generator of the cyclic group $\Lambda/\Lambda_R$. Then, by an action of the Weyl group we may assume that the coset representative $\lambda \in \Lambda$ is actually a dominant weight. Suppose that the root lattice has the following decomposition (corresponding to the simple components of $\overline H$): $$\Lambda_R = \bigoplus_{i=1}^\ell \Lambda^{i}\, .$$ Then, by possibly adding dominant weights from the $\Lambda^{i}$, we may assume that $(m \cdot \lambda) \in \Lambda_R$ has all its direct sum components $\lambda_i \neq 0$, $i\, \in\, [1\, ,\ell]$, where $\lambda$ is as above. For this choice of $\lambda \in \Lambda$ let $V_{\lambda}$ be the corresponding $H$-module given by $$\phi_{\lambda} : H {\longrightarrow}{\rm GL}(V_{\lambda})\, .$$ Then one knows that $\phi_{\lambda}$ is an irreducible representation of $H$. We claim that the representation $\phi_{\lambda}$ is even [*faithful*]{}. Suppose that this is not the case. Let $K_\lambda \,:= \,{\rm kernel}(\phi_{\lambda}) \,\neq\, e$. Firstly, $K_\lambda \subset Z(H)$. To see this, let $K'$ be the inverse image of $K_\lambda$ in the simply connected cover $H'$ of $H$. Then the choice of $\lambda$ so made that its simple components are non-zero in fact forces the following: Suppose that $H' = H_1 \times \cdots \times H_\ell$ is the decomposition of $H'$ into its almost simple factors. (We recall that a semisimple algebraic group is called almost simple if the quotient of it by its centre is simple.) Then the normal subgroup $K'$ in its decomposition in $H'$ is such that $K_i\, :=\, K'\bigcap H_i$ are proper normal subgroups of $H_i$. In particular, $K_i \subset Z(H_i)$ for all $i\, \in\, [1\, ,\ell]$. This implies that $K' \subset Z(H')$ and hence, $K_\lambda \subset Z(H)$. Note that the dominant character $\lambda$ is non-trivial on the generator of the centre $Z(H)$ because ${\mathcal X}(Z(H)) = \Lambda/\Lambda_R$. Now $K_\lambda \subset Z(H)$ and $Z(H)$ cyclic implies that $\lambda$ is non-trivial on the generator of $K_\lambda$ as well. This contradicts the fact that $K_\lambda = {\rm kernel}(\phi_{\lambda})$. This proves the claim. Therefore, the proof of the proposition is complete. If $H$ is of adjoint type, then by Proposition \[faithrep\] we can choose the inclusion $\phi: H {\hookrightarrow}G$ in Remark \[hb\](2) to be an irreducible representation. Henceforth, $\phi$ will be assumed to be irreducible. Let $E$ be a stable $H$-bundle of trivial topological type. Recall that one can realise $E$ from a unique, up to an inner conjugation, irreducible representation of $\pi_1(X)$ in a maximal compact subgroup of $H$ (see [@Ra1]). For notational convenience, we will always suppress the base point in the notation of fundamental group. Denote by $M(E)$ the Zariski closure of the image of $\pi_1(X)$ in $H$. \[codim\] Let $H$ be of adjoint type and let $\phi$ be a faithful irreducible representation of $H$ in $V$ (see Proposition \[faithrep\]). Let $U \subset Q(\phi)^{s}$ be the subset defined as follows: $$U = \{E \in Q(\phi)^{s}\, |\, M(E) = H\}.$$ Then $U$ is non-empty and is contained in the subset $Q_H'$ of stable $H$-bundles which have trivial automorphism group. (Since $H$ is of adjoint type, $Z(H)=e$.) Since $\phi$ is irreducible, using Lemma 2.1 of [@Su] we conclude that there is an irreducible representation $$\rho\, :\, \pi_1(X)\,{\longrightarrow}\, H$$ which have the property that the composition $\phi \circ \rho :\pi_1(X){\longrightarrow}G := \text{GL}(V)$ continues to remain irreducible. This implies that $M(E_{\rho}) = H$ by the construction in [@Su Lemma 2.1] and hence $E_{\rho} \in U$, i.e., $U$ is non-empty. Let $E \in U$. Then by the definition of $U$, there exists a representation $\rho$ of $\pi_1(X)$ in $H$ such that $E \simeq E_{\rho}$, where $E_{\rho}$ is the flat principal $H$-bundle given by $\rho$. Observe that $E_{\rho}$ is a stable $H$-bundle and the associated $G$-bundle is also stable. Hence all the automorphisms of this associated $G$-bundle lie in $Z(G)$. Since $H$ is of adjoint type, it follows that the $H$-bundle $E_{\rho}$ has no non-trivial automorphisms. Hence it follows that $U \subset Q_H'$. We have the following theorem on existence of universal families. \[family\] Let $H$ be a group of adjoint type and ${\mathcal M}'_X(H)$ the Zariski open subset of ${\mathcal M}_X(H)$ defined in the introduction. Then there exists a universal family of principal $H$-bundles on $X \times {\mathcal M}'_X(H)$. We first observe that the variety ${\mathcal M}'_X(H)$ is precisely the image of $Q_H'$ under the quotient map for the action of $\mathcal G$. We recall that $Q_H'$ is nonempty by Proposition \[codim\]. Since $Z(H) = e$, it follows that $H \subset {{\overline}G} \,:=\, G/Z(G)$, where $G ={\rm GL}(V)$ is as in Remark \[hb\](2). Consider the tautological $G$-bundle ${{\mathcal E}}$ on $X \times Q^s$, and let ${\overline}{{{\mathcal E}}}$ be the corresponding ${\overline}G$-bundle obtained by extending the structure group. Then it is well-known that the adjoint universal bundle ${\overline}{{{\mathcal E}}}$ descends to the quotient ${\mathcal M}_X(G)^s$ (see [@drezet]). We follow the same strategy for ${\mathcal M}'_X(H)$ as well. Consider the pulled back ${{\overline}G}$-bundle $(\text{Id}_X \times {\psi})^{*} {\overline}{{{\mathcal E}}}$ on $X\times Q_H'$, where $\psi$ is the map in and ${{\overline}G} \,:=\, G/Z(G)$. The action of ${{\mathcal G}}$ on $Q_H'$ is free by Lemma \[free\]. Therefore, the quotient $Q_H' {\longrightarrow}{\mathcal M}'_X(H)$ is a principal ${{{\mathcal G}}}$-bundle. Further, the action of ${{\mathcal G}}$ lifts to the tautological bundle $(\text{Id}_X \times {\psi})^{*} {\overline}{{{\mathcal E}}}$. In particular, the principal ${\overline}G$-bundle $(\text{Id}_X\times {\psi})^{*} {\overline}{{{\mathcal E}}}$ descends to a principal ${\overline}G$-bundle over $X \times {\mathcal M}'_X(H)$. Let us denote this descended ${\overline}G$-bundle over $X \times {\mathcal M}'_X(H)$ by ${\overline}{{{\mathcal E}}}_0$. Let $\pi \,:\, (\text{Id}_X\times \psi)^*{{{\mathcal E}}}(G/H)\,{\longrightarrow}\, (\text{Id}_X\times \psi)^*{{\overline}{{{\mathcal E}}}}({{\overline}G}/H)$ be the natural map induced by the projection $G/H {\longrightarrow}{{\overline}G}/H$, where $\psi$ is the map in . We note that the universal $H$-bundle over $X \times U$, where $U$ is defined in Proposition \[codim\], is a reduction of structure group of the pulled back $G$-bundle $(\text{Id}_X\times \psi)^*{{\mathcal E}}$. Let $$\sigma \,:\, X\times U\, \longrightarrow\, (\text{Id}_X\times \psi)^*{{{\mathcal E}}}(G/H)$$ be the section giving this reduction of structure group. Then the composition ${\pi}\circ \sigma$ is a section of $(\text{Id}_X\times \psi)^*{{\overline}{{{\mathcal E}}}}({{\overline}G}/H)$ over $X\times U$. Since $H$ is semisimple, a lemma of Chevalley says that there is a ${{\overline}G}$-module $W$ and an element $w\in W$ such that $H$ is precisely the isotropy subgroup for $w$ (see [@Bo p. 89, Theorem 5.1]). Therefore, ${{\overline}G}/H$ is identified with the closed ${{\overline}G}$-orbit in $W$ defined by $w$. Then we see that ${{\overline}{{{\mathcal E}}}}({{\overline}G}/H) {\hookrightarrow}{{\overline}{{{\mathcal E}}}}(W)$. We may therefore view the section ${\pi}\circ \sigma: X\times U\longrightarrow (\text{Id}_X\times \psi)^*{{\overline}{{{\mathcal E}}}}({{\overline}G}/H)$ as a section of the vector bundle $(\text{Id}_X\times \psi)^*{{\overline}{{{\mathcal E}}}}(W)$ over $X\times U$. Since $(\text{Id}_X \times {\psi})^{*} {\overline}{{{\mathcal E}}}$ descends to the ${{\overline}G}$-bundle ${{\overline}{{{\mathcal E}}}}_{0}$ over $X \times {\mathcal M}'_X(H)$, it follows that the associated vector bundle $(\text{Id}_X \times {\psi})^{*} {\overline}{{{\mathcal E}}}(W)$ also descends to $X \times {\mathcal M}'_X(H)$. Clearly this vector bundle is nothing but the associated vector bundle ${{\overline}{{{\mathcal E}}}}_{0}(W)$, associated to the ${\overline}G$-bundle ${{\overline}{{{\mathcal E}}}}_{0}$ on $X \times {\mathcal M}'_X(H)$ for the ${\overline}G$-module $W$. Since each point of ${\mathcal M}'_X(H)$ represents an isomorphism class of stable $H$-bundle, it follows that set theoretically the reduction section ${\pi}\circ \sigma \,\in \, {\Gamma}((\text{Id}_X\times \psi)^*{{\overline}{{{\mathcal E}}}}(W))$ descends to a section on $X \times {\mathcal M}'_X(H)$ of the descended vector bundle ${{\overline}{{{\mathcal E}}}}_{0}(W)$. We now appeal to [@knr Proposition 4.1], which implies that the section ${\pi}\circ \sigma$ in fact descends to give a holomorphic section of ${{\overline}{{{\mathcal E}}}}_{0}(W)$ over $X \times {\mathcal M}'_X(H)$. Again set-theoretically, the image of this section of ${{\overline}{{{\mathcal E}}}}_{0}(W)$ lies in ${{\overline}{{{\mathcal E}}}}_{0}({{\overline}G}/H)\, \subset\, {{\overline}{{{\mathcal E}}}}_{0}(W)$. As before, from [@knr Proposition 4.1] it follows that ${\pi}\circ \sigma$ gives a reduction of structure group to $H$ of the descended ${{\overline}G}$-bundle ${{\overline}{{{\mathcal E}}}}_{0}$. The $H$-bundle over $X \times {\mathcal M}'_X(H)$ obtained this way is the required universal $H$-bundle. This completes the proof of the theorem. Nonexistence of universal bundle ================================ Let $H$ be a complex semisimple linear algebraic group, and let $K\, \subset\, H$ be a maximal compact subgroup. The Lie algebra of $K$ will be denoted by $\mathfrak k$. A homomorphism $$\rho\, :\, \pi_1(X)\, \longrightarrow\, K$$ is called *irreducible* if no nonzero vector in $\mathfrak k$ is fixed by the adjoint action of the subgroup $\rho (\pi_1(X))\, \subset \, K$ on $\mathfrak k$. Let $\text{Hom}^{\text{irr}}(\pi_1(X), K)$ denote the space of all irreducible homomorphisms such that the corresponding $K$-bundle is topologically trivial. So any homomorphism in $\text{Hom}^{\text{irr}}(\pi_1(X), K)$ is induced by a homomorphism from $\pi_1(X)$ to the universal cover of $K$. Let $g$ denote the genus of $X$. Assume that $g\, \geq\, 3$. If we choose a basis $$\{a_1, \cdots ,a_g, b_1, \cdots ,b_g\} \,\subset\, \pi_1(X)$$ such that $$\label{pi-rep.} \pi_1(X) \,=\, {\langle} a_1,\ldots ,a_g, b_1,\ldots ,b_g : \prod_{i=1}^na_ib_ia_i^{-1}b_i^{-1} \rangle\, ,$$ then $\text{Hom}^{\text{irr}}(\pi_1(X), K)$ gets identified with a real analytic subspace of $K^{2g}$. For any $\rho\, \in\, \text{Hom}^{\text{irr}}(\pi_1(X), K)$, the principal $H$-bundle obtained by extending the structure group of the principal $K$-bundle given by $\rho$ is stable. A theorem of Ramanathan says that all topologically trivial stable $H$-bundles arise in this way, that is, the space of all equivalence classes of irreducible homomorphisms of $\pi_1(X)$ to $K$ is in bijective correspondence with the space of all stable $H$-bundles over $X$ [@Ra1 Theorem 7.1]. More precisely (see the proof of [@Ra1 Theorem 7.1]), the real analytic space underlying the moduli space ${\mathcal M}_X(H)$ parametrising topologically trivial stable $H$-bundles is analytically isomorphic to the quotient space $$R(\pi_1(X),H)\, :=\, \text{Hom}^{\text{irr}}(\pi_1(X), K)/K\, ,$$ for the action constructed using the conjugation action of $K$ on itself. Let $$\label{q-pr.-b00.} q\, :\, \text{Hom}^{\text{irr}}(\pi_1(X), K) \, \longrightarrow\, R(\pi_1(X),H) \,\cong\, {\mathcal M}_X(H)$$ be the quotient map. The open subset ${\mathcal M}'_X(H)\, \subset\, {\mathcal M}_X(H)$ is a smooth submanifold and the restriction of the map in $$\label{q-pr.-b.} q\vert_{q^{-1}({\mathcal M}'_X(H))}\, :\, q^{-1}({\mathcal M}'_X(H))\, \longrightarrow\, {\mathcal M}'_X(H)$$ is a smooth principal $K/Z$-bundle, where $Z$ is the centre of $K$. Note that $Z$ coincides with the centre of $H$ (as $H$ is semisimple, its centre is a finite group). If ${\mathcal U}_H$ is a universal $H$-bundle over $X\times {\mathcal M}'_X(H)$, then we have a reduction of structure group of $\mathcal U_H$ to $K$ which is constructed using the correspondence established in [@Ra1] between stable $H$-bundles and irreducible flat $K$-bundles. Let $$\label{d.uk} {\mathcal U}_K \,{\longrightarrow}\, X\times {\mathcal M}'_X(H)$$ be the smooth principal $K$-bundle obtained from ${\mathcal U}_H$ this way. Consider the principal $K/Z$-bundle over $X\times {\mathcal M}'_X(H)$, where $Z\, \subset\, K$ is the centre, obtained by extending the structure group of ${\mathcal U}_K$ using the natural projection of $K$ to $K/Z$. The restriction of this $K/Z$-bundle to $p\times {\mathcal M}'_X (H)$, where $p$ is the base point in $X$ used for defining $\pi_1(X)$, is identified with the $K/Z$-bundle in . To see this first note that the universal $K$-bundle over $X\times {\rm Hom}(\pi_1(X),K)$ is obtained as a quotient by the action of $\pi_1(X)$ on $\widetilde{X}\times {\rm Hom}(\pi_1(X),K)\times K$, where $\widetilde{X}$ is the pointed universal cover of $X$ for the base point $p$; the action of $z\in \pi_1(X)$ sends any $(\alpha,\beta,\gamma)$ to $(\alpha ,z, \beta , \beta(z)^{-1}\gamma)$. From this it follows that the restriction of this universal bundle to $p\times {\rm Hom}(\pi_1(X),K)$ is canonically trivialized. The above mentioned identification is constructed using this trivialization. Our aim is to show that no $K$-bundle over ${\mathcal M}'_X(H)$ exists that produces the $K/Z$-bundle in by extension of structure group, provided the centre $Z$ is non-trivial. Let $F_3$ denote the free group on three generators. Fix a surjective homomorphism $$\label{def.f} f\, :\, \pi_1(X) \,\longrightarrow\, F_3$$ that sends $a_i$, $1\leq i\leq 3$, to the $i$-th generator of $F_3$ and sends $a_i$, $4\leq i\leq g$, and $b_i$, $1\leq i\leq g$, to the identity element, where $a_j$, $b_j$ are as in (recall that $g\geq 3$). Set $$\label{qmap} R(F_3,H) \,:=\, \text{Hom}^{\text{irr}}(F_3, K)/K$$ to be the equivalence classes of irreducible representations. Note that $\text{Hom}(F_3, K) \simeq K^3$, and under this identification the action $$\mu \,:\, \text{Hom}(F_3, K)\times K \,\longrightarrow\, \text{Hom}(F_3, K)$$ given by $\mu(\rho, A) = A^{-1}\rho A$ corresponds to the simultaneous diagonal conjugation action of $K$ on the three factors. Since $F_3$ is a free group and $K$ is connected, any homomorphism $\rho$ from $F_3$ to $K$ can be deformed to the trivial homomorphism. This implies that the principal $H$-bundle corresponding to the homomorphism $$\rho\circ f\, :\, \pi_1(X) \,\longrightarrow\, K$$ is topologically trivial, where $f$ is defined in . Therefore, we have an embedding $$\label{i.map} R(F_3,H) \,\longrightarrow\, R(\pi_1(X),H)\, \cong\, {\mathcal M}_X(H)$$ that sends any $\rho$ to $\rho\circ f$. Let $R'(F_3,H)\, \subset\, R(F_3,H)$ be the inverse image of the open subset ${\mathcal M}'_X(H)$ under the above map. By Proposition \[codim\] it follows that $R'(F_3,H)$ is a [*non-empty*]{} open subset of $R(F_3,H)$. Fix a point $p \in X$, which will also be the base point for the fundamental group. Consider the restriction of the principal $K$-bundle ${\mathcal U}_K$ in to $p\times {\mathcal M}'_X(H)\, {\hookrightarrow}\, X\times {\mathcal M}'_X(H)$ and denote it by ${\mathcal U}_{K,p}$. Let $$\label{1ga} \gamma \,:\, p\times R'(F_3,H))\,{\longrightarrow}\, p\times{\mathcal M}'_X(H)$$ be the map given by the embedding $R'(F_3,H) {\longrightarrow}{\mathcal M}'_X(H)$ constructed in . Taking the pull back of the above defined principal $K$-bundle ${\mathcal U}_{K,p}$ over $p\times {\mathcal M}'_X(H)$ under the morphism $\gamma$ in , $$\label{eqdig} \begin{matrix} \widetilde M && {\mathcal U}_{K,p} \\ \Big\downarrow & & \Big\downarrow\\ p\times R'(F_3,H)) & \stackrel{\gamma}{\longrightarrow}& p\times {\mathcal M}'_X(H) \\ \end{matrix}$$ we obtain a principal $K$-bundle $\widetilde{M} \, :=\, {\gamma}^{*}{\mathcal U}_{K,p}$ on $R'(F_3,H)$ whose associated $K/Z$-bundle (as before, $Z$ is the centre of $K$) is precisely the $K/Z$-bundle $q^{-1}( R'(F_3,H)) \,\longrightarrow\, R'(F_3,H)$ constructed in . In particular, $q^{-1}( R'(F_3,H))\, \simeq\, {\widetilde M}/Z$. Let $\text{Hom}^{\text{irr}}(F_3, K)'\, \subset\, \text{Hom}^{\text{irr}}(F_3, K)$ be the inverse image of $R'(F_3,H)$ under the quotient map in . Let $$\label{map.q0} q_0\, :\, \text{Hom}^{\text{irr}}(F_3, K)' \,\longrightarrow\, R'(F_3,H)$$ be the restriction of the quotient map in . So $q_0$ is the quotient for the conjugation action of $K$. This map $q_0$ defines a principal $K/Z$-bundle over $R'(F_3,H)$ which is evidently identified with the $K/Z$-bundle ${\widetilde M}/Z\,\longrightarrow\, R'(F_3,H)$ obtained from . We will prove (by contradiction) that such a $K$-bundle $\widetilde{M}$ does not exist. In other words, we will show that there is no principal $K$-bundle over $R'(F_3,H)$ whose extension of structure group is the $K/Z$-bundle in . Assume the contrary, then we get the following diagram of topological spaces and morphisms: $$\label{eqdig2} \begin{matrix} Z & \longrightarrow & K & \longrightarrow & K/Z \\ || & {} & \Big\downarrow & & \Big\downarrow\\ Z & \longrightarrow & \widetilde{M} & \longrightarrow &q^{-1}_0( R'(F_3,H))\\ & & \Big\downarrow & & \Big\downarrow\\ & & R'(F_3,H) & = & R'(F_3,H) \end{matrix}$$ We will need a couple of results. The proof of the non-existence of the bundle $\widetilde{M}$ will be completed after establishing Proposition \[claim\]. \[newstead\] There exist finitely many compact differentiable manifolds $N_i$ and differentiable maps $$f_i \,:\, N_i\, {\longrightarrow}\, K^3\, ,$$ $1\,\leq\, i\,\leq\,d$, such that if $N \,:= \, \bigcup_{i=1}^d f_i(N_i)$ then 1. the complement $K^3 \setminus N$ is contained in $q^{-1}_0(R'(F_3,H))$, where $q_0$ is the projection in . 2. Furthermore, $\dim K^3 - \dim N_i \,\geq\, 4$ for all $i\,\in\, [1\, ,d]$. Take any homomorphism $\rho\, :\, \pi_1(X)\, \longrightarrow\, K$. Let $E_\rho$ denote the corresponding polystable principal $H$-bundle over $X$ [@Ra1]. The automorphism group of the polystable $H$-bundle $E_\rho$ coincides with the centraliser of $\rho(\pi_1(X))$ in $H$. Hence if $\rho(\pi_1(X))$ is dense in $K$, then the automorphism group of $E_\rho$ is the centre $Z(H)\, \subset\, H$. Therefore, if the topological closure $\overline{\rho(\pi_1(X))}$ of $\rho(\pi_1(X))$ is $K$, and $E_\rho$ is topologically trivial, then $E_\rho\, \in\, {\mathcal M}'_X(H)$. Now assume that the centraliser $C(\rho)\, \subset\, H$ of $\rho(\pi_1(X))$ in $H$ properly contains $Z(H)$. Since the complexification of $\overline{\rho(\pi_1(X))}$ is reductive, and the centraliser of a reductive group is reductive, we conclude that $C(\rho)$ is reductive. Take a semisimple element $z\, \in\, (H\setminus Z(H)) \bigcap C(\rho)$ (since $C(\rho)$ is reductive and larger than $Z(H)$ such an element exists). Let $C_z\, \subset\, K$ be the centraliser of $z$ in $K$. We have $\overline{\rho(\pi_1(X))}\, \subset\, C_z$, and $C_z$ is a proper subgroup of $K$ as $z\, \notin\, Z(H)$. Also, $C_z$ contains a maximal torus of $K$. Fix a maximal torus $T\, \subset\, K$. (Since any two maximal tori are conjugate, any subgroup $C_z$ of the above type would contain $T$ after an inner conjugation.) Consider all proper Lie subgroups $M$ of $K$ satisfying the following two conditions: 1. $T\, \subseteq\, M$, and 2. there exists a semisimple element $z\, \in\, H\setminus Z(H)$ such that $M$ is the centraliser of $z$ in $K$. Let $\mathcal S$ denote this collection of Lie subgroups of $K$. The connected ones among $\mathcal S$ are precisely the maximal compact subgroups of Levi subgroups of proper parabolic subgroups of $H$ containing $T$. Note that if $P$ is a proper parabolic subgroup of $H$, then $P\bigcap K$ is a maximal compact subgroup of a Levi subgroup of $P$. All the connected ones among $\mathcal S$ arise as $P\bigcap K$ for some proper parabolic subgroup $P\, \subset\, H$ containing $T$. This collection $\mathcal S$ is a finite set. To see this, we first note that there are only finitely many parabolic subgroups of $H$ that contain $T$. For a proper parabolic subgroup $P\,\subset\, H$ containing $T$, there are only finitely many Lie subgroups of $K$ that have $P\bigcap K$ as the connected component containing the identity element. Thus $\mathcal S$ is a finite set. Let $M_1, \cdots , M_d$ be the subgroups of $K$ that occur in $\mathcal S$. We will show that the codimension of each $M_i$ in $K$ is at least two. It suffices show that for a maximal proper parabolic subgroup $P$ of $G$ containing $T$, the codimension of $M\, :=\, P\bigcap K$ in $K$ is at least two. To prove that the codimension of $M\, =\, P\bigcap K$ in $K$ is at least two, let $\mathfrak k$ be the Lie algebra of $K$, and let $\mathfrak h$ be the Lie algebra of $T$ (the Cartan subalgebra). Let $${\mathfrak k} = {\mathfrak h} + \sum_{\alpha~{\rm a}~{\rm root}} {\mathfrak k}^{\alpha}$$ be the root space decomposition of $\mathfrak k$. Let $\mathfrak m$ be a Lie algebra of $M$ which is a Lie subalgebra of $\mathfrak k$ containing $\mathfrak h$. Then we have the decomposition $${\mathfrak m} \,=\, {\mathfrak h} + \sum_{\alpha~{\rm a}~{\rm root}} {\mathfrak m}^{\alpha}$$ for $\mathfrak m$ where each ${\mathfrak m}^{\alpha} $ is irreducible for $\mathfrak h$, and hence has to coincide with one of the ${\mathfrak k}^{\alpha}$. Then $\mbox{codim}_{_{\mathfrak k}}({\mathfrak m}) \geq 2$ since if $\alpha$ is a root for the subalgebra $\mathfrak m$ then so is $-{\alpha}$ (see [@Ad p. 83, Corollary 4.15]). Consequently, the codimension of $M$ in $K$ is at least two. Thus the codimension of each $M_i$, $i\, \in\, [1\, ,d]$, in $K$ is at least two. For each $i\, \in\, [1\, ,d]$, let $$\label{dbfi} \overline{f}_i \,:\, K \times M^3_i \,{\longrightarrow}\, K^3$$ be the map defined by $(x,(y_1, y_2,y_3)) \,\longmapsto\, (x y_1 {x}^{-1}, x y_w {x}^{-1}, x y_3 {x}^{-1})$. Consider the free action of $M_i$ on $K \times M^3_i$ defined by $$z \cdot (x,(y_1,y_2,y_3)) = (x z^{-1}, (z y_1 z^{-1}, z y_2 z^{-1}, z y_3 z^{-1}))\, ,$$ where $x\in K$ and $z,y_1,y_2, y_3\in M_i$. The map $f_i$ in clearly factors through the quotient $$\frac {K \times {M_i}^t}{M_i}$$ for the above action. Therefore, we have $$\label{dfi} f_i \,:\, N_i\, :=\, \frac {K \times {M_i}^t}{M_i}\, {\longrightarrow}\, K^t$$ induced by $f_i$. To prove part (1) of the lemma, we recall the earlier remark that for any homomorphism $\rho' \,\in\, q^{-1}_0(R'(F_3,H))$ (the map $q_0$ is defined in ), the automorphism group of the principal $H$-bundle corresponding to $\rho'\circ f$ (the homomorphism $f$ is defined in ) coincides with $Z(H)$ if the image of $\rho'$ is dense in $K$. From the properties of the collection $\{M_1,\cdots , M_d\}$ we conclude that $$N\, :=\, \bigcup_{i=1}^d f_i(N_i)\, \supseteq\, q^{-1}_0(R'(F_3,H))^c\, ,$$ where $f_i$ are defined in and $$q^{-1}_0(R'(F_3,H))^c\, \subset\, K^3$$ is the complement of $q^{-1}_0(R'(F_3,H))$ in $K^3$. Therefore, proof of part (1) is complete. To prove part (2), we note that $$\dim N_i \,=\, \dim K + 2\cdot \dim M_i$$ for all $i\,\in\, [1\, ,d]$. It was shown earlier that $\dim M_i \, \leq\, \dim K -2$. Therefore, $$\dim K^3 - \dim N_i \, =\, 3\cdot \dim K - \dim N_i \, \geq\, 3\cdot \dim K - 3\cdot \dim K +4 \, =\, 4\, .$$ This completes the proof of the lemma. \[claim\] Consider the $K/Z$-principal bundle $q^{-1}_0( R'(F_3,H)) \, \longrightarrow\, R'(F_3,H)$, where $q_0$ is the projection in . The induced homomorphism on fundamental groups $$\pi_1(K/Z) \,\longrightarrow\,\pi_1(q^{-1}(R'(F_3,H)))$$ obtained from the homotopy exact sequence is trivial. Let $x_1, x_2, x_3 \in K$ be regular elements; we recall that $x\in K$ is called regular if the centralizer $C(x)= \{y \in K \,\vert\, yx=xy \}$ is a maximal torus in $K$. Since the set of regular elements is dense in $K$, and $q^{-1}_0(R'(F_3,H)) \,\subset\, K^3$ is a nonempty open dense subset (this follows from Lemma \[newstead\]), we may choose these $x_i$, $i=1,2,3$, to lie in $q^{-1}_0(R'(F_3,H))$. Consider the orbit $\text{Orb}_{_{K^3}}(x_1\, , x_2\, , x_3)$, of $(x_1\, , x_2\, , x_3)\,\in\, K^3$ for the adjoint action of the group $K^3$ on itself. Clearly we have the following identification of this orbit: $$\text{Orb}_{_{K^3}}(x_1\, , x_2\, , x_3) \,= \, (K/{C(x_1)}) \times (K/{C(x_2)}) \times (K/{C(x_3)})$$ with $C({x_i})\, \subset\, K$ being the centralizer of $x_i$. The conjugation action of $K$ on $\text{Hom}(F_3,K)$ coincides with the restriction of above action of $K^3$ to the image of the diagonal map $K \, \hookrightarrow\, K^3$. Therefore, the fibre $K/Z$ through the point $(x_1, x_2, x_3)\,\in\, q^{-1}_0(R'(F_3,H))$ of the $K/Z$-bundle $$q^{-1}_0( R'(F_3,H)) \,\longrightarrow\, R'(F_3,H)$$ is contained in the orbit $$(K/{C(x_1)}) \times (K/{C(x_2)}) \times (K/{C(x_3)})\, .$$ Now if we choose a point $(x_1, x_2, x_3) \in K^3$ which is general enough, then by the definition of the inverse image $q^{-1}_0( R'(F_3,H))$ and Lemma \[newstead\](2) it follows immediately that the complement of the open dense subset $$q^{-1}_0(R'(F_3,H))\bigcap (\frac{K}{C(x_1)} \times \frac{K}{C(x_2)} \times \frac{K}{C(x_3)})\,\subset\, \frac{K}{C(x_1)} \times\frac{K}{C(x_2)}\times\frac{K}{C(x_3)}$$ is of codimension at least four. Since the image of $K/Z$ in $q^{-1}_0( R'(F_3,H))$ lies in $(q^{-1}( R'(F_3,H)))\bigcap \text{Orb}_{_{K^3}}(x_1, x_2, x_3)$, whose complement is of codimension at least four in $$\text{Orb}_{_{K^3}}(x_1, x_2, x_3) \,=\, (K/{C(x_1)}) \times (K/{C(x_2)}) \times (K/{C(x_3)})$$ it follows that the homomorphism $ \pi_1(K/Z) \,\longrightarrow\, \pi_1(q^{-1}_0 (R'(F_3,H)))$ in the proposition factors through $$\pi_1((q^{-1}_0( R'(F_3,H))) \cap \text{Orb}_{_{K^3}}(x_1, x_2, x_3))$$ $$=\, \pi_1((K/{C(x_1)}) \times (K/{C(x_2)}) \times (K/{C(x_3)})) \,= \,\pi_1(K/T)^3$$ where $T$ is a maximal torus of $K$. For any maximal torus $T\, \subset\, K$, the quotient $K/T$ is diffeomorphic to $H/B$, where $B$ is a Borel subgroup of $H$. Since $H/B$ is simply connected, we conclude that $(K/{C(x_1)}) \times (K/{C(x_2)}) \times (K/{C(x_3)})$ is simply connected. This completes the proof of the proposition. [*We now complete the proof of the non-existence of the covering ${\widetilde M} \longrightarrow q^{-1}( R'(F_3,H))$ as in .*]{} Since the homomorphism of fundamental groups $ \pi_1(K/Z) \,\longrightarrow\, \pi_1(q^{-1}_0( R'(F_3,H)))$ is trivial (see Proposition \[claim\]), the induced covering $K\,\longrightarrow\, K/Z$ is trivial (see the diagram ). But this is a contradiction to the fact that $K$ is connected. Therefore, we have proved the following theorem: \[nonexist\] If the centre $Z(H)$ is non-trivial, then there is no universal bundle over $X\times {\mathcal M}'_X(H)$. [11111]{} J. F. Adams, *Lectures on Lie Groups*, W. A. Benjamin, Inc, Amsterdam, 1969. V. Balaji and C. S. Seshadri, *Semistable Principal Bundles-I (in char 0)*, Jour. Alg. **258** (2002), 321–347. A. Borel, *Linear algebraic groups,* Second edition, Graduate Texts in Mathematics, 126, Springer–Verlag, New York, 1991. J.-M. Drézet and M. S. Narasimhan, *Groupes de Picard des variétés de modules de fibrés semi-stables sur les courbes algébriques*, Invent. Math. [**97**]{} (1989), 53–94. G. Faltings, *Stable $G$-bundles and projective connections*, Jour. Alg. Geom. [**2**]{} (1993), 507–568. S. Kumar, M. S. Narasimhan and A. Ramanathan, *Infinite Grassmannians and moduli spaces of $G$-bundles*, Math. Ann. [**300**]{} (1994), 41–75. P. E. Newstead, *A non-existence theorem for families of stable bundles*, Jour. Lond. Math. Soc. **6** (1973), 259–266. S. Ramanan, *The moduli spaces of vector bundles over an algebraic curve*, Math. Ann. **200** (1973), 69–84. A. Ramanathan, *Stable principal bundles on a compact Riemann surface*, Math. Ann. **213** (1975), 129–152. A. Ramanathan, *Moduli of principal bundles over algebraic curves: I*, Proc. Ind. Acad. Sci. (Math. Sci.) **106** (1996), 301–328. A. Ramanathan, *Moduli of principal bundles over algebraic curves: II*, Proc. Ind. Acad. Sci. (Math. Sci.) **106** (1996), 421–449. S. Subramanian, *Mumford’s example and a general construction*, Proc. Ind. Acad. Sci. (Math. Sci.) **99** (1989), 197–208. A. Tyurin, *Analogues of Torelli’s theorem for multidimensional vector bundles over an arbitrary algebraic curve*, Izv. Akad. Nauk SSSR Ser. Mat. **34** (1970), 338–365. English translation: Math. USSR – Izv. **4** (1970), 343–370. [^1]: All authors are members of the international research group VBAC (Vector Bundles on Algebraic Curves), which is partially supported by EAGER (EC FP5 Contract no. HPRN-CT-2000-00099) and by EDGE (EC FP5 Contract no. HPRN-CT-2000-00101). The fourth author would also like to thank the Royal Society and CSIC, Madrid, for supporting a visit to CSIC during which his contribution to this work was completed.

--- abstract: 'For $C^*$-algebras $A_1, A_2$ the map $(I_1,I_2)\to \mathrm{ker}(q_{I_1}\otimes q_{I_2})$ from $\mathrm{Id}^{\prime}(A_1)\times \mathrm{Id}^{\prime}(A_2)$ into $\mathrm{Id}^{\prime}(A_1\otimes_{\mathrm{min}} A_2)$ is a homeomorphism onto its image which is dense in the range. Here, for a $C^*$-algebra $A$, the space of all proper closed two sided ideals endowed with an adequate topology is denoted $\mathrm{Id}^{\prime}(A)$ and $q_I$ is the quotient map of $A$ onto $A/I$. New proofs of the equivalence of the property (F) of Tomiyama for $A_1\otimes_{\mathrm{min}} A_2$ with certain other properties are presented.' address: | School of Mathematical Sciences\ Tel Aviv University\ Tel Aviv 69778, Israel author: - 'Aldo J. Lazar' date: 'March 5, 2009' title: 'The space of ideals in the minimal tensor product of $C^*$-algebras' --- [^1] ———————————– Introduction {#S:I} ============ The relationship between the representations and the kernels of representations of the minimal tensor product of two $C^*$-algebras and those of the factors has been a substantial topic of study since the inception of this method of building new algebras, as one can see for instance from [@W], [@T], or [@G]. Starting with closed two sided ideals $I_1, I_2$ of the $C^*$-algebras $A_1, A_2$, respectively, one looks at the ideals $I := I_1\otimes_{\mathrm{min}} A_2 + A_1\otimes_{\mathrm{min}} I_2$ of $A_1\otimes_{\mathrm{min}} A_2$ and $\mathrm{ker}(q_{I_1}\otimes q_{I_2})$ which obviously contains $I$. In particular, it is interesting to see what happens when $I_1$ and $I_2$ are primitive or prime ideals. It is also of interest to know if the map between the pairs of ideals and the ideals of the tensor product has any continuity properties. When the primitive ideal spaces are given their hull-kernel topologies it was proved in [@W] that the map $(I_1,I_2)\to \mathrm{ker}(q_{I_1}\otimes q_{I_2})$ is a homeomorphism onto its image. Similar results, under the assumption that the tensor product enjoys the property (F) of Tomiyama, were obtained in [@H] and [@BK] for the spaces of the kernels of factorial representations and the spaces of prime ideals. We consider here the spaces of all the proper closed two-sided ideals with their $\tau_w$ topologies, as defined in [@A], and show in Theorem \[T:home\] that the same map is a homeomorphism. This establishes, in the most general situation, that the maps between the spaces of kernels of factorial representations and between the spaces of prime ideals, with their hull-kernel topologies, are also homeomorphisms. In section \[S:F\] we give new proofs, possibly simpler, to the equivalence of the property (F) with various other properties of the minimal tensor product. For instance, we show that property (F) is equivalent to having every closed two sided ideal of the minimal tensor product equal to the closed linear span of the elementary tensors contained in the ideal, a fact first proved in [@BK Proposition 2.16]. In [@Al Proposition 4.5] it is proved that every non-zero closed two sided ideal of any minimal tensor product of two $C^*$-algebras must contain a non trivial elementary tensor and this fact plays an important role in our proofs, via [@BK Lemma 2.12]. By an ideal of a Banach algebra we shall always mean a closed two sided ideal. The set of all the ideals of the Banach algebra $A$ is denoted by $\mathrm{Id}(A)$ and $\mathrm{Id}^{\prime}(A) := \mathrm{Id}(A)\setminus \{A\}$. For an ideal $I$ of a Banach algebra $A$ we denote by $q_I$ the quotient homomorphism of $A$ onto $A/I$. Given a $C^*$-algebra $A$, the set of its primitive ideals, the set of the kernels of its factorial representations, and the set of its prime ideals are denoted ${\mathrm{Prim}}(A)$, ${\mathrm{Fac}}(A)$, and ${\mathrm{Prime}}(A)$, respectively. Their topology will everywhere be the hull-kernel topology. Given a topological space $X$, we shall endow the space of all its closed subsets, ${\mathcal{F}}(X)$, with the topology for which a subbase is the collection of all the families $\{F\in {\mathcal{F}}(X)\mid F\cap O\neq {\emptyset}\}$, when $O$ runs through all the open subsets of $X$. If $A$ is a $C^*$-algebra the $\tau_w$ topology of $\mathrm{Id}(A)$ is defined by transporting the topology of ${\mathcal{F}}({\mathrm{Prim}}(A))$ to $\mathrm{Id}(A)$ via the well known correspondence between the closed subsets of ${\mathrm{Prim}}(A)$ and the ideals of $A$. Thus a subbase for this topology is given by all the sets $\{I\in \mathrm{Id}(A)\mid J\nsubseteq I\}$ where $J$ is an ideal of $A$, see [@A p.525]. $A_1$, $A_2$ being $C^*$-algebras, $A_1\odot A_2$ stands for their algebraic tensor product and $A_1\otimes A_2$ for their minimal tensor product which will be the only $C^*$-tensor product discussed here. If $B_1$, $B_2$ are also $C^*$-algebras and $\pi_1 : A_1\to B_1$, $\pi_2 : A_2\to B_2$ are homomorphisms then $\pi_1\otimes \pi_2 ; A_1\otimes A_2\to B_1\otimes B_2$ is the unique homomorphism determined by $(\pi_1\otimes \pi_2)(a_1\otimes a_2) = \pi_1(a_1)\otimes \pi_2(a_2)$ for every $a_1\in A_1$, $a_2\in A_2$. The map $\Phi : \mathrm{Id}(A_1)\times \mathrm{Id}(A_2) \to \mathrm{Id}(A_1\otimes A_2)$ is defined by $\Phi(I_1,I_2) := \mathrm{ker}(q_{I_1}\otimes q_{I_2})$. Very useful for the study of the structure of $A_1\otimes A_2$ are the slice maps defined by Tomiyama in [@T]. For a bounded linear functional $f$ on $A_2$, the bounded linear left slice map $L_f : A_1\otimes A_2\to A_1$ is defined by requiring $L_f(a_1\otimes a_2) := f(a_2)a_1$ for each $a_1\in A_1$, $a_2\in A_2$. Following [@T] one says that $A_1\otimes A_2$ has the property (F) if the set of all product states $f_1\otimes f_2$, $f_1$ and $f_2$ being pure states of $A_1$ and $A_2$, respectively, separates the ideals of $A_1\otimes A_2$. Equivalently, by [@T Theorem 5], for every ideal $I_1$ of $A_1$ and $I_2$ of $A_2$ one has $\Phi(I_1,I_2) = I_1\otimes A_2 + A_1\otimes I_2$. It is this property of the ideals that we shall use everywhere in the paper when we refer to the property (F). The author is very grateful to Professor R. J. Archbold for pointing out a mistake in an earlier version of this paper at the same time with providing the way to correct it and for suggesting many other improvements. The space of ideals of $A_1\otimes A_2$ {#S:Id} ======================================= \[L:hyper\] Let $X_1, X_2$ be topological spaces. The map $\varphi : {\mathcal{F}}(X_1)\times {\mathcal{F}}(X_2)\to {\mathcal{F}}(X_1\times X_2)$ given by $\varphi(F_1,F_2) := F_1\times F_2$ is continuous. Let $U$ be an open subset of $X_1\times X_2$, $U = \cup (U^1_{\iota}\times U^2_{\iota})$ with $U^j_{\iota}$ open subsets of $X_j$. Then $$\begin{gathered} \{(F_1,F_2)\in {\mathcal{F}}(X_1)\times {\mathcal{F}}(X_2) \mid \varphi(F_1,F_2)\cap U\neq {\emptyset}\} =\\ \cup (\{F_1\in {\mathcal{F}}(X_1) \mid F_1\cap U^1_{\iota} \neq {\emptyset}\}\times \{F_2\in {\mathcal{F}}(X_2) \mid F_2\cap U^2_{\iota} \neq {\emptyset}\}). \end{gathered}$$ This is an open subset of ${\mathcal{F}}(X_1)\times {\mathcal{F}}(X_2)$ and the continuity of $\varphi$ is established. From here on we let $A_1, A_2$ be $C^*$-algebras and $\Phi$ will be as defined in the introduction. Then $\Phi$ maps ${\mathrm{Prim}}(A_1)\times {\mathrm{Prim}}(A_2)$ into ${\mathrm{Prim}}(A_1\otimes A_2)$ by [@W], $\Phi({\mathrm{Fac}}(A_1)\times {\mathrm{Fac}}(A_2))\subset {\mathrm{Fac}}(A_1\otimes A_2)$ by [@G p. 6], and $\Phi({\mathrm{Prime}}(A_1)\times {\mathrm{Prime}}(A_2))\subset {\mathrm{Prime}}(A_1\otimes A_2)$ by [@BK Lemma 2.13(v)]. \[L:intersect\] Let $\mathcal{M}_k$ be a family of ideals of $A_k$ and $$I_k := \cap_{J\in \mathcal{M}_k} J$$ for $k = 1,2$. Then $\Phi(I_1,I_2) = \cap \{\Phi(J_1,J_2) \mid (J_1,J_2)\in \mathcal{M}_1\times \mathcal{M}_2\}$. For $J_1\in \mathcal{M}_1$ and $J_2\in \mathcal {M}_2$ there exists a natural homomorphism $\theta_{J_1,J_2}$ of $(A_1/I_1)\otimes (A_2/I_2)$ onto $(A_1/J_1)\otimes (A_2/J_2)$; one defines it in the obvious way on $(A_1/I_1)\odot (A_2/I_2)$ and then one extends it by continuity. This homomorphism satisfies $\theta_{J_1,J_2}\circ (q_{I_1}\otimes q_{I_2}) = q_{J_1}\otimes q_{J_2}$ and from this we infer $\Phi(I_1,I_2)\subset \Phi(J_1,J_2)$. Define a seminorm on $(A_1/I_1)\otimes (A_2/I_2)$ by $$N(x) := sup\{\|\theta_{J_1,J_2}(x)\| \mid (J_1,J_2)\in \mathcal{M}_1\times \mathcal{M}_2\}.$$ Then, if $x_1\in A_1/I_1$ and $x_2\in A_2/I_2$ with $x_1\otimes x_2\neq 0$ we have $N(x_1\otimes x_2) > 0$. By [@BK Lemma 2.12(i)] we have $N(x)\geq \|x\|$ for $x\in (A_1/I_1)\otimes (A_2/I_2)$. Hence, if for $a\in A_1\otimes A_2$ we have $(q_{J_1}\otimes q_{J_2})(a) = 0$ for each $(J_1,J_2)\in \mathcal{M}_1\times \mathcal{M}_2$ then $a\in \Phi(I_1,I_2)$ and we have proved $\Phi(I_1,I_2)\supset \cap \{\Phi(J_1,J_2) \mid (J_1,J_2)\in \mathcal{M}_1\times \mathcal{M}_2\}$. \[C:hulls\] Let $I_1, I_2$ be ideals in $A_1, A_2$ respectively. Then $\Phi(\mathrm{hull} I_1\times \mathrm{hull} I_2)$ is dense in $\mathrm{hull} \Phi(I_1,I_2)$. $I_k = \cap_{J\in \mathrm{hull} I_k} J$ hence, by Lemma \[L:intersect\], $$\Phi(I_1,I_2) = \cap \{\Phi(J_1,J_2) \mid (J_1,J_2)\in \mathrm{hull} I_1\times \mathrm{hull} I_2\}$$ and from this the conclusion follows. Let us denote by $\Phi^{\prime}$ the restriction of $\Phi$ to ${\mathrm{Prim}}(A_1)\times {\mathrm{Prim}}(A_2)$. Recall that $\Phi^{\prime}$ is a homeomorphism of ${\mathrm{Prim}}(A_1)\times {\mathrm{Prim}}(A_2)$ onto its image which is dense in ${\mathrm{Prim}}(A_1\otimes A_2)$, see [@W Lemme 16]. \[R:Phi\] Corollary \[C:hulls\] tells us that for $(I_1,I_2)\in \mathrm{Id}(A_1)\times \mathrm{Id}(A_2)$ we have $\Phi(I_1,I_2) = \cap \{\Phi^{\prime}(P_1,P_2) \mid (P_1,P_2)\in \mathrm{hull} I_1\times \mathrm{hull} I_2\}$. Now we are ready to prove the continuity of the map $\Phi$. This was proved in [@K Lemma 1.5] under the assumption that $A_1\otimes A_2$ has property $(F)$. \[L:cont\] The map $\Phi$ is continuous. Let $U$ be an open subset of $\mathrm{Prim}(A_1\otimes A_2)$. Then, by Corollary \[C:hulls\], $$\begin{gathered} \{(I_1,I_2)\in \mathrm{Id}(A_1)\times \mathrm{Id}(I_2) \mid \mathrm{hull} \Phi(I_1,I_2)\cap U \neq {\emptyset}\} =\\ \{(I_1,I_2)\in \mathrm{Id}(A_1)\times \mathrm{Id}(A_2) \mid \Phi(\mathrm{hull} I_1\times \mathrm{hull} I_2)\cap U \neq {\emptyset}\} =\\ \{(I_1,I_2)\in \mathrm{Id}(A_1)\times \mathrm{Id}(A_2) \mid (\mathrm{hull} I_1\times \mathrm{hull} I_2)\cap \Phi^{\prime-1}(U) \neq {\emptyset}\} \end{gathered}$$ and the latter set is open in $\mathrm{Id}(A_1)\times \mathrm{Id}(A_2)$ by the continuity of $\Phi^{\prime}$ and Lemma \[L:hyper\]. It has been pointed out to the author by Professor R. J. Archbold that Lemma \[L:cont\] yields an alternative proof of Lemma 1.4 in [@ASKS] that asserts the continuity of the canonical map from $\mathrm{Id}(A_1)\times \mathrm{Id}(A_2)$ to the space of ideals of $A_1\otimes_h A_2$, the Haagerup tensor product of the $C^*$-algebras $A_1$ and $A_2$. It is shown in [@ASKS Lemma 1.1] that this map is obtained by following the map $\Phi$ with the map $K\to K\cap (A_1\otimes_h A_2)$ from $\mathrm{Id}(A_1\otimes A_2)$ to $\mathrm{Id}(A_1\otimes_h A_2)$. The continuity of the latter map is established on p. 5 of [@ASKS] and the conclusion of [@ASKS Lemma 1.4] follows. \[T:home\] The map $\Phi$ is a homeomorphism of $\mathrm{Id}^{\prime}(A_1)\times \mathrm{Id}^{\prime}(A_2)$ onto its image which is dense in $\mathrm{Id}^{\prime}(A_1\otimes A_2)$. There is a map $\Psi$ from $\mathrm{Id}(A_1\otimes A_2)$ to $\mathrm{Id}(A_1)\times \mathrm{Id}(A_2)$ defined as follows: for $I\in \mathrm{Id}(A_1\otimes A_2)$ set $$I_{A_1} := \{a_1\in A_1 \mid a_1\otimes A_2\subset I\}, \quad I_{A_2} := \{a_2\in A_2 \mid A_1\otimes a_2\subset I\}$$ and $\Psi(I) := (I_{A_1},I_{A_2})$, see [@BK Lemma 2.13] for instance. Then $\Psi\circ \Phi$ restricted to $\mathrm{Id}^{\prime}(A_1)\times \mathrm{Id}^{\prime}(A_2)$ is $\mathrm{id}$, the identity map of $\mathrm{Id}^{\prime}(A_1)\times \mathrm{Id}^{\prime}(A_2)$. Indeed, let $(I_1,I_2)\in \mathrm{Id}^{\prime}(A_1)\times \mathrm{Id}^{\prime}(A_2)$ and set $I := \Phi(I_1,I_2)$. Clearly $I_1\subset I_{A_1}$ and $I_2\subset I_{A_2}$. Let now $a\in I_{A_1}$ and pick $b\in A_2\setminus I_2$. Then $a\otimes b\in I$ hence by [@BK Lemma 2.12(iii)], $a\otimes b\in I_1\otimes A_2 + A_1\otimes I_2$. Choose a bounded linear functional $f$ on $A_2$ that vanishes on $I_2$ and satisfies $f(b) = 1$. By using the left slice map $L_f$ we get $a = L_f(a\otimes b)\in I_1$. We obtained $I_1 = I_{A_1}$ and $I_2 = I_{A_2}$ is proved similarly, thus the claim is established. In particular, it follows that the restriction of $\Phi$ to the space considered is one to one. It is well known that $\Psi$ is continuous but we include a proof for completeness since we could find in print only proofs of the continuity of some restrictions of $\Psi$. Let $V_n$ be open subsets of ${\mathrm{Prim}}(A_n)$ and $K_n$ the corresponding ideals of $A_n$, and denote $\mathcal{U}_n := \{I_n\in \mathrm{Id}(A_n) \mid \mathrm{hull} I_n\cap V_n \neq {\emptyset}\}$, $n = 1,2$. Then $$\begin{gathered} \Psi^{-1}(\mathcal{U}_1\times \mathcal{U}_2) = \{I\in \mathrm{Id}(A_1\otimes A_2) \mid (\mathrm{hull} I_{A_n})\cap V_n \neq {\emptyset}, n = 1,2\} =\\ \{I\in \mathrm{Id}(A_1\otimes A_2) \mid I_{A_n}\nsupseteq K_n, n = 1,2\} =\\ \{I\in \mathrm{Id}(A_1\otimes A_2) \mid I\nsupseteq K_1\otimes A_2\}\cap \{I\in \mathrm{Id}(A_1\otimes A_2) \mid I\nsupseteq A_1\otimes K_2\} \end{gathered}$$ and the latter is an open subset of $\mathrm{Id}(A_1\otimes A_2)$. From $\Psi\circ \Phi = id$ and the continuity of $\Psi$ it follows that $\Phi$ maps open subsets of $\mathrm{Id}^{\prime}(A_1)\times \mathrm{Id}^{\prime}(A_2)$ onto relatively open sets and the proof that $\Phi$ is a homeomorphism onto its image is complete. Let now $\mathcal{W}_r$, $1\leq r\leq m$ be open subsets of ${\mathrm{Prim}}(A_1\otimes A_2)$. There are $\{P^1_r\}^m_1\subset {\mathrm{Prim}}(A_1)$ and $\{P^2_r\}^m_1\subset {\mathrm{Prim}}(A_2)$ such that $\Phi^{\prime}(P^1_r,P^2_r)\in \mathcal{W}_r$, $1\leq r\leq m$. If we set $I_1 := \cap^m_{r=1} P^1_r$ and $I_2 := \cap^m_{r=1} P^2_r$ then $$\Phi(I_1,I_2)\in \{I\in \mathrm{Id}(A_1\otimes A_2) \mid \mathrm{hull} I\cap \mathcal{W}_r\neq {\emptyset}, \quad 1\leq r\leq m\}$$ by Remark \[R:Phi\] and this shows that the image of $\Phi$ is dense in $\mathrm{Id}^{\prime}(A_1\otimes A_2)$. \[C:Fac\] The restriction of $\Phi$ to ${\mathrm{Fac}}(A_1)\times {\mathrm{Fac}}(A_2)$ is a homeomorphism onto a dense subset of ${\mathrm{Fac}}(A_1\otimes A_2)$. \[C:Pr\] The restriction of $\Phi$ to ${\mathrm{Prime}}(A_1)\times {\mathrm{Prime}}(A_2)$ is a homeomorphism onto a dense subset of ${\mathrm{Prime}}(A_1\otimes A_2)$. The density of the images in the above corollaries follows from the density of $\Phi^{\prime}({\mathrm{Prim}}(A_1)\times {\mathrm{Prim}}(A_2))$ in ${\mathrm{Prim}}(A_1\times A_2)$. Under the assumption that $A_1\otimes A_2$ has property $(F)$, Corollary \[C:Fac\] appears as Lemma 2.5 in [@H] and Corollary \[C:Pr\] is part of Proposition 2.16 of [@BK]. Property $(F)$ ensures that the restrictions of $\Phi$ are onto maps, see also Proposition \[P:F\] below. The property $(F)$ {#S:F} ================== We continue to use the same notation as in the previous section: $A_1$ and $A_2$ are $C^*$-algebras, $\Phi : \mathrm{Id}(A_1)\times \mathrm{Id}(A_2)\to \mathrm{Id}(A_1\otimes A_2)$ is the map defined in the introduction and $\Psi : \mathrm{Id}(A_1\otimes A_2)\to \mathrm{Id}(A_1)\times \mathrm{Id}(A_2)$ is the map defined in the proof of Theorem \[T:home\]. Then $\Psi({\mathrm{Fac}}(A_1\otimes A_2))\subset {\mathrm{Fac}}(A_1)\times {\mathrm{Fac}}(A_2)$ by [@G Proposition 1] and $\Psi({\mathrm{Prime}}(A_1\otimes A_2))\subset {\mathrm{Prime}}(A_1)\times {\mathrm{Prime}}(A_2)$ by [@BK Lemma 2.13(i)]. It is not known if $\Psi$ maps ${\mathrm{Prim}}(A_1\otimes A_2)$ into ${\mathrm{Prim}}(A_1)\times {\mathrm{Prim}}(A_2)$, even if $A_1\otimes A_2$ has the property $(F)$. We shall use also the map $\Delta : \mathrm{Id}(A_1)\times \mathrm{Id}(A_2)\to \mathrm{Id}(A_1\otimes A_2)$ defined by $\Delta(I_1,I_2) := I_1\otimes A_2 + A_1\otimes I_2$. By using slice maps one easily sees that $\Delta(\mathrm{Id}^{\prime}(A_1)\times \mathrm{Id}^{\prime}(A_2))\subset \mathrm{Id}^{\prime}(A_1\otimes A_2)$. From [@BK Lemma 2.12(iv)] we infer that the map $\Delta$ is one to one on $\mathrm{Id}^{\prime}(A_1)\times \mathrm{Id}^{\prime}(A_2)$. In the the result below we collect several conditions that are equivalent to the property $(F)$. All of them appear, possibly in a slightly different form, either in [@H Proposition 2.2] or in [@BK Proposition 2.16] and we only provide new proofs to various implications. \[P:F\] The following properties of $A_1\otimes A_2$ are equivalent: - $A_1\otimes A_2$ has property $(F)$. - For every prime ideal $I$ of $A_1\otimes A_2$ we have $I = \Phi(\Psi(I))$. - $\Phi({\mathrm{Prime}}(A_1)\times {\mathrm{Prime}}(A_2)) = {\mathrm{Prime}}(A_1\otimes A_2)$. - $\Delta({\mathrm{Prime}}(A_1)\times {\mathrm{Prime}}(A_2)) = {\mathrm{Prime}}(A_1\otimes A_2)$. - For every $I\in {\mathrm{Fac}}(A_1\otimes A_2)$ we have $I = \Phi(\Psi(I))$. - $\Phi({\mathrm{Fac}}(A_1)\times {\mathrm{Fac}}(A_2)) = {\mathrm{Fac}}(A_1\otimes A_2)$. - $\Delta({\mathrm{Fac}}(A_1)\times {\mathrm{Fac}}(A_2)) = {\mathrm{Fac}}(A_1\otimes A_2)$. - Every ideal $I$ of $A_1\otimes A_2$ is the closure of the linear span of the elementary tensors contained in $I$. \(1) $\Rightarrow$ (2). Let $I\in {\mathrm{Prime}}(A_1\otimes A_2)$ and set $(I_1,I_2) = \Psi(I)$. Then $I\subset \Phi(I_1,I_2) = I_1\otimes A_2 + A_1\otimes I_2$ by [@BK Lemma 2.13(iv)] and our assumption. From the definition of $\Psi$ follows $I_1\otimes A_2 + A_1\otimes I_2\subset I$ so we obtained $I = \Phi(\Psi(I))$. \(2) $\Rightarrow$ (3). $\Phi({\mathrm{Prime}}(A_1)\times {\mathrm{Prime}}(A_2))\subset {\mathrm{Prime}}(A_1\otimes A_2)$ by [@BK Lemma 2.13(v)] and $\Psi({\mathrm{Prime}}(A_1\otimes A_2))\subset {\mathrm{Prime}}(A_1)\times {\mathrm{Prime}}(A_2)$ by [@BK Lemma 2.13(i)] so (3) follows immediately from (2). \(3) $\Rightarrow$ (2). Let $I\in {\mathrm{Prime}}(A_1\otimes A_2)$. Then $I = \Phi(I_1,I_2)$ with $(I_1,I_2)\in {\mathrm{Prime}}(A_1)\times {\mathrm{Prime}}(A_2)$. As shown in the proof of Theorem \[T:home\], $\Psi\circ \Phi$ is the identity map of $\mathrm{Id}^{\prime}(A_1)\times \mathrm{Id}^{\prime}(A_2)$ hence $\Psi(I) = (I_1,I_2)$ and (2) obtains. \(2) $\Rightarrow$ (4) is trivial and (4) $\Rightarrow$ (1) is part of [@T Theorem 5]. Now, clearly (1) and (3) imply (3a). (3a) $\Rightarrow$ (2). Let $I$ be a prime ideal of $A_1\otimes A_2$ and set $\Psi(I) = (I_1,I_2)\in {\mathrm{Prime}}(A_1)\times {\mathrm{Prime}}(A_2)$, by [@BK Lemma 2.13(i)]. Then $J := \Phi(I_1,I_2)\in {\mathrm{Prime}}(A_1\otimes A_2)$ by [@BK Lemma 2.13(v)] and there exist $J_1\in {\mathrm{Prime}}(A_1), \; J_2\in {\mathrm{Prime}}(A_2)$ such that $\Delta(J_1,J_2) = J = \Phi(I_1,I_2)$. Then, again by [@BK Lemma 2.13(v)], $J_1 = I_1, \; J_2 = I_2$. Now, $\Delta(I_1,I_2)\subset I\subset \Phi(I_1,I_2) = \Delta(I_1,I_2)$, the second inclusion following from [@BK Lemma 2.13(iv)]. Thus we got $I = \Phi(I_1,I_2) = \Phi(\Psi(I))$. \(4) $\Leftrightarrow$ (5) is obtained as (2) $\Leftrightarrow$ (3) above by using [@G p. 6 and Proposition 1]. (1), (5) $\Rightarrow$ (5a) is obvious. (5a) $\Rightarrow$ (4) can be deduced, with obvious changes, as (3a) $\Rightarrow$ (2). (1), (3a) $\Rightarrow$ (6). For a $C^*$-algebra $A$ and an ideal $K$ of $A$ we shall denote by ${\mathrm{Prime}}(K)$ the open subset of ${\mathrm{Prime}}(A)$ consisting of all the prime ideals that do not contain $K$. Let $I$ be an ideal of $A_1\otimes A_2$ and pick $P\in {\mathrm{Prime}}(I)$. We claim that there exist ideals $J_1$ of $A_1$ and $J_2$ of $A_2$ such that $P\in {\mathrm{Prime}}(J_1\otimes J_2)$ and ${\mathrm{Prime}}(J_1\otimes J_2)\subset {\mathrm{Prime}}(I)$. Once this claim is established then we can assert that the ideal $I^{\prime}$ generated by all the elementary tensors contained in $I$ satisfies ${\mathrm{Prime}}(I^{\prime}) = {\mathrm{Prime}}(I)$ hence $I^{\prime} = I$ and we are done. From the assumption on $A_1\otimes A_2$ and Corollary \[C:Pr\] it follows that $\Phi = \Delta$ and there are open subsets $U_m$ of ${\mathrm{Prime}}(A_m)$ such that $P\in \Delta(U_1\times U_2)$ and $\Delta(U_1\times U_2)\subset {\mathrm{Prime}}(I)$. Denote by $J_m$ the ideal of $A_m$ that satisfies ${\mathrm{Prime}}(J_m) = U_m$. To substantiate the above claim it remains to show that ${\mathrm{Prime}}(J_1\otimes J_2) = \Delta(U_1\times U_2)$. Now, if $Q\in {\mathrm{Prime}}(J_1\otimes J_2)$ that is, $Q\in {\mathrm{Prime}}(A_1\otimes A_2)$ but $Q\nsupseteq J_1\otimes J_2$, and $Q = \Delta(Q_1,Q_2) = Q_1\otimes A_2 + A_1\otimes Q_2$ with $(Q_1,Q_2)\in {\mathrm{Prime}}(A_1)\times {\mathrm{Prime}}(A_2)$ then $Q_1\nsupseteq J_1$ and $Q_2\nsupseteq J_2$. Hence ${\mathrm{Prime}}(J_1\otimes J_2)\subseteq \Delta(U_1\times U_2)$. On the other hand, we are going to show that if $Q\in {\mathrm{Prime}}(A_1\otimes A_2)$ contains $J_1\otimes J_2$ and $Q_1, Q_2$ are the prime ideals of $A_1, A_2$, respectively, that satisfy $Q = \Delta(Q_1,Q_2)$ then $J_1\subset Q_1$ or $J_2\subset Q_2$. So suppose $J_1\otimes J_2\subset Q_1\otimes A_2 + A_1\otimes Q_2$ and assume that there exists $a_2\in J_2\setminus Q_2$. Choose a bounded linear functional $f$ on $A_2$ that vanishes on $Q_2$ but $f(a_2) = 1$. Then if $a_1\in J_1$, we have $a_1 = L_f(a_1\otimes a_2)\subset Q_1$ hence $J_1\subset Q_1$. \(6) $\Rightarrow$ (1). Let $I_m$ be an ideal of $A_m$. By our assumption on the ideals of $A_1\otimes A_2$ and [@BK Lemma 2.12(iii)] we have $\Phi(I_1,I_2) = \Delta(I_1,I_2)$. [99]{} S. D. Allen, A. M. Sinclair, and R. R. Smith, *The ideal structure of the Haagerup tensor product of $C^*$-algebras*, J. Reine Angew. Math. **442** (1993), 111–148. R. J. Archbold, *Topologies for primal ideals*, J. London Math. Soc.(2) **37** (1987), 524–542. R. J. Archbold, D. W. B. Somerset, E. Kaniuth, and G. Schlichting, *Ideal spaces of the Haagerup tensor product of $C^*$-algebras*, International J. Math. **8** (1995), 1–29. E. Blanchard and E. Kirchberg, *Non-simple purely infinite $C^*$-algebras: the Hausdorff case*, J. Funct. Anal. **207** (2004), 461–513. A. Guichardet, *Tensor products of $C^*$-algebras, part I. Finite tensor products*, Math. Inst. Aarhus Univ. Lecture Notes **12** (1969). W. Hauenschild, E. Kaniuth, and A. Voigt, *\*-Regularity and uniqueness of $C^*$-norm for tensor products of \*-algebras*, J. Funct Anal. **89** (1990), 137–149. E. Kaniuth, *Minimal primal ideal spaces and norms of inner derivations of tensor products of $C^*$-algebras*, Math. Proc. Camb. Phil. Soc. **119** (1996), 297–308. J. Tomiyama, *Applications of Fubini type theorems to the tensor product of $C^*$-algebras*, Tôhoku Math. J. **19** (1967), 213–226. A. Wulfsohn, *Produit tensoriel de $C^*$-algèbres*, Bull. Sci. Math. **87** (1963), 13–27. [^1]:

--- abstract: '$\pi$-$N$ and $\pi$-$\pi$ interactions near threshold are uniquely sensitive to the chiral symmetry breaking part of the strong interaction. The $\pi$-$N$ $\sigma$-term value with its implications for nucleon quark structure and the recent controversy concerning the size of the scalar quark condensate have renewed the experimental interest in these two fundamental systems. We report new differential cross sections for the reaction $\pi^-p$$\to$$\pi^0n$ at 27.5 MeV pion incident kinetic energy, measured between $\theta_{\rm CM}$ = 0$^\circ$ and 55$^\circ$. Our results are in excellent agreement with the existing comprehensive $\pi N$ phase shift analysis. We also report on a Chew-Low analysis of exclusive $\pi^+p$$\to$$\pi^+\pi^0p$ data at 260 MeV pion incident energy.' author: - | D. Počanić and E. Frlež\ Department of Physics, University of Virginia, Charlottesville, VA 22901, U.S.A. title: '$\pi$-$N$ charge exchange and $\pi^+$-$\pi^0$ scattering at low energies[^1]' --- $\pi$-$N$ CHARGE EXCHANGE AT 27.5 MeV ===================================== While the basic mechanism of spontaneous breaking of chiral symmetry is reasonably well in hand, certain aspects of the explicit breaking of chiral symmetry ($\chi$SB), due to nonzero quark masses, remain not fully resolved to date. In the $\pi$-$N$ system at low energies the quantities of interest are the chiral symmetry breaking “sigma term” and the scattering lengths. In particular, the $\sigma$-term has been found to have an unexpectedly large value (for the most recent comprehensive analysis see Ref. [@Gas-91]). The discrepancy between the $\sigma$-term values obtained from the baryon mass splitting and from extrapolation of the isospin-even $\pi$-$N$ scattering amplitude has been attributed to a nonzero $\bar{s}s$ content of the nucleon [@Gas-82]. $\pi$-$N$ scattering lengths are related quantities that provide an independent check of the chiral lagrangians. Thus, low energy $\pi$-$N$ interactions have retained a fundamental significance and interest over the years. Unfortunately, inconsistencies in the existing $\pi$-$N$ data set have given rise to significant uncertainties of the low energy $\pi$-$N$ amplitudes. These, in turn, are reflected in the error limits of the extracted “experimental” value of the $\sigma$-term. This situation has led to an effort to remeasure all low energy $\pi$-$N$ observables at the remaining meson facilities. In this work we focus on the charge exchange reaction below 30 MeV pion incident energy. Absolute measurements of the pion-nucleon charge exchange reaction $\pi^-p$$\to$$\pi^0n$ below 50 or even 100 MeV are sparse. The difficulties stem from the requirement that the beam composition, beam flux, and the $\pi^0$ detection efficiency all have to be measured or determined accurately in an absolute way. Early published data below 50 MeV were measured by detecting the neutron at a single angle, 0$^\circ$, corresponding to the $\pi^0$ angle of 180$^\circ$ [@Duc-73]. Another set of measurements [@Sal-84] used a large NaI(Tl) crystal counter to detect single photons from the final state $\pi^0$ decay, at 27.4 and 39.3 MeV incident pion energy, covering a wide angular range. Due to the nature of this method, yields from a broad range of $\pi^0$ angles were mixed in at any given laboratory angle of a detected single photon. Hence, the authors could only report a Legendre polynomial decomposition of the $\pi^-p$$\to$$\pi^0n$ angular distribution, up to order 2. First published direct measurements of angular distribution data of the $\pi^-p$$\to$$\pi^0n$ reaction below 50 MeV were made using the LAMPF $\pi^0$ spectrometer [@Bae-81] at seven energies between 32.5 and 63.5 MeV for $\theta_{\rm lab}(\pi^0) = 0 - 30^\circ$ [@Fit-86]. A device such as the $\pi^0$ spectrometer detects the two photons following the $\pi^0$ decay in coincidence. This, in turn, enables a full reconstruction of the neutral pion’s momentum four-vector. In our work we used the same technique. Experimental method and normalization ------------------------------------- The present measurements of the differential cross sections for the reaction $\pi^-p$$\to$$\pi^0n$ at $27.5 \pm 0.2$ MeV were carried out in the LEP secondary beam channel at the Clinton P. Anderson Meson Physics Facility (LAMPF) in Los Alamos. We used a weakly focusing 30 MeV $\pi^-$ beam tune with 12 mr divergence (both horizontal and vertical), a beam spot size of 9 mm FWHM, momentum spread $\Delta p/p = 3$% and pion flux averaging $5\times 10^5$ $\pi^-$/sec. Relative on-target beam intensity was monitored with a sealed ion gas chamber in combination with a high precision charge integrator. Absolute cross-calibration of chamber ionization counts was obtained through activation measurements of the $^{12}$C($\pi^-,\pi^-n)^{11}$C reaction using cylindrical plastic scintillator targets [@But-82]. The $^{11}$C activation measurements were reproducible to better than $\pm 2$%, while the $^{11}$C activation cross section used for normalization has an uncertainty of 4.7% [@Lei-90]. The electron and muon contaminations in the beam were determined by a combination of direct measurement and constraints using the integrated energy deposited in the sealed ion chamber that was calibrated in absolute terms independently. The associated uncertainty of the pion flux amounted to $\pm 2.4$%. Our measurements were carried out using a $711 \pm 2$ mg/cm$^2$ polyethylene (CH$_2$) target, with a suitable $^{12}$C target for background subtraction. In addition, we recorded charge exchange data using a $267 \pm 7$ mg/cm$^2$ liquid hydrogen target as a check. We used the LAMPF $\pi^0$ spectrometer to detect coincident photons following $\pi^0$ decay in $\pi^-p$$\to$$\pi^0n$. The spectrometer multiwire proportional chamber and veto counter efficiencies were calibrated independently using cosmic muons. All tracking efficiencies were also evaluated from data and compared with a detailed simulation using [GEANT]{} [@Bru-87]. The resulting uncertainty of the integral $\pi^0$ detection efficiency was 4.6%. The measured detector response to $\pi^0$’s from the charge exchange reaction under study was compared to simulations using [GEANT]{} and [PIANG]{} [@Gil-79], with excellent agreement. The rms angular resolution of the spectrometer was 2$^\circ$. Results and discussion ---------------------- Results of our measurements of the $\pi^-p$$\to$$\pi^0n$ angular distribution between 0$^\circ$ and 55$^\circ$ (c.m.), binned into $9^\circ$ wide bins, are plotted in Fig. \[f:ang-dist\] as full circles. Error bars shown in the figure reflect only statistical uncertainties; in addition, an overall normalization uncertainty of 7.5% applies to the data, as discussed above. For the sake of comparison we have also included in Fig. \[f:ang-dist\] the angular distribution predicted by the comprehensive $\pi$-$N$ phase shift analysis SM95 by the VPI group [@said] (solid curve). The agreement between our data and the VPI phase shift prediction is excellent. Older data at this energy from Ref. [@Sal-84] are available only in the form of a Legendre polynomial decomposition (fit) of the angular distribution. That fit is represented in Fig. \[f:ang-dist\] by a dashed line, while dotted lines denote the associated error limits. -12mm \[f:ang-dist\] -5mm In summary, our results provide a new stringent constraint on the low energy $\pi$-$N$ phase shifts, and are in excellent agreement with the existing body of $\pi$-$N$ data. Reaction $\pi^+p\to\pi^+\pi^0p$ at 260 MeV ========================================== Low energy $\pi$-$\pi$ scattering has enjoyed longstanding attention as a window into the mechanism of chiral symmetry breaking. Pion-pion scattering lengths have recently come sharply into focus due to the controversy regarding $\langle 0|\bar{q}q|0\rangle$, the scalar quark condensate, and the two radically different and far-reaching scenarios of $\chi$SB [@Kne-95]. The current most reliable value of $a_0^0 = 0.26 \pm 0.05\ \mu^{-1}$ (where $\mu \equiv m_\pi$), extracted mainly from $K_{e4}$ decay data [@Poc-95], is not accurate enough to make the required distinction. We report here on preliminary results of a Chew-Low analysis [@Che-59] of exclusive $\pi^+p \to \pi^+\pi^0p$ data measured at 260 MeV $\pi^+$ incident energy. The experimental apparatus and the total cross section analysis are described in Ref. [@Poc-94]. The Chew-Low method evaluates $\pi\pi$ cross sections by extrapolating to the pion pole the function $F(s,t,m_{\pi\pi})$: $$\sigma_{\pi\pi}(m_{\pi\pi}) = \lim_{t\to\mu^2} F(s,t,m_{\pi\pi}) = \lim_{t\to\mu^2} {\partial^2 \sigma_{\pi\pi N}(s) \over \partial t \, \partial m_{\pi\pi}} \cdot {\pi \over f_\pi^2} \cdot {p^2(t-\mu^2)^2 \over t\, m_{\pi\pi}(m_{\pi\pi}^2 - 4\mu^2)^{1/2}} ~, \label{e:chew-low}$$ where $m_{\pi\pi}$ is the dipion invariant mass, $t$ is the squared 4-momentum transfer to the proton, $\sqrt{s}$ is the total c.m. energy, $p$ is the incident pion momentum, and $f_\pi$ the pion decay constant. In this work, we had to perform a deconvolution of the instrumental resolution function from the data before we could construct an interpretable $F(s,t,m_{\pi\pi})$. Preliminary values of $F$, the Chew-Low extrapolation function, calculated from our $\pi^+\pi^0p$ data at 260 MeV are plotted against $t$ in Fig. \[f:chew-low\], alongside a linear fit. Points with $|t| > 7 \mu$ were excluded from the fit due to the diminishing contribution of the one pion exchange process; the lowest $t$ point was excluded due to deconvolution uncertainties. -10mm \[f:chew-low\] -5mm Using the new $\pi^+\pi^0$ cross section datum we can deduce $a_0^2 \simeq 0.55 \pm 0.24\ \mu^{-1}$. However, further work is required in order to extract a more reliable extrapolated value of $\sigma(\pi\pi)$. That result, in turn, will be added to the existing $\pi\pi$ data set for a comprehensive dispersion-relation analysis. We gratefully acknowledge valuable contributions by S. Bruch and R.C. Minehart in the analysis, and by the other members of the E1179 collaboration [@Poc-94] in data taking. [99]{} J. Gasser, H. Leutwyler and M. Sainio, Phys. Lett. B 253 (1991) 252. J. Gasser and H. Leutwyler, Phys. Rep. 87 (1982) 77. J. Duclos [*et al.,*]{} Phys. Lett. 43B (1973) 245. M. Salomon [*et al.,*]{} Nucl. Phys. A414 (1984) 493. H.W. Baer [*et al.*]{} Nucl. Inst. Meth., 180 (1981) 445. D.H. Fitzgerald [*et al.,*]{} Phys. Rev. C 34 (1986) 619. G.W. Butler [*et al.*]{}, Phys. Rev. C 26 (1982) 1737. M.J. Leitch, private communication, 1990. R. Brun [*et al*]{}., [GEANT3]{}, publication DD/EE/84-1, CERN, Geneva, 1987 S. Gilad, Ph. D. Thesis, Tel Aviv University, 1979 (unpublished). . M. Knecht, B. Moussallam, J. Stern and N. H. Fuchs, Nucl. Phys. B457 (1995) 513. for a recent review of $a(\pi\pi)$ see, e.g., D. Počanić, Lect. Notes in Phys. 452 (1995) 95. G.F. Chew and F.E. Low, Phys. Rev. 113 (1959) 1640. D. Počanić [*et. al.*]{}, Phys. Rev. Lett. 72 (1994) 1156. [^1]: Work supported by the U.S. National Science Foundation.

--- abstract: 'We derive the flow equations for the symmetry unbroken phase of a dilute 3-dimensional Bose gas. We point out that the flow equation for the interaction contains parts which are non-analytic at the origin of the frequency-momentum space. We examine the way this non-analyticity affects the fixed point of the system of the flow equations and shifts the value of the critical exponent for the correlation length closer to the experimental result in comparison with previous work where the non-analyticity was neglected. Finally, we emphasize the purely thermal nature of this non-analytic behaviour comparing our approach to a previous work where non-analyticity was studied in the context of renormalization at zero temperature.' --- G. Metikas$^{1}$ and G. Alber$^{2}$ $^{1}$ Abteilung für Quantenphysik, Universität Ulm, Albert-Einstein Allee 11, D-89069, Ulm, Germany $^{2}$ Institut für Angewandte Physik, Technische Universität Darmstadt, D-64289, Darmstadt, Germany Introduction ============ Thermal effective actions are in general non-local in coordinate space because the temperature-dependent Green functions contain parts which are non-analytic at the origin of the momentum-frequency space[@Kapusta; @LeBellac]. In a theory with two interacting scalar fields, one integrates out one of them to find the effective action for the other. At finite temperature, provided the coupling is weak, one usually proceeds by applying perturbation theory, and then making an expansion in powers of frequency and momentum in order to obtain a local effective Lagrangian. It is this latter expansion which leads to results which are not uniquely defined but depend on the path on the frequency-momentum plane through which the origin is approached. For example, when the perturbation is truncated at the self-energy level, the self-energy is non-analytic at the origin. The reason is that the expansion is around a singularity [@WeldonRules]. This effect was noticed for the first time by Abrahams and Tsuneto in the 60’s, in the context of BCS theory, while they were studying time-dependent Ginzburg-Landau theory near zero temperature and near the critical temperature [@Tsuneto]. Later it became clear that it is the origin of Debye screening and of plasma oscillations in QED [@Baym; @Petitgirard]. These two different physical phenomena correspond to two different ways of approaching the origin of the momentum-frequency plane. The effects of the non-analyticity have also been studied in QED$_{3}$ [@Zuk] and in QCD [@Kalashnikov; @WeldonCovariant; @Klimov; @WeldonQuark; @Braaten; @FrenkelHigh]. The non-analyticity is also present in the graviton self-energy [@RebhanAnalytical; @RebhanCollective] and in higher-order graviton diagrams [@FrenkelHard]. Even in the much simpler case of interacting scalars the non-analyticity of the self-energy persists [@Hott; @Fujimoto; @WeldonMishaps; @EvansZero]. In the case of interacting scalars, an interesting remark is that, whenever the internal propagators in a loop have different masses, the self-energy is analytic at the origin [@Vokos]. The reason is that the singularity is not at the origin anymore, allowing thus a uniquely defined expansion around the origin. This paper is based on the simple observation that an essential step of the renormalization group method (RG) applied to a theory with a self-interacting field, is to split the field in slow and fast components, and integrate out the fast field obtaining thus an effective action for the slow field [@WilsonKogut; @CreswickWiegel]. Therefore, according to the above discussion about thermal effective actions, when RG techniques are applied in the context of thermal field theory, we anticipate that effects originating in the non-analyticity will arise. We choose to examine this aspect of RG in the context of a 3-dimensional homogeneous self-interacting bosonic gas with weak repulsive interactions and discuss its possible physical significance in this case. However our analysis and conclusions should hold whenever RG is used at finite temperature. This choice of system was motivated by the renewed interest in the Bose-Einstein condensation due to its recent experimental realization. For the interacting gas, the approach which is most often used is that of Bogoliubov. However, this is just a mean-field type method and, in principle, one can improve upon it by using more sophisticated techniques. One possibility near the critical region is the renormalization group [@StoofBijlsmaRen; @Alber; @AlberMetikas; @Andersen]. In the case of the homogeneous gas, there is an extra, more important reason for looking for alternatives to the Bogoliubov approach. In the critical region, the Bogoliubov theory simply does not work because there are fluctuations around the mean-field that cannot be treated perturbatively. This happens because, as the temperature approaches the critical temperature $T_{c}$, the thermal cloud density develops an infrared singularity and thus diverges as the momentum tends to zero [@RSFW; @FermiBurnett]. In section 2, we introduce the basics of the BEC formalism above the critical region. We then apply Wilsonian renormalization and derive the flow equations for the parameters of the Lagrangian. We point out the non-analytic structure of the RG correction to the interaction term (vertex) and follow this non-analyticity as it propagates to the flow equation for the interaction. In section 3, we calculate the non-trivial fixed pont of the system of the flow equations and find the critical exponent for the correlation length. We examine the way the non-analyticity affects the fixed point. We note that taking the non-analyticity into account shifts the value of the critical exponent closer to the experimental result. In section 4, we compare our work with [@Shankar] where the issue of non-analyticity in the context of renormalization is also discussed. We point out that the conclusions of [@Shankar] hold only at $T=0$ whereas the non-analytic behaviour which we are investigating in this paper is purely thermal and vanishes at zero temperature, thus being completely independent of the non-analyticity discussed in [@Shankar]. In section 5, we present our conclusions. Non-analyticity in the uncondensed phase ======================================== When the two-body collisions between bosons are taken to be low-momentum or s-wave, the path integral representation of the partition function of the homogeneous interacting Bose-gas is given by $$Z(\mu,\beta,V,g) \equiv {\rm Tr} e^{-\beta(\hat{H} - \mu \hat{N})} = \int \delta[\phi,\phi^*] e^{-S[\phi,\phi^*]} \label{partition}$$ with the action $$S[\phi, \phi^*] = \frac{1}{\hbar} \int_0^{\hbar \beta } d\tau \int_{V} d^D {\bf x} \left[ \phi^* (\tau, {\bf x}) [\hbar\frac{\partial}{\partial \tau} - \frac{\hbar^2}{2m} \nabla^{2} - \mu] \phi(\tau, {\bf x}) + \frac{1}{2} g |\phi(\tau, {\bf x})|^4 \right]. \label{action}$$ In the low-momentum approximation the interparticle interaction can be described by the zero-momentum component of the Fourier transform of the two-body interaction potential. Thus, within this approximation, a repulsive, short-range potential can be characterized by a positive interaction strength $g$. In three spatial dimensions this interaction strength is related to the positive scattering length $a$ of the interparticle interaction by the familiar relation $g = 4\pi\hbar^2 a/m$. The chemical potential is denoted by $\mu$. The case $\mu < 0$ holds for $T>T_{c}$ and corresponds to the uncondensed phase whereas $\mu >0$ describes the condensate which is formed when $T<T_{c}$ [@FetterWalecka]. In this paper we will deal only with the uncondensed phase. Starting from ($\ref{action}$) we can derive the renormalization group equations for $g$ and $\mu$. This set of coupled differential equations can then be used for the study of universal as well as non-universal properties of the gas [@StoofBijlsmaRen; @Alber; @AlberMetikas; @Andersen]. In the following we will set $\hbar=1$. In order to implement the first step of the RG procedure (Kadanoff transformation), we split the field $\phi(x)$ into a long-wavelength component $\phi_<(x)$ and into a short-wavelength component $\delta \phi_>(x)$. The short-wavelength field involves Fourier components which are contained only in an infinitesimally thin shell in momentum space of thickness $\Lambda(1 - dl)\leq | {\bf p}|\leq \Lambda$ near the cutoff $\Lambda $, whereas the long-wavelength field has all its Fourier components in the sphere whose center is at the origin of the momentum space and its radius is $\Lambda(1 - dl)$. We impose no cutoff on the frequency and apply the Wilsonian technique of consecutive infinitesimal shell integration only to the momentum and not to the frequency [@Hertz]. We denote the volume of the shell by $\delta V_{{\bf p}}$, the volume of the sphere by $V_{{\bf p}}$. The coordinate space volume is denoted by $V$. For simplicity we will be referring to $\phi_<(x)$ as the lower or slow field and to $\delta \phi_>(x)$ as the upper or fast field. Whenever more compact notation is required we will be making use of the following: $$x =(\tau,{\bf x}),\hspace{1cm} p=(p_{0}^{n},{\bf p}) \hspace{1cm} {\rm with} \hspace{1cm} p_{0}^{n} = 2 \pi n /\beta,$$ $$\int dx = \int_{0}^{\beta} d\tau \int_{V} d^{3} {\bf x}, \hspace{1cm} \int dp = \frac{1}{\beta} \; \sum_{n=-\infty}^{\infty} \int \frac{d^{3}{\bf p}}{(2 \pi)^3}.$$ We integrate out the upper field and are left with an effective action for the lower field $$S_{\rm eff}[\phi_{<}, \phi_{<}^{*}] = S[\phi_{<}, \phi_{<}^{*}] + \frac{1}{2} {\rm Tr}{\rm Ln}[1 - \hat{G}^{>} \hat{\Sigma}]. \label{kadaction}$$ For details on the derivation of this result and the approximations involved see [@CreswickWiegel]. The ${\rm Tr}$ denotes the trace in both the functional and the internal space of $ \hat{G}^{>} \hat{\Sigma} $ whereas the ${ \rm tr}$ denotes the trace only in the internal space (see below). The hat denotes that the corresponding quantity is a Schwinger-Fock operator [@SchwingerFock], $$\hat{G}^{>}(\hat{p}) = \left( \begin{array}{cc} B(\hat{p}) & 0 \\ 0 & B^{*}(\hat{p}) \end{array} \right)$$\ and $$\hat{\Sigma}(\hat{x}) = \frac{g}{2} \left( \begin{array}{cc} 4 \phi^{*}_{<}(\hat{x}) \phi_{<}(\hat{x}) & 2 \phi_{<}(\hat{x}) \phi_{<}(\hat{x}) \\ 2 \phi_{<}^{*}(\hat{x}) \phi_{<}^{*}(\hat{x}) & 4 \phi^{*}_{<}(\hat{x}) \phi_{<}(\hat{x}) \end{array} \right)$$\ with $$B(\hat{p})= B(\hat{p}_{0},\hat{{\bf p}})=\frac{1}{i \hat{p}_{0} + E(\hat{{\bf p}})}, \hspace{2cm} E(\hat{{\bf p}}) = \frac{\hat{{\bf p}}^2}{2m} - \mu.$$ Note that the expression for ${\hat\Sigma}$ contains the coupling $g$. This enables us to perform a perturbative expansion over $g$ in (\[kadaction\]) in order to calculate it explicitly. We truncate this expansion to second order in $g$ $${\rm Tr} {\rm Ln} [1-\hat{G}^{>} \hat{\Sigma}] \approx {\rm Tr} [ -\hat{G}^{>} \hat{\Sigma} - \frac{1}{2} (\hat{G}^{>} \hat{\Sigma})^2].$$ The first trace is: $$\begin{aligned} { \rm Tr }[ \hat{G}^{>} \hat{\Sigma}]&&= \int dx~\int dp~ {\rm tr}[G^{>}(p) \Sigma(x)] \nonumber \\ && = \int dx~ |\phi_{<} (x) |^2 2g \int_{\delta V_{{\bf p}}} \frac{ d^{3} {\bf p} }{ (2 \pi )^3 } [1+2N[{E({\bf p})}]]\end{aligned}$$ where $N[{E({\bf p})}] = [e^{\beta [{E({\bf p})}]} -1]^{-1}$ is the Bose-Einstein distribution. We note that the first trace is quadratic in the modulus of the lower field and can therefore be interpreted as a correction to the chemical potential $$d\mu = - g \int_{\delta V_{{\bf p}}} \frac{ d^{3} {\bf p} }{ (2 \pi )^3 } [1+2N[{E({\bf p})}]].$$ The second trace is: $$\begin{aligned} && {\rm Tr}[\hat{G}^{>} \hat{\Sigma} \hat{G}^{>} \hat{\Sigma}] =\int dp \int dk \int dx \int dy \; e^{i(p-k)(y-x)} \; {\rm tr} [ G^{>}(p) \Sigma(x) G^{>}(k) \Sigma(y) ] = \nonumber \\ && \int dp \int dk \int dx \int dy \; e^{i(p-k)(y-x)} \; \frac{g^2}{4} \nonumber \\ && \left[ 16\; B(p)\; B(k)\; \phi_{<}^{*}(x)\; \phi_{<}(x)\; \phi_{<}^{*}(y) \phi_{<}(y)\; + 4\; B(p)\; B^{*}(k)\; \phi_{<}(x)\; \phi_{<}(x)\; \phi_{<}^{*}(y)\; \phi_{<}^{*}(y) \; + \right. \nonumber \\ && \left. 4\; B^{*}(p)\; B(k)\; \phi_{<}^{*}(x)\; \phi_{<}^{*}(x)\; \phi_{<}(y)\; \phi_{<}(y)\; + 16 \;B^{*}(p)\; B^{*}(k)\; \phi_{<}^{*}(x)\; \phi_{<}(x)\; \phi_{<}^{*}(y)\; \phi_{<}(y)\; \right].\nonumber \\\end{aligned}$$ In order to simplify the above expression we change variables as follows: 1. In the second and fourth terms in the square brackets, $p \rightarrow -p$ and $ k \rightarrow -k$. 2. In the second term, $ x \rightarrow y$ and $ y \rightarrow x$. The second trace now becomes: $$\begin{aligned} &&{\rm Tr} [ \hat{G}^{>} \hat{\Sigma} \hat{G}^{>} \hat{\Sigma}]= \int dp \int dk \int dx \int dy \; \frac{g^2}{4} \nonumber \\ &&\left\{ \left[ e^{i(p-k)(y-x)} + e^{i(k-p)(y-x)} \right]\; B(p)\; B(k)\; 16\; \phi_{<}^{*}(x)\; \phi_{<}(x)\; \phi_{<}^{*}(y)\; \phi_{<}(y)\; \right. \nonumber \\ &&\left. + 2 e^{i(p-k)(y-x)}\; B^{*}(p)\; B(k)\; 4\; \phi_{<}^{*}(x)\; \phi_{<}^{*}(x)\; \phi_{<}(y)\; \phi_{<}(y) \right\}. \end{aligned}$$ Changing variables again, $k \rightarrow q=k-p$, yields: $$\begin{aligned} &&{\rm Tr} [ \hat{G}^{>} \hat{\Sigma} \hat{G}^{>} \hat{\Sigma}]= \int dq \int dx \int dy \; \frac{g^2}{4} \nonumber \\ && \left\{ 2\; e^{-iq(y-x)}\; J_{1}(q)\; 4\; \phi_{<}^{*}(x)\; \phi_{<}^{*}(x)\; \phi_{<}(y)\; \phi_{<}(y)\; \right. \nonumber \\ && \left. + \left[ e^{-iq(y-x)} + e^{iq(y-x)} \right]\; J_{2}(q)\; 16\; \phi_{<}^{*}(x)\; \phi_{<}(x)\; \phi_{<}^{*}(y)\; \phi_{<}(y)\; \right\} \label{inq}\end{aligned}$$ where $$J_{1}(q_{0}^{m}, {{\bf q}})=\int dp \; B^{*}(p) \; B(p+q) \hspace{1cm} {\rm and} \hspace{1cm} J_{2}(q_{0}^{m}, {{\bf q}})=\int dp \; B(p) \; B(p+q).$$ We note that ${\bf p}$ is the momentum of the upper field (integrated over the infinitesimal shell around the cutoff) whereas ${\bf q}$ is the momentum of the lower field, Fig.\[r0\]. It is essential in the RG procedure, and in particular in the Kadanoff transformation, to recast the effective action obtained after integrating out the upper field in the form of the original action (\[action\]). The first trace is in a form that can be interpreted as a correction to the chemical potential. This is not the case however for the second trace; there are quartic products of fields but, unlike the four-field coupling term in the original action (\[action\]), these are non-local in coordinate space, thus not allowing the effective action to be recast in the form of the original action. In other words, though we start from an action containing interactions which are local in coordinate space, the RG procedure generates more general, non-local interactions. This is a well-known feature of RG, namely to generate extra terms that do not appear in the original action and have a more general form in comparison to what we started with [@WilsonKogut; @Fisher; @Cardy]. Provided that these extra terms are not relevant they can, in most cases, for the purpose of calculating universal properties, be discarded. In our case we can Taylor expand $\phi_{<}(y)$ around $\phi_{<}(x)$. If we truncate this expansion at leading order, $\phi_{<}(y) \approx \phi_{<}(x)$, we remain within the family of local interactions we started with. This procedure is usually called derivative expansion and is physically relevant only when the lower field is slowly varying both in space and in time. Wilsonian renormalization is compatible with the derivative expansion. The reason is that in Wilsonian renormalization we are interested in constructing an effective action for the slow field. This compatibility can also be seen from a more technical point of view; the derivative expansion of the lower field is equivalent to an expansion of $J_{1}$ and $J_{2}$ in powers of $q_{0}$ and ${\bf q}$ (this is easily seen from (\[inq\]) doing integration by parts). This means that truncating the derivative expansion at higher than the leading order would give momentum and frequency dependent corrections to the interaction. However such terms are known to be irrelevant (see for example [@Fisher], page 128) and can therefore be omitted. The second trace becomes: $$\begin{aligned} && {\rm Tr} [ \hat{G}^{>} \hat{\Sigma} \hat{G}^{>} \hat{\Sigma}]= \int dx \; \frac{g^2}{4} \nonumber \\ && \left\{ \lim_{q \rightarrow 0}[ 2 e^{iqx} J_{1}(q) ] \ 4 \phi_{<}^{*}(x) \phi_{<}^{*}(x) \phi_{<}(x) \phi_{<}(x) + \lim_{q \rightarrow 0}[(e^{iqx} + e^{-iqx}) J_{2}(q) ] \ 16 \phi_{<}^{*}(x) \phi_{<}(x) \phi^{*}_{<}(x) \phi_{<}(x) \right\} \nonumber \\\end{aligned}$$ which is a local expression that can be interpreted as a correction to the coupling term of the original action $$dg = - g^2 \left\{ \lim_{(q_{0}^{m},|{{\bf q}}|) \rightarrow (0,0)}[J_{1}(q_{0}^{m},{{\bf q}})] + 4 \lim_{(q_{0}^{m},|{{\bf q}}|) \rightarrow (0,0)}[J_{2}(q_{0}^{m}, {{\bf q}})] \right\}. \label{dg}$$ It is at this point that the non-analyticity enters our discussion. Because we are at finite temperature, the integrals over frequencies in $J_{1}$ and $J_{2}$, $I_{1}$ and $I_{2}$ respectively, become sums which can be easily calculated when we turn them into integrals on the complex plane through Poisson summation $$J_{1}(q_{0}^{m},{\bf q}) = \int_{\delta V_{{\bf p}}} \frac{d^{3}{\bf p}}{(2 \pi)^3} \; I_{1}(q_{0}^{m}, {\bf q}, {\bf p}) \hspace{1cm} {\rm and} \hspace{1cm} J_{2}(q_{0}^{m}, {\bf q} ) = \int_{\delta V_{{\bf p}}} \frac{d^{3}{\bf p}}{(2 \pi)^3} \; I_{2}(q_{0}^{m}, {\bf q}, {\bf p})$$ where $$I_{1}(q_{0}^{m}, {\bf q}, {\bf p})= \frac{1}{\beta} \sum_{n=- \infty}^{ \infty } B^{*}(p_{0}^{n},{\bf p}) B(p_{0}^{n}+q_{0}^{m}, {\bf p+q})) = \frac{1+ N[E({{\bf p}})] + N[E({{\bf p}}+ {{\bf q}})]}{E({{\bf p}}+ {{\bf q}}) + E({{\bf p}}) -i q_{0}^{m}},$$ $$I_{2}(q_{0}^{m}, {\bf q}, {\bf p})= \frac{1}{\beta} \sum_{n=- \infty}^{ \infty } B(p_{0}^{n}, {{\bf p}}) B(p_{0}^{n}+q_{0}^{m}, {{\bf p}}+ {{\bf q}})) = \frac{N[E({{\bf p}})] - N[E({{\bf p}}+ {{\bf q}})]}{E({{\bf p}}+ {{\bf q}}) - E({{\bf p}}) + i q_{0}^{m}}.$$ We have set $\exp{i \beta q_{0}^{m}} =1$ because $q_{0}^{m}=2 \pi m/ \beta$. In the following, we will suppress the superscript of $q_{0}^{m}$ for simplicity. The first sum, $I_{1}$, is non-vanishing at $T=0$ and is known as the regular term. The second sum, $I_2$, is purely thermal and is usually called Landau term in the context of thermal field theory [@LeBellac]. We observe that the successive limits of $J_{1}(q_{0},{{\bf q}})$ coincide, i.e. $$\lim_{q_{0} \rightarrow 0} \lim_{|{{\bf q}}| \rightarrow 0} J_{1}(q_{0},{{\bf q}}) = \int_{\delta V_{{\bf p}}} \frac{d^{3} {{\bf p}}}{(2 \pi)^3} \frac{1}{2 {E({\bf p})}}\left[1+2N[{E({\bf p})}] \right] = \lim_{|{{\bf q}}| \rightarrow 0} \lim_{ q_{0} \rightarrow 0} J_{1}(q_{0},{{\bf q}}) \label{J1 limits}$$ whereas the successive limits of $J_{2}(q_{0},{{\bf q}})$ do not $$\lim_{q_{0} \rightarrow 0} \lim_{|{{\bf q}}| \rightarrow 0} J_{2}(q_{0},{{\bf q}}) = 0 \neq \lim_{|{{\bf q}}| \rightarrow 0} \lim_{q_{0} \rightarrow 0} J_{2}(q_{0},{{\bf q}}) = \int_{\delta V_{{\bf p}}} \frac{d^{3} {{\bf p}}}{(2 \pi)^3} \beta e^{\beta {E({\bf p})}} N^{2}[{E({\bf p})}] . \label{J2 limits}$$ The reason these two limits do not commute is that $J_{2}$ has a singularity at the origin of the momentum-frequency space [@WeldonRules]. Of course, in the evaluation of the above limits, we interchanged the limits with both the integration over the momentum $|{{\bf p}}| $ and with the angular integration over $\theta$, so our conclusion is not entirely reliable so far. In principle, one should perform the integrations over $|{{\bf p}}|$ and $\theta$ first, and then take the limit. Unfortunately, in our case, the integration over $|{{\bf p}}|$ cannot be done analytically. However, we can perform the angular integration over $\theta $ analytically before evaluating the limits, provided that we split the integral as follows $$J_{2}(q_{0}, {{\bf q}}) = \int_{\delta V_{{\bf p}}} \frac{d^{3} {{\bf p}}}{(2 \pi)^3} \frac{N[E({{\bf p}})]}{E({{\bf p}}+ {{\bf q}}) - E({{\bf p}}) + i q_{0}} - \int_{\delta V_{{\bf p}}} \frac{d^{3} {{\bf p}}}{(2 \pi)^3} \frac{N[E({{\bf p}}+ {{\bf q}})]}{E({{\bf p}}+ {{\bf q}}) - E({{\bf p}}) + i q_{0}} \label{split}$$ and perform the change of variables ${{\bf p}}\rightarrow -{{\bf p}}-{{\bf q}}$ in the second term, eliminating thus the dependence of the Bose-Einstein distribution on the angle $\theta$ and making the angular integration possible. This procedure yields the result $$J_{2}(q_{0}, {{\bf q}}) = \int_{\delta V_{{\bf p}}} \frac{d^{3} {{\bf p}}}{(2 \pi)^3} \left[ \frac{1}{{E({\bf p}+{\bf q})}- {E({\bf p})}+ i q_{0}} + \frac{1}{{E({\bf p}+{\bf q})}- {E({\bf p})}- i q_{0}} \right] N[{E({\bf p})}].$$ It is crucial to note that this change of variables is not permissible in case we interchange the limit $ \lim_{|{{\bf q}}| \rightarrow 0} \lim_{q_{0} \rightarrow 0} $ with the integrations, because this causes both the terms in (\[split\]) to diverge. These two divergencies canceled each other before the change of variables ${{\bf p}}\rightarrow -{{\bf p}}-{{\bf q}}$ [@Zuk]. Keeping this remark in mind we now perform the angular integration and find $$J_{2}(q_{0}, {{\bf q}})= \frac{m}{4 \pi^2} \int_{\Lambda - d\Lambda}^{\Lambda} d|{{\bf p}}| \frac{|{{\bf p}}| N[{E({\bf p})}]}{|{{\bf q}}|} \ln{\left[ \frac{(E_{+} - {E({\bf p})})^2 + q_{0}^2}{(E_{-} - {E({\bf p})})^2 + q_{0}^2} \right] } \label{J2 angular}$$ where $E_{+}=E(|{\bf p}|+|{\bf q}|)$ and $E_{-}=E(|{\bf p}|- |{\bf q}|)$. Instead of just taking the two successive limits in $J_{2}$ as we did before, which in the momentum-frequency plane corresponds to approaching the origin in the direction of the one or the other axis, we could approach the origin through any other curve, for example in the direction of any straight line $q_{0}= a |{{\bf q}}|$. Here, of course, we should not forget that the frequency is discrete whereas the momentum is continuous. However, for the purpose of better illuminating the structure of $J_{2}$ around the origin, we shall make the approximation that the frequency is continuous so that $q_{0}= a |{{\bf q}}|$ can hold for any real $a$. Applying this parameterization to (\[J2 angular\]) and then taking the limit $|{{\bf q}}| \rightarrow 0$ yields $$\lim_{|{{\bf q}}| \rightarrow 0} J_{2}(a |{{\bf q}}|, |{{\bf q}}|) = \frac{m}{4 \pi^2} \int_{\Lambda - d\Lambda}^{\Lambda} d|{{\bf p}}| N[{E({\bf p})}] \frac{2 |{{\bf p}}|^2}{|{{\bf p}}|^2 + m^2 a^2}$$ which reproduces the first limit of (\[J2 limits\]) for $a \rightarrow \infty$. This result was derived from (\[J2 angular\]) and therefore is also not valid when the limit $a \rightarrow 0$ is interchanged with the integration over $|{{\bf p}}|$. However, if we do an integration by parts, we find $$\begin{aligned} && \lim_{|{{\bf q}}| \rightarrow 0} J_{2}(a |{{\bf q}}|, |{{\bf q}}|) = \frac{m}{2 \pi^2} \left[N[{E({\bf p})}] \frac{|{{\bf p}}|^3}{|{{\bf p}}|^2 + m^2 a^2} \right]_{\Lambda - d \Lambda}^{\Lambda} \nonumber \\ && + \frac{1}{2 \pi^2} \int_{\Lambda - d\Lambda}^{\Lambda} d|{{\bf p}}| |{{\bf p}}|^2 \left[\frac{|{{\bf p}}|^2}{|{{\bf p}}|^2 + m^2 a^2} \beta e^{\beta {E({\bf p})}} N^{2}[{E({\bf p})}] - \frac{2 m^3 a^2}{[|{{\bf p}}|^2 + m^2 a^2]^2} N[{E({\bf p})}] \right]. \label{J2 by parts}\end{aligned}$$ We note that the surface term vanishes, as it is evaluated at the cutoff. This result not only reproduces the first limit of (\[J2 limits\]) for $a \rightarrow \infty$ but also agrees with the second limit of (\[J2 limits\]) for $a \rightarrow 0$. If we perform the angular integration and apply the same parameterization to $J_{1}$, at the limit $|{{\bf q}}| \to 0$, $J_{1}$ is independent of $a$ and given by (\[J1 limits\]). Expressions (\[J1 limits\]) and (\[J2 by parts\]) are to be substituted in the correction for the coupling constant (\[dg\]). Before we proceed to the second step of the RG formalism, we parameterize the momentum according to $|{{\bf p}}(l)| = \Lambda e^{-l}$. The purpose this change of variables serves is simply to make the flow equations more elegant. So far, the flow equation for the chemical potential is: $$\frac{d\mu}{dl}= -g \frac{\Lambda^3 e^{-3l}}{2 \pi^2} [1+2N[\epsilon_{\Lambda} e^{-2l} - \mu ]] \label{dmudl}$$ and the flow equation for the coupling is: $$\begin{aligned} &&\hspace{-1cm} \frac{dg}{dl} = - g^2 \left\{ \frac{\Lambda^{3} e^{-3l}}{2 \pi^2} \frac{1}{2 [\epsilon_{\Lambda} e^{-2l} - \mu]} [1+2 N[\epsilon_{\Lambda} e^{-2l} - \mu]] + \right. \nonumber \\ &&\hspace{-1cm} \left. 4 \frac{\Lambda^{3} e^{-3l}}{2 \pi^2} \left[ \frac{\Lambda^2 e^{-2l}}{\Lambda^{2} e^{-2l} + m^2 a^2} \beta N[\epsilon_{\Lambda} e^{-2l} - \mu] [ 1+ N[ \epsilon_{\Lambda} e^{-2l} - \mu]] - \frac{2 m^3 a^2}{[\Lambda^2 e^{-2l} + m^2 a^2]^2} N[\epsilon_{\Lambda} e^{-2l} - \mu] \right] \right\} \nonumber \\ \label{dgdl}\end{aligned}$$ where $\epsilon_{\Lambda}=\Lambda^2/2m$. At this point we apply the second step of the RG procedure, namely the trivial rescaling, whose purpose is to bring the effective action in the form of the original one. There are two stages, first, we rescale the momentum according to $|{{\bf q}}| \rightarrow |{{\bf q}}(l) | = |{{\bf q}}| e^{l}$ in order to re-establish the original cutoff $\Lambda $. Then we require that the effective Lagrangian has the same form as the original Lagrangian. This induces the trivial rescaling of the parameters of the effective Lagrangian. $$\begin{aligned} V & \to & V(l)= V~e^{-3l}, \nonumber \\ \beta & \to & \beta(l)=\beta~e^{-2l}, \nonumber \\ \phi & \to & \phi(l) = \phi~e^{3l/2}, \nonumber \\ \mu &\to & \mu(l)=(\mu+ \Delta \mu)~e^{2l}, \nonumber \\ g &\to & g(l)= (g+\Delta g)~e^{-l}. \label{trivialscaling} \end{aligned}$$ The trivial rescaling of $\beta $ implies that the frequency is rescaled as $ q_{0} \rightarrow q_{0}(l) e^{-2l} $ and therefore $$a = q_{0} / |{{\bf q}}| \rightarrow a(l) \; e^{-l}. \label{a trivialscaling}$$ Recasting (\[dmudl\]) and (\[dgdl\]) in terms of rescaled variables yields the flow equations for the corresponding running quantities $$\frac{ d\mu (l)}{dl}= 2 \mu(l) - g(l) \frac{\Lambda^3}{2 \pi^2} [ 1+2 N_{l}] \label{redmudl}$$ and $$\frac{dg(l)}{dl}=- g(l)- g^2(l) \frac{\Lambda^3}{2 \pi^2} \left\{\frac{1+2N_{l}}{2[\epsilon_{\Lambda} - \mu(l)]} + 4 \left[ \frac{\Lambda^2}{\Lambda^2+m^2 a^2(l)} \beta(l) N_{l} [1+N_{l}] - \frac{2m^3 a^2(l)}{[\Lambda^2+m^2 a^2(l)]^2} N_{l} \right] \right\} \label{redgdl}$$ where $ N_{l} = [e^{\beta (l) [\epsilon_{\Lambda} - \mu (l)]}-1]^{-1}$ is the Bose-Einstein distribution in terms of the rescaled variables. Fixed Point =========== It is important to investigate whether the path-dependence of the flow equation for the coupling has any consequences on quantities of physical interest. We look at a universal property, the critical exponent for the correlation length. This is calculated from the coupled system of (\[redmudl\]) and (\[redgdl\]). In fact, in order to have an autonomous system, we should also take into account the flow equation for the inverse temperature, $$\frac{d\beta(l)}{dl}= - 2 \beta (l)$$ which is just a differential expression of the trivial scaling of $\beta$ (\[trivialscaling\]). We observe that, although $\beta $ appears in the equations for $\mu$ and $g$, these do not couple back to the equation for $\beta $. We also note that the fixed point for $\beta $ is zero, $\beta_{*}=0$. The fact that $\beta_{*} =0$ complicates things because $\beta(l) $ appears in the flow equations not only explicitly but also through $N_{l}$. This means that, when we evaluate the fixed point for the system of the flow equations, the right-hand side of (\[redmudl\]) and (\[redgdl\]) will diverge, because $N_{l}$ diverges for $\beta_{*}=0$. This problem is circumvented when we define a scaled running chemical potential $M(l)$ and a scaled running coupling constant $\tilde{G}(l)$ such that the set of equations for these new parameters decouples from the equation for $\beta $; $$M(l) = \beta_{\Lambda} \mu (l) \hspace{1cm} {\rm and} \hspace{1cm} \tilde{G} (l) = \Lambda^{3} \beta_{\Lambda} g(l) / b(l)$$ where $\beta_{\Lambda}= m/\Lambda^2$, $\epsilon_{>}=1/2$ and $b(l)=\beta(l)/\beta_{\Lambda}$ is the scaled inverse temperature. In terms of these new, dimensionless parameters $$\begin{aligned} &&\hspace{-1.3cm} \frac{dM(l)}{dl}=2 M(l) - \frac{1}{2 \pi^2} \tilde{G}(l) b(l) [1+2N_{l}], \nonumber \\ &&\hspace{-1.3cm} \frac{d\tilde{G}(l)}{dl}= \tilde{G}(l) - \frac{1}{2\pi^2} \tilde{G}^{2}(l) b(l) \left\{ \frac{1+ 2N_{l}}{2[\epsilon_{>} - M(l)]} + 4 \left[\frac{\Lambda^2}{\Lambda^2+ m^2 a^2(l)} b(l) N_{l}[1+N_{l}] - \frac{1}{\beta_{\Lambda}} \frac{2 m^3 a^2(l)}{[\Lambda^2+m^2 a^2(l)]^2} N_{l} \right] \right\}, \nonumber \\ &&\hspace{-1.3cm} \frac{db(l)}{dl}= - 2 b(l). \label{scaled system}\end{aligned}$$ In the neighbourhood of the fixed point, the rescaled temperature is high and the approximation $N_{l} \approx [\beta(l)[\epsilon_{\Lambda} - \mu(l)]]^{-1} = [b(l) [\epsilon_{>} - M(l)]]^{-1} $ holds [@StoofBijlsmaRen]. This yields the equations $$\begin{aligned} &&\frac{dM(l)}{dl}=2 M(l) - \frac{1}{2 \pi^2} \tilde{G}(l) \; b(l) \; [1 + 2 \frac{1}{b(l)[\epsilon_{>} - M(l)]}], \nonumber \\ && \frac{d\tilde{G}(l)}{dl}= \tilde{G}(l) - \frac{1}{2\pi^2} \tilde{G}^{2}(l) \left\{ \frac{b(l)}{2[\epsilon_{>} - M(l)]} + \frac{1}{[\epsilon_{>} - M(l)]^2} + 4 \frac{\Lambda^2}{\Lambda^2 + m^2 a^2(l)} \; \frac{b(l)}{\epsilon_{>} - M(l)} \right. \nonumber \\ &&\hspace{1.2cm} \left. + 4 \frac{\Lambda^2}{\Lambda^2 + m^2 a^2(l)} \; \frac{1}{[\epsilon_{>} - M(l)]^2} - 4 \frac{2 \Lambda^2 m^2 a^2(l)}{[\Lambda^2 + m^2 a^2(l)]^2} \; \frac{1}{\epsilon_{>} - M(l)} \right\}, \nonumber \\ && \frac{db(l)}{dl}= - 2 b(l). \label{approx system} \end{aligned}$$ At the fixed point $(M_{*}, \tilde{G}_{*}, b_{*}=0)$ the left hand side of (\[approx system\]) is zero by definition. The form of the right hand side depends subtly on whether $a$ is zero or not as we will see. Near the fixed point, the second term is dominant in the square brackets of the flow equation for the chemical potential in (\[approx system\]), $$\frac{dM(l)}{dl}=2 M(l) - \frac{1}{\pi^2} \tilde{G}(l) \frac{1}{\epsilon_{>} - M(l)}.$$ For $a=0$, we recall (\[a trivialscaling\]), which means that $a(l)=0$ for any $l$. Consequently, the equation for the coupling reduces to $$\frac{d\tilde{G}(l)}{dl}= \tilde{G}(l) - \frac{1}{2\pi^2} \tilde{G}^{2}(l) \left\{ \frac{b(l)}{2[\epsilon_{>} - M(l)]} + \frac{1}{[\epsilon_{>} - M(l)]^2} + 4 \frac{b(l)}{\epsilon_{>} - M(l)} + 4 \frac{1}{[\epsilon_{>} - M(l)]^2} \right\}.$$ Recalling the trivial scaling of $b$ (\[trivialscaling\]), we see that the second and fourth terms in the curly brackets above are dominant near the fixed point, $$\frac{d\tilde{G}(l)}{dl}= \tilde{G}(l) - \frac{1}{2\pi^2} \tilde{G}^{2}(l) \frac{5}{[\epsilon_{>} - M(l)]^2}$$ and calculate the non-trivial fixed point $$(M_{*},\tilde{G}_{*},b_{*}) =(\frac{1}{12},\frac{5}{4} \frac{\pi^2}{18}, 0).$$ We linearize the system of (\[approx system\]) around the fixed point and find the largest eigenvalue $\lambda_{+}=1.878$. Therefore the critical exponent for the correlation length is $\nu=0.532$, which agrees with the one found in [@StoofBijlsmaRen]. For $a \neq 0$, according to the trivial scaling of $a$ (\[a trivialscaling\]), $ a_{*} = \infty $ and therefore the first and the third terms in the curly brackets of (\[approx system\]) vanish near the fixed point, as in the case of $a=0$. From the remaining terms the fifth vanishes and, significantly, the fourth is also vanishing near the fixed point, leaving as dominant contribution only the second term. Consequently the flow equation for the coupling reduces to $$\frac{d\tilde{G}(l)}{dl}= \tilde{G}(l) - \frac{1}{2\pi^2} \tilde{G}^{2}(l) \frac{1}{[\epsilon_{>} - M(l)]^2}$$ and the non-trivial fixed point is $$(M_{*},\tilde{G}_{*},b_{*})=(\frac{1}{4},\frac{\pi^2}{8}, 0).$$ Linearizing around the fixed point we find that $\lambda_{+}=1.561$ and therefore $\nu = 0.640$. It is interesting to note that we could have found the same fixed point and critical exponent directly from (\[approx system\]), had we set $a = \infty$ and therefore $a(l) = \infty $ for any $l$. This situation is similar to what happens, for example, in the case of thermal QED for the photon propagator. Because the photon self-energy is non-analytic at the origin, different ways of approaching the origin lead to different dispersion relations and give rise to different types of excitations [@Kapusta; @LeBellac]. For short wavelengths, the dispersion relation is $q_{0}^2 = {{\bf q}}^2 + m_{P}^2$, where $ m_{P}^2 = (e^2/2) (T^2/3 + \mu^2/\pi^2)$ is the thermal mass for the transverse photons whereas the longitudinal photons do not propagate. However, for long wavelengths, the transverse photons have the dispersion $q_{0}^2 = \omega_{P}^2 + \frac{6}{5} {{\bf q}}^2 $ and the longitudinal photons have the dispersion $q_{0}^2 = \omega_{P}^2 +\frac{3}{5} {{\bf q}}^2$, where $\omega_{P}^2 = \frac{2}{3} m_{P}^2$ is the plasma frequency at order $e^2$. The phenomenon which we are describing here is of the same mathematical nature, the difference being that it is occurring not in the propagator but in the vertex between four bosons. To be more precise it is the vertex graph corresponding to $J_{2}$, Fig.\[r0\], that exhibits the same singular behaviour as the photon self-energy in QED. Comparing the two critical exponents derived above with the known experimental value $\nu=0.670$ [@Zinn], it becomes clear that the second procedure gives an improved estimate. In fact, our estimate is better even than the value $\nu =0.613$ calculated in [@StoofBijlsmaRen] with the inclusion of the marginal three-body scattering term in the action. Flow equations and non-analyticity at zero and finite temperature ================================================================= Non-analyticity in the context of RG has been discussed before by Shankar in [@Shankar]. In this work, the author gives a detailed overview of the RG approach to interacting, non-relativistic fermions in one, two and three dimensions and in certain instances (pages 161, 166, 170, 178) refers to the non-analyticity (or lack thereof) which appears in the one-loop RG corrections to the quartic interaction among fermions. This is highly reminiscent of the case we are studying, the essential difference being that we are dealing with bosons instead of fermions. This would render the main point of this paper - the study of a non-analyticity in the flow equation for the interaction - rather trivial and expected by extending [@Shankar] to bosons. This is not the case however, the non-analyticity we are studying is of completely different nature from the one studied in [@Shankar]. The RG calculations in [@Shankar] are at $T=0$ whereas ours at $T \neq 0$ and the non-analyticity we are referring to is essentially thermal and vanishes at $T=0$. To further clarify this point, let us consider the ”zero sound” (ZS) graph which is studied in Eq.(315), page 161 of [@Shankar]. Zero Temperature Non-Analyticity -------------------------------- ### Zero Sound Integral In the zero-sound calculation of [@Shankar], the following integral appears: $$S_{1}[\Omega, q] = \int_{-\Lambda}^{\Lambda} \frac{dk}{2 \pi} \int_{- \infty}^{\infty} \frac{d\omega }{2 \pi} \ \frac{1}{[ i \omega - k ][i (\omega + \Omega) - k - q ]}$$ where $\Omega$, $q$ are the external frequency and momentum and $\omega$, $k$ are the internal frequency and momentum respectively. The external momentum is constrained by the same cutoff as the internal momentum, $-\Lambda \leq q \leq \Lambda.$ We focus on the integral over $\omega$. When the external momentum and frequency are zero, the integrand has a double pole and therefore $S_{1}[0,0]= 0$, Fig.\[r1\]. For non-zero external frequency and momentum, the integrand has two single poles. Let us assume that $k > 0$. If $k+q > 0$, both poles are in the lower half-plane and closing the contour from above yields $S_{1}[\Omega,q]= 0$. However, if $k+q <0$, the two single poles are in different half-planes and closing the contour either from above or below yields $$S_{1}[\Omega, q] = \int_{-\Lambda}^{\Lambda} \frac{dk}{2 \pi} \frac{i}{\Omega + iq}.$$ Assuming $k<0$ we can argue the same way, so, only when $k$ and $k+q$ have different signs, the $\omega$-integral is non-zero, Fig.\[r2\]. Because of the $k$-integration, there is always a range of $k$ values for which $k$ and $k+q$ have different signs. Therefore any non-zero $q$ results in a non-vanishing $$S_{1}[\Omega,q]= \int_{- \Lambda}^{\Lambda} \frac{dk}{2 \pi} \frac{i}{\Omega + i q} [\theta(k) - \theta(k+q)]$$ which depends sensitively on how the limit $ \{ \Omega,q \} \to \{0,0 \}$ is taken [@Shankar], $$\lim_{\Omega \to 0} \lim_{q \to 0} S_{1}[\Omega,q] = 0 \neq \lim_{q \to 0} \lim_{\Omega \to 0} S_{1}[\Omega,q]= - 1.$$ ### Renormalization Integral In the RG calculation of [@Shankar], a one-loop correction to the interaction is of the form: $$S_{2}[\Omega, q] = \left[ \int_{-\Lambda}^{-\Lambda + d\Lambda} + \int_{\Lambda-d\Lambda}^{\Lambda} \right] \frac{dk}{2 \pi} \int_{- \infty}^{\infty} \frac{d\omega }{2 \pi} \ \frac{1}{[ i \omega - k ][i (\omega + \Omega) - k - q ]}.$$ As before, for zero external frequency and momentum, $S_{2}[0,0]=0$. For non-zero external frequency and momentum, we note that $S_{2}$ differs from $S_{1}$ only in the range of integration of the momenta, $ k \in [- \Lambda, -\Lambda + d\Lambda ] \cup [ \Lambda - d\Lambda, \Lambda]$ and $q \in [ -\Lambda + d\Lambda, \Lambda - d\Lambda]$. This means that $|k|>|q|$ and consequently $k,k+q$ have always the same sign throughout the integration over $k$. Therefore $S_{2}[\Omega,q]=0$ which has no dependence on how the limits of the external frequency and momentum are taken [@Shankar]. The conclusion is that, when performing RG calculations, non-analyticities at the origin of external frequency and momentum space vanish, even when they are present in the corresponding zero-sound calculations. Thermal Non-Analyticity ----------------------- ### Zero Sound Integral Now consider what happens at $T \neq 0$. In zero-sound calculations one has to perform the integral $$S_{1}^{\rm{T}}[\Omega_{m}, q] = \int_{-\Lambda}^{\Lambda} \frac{dk}{2 \pi} \frac{1}{\beta} \sum_{n=-\infty}^{n=\infty} \ \frac{1}{[ i \omega_{n} - k ][i (\omega_{n} + \Omega_{m}) - k - q ]}$$ where $\omega_{n}=(2 n + 1) \pi / \beta$, $\Omega_{m}=(2m +1) \pi / \beta$ are the discretized internal and external frequencies respectively. When the external frequency and momentum are zero, we find $$S_{1}^{\rm{T}}[0,0]= - \int_{-\Lambda}^{\Lambda} \frac{dk}{2 \pi} \frac{\beta}{4} \cosh^{-2}[\frac{k \beta}{2}] \stackrel{\rm{T} \to 0 }{\longrightarrow} -1.$$ For non-zero external frequency and momentum, we obtain $$S_{1}^{\rm{T}}[\Omega_{m}, q] = \int_{-\Lambda}^{\Lambda} \frac{dk}{2 \pi} \frac{i}{\Omega_{m}+iq} \frac{1}{2} \left\{ \tanh{[\frac{k \beta}{2}]} - \tanh{[\frac{(k+q) \beta}{2}]} \right\}$$ which gives the correct zero temperature limit $ S_{1}^{\rm{T} \to 0}[\Omega_{m}, q] = S_{1}[\Omega, q]$. At $T \neq 0$, $$\lim_{\Omega_{m} \to 0} \lim_{q \to 0} S_{1}^{\rm{T}}[\Omega_{m}, q]=0 \stackrel{ \rm{T} \to 0 }{ \longrightarrow } 0,$$ $$\lim_{q \to 0} \lim_{\Omega_{m} \to 0} S_{1}^{\rm{T}}[\Omega_{m}, q]= - \int_{-\Lambda}^{\Lambda} \frac{dk}{2 \pi} \frac{\beta}{4} \cosh^{-2}[\frac{k \beta}{2}] \stackrel{\rm{T} \to 0 }{\longrightarrow} -1$$ and therefore the non-analyticity of the zero-temperature zero-sound calculation persists at finite temperatures, Fig.\[r3\]. ### Renormalization Integral As at $T=0$, the integral appearing in RG calculations, differs from $S_{1}^{\rm{T}}$ only in the range over which the internal and external momentum are integrated. When the external frequency and momentum are zero, we find $$S_{2}^{\rm{T}}[0,0]= - \left[ \int_{-\Lambda}^{-\Lambda + d\Lambda} + \int_{\Lambda-d\Lambda}^{\Lambda} \right] \frac{dk}{2 \pi} \frac{\beta}{4} \cosh^{-2}[\frac{k \beta}{2}] \stackrel{\rm{T} \to 0 }{\longrightarrow} 0.$$ For non-zero external frequency and momentum, we obtain $$S_{2}^{\rm{T}}[\Omega_{m}, q] = \left[ \int_{-\Lambda}^{-\Lambda + d\Lambda} + \int_{\Lambda-d\Lambda}^{\Lambda} \right] \frac{dk}{2 \pi} \frac{1}{\beta} \sum_{n=-\infty}^{\infty} \ \frac{1}{[ i \omega_{n} - k ][i (\omega_{n} + \Omega_{m}) - k - q ]}.$$ Poisson summation yields, Fig.\[r3\], $$S_{2}^{\rm{T}}[\Omega_{m}, q] = \left[ \int_{-\Lambda}^{-\Lambda + d\Lambda} + \int_{\Lambda-d\Lambda}^{\Lambda} \right] \frac{dk}{2 \pi} \frac{i}{\Omega_{m}+iq} \frac{1}{2} \left\{ \tanh{[\frac{k \beta}{2}]} - \tanh{[\frac{(k+q) \beta}{2}]} \right\}$$ which gives the correct zero-temperature limit $$S_{2}^{\rm{T} \to 0}[\Omega_{m}, q]=0.$$ However, unlike the zero temperature renormalization integral, the finite temperature renormalization integral is non-analytic, because $$\lim_{\Omega_{m} \to 0} \lim_{q \to 0} S_{2}^{\rm{T}}[\Omega_{m}, q]=0 \stackrel{ \rm{T} \to 0 }{ \longrightarrow } 0,$$ $$\lim_{q \to 0} \lim_{\Omega_{m} \to 0} S_{2}^{\rm{T}}[\Omega_{m}, q]= - \left[ \int_{-\Lambda}^{-\Lambda + d\Lambda} + \int_{\Lambda-d\Lambda}^{\Lambda} \right] \frac{dk}{2 \pi} \frac{\beta}{4} \cosh^{-2}[\frac{k \beta}{2}] \stackrel{\rm{T} \to 0 }{\longrightarrow} 0.$$ Therefore, at finite temperature, the non-analyticity of the zero-sound one-loop graph persists but there is also an extra, purely thermal non-analyticity appearing in the finite temperature renormalization one-loop graph. The zero temperature non-analyticities are due to the splitting of a double pole into two single poles residing in different half-planes. The finite temperature non-analyticities are due only to the splitting of a double pole and appear even if the resulting two single poles reside in the same half-plane. Therefore it is only natural that extra non-analyticities appear at finite temperature in addition to those existing at zero-temperature. Conclusions =========== We have investigated thermal non-analyticities at the origin of the momentum-frequency space in the context of Wilsonian renormalization. We have shown that taking them into account leads to an improved estimate for the critical exponent of the correlation length, when approaching the critical region from the symmetry unbroken phase. It is well-known that the $\epsilon$-expansion gives values closer to the experiment than ours, however one should remember that the $\epsilon$-expansion is asymptotic [@Zinn]. It is therefore crucial to find safer ways of calculating the critical exponent, such as the improved momentum-shell method which was used here. When approaching the critical region from the symmetry-broken phase there is an excellent estimate $\nu=0.685$ [@StoofBijlsmaRen]. It may be possible to improve this value by taking into account the non-analyticity as we did for the approach from the symmetry unbroken phase. Work on this issue is currently under progress. Finally, we have pointed out that when one applies Wilsonian renormalization at finite temperature, one may encounter non-analytic behaviour which is different in nature from the non-analyticity (or lack thereof) discussed in [@Shankar]. Acknowledgement {#acknowledgement .unnumbered} =============== The authors wish to thank W. Wonneberger for bringing reference [@Shankar] to our attention. G.M. wishes to thank I.J.R. Aitchison for many discussions on the subtleties of thermal field theory as well as V. Yudson for a useful discussion on renormalization and statistical physics. This work was supported by the Deutsche Forschungsgemeinschaft within the Forschergruppe “Quantengase”. [10]{} J. I. Kapusta, [*Finite-Temperature Field Theory*]{} (Cambridge University Press, England, 1989). M. LeBellac, [*Thermal Field Theory*]{} (Cambridge University Press, England, 1996). H. A. Weldon, Phys.Rev. [**D28**]{}, 2007 (1983). E. Abrahams and T. Tsuneto, Phys.Rev. [**152**]{}, 416 (1966). G. Baym, J. P. Blaizot, and B. Svetitsky, Phys.Rev. [**D46**]{}, 4043 (1992). E. Petitgirard, Z.Phys. [**C54**]{}, 673 (1992). I. J. R. Aitchison and J. A. Zuk, Ann.Phys. [**242**]{}, 77 (1995). O. K. Kalashnikov and V. V. Klimov, Sov. J. Nucl. Phys. [**31**]{}, 699 (1980). H. A. Weldon, Phys.Rev. [**D26**]{}, 1394 (1982). V. V. Klimov, Sov. J. Nucl. Phys. [**33**]{}, 934 (1981). H. A. Weldon, Phys.Rev. [**D26**]{}, 2789 (1982). E. Braaten and R. D. Pisarski, Nucl.Phys. [**B337**]{}, 569 (1990). J. Frenkel and J. C. Taylor, Nucl. Phys. [**B334**]{}, 199 (1990). A. Rebhan, Nucl. Phys. [**B368**]{}, 479 (1992). A. Rebhan, Nucl. Phys. [**B351**]{}, 706 (1991). J. Frenkel and J. C. Taylor, Z.Phys. [**C49**]{}, 515 (1991). A. Das and M. Hott, Phys.Rev. [**D50**]{}, 6655 (1994). Y. Fujimoto and H. Yamada, Z. Phys. [**C37**]{}, 265 (1988). H. A. Weldon, Phys.Rev. [**D47**]{}, 594 (1993). T. S. Evans, Z. Phys. [**C36**]{}, 153 (1987). P. Arnold, S. Vokos, P. Bedaque, and A. Das, Phys.Rev. [**D47**]{}, 4698 (1993). K. G. Wilson and J. Kogut, Phys.Rep. [**12**]{}, 75 (1974). R. J. Creswick and F. W. Wiegel, Phys.Rev. [**A28**]{}, 1579 (1983). H. T. C. Stoof and M. Bijlsma, Phys.Rev. [**A54**]{}, 5085 (1996). G. Alber, Phys.Rev. [**A63**]{}, 023613 (2001). G. Alber and G. Metikas, Appl.Phys. [**B73**]{}, 773 (2001). J. O. Andersen and M. Strickland, Phys.Rev. [**A60**]{}, 1442 (1999). M. Rasolt, M. J. Stephen, M. Fischer, and P. Weichman, Phys.Rev.Lett [**53**]{}, 798 (1984). K. Burnett, in [*Bose-Einstein Condensation in Atomic Gases*]{}, Societá Italiana di Fisica, edited by M. Inguscio, S. Stringari, and C. E. Wieman (IOS, Varenna, 1998), Vol. CXL. R. Shankar, Rev.Mod.Phys. [**66**]{}, 129 (1994). A. L. Fetter and J. D. Walecka, [*Quantum Theory of Many-Particle Systems*]{} (McGraw-Hill, N.Y., 1971). J. A. Hertz, Phys.Rev [**B14**]{}, 1165 (1976). J. Schwinger, Phys.Rev. [**82**]{}, 664 (1951). M. E. Fisher, in [*Lecture Notes in Physics*]{} (Springer, Heidelberg, 1983), Vol. 186, p. 1. J. Cardy, [*Scaling and Renormalization in Statistical Physics*]{} (Cambridge University Press, England, 1996). J. Zinn-Justin, [*Quantum Field Theory and Critical Phenomena*]{} (Oxford University Press, England, 1989). ![The two contributions of the RG correction to the interaction g, $J_{1}$ and $J_{2}$; $p$ is the frequency-momentum of the upper field (the momentum ${\bf p}$ is integrated over the infinitesimal shell around the cutoff), $q$ is the frequency-momentum of the lower field. \[r0\]](silly3.ps "fig:"){width="4in"} ![The two contributions of the RG correction to the interaction g, $J_{1}$ and $J_{2}$; $p$ is the frequency-momentum of the upper field (the momentum ${\bf p}$ is integrated over the infinitesimal shell around the cutoff), $q$ is the frequency-momentum of the lower field. \[r0\]](silly4.ps "fig:"){width="4in"} ![ The integrand of the $\omega$-integration has one double pole and the integral is zero. Any non-zero external momentum splits the double pole into two single poles. There is always a range of $k$ for which the two single poles reside in different half-planes and consequently the integral is non-vanishing. \[r1\]](r1.ps){width="6in"} ![ The integrand of the $\omega$-integration has one double pole and the integral is zero. The introduction of non-zero external momentum splits the double pole into two single poles. However, because the $k$-integration is over the infinitesimal shell near the cutoff and the external momentum $q$ takes value below the infinitesimal shell, the two single poles are always on the same half-plane and the integral remains zero. \[r2\]](r2.ps){width="6in"} ![ The integrand of the $\omega$-integration has one double pole. The introduction of non-zero external momentum splits the double pole into two single poles and the integral is non-vanishing regardless of whether the two single poles are on different half-planes or not. \[r3\]](r3.ps "fig:"){width="5in"} ![ The integrand of the $\omega$-integration has one double pole. The introduction of non-zero external momentum splits the double pole into two single poles and the integral is non-vanishing regardless of whether the two single poles are on different half-planes or not. \[r3\]](r4.ps "fig:"){width="5in"}

--- abstract: 'We investigate the phase structure of $SU(3)$ gauge theory in four and five dimensions with one compact dimension by using perturbative one-loop and PNJL-model-based effective potentials, with emphasis on spontaneous gauge symmetry breaking. When adjoint matter with the periodic boundary condition is introduced, we have rich phase structure in the quark-mass and compact-size space with gauge-symmetry-broken phases, called the $SU(2)\times U(1)$ split and the $U(1)\times U(1)$ re-confined phases. Our result is qualitatively consistent with the recent lattice calculations. When fundamental quarks are introduced in addition to adjoint quarks, the split phase becomes more dominant and larger as a result of explicit center symmetry breaking. We also show that another $U(1)\times U(1)$ phase (pseudo-reconfined phase) with negative vacuum expectation value of Polyakov loop exists in this case. We study chiral properties in these theories and show that chiral condensate gradually decreases and chiral symmetry is slowly restored as the size of the compact dimension is decreased.' author: - Kouji Kashiwa - Tatsuhiro Misumi title: | Phase structure and Hosotani mechanism\ in gauge theories with compact dimensions revisited --- Introduction {#sec:intro} ============ Now that the Higgs-like particle has been discovered in Large Hadron Collider (LHC) [@AT1; @CMS1], one of primary interests of particle physics is to understand mechanism of dynamical electroweak symmetry breaking. To settle down the unsolved issues including the hierarchy problem, there have been proposed several promising courses including supersymmetric, composite-Higgs and extra-dimensional models. Among them, the gauge-Higgs unification with extra dimensions [@Man1; @H1] produces brilliant dynamics of gauge symmetry breaking, called the Hosotani mechanism [@H1; @H2; @DMc1; @HIL1; @Hat1; @H3; @H4], with the Higgs particle as an extra-dimensional component of the gauge field: when adjoint fermions are introduced with periodic boundary condition (PBC) in gauge theory with a compact dimension, the compact-space component of the gauge field can develop a non-zero vacuum expectation value (VEV), which breaks the gauge symmetry to its subgroup spontaneously. This phenomenon originates in the non-abelian Aharonov-Bohm effect. There have been proposed a number of BSM models directly and indirectly based on this mechanism [@ACP1; @MT1; @HMTY1; @ST1; @FOOS1; @ST2; @HKT1; @IMST1; @HNU1; @HTU1; @HTU2; @HH1; @IKY0] Recently, the same phenomenon has been observed in a different context. As well-known, the imaginary time dimension is compacted in the finite-temperature (Quantum Chromodynamics) QCD. When the adjoint fermion is introduced with PBC in the theory, it was shown that exotic phases appear [@U1; @U2; @O1; @MO1; @MO2; @MO3; @NO1; @CD1], where the color trace of Polyakov line takes non-trivial VEV leading to the dynamical gauge symmetry breaking. This can be interpreted as $R^{3}\times S^{1}$ realization of the Hosotani mechanism. Further study is now required to understand detailed properties and phase diagram for this case. As for the five-dimensional gauge theory on $R^{4}\times S^{1}$, lattice simulation is still at the stage to understand the nature of 5D pure Yang-Mills theory. First attempt was done in Ref. [@C1] and recent progress was shown in Ref. [@KNO1; @JN1; @BBB1; @LPS1; @CW1; @BCW1; @EKM1; @EFK1; @DFKK1; @DFKKN1; @DFK1; @FFKLV1; @IK1; @IK2; @IK3; @IKL1; @KIL1; @IK4; @IK5; @DKR1; @dFKP1; @KLR1; @Del1; @IKY1]. At the present, we have no detailed investigation on the phase structure of the five-dimensional gauge theory with dynamical quarks. The purpose of our work is to understand properties of gauge theories with quarks on $R^{3}\times S^{1}$ and $R^{4}\times S^{1}$ by using effective theories and show a guideline for further lattice simulations. We focus on the phase structure in $SU(3)$ gauge theory by mainly using the perturbative one-loop effective potential. When we look into chiral properties, the Polyakov-loop-extended Nambu–Jona-Lasinio (PNJL) model [@GO1; @Fuku1; @K1; @HK1] is introduced. We obtain the phase diagram in the quark-mass and compacted-size space for several different choices of fermion representations and boundary conditions. In the theory with adjoint fermions with PBC, we find rich phase diagram with unusual phases where $SU(3)$ gauge symmetry is spontaneously broken to $SU(2)\times U(1)$ (split phase) or $U(1)\times U(1)$ (re-confined phase). This result is consistent with the recent lattice simulations [@MO1; @CD1] and the perturbative calculations for massless quarks [@H3; @H4]. We show that chiral condensate is gradually decreased and the chiral symmetry is slowly restored with the compacted size being decreased although small chiral transitions coincide with the deconfined/split and split/reconfined transitions. We also study phase diagram for the case with both fundamental and adjoint quarks, and show that the split phase becomes more dominant as a result of the explicit breaking of $Z_{3}$ center symmetry. This result indicates that the $SU(3)\to SU(2)\times U(1)$ phase, which is of significance in terms of phenomenology, can be controlled by adjusting the number of fundamental flavors. In this case we point out that another unusual ($U(1)\times U(1)$) phase with negative VEV of color-traced Polyakov loop emerges, which we call “pseudo-reconfined phase". We note similar investigation on the phase diagram of compact-space gauge theory with PBC fermions were shown in [@U1; @U2; @O1; @MO2; @NO1], where the authors focus on the volume independence of the vacuum structure from the viewpoint of the large N reduction [@EK1; @KUY1; @U1; @U2; @UY1; @BS1; @B1; @BS2; @PU1]. In order to reproduce the confined phase at low temperature, they used the gluon effective potentials with the mass-dimension parameter, which leads to explicit gauge symmetry breaking. Since these effective potentials do not suit our purpose of studying spontaneous gauge symmetry breaking due to the Hosotani mechanism, we mainly use setups with manifest gauge symmetry. It is also notable that gauge symmetry breaking for all clasical Lie groups with PBC matters are classified from topological viewpoints in [@AU1]. This paper is organized as following. In Sec. \[sec:EP\], we show our setups including the one-loop potential and the PNJL model. Sec. \[sec:PS4\] shows our numerical results for four dimensional cases. Sec. \[sec:PS5\] shows the results for five-dimensional cases. In Sec. \[sec:OB\] we discuss how our predictions can be checked in the lattice simulations. Sec. \[sec:sum\] is devoted to summary. Effective potential {#sec:EP} =================== In this section we discuss our setups for SU(N) gauge theory. The one-loop effective potential is composed of two parts, gluonic and quark contributions. Since our goal is to study the phase diagram in terms of Wilson-line phases (Polyakov-line phases), $q_{i}$ $(i=1,2,...,N)$, we will write the effective potential as a function of these variables. Besides the one-loop potential, we consider the contribution from the chiral part, which originates in the NJL-type four-point interactions. The total effective potential is then regarded as that of the PNJL model [@Fuku1] as a function of the $q_{i}$ and the chiral condensate $\sigma$. Depending on our purposes, we in some cases use only the perturbative one-loop effective potential, and in other cases use the PNJL effective potential. We also comment on possible deformation reflecting non-perturbative effects in the gluonic contribution. SU(N) in four dimensions {#subsec:D4} ------------------------ We begin with one-loop effective potential for $SU(N)$ gauge theory. In four dimensions, the finite-temperature one-loop effective potential for gauge bosons and fermions is well studied in Ref. [@GPY1; @W1]. What we do in the present study is just to regard the compacted imaginary-time direction as one of the spatial direction, and to replace $T$ by $1/L$ where $L$ is a size of the compacted dimension. Firstly, we rewrite the $SU(N)$ gauge boson field as $$\begin{aligned} A_\mu &= \langle A_y \rangle + \tilde{A}_\mu,\end{aligned}$$ where $y$ stands for a compact direction and $\langle A_y \rangle$ is a vacuum expectation value (VEV) while $\tilde{A}_\mu$ is fluctuation from it. The VEV can be replaced by $$\begin{aligned} \langle A_y \rangle &=\frac{2 \pi}{gL} q,\end{aligned}$$ where $q$’s color structure is $\mathrm{diag}(q_1,q_2,...,q_{N})$ with $q_1+q_2+\cdot\cdot\cdot+q_{N-1}+q_{N}=0$. We note that eigenvalues of $q_{i}$ are invariant under all gauge transformations preserving boundary conditions. Then we can easily observe spontaneous gauge symmetry breaking from values of $q_{i}$ in the vacuum. For detailed argument on gauge transformation for this topic, see [@H3] for example. The gluon one-loop effective potential ${\cal V}_g$ is expressed as $$\begin{aligned} {\cal V}_g &= - \frac{2}{L^4 \pi^2} \sum_{i,j=1}^N \sum_{n=1}^{\infty} \Bigl( 1 - \frac{1}{N} \delta_{ij} \Bigr) \frac{\cos( 2 n \pi q^{ij})}{n^4} \end{aligned}$$ where $q_{ij} = ( q_i - q_j )$ and $N$ is the number of color degrees of freedom. The contribution from massless fundamental quarks ${\cal V}_f$ is given by $$\begin{aligned} {\cal V}_f &= \frac{4 N_f}{L^4 \pi^2} \sum_{i=1}^N \sum_{n=1}^\infty \frac{\cos[2 \pi n (q_i + 1/2)]}{n^4},\end{aligned}$$ where $N_f$ is the number of fundamental flavors. Depending on boundary conditions, we should replace $q_i + 1/2$ by $q_i + \phi$. For example, the choice of $\phi=0$ describes quarks with periodic boundary conditions. From here, we denote ${\cal V}_f^\phi$ as effective potential of the fundamental fermion with boundary angle $\phi$. The contribution from massive fundamental quarks is expressed by using the second kind of the modified Bessel function $K_2(x)$ as $$\begin{aligned} {\cal V}_f^\phi(N_{f},m_f) &= \frac{ 2 N_f m_{f}^{2}}{\pi^2L^2} \sum_{i=1}^N \sum_{n=1}^\infty \frac{K_2 ( n m_{f} L )}{n^2} \cos [2 \pi n (q_i + \phi)],\end{aligned}$$ where $m_{f}$ is the fundamental fermion mass. (We assume the same mass for all flavors.) Here the second kind of the modified Bessel function $K_\nu(x)$ is defined as $$\begin{aligned} K_\nu (x) &= \frac{\sqrt{\pi} (x/2)^\nu}{\Gamma(\nu+1/2)} \int_1^\infty e^{-xt} (t^2-1)^{\nu-\frac{1}{2}} dt,\end{aligned}$$ where $\Gamma(x)$ is the gamma function. The adjoint quark contribution ${\cal V}_a^\phi$ can be easily obtained by the following replacement, $$\begin{aligned} {\cal V}_a^\phi (N_{a}, m_{a}) &= \frac{ 2 N_a m_{a}^{2}}{\pi^2L^2} \sum_{i,j=1}^N \sum_{n=1}^\infty \Bigl( 1 - \frac{1}{N} \delta_{ij} \Bigr) \frac{K_2 ( n m_{a} L )}{n^2} \cos [2 \pi n (q_{ij} + \phi)], \label{Re}\end{aligned}$$ where $N_a$ and $m_{a}$ are the number of flavors and the mass for adjoint fermions. For the gauge theory with $N_{f}$ fundamental and $N_{a}$ adjoint fermions with arbitrary boundary conditions, the total one-loop effective potential is given by $${\cal V} = {\cal V}_{g}+{\cal V}_{f}^{\phi}(N_{f}, m_f)+{\cal V}_a^{\phi}(N_{a}, m_a). \label{eq_ep_pert}$$ This total one-loop effective potential contains eight parameters including the compactification scale $L$, the number of colors $N$, the fermion masses $m_{f}$, $m_{a}$, the number of flavors $N_{f}$, $N_{a}$, and the boundary conditions $\phi$ for two kinds of matter fields. All through the present study we keep $N=3$, then obtain phase diagram in $1/L$-$m_{a}$ space with $m_{f}$, $N_{f}$, $N_{a}$ and $\phi$ fixed to several values. The reason we change $m_{a}$ while fixing $m_{f}$ is that gauge symmetry phase diagram is more sensitive to the former than the latter. Up to this point, we work in first-principle perturbative calculations and have no model-parameter-fixing process dictated by physical results or numerical calculations. We here comment on ability and limitation of the perturbative one-loop effective potential. This weak-coupling-limit potential at least works to investigate weak-coupling physics. In this study a small compactification scale $L$ (or high temperature $T\sim 1/L$) corresponds to weak coupling, and the $1/L$-$m_{a}$ phase diagram obtained from the effective potential is valid at smaller $L$ while the strong-coupling physics near larger $L$ or small temperature cannot be reproduced. We note that gauge symmetry breaking due to Hosotani mechanism takes place even at weak-coupling due to the compact dimension topology [@H1; @H2; @H3; @H4]. Indeed, the exotic phases in the recent lattice simulations [@CD1] have been shown to be at weak-coupling regime. Our potential thus work to reveal properties of these exotic phases. Now we consider contribution from the chiral sector. We introduce the four-point interaction at the action level as $$g_{S}[(\bar{\psi} \psi)^{2} +(\bar{\psi} i \gamma_{5} {\vec \tau} \psi)^{2}], \label{gs}$$ where $\psi$ is a two-flavor fermion field and $g_{S}$ is the effective coupling constant with the mass dimension minus two. (We here concentrate on the two-flavor case.) The introduction of the above term (\[gs\]) leads to addition of the following zero-temperature contribution to the one-loop effective potential (\[eq\_ep\_pert\]), $${\cal V}_{\chi} = g_{S} \sigma^{2}-{d_{R}\Lambda^{4}\over{4\pi^{2}}}\left[ \left(2+{m_{c}^{2}\over{\Lambda^2}}\right)\sqrt{1+{m_{c}^{2}\over{\Lambda^2}}}+{m_{c}^{4}\over{\Lambda^4}} \log\left({m_{c}/\Lambda\over{1+\sqrt{1+m_{c}^2 /\Lambda^{2}}}}\right) \right],$$ where $\Lambda$ is a cutoff scale of the effective theory and $m_{c}$ stands for constituent quark mass $m_{c}=m-2g_{S}\sigma$ with $\sigma$ being chiral condensate [@NO1]. Depending on fundamental or adjoint representations, $d_{R}$ takes $N$ or $N^2 -1$ respectively. In the case with both fundamental and adjoint quarks, we need two different sets of chiral sectors. The total effective potential combined with the chiral contribution is given by $${\cal V}_\mathrm{total} = {\cal V} + {\cal V}_{\chi}. \label{PNJL-like}$$ We note that this total effective potential is identical to that of the PNJL model [@Fuku1] adopting the one-loop potential as the gluon contribution, $${\cal L}_{\rm PNJL}= \bar{\psi}(\gamma_{\mu}D_{\mu}+m)\psi -g_{S}[(\bar{\psi} \psi)^{2}+(\bar{\psi} i\gamma_{5}{\vec \tau}\psi)^{2}] +{\cal V}_{g},$$ with $D_{j}=\partial_{j}, \, D_{4}=\partial_{4}+iA_{4}$. We thus call the total effective potential (\[PNJL-like\]) PNJL or PNJL-based effective potential. In the present study we will use this PNJL-based potential to study gauge theory with adjoint or fundamental quarks with periodic boundary conditions(PBC). In this model, we have two model parameters, the cutoff scale $\Lambda$ and the effective-coupling $g_{S}$ in addition to the eight parameters in the one-loop perturbative potential shown below (\[eq\_ep\_pert\]). These two parameters are the very model parameters and should be fixed to reproduce physical results or first-principle lattice calculations. On the other hand, the lattice QCD simulation on the compactified space has been done as finite-temperature lattice QCD, but for only anti-periodic boundary conditions (aPBC) for quarks. Thus, we choose the following criterion for fixing the two parameters for PBC quarks; we first fix them so as to reproduce the zero-temperature constituent mass and the chiral critical temperature in the lattice finite-temperature QCD with aPBC matters [@KL1], then we use the same parameter set for the PBC case. According to [@KL1], the zero-temperature constituent mass for aPBC adj. QCD is given by $$m_{c}(T=0, m_{a}=0)=2.322{\rm GeV},$$ and the chiral phase transition takes place at $$1/L_{CT}=T_{CT}\sim 2 {\rm GeV},$$ where $L_{CT}$ and $T_{CT}$ are critical values for chiral transition. We can fix $\Lambda$ and $g_{S}$ by using these lattice results. This method is also applied to the fundamental quarks. Apart from the chiral contribution, the PNJL potential we obtained is based on perturbative calculations, and chiral properties we will obtain from it is again valid for a small $L$ region. In addition to this limitation, we have the cutoff scale $\Lambda$ in this model. It means that we cannot apply the model beyond this cutoff. The model is thus valid for intermediate compactification scale as $0\ll 1/L <\Lambda$. In the next section we will find specific chiral properties within this limitation. We here comment on non-perturbative modification of the PNJL model. In the standard use of the PNJL model, the gluonic contribution ${\cal V}_{g}$ is replaced by the “non-perturbatively"-deformed ones: in order to mimic the confinement/deconfinement phase transition in the study on QCD phase diagram,, the one-loop gluon potential should be replaced by some nonperturbative versions. We have several schemes including the simple scale introduction in Ref. [@NO1], center-stabilized potential in [@MO3; @MeMO1; @MeO1], the one-loop [*ansatz*]{} in Ref. [@Rob1] and the strong-coupling lattice potential in Ref. [@Fuku1; @RRTW1]. The first one in [@NO1] is given by $$\begin{aligned} {\cal V}_{g}^\mathrm{np1} &=- \frac{2}{L^4 \pi^2} \sum_{i,j=1}^N \sum_{n=1}^{\infty} \Bigl( 1 - \frac{1}{N} \delta_{ij} \Bigr) \frac{\cos( 2 n \pi q^{ij})}{n^4} \,+\, \frac{M^2}{2\pi^2L^2} \sum_{i,j=1}^N \sum_{n=1}^\infty \Bigl( 1 - \frac{1}{N} \delta_{ij} \Bigr) \frac{\cos(2n\pi q^{ij})}{n^2}, \label{g_np}\end{aligned}$$ where the second term is the modification incorporating non-perturbative effect and the mass scale $M$ works as a scale for confinement/deconfinement transtion. The center-stabilized potential in [@UY1; @MO3; @MeMO1; @MeO1] has a similar form given as $$\begin{aligned} {\cal V}_{g}^\mathrm{np2} &=- \frac{2}{L^4 \pi^2} \sum_{i,j=1}^N \sum_{n=1}^{\infty} \Bigl( 1 - \frac{1}{N} \delta_{ij} \Bigr) \frac{\cos( 2 n \pi q^{ij})}{n^4} \,+\,{1\over{L}}\sum^{[N/2]}_{n=1} \sum_{i,j=1}^N a_{n} \Bigl( 1 - \frac{1}{N} \delta_{ij} \Bigr)\cos( 2 n \pi q^{ij}), \label{g_np2}\end{aligned}$$ where $a_{n}$ in the deformation term is a mass dimension-3 model parameter, which works as the phase transition scale in this case. The one-loop ansatz potential is based on simple introduction of the scale $M$ to the original gluon potential as $$\begin{aligned} {\cal V}_{g}^\mathrm{np3} &=- \frac{2M^2}{L^2 \pi^2} \sum_{i,j=1}^N \sum_{n=1}^{\infty} \Bigl( 1 - \frac{1}{N} \delta_{ij} \Bigr) \frac{\cos( 2 n \pi q^{ij})}{n^4}, \label{g_np3}\end{aligned}$$ where some of compactification scale $L$ is just replaced by $M$ in the original potential. The potential from the lattice strong-coupling expansion is non-perturbative deformation adopted in the original PNJL model [@Fuku1], which is given by $$\begin{aligned} {\cal V}_{g}^\mathrm{np4} = -{4\over{a^{3}L}}e^{-\sigma aL}({\rm Tr}P)^{2}- {1\over{a^{3}L}}\log\Big[ -({\rm Tr}P)^{4}+8{\rm Re}({\rm Tr}P)^{3}-18({\rm Tr}P)^{2}+27 \Big], \label{g_np4}\end{aligned}$$ where $1/a$ stands for the cutoff scale and $\sigma$ for the string tension for confinement. $P=\mathcal{P}\exp [ig\int dy A_{y}]\sim{\rm diag}[e^{2\pi i q_{1}},...,e^{2\pi i q_{N}}]$ is a compactified dimension Polyakov (Wilson) loop. These two model parameters work as the scale for the phase transition. All the modifications able to reproduce the phase transition, however, it is notable that the $SU(N)$ gauge symmetry is explicitly broken due to the mass-dimensionful model parameters in them. It means that they do not suit our purpose of classifying the gauge-broken phases. We also emphasize that, in the first place, the exotic phases including gauge-broken phases are likely to emerge at weak-coupling region [@CD1] and we do not need the above deformations for the strong-coupling region for our purpose. Thus, we will mainly adopt one-loop effective potential as the gluonic contribution in our PNJL model and just discuss how the phase diagram is changed by the deformation in the next section. SU(N) in five dimensions {#subsec:D5} ------------------------ In the five-dimensional case, most of the setup is parallel to the four-dimensional case except for difference of mass dimensions of fields and parameters. We here show the one-loop effective potential of gluon and quarks below. The five-dimensional one-loop effective potential in the gluon sector is given by $$\begin{aligned} {\cal V}_g &= - \frac{9}{4\pi^2L^5} \sum_{i,j=1}^N \sum_{n=1}^{\infty} \Bigl( 1 - \frac{1}{N} \delta_{ij} \Bigr) \frac{\cos( 2 \pi n q_{ij})}{n^5},\end{aligned}$$ which has $1/L^5$ dimension. The effective potential for massless fermions in this case is given by $$\begin{aligned} {\cal V}_f^{1/2} &= -2 T \int \frac{dk}{ 8\pi^2} k^3 \Bigl[ \ln \Bigl( 1 + e^{-L|{\bf k}| + 2 i \pi q^{ij}} \Bigr) + \ln \Bigl( 1 + e^{-L|{\bf k}| - 2 i \pi q^{ij}} \Bigr) \Bigr] \nonumber\\ &= \frac{3}{L^5 \pi^2} \sum_{n=1}^\infty \frac{\cos [2 \pi n (q^{ij}+1/2) ]}{n^5}. \label{Integral}\end{aligned}$$ The effective potential of the massless fundamental and adjoint fermions with arbitrary boundary condition is expressed as $$\begin{aligned} {\cal V}_f^\phi &= \frac{3 N_f}{\pi^2L^5} \sum_{i=1}^N \sum_{n=1}^\infty \frac{\cos[ 2 \pi n (q_i + \phi ) ]}{n^5}, \\ {\cal V}_a^\phi &= \frac{3 N_a}{\pi^2L^5} \sum_{i,j=1}^N \sum_{n=1}^\infty \Bigl( 1 - \frac{1}{N} \delta_{ij} \Bigr) \frac{\cos[ 2 \pi n (q_{ij} + \phi ) ]}{n^5}.\end{aligned}$$ For the massive fermion, we should replace $|{\bf k}|$ by $E_p = \sqrt{{\bf p}^2+m^2}$ in Eq. (\[Integral\]). To evaluate the integration, we should expand the logarithm in the same way as the four dimesional case. Individual integration is done by using the Bessel function as $$\begin{aligned} \frac{e^{\pm 2i \pi nq^{ij}}}{n} \int_0^\infty p^3 e^{-n L E_\mathrm{p}} dp &= - \frac{e^{\pm 2i \pi nq^{ij}}}{n^2} \frac{\partial}{\partial L} \int_1^\infty m^3 (t^2-1) e^{-nLmt} dt. \nonumber\\ &= - \frac{e^{\pm 2i \pi nq^{ij}}}{n^2} \frac{2^{3/2}}{\sqrt{\pi}} \frac{\partial}{\partial L} \Bigl[ \Bigl( \frac{m}{n L} \Bigr)^{3/2} K_{3/2} (n L m) \Bigr] \nonumber\\ &= \frac{e^{\pm 2i \pi nq^{ij}}}{n} \frac{ 2^{3/2} m}{\sqrt{\pi}} \Bigl( \frac{m}{n L} \Bigr)^{3/2} K_{5/2}(n L m).\end{aligned}$$ Then, the potential contributions from massive fermions are obtained by using the $K_{5/2}(x)$ as $$\begin{aligned} {\cal V}_f^\phi (N_{f}, m_{f}) &= \frac{ \sqrt{2} N_f ( m_{f}/L )^{5/2} }{\pi^{5/2}} \sum_{i=1}^N \sum_{n=1}^\infty \frac{K_{5/2} (n m_{f} L)}{n^{5/2}} \cos[ 2 \pi n (q_i + \phi )], \\ {\cal V}_a^\phi (N_{a}, m_{a}) &= \frac{ \sqrt{2} N_a ( m_{a}/L )^{5/2} }{\pi^{5/2}} \sum_{i,j=1}^N \sum_{n=1}^\infty \Bigl( 1 - \frac{1}{N} \delta_{ij} \Bigr) \frac{K_{5/2} (n m_{a} L)}{n^{5/2}} \cos[ 2 \pi n (q_{ij} + \phi )].\end{aligned}$$ This total one-loop effective potential in five dimensions again contains eight parameters including the compactification scale $L$, the number of colors $N$, the fermion masses $m_{f}$, $m_{a}$, the number of flavors $N_{f}$, $N_{a}$, and the boundary conditions $\phi$ for two kinds of matter fields. As with the four-dimensional case we keep $N=3$, then obtain phase diagram in $1/L$-$m_{a}$ space with $m_{f}$, $N_{f}$, $N_{a}$ and $\phi$ fixed to several values. The $1/L$-$m_{a}$ phase diagram obtained from this effective potential is valid at smaller $L$ or the weak coupling regime, and thus works to reveal properties of the exotic phases at weak coupling. We here comment on treatment of parity pairs. When we need a parity-even mass term in odd dimensions (e.g. $Z_2$ orbifolded theory), we need a set of parity pairs $\psi^{\pm}(-y)=\pm\psi(y)$ in the action [@MT1]. Although our study does not take much care about such a case, we note that the factor two appears in front of the fermion potentials if the parity pair is introduced. It is thus the same situation as doubled flavors within this setup. In Sec. \[sec:PS5\], we will see that the phase diagram changes depending on the choice of the parity pair in five dimensions. Phase structure in four dimensions {#sec:PS4} ================================== In this section, we investigate the phase structure in the compacted-size $L^{-1}$ and quark-mass $m$ space for $SU(3)$ gauge theories on $R^{3}\times S^{1}$. Because of the relation $-q_{3}=q_{1}+q_{2}$, the vacuum structure is discussed just from the $(q_{1},q_{2})$ potential. In the following we begin with investigation on vacuum and phase structures based on the one-loop potential (\[eq\_ep\_pert\]) for several choices of matters and boundary conditions. By far, lattice QCD simulations in four dimensions with one compacted dimension have been done for aPBC fundamental, PBC fundamental, aPBC adjoint [@KL1] and PBC adjoint [@CD1] matters. There are no lattice study on systems including both fundamental and adjoint matters. We also have no lattice corresponding results for five dimensional cases. In the present study we mainly focus on the case with PBC adjoint, and with PBC adjoint and fundamental matters, and make comparison to the lattice result if exists. As shown in the above section, we make use of the other lattice results to fix model parameters in PNJL models. We note that we will sometimes consider one-flavor cases as $(N_{f},N_{a})=(0,1)$ or $(N_{f},N_{a})=(1,1)$. Since the anomaly is not implemented and difference of one- and multiple-flavors just appears as an overall factor in the effective potential, our four-dimensional results are qualitatively unchanged even for the two-flavor cases as $(N_{f},N_{a})=(0,2)$ or $(N_{f},N_{a})=(2,2)$. Thus, these one-flavor examples are sufficient for our purpose of studying phase diagram qualitatively. (Things are different in five dimensions as shown in Sec. \[sec:PS5\].) In the appendix. \[sec:VSF\], we summarize the potential minima for gauge-symmetry-preserved cases, which are not our main results. We show the contour plots of the effective potentials for $(N_{f},N_{a})=(0,0)$ in Fig. \[Fig\_p\_g\_3D\], $(N_{f},N_{a})=(1,0)$ with the anti-periodic boundary condition (aPBC) in Fig. \[Fig\_p\_gfa\_3D\], $(N_{f},N_{a})=(1,0)$ with the periodic boundary condition (PBC) in Fig. \[Fig\_p\_gfp\_3D\], and $(N_{f},N_{a})=(0,1)$ with aPBC in Fig. \[Fig\_p\_gaa\_3D\]. We set $m_{f}=m_{a}=0$ in these cases. In the pure-gauge and gauge+adjoint theories, the three degenerate vacua clearly reflects the $Z_{3}$ center symmetry while the introduction of fundamental fermions lift of the degeneracy, which means breaking of the center symmetry. We note that the global minima are given by any of $(q_{1},q_{2})=(0,0),(1/3,1/3), (-1/3,-1/3)$, and we have $q_{1}=q_{2}=q_{3}$ (mod 1) in the vacuum. It obviously shows that the $SU(3)$ symmetry remains in all these cases. ![ The one-loop effective potential of $SU(3)$ gauge theory on $R^{3}\times S^1$ with one adjoint fermion with PBC $[ {\cal V}_g + {\cal V}_a^{0}(N_{a}=1, m_a =0)] L^4$. (Right) The contour plot as a function of $q_1$ and $q_2$. Thicker region stands for deeper region of the potential. (Left) The effective potential as a function of $q_1$ with $q_2=0$. The global minima are located at $(q_{1},q_{2})=(\pm1/3,0)$. []{data-label="Fig_p_gap_3D"}](potential_gap_CP.eps "fig:"){width="40.00000%"} ![ The one-loop effective potential of $SU(3)$ gauge theory on $R^{3}\times S^1$ with one adjoint fermion with PBC $[ {\cal V}_g + {\cal V}_a^{0}(N_{a}=1, m_a =0)] L^4$. (Right) The contour plot as a function of $q_1$ and $q_2$. Thicker region stands for deeper region of the potential. (Left) The effective potential as a function of $q_1$ with $q_2=0$. The global minima are located at $(q_{1},q_{2})=(\pm1/3,0)$. []{data-label="Fig_p_gap_3D"}](potential_gap_2D.eps "fig:"){width="40.00000%"} Now, we shall look into the case with spontaneous gauge symmetry breaking. We consider $(N_{f},N_{a})=(0,1)$ with PBC. We note that this theory has exact center symmetry, and all the phases, even the gauge-broken phase, should reflect this symmetry. Figure \[Fig\_p\_gap\_3D\] shows the effective potential $[ {\cal V}_g +{\cal V}_a^{0} ] L^4$. The left contour plot is obviously different from the gauge-symmetric cases. Careful search shows that the minima are located at $(q_1,q_2)=(0,1/3)$, $(1/3,0)$, $(-1/3,1/3)$, $(-1/3,0)$, $(0,-1/3)$, $(1/3,-1/3)$. It means the vacua are given by permutations of $(q_1,q_2,q_3)=(0, 1/3, -1/3)$, and $SU(3)$ gauge symmetry is broken into $U(1)\times U(1)$. This is the famous result, known as the Hosotani mechanism, where the Aharonov-Bohm effect in the compacted dimension nontrivially breaks gauge symmetry [@H1; @H2; @H3; @H4]. We note that this situation is sometimes called “re-confined phase" [@CD1] since the color fundamental trace of the Polyakov loop $\Phi \equiv ({\rm Tr}_F~P) / N$ becomes zero. To study the phase diagram, we introduce nonzero quark mass. Figure \[Fig\_p\_gapm\_2D\] shows the effective potential $[{\cal V}_g+{\cal V}_a^{0}(N_{a}, m_{a})]L^4$ as a function of $q_{1}$ with $q_{2}=0$ for $m L=1.2$, $1.6$, $2.0$ and $3.0$ from left to right panels ($m \equiv m_{a}$). It is clearly seen that there is the first-order phase transition in the vicinity of $m L=1.6$. This is a transition between the re-confined phase and the other gauge-broken phase, which we call the “split phase" [@CD1]. The contour plots for $mL=1.6$ and $mL=1.8$ are shown in Fig. \[Fig\_p\_gapm\_3D\]. The $mL=1.8$ case corresponds to the split phase. The global minima in the split phase are given by - $\mathrm{Im}~\Phi=0$ : $(q_1,q_2)=(0,0.5)$, $(0.5,0.5)$, $(0.5,0)$, - $\mathrm{Im}~\Phi>0$ : $(q_1,q_2)=(1/6,-1/3)$, $(-1/3, 1/6)$, $(1/6,1/6)$, - $\mathrm{Im}~\Phi<0$ : $(q_1,q_2)=(1/3,-1/6)$, $(-1/6,-1/6)$, $(-1/6,1/3)$. These results indicate that the vacuum in the first set is given by permutations of $(q_1,q_2,q_3)=(0, 0.5, 0.5)$, and $SU(3)$ gauge symmetry is broken into $SU(2)\times U(1)$. The vacua for other two sets are derived by the $Z_3$ transformation of $(q_1,q_2,q_3)=(0, 0.5, 0.5)$. From the viewpoint of phenomenological symmetry breaking, this phase and similar phases for large color numbers are the most preferable. In Fig. \[4d\_phase\_p\] we depict the $L^{-1}$-$m$ phase diagram with one PBC adjoint quark $(N_{f}, N_{a})=(0,1)$ based on the one-loop effective potential. We note that, as $m$ appears as $m L$ in this one-loop effective potential, we have scaling in the phase diagram and we can choose arbitrary mass-dimension unit. Since we drop the non-perturbative effect in the gluon potential, we have no confined phase at small $L^{-1}$ (at low temperature). The order of the three phases in Fig. \[4d\_phase\_p\] (deconfined $SU(3)$ $\to$ split $SU(2)\times U(1)$ $\to$ reconfined $U(1)\times U(1)$ from small to large $L^{-1}$.) is consistent with that of the lattice simulation except that we have no confined phase. We note that all the critical lines in the figure are first-order. In Fig. \[Fig\_pl\_distribution\] we depict a schematic distribution plot of $\Phi$ in the complex plane. It is obvious that each phase reflects $Z_{3}$ symmetry. In the split phase, $\Phi$ takes nonzero values but in a different manner from the deconfined phase. In the re-confined phase, we exactly have $\Phi=0$ with the vacuum breaking the gauge symmetry. ![ The one-loop effective potential of $SU(3)$ gauge theory on $R^{3}\times S^1$ with one PBC adjoint quark as a function of $q_1$ with $q_2=0$ $[ {\cal V}_g + {\cal V}_a^{0}] L^4$, for $m L=1.2$ (reconfined), $1.6$ (reconfined$\leftrightarrow$split), $2.0$ (split$\leftrightarrow$deconfined) and $3.0$ (deconfined). []{data-label="Fig_p_gapm_2D"}](2D_potential_gap_t10_m12.eps "fig:"){width="23.00000%"} ![ The one-loop effective potential of $SU(3)$ gauge theory on $R^{3}\times S^1$ with one PBC adjoint quark as a function of $q_1$ with $q_2=0$ $[ {\cal V}_g + {\cal V}_a^{0}] L^4$, for $m L=1.2$ (reconfined), $1.6$ (reconfined$\leftrightarrow$split), $2.0$ (split$\leftrightarrow$deconfined) and $3.0$ (deconfined). []{data-label="Fig_p_gapm_2D"}](2D_potential_gap_t10_m16p05.eps "fig:"){width="23.00000%"} ![ The one-loop effective potential of $SU(3)$ gauge theory on $R^{3}\times S^1$ with one PBC adjoint quark as a function of $q_1$ with $q_2=0$ $[ {\cal V}_g + {\cal V}_a^{0}] L^4$, for $m L=1.2$ (reconfined), $1.6$ (reconfined$\leftrightarrow$split), $2.0$ (split$\leftrightarrow$deconfined) and $3.0$ (deconfined). []{data-label="Fig_p_gapm_2D"}](2D_potential_gap_t10_m20p01.eps "fig:"){width="23.00000%"} ![ The one-loop effective potential of $SU(3)$ gauge theory on $R^{3}\times S^1$ with one PBC adjoint quark as a function of $q_1$ with $q_2=0$ $[ {\cal V}_g + {\cal V}_a^{0}] L^4$, for $m L=1.2$ (reconfined), $1.6$ (reconfined$\leftrightarrow$split), $2.0$ (split$\leftrightarrow$deconfined) and $3.0$ (deconfined). []{data-label="Fig_p_gapm_2D"}](2D_potential_gap_t10_m30.eps "fig:"){width="23.00000%"} ![ Contour plot of the one-loop effective potential of $SU(3)$ gauge theory on $R^{3}\times S^1$ with one PBC adjoint quark $[ ( {\cal V}_g )_{pert} + {\cal V}_a^{0} ] L^4$, for $m L=1.6$ and $1.8$ ($SU(2)\times U(1)$ split phase) as a function of $q_1$ and $q_2$. Thicker region indicates deeper region of the potential.[]{data-label="Fig_p_gapm_3D"}](potential_gap_m16p05_CP.eps "fig:"){width="40.00000%"} ![ Contour plot of the one-loop effective potential of $SU(3)$ gauge theory on $R^{3}\times S^1$ with one PBC adjoint quark $[ ( {\cal V}_g )_{pert} + {\cal V}_a^{0} ] L^4$, for $m L=1.6$ and $1.8$ ($SU(2)\times U(1)$ split phase) as a function of $q_1$ and $q_2$. Thicker region indicates deeper region of the potential.[]{data-label="Fig_p_gapm_3D"}](potential_gap_m18_CP.eps "fig:"){width="40.00000%"} ![$L^{-1}$-$m$ phase diagram for $SU(3)$ gauge theory on $R^{3}\times S^1$ with one PBC adjoint quark based on one-loop effective potential. D stands for “deconfined ($SU(3)$)", S for “split ($SU(2)\times U(1)$)" and R for “re-confined ($U(1)\times U(1)$)" phases. Phase transitions are first-order.[]{data-label="4d_phase_p"}](phase_diagra_gap_4D.eps){width="40.00000%"} ![ Schematic distribution plot of Polyakov loop $\Phi$ as a function of $\mathrm{Re}~\Phi$ and $\mathrm{Im}~\Phi$ for $SU(3)$ gauge theory on $R^{3}\times S^1$ with one PBC adjoint quark. []{data-label="Fig_pl_distribution"}](pl-distribution.eps){width="40.00000%"} We next turn on the chiral sector and consider the PNJL effective potential (\[PNJL-like\]). We investigate chiral properties of $(N_{f},N_{a})=(0,2)$ $SU(3)$ gauge theory on $R^{3}\times S^{1}$. [^1]. Before proceeding to the main topic, we discuss validity of the effective model (\[PNJL-like\]) for the purpose of studying chiral properties at weak-coupling region. Although we have no confined phase nor confined/deconfined phase transition in our model, the chiral restoration associated with the phase transition is correctly reproduced in this model for the known cases: $(N_{f},N_{a})=(2,0)$ with aPBC, $(N_{f},N_{a})=(2,0)$ with PBC and $(N_{f},N_{a})=(0,2)$ with aPBC. We depict behavior of the constituent mass for these cases in Fig. \[Fig\_chiral\], where the chiral phase transition takes place at some point for the three cases. The parameters are chosen so as to have the correct critical temperatures in finite-temperature $SU(3)$ gauge theory with aPBC quarks [@KL1; @EHS1], $\Lambda=0.63$ GeV and $g_{S}\Lambda^{2}=2.19$ for fundamental quarks and $\Lambda=23.22$ GeV and $g_{S}\Lambda^{2}=0.63$ for adjoint quarks [@NO1]. We are now convinced that the model can work to study chiral properties, and we go on to the main topic, $(N_{f},N_{a})=(0,2)$ with PBC. We calculate the PNJL effective potential in Eq. (\[PNJL-like\]) and search for the vacua for $q_{1}$, $q_{2}$ and $\sigma$. We depict the phase diagram for this case in Fig. \[Fig\_chi\_phase\]. Due to nonzero constituent mass, the whole phase diagram is shifted, and we have phase transitions even for a $m_{a}=0$ massless case. We also note that the critical lines are curved due to the dimensionful parameters introduced. Although our analysis is valid only for the weak-coupling regime, the phase diagram is qualitatively consistent with that of the lattice simulations [@MO1; @CD1] except that ours have no confinement/deconfinement phase transition. ![Constituent mass $m_{c}$ as a function of $L^{-1}$ for $(N_{f},N_{a})=(2,0)$ with aPBC, $(N_{f},N_{a})=(2,0)$ with PBC, $(N_{f},N_{a})=(0,2)$ with aPBC and $(N_{f},N_{a})=(2,0)$ with PBC with the bare mass fixed as $m=0$ GeV. We choose the parameter set as $\Lambda=0.63$ GeV and $g_{S}\Lambda^{2}=2.19$ for fundamental cases and $\Lambda=23.22$ GeV and $g_{S}\Lambda^{2}=0.63$ for adjoint cases. []{data-label="Fig_chiral"}](Fig_chiral.eps){width="40.00000%"} To look into chiral properties, we simultaneously depict the real part of VEV of Polyakov loop $\Phi$ and the constituent mass $m_{c}$ as a function of $L^{-1}$ with the bare mass fixed as $m=0$ GeV and $m=1$ GeV in Fig. \[Fig\_chi\_P\]. We here normalize the constituent mass as $m_{c}(L^{1})/m_{c}(L^{-1}=0)$. It is notable that chiral condensate, or equivalently, constituent quark mass does not undergo a clear transition even for large values of $L^{-1}$, and it gradually decreases. We thus have no clear chiral restoration transition while chiral symmetry is gradually restored in this theory. This result and the standard-PNJL result in [@NO1] are consistent with those of the lattice simulation [@CD1], which argues that the chiral restoration at weak coupling should occur at a quite small value of the compacted size $L$. The other notable point is that the chiral condensate undergoes quite small transitions coinciding with the deconfined/split and split/re-confined phase transitions. (It can be seen better in Fig. \[Fig\_chiral\] or the right panel for $m=1$ GeV in Fig. \[Fig\_chi\_P\].) This kind of the transition propagation is well studied in [@BCPG1; @Kashiwa1], but they may be too small to be observed in the lattice simulations. ![$L^{-1}$-$m$ phase diagram for $SU(3)$ gauge theory on $R^{3}\times S^1$ with two-flavor PBC adjoint quarks $(N_{f},N_{a})=(0,2)$ based on the PNJL-type effective potential. D stands for “deconfined", S for “split" and R for “re-confined" phases. Critical lines are first-order.[]{data-label="Fig_chi_phase"}](Fig_PD_m0.eps){width="40.00000%"} ![ VEV of Polyakov loop $\Phi$ (blue dashed) with $q_2=0$ and the constituent mass $m_{c}$ (red solid) as a function of $L^{-1}$ with the bare mass fixed as $m=0$ GeV (left) and $m=1$ GeV (right). The constituent mass is normalized as $m_{c}(L^{-1})/m_{c}(L^{-1}=0)$. The result indicates that the chiral symmetry is gradually restored without the clear phase transition. []{data-label="Fig_chi_P"}](Fig_T-dep_m0000.eps "fig:"){width="40.00000%"} ![ VEV of Polyakov loop $\Phi$ (blue dashed) with $q_2=0$ and the constituent mass $m_{c}$ (red solid) as a function of $L^{-1}$ with the bare mass fixed as $m=0$ GeV (left) and $m=1$ GeV (right). The constituent mass is normalized as $m_{c}(L^{-1})/m_{c}(L^{-1}=0)$. The result indicates that the chiral symmetry is gradually restored without the clear phase transition. []{data-label="Fig_chi_P"}](Fig_T-dep_m1000.eps "fig:"){width="40.00000%"} Let us briefly comment on non-perturbative deformation of the gluon potential discussed in Sec. \[sec:EP\]. As a result, in Appendix. \[sec:VSF\]2 we show a result for the deformation (\[g\_np\]). Indeed, with introducing the dimension parameter, for example $M=596$ MeV, the phase diagram in Fig. \[4d\_phase\_p\] is modified as Fig. \[4d\_phase\_np\]. Here the confined phase emerges at small $L^{-1}$ region, but it is connected with the re-confined phase through the small mass region. This result clearly shows that the gauge symmetry is broken as $SU(3) \to U(1) \times U(1)$ even at zero-temperature or infinite-$L$. We note that the similar result with the same deformation in the gluon potential is shown in Ref. [@NO1], where it is argued that the unified confined phase implies the volume-independence of the confined phase structure. For the other deformations, we have the same situation with explicit breaking of gauge symmetry. For our purpose of clarifying and classifying phases of gauge symmetry, the deformations are not appropriate although they may work as more phenomenological means. We next consider $(N_{f},N_{a})=(1,1)$ with PBC. We concentrate on the case with a massless fundamental quark, and the potential is given by ${\cal V}_{g}+{\cal V}_{f}^{\phi}(N_{f}=1, m_f=0)+{\cal V}_a^{\phi}(N_{a}=1, m_a=m)$. In this case, since the fundamental quark breaks the $Z_{3}$ center symmetry, the minima at ${\mathrm Re}~\Phi<0$ become true vacua in the deconfined and split phases as shown in Fig. \[Fig\_pl\_distribution2\]. (Fundamental matter with PBC moves the vacua to ${\mathrm Re}~\Phi<0$ direction.) In addition, we have no exact re-confiend phase since $\Phi=0$ vacuum cannot be chosen because of the center symmetry breaking in Fig. \[Fig\_pl\_distribution2\]. We term this unusual phase as “pseudo-reconfined phase", where the $SU(3)$ is broken to $U(1)\times U(1)$ and $\Phi$ takes a nonzero and negative value. The existence of this phase is consistent with the research on flavor-number dependence of gauge-symmetry-broken manners in Ref. [@Hat1]. In Fig. \[4d\_phase\_fp\], we depict the phase diagram for the same case. We emphasize that the split phase gets larger by introducing PBC fundamental quarks, compared to Fig. \[4d\_phase\_p\]. It is because the center symmetry breaking chooses $(q_{1},q_{2},q_{3})=(0,0.5,0.5)$ and its permutations as true vacua among all the possible minima Fig. \[Fig\_pl\_distribution2\], and makes it more stable than the center-symmetric case in Fig. \[Fig\_pl\_distribution\]. ![ Schematic distribution plot of Polyakov loop $\Phi$ for $SU(3)$ gauge theory on $R^{3}\times S^1$ with one PBC adjoint and one PBC massless fundamental quarks $(N_{f},N_{a})=(1,1)$. Points painted over stand for vacua in this case. $Z_{3}$ symmetry is broken, and some of the three minima are chosen as true vacua in deconfined and split phase. The Polyakov loop $\Phi$ in the pseudo-reconfined phase (PC) takes a nonzero and negative value.[]{data-label="Fig_pl_distribution2"}](pl-distribution_fp.eps){width="40.00000%"} ![ Phase diagram for $R^{3}\times S^1$ $SU(3)$ gauge theory with one PBC adjoint and one massless PBC fundamental quarks $(N_{f},N_{a})=(1,1)$ based on the one-loop effective potential. $m$ is the adjoint quark mass $m=m_{a}$ and $L^{-1}$ is the inverse of the compacted size. PC stands for “pseudo-reconfined" phase.[]{data-label="4d_phase_fp"}](phase_diagra_gapfp_00_4D.eps){width="40.00000%"} To look into the details of pseudo-reconfined phase and the phase transition to the split phase, we depict the expanded effective potential near the global minima as a function of $q_1$ with $q_2 =0$ in Fig. \[skew-rec4D\]. The left panel shows the result for the massless adjoint quark ($m_{a}=m=0$), which corresponds to the pseudo-confined phase. The minimum is not located at $(q_1,q_2)=(0.5,0)$(split case) nor at $(q_1,q_2)=(1/3,0)$(re-confined case). In the pseudo-confined phase we totally have six minima for $q_1$ and $q_2$ as $(q_{1},q_{2})\sim(0,0.4)$, $(0.4,0)$, $(-0.4,0.4)$, $(-0.4,0)$, $(0,-0.4)$, $(0.4,-0.4)$, which means that the vacua are given by the permutation of $(q_{1},q_{2},q_{3})\sim(0,0.4,-0.4)$. The right panel shows the first-order phase transition between pseudo-confined and split phases. Since the potential barrier at the phase transition is quite low, the fluctuation could break down clear phase transition. For the cases with $(N_{f}, N_{a})=(1,2)$ and $(N_{f}, N_{a})=(1,3)$ where the flavor of adjoint quarks is larger than that of fundamental quarks, the minima of the potential in the pseudo-reconfined phase becomes deeper, and we can observe the first-order phase transition more distinctly. ![Expanded effective potential of $R^{3}\times S^1$ $SU(3)$ gauge theory with one PBC adjoint and one massless PBC fundamental quarks as a function of $q_1$ with $q_2=0$. Left one shows the case with the pseudo-reconfined phase ($m=m_{a}=0$), where we have the minimum at $(q_{1},q_{2})\sim(0.4,0)$. Right one shows the first-order phase transition between the pseudo-reconfined and split phase at $m_{a}L=0.77$.[]{data-label="skew-rec4D"}](potential_4D_PC_m00.eps "fig:"){width="40.00000%"} ![Expanded effective potential of $R^{3}\times S^1$ $SU(3)$ gauge theory with one PBC adjoint and one massless PBC fundamental quarks as a function of $q_1$ with $q_2=0$. Left one shows the case with the pseudo-reconfined phase ($m=m_{a}=0$), where we have the minimum at $(q_{1},q_{2})\sim(0.4,0)$. Right one shows the first-order phase transition between the pseudo-reconfined and split phase at $m_{a}L=0.77$.[]{data-label="skew-rec4D"}](potential_4D_PC_to_S.eps "fig:"){width="40.00000%"} The chiral sector can be introduced by extending to $(N_{f}, N_{a})=(2,2)$. In this case we have two chiral sectors for fundamental and adjoint quarks, and we have arbitrariness how to implement four-point interactions and choose relative parameters. For example, we may consider the following forms of the four-point and eight-point interactions. $$\begin{aligned} (g_{S})_{f}[(\bar{\psi}_{f}\psi_{f})^{2}& +(\bar{\psi}_{f}i\gamma_{5}{\vec \tau}\psi_{f})^{2}] +(g_{S})_{a}[(\bar{\psi}_{a}\psi_{a})^{2} +(\bar{\psi}_{a}i\gamma_{5}{\vec \tau}\psi_{a})^{2}] \nonumber\\ \,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\, &+(g_{S})_{fa}[\{(\bar{\psi}_{f}\psi_{f})^2 +(\bar{\psi}_{f}i\gamma_{5}{\vec \tau}\psi_{f})^2 \}^2 \{(\bar{\psi}_{a} \psi_{a})^2 +(\bar{\psi}_{a}i\gamma_{5}{\vec \tau}\psi_{a})^2 \}^2],\end{aligned}$$ where $\psi_{f}$ and $\psi_{a}$ stand for fundamental and adjoint quark fields, and $(g_{S})_{f}, (g_{S})_{a}, (g_{S})_{fa}$ stand for fundamental, adjoint and mixing effective coupling. Even if we fix a form of the four-point interactions, we still have no criterion on how to set the parameters since there is no lattice study on this case either for aPBC or PBC. Thus we just show results of the chiral properties for two representative sets of the parameters. In either set, $g_a$ and $\Lambda$ are set the same value used in Fig. \[Fig\_chiral\] and the mixing term is set as $(g_{S})_{fa}= g_a/\Lambda^6$. As the first case we consider $(g_{S})_f = R_{\it Fierz} \times (g_{S})_a$, where $R_{\it Fierz}$ is the coefficient obtained from the Fierz transformation. We call it ”scenario A", where the fundamental chiral condensate becomes zero for all the region of $1/L$ while the adjoint chiral condensate has the qualitatively similar behavior to $(N_{f}, N_{a})=(0,2)$ case as shown in left-panel of Fig. \[Fig\_chi\_P1\]. As the second case we consider $(g_{S})_f = (8/3) \times (g_{S})_a$, which we call scenario B. In this case the fundamental chiral condensate has nonzero value at $1/L=0$ as shown in the right-panel of Fig. \[Fig\_chi\_P1\]. In this case, unlike the adjoint chiral condensate, the fundamental chiral condensate never reacts to the gauge symmetry phase transitions and the decreasing behavior becomes relatively gentle at large $1/L$ ($1/L \sim 5$ GeV). We consider that either of the scenarios of chiral properties for $(N_{f}, N_{a})=(2,2)$ will be detected in the on-going lattice simulation. ![ VEV of Polyakov loop $\Phi$ (black dotted), the adjoint fermion constituent mass $m_{a}$ (red solid) and the fundamental fermion constituent mass $m_{f}$ (blue dashed) as a function of $L^{-1}$ for the scenario A (left) and B (right). The constituent masses are normalized as by $2.32$ GeV[]{data-label="Fig_chi_P1"}](Fig_T-dep_m0000-8f.eps "fig:"){width="40.00000%"} ![ VEV of Polyakov loop $\Phi$ (black dotted), the adjoint fermion constituent mass $m_{a}$ (red solid) and the fundamental fermion constituent mass $m_{f}$ (blue dashed) as a function of $L^{-1}$ for the scenario A (left) and B (right). The constituent masses are normalized as by $2.32$ GeV[]{data-label="Fig_chi_P1"}](Fig_T-dep_m0000-8fS.eps "fig:"){width="40.00000%"} Besides the cases we have shown above, we also have other interesting choices of matters. As shown in [@Kouno1], the $SU(3)$ gauge theory with three fundamental flavors with flavored twisted boundary conditions $\phi=0, \,1/3,\, 2/3$ can accidentally keep the $Z_{3}$ center symmetry. For example, in the case of $(N_{f},N_{a})=(3,1)$ with the flavored twisted boundary conditions $\phi=0,\, 1/3,\, 2/3$ for fundamental fermions and PBC for the adjoint fermion, the distribution plot of $\Phi$ becomes symmetric as Fig. \[Fig\_pl\_distribution\] and we have exact re-confined phase although the theory contains fundamental quarks. Moreover, in [@Kouno2], the present authors and collaborators discuss the $SU(3)$ gauge theory with $(N_{f},N_{a})=(3,0)$ with the flavored twisted boundary conditions $\phi=0,\, 1/3,\, 2/3$ can lead to the spontaneous gauge symmetry breaking without adjoint quarks. This case is also fascinating as a future research topic. Phase structure in five dimensions {#sec:PS5} ================================== In this section, we discuss the vacuum and phase structure in $SU(3)$ gauge theories on $R^{4}\times S^{1}$. Since the five-dimensional gauge theory is not renormalizable, we cannot discuss its non-perturbative aspects in a parallel way to the four-dimensional case. Thus we will concentrate on the one-loop part of the effective theory, and will not consider the contribution from the chiral sector. It is sufficient for our purpose of investigating the phase structure at weak-coupling. We consider the five-dimensional one-loop effective potential $[ {\cal V}_g + {\cal V}_{f}^{0}(N_{f},m_{f}) + {\cal V}_a^{0}(N_{a},m_{a}) ] L^5$. We first look at the case with $(N_{f},N_{a})=(0,1)$ with PBC. The qualitative properties are common with the four-dimensional case. Figure \[Fig\_p\_gapm\_D5\_2D\] shows the effective potential $[{\cal V}_g+{\cal V}_a^{0}(N_{a}=1, m)]L^5$ as a function of $q_{1}$ with $q_2=0$ fixed. The adjoint quark mass is set as $m L=0.5$, $1.3$ and $2.0$ from left to right panels. The three cases correspond to re-confined (left), split (center), and deconfined (right) phase. The contour plots at $m L=0.5$(reconfined) and $mL =1.3$(split) are depicted in Fig. \[Fig\_p\_gfapm\_D5\_3D\]. The manners of $SU(3)$ symmetry breaking and the distribution of Polyakov line $\Phi$ in each phase are the same as the four-dimensional case; $SU(3)\to U(1)\times U(1)$ in the re-confined phase and $SU(3)\to SU(2)\times U(1)$ in the split phase. We depict the phase diagram in $L^{-1}$-$m$ for this case in Fig. \[Fig\_PD\_gap\]. The qualitative configuration of the three phases in the phase diagram is indifferent from the four-dimensional case although the split phase is smaller in five dimensions. ![The one-loop effective potential of $SU(3)$ gauge theory on $R^{4}\times S^1$ with one PBC adjoint quark as a function of $q_1$ with $q_2=0$ $({\cal V}_g+{\cal V}_a^{0}(1, m)) L^5$. Depicted for $m L=0.5$ (re-confined), $1.3$ (split) and $2.0$ (deconfined). []{data-label="Fig_p_gapm_D5_2D"}](2D_potential_gap_t10_m05_5D_c.eps "fig:"){width="30.00000%"} ![The one-loop effective potential of $SU(3)$ gauge theory on $R^{4}\times S^1$ with one PBC adjoint quark as a function of $q_1$ with $q_2=0$ $({\cal V}_g+{\cal V}_a^{0}(1, m)) L^5$. Depicted for $m L=0.5$ (re-confined), $1.3$ (split) and $2.0$ (deconfined). []{data-label="Fig_p_gapm_D5_2D"}](2D_potential_gap_t10_m13_5D_c.eps "fig:"){width="30.00000%"} ![The one-loop effective potential of $SU(3)$ gauge theory on $R^{4}\times S^1$ with one PBC adjoint quark as a function of $q_1$ with $q_2=0$ $({\cal V}_g+{\cal V}_a^{0}(1, m)) L^5$. Depicted for $m L=0.5$ (re-confined), $1.3$ (split) and $2.0$ (deconfined). []{data-label="Fig_p_gapm_D5_2D"}](2D_potential_gap_t10_m20_5D_c.eps "fig:"){width="30.00000%"} ![ Contour plot of the one-loop effective potential of $SU(3)$ gauge theory on $R^{4}\times S^1$ with one PBC adjoint quark $({\cal V}_g+{\cal V}_a^{0}(N_{a}=1,m_{a}=m)) L^5$ for $mL=0.5$ and $1.3$ as a function of $q_1$ and $q_2$. []{data-label="Fig_p_gfapm_D5_3D"}](potential_gap_m05_CP_5D_c.eps "fig:"){width="40.00000%"} ![ Contour plot of the one-loop effective potential of $SU(3)$ gauge theory on $R^{4}\times S^1$ with one PBC adjoint quark $({\cal V}_g+{\cal V}_a^{0}(N_{a}=1,m_{a}=m)) L^5$ for $mL=0.5$ and $1.3$ as a function of $q_1$ and $q_2$. []{data-label="Fig_p_gfapm_D5_3D"}](potential_gap_m13_CP_5D_c.eps "fig:"){width="40.00000%"} ![ $L^{-1}$-$m_{a}$ phase diagram for $SU(3)$ gauge theory on $R^{4}\times S^1$ with one PBC adjoint quark based on the one-loop effective potential. []{data-label="Fig_PD_gap"}](phase_diagra_gap_5D.eps){width="40.00000%"} We next consider the case for $(N_{f},N_{a})=(1,1)$ with PBC. We again concentrate on $m_{f}=0$. Due to explicit breaking of $Z_{3}$ center symmetry, $\Phi<0$ minima are chosen as vacua in the deconfined and split phases in the same way as the four dimensional case (Fig. \[Fig\_pl\_distribution2\]). It is notable that the split phase is again enlarged by introducing fundamental fermions, but more effectively in five dimensions. Enhancement of the split phase due to the fundamental matter in five dimensions is sensitive to choice of parity pairs, or equivalently, choice of the number of flavors. For $(N_{f},N_{a})=(1,1)$ without the parity pair, the pseudo-confined phase disappears and the split phase becomes the unique gauge-broken phase as shown in Fig. \[Fig\_PD\_gapfp\]. This result is consistent with that of the massless case ($m_{a}=0$) in Ref. [@H3; @H4]. On the other hand, when we introduce parity pairs for $(N_{f}, N_{a})=(1,1)$ or equivalently consider $(N_{f},N_{a})=(2,2)$, the split phase becomes wider than that in Fig. \[Fig\_PD\_gap\] but the pseudo-reconfined phase still survives as shown in Fig. \[Fig\_PD\_gapfp2\]. In the five-dimensional pseudo-confined phase we again have six minima for $q_1$ and $q_2$ as $(q_{1},q_{2})\sim(0,0.4)$, $(0.4,0)$, $(-0.4,0.4)$, $(-0.4,0)$, $(0,-0.4)$, $(0.4,-0.4)$, which indicates that the minima are given by the permutation of $(q_{1},q_{2},q_{3})\sim(0,0.4,-0.4)$. We depict the expanded effective potential in Fig. \[skew-rec5D\]. The left panel shows the massless case ($m_{a}L=0$), which corresponds to the pseudo-reconfined phase. The right panel shows the first-order phase transition between the pseudo-reconfined and split phases ($m_{a}L=1.18$). In the cases with $(N_{f},N_{a})=(1,2)$ or $(N_{f},N_{a})=(1,3)$, the potential minima in the pseudo-reconfined phase becomes deeper and the phase transition gets more distinct. ![ $L^{-1}$-$m_{a}$ phase diagram for $SU(3)$ gauge theory on $R^{4}\times S^{1}$ with one adjoint and one massless fundamental quarks ($(N_{f},N_{a})=(1,1)$, $m_{f}=0$, PBC) based on the one-loop effective potential. []{data-label="Fig_PD_gapfp"}](phase_diagra_gapfp_5D.eps){width="40.00000%"} ![ $L^{-1}$-$m_{a}$ phase diagram for $SU(3)$ gauge theory on $R^{4}\times S^{1}$ with a set of parity pairs of $(N_{f},N_{a})=(1,1)$, or equivalently $(N_{f},N_{a})=(2,2)$, based on the one-loop effective potential. []{data-label="Fig_PD_gapfp2"}](phase_diagra_gapfp_2times_5D.eps){width="40.00000%"} ![Expanded 5D effective potential of $SU(3)$ gauge theory on $R^{4}\times S^{1}$ with a set of parity pairs of $(N_{f},N_{a})=(1,1)$ or $(N_{f},N_{a})=(2,2)$ as a function of $q_1$ with $q_2=0$. Left one shows the pseudo-confined phase ($m=m_{a}=0$), where we have the minimum at $(q_{1},q_{2})\sim(0.4,0)$. Right one shows the first-order phase transition between the pseudo-reconfined and split phase at $m_{a}L=1.18$.[]{data-label="skew-rec5D"}](potential_5D_PC_m00.eps "fig:"){width="40.00000%"} ![Expanded 5D effective potential of $SU(3)$ gauge theory on $R^{4}\times S^{1}$ with a set of parity pairs of $(N_{f},N_{a})=(1,1)$ or $(N_{f},N_{a})=(2,2)$ as a function of $q_1$ with $q_2=0$. Left one shows the pseudo-confined phase ($m=m_{a}=0$), where we have the minimum at $(q_{1},q_{2})\sim(0.4,0)$. Right one shows the first-order phase transition between the pseudo-reconfined and split phase at $m_{a}L=1.18$.[]{data-label="skew-rec5D"}](potential_5D_PC_to_S.eps "fig:"){width="45.00000%"} In the end of this section we comment on the other aspect of gauge theory with a compacted dimension. If we regard the compacted direction as time direction, the boundary condition for the Polyakov-loop phases can be seen as imaginary chemical potential. The periodic and anti-periodic boundary conditions correspond to different Roberge-Weiss transition points on the QCD phase diagram. From this viewpoint, it is clear that fundamental fermions with PBC works to move the vacua to $\mathrm{Re}~\Phi<0$ direction as shown in Fig. \[Fig\_pl\_distribution2\] while those with aPBC move it to $\mathrm{Re}~\Phi>0$ direction as shown in Fig. \[Fig\_pl\_distribution3\] ![ Schematic distribution plot of Polyakov loop $\Phi$ for SU(3) gauge theory with one PBC adjoint and one aPBC fundamental quarks. Points painted over stand for vacua in this case. $Z_{3}$ symmetry is broken in the different manner from Fig. \[Fig\_pl\_distribution2\].[]{data-label="Fig_pl_distribution3"}](pl-distribution_fa.eps){width="40.00000%"} Observables comparable to lattice {#sec:OB} ================================= In this section we discuss observables quantitatively in our study, which can be compared to existing and on-going lattice simulations. [*Mass spectrum*]{} : We fist consider the mass spectrum in the gauge-broken phases, $SU(2)\times U(1)$ and $U(1)\times U(1)$ phases. As shown in Ref. [@H1; @H2; @H3; @H4], the Kalza-Klein spectrum for gauge bosons is given by $$M_{n}^{2}\,=\, {1\over{L^{2}}}(n+q_{i}-q_{j})^{2},$$ where $n$ stands for KK index. We here focus on $n=0$ modes and the case with zero quark mass. As long as $q_{i}=q_{j}$ $(i\not= j)$, these modes are massless. On the other hand, when $q_{i}\not =q_{j}$ is realized at the vacuum, some or all of $n=0$ modes become massive. This phenomenon is consistent with the Higgs mechanism with the gauge boson obtaining mass due to gauge symmetry breaking. In our study, we find the two gauge-broken phase $SU(2)\times U(1)$ and $U(1)\times U(1)$ for $SU(3)$ gauge theory on $R^{3}\times S$. In the $SU(2)\times U(1)$ phase of $SU(3)$ gauge theory, where we originally have 8 massless gauge bosons, we have 4 massive modes, whose mass is given by $$M_{SU(2)\times U(1)}^{2}\,=\,{1\over{4L^{2}}}, \label{su2u1m}$$ where we substitute $(q_{1}, q_{2},q_{3})=(0, 0.5,0.5)$ to $q_{i}$. This result of the gauge boson mass is common for the adj. case and the adj.-fund. case. Irrespective of the matter field, $SU(2)\times U(1)$ phase has the five gauge bosons with the mass (\[su2u1m\]). Things change in $U(1)\times U(1)$ phase. In $U(1)\times U(1)$ phase (re-confined phase) for $(N_{f}, N_{a})=(0,1)$, we have 6 massive modes, whose masses are given by $$\begin{aligned} M_{U(1)\times U(1)}^{2}(N_f =0, N_a =1)=\left\{ \begin{array}{ll} &{1\over{9L^{2}}}\,\,\,\,\,\,\,({\rm 4\,\, modes}) \\ & {4\over{9L^{2}}}\,\,\,\,\,\,\,({\rm 2\,\, modes}) \\ \end{array} \right.\end{aligned}$$ For $(N_{f}, N_{a})=(1,1)$, the $U(1)\times U(1)$ phase (pseudo-confined phase) again has 6 massive modes, but the masses are different from the above as $$\begin{aligned} M_{U(1)\times U(1)}^{2}(N_f =1, N_a =1)=\left\{ \begin{array}{ll} &{4\over{25L^{2}}}\,\,\,\,\,\,\,({\rm 4\,\, modes}) \\ & {16\over{25L^{2}}}\,\,\,\,\,\,\,({\rm 2\,\, modes}) \\ \end{array} \right.\end{aligned}$$ Thus, if we look into gauge mass as a function of the compactification scale, there should be clear difference between different choices of the matter field even in the same symmetry phase. It is good indication not only for the gauge symmetry breaking but also for specifying the phases. These results are the case with both four-dimensional and five-dimensional cases. We note that our results of mass spectrum are valid for small $L$ (weak-coupling), and the lattice simulation for relatively weak-coupling can reproduce our results. [*Polyakov-loop*]{} : As we have discussed, the trace of compact-dimensional Polyakov loop is also a good indication of the exotic phases [@CD1]. We first show how the exotic phases seen in the lattice QCD with PBC adjoint fermions [@CD1] is interpreted from the gauge symmetry breaking phases. Fig. \[PoD\] is the distribution plot of the Polyakov loop in the lattice simulations. Each point corresponds to each of the results for different gauge configurations. We also depict corresponding results in our analytical calculation of Fig. \[Fig\_pl\_distribution\] on the figure. By comparing them, we find that each of the cases can be understood as one of $SU(3)$ deconfined, $SU(2)\times U(1)$ split or $U(1)\times U(1)$ re-confined phases. (The strong-coupling confined phase cannot be reproduced in our weak-coupling study.) The lattice artifacts makes $Z_{3}$ symmetry no-exact, and one of $Z_{3}$ minima is selected depending on $\beta$. This result means that we can predict distribution of Polyakov-loop for other choices of matter fields in the on-going lattice simulations. For example, $(N_{f},N_{a})=(1,1)$ case should have the distribution of Polyakov-loop depending on $\beta$ as shown in Fig. \[PoD2\]. The explicit $Z_{3}$ symmetry breaking shifts the true vacua to the ${\rm Re}~\Phi<0$ side in the complex space. As a result, the vacuum with ${\rm Re}~\Phi<0$ should be chosen in $SU(2)\times U(1)$ phase while the $U(1)\times U(1)$ phase also has ${\rm Re}~\Phi<0$. This behavior of the Polyakov-loop distribution will be observed in the on-going simulation for the case with both adj. and fund. quarks. As shown in Fig. \[Fig\_pl\_distribution3\], when we take aPBC for fundamental quarks instead of PBC, the behavior is changed as the vacua at ${\rm Re}~\Phi>0$ are chosen. These predictions are valid for both four-dimensional and five-dimensional theories. We note that our results catch physical properties at least at small $L$ (weak-coupling), and the lattice simulation for relatively weak-coupling can reproduce our results. ![Comparison between distribution plot of Polyakov loop $\Phi$ on the lattice [@CD1] and that of the one-loop effective potential for $SU(3)$ gauge theory on $R^{3}\times S^1$ with PBC adjoint quarks. (Watermarks stand for distribution of plot points on the lattice [@CD1].) Apart from the strong-coupling confined phase, all of the specific behavior can be interpreted as the phases we found in our analytical calculations. []{data-label="PoD"}](PoD.eps){width="60.00000%"} ![Prediction of distribution plot of Polyakov loop $\Phi$ based on the one-loop effective potential for $SU(3)$ gauge theory on $R^{d}\times S^1$ with PBC adjoint and fundamental quarks.[]{data-label="PoD2"}](PoD2.eps){width="80.00000%"} [*Chiral condensate and chiral susceptibility*]{} : The special behavior of constituent mass (chiral condensate) as a function of $1/L$, which we calculated in the PNJL model, is also peculiar to the theories with spontaneous gauge symmetry breaking. And it can be detected on the lattice. Even if it is not easy to look into exact value and details of constituent mass on the lattice, the chiral susceptibility works to investigate the subtle behavior. Chiral susceptibility in our model is defined as $$\chi={C_{ll}\over{C_{qq}C_{ll}-C_{ql}^{2}}},$$ with $$C_{qq} = {L^{4}\over{\Lambda}} {\partial^{2}\mathcal{V}\over{\partial m^{2}}},\,\,\,\,\,\,\, C_{ll} = {L^{4}\over{\Lambda^{3}}} {\partial^{2}\mathcal{V}\over{\partial \Phi^{2}}},\,\,\,\,\,\,\, C_{qq} = {L^{4}\over{\Lambda^2}} {\partial^{2}\mathcal{V}\over{\partial m \partial \Phi}},$$ where $m$ stands for the adjoint or fundamental quark mass. For example, the chiral behavior in the $SU(3)$ gauge theory with PBC adjoint matter in Fig. \[Fig\_chi\_P\] indicates that the chiral condensate reacts to the gauge-symmetry phase transitions slightly and it slowly decreases in the re-confined phase with $1/L$ getting large. Chiral susceptibility in this case should have two discrete changes at small $1/L$. The chiral properties in the $SU(3)$ gauge theory with PBC adjoint and fundamental matters in Fig. \[Fig\_chi\_P1\] can have more special behavior. Whichever of the scenarios A and B are realized on the lattice, peculiar behaviors of chiral condensate can be detected on the lattice. We note that our results from the PNJL model are valid for $0\ll 1/L< \Lambda$, which roughly means the region for $SU(2)\times U(1)$ phase and $U(1)\times U(1)$ phases at $1/L<\Lambda$ GeV. We expect the lattice results for these two phases reproduce our results. Summary {#sec:sum} ======= In this paper we have studied the phase diagram for $SU(3)$ gauge theories with a compact spatial dimension by using the effective models, with emphasis on the dynamical gauge symmetry breaking. We show that introduction of adjoint matter with periodic boundary condition leads to the unusual phases with spontaneous gauge symmetry breaking, whose ranges are controlled by introducing fundamental matter. We also study chiral properties in these theories and show that the chiral condensate remains nonzero even at a small compacted size. In Sec. \[sec:EP\], we developed our setup based on one-loop effective potential and four-point fermion interactions. The effective potential is composed of the gluon, quark and chiral sectors. The total effective potential corresponds to that of the PNJL model with one-loop gluon potential. This setup is effective enough to investigate vacuum structure and chiral properties at weak-coupling or small size of the compact dimension. In Sec. \[sec:PS4\], we elucidated the vacuum and phase structure in $SU(3)$ gauge theory on $R^{3}\times S^{1}$. The theory with PBC adjoint quarks has three different phases in the $L^{-1}$-$m$ space, including the deconfined phase ($SU(3)$, nonzero $\Phi$), the split phase ($SU(2)\times U(1)$, nonzero $\Phi$) and the re-confined phase ($U(1)\times U(1)$, zero $\Phi$ with nontrivial global minima). The configuration of these phases in the phase diagram is consistent with that of the lattice calculations although we have no confined phase. By using the PNJL effective potential, we argued that chiral symmetry is slowly restored without clear phase transition when the size of the compact dimension is decreased. In this section we also studied vacuum and phase structure for the case with both fundamental and adjoint matters, and showed that the split phase is generically widened by adding PBC fundamental quarks. We consider that it is because one of the three possible minima for the split phase becomes more stable due to the center symmetry breaking. We showed that another $U(1)\times U(1)$ phase with a negative value of $\Phi$ emerges in this case (pseudo-reconfined phase). In Sec. \[sec:PS5\], we studied the vacuum and phase structure in $SU(3)$ gauge theory on $R^{4}\times S^{1}$. In the five-dimensional case we concentrate on the one-loop part of the effective potential. The theory with PBC adjoint quarks again has the split ($SU(2)\times U(1)$) and re-confined ($U(1)\times U(1)$) phases. Introduction of fundamental quarks works to enlarge the split phase more effectively than the four-dimensional case. Especially in the case with one adjoint and one fundamental quarks without parity pairs, the split phase overcomes the pseudo-reconfined phase and becomes a unique gauge-symmetry-broken phase. In Sec. \[sec:OB\], we discuss observables comparable to the lattice simulations. We list up observables including particle mass spectrum, in particular gauge boson mass (Lowest KK spectrum), Polyakov-loop in the compact dimension, constituent mass. They can be good indications of exotic phase and properties both in the existing and on-going simulations. All through this paper, we treat PBC and aPBC cases in a parallel manner. We note that the reference [@U1; @U2] argues that gauge theory with PBC fermions has no thermal fluctuation, and thermal interpretation is inappropriate in such a case. This means that all the results here should be understood as topological phenomena. In the end of the paper, we discuss whether the lattice simulation can check our predictions. One of our main results is that enlargement of the split phase in the presence of fundamental fermions. This property would be observed in the on-going lattice simulations on four-dimesional gauge theory with a compact direction [@Hgroup1]. The pseudo-reconfined phase can be also observed where the VEV of Polyakov loop becomes negative but different from that of the split phase. On the other hand, it seems difficult to show the first-order phase transition between the pseudo-reconfined and split phases. The lattice calculation can check whether chiral condensate remains finite in the re-confined phase even at a very small compacted size. The small chiral transitions coinciding with the gauge-symmetry phase transitions are subtle. If the lattice simulation succeeds to measure chiral susceptibility with high precision, this phenomena may be able to be observed. During preparing this draft, the collaboration [@Hgroup1] kindly informed us that it also performed perturbative calculations on the phase diagram in gauge theory on $R^{3}\times S^{1}$ and $R^{4}\times S^{1}$. It could be valuable for readers to compare the results from the two independent groups. The authors are grateful to E. Itou for giving them chances to have interest in the present topics and reading the draft carefully. They are thankful to Y. Hosotani, H. Hatanaka and H. Kouno for their careful reading of this manuscript and valuable comments. They also thank G. Cossu and J. Noaki for the kind co-operation. K. K. thanks H. Nishimura for useful comments. K.K. is supported by RIKEN Special Postdoctoral Researchers Program. T.M. is supported by Grand-in-Aid for the Japan Society for Promotion of Science (JSPS) Postdoctoral Fellows for Research Abroad(No.24-8). Vacua and phase structures for other cases {#sec:VSF} ========================================== Gauge symmetric cases --------------------- In this appendix, we discuss the vacuum structure for the several gauge-symmetric cases in $R^{3}\times S^{1}$ $SU(3)$ gauge theory. We concentrate on massless cases as $m_a = m_{f} =0$. The potential contour plot for $(N_f, N_a)=(0,0)$ (pure gauge theory) is shown in Fig. \[Fig\_p\_g\_3D\]. The effective potential has the minima at $(q_1,q_2)=(0,0)$, $(1/3,1/3)$, $(-1/3,-1/3)$. Since it means that the three Polyakov-line phases are equivalent in the vacuum $(q_{1}=q_{2}=q_{3})$, the $SU(3)$ gauge symmetry is intact. The case with one fundamental fermion with anti-periodic boundary condition is shown in Fig. \[Fig\_p\_gfa\_3D\]. The fundamental quark breaks $Z_3$ symmetry explicitly and thus, two of the three minima become the meta-stable. If the boundary condition of fermion is changed to the periodic one, the global minima move to $(\pm 1/3, \pm 1/3)$ as shown in Fig. \[Fig\_p\_gfp\_3D\]. On the other hand, the case for an adjoint fermion with anti-periodic boundary condition is similar to the pure gauge case as shown in Fig. \[Fig\_p\_gaa\_3D\] since it keeps the $Z_{3}$ center symmetry. ![ The contour plot of one-loop effective potential for pure $SU(3)$ gauge theory, ${\cal V}_g L^4$ as a function of $q_1$ and $q_2$. Thicker region stands for deeper region of the effective potential. []{data-label="Fig_p_g_3D"}](potential_g_CP.eps){width="40.00000%"} ![ The contour plot of one-loop effective potential for $R^{3}\times S^{1}$ $SU(3)$ gauge theory with one massless fundamental fermion with anti-periodic boundary condition ($(N_{f},N_{a})=(1,0)$ with aPBC) $({\cal V}_g+{\cal V}_f^{1/2}) L^4$. []{data-label="Fig_p_gfa_3D"}](potential_gfa_CP.eps){width="40.00000%"} ![ The contour plot of one-loop effective potential for $R^{3}\times S^{1}$ $SU(3)$ gauge theory with one fundamental fermion with periodic boundary condition ($(N_{f},N_{a})=(1,0)$ with PBC) $({\cal V}_g+{\cal V}_f^{0}) L^4$. []{data-label="Fig_p_gfp_3D"}](potential_gfp_CP.eps){width="40.00000%"} ![ The contour plot of one-loop effective potential for $R^{3}\times S^{1}$ $SU(3)$ gauge theory with the $N_a = 1$ adjoint fermion with anti-periodic boundary condition ($(N_{f},N_{a})=(0,1)$ with aPBC) $({\cal V}_g+{\cal V}_a^{1/2}) L^4$. []{data-label="Fig_p_gaa_3D"}](potential_gaa_CP.eps){width="40.00000%"} Phase diagram with non-perturbative deformation ----------------------------------------------- In Fig. \[4d\_phase\_np\], we depict the phase diagram for $R^{3}\times S^{1}$ $SU(3)$ gauge theory with $(N_{f}, N_{a})=(0,1)$ with PBC based on the nonperturbatively deformed gluonic potential. As an example, we consider the following modification from the perturbative potential in [@MO3; @MeMO1; @MeO1; @NO1]: $$\begin{aligned} {\cal V}_{g}^\mathrm{np} &=- \frac{2}{L^4 \pi^2} \sum_{i,j=1}^N \sum_{n=1}^{\infty} \Bigl( 1 - \frac{1}{N} \delta_{ij} \Bigr) \frac{\cos( 2 n \pi q^{ij})}{n^4} \,+\, \frac{M^2}{2\pi^2L^2} \sum_{i,j=1}^N \sum_{n=1}^\infty \Bigl( 1 - \frac{1}{N} \delta_{ij} \Bigr) \frac{\cos(2n\pi q^{ij})}{n^2}\end{aligned}$$ where $M$ is the mass-dimension 1 parameter. We set the scale parameter as $M=596$ MeV. The confined phase and the first-order phase transition show up, but it is merged into the gauge-broken re-confined phase. We note that the $SU(3)$ gauge symmetry is explicitly broken in the confined (re-confined) phase. ![Phase diagram for $R^{3}\times S^{1}$ SU(3) theory with one PBC adjoint quark based on the effective potential with non-perturbative deformation ($M=596$ MeV). C stands for “confined", D stands for “deconfined" and S for “split" phases. Compared to the perturbative one-loop case, the confined phase emerges but merges into the re-confined phase.[]{data-label="4d_phase_np"}](phase_diagra_gap_4D_with_NPE.eps){width="40.00000%"} [99]{} ATLAS Collaboration, Phys. Lett. B [**716**]{} (2012) 1. CMS Collaboration, Phys. Lett. B [**716**]{} (2012) 30. N. Manton, Nucl.Phys.B [**158**]{} (1979) 141. Y. Hosotani, Phys.Lett.B [**126**]{}, 309 (1983). Y. Hosotani, Annals Phys [**190**]{}, 233 (1989). A. T. Davies and A. McLachlan, Phys. Lett. B [**200**]{} (1988) 305; Nucl. Phys. B [**317**]{} (1989) 237. H. Hatanaka, T. Inami and C. S. Lim, Mod. Phys. Lett. A [**13**]{} (1998) 2601. H. Hatanaka, Prog. Theor. Phys. [**102**]{} (1999) 407 \[arXiv:hep-th/9905100\]. Y. Hosotani, Proceedings for DSB 2004, Nagoya (2005) \[arXiv:hep-ph/0504272\]. Y. Hosotani, Proceedings for GUT 2012, Kyoto (2012) \[arXiv:1206.0552\]. K. Agashe, R. Contino and A. Pomarol, Nucl. Phys. B [**719**]{} (2005) 165. N. Maru, K. Takenaga, Phys.Lett.B [**637**]{}, 287 (2006) \[arXiv:hep-ph/0602149\]. Y. Hosotani, N. Maru, K. Takenaga and T. Yamashita, Prog. Theor. Phys. [**118**]{} (2007) 1053 \[arXiv:0709.2844\]. M. Sakamoto, K. Takenaga, Phys.Rev.D [**76**]{}, 085016 (2007) \[axXiv:0706.0071\]. Y. Hosotani, K. Oda, T. Ohnuma and Y. Sakamura, Phys. Rev. D[**78**]{} (2008) 096002; [**79**]{} (2009) 079902. M. Sakamoto and K. Takenaga, Phys. Rev. D [**80**]{} (2009) 085016 \[arXiv:0908.0987\]. Y. Hosotani, P. Ko and M. Tanaka, Phys. Lett. B [**680**]{} (2009) 179. K. Ishiyama, M. Murata, H. So and K. Takenaga, Prog. Theor. Phys. [**123**]{} (2010) 257 \[arXiv:0911.4555\] Y. Hosotani, S. Noda and N. Uekusa, Prog. Theor. Phys. [**123**]{} (2010) 757. Y. Hosotani, M. Tanaka and N. Uekusa, Phys. Rev. D [**82**]{} (2011) 115024. Y. Hosotani, M. Tanaka and N. Uekusa, Phys. Rev. D [**84**]{} (2011) 075014. H. Hatanaka and Y. Hosotani, Phys. Lett. B [**713**]{} (2012) 481. N. Irges and F. Knechtli, K. Yoneyama, \[arXiv:1212.5514\]. M. Unsal, Phys. Rev. Lett. [**100**]{}, 032005 (2008) \[arXiv:0708.1772\]. M. Unsal, Phys. Rev. D [**80**]{}, 065001 (2009) \[arXiv:0709.3269\]. M. C. Ogilvie, \[arXiv:1211.2843\]. J. C. Myers, M. C. Ogilvie, Phys.Rev.D [**77**]{}, 125030 (2008) \[arXiv:0707.1869\]. J. C. Myers, M. C. Ogilvie, PoS Lattice2008, 201 (2008) \[arXiv:0809.3964\]. J. C. Myers, M. C. Ogilvie, JHEP [**07(2012)**]{}, 095 (2009) \[arXiv:0903.4638\]. H. Nishimura, M. C. Ogilvie, Phys.Rev.D [**81**]{}, 014018 (2010) \[arXiv:0911.2696\]. G. Cossu, M. D’Elia, JHEP [**07(2009)**]{}, 048 (2009) \[arXiv:0904.1353\]. M. Creutz, Phys.Rev.Lett [**43**]{}, 553 (1979). H. Kawai, M. Nio and Y. Okamoto, Prog. Theor. Phys. [**88**]{} (1992) 341. J. Nishimura, Mod. Phys. Lett. A [**11**]{} (1996) 3049 \[arXiv:hep-lat/9608119\]. B. B. Beard et al., Nucl. Phys. Proc. Suppl. [**63**]{} (1998) 775 \[arXiv:hep-lat/9709120\]. C. B. Lang, M. Pilch and B. S. Skagerstam, Int. J. Mod. Phys. A [**3**]{} (1988) 1423. S. Chandrasekharan and U. J. Wiese, Nucl. Phys. B [**492**]{} (1997) 455 \[arXiv:hep-lat/9609042\]. R. Brower, S. Chandrasekharan and U. J. Wiese, Phys. Rev. D [**60**]{} (1999) 094502 \[arXiv:hep-th/9704106\]. S. Ejiri, J. Kubo and M. Murata, Phys. Rev. D [**62**]{} (2000) 105025 \[arXiv:hep-ph/0006217\]. S. Ejiri, S. Fujimoto and J. Kubo, Phys. Rev. D [**66**]{} (2002) 036002 \[arXiv:hep-lat/0204022\]. P. Dimopoulos, K. Farakos, A. Kehagias and G. Koutsoumbas, Nucl. Phys. B [**617**]{} (2001) 237 \[arXiv:hep-th/0007079\]. P. Dimopoulos, K. Farakos, C. P. Korthals-Altes, G. Koutsoumbas and S. Nicolis, JHEP [**02(2001)**]{} 005 \[arXiv:hep-lat/0012028\]. P. Dimopoulos, K. Farakos and G. Koutsoumbas, Phys. Rev. D [**65**]{} (2002) 074505 \[arXiv:hep-lat/0111047\]. K. Farakos, P. de Forcrand, C. P. Korthals Altes, M. Laine and M. Vettorazzo, Nucl. Phys. B 655 (2003) 170 \[arXiv:hep-ph/0207343\]. N. Irges and F. Knechtli, Nucl. Phys. B [**719**]{} (2005) 121 \[arXiv:hep-lat/0411018\]. N. Irges and F. Knechtli, hep-lat/0604006. N. Irges and F. Knechtli, Nucl. Phys. B [**775**]{} (2007) 283 \[arXiv:hep-lat/0609045\]. N. Irges, F. Knechtli and M. Luz, JHEP [**08(2007)**]{} 028 \[arXiv:0706.3806\]. F. Knechtli, N. Irges, M. Luz, J.Phys.Conf.Ser. [**110**]{},102006 (2008) \[arXiv:0711.2931\]. N. Irges and F. Knechtli, Nucl. Phys. B [**822**]{} (2009) 1 \[arXiv:0905.2757\]. N. Irges and F. Knechtli, Phys. Lett. B [**685**]{} (2010) 86 \[arXiv:0910.5427\]. L. Del Debbio, E. Kerrane and R. Russo, Phys. Rev. D [**80**]{} (2009) 025003 \[arXiv:0812.3129\]. P. de Forcrand, A. Kurkela, M. Panero, JHEP [**06(2010)**]{}, 050 (2010) \[arXiv:1003.4643\]. F. Knechtli, M. Luz, A. Rago, Nucl. Phys. B [**856**]{} (2012) 74 \[arXiv:1110.4210\]. L. Del Debbio, A. Hart, E. Rinaldi, JHEP [**07(2012)**]{} 178 \[arXiv:1203:2116\]. N. Irges and F. Knechtli, K. Yoneyama, Nucl.Phys.B [**865**]{} (2012) 541 \[arXiv:1206.4907\]. A. Gocksch and M. Ogilvie, Phys. Rev. D31, 877 (1985). K. Fukushima, Phys. Lett. B591, 277 (2004), \[arXiv:hep-ph/0310121\]. S. P. Klevansky, Rev. Mod. Phys. 64, 649 (1992). T. Hatsuda and T. Kunihiro, Phys. Rept. 247, 221 (1994), \[arXiv:hep-ph/9401310\]. C. Ratti, S Roessner, M. A. Thaler, W. Weise, Eur.Phys.J.C [**49**]{}, 213 (2007) \[arXiv:hep-ph/0609218\]. T. Eguchi and H. Kawai, Phys. Rev. Lett. 48, 1063 (1982). P. Kovtun, M. Unsal, L. G. Yaffe, JHEP [**06(2007)**]{}, 019 (2007), \[arXiv:hep-th/0702021\]. M. Unsal and L. G. Yaffe, Phys. Rev. D [**78**]{}, 065035 (2008) \[arXiv:0803.0344\]. B. Bringoltz and S. R. Sharpe, Phys. Rev. D[**78**]{}, 034507 (2008) \[arXiv:0805.2146\]. B. Bringoltz, JHEP [**06(2009)**]{}, 091 (2009) \[arXiv:0905.2406\]. B. Bringoltz and S. R. Sharpe, Phys. Rev. D [**80**]{}, 065031 (2009) \[arXiv:0906.3538\]. E. Poppitz and M. Unsal, JHEP [**01(2010)**]{}, 098 (2010) \[arXiv:0911.0358\]. P. C. Argyres and M. Unsal, JHEP [**08(2012)**]{}, 063 (2012) \[arXiv:1206.1890\]. D. Gross, R. Pisarski, L. Yaffe, Rev.Mod.Phys [**53**]{}, 43 (1981). N. Weiss, Phys.Rev.D [**24**]{}, 475 (1981). P. N. Meisinger, T. R. Miller, and M. C. Ogilvie, Phys. Rev. D [**65**]{}, 034009 (2002) \[arXiv:hep-ph/0108009\]. P. N. Meisinger and M. C. Ogilvie, Phys.Rev.D [**81**]{}, 025012 (2010) \[arXiv:0905.3577\]. A. Dumitru, et.al., Phys.Rev.D [**83**]{}, 034022 (2011) \[arXiv:1011.3820\]; A. Dumitru, et.al., Phys.Rev.D [**86**]{}, 105017 (2012) \[arXiv:1205.0137\]. F. Karsch and M. Lutgemeier, Nucl. Phys. B [**550**]{}, 449 (1999) \[arXiv:hep-lat/9812023\]. J. Engels, S. Holtmann and T. Schulze, Nucl. Phys. B [**724**]{}, 357 (2005) \[arXiv:hep-lat/0505008\]. A. Barducci, R. Casalbuoni, G. Pettini, R. Gatto, Phys.Lett.B [**301**]{}, 95 (1993). K. Kashiwa, Y. Sakai, H. Kouno, M. Matsuzaki, M. Yahiro, J.Phys.G [**36**]{}, 105001 (2009) \[arXiv:0804.3557\]. H. Kouno, Y. Sakai, T. Makiyama, K. Tokunaga, T. Sasaki, M. Yahiro, J.Phys.G [**39**]{} (2012) 085010-1 085010-21; Y. Sakai, H. Kouno, T. Sasaki and M. Yahiro, Phys.Lett.B [**718**]{} (2012) 130. H. Kouno, T. Misumi, K. Kashiwa, T. Makiyama, T Sasaki, M. Yahiro, \[arXiv:1304.3274\]. G. Cossu, H. Hatanaka, Y. Hosotani, E. Itou, J. Noaki, work in progress. [^1]: One may ask whether $N_{a}=2$ corresponds to the conformal window. However, since we introduce nonzero quark mass in our study, the conformality, even if exists, is broken.

--- abstract: 'The Zarankiewicz problem asks for an estimate on $z(m, n; s, t)$, the largest number of $1$’s in an $m \times n$ matrix with all entries $0$ or $1$ containing no $s \times t$ submatrix consisting entirely of $1$’s. We show that a classical upper bound for $z(m, n; s, t)$ due to Kővári, Sós and Turán is tight up to the constant for a broad range of parameters. The proof relies on a new quantitative variant of the random algebraic method.' address: 'Department of Mathematics, California Institute of Technology, Pasadena, CA 91125, USA' author: - David Conlon title: Some remarks on the Zarankiewicz problem --- Introduction ============ The classical Zarankiewicz problem [@Z51] asks for an estimate on $z(m, n; s, t)$, the maximum number of edges in a bipartite graph $G = (U, V; E)$ with $|U| = m$ and $|V| = n$ containing no copy of $K_{s,t}$. The order here is important, in that the set of size $s$ in $K_{s,t}$ must be embedded in $U$, while the set of size $t$ must be in $V$. Equivalently, $z(m, n; s, t)$ is the largest number of $1$’s in an $m \times n$ matrix with all entries $0$ or $1$ containing no $s \times t$ submatrix consisting entirely of $1$’s. A result of Kővári, Sós and Turán [@KST54] says that $$z(m, n; s, t) = O(m n^{1-1/s} + n),$$ where the implied constant depends only on $s$ and $t$ (a convention that we adopt throughout). If we swap the roles of $m$ and $n$ and of $s$ and $t$, we also obtain the bound $$z(m, n; s, t) = O(n m^{1-1/t} + m).$$ Assuming that $s \leq t$, the point where the second bound becomes better than the first is when $m$ is on the order of $n^{t/s}$. It turns out that this crossover point is critical to the problem, in that a lower bound for $z(n^{t/s}, n; s, t)$ which matches the upper bound up to a constant implies, by a simple sampling argument, a lower bound for $z(m, n; s, t)$ which is tight up to the constant for all $m$. These observations give rise to the following attractive question. \[qn:main\] Is it the case that for any fixed $s$ and $t$ with $2 \leq s \leq t$ and any $m \leq n^{t/s}$, $$z(m, n; s, t) = \Omega(m n^{1-1/s})?$$ In light of this question, our current state of knowledge about lower bounds for the Zarankiewicz problem seems rather weak. The classic result in the area is due to Kollár, Rónyai and Szabó [@KRSz], who showed[^1] that for any $s$ there exists $t$ such that $$z(n, n; s, t) = \Omega(n^{2-1/s}).$$ In their work, it suffices to take $t \geq s! + 1$, a bound that was subsequently improved by Alon, Rónyai and Szabó [@ARSz] to $t \geq (s-1)! + 1$. Unfortunately, this only furnishes a complete answer to Question \[qn:main\] in two cases, when $s = t = 2$ and when $s = t = 3$ (and both of these cases were well understood much earlier [@Bro66; @Erd38]). However, a result of Alon, Mellinger, Mubayi and Verstraëte [@AMMV12] shows that Question \[qn:main\] also has a positive answer for $s = 2$ and arbitrary $t$. \[thm:2case\] For any fixed $t \geq 2$ and any $m \leq n^{t/2}$, $$z(m, n; 2, t) = \Omega(m n^{1/2}).$$ In recent years, alternative proofs of the Kollár–Rónyai–Szabó theorem, though with weaker control on $t$, were found by Blagojević, Bukh and Karasev [@BBK] and by Bukh [@B15] using constructions where adjacency is determined by a randomly chosen algebraic variety. The first traces of this random algebraic method go back some way, to work of Matoušek [@Mat] in discrepancy theory, but it is the variant originating with Bukh [@B15], and developed further by the author [@Con], that has proved most useful. For instance, it has led to considerable progress [@BC; @CJL; @Jan; @JJM; @JMY; @JQ; @KKL18] on the celebrated rational exponents conjecture of Erdős and Simonovits [@E81], amongst other applications [@BG; @CT]. Our main result, a general lower bound for $z(m, n; s, t)$ valid over a broad range of $m$, is another application of the random algebraic method, though in a new, arguably simpler, form that returns quantitative estimates not at present available through the application of Bukh’s method. \[thm:main\] For any fixed $2 \leq s \leq t$ and any $m \leq n^{t^{1/(s-1)}/s(s-1)}$, $$z(m, n; s, t) = \Omega(m n^{1-1/s}).$$ This may be seen as partial progress on Question \[qn:main\], even if it leaves considerable room for improvement (except in the $s = 2$ case, where it agrees with Theorem \[thm:2case\]). On the other hand, we note that the only previous results dealing with lower bounds when $m = \omega(n)$ are a result of Matoušek [@Mat] and a result of Alon, Rónyai and Szabó [@ARSz], which subsumes that of Matoušek, saying that the conclusion of Theorem \[thm:main\] holds for any fixed $s \geq 2$ and $t \geq s! + 1$ and any $m \leq n^{1+1/s}$. Our result only begins to match the Alon–Rónyai–Szabó result when $t$ is roughly $s^{2s}$, but rapidly improves on it for larger $t$. Random polynomials and varieties ================================ Let $q$ be a prime power and let $\mathbb{F}_q$ be the finite field of order $q$. We will consider polynomials in $t$ variables over $\mathbb{F}_q$, writing any such polynomial as $f(X)$, where $X = (X_1, \dots, X_t)$. We let $\mathcal{P}_d$ be the set of polynomials in $X$ of degree at most $d$, that is, the set of linear combinations over $\mathbb{F}_q$ of monomials of the form $X_1^{a_1} \cdots X_t^{a_t}$ with $\sum_{i=1}^t a_i \leq d$. By a random polynomial, we just mean a polynomial chosen uniformly from the set $\mathcal{P}_d$. One may produce such a random polynomial by choosing the coefficients of the monomials above to be random elements of $\mathbb{F}_q$. The next lemma estimates the probability that a randomly chosen polynomial from $\mathcal{P}_d$ passes through each of $m$ distinct points. This is very similar to a result of Bukh [@B15], but there is a crucial difference, in that we are interested in whether our random polynomial passes through points in ${\overline{{\mathbb{F}}}}_q^t$, where ${\overline{{\mathbb{F}}}}_q$ is the algebraic closure of ${\mathbb{F}}_q$, and not just ${\mathbb{F}}_q^t$. \[prob\] Suppose that $q > \binom{m}{2}$ and $d \geq m - 1$. If $f$ is a random $t$-variate polynomial of degree $d$ over $\mathbb{F}_q$ and $x_1, \dots, x_m$ are $m$ distinct points in ${\overline{{\mathbb{F}}}}_{q}^t$, then $$\mathbb{P}[f(x_i) = 0 \mbox{ for all } i = 1, \dots, m] \leq 1/q^{m}.$$ Let $x_i = (x_{i,1}, \dots, x_{i,t})$ for each $i = 1, \dots, m$. We choose elements $a_2, \dots, a_t \in \mathbb{F}_q$ such that $x_{i,1} + \sum_{j=2}^t a_j x_{i, j}$ is distinct for all $i = 1, \dots, m$. To see that this is possible, note that there are exactly $\binom{m}{2}$ equations $$x_{i,1} + \sum_{j=2}^t a_j x_{i, j} = x_{i',1} + \sum_{j=2}^t a_j x_{i', j},$$ each with at most $q^{t-2}$ solutions $(a_2, \dots, a_t)$. Therefore, since the total number of choices for $(a_2, \dots, a_t)$ is $q^{t-1}$ and $q^{t-1} > q^{t-2} \binom{m}{2}$, we can make an appropriate choice. We now consider $\mathcal{P}'_d$, the set of polynomials of degree at most $d$ in $Z$, where $Z_1 = X_1 + \sum_{j=2}^t a_j X_j$ and $Z_j = X_j$ for all $2 \leq j \leq t$. Since this change of variables is an invertible linear map, $\mathcal{P}'_d$ is identical to $\mathcal{P}_d$. It will therefore suffice to show that a randomly chosen polynomial from $\mathcal{P}'_d$ passes through all of the points $z_1, \dots, z_m$ corresponding to $x_1, \dots, x_m$ with probability at most $q^{-m}$. For this, we will need the fact that, by our choice above, $z_{i,1} \neq z_{i',1}$ for any $1 \leq i < i' \leq m$. For any $f$ in $\mathcal{P}'_d$, we may write $f = g + h$, where $h$ contains all monomials of the form $Z_1^j$ for $j = 0, 1, \dots, m-1$ and $g$ contains all other monomials. For any fixed choice of $g$, there is exactly one choice of $h$ with coefficients in ${\overline{{\mathbb{F}}}}_{q}$ such that $f(z_i) = 0$ for all $i = 1, \dots, m$, namely, the unique polynomial of degree at most $m-1$ which takes the value $-g(z_i)$ at $z_{i,1}$ for all $i = 1, 2, \dots, m$, where uniqueness follows from the fact that the $z_{i,1}$ are distinct. Therefore, the number of choices for $h$ with coefficients in $\mathbb{F}_q$ is either $0$ or $1$. Since this is out of a total of $q^m$ possibilities, we see that the probability $f$ passes through all of the $z_i$ is at most $q^{-m}$, as required. Despite yielding quantitative results that were unavailable to earlier versions of the random algebraic method, our method relies on rather less input from algebraic geometry. Recall that a variety over an algebraically closed field ${\overline{{\mathbb{F}}}}$ is a set of the form $$W = \{x \in {\overline{{\mathbb{F}}}}^t : f_1(x) = \dots = f_s(x) = 0\}$$ for some collection of polynomials $f_1, \dots, f_s : {\overline{{\mathbb{F}}}}^t \rightarrow {\overline{{\mathbb{F}}}}$. The variety is irreducible if it cannot be written as the union of two proper subvarieties. The dimension $\dim W$ of $W$ is then the maximum integer $d$ such that there exists a chain of irreducible subvarieties of $W$ of the form $$\emptyset \subsetneq \{p\} \subsetneq W_1 \subsetneq W_2 \subsetneq \dots \subsetneq W_d \subset W,$$ where $p$ is a point. The following three standard lemmas about varieties will suffice for our purposes. \[infinite\] Every variety $W$ over an algebraically closed field ${\overline{{\mathbb{F}}}}$ with $\dim W \geq 1$ has infinitely many solutions. \[dimdrop\] Suppose that $W$ is an irreducible variety over an algebraically closed field ${\overline{{\mathbb{F}}}}$. Then, for any polynomial $g: {\overline{{\mathbb{F}}}}^t \rightarrow {\overline{{\mathbb{F}}}}$, $W \subseteq \{x : g(x) = 0\}$ or $W \cap \{x : g(x) = 0\}$ is a variety of dimension less than $\dim W$. \[bezout\] If, for a collection of polynomials $f_1, \dots, f_t : {\overline{{\mathbb{F}}}}^t \rightarrow {\overline{{\mathbb{F}}}}$, the variety $$W = \{x \in {\overline{{\mathbb{F}}}}^t : f_1(x) = \dots = f_t(x) = 0\}$$ has $\dim W = 0$, then $$|W| \leq \prod_{i=1}^t \deg(f_i).$$ Moreover, for a collection of polynomials $f_1, \dots, f_s : {\overline{{\mathbb{F}}}}^t \rightarrow {\overline{{\mathbb{F}}}}$, the variety $$W = \{x \in {\overline{{\mathbb{F}}}}^t : f_1(x) = \dots = f_s(x) = 0\}$$ has at most $\prod_{i=1}^s \deg(f_i)$ irreducible components. Proof of Theorem \[thm:main\] ============================= Fix $d = \lceil t^{1/(s-1)}\rceil - 1$ and $\ell = \lfloor \frac{1}{2d} q^{(d+1)/(s-1)} \rfloor$. Consider the bipartite graph between sets $U$ and $V$, where $V$ may be viewed as a copy of ${\mathbb{F}}_q^s$ for some prime power $q$ and $U$ has order $\ell$, each vertex $u_i$ of which is associated to an $(s-1)$-variate polynomial $f_i$ of degree at most $d$ with coefficients in ${\mathbb{F}}_q$. Each $u_i$ is then joined to the set of points $$S_i = \{(x_1, \dots, x_{s-1}, f_i(x_1, \dots, x_{s-1})) : x_1, \dots, x_{s-1} \in {\mathbb{F}}_q\}$$ in $V$. Note that, for any $1 \leq j \leq s$ and $1 \leq i_1 < \dots < i_j \leq k$, $$S_{i_1} \cap \cdots \cap S_{i_j} = \{(x_1, \dots, x_s) : x_s = f_{i_1}(x_1, \dots, x_{s-1}) = \dots = f_{i_j}(x_1, \dots, x_{s-1}) \}.$$ This intersection therefore has the same size as $T_{i_1, i_2} \cap \dots \cap T_{i_1, i_j}$, where $$T_{i, i'} = \{(x_1, \dots, x_{s-1}) : (f_i - f_{i'})(x_1, \dots, x_{s-1}) = 0\}.$$ Our aim now is to show that there is a choice of $f_i$ for $i = 1, \dots, \ell$ such that, for any $1 \leq j \leq s$ and $1 \leq i_1 < \dots < i_j \leq \ell$, the intersection $T_{i_1, i_2} \cap \dots \cap T_{i_1, i_j}$ has dimension at most $s-j$. To do this, we will pick the $f_i$ in sequence and show, by induction, that for every $1 \leq k \leq \ell$, there exist $f_1, \dots, f_k$ such that $T_{i_1, i_2} \cap \dots \cap T_{i_1, i_j}$ has dimension at most $s - j$ for any $1 \leq j \leq s$ and $1 \leq i_1 < \dots < i_j \leq k$. To begin the induction, we let $f_1$ be any $(s-1)$-variate polynomial of degree $d$. In this case, the condition that the intersection $T_{i_1, i_2} \cap \dots \cap T_{i_1, i_j}$ have dimension at most $s-j$ for all $1 \leq j \leq s$ and $1 \leq i_1 < \dots < i_j \leq k$ is degenerate, but can be meaningfully replaced by the observation that the set of all $(x_1, \dots, x_{s-1})$, corresponding to the trivial intersection, equals $\mathbb{F}_q^{s-1}$, which has dimension $s-1$, as required. Suppose now that $f_1, \dots, f_{k-1}$ have been chosen consistent with the induction hypothesis. We would like to pick $f_k$ so that for any $1 \leq j \leq s$ and $1 \leq i_1 < \dots < i_{j-1} < k$, the intersection $T_{i_1, i_2} \cap \dots \cap T_{i_1, i_{j-1}} \cap T_{i_1, k}$ has dimension at most $s-j$. For now, fix $1 \leq j \leq s$ and $1 \leq i_1 < \dots < i_{j-1} < k$ and note, by the induction hypothesis, that $T_{i_1, i_2} \cap \dots \cap T_{i_1, i_{j-1}}$ has dimension at most $s - j + 1$. Split the variety $T_{i_1, i_2} \cap \dots \cap T_{i_1, i_{j-1}}$ into irreducible components $W_1, \dots, W_r$ and suppose that $W_a$ is a component of dimension $s - j + 1 \geq 1$. By Lemma \[infinite\], $W_a$ has infinitely many points when considered as a variety over ${\overline{{\mathbb{F}}}}$. Fix $d+1$ points $w_1, \dots, w_{d+1}$ on $W_a$. For any $(s-1)$-variate polynomial $f$, write $$T_{i_1, f} = \{(x_1, \dots, x_{s-1}) : (f - f_{i_1})(x_1, \dots, x_{s-1}) = 0\}.$$ By Lemma \[dimdrop\], we see that if $\dim W_a \cap T_{i_1, f} = \dim W_a$, then $T_{i_1, f}({\overline{{\mathbb{F}}}})$ must contain all of $W_a$ and, in particular, each of $w_1, \dots, w_{d+1}$. Therefore, for a random $(s-1)$-variate polynomial $f$ of degree $d$, the probability that $W_a \cap T_{i_1, f}$ does not have dimension at most $s - j$ is at most the probability that the polynomial $f - f_{i_1}$ passes through all of $w_1, \dots, w_{d+1}$, which, by Lemma \[prob\], is at most $q^{-(d+1)}$. Since, by Lemma \[bezout\], the number of irreducible components of $T_{i_1, i_2} \cap \dots \cap T_{i_1, i_{j-1}}$ is at most $d^{s-1}$, this implies that the probability $T_{i_1, i_2} \cap \dots \cap T_{i_1, i_{j-1}} \cap T_{i_1,f}$ does not have dimension at most $s - j$ is at most $d^{s-1} q^{-(d+1)}$. By taking a union bound over the at most $\ell^{s-1}$ choices for $j$ and $i_1, \dots, i_{j-1}$, we see that the probability there exists $1 \leq j \leq s$ and $1 \leq i_1 < \dots < i_{j-1} < k$ such that $T_{i_1, i_2} \cap \dots \cap T_{i_1, i_{j-1}} \cap T_{i_1,f}$ does not have dimension at most $s- j$ is at most $\ell^{s-1} d^{s-1} q^{-(d+1)} < 1$ for $q$ sufficiently large. Therefore, there exists an $(s-1)$-variate polynomial $f$ of degree at most $d$ such that $T_{i_1, i_2} \cap \dots \cap T_{i_1, i_{j-1}} \cap T_{i_1,f}$ has dimension at most $s - j$ for any $1 \leq j \leq s$ and $1 \leq i_1 < \dots < i_{j-1} < k$, so taking $f_k = f$ completes the induction. To conclude the proof of Theorem \[thm:main\], we note that for any $1 \leq i_1 < \dots < i_s \leq \ell$ the intersection $T_{i_1, i_2} \cap \dots \cap T_{i_1, i_{s}}$ has dimension zero, so, by Bézout’s theorem, Lemma \[bezout\], the number of points in the intersection is at most $d^{s-1} < t$. That is, for any $1 \leq i_1 < \dots < i_s \leq \ell$, the intersection $S_{i_1} \cap \dots \cap S_{i_s}$ has at most $t-1$ points, so there is no copy of $K_{s,t}$ with $s$ vertices in $U$ and $t$ vertices in $V$. Since $|U| = \ell = \Omega(q^{(d+1)/(s-1)})$, $|V| = q^s$ and $|E| = \ell q^{s-1}$, we therefore have, for $m_0 \coloneqq m_0(n) = n^{(d+1)/s(s-1)} \geq n^{t^{1/(s-1)}/s(s-1)}$, that $$z(m_0, n; s, t) = \Omega(m_0 n^{1 - 1/s})$$ when $n$ is of the form $q^s$ with $q$ a prime power. By interpolating between prime powers, we can easily extend this result to all $n$. Finally, for any $m \leq m_0$, we can verify that $z(m,n;s, t) = \Omega(m n^{1-1/s})$ by choosing a random subset $U'$ of $U$ of order $m$ and noting that the expected number of edges between $U'$ and $V$ is $\Omega(m n^{1 - 1/s})$. Therefore, there must exist some choice for $U'$ such that the number of edges between $U'$ and $V$ is $\Omega(m n^{1 - 1/s})$, demonstrating the required lower bound. [**Remark.**]{} It is worth remarking that the method becomes a little simpler if we instead consider $$S_i = \{x \in {\mathbb{F}}_q^s : f_i(x) = 0\}$$ for random $s$-variate polynomials $f_i$ of degree at most $d$ with coefficients in ${\mathbb{F}}_q$. There is, however, a small tradeoff in the bound, in that it only allows us to determine a tight bound for $z(m, n; s,t)$ for $m \leq n^{t^{1/s}/s(s-1)}$ rather than $m \leq n^{t^{1/(s-1)}/s(s-1)}$. The difference is quite similar to the difference between the norm graphs of [@KRSz] and the projective norm graphs of [@ARSz]. Concluding remarks ================== The result of Alon, Mellinger, Mubayi and Verstraëte, Theorem \[thm:2case\], implies that $z(n, n^{t/2}; t, t) \geq z(n, n^{t/2}; t, 2) = \Omega(n^{(t+1)/2})$. By replacing $n$ with $n^{2/t}$, we see that $z(n^{2/t}, n; t, t) = \Omega(n^{1 + 1/t})$, which agrees with the upper bound up to a constant. By the usual sampling argument, this yields the following corollary. \[cor:2case\] For any fixed $t \geq 2$ and any $m \leq n^{2/t}$, $$z(m, n; t, t) = \Omega(m n^{1 - 1/t}).$$ That is, there is an asymptotically tight estimate for the Zarankiewicz problem when $s = t$ and one side is much larger than the other (and, crucially, in the non-trivial range where $m = \omega(n^{1/t})$ and $z(m, n; t, t) = \omega(n)$). A tight lower bound for $z(m, n; s, t)$ for all $m \leq n^{t/s}$ would similarly provide a tight lower bound for $z(m, n; t, t)$ for all $m \leq n^{s/t}$. However, we believe this will be difficult, or perhaps even impossible, to achieve for $s \geq 3$. In particular, the first open case of Question \[qn:main\], where $s = 3$ and $t = 4$, seems hard. [**Acknowledgements.**]{} I would like to thank Cosmin Pohoata for helpful discussions. I am also grateful to Dhruv Mubayi for drawing my attention to his work with Alon, Mellinger and Verstraëte [@AMMV12]. N. Alon, K. E. Mellinger, D. Mubayi and J. Verstraëte, The de Bruijn–Erdős theorem for hypergraphs, [*Des. Codes Cryptogr.*]{} [**65**]{} (2012), 233–245. N. Alon, L. Rónyai and T. Szabó, Norm-graphs: variations and applications, [*J. Combin. Theory Ser. B*]{} [**76**]{} (1999), 280–290. P. V. M. Blagojević, B. Bukh and R. Karasev, Turán numbers for $K_{s,t}$-free graphs: topological obstructions and algebraic constructions, [*Israel J. Math.*]{} [**197**]{} (2013), 199–214. W. G. Brown, On graphs that do not contain a Thomsen graph, [*Canad. Math. Bull.*]{} [**9**]{} (1966), 281–285. B. Bukh, Random algebraic construction of extremal graphs, [*Bull. Lond. Math. Soc.*]{} [**47**]{} (2015), 939–945. B. Bukh and D. Conlon, Rational exponents in extremal graph theory, [*J. Eur. Math. Soc.*]{} [**20**]{} (2018), 1747–1757. B. Bukh and X. Goaoc, Shatter functions with polynomial growth rates, [*SIAM J. Discrete Math.*]{} [**33**]{} (2019), 784–794. D. Conlon, Graphs with few paths of prescribed length between any two vertices, [*Bull. Lond. Math. Soc.*]{} [**51**]{} (2019), 1015–1021. D. Conlon, O. Janzer and J. Lee, More on the extremal number of subdivisions, to appear in [*Combinatorica*]{}. D. Conlon and M. Tyomkyn, Repeated patterns in proper colourings, preprint available at arXiv:2002.00921 \[math.CO\]. P. Erdős, On sequences of integers no one of which divides the product of two others and on some related problems, [*Mitt. Forsch.-Inst. Math. Mech. Univ. Tomsk*]{} [**2**]{} (1938), 74–82. P. Erdős, On the combinatorial problems which I would most like to see solved, [*Combinatorica*]{} [**1**]{} (1981), 25–42. O. Janzer, The extremal number of the subdivisions of the complete bipartite graph, [*SIAM J. Discrete Math.*]{} [**34**]{} (2020), 241–250. T. Jiang, Z. Jiang and J. Ma, Negligible obstructions and Turán exponents, preprint available at arXiv:2007.02975 \[math.CO\]. T. Jiang, J. Ma and L. Yepremyan, On Turán exponents of bipartite graphs, preprint available at arXiv: 1806.02838 \[math.CO\]. T. Jiang and Y. Qiu, Many Turán exponents via subdivisions, preprint available at arXiv:1908.02385 \[math.CO\]. D. Y. Kang, J. Kim and H. Liu, On the rational Turán exponents conjecture, preprint available at arXiv:1811.06916 \[math.CO\]. J. Kollár, L. Rónyai and T. Szabó, Norm-graphs and bipartite Turán numbers, [*Combinatorica*]{} [**16**]{} (1996), 399–406. T. Kővári, V. T. Sós and P. Turán, On a problem of K. Zarankiewicz, [*Colloq. Math.*]{} [**3**]{} (1954), 50–57. J. Matoušek, On discrepancy bounds via dual shatter function, [*Mathematika*]{} [**44**]{} (1997), 42–49. K. Zarankiewicz, Problem 101, [*Colloq. Math.*]{} [**2**]{} (1951), 301. [^1]: Their result actually gives something considerably stronger, namely, a graph with $n$ vertices and $\Omega(n^{2-1/s})$ edges containing no copy of $K_{s,t}$ with no regard for how it is oriented in the graph. However, the corollary stated here is all that is relevant to our discussion.

--- abstract: 'The Dicke effect is examined in an all-optical system of an optical waveguide side coupled to two interacting ring resonators in a liquid crytal environment. The system is shown to exhibit all the signatures of the Dicke effect under an active and reversible control by an applied voltage.' author: - 'A. E. Çetin' - 'Ö. E. Müstecapl[i]{}oğlu' bibliography: - 'references.bib' title: 'Electrically Tunable Dicke Effect in Double-Ring Resonator' --- Substantial narrowing of the spectral line shapes due to collisions of radiating and non-radiating atoms is first described by R. H. Dicke in 1953 and called as the Dicke Effect [@Dicke53]. The overall line shape consists of a superposition of a broad and a narrow line shapes centered at the transition frequency. Such a splitting of atomic decay into a pair of fast and slow decay channels is closely related to the superradiance phenomenon predicted shortly after the Dicke effect [@Dicke54]. Superradiance is cooperative spontaneous emission of radiation from an initially excited coherent ensemble of atoms. The slow and fast decay channels are respectively named as subradiant and superradiant decays of the system. Collective symmetric or anti-symmetric states of an ensemble of atoms are respectively in superradiant or subradiant phases. In addition to the atomic ensembles, The Dicke effect has been extensively studied in other systems, such as photonic crystals [@photonic_xtal_dicke], plasmonic lattices [@plasmon_dicke], electronic mesoscopic systems, [@brandes; @ApelPacheco08; @OrellanaDiez06; @OrellanaClaro04; @PodivilovShapiro92; @VorrathBrandes03; @brandes99; @shahbazyan] and in electron waveguides [@LeeReichl08; @LeeReichl09]. Not all the signatures of the Dicke effect can be found in many of these systems. Furthermore, control of the Dicke effect is too challenging. Our purpose is to examine the Dicke effect in an all-optical device with active and reversible control. We find that a pair of microring resonators coupled to a waveguide can exhibit all the key signatures of the Dicke effect in a controllable way with the help of a nematic liquid crystal (NLC). Tunable lifetimes of quasibound states in the resonators can be translated into reversible active control of optical signals in multiple ring resonator configurations. Such systems are used for many optical communication and signal processing applications such as all-optical logic gates [@ring_logic] and all-optical memory elements [@ring_memory]. From a fundamental perspective, the system allows for controlled investigations of quantum interference and decoherence by providing an all-optical analog of an Anderson-Fano model which is the prototype description of interaction between (quasi) bound and (quasi) continuum states [@anderson; @fano]. In addition, extending the system to multiple ring configuration, quantum phase transitions in the context of superradiance can be systematically examined and probed (For a review see [@brandes] and references therein). It has been shown that NLC can be used to tune the resonances of a single ring resonator coupled to a waveguide [@MauneDalton03]. To control linewidths, we introduce another ring to that system as depicted in Fig. \[figure1\], where two identical microring resonators at a distance $d$ apart are side coupled to a waveguide. A TE polarized Gaussian beam (electric field is perpendicular to the plane of the resonators) is sent from the input port of the waveguide. Two electrodes provide a voltage to change orientation of the NLC molecules used for cladding the resonators. Similar set up but without NLC has been examined for its reflective properties [@uzunoglu; @dagli]. NLC allows for controllable coupling coefficients between the resonators and the waveguide. ![The dimension of the system consisting of two ring-resonators on top of a waveguide feeded by two electrodes.[]{data-label="figure1"}](figure1.eps){width="8.3cm"} For a single ring, the resonance wavelength is determined by the Fabry-Perot etalon resonance condition, $m\lambda_{m}=2\pi Rn_{eff}$ where $m=1,2,3...$, $\lambda_{m}$ is the wavelength of the mth resonator mode, $R$ is the radius of the microring and $n_{eff}$ is the effective refractive index for the waveguide mode [@MauneDalton03; @SalehTeich91]. Proximity coupling by the evanescent tails makes $n_{eff}$ depending on the refractive index of the NLC cladding, $n_{clad}$ which is determined by $$\frac{1}{n_{clad}^{2}}=\frac{\cos^{2}(\theta)} {n_{e}^{2}}+\frac{\sin^{2}(\theta)}{n_{o}^{2}}, \label{eq1}$$ where $n_{e}=1.744$ and $n_{o}=1.517$ at $589$ nm [@MauneDalton03], and $\theta$ is the angle of the NLC directors (a local pseudo vector in the mean molecular long axis direction) relative to the radial axis from the origin in the middle of the electrodes [@MauneDalton03; @Wiley98]. When there is no applied field, assuming the optical field is too weak to induce any reorientation (optical Frederick’s effect) of the NLC directors, the NLC is in the isotropic phase, so that $n_{clad}=1.596$. When sufficiently strong potential is applied from the electrodes, the directors are deformed as shown in Fig. \[figure2\]. $\theta$ is locally determined by the Euler-Lagrange equations, to minimize the free energy density of the NLC with given elastic properties. For a potential field much stronger than the elastic contribution, the directors will be fully polarized in the applied electric field direction [@Collings02]. This permits local modulation of the $n_{eff}$, analogous to electrooptic effect. The electric field lines in Fig. \[figure2\], which is numerically determined by the Poisson equation, indicate the director alignments that can be controlled by the boundary conditions at the electrodes [@MauneDalton03]. After $\theta$ distribution is found, spatially inhomogeneous refractive index of NLC cladding is calculated by Eq. \[eq1\], which is used to propagate the Gaussian beam in the waveguide and to evaluate its transmission. We have repeated this process for different potential values and different separations between the rings. ![(Color) Angular difference $[^{o}]$ between the directors of the NLC molecules and the radial axis. (Arrow) The direction of the electric field generated by the electrodes at 5 V.[]{data-label="figure2"}](figure2.eps){width="8.3cm"} Using transfer matrix formalism [@transfer_mat1; @transfer_mat2], reflection characteristics of coupled ring resonators have been examined [@uzunoglu; @dagli]. In order to examine the coupling between the electromagnetic modes as realistically as possible, and to treat NLC cladding correctly, we follow a computational approach based upon finite element methods [@numerics]. We verify that our numerical analysis yields the similar lineshape structures obtained by the transfer matrix method. We calculate the resonances by evaluating the intensity of the wave at the output port of the waveguide, $I_{out}$. The spatially inhomogeneous $n_{clad}$ is used in the coupled wave equations to solve for the resonator and the waveguide modes and the evanescent waves in the NLC. Different computational grids are used for every different geometry arising when the ring separation is varied. The resonances of the system for different $d$ values without an applied potential are shown in Fig. \[figure3\]. Characteristic splitting of the single ring resonance into four peaks is observed. Four peaks arise due to simultaneous presence of direct proximity coupling together with the additional bus waveguide mediated coupling between the rings [@dagli]. In other systems such as ballistic electron guides, the resonators are coupled only through the bus waveguide and two peak splitting occurs. The asymmetry of the lineshapes in Fig. \[figure3\] is also a characteristic signature of the quantum interference. The intereference is of Fano type, due to spectrally overlapping subradiant and superradiant decay channels. When the gap between two ring resonators is widen, the proximity coupling between the rings reduce, while the bus-ring coupling remains unchanged. The splitting between the symmetrically placed symmetric and antisymmetric modes about the isolated ring resonance decreases. The further splitting of these modes due to the ring-bus interaction is independent of $d$. On the other hand, their width ($\Gamma$, Full-Width Half-Maximum (FWHM)) varies with the gap between the rings. In addition to the splitting into subradiant and superradiant channels, we have found the oscillatory behavior of the linewidths with the distance between the resonators. This signature of the Dicke effect is due to the spatial interference of radiation from decaying quasibound states of the rings coupled to the waveguide at separated locations, inherent to the collective nature of the system. Dicke oscillations of symmetric and antisymmetric modes are translated into further split modes in our case. The effect can be understood analogous to the level repulsion between coherently coupled degenerate bound states [@LeeReichl09]. Together with the interaction caused splitting of the resonances, interacting (interfering) decay channels also split. Representing linewidths as imaginary parts of the resonances, the symmetric and antisymmetric modes lie $180^o$ out of phase in the complex plane. The interaction channel through the bus waveguide introduces $d$ dependent phase accumulation to the mode freely propagating in between the rings along the waveguide. This is translated as a sinusoidal coupling between the rings or a rotation operation in the complex energy plane. By increasing the gap between the rings, the resonances collapse back onto the isolated ring value in a spiraling fashion. During this spiral motion, their imaginary parts periodically become vanishing and finite and hence oscillates. ![The resonance of the system as a function of wavelength, for a single resonator (dashed line) and two resonators (solid line) for different $d$ values, (a) 25 nm, (b) 50 nm, (c) 75 nm and (d) 100 nm at 0 V. []{data-label="figure3"}](figure3.eps){width="8.3cm"} Choosing a resonance, the most energetic one, labeled with a blue box as in Fig. \[figure3\](a) at $d=25$ nm, we investigate the dependence of the resonance width and resonance wavelength on the spatial separation between the rings. Fig. \[figure4\](a) shows the variation in the width of the resonance as a function of the gap between two ring resonators. The coupling between the rings decreases with distance and the oscillations eventually disappear. The width saturates at the single ring value at long distances. In one-dimensional system of bi-ripple ballistic electron waveguide, sinusoidal periodic behavior of the resonance width is found as the coupling does not decay with the distance [@LeeReichl09]. Our situation is similar to traditional Dicke systems in higher dimensions, where the coupling decreases with the distance [@brandes99; @shahbazyan]. Approach to the single ring resonance (resonance tuning) as we increase $d$ by an amount $\Delta$ is shown in Fig. \[figure4\](b). As the coupling decreases with distance, so does the splitting of the collective modes, and thus, the most energetic mode becomes less and less different from the single ring resonance.  vs the gap between two ring resonators (b) resonance tuning vs additional gap distance to $d=25$ nm at 0 V.[]{data-label="figure4"}](figure4.eps){width="8.3cm"} Considering now another most energetic resonance found at a larger separation between the rings, labeled with a blue box as in Fig. \[figure3\](d) at $d=100$ nm, we examine possible active control of these signatures of the Dicke effect demonstrated in Fig. \[figure4\]. When the potential applied from the electrodes is turned on, the linewidth and the resonance are changed as shown in Fig. \[figure5\]. The linewidth decreases with increasing potential, which aligns the directors such that in the coupler zones, the cladding index increases to $n_e$. This locally reduces the index contrast with the silicon guides and enhances the field penetration, and thus the proximity coupling between the resonators increases. The chosen most energetic mode becomes further split from the single ring resonance with the increasing coupling coefficient. The applied voltage and the spatial separation between the rings have opposite effects on the coupling resonators.  at $d=100$ nm vs applied voltage from the electrodes.[]{data-label="figure5"}](figure5.eps){width="8.3cm"} Finally, we analyze the influence of the external potential on the periodicity of $\Gamma$ as a function $d$. Fig. \[figure6\](a) is in agreement with the earlier observations in Fig. \[figure5\](a) that the resonance width decreases with the applied voltage. The periodicity of $\Gamma$ when no voltage is applied (see Fig. \[figure3\](a)) is determined to be $\lambda_{D}=48.92$ nm. Change of the period from this value when the potential is applied is denoted by $\delta_{D}$. Fig. \[figure6\](b), shows that the periodicity of the $\Gamma$ increases steadily with the potential applied from the electrodes. Enhanced coupling of the rings increases the splitting of symmetric and antisymmetric modes. For the chosen most energetic resonance, this makes it closer to the waveguide mode. Thus the free propagation and the associated phase accumulation between the rings occurs at smaller frequency or at larger period. In the complex energy plane, that means the rate of rotation with $d$ is reduced. Summarizing, we have examined the resonances and their widths in an all-optical system of a pair of ring resonators side-coupled to an optical waveguide in a NLC environment. We have found that the system exhibits all the key signatures of the Dicke effect, splitting of the lifetimes into slow and fast decaying channels in an oscillatory manner with the separation of the resonators. The energies and the lifetimes of the quasibound states of the resonators can be controlled with the applied voltage, which allows for tunable longer range interactions between the resonators. Such a reversible active control can be used to examine coherent collective effects, decoherence, and superradiant phase transitions. Besides, fine tuning of finesse and full spectral range can be exploited for multi-ring systems, such as CROW [@crow] and SCISSOR [@scissor] configurations, as well as rings coupled to multiple waveguides. These systems are subject to many applications, particularly optical logic and memory, filtering, reflectivity and optical switching. The authors acknowledge useful comments by A. Serpenguzel, H. Altug and H. Tureci.  vs the gap between two ring resonators for different voltage values applied from the electrodes (b) Increase in the periodicity of the resonance width function for different voltage values.[]{data-label="figure6"}](figure6.eps){width="8.3cm"}

--- abstract: | Time profiles of gamma ray bursts (GRBs) are extremely diverse in their durations, morphologies, and complexity. Nevertheless, the average peak-aligned profile of all bursts detected by BATSE with sufficient data quality has a simple “stretched” exponential shape, $F \propto \exp[-(t/t_0)^{1/3}]$, where $t$ is the time measured from the time for the peak flux, $F_p$, of the event, and $t_0$ is a time constant. We study the behaviour of $t_0$ of both the post-peak and the pre-peak slopes of the average time profile as a function of the peak brightness range of the burst sample. We found that the post-peak slope shows time dilation when comparing bright and dim bursts, while the pre-peak slope hardly changes. Thus dimmer bursts have a different shape – they are more asymmetric. This shape-brightness correlation is observed at a 99.6% confidence level. Such a correlation has a natural explanation within the pulse avalanche model, which is briefly described. Complex events, consisting of many pulses are more symmetric and are intrinsically brighter. Bursts consisting of one or a few pulses are intrinsically weaker and more asymmetric. For such a correlation to be observable requires that the luminosity distance distribution of GRBs to be different from a power-law. Keywords: Gamma-ray bursts, Methods: Data analysis. author: - | [**Boris Stern$^{1,2}$, Roland Svensson$^2$ and Juri Poutanen$^{2,3}$**]{}\ $^1$Institute for Nuclear Research, Moscow\ $^2$Stockholm Observatory, Saltsjöbaden, Sweden\ $^3$Uppsala Observatory, Uppsala, Sweden title: '**AVERAGE PROPERTIES OF THE TIME BEHAVIOUR OF GAMMA-RAY BURSTS**' --- ł 218[\_[2-18]{}]{} 218[\_[[int]{},2-18]{}]{} = ==1=1=0pt =2=2=0pt =2=2=0pt INTRODUCTION ============ Stern (1996) found that the average peak-aligned time profile of GRBs (the procedure of peak-alignment was pioneered by Mitrofanov et al. 1994, 1995) in the BATSE-2 catalog has a simple “stretched” exponential shape,\ $<F/F_p>=\exp[-(t/t_0)^{1/3}]$, where $t$ is the time since the peak flux, $F_p$, of the event, and $t_0$ is a constant ranging from 0.3 s for strong bursts to $\sim$ 1 s for dim bursts. This dependence of $t_0$ on brightness can be interpreted as a cosmological time dilation (e.g., [@pac92; @pir92]). Such a simple average time profile is remarkable considering the diverse and chaotic behavior of the individual time profiles of GRBs. On the other hand, the simple shape of the average time profile gives an excellent opportunity to study effects such as time dilation. Here, we study the two slopes of the average time profile for a larger sample of GRBs and with a more accurate treatment of the background than was done in Stern (1996). Another advantage we now have is the access to the pulse avalanche model developed by Stern & Svensson (1996) which successfully describes many statistical properties of GRBs including the stretched exponential shape of the profile and, of particular importance for the present work, the rms variance of individual time profiles. This means that we can rely on this model when estimating the errors of stretched exponential fits, which in turn gives us reliable estimates of the significance levels of the observed effects. DATA PROCESSING =============== This work is based on data obtained from the publicly available BATSE database at Goddard Space Flight Center. Our sample includes bursts up to trigger number 3745. We used the 0.064 s and 1.024 s time resolution data from the Large Area Detectors (LAD). All time profiles were constructed with 64 ms time resolution together with pre and post-burst extensions of 1024 ms time resolution. All background fits were done with the 1024 ms data as they cover a wider time interval including the pre-trigger history. The time profiles were studied using count rates in all 4 LAD’s energy channels. The procedure of background fitting included: – A visual examination of all bursts including both 64ms and 1 s resolution data. – All doubtful peaks and count rate variations with $\chi^2$ exceeding that of Poisson noise were analyzed in order to see whether they came from the same direction as the main peak of the burst. This was accomplished by comparing the direction of the eight-component vector consisting of the $\chi^2$ from the eight LADs for the feature with the direction of the corresponding vector for the main peak. – Discarding all events with highly variable background as well as those where we were unable to confidently extract sufficiently long intervals of pre-peak and post-peak histories. – Possible use of widely separated fitting windows to avoid losses of weak GRB signals. When sorting bursts into brightness groups we used peak fluxes for 64 ms time resolution from the BATSE-3 catalog. The errors were calculated using the pulse avalanche model. What we need are the statistical errors of the stretched exponential fits to the data. To extract this error directly from the data is very difficult because of strong correlations along the time profile. Instead, we simulated many samples of $N$ events, determined the average time profile for each sample, and then made a stretched exponential fit to each simulated average profile: $F(t)=\beta \exp[-(t/t_0)^{1/3}]$, where $t_0$ and $\beta$ are fitting parameters. The same procedure was used as for real data. Finally, the rms errors of the fitting parameters was calculated. We have two slopes of the profile - the pre-peak (rising) slope and post-peak (decaying) slope. We fitted them simultaneously with different $t_0$, denoted here as $t_r$ and $t_d$, and with a common $\beta$. Fitting the simulated samples gave the following relative standard deviations: $\sigma (t_{r,d})/ t_{r,d} = 0.20 \sqrt{100/N}$ $\sigma (t_r + t_d) / (t_r + t_d) = 0.19 \sqrt{100/N}$ $\sigma (t_d/t_r) / (t_d/t_r) = 0.13 \sqrt{100/N}$ Note, that the relative accuracy for the sum of the two time constants is close to that for one time constant, while the accuracy of the asymmetry ratio, $t_d/t_r$, is markedly better. This is a consequence of a strong correlation between the two slopes – a circumstance that favors the measurement of shape vs. brightness correlations and that complicates the measurement of the time dilation effect. The errors are robust against variations of the parameters of the pulse avalanche model as long as the model gives approximately the correct stretched exponential average profile. The procedure for the data analysis will be described in greater detail in Stern et al. (1997). THE AVERAGE TIME PROFILE FOR DIFFERENT BRIGHTNESS GROUPS ======================================================== The average peak-aligned profiles for three brightness groups are shown in Figure 1. Both the rising and and the decaying profiles are well shaped stretched exponentials for the bright and the medium group, while both profiles are quite deformed for the weakest group. The rising (pre-peak) slope is steeper for all brightness groups, but the asymmetry is increasing when going from the brightest to the dimmest group. The results of our stretched exponential fits are summarized in Table 1 and in Figure 2. =8.0cm =8.0cm \# Peak flux $N$ $t_d$ $t_r$ $t_r$ + $t_d$ $t_d/t_r$ ---- ------------ ----- ---------------- --------------- ----------------- ---------------- 1 12.5 – 200 64 0.42$\pm$ 0.10 0.36$\pm$0.09 0.78 $\pm$ 0.18 1.17 $\pm$0.19 2 3 – 12.5 193 0.60$\pm$ 0.08 0.40$\pm$0.06 1.00 $\pm$ 0.13 1.50 $\pm$0.11 3 1.75 – 3 159 0.75$\pm$ 0.11 0.37$\pm$0.06 1.12 $\pm$ 0.16 2.02 $\pm$0.21 4 1 – 1.75 241 0.72$\pm$ 0.09 0.46$\pm$0.06 1.18 $\pm$ 0.14 1.56 $\pm$0.13 5 .7 – 1 139 0.85$\pm$ 0.14 0.48$\pm$0.08 1.33 $\pm$ 0.21 1.89 $\pm$0.18 6 0 – .7 116 0.80$\pm$ 0.14 0.37$\pm$0.07 1.17 $\pm$ 0.20 2.16 $\pm$0.26 7 7.5 – 200 111 0.47$\pm$ 0.08 0.36$\pm$0.07 0.83$\pm$ 0.14 1.30 $\pm$0.16 8 5 – 200 157 0.50$\pm$ 0.07 0.35$\pm$0.06 0.85 $\pm$ 0.12 1.56 $\pm$0.13 9 .75 – 2.5 463 0.78$\pm$ 0.07 0.45$\pm$0.06 1.22 $\pm$ 0.11 1.73 $\pm$0.10 Fits of all time profiles have a good $\chi^2$ except that for the weakest sample (\# 6). Its profile is apparently deformed. This deformation is partially associated with the low trigger efficiency of this sample and with effects of Poisson noise. Nevertheless these biases are insufficient to explain this deformation and a real deformation effect could be present in the shape of the weakest sample. An analysis of this effect is beyond the scope of this poster paper. The strongest effect in the remaining samples is the time dilation of the decaying (post-peak) slope of the time profile. Comparing $t_d$ for samples 1 and 9 we find the time dilation to be a factor 1.86$^{+0.82}_{-0.57}$, where the 90% confidence interval is given. Comparing samples 7 and 9 gives a factor 1.66$^{+0.68}_{-0.42}$. Varying the lower flux limit of the brightest sample from $ F_p$ = 5 ph cm$^{-2}$ s$^{-1}$ and upwards does not significantly change the magnitude of the time dilation effect. A careful estimate of the significance level of the time dilation using model simulations gives 0.985. The rising (pre-peak) slope is, however, surprisingly stable. The variations of $t_r$ do not exceed statistical errors. This leads to an increasing asymmetry ratio, $t_d/t_r$, for the weaker samples. The effect is significant comparing samples 1 and 9: $(t_d/t_r)_{\rm dim}/(t_d/t_r)_{\rm bright}$ = 1.48$^{+0.55}_{-0.32}$ ($90\%$ confidence interval). The probability of sampling such ratios if the asymmetry was constant is 4 $\cdot$ 10$^{-3}$. The time dilation of $t_r + t_d$ has a smaller, but still acceptable significance when comparing samples 7 and 9. The “dim/bright” ratio becomes 1.47$^{+0.63}_{-0.36}$. The probability for zero time dilation is 0.01. INTERPRETATION OF CORRELATIONS USING THE PULSE AVALANCHE MODEL ============================================================== The pulse avalanche model is based on the assumption that the well-shaped stretched exponential time profile results from a simple stochastic process responsible for the generation of time histories of GRBs. Then the diversity of GRB’s time histories arise as different random realizations of the same stochastic process at approximately the same parameters. The important requirement is a near-criticality of the process - then it provides a large variety of individual bursts behaviours. This idea was implemented as a near-critical chain reaction of events (of still unknown nature, it could, e.g., be reconnections of turbulent magnetic field), where each event is associated with one pulse of gamma-ray emission. Then, in a near critical regime, one spontaneous pulse can give rise to a long cascade of secondary pulses piling up into a complex chaotic event, or, depending on chance, no further pulses may result and we will instead see a simple single pulse event. With a proper scaling of time delays between pulses and a proper spectrum of spontaneous pulses (flicker noise), the model successfully reproduces the stretched exponential profile (see Figure 3) as well as the autocorrelation function. There is also qualitative agreement with the duration distribution, the ratio of the number of simple and complex bursts is reproduced, and the model even produces the visual impression of real bursts. For more details see Stern & Svensson (1996). Just two simple assumptions of those implemented in the pulse avalanche model are needed to demonstrate that all correlations described above are very natural. GRBs consist of a number of pulses of different durations but similar shapes. Let all these pulses have locally independent sources of energy. Then if two pulses coincide in time, their amplitudes sum up. In a complex event, hundreds of pulses are piling up, increasing the peak brightness by up to an order of magnitude. Then, if a pulse is a kind of standard candle, simple events are intrinsically weak and complex events are intrinsically bright. At the same time, simple events are asymmetric just because a single pulse is asymmetric (e.g., Norris et al. 1996) with a sharp rise and a slower decay. This asymmetry is washed out in complex events where the position of the highest peak is more or less random among many overlapping pulses. To demonstrate this we simulated a large sample of “bursts” using model parameters which gave an approximate agreement with the stretched exponential average time profile for the sample of all real bursts. The results are summarized in Table 2. We see that the asymmetry ratio, $t_d/t_r$, increases with decreasing peak flux. Peak flux $t_d$ $t_r$ $t_r$ + $t_d$ $t_d/t_r$ -------------- -- ------- ------- --------------- ----------- 0 – $\infty$ 0.66 0.41 1.07 1.61 3 – $\infty$ 0.75 0.68 1.43 1.10 0.8 – 3 0.72 0.50 1.22 1.44 0 – 0.8 0.50 0.19 0.69 2.63 : Time constants of the simulated bursts in different “intrinsic brightness” intervals. The amplitude of the single pulse is sampled uniformly from the interval $[0,1]$. =7.9cm CONCLUSIONS =========== Besides the time dilation effect, observed in many previous works (e.g., Norris et al. 1994) we also see a dependence of the profile asymmetry on brightness and this effect is of the same order of magnitude as the time dilation itself. A correlation of such a kind cannot be due to spectral redshifts. We find that strong events have slightly smaller asymmetry in the higher energy bands, (LAD channels 4 and 3) than in the lower energy bands (LAD channels 1 and 2). Redshifting the softer part of the spectrum below the detector threshold would then give rise to more symmetric, rather than asymmetric profiles. Details will be published elsewhere. The effect of trigger efficiency is negligible for the brightness groups considered here. Also one can hardly find an evolutionary factor that would change the asymmetry. One can suggest that there exist two separate classes of GRBs with different degree of asymmetry, which are differently distributed in space (separate classes of [*long*]{} bursts are required as short bursts do not contribute to the asymmetry). But such a suggestion seems too arbitrary, too radical, and unnecessary as there exists a much simpler explanation. The simplest explanation is that the observed correlation is a consequence of an intrinsic correlation between shape and brightness as described above. The necessary condition for such a correlation to be observable is a significant deviation from a power law for the GRB distribution over luminosity distance. This would allow intrinsically strong events to dominate in the brightest observational range. The observed log $N$ - log $P$ distribution is actually curved (Meegan et al. 1996) and this is natural if the distance distribution covers both Euclidean and $z \sim 1$ regions which have different luminosity distance scalings. Maybe our detection of a strong shape - brightness correlation imposes a stronger constraint on the curvature of the true radial distribution of GRBs than what follows from the observed log $N$ - log $P$ distribution. Detailed studies are, however, required to formulate this intuitive conclusion at a quantitative level. As one kind of correlation has been observed, other kinds of intrinsic correlations may also be observable and this causes a problem for the cosmological interpretation of the time dilation effect. This problem, considered by Brainerd (1994), arises from unavoidable correlations between peak luminosity and time scales caused by different bulk Lorentz factor in the sources of different bursts. This effect can mimic both cosmological time dilation and spectral redshift. However, if our interpretation using the pulse avalanche model is valid we must conclude that the real time dilation is larger than that obtained from Table 1. Actually, within the pulse avalanche framework, intrinsically weak events are not only more asymmetric, but they are also narrower (see column $t_r+t_d$ in Table 2). This is a new correction that increases the real time dilation more than the correction arising from spectral redshift (see Norris et al. 1994). The corrected time dilation could exceed a factor 2 and it could be caused by different effects, including the cosmological one. Unfortunately, the task of extracting the cosmological component from the total time dilation seems extremely difficult. ACKNOWLEDGMENTS {#acknowledgments .unnumbered} =============== We acknowledge support from the Swedish Royal Academy of Sciences, the Swedish Natural Science Research Council (in particular a postdoctoral fellowship for J.P.), and a NORDITA Nordic Project grant. This research used data obtained from the HEASARC Online Service provided by NASA/GSFC. Brainerd, J. J. 1994, ApJ, 428, L1 Meegan, C. A., et al. 1996, ApJS, 106, 65 Mitrofanov, I. G., Chernenko, A. M., Pozanenko, A. S., Paciesas, W. S., Kouveliotou, C., Meegan, C. A., Fishman, G. J., Sagdeev, R. Z. 1994, in G. J. Fishman, J. J. Brainerd, K. C. Hurley (eds), Proc. of the Second Huntsville workshop on Gamma Ray Bursts, AIP, New York, 187 Mitrofanov, I. G., et al. 1995, Astronomy Reports, 39, 305 Norris, J. P., Nemiroff, R. J., Scargle, J. D., Kouveliotou, C., Fishman, G. J., Meegan, C. A., Paciesas, W. S., Bonnell, J. T. 1994, ApJ, 424, 540 Norris, J. P., Nemiroff, R. J., Bonnell, J. T., Scargle, J. D., Kouveliotou, C., Paciesas, W. S., Meegan, C. A., Fishman, G. J. 1996, ApJ, 459, 393 Paczyński, B. 1992, Nature, 355, 521 Piran, T. 1992, ApJ, 389, L45 Stern, B. E. 1996, ApJ, 464, L111 Stern, B. E., Svensson, R. 1996, ApJ, 469, L109 Stern, B. E., Poutanen, J., Svensson, R. 1997, in preparation

--- author: - 'Jun Xu, Yuan Huang, Ming-Ming Cheng,  Li Liu, Fan Zhu, Zhou Xu, Ling Shao[^1] [^2]' bibliography: - 'NAC.bib' title: 'Noisy-As-Clean: Learning Self-supervised Denoising from the Corrupted Image' --- Introduction {#sec:introduction} ============ denoising is an ill-posed inverse problem to recover a *clean* image $\mathbf{x}$ from the *observed* noisy image $\mathbf{y}=\mathbf{x}+\mathbf{n}_{o}$, where $\mathbf{n}_{o}$ is the *observed* corrupted noise.  One popular assumption on $\mathbf{n}$ is the additive white Gaussian noise (AWGN) with standard deviation (std) $\sigma$, which serves as a perfect test bed for supervised networks in the deep learning era [@vggnet; @googlenet; @resnet].Supervised networks [@nlnet; @dncnn; @n3net] learn the image priors and noise statistics on plenty pairs of clean and corrupted images, and achieve promising denoising performance on the images with similar priors and noise statistics (e.g., AWGN). \[f-example\] With advances on AWGN noise removal [@mlp; @dncnn; @n3net], a natural question arises is how these denoising networks can exert their effect on real noisy photographs.Realistic noise is signal dependent and more complex than AWGN [@crosschannel2016; @dnd2017; @PolyUdataset].  Thus, previous supervised denoising networks unavoidably suffer from a *domain gap problem*: both the image priors and noise statistics in training are different from those of the real-world test images.  Recently, several unsupervised [@gcbd; @noise2noise; @dip; @noise2void] and self-supervised [@noise2self; @ss2019] networks have been developed to get rid of the dependence on clean images, which are difficult to be obtained in real-world scenarios.However, unsupervised networks are subjected to the gap on either image priors or noise statistics, while self-supervised suffer from the gap on noise statistics, between the external images for training and the corrupted ones for test.  Besides, several networks [@noise2noise; @dip] succeed on the zero-mean noise.But the realistic noise in real-world images is not necessarily zero-mean [@crosschannel2016; @dnd2017; @sidd2018]. To alleviate the domain gap on image priors and noise statistics between training and test images, in this paper, we propose a “Noisy-As-Clean” (NAC) strategy for training self-supervised denoising networks.  In our NAC, we directly train an image-specific network by taking the corrupted image $\mathbf{y}=\mathbf{x}+\mathbf{n}_{o}$ as the “*clean*” target.  Thus, the domain gap on image priors are largely bridged by our NAC.  To reduce the gap on noise statistics, for the target corrupted image $\mathbf{y}$, we take as the input of our NAC a *simulated* noisy image $\mathbf{z}=\mathbf{y}+\mathbf{n}_{s}$ consisting of the corrupted image $\mathbf{y}$ and a *simulated* noise $\mathbf{n}_{s}$, which is statistically close to the corrupted noise $\mathbf{n}_{o}$ in $\mathbf{y}$.By this way, our NAC network learns to clean up the *simulated* noise $\mathbf{n}_{s}$ from the doubly corrupted image $\mathbf{z}$ during training, and thus is able to remove the noise $\mathbf{n}_{o}$ from the corrupted image $\mathbf{y}$ during test. A simple but useful observation about our NAC strategy is: *as long as the corrupted noise is “weak”, it is feasible to train a self-supervised denoising network only with the corrupted test image, and the learned parameters are very close to those of a *supervised* network trained with a pair of the corrupted image and its clean version*. Though being very simple, our NAC strategy is very effective for image denoising. In Figure \[f-example\], we compare the denoised images by the vanilla DnCNN [@dncnn] and the DnCNN trained with our NAC (DnCNN+NAC), on the image “House” corrupted by AWGN ($\sigma=15$).We observe that the “DnCNN+NAC” achieves better visual quality and higher PSNR/SSIM results than DnCNN [@dncnn], which is trained on plenty of noisy and clean image pairs. Experiments on diverse benchmarks demonstrate that, when trained with our NAC strategy, the DnCNN [@dncnn] and ResNet [@resnet] in Deep Image Prior (DIP) [@dip] achieve comparable or better performance than supervised denoising networks on synthetic and real-world noisy images.  Our work reveals that, *when the noise is “weak”*, a self-supervised network trained directly on the corrupted image can obtain comparable or even better performance than supervised networks on image denoising. In summary, our contribution are mainly three-fold: - We propose a “Noisy-As-Clean” (NAC) strategy for training self-supervised denoising networks. - We provide a theoretical background of our NAC strategy, and implement the DnCNN [@dncnn] and ResNet in DIP [@dip] into self-supervised networks by our NAC for effective image denoising. - Experiments on synthetic and real-world benchmarks show that, on weak noise, the DnCNN and ResNet in [@dip] trained by our NAC achieve comparable or even better performance than the comparison denoising networks. The remaining parts of this paper are organized as follows. In §\[sec:related\], we introduce the related work. In §\[sec:nas\], we present the theoretical background of our NAC strategy for self-supervised image denoising.In §\[sec:self-sup\], we implement the DnCNN [@dncnn] and ResNet used in [@dip] as self-supervised networks by our NAC.Extensive experiments are conducted in §\[sec:exp\] demonstrate that, the DnCNN and ResNet networks trained by our NAC achieve comparable or even better performance than previous supervised image denoising networks on benchmark synthetic and real-world datasets.Conclusion is given in §\[sec:con\]. Related Work {#sec:related} ============ In Table \[t1\], we summarize several state-of-the-art supervised [@dncnn; @cbdnet], unsupervised [@noise2noise; @gcbd; @noise2void] and self-supervised [@dip; @noise2self; @ss2019] networks, image priors, and noise statistics.  In this work, to bridge the *domain gap problem*, we propose a “Noisy-As-Clean” strategy to learn the image-specific internal prior and noise statistics directly from the corrupted test image.  **Supervised denoising networks** are trained with plenty pairs of noisy and clean images.This category of networks can learn external image priors and noise statistics from the training data.  Several methods [@dncnn; @n3net; @nlrn2018] have been developed with achieving promising performance on AWGN noise removal, where the statistics of training and test noise are similar.  However, due to the aforementioned *domain gap problem*, the performance of these networks degrade severely on real-world noisy images [@crosschannel2016; @dnd2017; @PolyUdataset]. **Unsupervised and self-supervised denoising networks** are developed to remove the need on plenty of clean images. Along this direction, Noise2Noise (N2N) [@noise2noise] trains the network between pairs of corrupted images with the same scene, but independently sampled noise. This work is feasible to learn external image priors and noise statistics from the training data. However, in real-world scenarios, it is difficult to collect large amounts of paired images with independent corruption for training. Noise2Void (N2V) [@noise2void] predicts a pixel from its surroundings by learning blind-spot networks, but it still suffers from the domain gap on image priors between the training images and test images. This work assumes that the corruption is zero-mean and independent between pixels. However, as mentioned in Noise2Self (N2S) [@noise2self], N2V [@noise2void] significantly degrades the training efficiency and denoising performance at test time. Recently, Deep Image Prior (DIP) [@dip] reveals that the network structure can resonate with the natural image priors, and can be utilized in image restoration without external images. However, it is not practical to select a suitable network and early-stop its training at right moments for each corrupted image. Self-supervised denoisers [@noise2self; @ss2019] employ explicit corruption models, and train the networks only with the corrupted image itself.  In this work, we utilize the helpful noise model to learn self-supervised denoising networks for real-world image denoising. -- ---------------------------- ------------ ------- ------- Image Noise Stat. DnCNN [@dncnn] 17’TIP Ext. CBDNet [@cbdnet] 19’CVPR Ext. Noise2Noise [@noise2noise] 18’ICML Ext. GAN-CNN [@gcbd] 18’CVPR Ext. Noise2Void [@noise2void] 19’CVPR Ext. Deep Image Prior [@dip] 18’CVPR Int. Noise2Self   [@noise2self] 19’ICML Ext. Self-Supervised [@ss2019] 19’NeurIPS Ext. Noisy-As-Clean (Ours) 20’Submit Int. -- ---------------------------- ------------ ------- ------- : **Summary of representative networks for image denoising**. **S.**: Supervised networks. **U.**: Unsupervised networks. **SS.**: Self-supervised networks. **Pub.**: Publication. **Int.**: Internal image priors. **Ext.**: External image priors. **Stat.**: Statistics. The networks with “” are able to learn the noise statistics from training data. \[t1\] \[f1\] **Internal and external image priors** are widely used for diverse image restoration tasks [@mcwnnm; @foe; @pgpd]. Internal priors are directly learned from the input test image itself, such as the multi-scale priors [@ksvd; @STAR2020; @cvid2020], image-specific details [@iraniinternal; @Liang_2018_CVPR], and non-local self similarity [@mcwnnm; @twsc; @NLH2020].  The external ones are learned on external natural images [@epll; @pgpd; @gid2018].  Internal priors are adaptive to its image contents, but somewhat affected by the corruptions [@ksvd; @iraniinternal].  By contrast, the external priors are effective for restoring images with general contents, but may not be optimal for specific test image [@epll; @pgpd; @chen2017trainable]. **Noise statistics** is of key importance for image denoising. The AWGN noise is one typical noise with widespread study. Recently, researchers shift more attention to the realistic noise produced in camera sensors [@dnd2017; @sidd2018], which is usually modeled as mixed Poisson and Gaussian distribution [@poissongaussian]. The Poisson component mainly comes from the irregular photons hitting the sensor [@liu2006noise], while Gaussian noise is majorly produced by dark current [@crosschannel2016]. Though performing well on the synthetic noise being trained with, supervised denoisers [@dncnn; @nlrn2018; @cbdnet] still suffer from the *domain gap problem* when processing the real-world noisy images. Theoretical Background of “Noisy-As-Clean” {#sec:nas} ========================================== Training a supervised network $f_{\theta}$ (parameterized by $\theta$) requires many pairs $\{(\mathbf{y}_{i},\mathbf{x}_{i})\}$ of noisy image $\mathbf{y}_{i}$ and clean image $\mathbf{x}_{i}$, by minimizing an empirical loss function $\mathcal{L}$ as $$\arg\min_{\theta}\sum_{i=1}\mathcal{L}(f_{\theta}(\mathbf{y}_{i}),\mathbf{x}_{i}).$$ Assuming that the probability of occurrence for pair $(\mathbf{y}_{i},\mathbf{x}_{i})$ is $p(\mathbf{y}_{i},\mathbf{x}_{i})$, then statistically we have $$\label{eqn:e2} \begin{split} \theta^{*} &=\arg\min_{\theta}\sum_{i=1}p(\mathbf{y}_{i},\mathbf{x}_{i})\mathcal{L}(f_{\theta}(\mathbf{y}_{i}),\mathbf{x}_{i}) \\ &= \arg\min_{\theta}\mathbb{E}_{(\mathbf{y},\mathbf{x})}[\mathcal{L}(f_{\theta}(\mathbf{y}),\mathbf{x})], \end{split}$$ where $\mathbf{y}$ and $\mathbf{x}$ are random variables of noisy and clean images, respectively.  The paired variables $(\mathbf{y},\mathbf{x})$ are dependent, and their relationship is $\mathbf{y}=\mathbf{x}+\mathbf{n}_{o}$, where $\mathbf{n}_{o}$ is the random variable of *observed* noise.  By exploring the dependence of $p(\mathbf{y}_{i},\mathbf{x}_{i})=p(\mathbf{x}_{i})p(\mathbf{y}_{i}|\mathbf{x}_{i})$, Eqn. (2) is equivalent to $$\label{e3} \begin{split} \theta^{*} &=\arg\min_{\theta}\sum_{i=1}p(\mathbf{x}_{i})p(\mathbf{y}_{i}|\mathbf{x}_{i})\mathcal{L}(f_{\theta}(\mathbf{y}_{i}),\mathbf{x}_{i}) \\ &= \arg\min_{\theta}\mathbb{E}_{\mathbf{x}}[\mathbb{E}_{\mathbf{y}|\mathbf{x}}[\mathcal{L}(f_{\theta}(\mathbf{y}),\mathbf{x})]] . \end{split}$$ This indicates that the network $f_{\theta}$ can minimize the loss function by solving Eqn. (3) separately for each clean image.  Different with the “zero-mean” assumption in [@noise2noise; @noise2void], here we study a more practical assumption on noise statistics, i.e., *the expectation $\mathbb{E}[\mathbf{x}]$ and variance* $\text{Var}[\mathbf{x}]$ *of signal intensity are much stronger than those of noise $\mathbb{E}[\mathbf{n}_{o}]$ and* $\text{Var}[\mathbf{n}_{o}]$ (negligible but not necessarily zero): $$\label{e4} \mathbb{E}[\mathbf{x}] \gg \mathbb{E}[\mathbf{n}_{o}] , \ \text{Var}[\mathbf{x}] \gg \text{Var}[\mathbf{n}_{o}] .$$ This is actually valid in real-world scenarios, since we can clearly observe the contents in most real photographs, *with little influence of the noise*.  The noise therein is often modeled by zero-mean Gaussian or mixed Poisson and Gaussian (for realistic noise).  Hence, the noisy image $\mathbf{y}$ should have similar expectation with the clean image $\mathbf{x}$: $$\label{e5} \mathbb{E}[\mathbf{y}] = \mathbb{E}[\mathbf{x}+\mathbf{n}_{o}] = \mathbb{E}[\mathbf{x}]+\mathbb{E}[\mathbf{n}_{o}] \approx \mathbb{E}[\mathbf{x}].$$ Now we add *simulated* noise $\mathbf{n}_{s}$ to the *observed* noisy image $\mathbf{y}$, and generate a new noisy image $\mathbf{z}=\mathbf{y}+\mathbf{n}_{s}$.  We assume that $\mathbf{n}_{s}$ is statisticly close to $\mathbf{n}_{o}$, i.e., $\mathbb{E}[\mathbf{n}_{s}]\approx\mathbb{E}[\mathbf{n}_{o}]$ and $\text{Var}[\mathbf{n}_{s}]\approx\text{Var}[\mathbf{n}_{o}]$.  Then we have $$\label{e6} \mathbb{E}[\mathbf{z}]\gg\mathbb{E}[\mathbf{n}_{s}], \ \text{Var}[\mathbf{z}]\gg\text{Var}[\mathbf{n}_{s}] .$$ Therefore, the *simulated* noisy image $\mathbf{z}$ has similar expectation with the *observed* noisy image $\mathbf{y}$: $$\label{e7} \mathbb{E}[\mathbf{z}] = \mathbb{E}[\mathbf{y}+\mathbf{n}_{s}] \approx \mathbb{E}[\mathbf{y}] .$$ By the *Law of Total Expectation* [@billingsley1995probability], we have $$\label{e8} \mathbb{E}_{\mathbf{y}}[\mathbb{E}_{\mathbf{z}}[\mathbf{z}|\mathbf{y}]] = \mathbb{E}[\mathbf{z}] \approx \mathbb{E}[\mathbf{y}] = \mathbb{E}_{\mathbf{x}}[\mathbb{E}_{\mathbf{y}}[\mathbf{y}|\mathbf{x}]] .$$ Since the loss function $\mathcal{L}$ (usually $\ell_{2}$) and the conditional probability density functions $p(\mathbf{y}|\mathbf{x})$ and $p(\mathbf{z}|\mathbf{y})$ are all *continuous everywhere*, the optimal network parameters $\theta^{*}$ of Eqn. (3) changes little with the addition of negligible noise $\mathbf{n}_{o}$ or $\mathbf{n}_{s}$. With Eqns. (4)-(8), when the $\mathbf{x}$-conditioned expectation of $\mathbb{E}_{\mathbf{y}|\mathbf{x}}[\mathcal{L}(f_{\theta}(\mathbf{y}),\mathbf{x})]$ are replaced with the $\mathbf{y}$-conditioned expectation of $\mathbb{E}_{\mathbf{z}|\mathbf{y}}[\mathcal{L}(f_{\theta}(\mathbf{z}),\mathbf{y})]$, $f_{\theta}$ obtains similar $\mathbf{y}$-conditioned optimal parameters $\theta^{*}$: $$\begin{split} \label{e9} &\arg\min_{\theta}\mathbb{E}_{\mathbf{y}}[\mathbb{E}_{\mathbf{z}|\mathbf{y}}[\mathcal{L}(f_{\theta}(\mathbf{z}),\mathbf{y})]] \\ \approx &\arg\min_{\theta}\mathbb{E}_{\mathbf{x}}[\mathbb{E}_{\mathbf{y}|\mathbf{x}}[\mathcal{L}(f_{\theta}(\mathbf{y}),\mathbf{x})]] = \theta^{*}. \end{split}$$ The network $f_{\theta}$ minimizes the loss function $\mathcal{L}$ for each input image pair separately, which equals to minimize it on all finite pairs of images. Through simple manipulations, Eqn. (9) is equivalent to $$\label{e10} \begin{split} &\arg\min_{\theta}\sum_{i=1}p(\mathbf{y}_{i})p(\mathbf{z}_{i}|\mathbf{y}_{i})\mathcal{L}(f_{\theta}(\mathbf{z}_{i}),\mathbf{y}_{i}) \\ =& \arg\min_{\theta}\mathbb{E}_{\mathbf{y}}[\mathbb{E}_{\mathbf{z}|\mathbf{y}}[\mathcal{L}(f_{\theta}(\mathbf{z}),\mathbf{y})]] \approx \theta^{*} . \end{split}$$ By exploring the dependence of $p(\mathbf{z}_{i},\mathbf{y}_{i})=p(\mathbf{y}_{i})p(\mathbf{z}_{i}|\mathbf{y}_{i})$, Eqn. (10) is equivalent to $$\begin{split} \label{e11} &\arg\min_{\theta}\mathbb{E}_{(\mathbf{z},\mathbf{y})}[\mathcal{L}(f_{\theta}(\mathbf{z}),\mathbf{y})] \\ =& \arg\min_{\theta}\sum_{i=1}p(\mathbf{z}_{i},\mathbf{y}_{i})\mathcal{L}(f_{\theta}(\mathbf{z}_{i}),\mathbf{y}_{i}) \approx \theta^{*} . \end{split}$$ **A simple but useful observation is**: *as long as the noise is weak, the optimal parameters of self-supervised network trained on noisy image pairs $\{(\mathbf{z}_{i},\mathbf{y}_{i})\}$ are very close to the optimal parameters of the supervised networks trained on pairs of noisy and clean images $\{(\mathbf{y}_{i},\mathbf{x}_{i})\}$*.In Figure \[fig:example\], we explain our NAC strategy and illustrate this observation through an example on the image “Test004” from the BSD68 dataset: The clean image $\mathbf{x}$ in (a) is firstly corrupted by observed AWGN noise $\mathbf{n}_{o}$ with $\sigma=5$.  Then we add simulated AWGN noise $\mathbf{n}_{s}$ also with $\sigma=5$ to the corrupted image $\mathbf{x}+\mathbf{n}_{o}$ in (b), and obtain a doubly corrupted image $\mathbf{x}+\mathbf{n}_{o}+\mathbf{n}_{s}$ in (c). The DnCNN [@dncnn] with our NAC strategy, named as DnCNN+NAC, is trained with the doubly corrupted image $\mathbf{x}+\mathbf{n}_{o}+\mathbf{n}_{s}$ in (c) as input and the corrupted image $\mathbf{x}+\mathbf{n}_{o}$ in (b) as target.  The final training output is plotted in (d), with similar PSNR and SSIM [@ssim] results as the corrupted image $\mathbf{x}+\mathbf{n}_{o}$ in (b).  Then the DnCNN+NAC network learned on (b) and (c) is directly employed to perform inference on the corrupted image $\mathbf{x}+\mathbf{n}_{o}$ in (b), and produces the testing output in (e).  When compared to DnCNN [@dncnn], our DnCNN+NAC achieves much higher PSNR and SSIM results on the corrupted image (b). The estimated simulated noise $\mathbf{n}_{s}$ and observed noise $\mathbf{n}_{o}$ in training and test stages are plotted in (g) and (h), respectively. One can see that they are visually in similar noise statistics. **Consistency of noise statistics**.  Since our contexts are the real-world scenarios, the noise can be modeled by mixed Poisson and Gaussian distribution [@poissongaussian]. Fortunately, both the two distributions are linear additive, i.e., the addition variable of two Poisson (or Gaussian) distributed variables are still Poisson (or Gaussian) distributed. Assume that the observed (simulated) noise $\mathbf{n}_{o}$ ($\mathbf{n}_{s}$) follows a mixed $\mathbf{x}$-dependent ($\mathbf{y}$-dependent) Poisson distribution parameterized by $\lambda_{o}$ ($\lambda_{s}$) and Gaussian distribution $\mathcal{N}(\bm{0}, \sigma_{o}^{2})$ ($\mathcal{N}(\bm{0}, \sigma_{s}^{2})$), i.e., $$\begin{split} \mathbf{n}_{o} &\sim \mathbf{x}\odot \mathcal{P}(\lambda_{o})+\mathcal{N}(\bm{0}, \sigma_{o}^{2}) , \\ \mathbf{n}_{s} &\sim \mathbf{y}\odot \mathcal{P}(\lambda_{s})+\mathcal{N}(\bm{0}, \sigma_{s}^{2}) \\ &\approx \mathbf{x}\odot \mathcal{P}(\lambda_{s})+\mathcal{N}(\bm{0}, \sigma_{s}^{2}) , \end{split}$$ where $\mathbf{x}\odot \mathcal{P}(\lambda_{o})$ and $\mathbf{y}\odot P(\lambda_{s})$ indicates that the noise $\mathbf{n}_{o}$ and $\mathbf{n}_{s}$ are element-wisely dependent on $\mathbf{x}$ and $\mathbf{y}$, respectively. The “$\approx$” is valid if we assume that the observed noise $\mathbf{n}_{o}$ is “weak” when compared to the signal $\mathbf{x}$.To this end, we have $$\label{e13} \mathbf{n}_{o}+\mathbf{n}_{s} \sim \mathbf{x}\odot \mathcal{P}(\lambda_{o}+\lambda_{s})+\mathcal{N}(\bm{0}, \sigma_{o}^{2}+\sigma_{s}^{2}+2\rho\sigma_{o}\sigma_{s}) ,$$ where $\rho$ is the correlation between $\mathbf{n}_{o}$ and $\mathbf{n}_{s}$ ($\rho=0$ if they are independent). This indicates that the summed noise variable $\mathbf{n}_{o}+\mathbf{n}_{s}$ still follows a mixed $\mathbf{x}$ dependent Poisson and Gaussian distribution, guaranteeing the consistency in noise statistics between the *observed* realistic noise and the *simulated* noise. As will be validated by the experiments (§\[sec:exp\]), this property makes our NAC strategy consistently effective on different noise removal tasks. ------------------------ ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- PSNR$\uparrow$ SSIM$\uparrow$ PSNR$\uparrow$ SSIM$\uparrow$ PSNR$\uparrow$ SSIM$\uparrow$ PSNR$\uparrow$ SSIM$\uparrow$ PSNR$\uparrow$ SSIM$\uparrow$ **BM3D** [@bm3d] 38.07 0.9580 34.40 0.9234 32.38 0.8957 31.00 0.8717 29.97 0.8503 **DnCNN** [@dncnn] 38.76 0.9633 34.78 0.9270 32.86 0.9027 31.45 0.8799 30.43 0.8617 **N2N** [@noise2noise] **39.72** 0.9665 36.18 0.9446 33.99 0.9149 **32.10** 0.8788 30.72 0.8446 **DIP** [@dip] 32.49 0.9344 31.49 0.9299 29.59 0.8636 27.67 0.8531 25.82 0.7723 **N2V** [@noise2void] 27.06 0.8174 26.79 0.7859 26.12 0.7468 25.89 0.7405 25.01 0.6564 **DnCNN+NAC** **0.9336** 33.64 0.8697 31.15 0.8024 29.22 0.7382 **Blind DnCNN+NAC** **0.9333** 33.63 0.8693 31.14 0.8018 29.21 0.7376 **ResNet+NAC** **36.55** **Blind ResNet+NAC** 38.48 **0.9805** ------------------------ ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- \[t-g12\] Learning Self-supervised Denoising Networks\ by “Noisy-As-Clean” {#sec:self-sup} ============================================ Here, we propose to learn self-supervised denoising networks with our “Noisy-As-Clean” (NAC) strategy. We employ the DnCNN [@dncnn] and ResNet in DIP [@dip] as our baseline, and call the self-supervised networks as DnCNN+NAC and ResNet+NAC, respectively.Note that we only need the *observed* noisy image $\mathbf{y}$ to generate noisy image pairs $\{(\mathbf{z},\mathbf{y})\}$ with *simulated* noise $\mathbf{n}_{s}$, as illustrated in Figure \[f1\]. **Training self-supervised networks by our NAC**. For real-world images captured by camera sensors, one can hardly distinguish the realistic noise from the signal. The signal intensity $\mathbf{x}$ is usually stronger than the noise intensity. That is, the expectation of the observed realistic noise $\mathbf{n}_{o}$ is usually much smaller than that of the latent clean image $\mathbf{x}$. If we train an image-specific network for the new noisy image $\mathbf{z}$ and regard the original noisy image $\mathbf{y}$ as the ground-truth image, then the trained image-specific network basically joint learn the image-specific prior and noise statistics. It has the capacity to remove the noise $\mathbf{n}_{s}$ from the new noisy image $\mathbf{z}$. Then if we perform denoising on the original noisy image $\mathbf{y}$, the observed noise $\mathbf{n}_{o}$ can be well-removed. Note that we *do not* use the clean image $\mathbf{x}$ as “ground-truth” in training the DnCNN+NAC and ResNet+NAC networks.  **Training blind denoising networks**. Most of existing supervised denoising networks train a specific model to process a fixed noise pattern [@crosschannel2016; @nlrn2018; @upi]. To tackle unknown noise, one feasible solution for these networks is to assume the noise as AWGN and estimate its noise deviation. The corresponding noise is removed by using the networks trained with the estimated level. But this strategy largely degrades the denoising performance when the noise deviation is not estimated accurately. Besides, this solution can hardly deal with realistic noise, which is usually not AWGN, captured on real photographs. In order to be effective on removing realistic noise, the self-supervised networks by our NAC are feasible to blindly remove the unknown noise from real photographs. Inspired by [@dncnn; @cbdnet], we propose to train a blind version of DnCNN+NAC and ResNet+NAC networks by using the AWGN noise within a range of levels (e.g., $[0, 55]$) for removing unknown AWGN noise. We also train blind ResNet+NAC with mixed AWGN and Poisson noise (both within a range of intensities) for removing the realistic noise. More details will be explained in §\[sec:blind\]. ------------------------ ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- PSNR$\uparrow$ SSIM$\uparrow$ PSNR$\uparrow$ SSIM$\uparrow$ PSNR$\uparrow$ SSIM$\uparrow$ PSNR$\uparrow$ SSIM$\uparrow$ PSNR$\uparrow$ SSIM$\uparrow$ **BM3D** [@bm3d] 37.59 0.9640 33.32 0.9163 31.07 0.8720 29.62 0.8342 28.57 0.8017 **DnCNN** [@dncnn] 38.07 33.88 31.73 0.8706 **30.27** **0.8563** **N2N** [@noise2noise] **38.58** 0.9627 34.07 0.9200 **31.81** **0.8770** 30.14 0.8550 28.67 0.8123 **DIP** [@dip] 29.74 0.8435 28.16 0.8310 27.07 0.7867 25.80 0.7205 24.63 0.6680 **N2V** [@noise2void] 26.70 0.7915 26.39 0.7621 25.77 0.7126 25.41 0.6678 24.83 0.6305 **DnCNN+NAC** **0.9674** **34.21** 0.8913 30.72 0.8044 28.25 0.7230 26.34 0.6515 **Blind DnCNN+NAC** **0.9674** **34.21** 0.8911 30.71 0.8041 28.24 0.7227 26.33 0.6511 **ResNet+NAC** **28.96** **0.8185** **Blind ResNet+NAC** 38.26 0.9605 **0.9266** ------------------------ ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- \[t-g68\] **Testing** is performed by directly regarding an *observed* noisy image $\mathbf{y}=\mathbf{x}+\mathbf{n}_{o}$ as input.  We only test the image $\mathbf{y}$ once.  The denoised image can be represented as $\hat{\mathbf{y}}=f_{\theta^{*}}(\mathbf{y})$, with which the objective metrics, e.g., PSNR and SSIM [@ssim], can be computed with the clean image $\mathbf{x}$. **Implementation details**. We employ the DnCNN [@dncnn] and ResNet in DIP [@dip] as the backbones, and turn them into self-supervised networks by our NAC strategy, which are named as DnCNN+NAC and ResNet+NAC, respectively. The DnCNN contains 17 layers of convolution, Batch Normalization (BN) [@bn2015], and Rectified Linear Units (ReLU) activation operator [@relu2010].  To accommodate DnCNN with our NAC strategy, we set the output of DnCNN+NAC as the denoised image, not the residual noise in DnCNN [@dncnn].  We observe no difference between the results on PSNR, SSIM [@ssim], and visual quality by employing these two types of outputs in our experiments. As DnCNN, the parameters of DnCNN+NAC are initialized from a pretrained ResNet. As used in [@dip], the ResNet in our ResNet+NAC includes $10$ residual blocks, each containing two convolutional layers followed by a BN [@bn2015] and a ReLU [@relu2010] after the first BN.  The parameters are randomly initialized without being pretrained.  For both baselines, the optimizer is Adam [@adam] with default parameters.  The learning rate is fixed at $0.001$ in all experiments.  We use the $\ell_{2}$ loss function.  For each test image, we only train the DnCNN+NAC in 100 epochs, while the original DnCNN is trained with 180 epochs.  The ResNet+NAC is trained in $1000$ epochs for each test image, the same as that in DIP [@dip]. As suggested by DnCNN [@dncnn] and DIP [@dip], we employ $4$ rotations {0, 90, 180, 270} combined with 2 mirror (vertical and horizontal) reflections, resulting in totally $8$ transformations for data augmentation.  We implement the DnCNN+NAC and ResNet+NAC networks in PyTorch. Experiments {#sec:exp} =========== In this section, we evaluate the performance of our “Noisy-As-Clean” (NAC) networks on image denoising. In all experiments, we train a denoising network using only the noisy test image $\mathbf{y}$ as the target, and using the *simulated* noisy image $\mathbf{z}$ as the input.  For all comparison methods, the source codes or trained models are downloaded from the corresponding authors’ websites. We use the default parameter settings, unless otherwise specified.The PSNR, SSIM [@ssim], and visual quality of different methods are used to evaluate the comparison.  We first test with synthetic noise such as additive white Gaussian noise (AWGN) in §\[sec:syn\], continue to perform blind image denoising in §\[sec:blind\], and finally tackle the realistic noise in §\[sec:real\].  In §\[sec:dis\], we conduct comprehensive ablation studies to gain deeper insights into our NAC strategy. -- ---------------- -------------------- --------------------- -------------------- ---------------------- ------------------ ------------------------ ---------------- ---------------------- Method **CBM3D** [@cbm3d] **NI** [@neatimage] **DnCNN+**[@dncnn] **CBDNet** [@cbdnet] **GCBD** [@gcbd] **N2N** [@noise2noise] **DIP** [@dip] **Blind ResNet+NAC** PSNR$\uparrow$ 35.19 35.33 35.40 NA 35.32 **35.69** SSIM$\uparrow$ 0.9063 0.9212 0.9115 NA 0.9160 **0.9259** PSNR$\uparrow$ 34.51 35.11 35.58 33.10 NA **36.20** SSIM$\uparrow$ 0.8507 0.8778 0.9217 0.8110 NA **0.9252** -- ---------------- -------------------- --------------------- -------------------- ---------------------- ------------------ ------------------------ ---------------- ---------------------- \[t-cc+dnd\] Synthetic Noise Removal With Known Noise {#sec:syn} ---------------------------------------- We evaluate the DnCNN+NAC and ResNet+NAC networks on images corrupted by synthetic AWGN noise.  More results on signal dependent Poisson noise and mixed Poisson-AWGN noise are provided in the *Supplementary File*. **Training self-supervised networks**. Here, we train an image-specific denoising network using the *observed* noisy test image $\mathbf{y}$ as the target, and the *simulated* noisy image $\mathbf{z}$ as the input. Each *observed* noisy image $\mathbf{y}=\mathbf{x}+\mathbf{n}_{o}$ is generated by adding the *observed* noise $\mathbf{n}_{o}$ to the clean image $\mathbf{x}$. The *simulated* noisy image $\mathbf{z}=\mathbf{y}+\mathbf{n}_{s}$ is generated by adding *simulated* noise $\mathbf{n}_{s}$ to *observed* noisy image $\mathbf{y}$. **Comparison methods**.We compare DnCNN+NAC and ResNet+NAC networks with state-of-the-art image denoising methods [@bm3d; @dncnn; @noise2noise]. On AWGN noise, we compare with BM3D [@bm3d], DnCNN [@dncnn], Noise2Noise (N2N) [@noise2noise], Deep Image Prior (DIP) [@dip], and Noise2Void (N2V) [@noise2void].  **Test datasets**. We evaluate the comparison methods on the *Set12* and *BSD68* datasets, which are widely tested by supervised denoising networks [@dncnn; @nlrn2018] and previous methods [@bm3d; @pgpd]. The *Set12* dataset contains 12 images of sizes $512\times512$ or $256\times256$, while the *BSD68* dataset contains 68 images of different sizes. **Results on AWGN noise** with noise levels (standard deviation, or std) of $\sigma\in\{5, 10, 15, 20, 25\}$ are provided here.  The *observed* noise $\mathbf{n}_{o}$ is AWGN with std of $\sigma$, while the *simulated* noise $\mathbf{n}_{s}$ is with the same $\sigma$ as that of $\mathbf{n}_{o}$.  The comparison results are listed in Tables \[t-g12\] and \[t-g68\].  It can be seen that, DnCNN+NAC achieves better PSNR and SSIM results than those of the original DnCNN when $\sigma=5,10$. Note that DnCNN are supervised networks trained offline on the *BSD400* dataset, while the variant DnCNN+NAC network is trained online for each corrupted image.Besides, the blind version of DnCNN+NAC achieves negligible performance drop when compared to the DnCNN+NAC, which is consistent with [@dncnn].  On the other side, the ResNet+NAC networks achieve comparable or better performance on PSNR and SSIM [@ssim] than BM3D [@bm3d] and DnCNN [@dncnn], especially when the noise levels are weak ($\sigma=5,10$).  Besides, our ResNet+NAC networks outperform the other unsupervised and self-supervised networks such as N2N [@noise2noise], DIP [@dip], and N2V [@noise2void] by a large margin on PSNR and SSIM [@ssim].In Figures \[fig:awgn15\] and \[fig:awgn5\], we provide the visual comparisons of the denoised images by the competing methods.  One can see that the ResNet+NAC networks produce better image quality and higher PSNR/SSIM results than the comparison methods. Synthetic Noise Removal With Unknown Noise {#sec:blind} ------------------------------------------ To deal with unknown noise, we propose to train blind versions of the DnCNN [@dncnn] and ResNet in [@dip] by our NAC strategy.  Here, we test the Blind DnCNN+NAC and Blind ResNet+NAC networks on AWGN noise with unknown noise deviation. We use the same training strategy, comparison methods, and test datasets as in §\[sec:syn\]. **Training blind networks**. We train the Blind DnCNN+NAC and Blind ResNet+NAC networks on the corrupted test image degraded *again* by AWGN noise with unknown noise levels (deviations).The noise levels are randomly sampled in Gaussian distribution within $[0, 55]$. We also test on noise levels in uniform distribution and obtain similar results. We repeat the training of DnCNN+NAC and ResNet+NAC networks on the test image with different deviations.  **Results on blind denoising**. For the same test image, we add to it the AWGN noise whose deviation is also in $\{5, 10, 15, 20, 25\}$. The blindly trained DnCNN+NAC and ResNet+NAC networks are directly utilized to denoise the test image without estimating its deviation. The results are also listed in Tables \[t-g12\] and \[t-g68\]. We observe that, the Blind ResNet+NAC networks trained on AWGN noise with unknown levels can achieve even better PSNR and SSIM [@ssim] results than the ResNet+NAC networks trained on specific noise levels. Note that on *BSD68*, the ResNet+NAC networks achieve higher PSNR and SSIM results than DnCNN [@dncnn]. This demonstrates the effectiveness of our ResNet+NAC networks on blind image denoising. With the success on blind image, next we will turn to real-world image denoising, in which the noise is also unknown and very complex. Practice on Real Photographs {#sec:real} ---------------------------- With the promising performance on blind image denoising, here we tackle the realistic noise for practical applications. The *observed* realistic noise $\mathbf{n}_{o}$ can be roughly modeled as mixed Poisson noise and AWGN noise [@poissongaussian; @cbdnet]. Hence, for each *observed* noisy image $\mathbf{y}$, we generate the *simulated* noise $\mathbf{n}_{s}$ by sampling the $\mathbf{y}$-dependent Poisson part and the independent AWGN noise. **Training blind ResNet+NAC networks** is also performed for each test image, i.e., the *observed* noisy image $\mathbf{y}$. In real-world scenarios, each *observed* noisy image $\mathbf{y}$ is corrupted without knowing the specific noise statistics of the *observed* noise $\mathbf{n}_{o}$. Therefore, the *simulated* noise $\mathbf{n}_{s}$ is directly estimated on $\mathbf{y}$ as mixed $\mathbf{y}$-dependent Poisson and AWGN noise. For each transformation image in data augmentation, the Poisson noise is randomly sampled with the parameter $\lambda$ in $0<\lambda\le25$, and the AWGN noise is randomly sampled with the noise level $\sigma$ in $0<\sigma\le25$. **Comparison methods**.We compare with state-of-the-art methods on real-world image denoising, including CBM3D [@cbm3d], the commercial software Neat Image [@neatimage], two supervised networks DnCNN+ [@dncnn] and CBDNet [@cbdnet], and two unsupervised networks GCBD [@gcbd] and Noise2Noise [@noise2noise], and the self-supervised network DIP [@dip]. Note that DnCNN+ [@dncnn] and CBDNet [@cbdnet] are two state-of-the-art supervised networks for real-world image denoising, and DnCNN+ is an improved extension of DnCNN [@dncnn] with better performance (the authors of DnCNN+ provide us the models/results of DnCNN+). {width="100.00000%"} {width="100.00000%"} {width="100.00000%"} {width="100.00000%"} {width="100.00000%"} {width="100.00000%"} {width="100.00000%"} {width="100.00000%"} \[fig:dnd\] **Test datasets**. We evaluate the comparison methods on the *Cross-Channel* (*CC*) dataset [@crosschannel2016] and *DND* dataset [@dnd2017].The *CC* dataset [@crosschannel2016] includes noisy images of 11 static scenes captured by Canon 5D Mark 3, Nikon D600, and Nikon D800 cameras. The noisy images are collected under a highly controlled indoor environment. Each scene is shot $500$ times using the same settings. The average of the $500$ shots is taken as “ground-truth”. We use the default $15$ images of size $512\times512$ cropped by the authors to evaluate different image denoising methods.  The *DND* dataset [@dnd2017] contains 50 scenarios captured by Sony A7R, Olympus E-M10, Sony RX100 IV, and Huawei Nexus 6P. Each scene is cropped to 20 bounding boxes of $512\times512$ pixels, generating totally 1000 test images. The noisy images are collected under higher ISO values with shorter exposure times, while the “ground truth” images are captured under lower ISO values with adjusted longer exposure times. The “ground truth” images are not released, but we can obtain the PSNR and SSIM results by submitting the denoised images to the [*DND*’s Website](https://noise.visinf.tu-darmstadt.de/benchmark/#results_srgb). **Comparison results on PSNR and SSIM** are listed in Table \[t-cc+dnd\].  As can be seen, the ResNet+NAC networks achieve better performance than all previous denoising methods, including the CBM3D [@cbm3d], the supervised networks DnCNN+ [@dncnn] and CBDNet [@cbdnet], and the unsupervised networks GCBD [@gcbd], N2N [@noise2noise], and DIP [@dip]. This demonstrates that the ResNet+NAC networks can indeed handle the complex, unknown, and realistic noise, and achieve better performance than supervised networks such as DnCNN+ [@dncnn] and CBDNet [@cbdnet]. **Qualitative results**.  In Figures \[fig:cc\] and \[fig:dnd\], we show the denoised images of our ResNet+NAC and the comparison methods on the images of “*5dmark3-iso3200-1*” from the *CC* dataset [@crosschannel2016] and “*0017\_3*” from the *DND* dataset [@dnd2017], respectively. We observe that our self-supervised Blind ResNet+NAC is very effective on removing realistic noise from the real photograph. Besides, the Blind ResNet+NAC networks achieve competitive PSNR and SSIM results when compared with the other methods, including the supervised DnCNN+ [@dncnn] and CBDNet [@cbdnet]. **Speed**. The work most similar to ours is Deep Image Prior (DIP) [@dip], which also trains an image-specific network for each test image. Averagely, DIP needs $603.9$ seconds to process a $512\times512$ color image, on which our ResNet+NAC network needs $583.2$ seconds (on an NVIDIA Titan X GPU). Ablation Study {#sec:dis} -------------- To further study our NAC strategy, we conduct more examination of our ResNet+NAC networks on image denoising. Specifically, we assess 1) differences of the ResNet+NAC from the ResNet in DIP [@dip]; 2) how the number of residual blocks and epochs influence the ResNet+NAC; 3) comparison with the “Oracle” performance of the ResNet+NAC networks; 4) performance of the ResNet+NAC on “strong” noise. **1) Differences from DIP [@dip]**. Though the basic network in our work is the ResNet used in DIP [@dip], our ResNet+NAC network is essentially different from DIP on at least two aspects. First, our ResNet+NAC is a novel strategy for self-supervised learning of *adaptive network parameters* for the degraded image, while DIP aims to investigate *adaptive network structure* without learning the parameters. Second, our ResNet+NAC learns a mapping from the synthetic noisy image $\mathbf{z}=\mathbf{y}+\mathbf{n}_{s}$ to the noisy image $\mathbf{y}$, which approximates the mapping from the noisy image $\mathbf{y}=\mathbf{x}+\mathbf{n}_{o}$ to the clean image $\mathbf{x}$. But DIP maps a random noise map to the noisy image $\mathbf{y}$, and the denoised image is obtained during the process. Due to the two reasons, DIP needs early stop for different images, while our ResNet+NAC achieves more robust (and better) denoising performance than DIP on diverse images. In Figure \[fig:curve\], we plot the curves of training loss and test PSNR of DIP (a) and ResNet+NAC (b) networks in 10,000 epochs, on two images of “Cameraman” and “House”. We observe that DIP needs early stop to select the best results, while our ResNet+NAC can stably achieve better denoising results within 1000 epochs. \[fig:curve\] \[f-oracle-strong\] **2) Influence on the number of residual blocks and epochs**. Our backbone network is the ResNet [@dip] with 10 residual blocks trained in 1000 epochs. Now we study how the number of residual blocks and epochs influence the performance of ResNet+NAC on image denoising. The experiments are performed on the *Set12* dataset corrupted by AWGN noise ($\sigma=15$). From Table \[t-block\], we observe that, with more residual blocks, the ResNet+NAC networks can achieve better PSNR and SSIM [@ssim] results. And 10 residual blocks are enough to achieve satisfactory results. With more (e.g., 15) blocks, there is little improvement on PSRN and SSIM. Hence, we use 10 residual blocks the same as [@dip]. Then we study how the number of epochs influence the performance of ResNet+NAC on image denoising. From Table \[t-epoch\], one can see that on the *Set12* dataset corrupted by AWGN noise ($\sigma=15$), with more training epochs, our ResNet+NAC networks achieve better PSNR and SSIM results, but with longer processing time. \# of Blocks 1 2 5 10 15 ---------------- -------- -------- -------- -------- -------- -- PSNR$\uparrow$ 33.58 33.85 34.14 34.24 34.26 SSIM$\uparrow$ 0.9161 0.9226 0.9272 0.9277 0.9272 : **Average PSNR (dB)/SSIM of ResNet+NAC with different number of blocks** on *Set12* corrupted by AWGN noise ($\sigma=15$). \[t-block\] **3) Comparison with Oracle**. We also study the “Oracle” performance of the ResNet+NAC networks. In “Oracle”, we train the ResNet+NAC networks on the pair of *observed* noisy image $\mathbf{y}$ and its clean image $\mathbf{x}$ corrupted by AWGN noise or signal dependent Poisson noise. The experiments are performed on *Set12* dataset corrupted by AWGN or signal dependent Poisson noise. The noise deviations are in $\{5,10,15,20,25\}$. Figure \[f-oracle-strong\] (a) shows comparisons of our ResNet+NAC and its “Oracle” networks on PSNR and SSIM. It can be seen that, the “Oracle” networks trained on the pair of noisy-clean images only perform slightly better than the original ResNet+NAC networks trained with the *simulated*-*observed* noisy image pairs $(\mathbf{z},\mathbf{y})$. With our NAC strategy, the ResNet networks trained only with noisy test images achieves similar promising performance on the weak noise. **4) Performance on strong noise**.Our NAC strategy is based on the assumption of “weak noise”. It is natural to wonder how well ResNet+NAC performs against strong noise. To answer this question, we compare the ResNet+NAC networks with BM3D [@bm3d] and DnCNN [@dncnn], on *Set12* corrupted by AWGN noise with $\sigma=50$. The PSNR and SSIM results are plotted in Figure \[f-oracle-strong\] (b). One can see that, our ResNet+NAC networks are limited in handling strong AWGN noise, when compared with BM3D [@bm3d] and DnCNN [@dncnn]. \# of Epochs 100 200 500 1000 5000 ------------------ -------- -------- -------- -------- -------- PSNR$\uparrow$ 31.80 32.79 33.77 34.24 34.28 SSIM$\uparrow$ 0.8714 0.9023 0.9189 0.9277 0.9280 Time$\downarrow$ 67.4 132.5 302.0 583.2 2815.6 : **Average PSNR (dB) and time (s) of ResNet+NAC with different number of epochs** on *Set12* corrupted by AWGN noise ($\sigma=15$). \[t-epoch\] Conclusion {#sec:con} ========== In this work, we proposed a “Noisy-As-Clean” (NAC) strategy for learning self-supervised image denoising networks. In our NAC, we trained an image-specific network by taking the corrupted image as the target, and adding to it the simulated noise to generate the doubly corrupted noisy input. The simulated noise is close to the observed noise in the noisy test image. This strategy can be seamlessly embedded into existing supervised denoising networks. We observed that *it is possible to learn a self-supervised network only with the corrupted image, approximating the optimal parameters of a supervised network learned with a pair of noisy and clean images*. Extensive experiments on synthetic and real-world benchmarks demonstrate that, the DnCNN [@dncnn] and ResNet in Deep Image Prior [@dip] trained with our NAC strategy achieved comparable or better performance on PSNR, SSIM, and visual quality, when compared to previous state-of-the-art image denoising methods, including supervised denoising networks. These results validate that our NAC strategy can learn effctive image-specific priors and noise statistics only from the corrupted test image. [^1]: This work was supported in part by the Major Project for New Generation of AI under Grant No. 2018AAA0100400, NSFC (61922046), and Tianjin Natural Science Foundation (18ZXZNGX00110). ($^*$Corresponding author: M.-M. Cheng) [^2]: J. Xu and M.-M. Cheng are with TKLNDST, College of Computer Science, Nankai University, Tianjin, China (E-mail: csjunxu@nankai.edu.cn, cmm@nankai.edu.cn). Y. Huang is with School of Electronic and Information Engineering, Xi’an Jiaotong University, Xi’an, China. L. Liu, F. Zhu, and L. Shao are with Inception Institute of Artificial Intelligence (IIAI) and Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), Abu Dhabi, UAE. Zhou Xu is with School of Big Data and Software Engineering, Chongqing University, Chongqing, China. The first two authors contribute equally.

--- abstract: | In this article the research of Sylows $p$-subgroups of ${{A}_{n}}$ and ${{S}_{n}}$, which was started in [@Dm; @Sk; @Paw] is continued. Let $Syl_2{A_{2^k}}$ and $Syl_2{A_{n}}$ be Sylow 2-subgroups of corresponding alternating groups $A_{2^k}$ and $A_{n}$. We find a least generating set and a structure for such subgroups $Sy{{l}_{2}}{{A}_{{{2}^{k}}}}$ and $Syl_2{A_{n}}$ and commutator width of $Syl_2{A_{2^k}}$ [@Mur]. The purpose of this paper is to research the structure of a Sylow 2-subgroups and to construct a minimal generating set for such subgroups. The main result is to prove minimality of this generating set for the above indicated subgroups and also the description of their structure. Key words: minimal set of generators; wreath product of group; Sylow subgroups; commutator subgroup, semidirect product. author: - Ruslan Skuratovskii title: 'Structure and minimal generating sets of Sylow 2-subgroups of alternating groups, properties of its commutator subgroup' --- Introduction ============= The aim of this paper is to research the structure of Sylow 2-subgroups of $A_{2^k}$, $A_n$ and to construct a minimal generating set for $Syl_2 {A_{2^k}}$. The case of Sylow subgroup where $p=2$ is very special because group $C_2\wr C_2\wr C_2 \ldots \wr C_2 $ admits odd permutations, this case was not fully investigated in [@Dm; @Sk]. The authors of [@Dm; @Paw] didn’t prove minimality of found by them system of generators for such Sylow 2-subgroups of $A_n$ and structure of a Sylow 2-subgroups was found by them not fully, for case $n=2^k$ structure of $Syl_2 A_n$ was not found. This question is up till now under consideration. There was a mistake in a statement about irreducibility of a set of $k+1$ elements for $Syl_2(A_{2^k})$ that was in abstract [@Iv] on Four ukraine conference of young scientists in 2015 year. These groups have applications in the automaton theory, because if all states of automaton $A$ have output function that can be presented as cycle $(1,2,...,p)$ then group $G_A(X)$ of this automaton, where $X$ is finite alphabet, is Sylows $p$-subgroup of the group of all automaton transformations $GA(X)$ [@GrNe]. Also in this case holds $Syl_p(AutX)>FGA(X)$ [@GrNe], where $ \mid X \mid \in \mathbb{N} $. Thus, finding the minimum size of generating set number is important. Let $X^*$ be the free monoid freely generated by $X$. All undeclared terms are from [@Ne; @Gr]. Preliminaries ============== The set $X^*$ is naturally a vertex set of a regular rooted tree, i.e. a connected graph without cycles and a designated vertex $v_0$ called the root, in which two words are connected by an edge if and only if they are of form $v$ and $vx$, where $v\in X^*$, $x\in X$. The set $X^n \subset X^*$ is called the $n$-th level of the tree $X^*$ and $X^0 = \{v_0\}$. We denote by $v_{j,i}$ the vertex of $X^j$, which has the number $i$. The subtree of $X^{*}$ induced by the set of vertices $\cup_{i=0}^k X^i$ is denoted by $X^{[k]}$. Note that the unique vertex $v_{k,i}$ corresponds to the unique word $v$ in alphabet $X$. For every automorphism $g\in Aut{{X}^{*}}$ and every word $v \in X^{*}$ define the section (state) $g_{(v)} \in AutX^{*}$ of $g$ at $v$ by the rule: $g_{(v)}(x) = y$ for $x, y \in X^*$ if and only if $g(vx) = g(v)y$. The restriction of the action of an automorphism $g\in AutX^*$ to the subtree $X^{[l]}$ is denoted by $g_{(v)}|_{X^{[l]}}$. A restriction $g_{(v)}|_{X^{[1]}} $ is called the vertex permutation (v.p.) of $g$ in a vertex $v$. Let us introduce conventional signs for a v.p. state value of $\alpha $ in $v_{ki}$ as ${{s}_{ki}}(\alpha )$ we put that ${{s}_{ki}}(\alpha )=1$ if $\alpha_{(v_{ki})} |_{X^{[1]}}(x)=y,\,\,\,x\ne y $ such state of v.p. is active, and $s_{ki}(\alpha )=0$ if $\alpha_{(v_{ki})} |_{X^{[1]}}(x)=x $ such state of v.p. is trivial. Let us label every vertex of ${{X}^{l}},\,\,\,0\le l<k$ by sign 0 or 1 in relation to state of v.p. in it. Obtained by such way a vertex-labeled regular tree is element of $Aut{{X}^{[k]}}$. We denote by $v_{j,i} X^{[k-j]} $ subtree of $X^{[k]}$ with a root in $v_{j,i}$. An automorphism of $X^{[k]}$ with non-trivial states of v.p. in some of $v_{1,1}$, $v_{1,{2}}$, $v_{2,{1}}$,..., $v_{2,4}$, ... ,$v_{m,1}$, ... ,$v_{m,j}$, $ m < k, j \leq 2^m$ is denoted by $\mathop{\beta}_{1,(i_{11},i_{12});...; l,(i_{l1},...,i_{l2^l});...;{m},(i_{m1},...,i_{m2^m})}$ where the index that stands straight before parentheses are number of level in parentheses we write a tuple of states of v.p. of this level. In other words we set $i_{mj}=0$ if v.p. in $v_{mj}$ is trivial, $i_{mj}=1$ in other case, i.e. $i_{mj} = {{s}_{mj}}(\mathop{\beta})$, where $\mathop{\beta}\in AutX^{[k]}$, $m<k$. If for some $l$ all $i_{lj}=0$ then $2^l$-tuple $l,(i_{l1} ,..., i_{l 2^l})$ does not figure in indexes of $\beta$. But if numbers of active vertices are certain, for example $v_{j,1}$ and $v_{j,s}$, we can use more easy notation $\mathop{\beta}_{j,(1,s);}$, where in parentheses numbers of vertices with active state of v.p. from level $j$. If in parentheses only one index then parentheses can be omitted for instance $\mathop{\beta}_{j,(s);}=\mathop{\beta}_{j,s;}$. Denote by $\tau_{i,...,j}$ the automorphism of $X^{[k]}$, which has a non-trivial v.p. only in vertices $v_{k-1,i}$, ... ,$v_{k-1, j}$, $j \leq 2^{k-1} $ of the level $X^{k-1}$. Denote by $\tau$ the automorphism $\tau_{1,2^{k-1}}$. Let us consider special elements such that: $ \mathop{\alpha}_{0}=\mathop{\beta}_{0}=\mathop{\beta}_{0,(1,0,...,0)}, \mathop{\alpha}_{1}=\mathop{\beta}_{1}=\mathop{\beta}_{1,(1,0,...,0)} , \ldots ,\mathop{\alpha}_{l}=\mathop{\beta}_{l}=\mathop{\beta}_{l,(1,0,...,0)} $. Main result ============= Recall that a wreath product of permutation groups is associative construction. We consider $C_2$ as additive group with two elements 0, 1. For constructing a wreath product we define an action of $C_2$ by shift on $X=\{0, 1\}$. As well known that $AutX^{[k-1]} \simeq \underbrace {C_2 \wr ...\wr C_2}_{k-1}$ [@GrNe]. \[even\] Every automorphism that has active v.p. only on $X^l$, $l<k-1$ acts by even permutation on $X^{k}$. Actually every transposition in vertex from $X^l$, $l<k-1$ acts on even number of pair of vertexes because of binary tree structure. More precisely it realize even permutation on the set $X^{k}$ with cyclic structure [@Sh] $(1^{2^{k-1}-2^{k-l-l}},2^{2^{k-l-l}})$ because it formed by the structure of binary tree. \[B\_k-1\] Due to Lemma \[even\] automorphisms from $Aut X^{[k-1]}=\langle \mathop{\alpha}_{0}. ... , \mathop{\alpha}_{k-2} \rangle$ form a group ${B_{k-1}}= \underbrace {C_2 \wr ...\wr C_2}_{k-1} $ which acts on $X^{k-1}$ by even permutations. Size of $B_{k-1}$ equal to $2^{2^{k-1}-1}$. Let us denote by $W_{k-1}$ the subgroup of $Aut X^{[k]}$ such that has active states only on $ X^{k-1}$ and number of such states is even, i.e. $W_{k-1} \vartriangleleft St_{G_k}(k-1)$ [@Ne]. \[ordW\] The size of ${{W}_{k-1}}$ is equal to ${{2}^{{{2}^{k-1}}-1}},\,\,k > 1$ and its structure is $(C_2)^{{{2}^{k-1}}-1}$. On ${{X}^{k-1}}$ we have ${{2}^{k-1}}$ vertices where can be elements of a group ${{V}_{k-1}}\simeq {{C}_{2}}\times {{C}_{2}}\times ...\times {{C}_{2}}\simeq {{({{C}_{2}})}^{k-1}}$, but as a result of the fact that ${X}^{k-1}$ contains only even number of non trivial v.p. from $X^{k-1}$, there are only half of all permutations from ${{V}_{k-1}} \simeq St_{G_k}(k-1)$ on $X^{k-1}$. So it is subgroup ${{W}_{k-1}}\simeq {}^{C_{2}^{{{2}^{k-1}}}}/{}_{{{C}_{2}}}$ of ${{V}_{k-1}}$. So we can state that $|{{W}_{k-1}}|=2^{k-1}-1$, $W_{k-1}$ has $k-1$ generators and we can consider ${W}_{k-1}$ as vector space of dimension $k-1$. For example let us consider the subgroup $W_{4-1}$ of $A_{2^4}$ its size is $2^{2^{4-1}-1}=2^7$ and $|A_{2^4}|=2^{14}$. Let us denote by $G_k$ the subgroup of $Aut X^{[k]}$ such that $G_k \simeq B_{k-1} \ltimes W_{k-1}$. \[gen\] The elements $\tau $ and ${{\alpha }_{0}},...,{{\alpha }_{k-1}}$ generate arbitrary element ${{\tau }_{ij}}$. According to [@Gr; @Sk] the set ${{\alpha }_{0}},...,{{\alpha }_{k-2}}$ is minimal generating set for group $Aut{{X}^{[k-1]}}$. Firstly, we shall prove the possibility of generating an arbitrary $\tau_{ij}$, from set $v_{(k-1,i)}$, $1\leq i \leq 2^{k-2} $. Since $Aut v_{1,1}X^{[k-2]} \simeq \left\langle {{\alpha }_{1}},...,{{\alpha }_{k-2}} \right\rangle $ acts on ${{X}^{k-1}}$ transitively from it follows existing of an ability to permute vertex with a transposition from automorphism $\tau $ and stands in $v_{k-1,1}$ in arbitrary vertex ${{v}_{k-1,j}},\,\,\,j\le {{2}^{k-2}}$ of $v_{1,1}X^{[k-1]}$. For this goal we act by $\alpha_{k-j}$ at $\tau $: $\alpha_{k-j} \tau \alpha_{k-j} ={{\tau }_{j, 2^{k-2}}}$. Similarly we act on $\tau$ by corespondent $\alpha_{k-i}$ to get $\tau_{i, 2^{k-2}}$ from $\tau $: $\alpha_{k-i} \tau \alpha_{k-i}^{-1}={{\tau }_{i, 2^{k-2}}}$. Note that automorphisms $\alpha_{k-j}$ and $\alpha_{k-i}, 1<i,j<k-1$ acts only on subtree $v_{1,1}X^{[k-1]}$ that’s why they fix v.p. in $v_{k-1, 2^{k-1}}$. Now we see that ${\tau }_{i, 2^{k-2}}{{\tau }_{j, 2^{k-2}}}={{\tau }_{i, j}}$, where $1 \leq i,j < 2^{k-2}$. To get ${\tau }_{m, l}$ from $v_{1,2}X^{[k-1]}$, i.e. $2^{k-2} < m,l \leq 2^{k-1} $ we use $\alpha_0$ to map ${\tau }_{i, j}$ in ${\tau }_{i+2^{k-2}, j+2^{k-2}}\in v_{1,2} AutX^{[k-1]}$. To construct arbitrary transposition ${\tau }_{i,m}$ from $W_{k-1}$ we have to multiply ${\tau }_{1,i} {\tau } {\tau }_{m,2^{k-1}}={\tau }_{i,m}$. Let us realize a natural number of $v_{k,l}$, $1<l<2^k$ in 2-adic set of presentation (binary arithmetic). Then $l={\delta_{{{l}_{1}}}}{{2}^{m_l}}+{\delta_{{{l}_{2}}}}{{2}^{m_l-1}}+...+{\delta_{{{l}_{m_l+1}}}},\,\, \delta_{l_i} \in \{0,1\}$ where is a correspondence between ${\delta_{{{l}_{i}}}}$ that from such presentation and expressing of automorphisms: $\tau_{l,2^{k-1}} = \prod_{i=1}^{m_l} \alpha_{k-2-(m_{l}-i)}^{\delta_{{l}_{i}}} \tau \prod_{i=1}^{m_l} \alpha_{k-2-(m_{l}-i)}^{\delta_{l_i}}, 1 \leq m_l \leq k-2$. In other words $\left\langle {{\alpha }_{0}},...,{{\alpha }_{k-2}},\tau \right\rangle \simeq {{G}_{k}}$. \[genG\_k-1\] The elements from condition of Lemma \[gen\] are enough to generate a basis of $W_{k-1}$. \[ordG\_k\] Sizes of groups $G_k = \langle \mathop{\alpha}_{0}, \mathop{\alpha}_{1}, \mathop{\alpha}_{2},...,\mathop{\alpha}_{k-2}, \mathop{\tau} \rangle $ and $Syl_2(A_{2^{k}})$ are equal to $2^{2^{k}-2}$. In accordance with Legendre’s formula, the power of 2 in ${{2}^{k}}!$ is $\left[ \frac{{{2}^{k}}}{2} \right]+\left[ \frac{{{2}^{k}}}{{{2}^{2}}} \right]+\left[ \frac{{{2}^{k}}}{{{2}^{3}}} \right]+...+\left[ \frac{{{2}^{k}}}{{{2}^{k}}} \right]=\frac{{{2}^{k}}-1}{2-1}$. We need to subtract 1 from it because we have only $\frac{n!} {2}$ of all permutations as a result: $\frac{{{2}^{k}}-1}{2-1}-1=2^{k}-2$. So $\left| Syl({{A}_{{{2}^{k}}}}) \right|={{2}^{{{2}^{k}}-2}}$. The same size has group $G_k=B_{k-1} \ltimes W_{k-1}$ and $|G_k|=|B_{k-1}|\cdot|W_{k-1}|= |Syl_2 A_{2^k}|$. Since size of groups $G_{k}$ according to Proposition \[ordW\] and the fact that $|B_{k-1}|=2^{2^{k-1}-1}$ is $2^{2^{k}-2}$. For instance the sizes of $Syl_2 (A_8)$, $B_{3-1}$ and $W_{3-1}$ are such $|W_{3-1}|= 2^{2^{3-1}-1}=2^3=8$, $|B_{3-1}|=|C_2\wr C_2| = 2 \cdot 2^2=2^3$ and according to Legendre’s formula, the power of 2 in ${{2}^{k}}!$ is $\frac{{2}^{3}}{2}+ \frac{{2}^{3}}{2^2}+\frac{{2}^{3}}{2^3} -1=6$ so $Syl_2 (A_8) = 2^6=2^{2^k-2}$, where $k=3$. Next example for $A_{16}$: $Syl_2 (A_{16}) =2^{2^4-2}= 2^{14}, k=4$, $|W_{4-1}|= 2^{2^{4-1}-1}=2^7$, $|B_{4-1}|=|C_2\wr C_2\wr C_2| = 2 \cdot 2^2\cdot 2^4 = 2^7$. So we have the $|A_{16}|=|W_{3}||B_{3}|$ equality which endorse the condition of this lemma. **A maximal 2-subgroup** of $Aut{{X}^{\left[ k \right]}}$ that acts by even permutations on ${{X}^{k}}$ has the structure of the semidirect product $G_k \simeq B_{k-1} \ltimes W_{k-1} $ and isomorphic to $Syl_2A_{2^k}$. A maximal 2-subgroup of $Aut{{X}^{\left[ k-1 \right]}}$ is isomorphic to ${{B}_{k-1}}\simeq \underbrace{{{C}_{2}}\wr {{C}_{2}}\wr ...\wr {{C}_{2}}}_{k-1}$ (this group acts on ${{X}^{k-1}}$). A maximal 2-subgroup which has elements with active states only on ${{X}^{k-1}}$ corresponds subgroup ${W}_{k-1}$. Since subgroups ${B}_{k-1}$ and ${{W}_{k-1}}$ are embedded in $Aut{{X}^{\left[ k \right]}}$, then define an action of ${{B}_{k-1}}$ on elements of ${{W}_{k-1}}$ as ${{\tau }^{\sigma }}=\sigma \tau {{\sigma }^{-1}},\,\,\,\sigma \in {{B}_{k-1}},\,\,\tau \in {{W}_{k-1}}$, i.e. action by inner automorphism (inner action) from $Aut{{X}^{\left[ k \right]}}$. Note that ${{W}_{k-1}}$ is subgroup of stabilizer of ${{X}^{k-1}}$ i.e. ${{W}_{k-1}}<St_{Aut{X}^{[k]}}(k-1)\lhd AutX^{[k]}$ and is normal too $W_{k-1}\lhd AutX^{[k]}$, because conjugation keeps a cyclic structure of permutation so even permutation maps in even. Therefore such conjugation induce automorphism of ${W}_{k-1}$ and $G_k \simeq B_{k-1}\ltimes W_{k-1}$. Since at ${{X}^{k-1}}$ is ${{2}^{{{2}^{k-1}}}}$ vertexes and half of combinations of active states from ${X}^{k-1}$ can form even permutation thus $\left| {{W}_{k-1}} \right|={{2}^{{{2}^{k-1}}-1}}$ that is proved in Proposition \[ordW\]. Using the Corollary \[B\_k-1\] about ${{B}_{k-1}}$ we get size of ${{G}_{k}}\simeq {{B}_{k-1}} \ltimes {{W}_{k-1}}$ is ${{2}^{{{2}^{k-1}}-1}}\cdot {{2}^{{{2}^{k-1}}-1}}={{2}^{{{2}^{k}}-2}}$. Since $G_k$ is the maximal 2-subgroup then $G_k \simeq Syl_2A_{2^k}$. \[isomor\] The set $S_{\mathop{\alpha}}= \{\mathop{\alpha}_{0}, \mathop{\alpha}_{1}, \mathop{\alpha}_{2}, ... ,\mathop{\alpha}_{k-2}, \mathop{\tau}\}$ of elements from subgroup of $AutX^{[k]}$ generates a group $G_k$ which isomorphic to $Syl_2(A_{2^{k}})$. As we see from Corollary \[B\_k-1\], Lemma \[gen\] and Corollary \[genG\_k-1\] group $G_k$ are generated by $S_{\mathop{\alpha}}$ and their sizes according to Lemma \[ordG\_k\] are equal. So according to Sylow’s theorems 2-subgroup $G_k<A_{2^k}$ is $Syl_2 (A_{2^k})$. Consequently, we construct a generating set, which contains $k$ elements, that is less than in [@Iv]. The structure of Sylow 2-subgroup of $A_{2^k}$ is the following: $\underset{i=1}{\overset{k-1}{\mathop{\wr }}}\,{{C}_{2}} \ltimes \prod_{i=1}^{2^{k-1}-1} C_2 $, where we take $C_2$ as group of action on two elements and this action is faithful. It adjusts with construction of normalizer for $Syl_p(S_n)$ from [@Weisner], where it was said that $Syl_2(A_{2^l})$ is self-normalized in $S_{2^l}$. Let us call the index of automorphism $\beta$ on $X^l$ as a number of active v.p. of $\beta$ on $X^l$. Define an **element of type** `T` as an automorphism ${{\tau }_{{{i}_{0}},...,{{i}_{{{2}^{k-1}}}};{{j}_{{{2}^{k-1}}}},...,{{j}_{{2}^k}}}}$, that has even index at ${{X}^{k-1}}$ and it has exactly $m_1$ active states, $ m_1\equiv 1 (mod 2)$, in vertexes ${{v}_{k-1,j}}, $ with number $1 \leq j \leq 2^{k-2}$ and $m_2, \, m_2\equiv 1 (mod 2)$ active states in vertices ${{v}_{k-1,l}}$, ${2^{k-2}} < l \leq {2^{k-1}}$. Set of such elements is denoted by `T`. A combined element is such an automorphism $ {{\beta }_{1,{{i}_{1}};2,{{i}_{2}};...;k-1,{{i}_{k-1}}; \tilde{\tau }}}$, that it’s restriction ${{\beta}_{1,{{i}_{1}};2,{{i}_{2}};...;k-1,{{i}_{k-1}};\tilde{\tau }}}\left|_{{{X}_{k-1}}} \right.$ coincide with one of elements that can be generated by ${{S}_{\alpha}}$ and $Rist_{<{{\beta}_{1,{{i}_{1}};2,{{i}_{2}};...;k-1,{{i}_{k-1}}; \tilde{ \tau} }}>}(k-1) =\left\langle \tau' \right\rangle $ [@Ne] where $\tau' \in $`T`. Set of such elements is denoted by `C`. In other word elements $g \in$`C` on level ${{X}^{k-1}}$ have such structure as element and generator of type `T`. As well ${{\tau }_{{{i}_{0}},...,{{i}_{{{2}^{k-1}}}};{{j}_{{{2}^{k-1}}}},...,{{j}_{{{2}^{k}}}}}}\in S{{t}_{Aut{{X}^{k}}}}(k-1)$. The minimum size of a generating set $S$ of G we denote by rk$G$ and call the rank of $G$ [@Bog]. By distance between vertices we shall understand usual distance at graph between its vertexes. By distance of automorphism $g$ (element) we shall understand maximal distance between two vertexes with active states of $g$. \[Lemma about keeping of distance\] A vertices permutations on ${X}^{k}$ that has distance ${{d}_{0}}$ can not be generated by vertex permutations with distance ${{d}_{1}}$ such that ${{d}_{1}}<{{d}_{0}}$. The element ${{\tau }_{ij}}$ with distance $\rho ({{\tau }_{ij}})={{d}_{0}},\,\,\,{{d}_{0}}<{{d}_{1}}$ can be mapped by automorphic mapping only in automorphism with distance ${{d}_{0}}$ because automorphic mapping keep incidence relation and so it possess property of isometry. Also multiplication of portrait (labeled graph) of automorphism ${{\tau }_{ij}}:\,\,\,\,\,\rho ({{\tau }_{ij}})={{d}_{1}}$ give us portrait of element with distance no greater than ${{d}_{1}}$, it follows from properties of group operation. For instance $\tau_{1i} \tau_{1j}=\tau_{ii}$, where $i, j > 2^{k-2}$ $\rho ({{\tau }_{1i}})=\rho ({{\tau }_{1j}})=2k-2$ but $\rho ({{\tau }_{ij}})< 2k-2$. \[about transposition\] An arbitrary automorphism $\tau{'} \in$`T` (or in particular $\tau $) can be generated only with using odd number of automorphisms from `C` or `T`. Let us assume that there is no such element ${{\tau }_{ij}}$ which has distance $2k-2$ then accord to Lemma \[Lemma about keeping of distance\] it is imposable to generate are pair of transpositions $\tau{'} $ with distance $\rho ({{\tau }_{ij}})=2k$ since such transpositions can be generated only by ${{\tau }_{ij}}$ that described in the conditions of this Lemma: $i\le {{2}^{k-2}},\,\,\,j>{{2}^{k-2}}$. Combined element can be decomposed in product $ \tau \dot {\beta }_{{i}_{l}} = {{\beta }_{{{i}_{l}};\tau }} $ so we can express by using $\tau$ or using a product where odd number an elements from `T` or `C`. If we consider product $P$ of even number elements from `T` then automorphism $P$ has even number of active states in vertexes ${{v}_{k-1,i}}$ with number $ i \leq 2^{k-2}$ so $P$ does not satisfy the definition of generator of type `T`. \[About generating distance\] Any element of type `T` can not be generated by ${{\tau }_{ij}}\in Aut {{v}_{1,1}}{{X}^{[k-1]}}$ and ${{\tau }_{ml}}\in Aut {{v}_{1,2}}{{X}^{[k-1]}}$. The same corollary is true for a combined element. It can be obtained from the Lemma \[Lemma about keeping of distance\] because such ${\tau_{ij}}\in Aut {{v}_{1,1}}{{X}^{[k-1]}}$ has distance less then $2k-2$ so it does not satisfy conditions of the Lemma \[Lemma about keeping of distance\]. I.e. $\tau{'}$ can not be generated by vertices automorphisms with distance between vertices less than $2k-2$ such distance has only automorphisms of type `T` and `C`. But elements from $ Aut {{v}_{1,1}}{{X}^{[k-1]}}$ do not belongs to type `T` or `C`. \[About not closed set of element of type T\] Sets of elements of types `T`, `C` are not closed by multiplication and raising to even power. Let $\varrho, \rho \in$ `T` (or `C`) and $\varrho, \rho = \eta$. The numbers of active states from $\varrho$ and $\rho$ in tuple of vertices $v_{k-1, i}$, $1 \leq i \leq 2^{k-2}$ sums by $mod 2$, numbers of active states from $\varrho$ and $\rho$ in vertices on vertices $v_{k-1, i}$, $2^{k-2} < i \leq 2^{k-1}$ sums by $mod 2$ too. Thus $\eta$ has even numbers of active states on these tuples. Hence $ RiS{{t}_{\left\langle \eta \right\rangle }}(k-1)$ doesn’t contain elements of type $\tau $ so $\eta \notin $`T`. Really if we raise the element ${{\beta }_{1,{{i}_{1}};2,{{i}_{2}};...;k-1,{{i}_{k-1}};\tau }}\in $`T` to even power or we evaluate a product of even number of multipliers from `C` corteges ${{\mu }_{0}}$ and ${{\mu }_{1}}$ permutes with whole subtrees ${{v}_{1,1}}{{X}^{[k-1]}}$ and ${{v}_{1,2}}{{X}^{[k-1]}}$, then we get an element $g$ with even indices of ${{X}^{k}}$ in $v_{1,1}{{X^{k-1}}}$ and $v_{1,2}{{X^{k-1}}}$. Thus $g \notin $`T`. Consequently elements of `C` do not form a group, and the set `T` as a subset of `C` is not closed too. Let ${S^{'}_{\alpha }}=\left\langle {{\alpha }_{0}},\,{{\alpha }_{1}},...,{{\alpha }_{k-2}} \right\rangle $ so as well known [@Gr] $\left\langle S_{\alpha }^{'} \right\rangle =Aut{{X}^{[k-1]}}$. The cardinality of a generating set $S$ is denoted by $\mid {{S}} \mid$ so $\mid {{S}^{'}_{\alpha}} \mid=k-1$. Recall that $rk\left( G \right)$ is the rank of a group $G$ [@Bog]. Let $S_{\beta }={{S}_{\alpha }^{'}}\cup \tau_{i...j} $, where $\tau_{i...j} \in$`T` and ${{S}_{\beta }^{'}}$ is generating system which contains combine elements, $\mid {{S}^{'}_{\beta}} \mid = k$. It’s known that $rk(Aut{{X}^{[k-1]}})= k-1 $ and $\mid {{S}^{'}_{\alpha}} \mid = k-1$ [@Gr]. So if we complete ${S}^{'}_{\alpha}$ by $\tau$ or element of type `T` we obtain set ${{S}_{\beta }}$ such that ${{G}_{k}}\simeq \left\langle {{S}_{\beta }} \right\rangle $ and $ |{S}_{\beta }|= k$. Hence to construct combined element $\beta $ we multiply generator ${{\alpha}_{i}}$ of ${{S}^{'}_{\alpha}}$ or arbitrary element that can be express from ${{S}^{'}_{\alpha}}$ on the element of type `T`, i.e., we take $\tau' \cdot {{\beta }_{i}}$ instead of ${{\beta }_{i}}$ and denote it ${\beta }_{i; \tau'}$. It’s equivalent that $Rist_{{\beta }_{i; \tau'} }(k-1)=\left\langle \tau' \right\rangle $, where $\tau' \in$`T`. Let us assume that $ S_{\beta }^{'}$ has a cardinality $k-1$. If in this case $ S_{\beta }^{'}$ is generating system again, then element $\tau $ can be expressed from it. There exist too ways to express element of type `T` from ${{S}_{\beta }^{'}}$. To express element of type `T` from ${{S}_{\beta }^{'}}$ we can use a word ${{\beta }_{i,\tau }}\beta _{i}^{-1}=\tau $ but if ${{\beta }_{i,\tau }}\in {{S}_{\beta }^{'}}$ then ${{\beta }_{{{i}}}}\notin S_{\beta }^{'}$ in contrary case $\mid S_{\beta }^{'} \mid = k$. So we can not express word ${\beta }_{i,\tau }\beta _{i}^{-1}\left| _{{X}^{[k-1]}} \right.=e$ to get ${{\beta }_{i,\tau }}\beta _{i}^{-1}=\tau $. For this goal we have to find relation in a group that is a restriction of the group ${{G}_{k}}$ on ${{X}^{[k-1]}}$. We have to take in consideration that $ {{G}_{k}}\left| _{{{X}^{[k-1]}}} \right.={{B}_{k-1}} $. Really in wreath product $\wr _{j=1}^{k}{{\mathcal{C}}^{(j)}_{2}}\simeq B_{k-1}$ holds a constitutive relations $\alpha_{i}^{2^m}=e$ and $\left[ \alpha _{m}^{i}{{\alpha }_{{{i}_{n}} }}\alpha_{m}^{-i},\,\alpha_{m}^{j}{\alpha }_{{{i}_{k}} }\alpha_{m}^{-j} \right]=e,\,\,\,i\ne j$, where $\alpha_{m} \in {{S}^{'}_{^{\alpha }}}$, $\alpha_{i_k} \in {{S}^{'}_{^{\alpha }}}$ are generators of factors of $\wr _{j=1}^{k}{{\mathcal{C}}^{(j)}_{2}}$ ($m<n$, $m<k$) [@Sk; @DrSku]. Such relations are a words $\left[ \beta _{m}^{i}{{\beta }_{{{i}_{n}},{\pi} }}\beta_{m}^{-i},\,\beta _{m}^{j}{\beta }_{{{i}_{k}}, {\pi} }\beta_{m}^{-j} \right],\,\,\,i\ne j$ or $\beta_{i}^{2^m}=e$, ${\beta }_{{{i}_{n}}}$, ${\beta_m},\,\,{{\beta }_{{{i}_{k}}}},\, {{\beta }_{{{i}_{n}}, \pi }}$ are generators of $S^{*}_{\beta}(k-1)$, that could be an automorphism $ \theta$. But $\left[ \beta _{m}^{i}{{\beta }_{{{i}_{n}},{\pi} }}\beta_{m}^{-i},\,\beta _{m}^{j}{\beta }_{{{i}_{k}}, {\pi} }\beta_{m}^{-j} \right],\,\,\,i\ne j$ does not belongs to `T` because this word has logarithm 0 by every element [@K]. According to Lemma \[about transposition\] and Lemma \[About not closed set of element of type T\] product of even number element of type C doesn’t equal to element of C or T. Let us assume an existence of generating set of cardinality $k-1$ for $Syl_2(A_{2^k})$ that in general case has form $S_{\beta }^{*}(k-1)=\left\langle \mathop {\beta}_0, \mathop {\beta}_{0;1,(i_{11},i_{12}); \pi_1 } , ... , \beta_{0;...;k-1,(i_{k-11},...,i_{k-1j});\pi_{k-1} } \right\rangle, \\ 0<i<k-1, j \leq 2^{i}, \pi_1 \in T $, where $\pi_i$ is the cortege of vertices from $X^{k-1}$ with non trivial v.p. which realize permutation with distance $2(k-1)$. In other word if element $\pi_i \in$ `T` then $\beta_{0;...;i,(i_{i1},...,i_{ij}); \pi_i } = \pi_i \alpha_{0;...;i,(i_{i1},...,i_{ij}) }$. From it follows $\mathop {\beta}_{0;1,(i_{11},i_{12}); ... ;{m},(i_{m1},...,i_{mj};\pi_m) }\mid_{X^{[k-1]}} = \mathop {\alpha}_{0;1,(i_{11},i_{12}); ... ,{m},(i_{m1},...,i_{mj}) } \in S^{'}_{\alpha}$. Note, that automorphisms from set $S^{*}_{\beta}(k-1)$ generate on truncated rooted tree [@Ne] $X^{[k-1]}$ group $\langle S^{*}_{\beta}(k-1) \rangle \mid_{X^{[k-1]}} \simeq \wr^{k-1}_{i=1} C_2^{(i)}\simeq B_{k}$. \[Th about general relation\] Any element of type `T` can not be expressed by elements of $S^{*}_{\beta}(k-1)$. It is necessary to express an automorphism $ \theta$ of type `T` express such automorphism which has zero indexes of $X^0, ... , X^{k-2}$, this conclusion follows from structure of elements from `T`. It means that word $w$ from letters of $S^{*}_{\beta}(k-1)$ such that $w = \theta$ is trivial in group $B_k$, that arise on restriction of $\langle S^{*}_{\beta}(k-1)\rangle$ on $X^{[k-1]}$, as well restriction $ G_k \mid_{X^{[k-1]}}\simeq B_k$. Every relation from $B_{k}$ can be expressed as a product of words from the normal closure $R^{B_{k}}$ of the set of constitutive relations of the group $B_{k}$ [@Bog]. But defined relations $r_i$ of $B_{k}$ have form of commutators [@DrSku; @Sk] so the number of inclusions of every multiplier is even and as follows from lemma \[About not closed set of element of type T\] that $r_i \mid_{[X^k]} \notin$ `T`. Really in wreath product $\wr _{j=1}^{k}{{\mathcal{C}}^{(j)}_{2}}\simeq B_{k-1}$ holds a constitutive relations $\alpha_{i}^{2^m}=e$ and $\left[ \alpha _{m}^{i}{{\alpha }_{{{i}_{n}} }}\alpha_{m}^{-i},\,\alpha_{m}^{j}{\alpha }_{{{i}_{k}} }\alpha_{m}^{-j} \right]=e,\,\,\,i\ne j$, where $n,k,m$ are number of groups in $\wr _{j=1}^{k}{{\mathcal{C}}^{(j)}_{l}}$ ($m<n$, $m<k$) [@Sk; @DrSku], where ${\alpha }_{{{i}_{n}}}$, ${\alpha }_{{{i}_{n}}}$, ${{\alpha }_{m}},\,\,{{\alpha }_{{{i}_{k}}}}$ are generators of the $B_{k-1}$ from ${{S}^{'}_{^{\alpha }}}$. So it give us a word $\left[ \beta _{m}^{i}{{\beta }_{{{i}_{n}},{\pi} }}\beta_{m}^{-i},\,\beta _{m}^{j}{\beta }_{{{i}_{k}}, {\pi} }\beta_{m}^{-j} \right],\,\,\,i\ne j$ that could be an automorphism $ \theta$ but does not belongs to `T` because the word has structure of commutator or belongs to normal closure $R^{B_{k-1}}$ so has logarithm 0 by every element, where ${\beta }_{{{i}_{n}}}$, ${\beta_m},\,\,{{\beta }_{{{i}_{k}}}},\, {{\beta }_{{{i}_{n}}, \pi }}$ are generators of $G_k$ from $S^{*}_{\beta}(k-1)$. Let ${{\beta}_{1,1,(i_{11},i_{12}), \ldots ,i,(i_{i1},...,i_{i2^i});\pi_i}}$ be an arbitrary element of type `C`, where $\pi_i$ is cortege of vertexes from $X^{k-1}$ having non trivial states which realize permutation with distance $2(k-1)$. The case where $\theta = \beta_{{k-1}; {\pi_i}} \in$ `T` can be expressed by multiplying arbitrary\ ${{\beta}_{1,(i_{11,12}), \ldots ,i,(i_{i1},...,i_{i2^i});\pi_i}}$ on $\beta_{1,(i_{11,12}), \ldots ,i,(i_{i1},...,i_{i2^i})}^{-1}$ means that such set has size more then $k-1$ what is contradiction. Really if ${{\beta}_{1,(i_{11,12}), \ldots ,i,(i_{i1},...,i_{i2^i})}}$ and ${{\beta}_{1,(i_{11,12}), \ldots ,i,(i_{i1},...,i_{i2^i});\pi_i}} \, \in S_{\beta}^{*}$ then ${\beta}_{1,(i_{11,12}), \ldots ,i,(i_{i1},...,i_{i2^i})} |_{X^{[k-1]}} = {{\beta}_{1,(i_{11,12}), \ldots ,i,(i_{i1},...,i_{i2^i});\pi_i}}|_{X^{[k-1]}} $ i.e. these elements are mutually inverse at restriction on $X^{[k-1]}$, but it means that in restriction of $S^{*}_{\beta}(k-1)$ to $X^{[k-1]}$ that corresponds to the generating set $S^{'}_{\alpha}(k-1)$ for $AutX^{[k-1]} \simeq B_{k-1} $ we use two equal generators. So it has at least $k$ generators, because $rk(Aut X^{[k-1]})=k-1$ according to lemma \[B\_k-1\]. The subcase of this case where ${{\beta}^{-1}_{1,(i_{11},i_{12}), \ldots ,l,(i_{l1},...,i_{l2^i})}}$ can be expressed from $S^{*}_{\beta}(k-1)$ as a product of its generators has the same conclusion. Really if we can generate arbitrary element from $B_{k-1}$ by generators from $S^{*}_{\beta}(k-1)$ then $k-1$ generators is contained in set $S^{*}_{\beta}(k-1)$ but we have a further ${{\beta}_{1,(i_{11},i_{12}), \ldots ,i,(i_{i1},...,i_{i2^i}); {\pi}}}$. In other words if arbitrary element ${{\beta}_{1,(i_{11},i_{12}), \ldots ,l,(i_{l1},...,i_{l2^i})}}$ of the $B_{k-1}$ does not contains in $S^{*}_{\beta}(k-1)$ but can be expressed from it then $S^{*}_{\beta}(k-1)$ has at least $k-1$ elements exclusive of ${{\beta}_{1,(i_{11},i_{12}), \ldots ,i,(i_{i1},...,i_{i2^i}); {\pi}}}$ then $S^{*}_{\beta}(k-1) \geq k $. \[generating pair\] A necessary and sufficient condition of expressing an element $\tau $ from $S_{\beta }^{'}$ is existing of pair: ${{\beta }_{{{i}_{m}};\tau }},\,\,\,\,{{\beta }_{{{i}_{_{m}}}}}$ in $\langle S_{\beta}^{'} \rangle$. So size of a generating set of $G_k$ which contains $S_{\beta }^{'}$ is at least $k$. Proof can be obtained from Theorem \[Th about general relation\] and Lemma \[Th about general relation\] from which we have that element of type `T` cannot be expressed from $\left[ \beta _{m}^{i}{{\beta }_{{{i}_{n}}.\tau }}\beta _{m}^{-i},\,\beta _{m}^{j}{{\beta }_{{{i}_{k}}.\tau }}\beta _{m}^{-j} \right]=e,\,\,\,i\ne j$ because such word has even number of elements from `C`. Sufficient condition follows from formula ${{\beta }_{{{i}_{m}};\tau }}\beta _{{{i}_{_{m}}}}^{-1}=\tau $. \[rk\] A generating set of ${{G}_{k}}$ contains $S_{\alpha}^{'}$ and has at least $k-1$ generators. The subgroup ${{B}_{k-1}}<{{G}_{k}}$ is isomorphic to $Aut{{X}^{k-1}}$ that has a minimal set of generators of $k-1$ elements [@Gr]. Moreover, the subgroup ${{B}_{k-1}} \simeq {}^{{{G}_{k}}}/{}_{{{W}_{k-1}}}$, because $G_{k} \simeq B_{k-1}\ltimes {W}_{k-1} $, where ${W}_{k-1} \vartriangleright G_k$. As it is well known that if $H\lhd G$ then $\text{rk}(G)\ge \text{rk}(H)$, because all generators of $G_{k}$ may belongs to different quotient classes [@Magn]. As a corollary of last lemma and Theorem \[Th about general relation\] we see that generating set of size $k-1$ does not exist. Let us sharpen and reformulate following theorem which is in [@Meld]. \[form of comm\] An element of form $(r_1, \ldots, r_{p-1}, r_p) \in W' = (C_p \wr B)'$ iff product of all $r_i$ (in any order) belongs to $B'$ and $cw(B)=1$. We deduce here a special form of commutator elements. If we multiply elements having form of a tuple $(r_1, \ldots, r_{p-1}, r_p)$, where $r_i={{h}_{i}}{{g}_{a(i)}}h_{ab(i)}^{-1}g_{ab{{a}^{-1}}(i)}^{-1}$, $h, \, g \in B$ and $a,b \in C_p$, then we obtain a product $$\label{Meld} \stackrel{p}{ \underset{\text{\it i=1}} \prod} r_i = \prod\limits_{i=1}^{p}{{{h}_{i}}{{g}_{a(i)}}h_{ab(i)}^{-1}g_{ab{{a}^{-1}}(i)}^{-1} \in B'}.$$ Recall that $ \stackrel{k}{ \underset{\text{\it i=1}}{\wr }}C_p \simeq Syl_p S_{p^k} $. \[comm B\_k old\] An element $g \in B_k \simeq \stackrel{k}{ \underset{\text{\it i=1}}{\wr }}C_p $ belongs to commutator subgroup $B'_k \simeq (\stackrel{k}{ \underset{\text{\it i=1}}{\wr }}C_p)'$ iff $g$ has all possible even indexes on ${{X}^{l}},\,\,l<k$. Let us prove the ampleness by induction by a number of level $l$. We first show that our statement for base of the induction is true. Actually, if $\alpha, \beta \in B_{0}$ then $(\alpha \beta {{\alpha }^{-1}}) \beta^{-1}$ determine a trivial v.p. on $X^{0}$. If $\alpha, \beta\in B_{1}$ and $\beta$ has an odd index on $X^1$, then $(\alpha \beta {{\alpha }^{-1}})$ and $\beta^{-1}$ have the same index on $X^1$. Consequently, in this case an index of the product $(\alpha \beta {{\alpha }^{-1}}) \beta^{-1}$ can be 0 or 2. Case where $\alpha, \beta \in B_{1}$ and has even index on $X^1$, needs no proof, because the product and the sum of even numbers is an even number. To finish the proof it suffices to assume that for $B_{l-1}$ statement holds and prove that it holds for $B_l$. Let $\alpha, \beta$ are an arbitrary automorphisms from $Aut X^{[k]}$ and $\beta$ has index $x$ on $X^{l}, \, l<k$, where $0 \leq x \leq 2^{l} $. A conjugation of an automorphism $\beta $ by arbitrary $\alpha \in Aut{{X}^{[k]}}$ gives us arbitrary permutations of $X^l$ where $\beta $ has active v.p. Thus following product $(\alpha \beta {{\alpha }^{-1}}) \beta ^{-1}$ admits all possible even indexes on $X^l, l<k$ from 0 to $2x$. In addition $[\alpha, \beta]$ can has arbitrary assignment (arrangement) of v.p. on $X^l$. Let us present $B_k$ as $B_k=B_l \wr B_{k-l}$, so elements $\alpha, \beta$ can be presented in form of wreath recursion $\alpha = (h_{1},...,h_{2^l })\pi_1, \, \beta = (f_{1},...,f_{2^l })\pi_2$, $h_{i}, f_{i} \in B_{k-l} ,\ 0<i \leq 2^l$ and $h_{i}, f_{j}$ corresponds to sections of automorphism in vertices of $X^l$ of isomorphic subgroup to $B_l$ in $Aut X^{[k]}$. Actually, the parity of this index are formed independently of the action of $Aut X^{[l]}$ on $X^l$. So this index forms as a result of multiplying of elements of commutator presented as wreath recursion $(\alpha \beta \alpha^{-1}) \cdot \beta ^{-1} = (h_{1},...,h_{2^l})\pi_1 \cdot (f_{1},...,f_{2^l })\pi_2= (h_{1},...,h_{2^l}) (f_{\pi_1(1)},...,f_{\pi_1(2^l)})\pi_1 \pi_2 $, where $h_{i}, f_{j} \in {B}_{k-l}$, $l<k$ and besides automorphisms corresponding to $h_{i}$ are $x$ automorphisms which has active v.p. on $X^l$. Analogous automorphisms $h_{i}$ has number of active v.p. equal to $x$. As a result of multiplication we have automorphism with index $2i:$ $0 \leq 2i \leq 2x$. Consequently, commutator $[\alpha, \beta]$ has arbitrary even indexes on $X^m$, $m<l$ and we showed by induction that it has even index on $X^l$. Let us prove the necessity by induction by number of level $l$. Let $g \in B_k$ and $g$ has all even indexes on $X^j$ $0 \leq j < k$ we need to show that $g \in B'_k$. According to condition of this Lemma $g_1 g_2$ has even indexes. An element $g$ has form $g=(g_1, g_2)$, where $g_1, g_2 \in B_{k-1}$, and products $g_1 g_2 = h, g_2 g_1= h' \in B'_{k-1}$ because $h, h' \in B_{k-1}$ and for $B_{k-1}$ induction assumption holds. Therefore all products of form $g_1 g_2$ indicated in formula \[Meld\] belongs to $B'_{k-1}$. Hence, from Theorem \[form of comm\] follows that $g= (g_1, g_2 ) \in B'_k$. An even easier Proposition, that needs no proof, is the following. \[B\_k\_criteria\] An element $(g_1, g_2)\sigma^i$, $i \in \{0, 1\} $ of wreath power $\stackrel{k}{ \underset{\text{\it i=1}}{\wr } }C_2 $ belongs to its subgroup $ G_k$, iff $g_1g_2 \in G_{k-1}$. \[B’\_k and B\^2\_k\] If $g $ is an element of wreath power $\stackrel{k}{ \underset{\text{\it i=1}}{\wr } }C_2 \simeq B_k $ then $g^2 \in B'_{k}$. As it was proved in Lemma \[comm B\_k old\] commutator $[\alpha, \beta]$ from $B_k$ has arbitrary even indexes on $X^m$, $m<k$. Let us show that elements of $B_k^2$ have the same structure. Let $\alpha, \beta \in B_k$ an index of the automorphisms $\alpha^2 $, $(\alpha \beta)^2 $ and $\alpha, \beta \in G_k$ on $X^l, \, l<k-1$ are always even. In more detail the indexes of $\alpha^2 $, $(\alpha \beta)^2 $ and $\alpha^{-2} $ on $X^l$ are determined exceptionally by the parity of indexes of $\alpha $ and $\beta $ on ${{X}^{l}}$. Actually, the parity of this index are formed independently of the action of $Aut X^l$ on $X^l$. So this index forms as a result of multiplying of elements $\alpha \in B_k$ presented as wreath recursion $ \alpha^2 = (h_{1},...,h_{2^l})\pi_1 \cdot (f_{1},...,f_{2^l })\pi_1= (h_{1},...,h_{2^l}) (f_{\pi_1(1)},...,f_{\pi_1(2^l)})\pi_1 \pi_1 $, where $h_{i}, f_{j} \in {B}_{k-l}, \, \pi_1 \in B_l$, $l<k$ and besides automorphisms corresponding to $h_{i}$ are $x$ automorphisms which has active v.p. on $X^l$. Analogous automorphisms $h_{i}$ has number of active v.p. equal to $x$. As a result of multiplication we have automorphism with index $2i:$ $0 \leq 2i \leq 2x$. Since $g^2 $ admits only an even index on $X^l$ of $Aut X^{[k]}$, $0<l<k$, then $g^2 \in B'_{k}$ according to lemma \[comm B\_k old\] about structure of a commutator subgroup. \[L\_k\_comm\_criteria\] An element $(g_1, g_2)\sigma^i \in G_k'$ iff $g_1,g_2 \in G_{k-1}$ and $g_1g_2\in B_{k-1}'$. Indeed, if $(g_1, g_2) \in G_k'$ then indexes of $g_1$ and $g_2$ on $X^{k-1}$ are even according to Lemma \[comm\] thus, $g_1,g_2 \in G_{k-1}$. A sum of indexes of $g_1$ and $g_2$ on $X^{l}$, $l<k-1$ are even according to Lemma \[comm\] too, so index of product $g_1 g_2$ on $X^{l}$ is even. Thus, $g_1g_2\in B_{k-1}'$. Let us prove the sufficiency via Lemma \[comm\]. Wise versa, if $g_1,g_2 \in G_{k-1}$ then indexes of these automorphisms on $X^{k-2}$ of subtrees $v_{11}X^{[k-1]}$ and $v_{12}X^{[k-1]}$ are even as elements from $ G_k'$ have. The product $g_1g_2$ belongs to $B_{k-1}'$ by condition of this Lemma so sum of indexes of $g_1, g_2$ on any level $X^l$, $0 \leq l<k-1 $ is even. Thus, the characteristic properties of $G_k'$ described in this Lemma \[comm\] holds. Let $X_1=\{v_{k-1,1}, v_{k-1,2},..., v_{k-1,2^{k-2}} \} $ and $X_2=\{v_{k-1,2^{k-2}+1}, ..., v_{k-1,2^{k-1}} \}$. Let group $Sy{{l}_{2}}{{A}_{{{2}^{k}}}}$ acts on $X^{[k]}$. \[comm\] An element $g$ belongs to $G_k' \simeq (Syl_2{A_{2^k}})'$ iff $g$ is arbitrary element from $G_k$ which has all even indexes. Let us prove the ampleness by induction by a number of level $l$. Recall that any authomorphism $\theta \in Syl_2 A_n$ has an even index on $X^{k-1}$ so number parities of active v. p. on ${{X}_{1}}$ and on ${{X}_{2}}$ are the same. Conjugation by automorphism $\alpha$ from $Aut{{v}_{11}}{{X}^{\left[ k-1 \right]}}$ of automorphism $\theta $, that has some number $x:$ $1 \leq x \leq 2^{k-2}$ of active v. p. on ${{X}_{1}}$ does not change $x$. Also automorphism $\theta^{-1} $ has the same number $x$ of v. p. on $X_{k-1}$ as $\theta $ has. If $\alpha$ from $Aut{{v}_{11}}{{X}^{\left[ k-1 \right]}}$ and $ \alpha \notin Aut{{X}^{\left[ k \right]}}$ then conjugation $(\alpha \theta {{\alpha }^{-1}})$ permutes vertices only inside $X_1$ ($X_2$). Thus, ${\alpha }\theta {\alpha^{-1} }$ and $\theta$ have the same parities of number of active v.p. on $X_1$ ($X_2$). Hence, a product ${\alpha }\theta {\alpha^{-1} } \theta^{-1}$ has an even number of active v.p. on $X_1$ ($X_2$) in this case. More over a coordinate-wise sum by `mod2` of active v. p. from $(\alpha \theta {{\alpha }^{-1}})$ and $\theta^{-1}$ on $X_1$ ($X_2$) is even and equal to $y:$ $0 \leq y \leq 2x$. If conjugation by $\alpha$ permutes sets $X_1$ and $X_2$ then there are coordinate-wise sums of no trivial v.p. from $\alpha \theta \alpha^{-1} \theta^{-1}$ on $X_1$ (analogously on $X_2$) have form:\ $({{s}_{k-1,1}}(\alpha \theta {{\alpha }^{-1}}),..., {{s}_{k-1, 2^{k-2}}}(\alpha \theta {{\alpha }^{-1}}) )\oplus ({{s}_ {k-1,1}}(\theta^{-1}), ..., {{s}_{k-1,{{2}^{k-2}}}}(\theta^{-1} ))$. This sum has even number of v.p. on $X_1$ and $X_2$ because $(\alpha \theta {{\alpha }^{-1}})$ and ${{\theta }^{-1}}$ have a same parity of no trivial v.p. on $X_1$ ($X_2$). Hence, $(\alpha \theta {{\alpha }^{-1}}){{\theta }^{-1}}$ has even number of v.p. on ${{X}_{1}}$ as well as on ${{X}_{2}}$. An authomorphism $\theta $ from $G_k$ was arbitrary so number of active v.p. $x$ on $X_1$ is arbitrary. And ${\alpha }$ is arbitrary from $AutX^{[k-1]}$ so vertices can be permuted in such way that the commutator $[{\alpha },\theta]$ has arbitrary even number $y$ of active v.p. on $X_1$, $0 \leq y \leq 2x$. A conjugation of an automorphism $\theta $ having arbitrary index $x$, $1 \leq x \leq 2^{l}$ on ${{X}^{l}}$ by different $\alpha \in Aut{{X}^{[k]}}$ gives us all permutations of active v.p. that $\theta $ has on ${{X}^{l}}$. So multiplication $(\alpha \theta {{\alpha }^{-1}})\theta $ generates a commutator having index $y$ equal to coordinate-wise sum by $mod 2$ of no trivial v.p. from vectors $({{s}_{l1}}(\alpha \theta {{\alpha }^{-1}}),{{s}_{l}}_{2}(\alpha \theta {{\alpha }^{-1}}),...,{{s}_{l{{2}^{l}}}}(\alpha \theta {{\alpha }^{-1}}))\oplus ({{s}_{l1}}(\theta ),{{s}_{l}}_{2}(\theta ),...,{{s}_{l{{2}^{l}}}}(\theta ))$ on ${{X}^{l}}$. A indexes parities of $\alpha \theta {{\alpha }^{-1}}$ and ${{\theta }^{-1}}$ are same so their sum by $mod 2$ are even. Choosing $\theta $ we can choose an arbitrary index $x$ of $\theta $ also we can choose arbitrary $\alpha $ to make a permutation of active v.p. on ${{X}^{l}}$. Thus, we obtain an element with arbitrary even index on ${{X}^{l}}$ and arbitrary location of active v.p. on ${{X}^{l}}$. Check that property of number parity of v.p. on ${{X}_{1}}$ and on ${{X}_{2}}$ is closed with respect to conjugation. We know that numbers of active v. p. on ${{X}_{1}}$ as well as on ${{X}_{2}}$ have the same parities. So action by conjugation only can permutes it, hence, we again get the same structure of element. Conjugation by automorphism $\alpha $ from $Aut{{v}_{11}}{{X}^{\left[ k-1 \right]}}$ automorphism $\theta $, that has odd number of active v. p. on ${{X}_{1}}$ does not change its parity. Choosing the $\theta $ we can choose arbitrary index $x$ of $\theta $ on ${{X}^{k-1}}$ and number of active v.p. on ${{X}_{1}}$ and ${{X}_{2}}$ also we can choose arbitrary $\alpha $ to make a permutation active v.p. on ${{X}_{1}}$ and ${{X}_{2}}$. Thus, we can generate all possible elements from a commutant. Also this result follows from Lemmas \[L\_k\_comm\_criteria\] and \[comm B\_k old\]. Let us check that the set of all commutators $K$ from $Syl_2 A_{2^k}$ is closed with respect to multiplication of commutators. Let $\kappa_1, \kappa_2 \in K$ then $\kappa_1 \kappa_2$ has an even index on $X^l$, $l<k-1$ because coordinate-wise sum $({{s}_{l,1}}(\kappa_1),..., {{s}_{k-1, 2^l}}(\kappa_1) )\oplus ({{s}_ {l,\kappa_1(1)}}(\kappa_2), ..., {{s}_{l,\kappa_1({{2}^{l}})}}(\kappa_2 ))$. of two $2^l$-tuples of v.p. with an even number of no trivial coordinate has even number of such coordinate. Note that conjugation of $\kappa $ can permute sets ${{X}_{1}}$ and ${{X}_{2}}$ so parities of $x_1$ and $X_2$ coincide. It is obviously index of $\alpha \kappa \alpha^{-1}$ is even as well as index of $\kappa $. Check that a set $K$ is a set closed with respect to conjugation. Let $\kappa \in K$, then $\alpha \kappa {{\alpha }^{-1}}$ also belongs to $K$, it is so because conjugation does not change index of an automorphism on a level. Conjugation only permutes vertices on level because elements of $Aut{{X}^{\left[ l-1 \right]}}$ acts on vertices of ${{X}^{l}}$. But as it was proved above elements of $K$ have all possible indexes on ${{X}^{l}}$, so as a result of conjugation $\alpha \kappa {{\alpha }^{-1}}$ we obtain an element from $K$. Check that the set of commutators is closed with respect to multiplication of commutators. Let $\kappa_1, \kappa_2 $ be an arbitrary commutators of $G_k$. The parity of the number of vertex permutations on $X^l$ in the product $\kappa_1 \kappa_2 $ is determined exceptionally by the parity of the numbers of active v.p. on ${{X}^{l}}$ in $\kappa_1$ and $\kappa_2$ (independently from the action of v.p. from the higher levels). Thus $\kappa_1 \kappa_2 $ has an even index on $X^l$. Hence, normal closure of the set $K$ coincides with $K$. \[comm\] Frattini subgroup $ \phi(G_k)= {{G_k}^{2}}\cdot [G_k,G_k]= {{G_k}^{2}} $ acts by all even permutations on ${{X}^{l}},\,\,\,0\le l<k-1$ and by all even permutations on $ X^{k}$ except for those from `T`. Index of the automorphism $\alpha^2 $, $\alpha \in {{S}_{\beta }}$ on $X^l$ is always even. Really the parity of the number of vertex permutations at $X^l$ in the product $({\alpha }_i {\alpha }_j)^2$, ${\alpha }_i, {\alpha }_j \in S_{\alpha }$, $i,j<k$ is determined exceptionally by the parity of the numbers of active states of v.p. on ${{X}^{l}}$ in $\alpha $ and $\beta $ (independently of the action of v.p. from the higher levels). On $X^{k-1}$ group $G^2$ contains all automorphisms of form $\tau_{1 i}, i\leq 2^{k-1}$ which can be generated in such way $({\alpha }_{k-2} \tau_{12})^2= \tau_{1234}$, $\tau_{12}\tau_{1234} = \tau_{34}$, $({\alpha }_{k-i} \tau_{12})^2= \tau_{1, 2, 1+2^{k-i}, 2+2^{k-i}}$ then $\tau_{1, 2, 1+2^{k-i}, 2+2^{k-i}} \tau_{12} =\tau_{1+2^{k-i}, 2+2^{k-i}}$. In such way we get set of form ${\tau_{12}, \tau_{23},, \tau_{34}, ... ,\tau_{2^{k-1}-1,2^{k-1}}}$. This set is the base for $W_{k-1}$. The parity of the number of vertex permutations at $X^l$ in the product ${\alpha }_i$ or ${\alpha }_i {\alpha }_j$, ${\alpha }_i,{\alpha }_j \in S_{\alpha }$) is determined exceptionally by the parity of the numbers of active v.p. on ${{X}^{l}}$ in $\alpha $ and $\beta $ (independently of the action of v.p. from the higher levels). Thus $[\alpha ,\beta ]=\alpha \beta {\alpha }^{-1}{\beta }^{-1}$ has an even number of v. p. at each level. Therefore, the commutators of the generators from ${{S}_{\alpha }}$ and elements from $G^2$ generate only the permutations with even number of v. p. at each ${{X}^{l}}$, ($0\le l\le k-2$). Let us consider ${{\left( {{\alpha }_{0}}{{\alpha }_{l}} \right)}^{2}}={{\beta }_{l({{1,2}^{l-1}}+1)}}$. Conjugation by the element ${{\beta }_{1(1,2)}}$ (or ${{\beta }_{i(1,2)}},\,\,0<i<l$) give us ability to express arbitrary coordinate $x:\,\,1\le x \leq {{2}^{l-1}}$ where $x=2^{k-1}-i$, i.e. from the element ${{\beta }_{l({{1,2}^{l-1}}+1)}}$ we can express ${{\beta }_{l(x{{,2}^{l-1}}+1)}}$. For instance $x={{2}^{j-1}}+1$, $j<l$: ${{\beta }_{l-j(1,2)}}{{\beta }_{l({{1,2}^{l-1}}+1)}}{{\beta }_{l-j(1,2)}}={{\beta }_{l({{2}^{j-1}+1},{{2}^{l-1}}+1)}}$. If $x={{2}^{l-j}}+2$ than to realize every shift on $x$ on set $X^l$ the element ${{\beta }_{l({{1,2}^{l-1}}+1)}}$ should to be conjugated by such elements ${{\beta }_{l-j(1,2)}} {{\beta }_{l-1(1,2)}}$. So in such way can be realized every ${{\beta }_{l(x{{,2}^{l-1}}+1)}}$ and analogously every ${{\beta }_{l({{2}^{l-1},y})}}$ and ${{\beta }_{l(x,y)}}$. Hence we can express from elements of $G^2$ every even number of active states of v.p. on $X^l$. Define the subgroup $G(l)<Aut{{X}^{[k]}}$, where $l< k$, as $Stab_{Aut{{X}^{[k]}}}(l)\left| _{{{X}^{l}}} \right.$. It is plain, that $G(l) \simeq ^{Stab_{Aut{{X}^{[k]}}}(l)} /_{ Stab_{Aut{{X}^{[k]}}}(l+1)}$ because $Stab_{Aut{{X}^{[k]}}}(l+1)$ is normal subgroup of finite index in $Aut{{X}^{[k]}}$ [@Ne]. Let us construct a homomorphism from $G(l)$ onto ${{C}_{2}}$ in the following way: $\varphi (\alpha )=\sum\limits_{i=1}^{{{2}^{l}}}{{{s}_{li}}(\alpha )}\bmod 2$. Note that $\varphi (\alpha \cdot \beta )=\varphi (\alpha )\circ \varphi (\beta )=(\sum\limits_{i=1}^{{{2}^{l}}}{{{s}_{li}}(\alpha )}+\sum\limits_{i=1}^{{{2}^{l}}}{{{s}_{li}}(\beta )})mod2$. Structure of subgroup $G_{k}^{2}{{G}_{k}}'\triangleleft \underset{1}{\overset{k}{\mathop{\wr }}}\,{{S}_{2}}\simeq Aut{{X}^{[k]}}$ can be described in next way. This subgroup contains the commutant ${G}_{k}'$. So it has on each ${{X}^{l}},\,\,\,0\le l<k-1$ all even indexes that can exists there. On $X^{k-1}$ it does not exist v.p. of type `T`, which has the distance $2k-2$, rest of even the indexes are present on ${X}^{k-1}$. It’s so, because the sets of elements of types `T` and `C` are not closed under operation of calculating the even power as it proved in Lemma \[About not closed set of element of type T\]. Thus, the squares of the elements don’t belong to `T` and `C` (because they have the distance, which is less than $2k-2$). This implies the following corollary. \[qoutient\] A quotient group ${}^{{{G}_{k}}}/{}_{G_{k}^{2}{{G}^{'}_{k}}}$ is isomorphic to $\underbrace{{{C}_{2}}\times {{C}_{2}}\times ...\times {{C}_{2}}}_{k}$. The proof is based on two facts $G_{k}^{2}{{G}^{'}_{k}}\simeq G_{k}^{2}\triangleleft {{G}_{k}}$ and $\left| G:G_{k}^{2}{{G}^{'}_{k}} \right|=2^k$. Construct a homomorphism from $G_k(l)$ onto ${{C}_{2}}$ in the following way: $\varphi (\alpha )=\sum\limits_{i=1}^{{{2}^{l}}}{{{s}_{li}}(\alpha )}\bmod 2$. Note that $\varphi (\alpha \cdot \beta )=\varphi (\alpha )\circ \varphi (\beta )=(\sum\limits_{i=1}^{{{2}^{l}}}{{{s}_{li}}(\alpha )}+\sum\limits_{i=1}^{{{2}^{l}}}{{{s}_{li}}(\beta )})mod2$, where $\alpha ,\,\,\beta \in Aut{{X}^{[n]}}$. Index of $\alpha \in G_{k}^{2}$ on ${{X}^{l}},\,l<k-1$ is even but index of $\beta \in {{G}_{k}}$ on ${{X}^{l}}$ can be both even and odd. Note that ${{G}_{k}}(l)$ is abelian group and $G_{k}^{2}(l)\trianglelefteq {{G}_{k}}$. Since words with equal logarithms to all bases [@K] belong to distinct cosets of the commutator, the subgroup $G_{k}^{2}(l)$ is the kernel of this mapping. Also we can use homomorphism $\varphi $ which is described above and denote it as ${{\varphi }_{l}}$, to map ${{G}_{k}}(l)$ onto ${{C}_{2}}$ the $\ker {{\varphi }_{l}}=G_{k}^{2}(l)$. Really if $\alpha $ from ${{G}_{k}}(l)$ has odd number of active states of v.p. on ${{X}^{l}},\,\,\,l<k-1$ than ${{\varphi }_{l}}(\alpha )=1$ in ${{C}_{2}}$ otherwise if this number is even than $\alpha $ from $\ker {{\varphi }_{i}}$ so ${{\varphi }_{l}}(\alpha )=0$ hence $\ker {{\varphi }_{l}}=G_{k}^{2}(l)$. So ${}^{{{G}_{k}}(l)}/{}_{G_{k}^{2}(l)}={{C}_{2}}$ analogously ${}^{{{B}_{k}}(l)}/{}_{B_{k}^{2}(l)}={{C}_{2}}$. Let us check that mapping $({{\varphi }_{0}},{{\varphi }_{1}},...,{{\varphi }_{k-2}},{{\phi }_{k-1}})$ is the homomorphism from ${{G}_{k}}$ to $\underbrace{{{C}_{2}}\times {{C}_{2}}\times ...\times {{C}_{2}}}_{k}$. By virtue of the fact that we can construct homomorphism ${{\varphi }_{i}}$ from every factor ${G}_{k}(i)$ of this direct product to ${{C}_{2}}$. The group ${}^{{{G}_{k}}}/{}_{G_{k}^{2}}$ is elementary abelian 2-group because ${{g}^{2}}=e,\,\,g\in G$. Parity of index of $\alpha \cdot \beta$ on $X^l$ is equal to sum by $\bmod 2$ of indexes of $\alpha$ and $\beta$ hence $\varphi_l (\alpha \cdot \beta )=\left( \varphi_l (\alpha )+\varphi_l (\beta ) \right)$ because multiplication $\alpha \cdot \beta $ in ${{G}_{k}}$ does not change a parity of index of $\beta $, $\beta \in {{G}_{k}}$ on ${{X}^{l}}$. Really action of element of active group $A=\underbrace{{{C}_{2}}\wr {{C}_{2}}\wr ...\wr {{C}_{2}}}_{l-1}$ from wreath power $(\underbrace{{{C}_{2}}\wr {{C}_{2}}\wr ...\wr {{C}_{2}}}_{l-1})\wr {{C}_{2}}$ on element from passive subgroup ${{C}_{2}}$ of second multiplier from product $gf,\,\,g,f\in (\underbrace{{{C}_{2}}\wr {{C}_{2}}\wr ...\wr {{C}_{2}}}_{l-1})\wr {{C}_{2}}$ does not change a parity of index of $\beta $ on ${{X}^{l}}$, if index of $\beta $ was even then under action it stands to be even and the sum $\varphi (\alpha )\bmod 2+\varphi (\beta )\bmod 2$ will be equal to $(\varphi (\alpha )+\varphi (\beta ))\bmod 2$, hence it does not change a $\varphi (\beta )$. Since words with equal logarithms to all bases [@K] belong to distinct cosets of the commutator, the subgroup $G_{k}^{2}(l)$ is the kernel of this mapping. Let us define the permutations of the type 2 that act on ${{X}_{1}}$ and ${{X}_{2}}$, where ${{X}_{1}}=\{{{v}_{k,1}},...,{{v}_{k{{,2}^{k-1}}}}\},\,\,{{X}_{2}}=\{{{v}_{k{{,2}^{k-1}}+1}},...,{{v}_{k{{,2}^{k}}}}\},\,{{X}_{1}}\cup {{X}_{2}}={{X}^{k}}$ only by even permutations. Subgroup $G_{k}^{2}G{{'}_{k}}$ acts only by permutations of type 2 on ${{X}_{1}}$, ${{X}_{2}}$, according to Statement \[comm\]. The restriction ${{\left. G_{k}^{2} \right|}_{{{X}^{[k-1]}}}}$ acts only by permutations of the second type (elements of it form a normal subgroup in ${{G}_{k}}$) by parity of permutation on sets ${{X}_{1}}$ and ${{X}_{2}}$. A permutation of Type 1, where on ${{X}_{1}}$ and ${{X}_{2}}$ the group ${{G}_{k}}$ can acts by odd as well as by even permutations but in such way to resulting permutation on ${{X}^{k}}$ is always even. The number of active states from subgroup ${{G}_{k}}(k-1)$ on ${{X}^{k-1}}$ can be even as well as odd. It means that on set of vertices of ${{X}^{k-1}}$ over ${{X}_{1}}$ i.e. vertices that are connected by edges with vertices of ${{X}^{k-1}}$ over ${{X}_{1}}$ automorphism of ${{G}_{k}}$ can contains odd number of active states (and ${{X}_{2}}$ analogously). Hence for a subgroup $G(k-1)\simeq W_{k-1}$ such that has the normal subgroup ${{G}_{k}}^2({k-1})\triangleleft {{G}_{k}}({k-1})$ we construct a homomorphism: $ {{\phi }_{k-1}} \left( {{G}_{k}}({{X}_{k}}) \right)\to {{C}_{2}} \simeq {}^{{G}_{k}}({k-1})/{}_{ G^2_k (k-1)} $ as product of sum by $mod2$ of active states (${{s}_{k-1,i}}\in \{0,1\}$, $0<j\le {{2}^{k-2}}$ if ${{v}_{ij}}\in {{X}_{1}}$ and $X_2$ corespondently) on each set ${{X}_{1}}$ and ${{X}_{2}}$: $\phi_{\alpha} ({{X}_{1}})\,\, \cdot \,\,\phi_{\alpha} ({{X}_{2}})=\sum\limits_{i=1}^{{{2}^{k-2}}}{{{s}_{k-1,i}}(\alpha)}(\bmod 2) \cdot \sum\limits_{i={{2}^{k-2}}+1}^{{{2}^{k-1}}}{{{s}_{k-1,i}}(\alpha)} (\bmod 2)$. Where ${{s}_{k-1,i}}(\alpha)=1$ if there is active state in ${{v}_{k-1,i}}, \, i<2^{k-1}+1$ and ${{s}_{k-1,i}}(\alpha)=0$ if there is no active state. It follows from structure of ${{G}_{k}}$ that $\phi_{\alpha} ({{X}_{1}})\,\,=\,\,\phi_{\alpha} ({{X}_{2}})$ so it is 0 or 1. But $G_{k}^{2}{{G}^{'}_{k}}$ admits only permutations of Type 2 on ${{X}^{k}}$ so $G_{k}^{2}{{G}_{k}}'({{X}_{k}})\triangleleft {{G}_{k}}({{X}_{k}})$ because it holds a conjugacy and it is a kernel of mapping from ${{G}_{k}}({{X}_{k}})$ onto ${{C}_{2}}$. Hence for a subgroup $W_{k-1}$ such that has the normal subgroup ${{G}_{k}}^{2}(k-1)\triangleleft W_{k-1}$ it was constructed a homomorphism: ${{\phi }_{k-1}}\left( W_{k-1} \right)\to {{C}_{2}}\simeq {{G}_{k}}(k-1){{/}_{G_{k}^{2}(k-1)}}$ as product of sum by $mod2$ of active states from $X_1$ and $X_2$. As the result we have ${}^{{{G}_{k}}}/{}_{G_{k}^{2}}\simeq \underbrace{{{C}_{2}}\times {{C}_{2}}\times ...\times {{C}_{2}}}_{k}$. Considering that it was proved in Theorem 1 and Theorem 2 that ${{G}_{k}}\simeq A_{2^k}$ we can formulate next Corollary. The group $Syl_2 A_{2^k}$ has a minimal generating set with $k$ generators. Since quotient group of ${{G}_{k}}$ by subgroup of Frattini $G_{k}^{2}{{G}^{'}_{k}}$ has minimal set of generators from $k$ elements because ${}^{{{G}_{k}}}/{}_{G_{k}^{2}{{G}^{'}_{k}}}$ is isomorphic to linear $p$-space $(p=2)$ of dimension $k$ (or elementary abelian group). Then according to theorems from [@Rot] $rk ({{G}_{k}})=k$. It means that $A_{2^k}$ is a group with fixed size of minimal generating set. The set $S_{\mathop{\beta}}=\{\mathop{\beta}_{0}, \mathop{\beta}_{1}, \mathop{\beta}_{2}, \ldots , \mathop{\beta}_{k-2}, \tau \}$, where $\mathop{\beta}_{i} = \alpha_i$, is a minimal generating set for a group $G_k$ that is isomorphic to Sylow 2-subgroup of $A_{2^{k}}$. We have isomorphism of $G_k$ and $Syl_2 (A_{2^k})$ from Theorem \[isomor\], the minimality of $S_{\mathop{\beta}}$ following from Theorem \[Th about general relation\] which said that $S_{\mathop{\beta}}$ has to contain an element of type `T`, Theorem 3 and Lemma \[rk\] about minimal rank. Another way to prove the minimality of $S_{\mathop{\beta}}$ is given to us by Corollary \[qoutient\] about quotient by Frattini subgroup. For example a minimal set of generators for $Syl_2(A_{8})$ can be constructed by following way, for convenience let us consider the next set: Consequently, in such way we construct second generating set for $A_{2^k}$ of $k$ elements that is less than in [@Iv], and this set is minimal. We will call ***diagonal base*** (${{S}_{d}}$) for $Syl_2 S_{2^k} \simeq Aut X^{[k]}$ such generating set that has the property $s_{jx}(\alpha_i)=0$ iff $ i \neq j$, (for $1\leq x\leq 2^j$) and every $\alpha_i, \, i<k$ has odd number of active v.p. A number of no trivial v.p. that can be on ${{X}^{j}}$ is odd the number of ways to chose tuple of no trivial v.p. on ${{X}^{j}}$ for generator from ${{S}_{d}}$ and equal to $2^ {2^{j}}:2=2^ {{2}^{j-1}}$. Thus, general cardinality of ${{S}_{d}}$ for $Sy{{l}_{2}}{{S}_{{{2}^{k}}}}$ is ${{2}^{{{2}^{k}}-k-1}}$. There is minimum one generator of type `T` in ${{S}_{d}}$ for $Sy{{l}_{2}}{{A}_{{{2}^{k}}}}$. This generator can be chosen not less than in $C_{{{2}^{k-2}}}^{1}C^1_{{2}^{k-2}}= {({{2}^{k-2}})}^{2} = 2^{2k-4}$ ways. Thus, total cardinality of ${{S}_{d}}$ for $Sy{{l}_{2}}{{A}_{{{2}^{k}}}}$ is ${{2}^{{{2}^{k-1}}-k-2}}{{({{2}^{k-2}})}^{2}}$. And there are $k$ generators in a minimal set of generators, therefore $\left|\phi(G_k)\right| = \left| {{G}_{k}} \right|:{{2}^{k}}$ should be raised to the power of $k$. It equals to ${{(\left| {{G}_{k}} \right|:{{2}^{k}})}^{k}}={{({{2}^{{{2}^{k}}-2}}:{{2}^{k}})}^{k}}={{2}^{k({{2}^{k}}-k-2)}}$. As a result, we have ${{2}^{k({{2}^{k}}-k-1)}} \cdot ({{2}^{k}}-1)({{2}^{k}}-2)({{2}^{k}}-{{2}^{2}})...({{2}^{k}}-{{2}^{k-1}})$. Let us consider an examples of $Syl_2 A_n$ for a cases $n=4k+r$, where $r\leq 3$. The structure of $Syl_2A_{12}$ is the same as of the subgroup $H_{12} < Syl_2(S_8) \times Syl_2(S_4)$, for that $[Syl_2(S_8) \times Syl_2(S_4):H_{12}]=2$, $|Syl_2(A_{12})|= 2^{[12/2] + [12/4]+ [12/8]-1} = 2^9$. Also $|Syl_2(S_8)|=2^7$, $|Syl_2(S_4)|=2^3$, so $|Syl_2(S_8) \times Syl_2(S_4)|=2^{10}$ and $|H_{12}|=2^9$, because its index in $Syl_2(S_8) \times Syl_2(S_4)$ is 2. The structure of $Syl_2(A_6)$ is the same as of $H_6 < Syl_2(S_4) \times (C_2)$. Here $H_6 = \{(g,h_g)|g \in Syl_2(S_4), h_g \in C_2\}$, where $$\label{H} \begin{cases} h_g = e, \ \ if \ g|_{X_2} \in Syl_2(A_6), \\ h_g = (5,6), \ if \, g|_{X^2} \in {Syl_2(S_6) \setminus Syl_2 A_6}. \end{cases}$$ The structure of $Syl_2(A_{6})$ is the same as subgroup $H_6:$ $H_6 < Syl_2(S_4) \times (C_2)$ where $H_6= \{ (g, h) | g\in Syl_2(S_4), h \in AutX \}$. So last bijection determined by (\[H\]) giving us $Syl_2 A_{6} \simeq Syl_2 S_{4} $. As a corollary we have $Syl_2 A_{{2^k}+2} \simeq Syl_2 S_{2^k} $. The structure of $Syl_2(A_{7})$ is the same as of the subgroup $H_7:$ $H_7 < Syl_2(S_4) \times S_2$ where $H_6= \{ (g, h) | g\in Syl_2(S_4), h \in S_2 \}$ and $h$ depends of $g$: $$\label{HH} \begin{cases} h_g = e, \ \ if \ g|_{X^2}\in Syl_2 A_7, \\ h_g = (i,j), i,j \in\{ 5,6,7 \}, \ if \, g|_{X^2}\in {Syl_2 S_7\setminus Syl_2A_7}. \end{cases}$$ The generators of the group $H_7$ have the form $(g,h), \, \, g\in Syl_2(S_4), \, h\in C_2$, namely: $ \{ {\beta_{0}; \beta_{1}, \tau} \} \cup \{ (5,6) \}$. An element $h_g$ can’t be a product of two transpositions of the set: ${(i,j), (j,k), (i,k)}$, where $i,j,k$ $\in\{ 5,6,7 \} $, because $(i,j)(j,k)=(i,k,j)$ but $ord(i,k,j) =3$, so such element doesn’t belong to 2-subgroup. In general elements of $Syl_2 A_{4k+3}$ have the structure (\[HH\]), where $h_g = (i,j), \,\, i,j \in\{ 4k+1, 4k+2, 4k+3 \}$ and $g\in Syl_2 S_{4k}$. Also $|Syl_2(S_8)|=2^7$, $|Syl_2(S_4)|=2^3$, so $|Syl_2(S_8) \times Syl_2(S_4)|=2^{10}$ and $|H_{12}|=2^9$, because its index in $Syl_2(S_8) \times Syl_2(S_4)$ is 2. The structure of $Syl_2(A_6)$ is the same as of $H_6 < Syl_2(S_4) \times (C_2)$. Here $H_6 = \{(g,h_g)|g \in Syl_2(S_4), h_g \in C_2\}$. The sizes of this groups are equal, really $|Syl_2(A_7)|= 2^{[7/2] + [7/4]-1} = 2^3= |H_7|$. In case $g|_{L_2}\in {S_7\setminus A_7}$ we have $C_3^2$ ways to construct one transposition that is direct factor in $H$ which complete $Syl_2 S_4$ to $H_7$ by one transposition $: \{(5,6); (6,7); (5,7) \}$. The structure of $Syl_2(A_{2^k+2^l})$ $(k>l)$ is the same as of the subgroup $H_{2^k+2^l} < Syl_2(S_{2^k}) \times Syl_2(S_{2^l})$, for that $[Syl_2(S_{2^k}) \times Syl_2(S_{2^l}):H]=2$. $|Syl_2(A_{2^k+2^l})|= 2^{[(2^k+2^l)!/2] + [(2^k+2^l)!/4]+ .... -1} $. Here $H = \{(g,h_g)|g \in Syl_2(S_2^k), h_g \in Syl_2(S_2^l\}$, where $$\label{HHH} \begin{cases} h \in A_{2^l}, \ \ if \ g|_{X^{k-1}}\in A_{2^k}, \\ h: h|_{X^2}\in {Syl_2 (S_{2^l}) \setminus Syl_2 A_{2^l}}, \ if \, g|_{X^k}\in {Syl_2 S_{2^k} \setminus Syl_2 A_{2^k}}. \end{cases}$$ The generators of the group $H_7$ have the form $(g,h), \, \, g\in Syl_2(S_4), \, h\in C_2$, namely: ${\beta_{0}; \beta_{1}, \tau} \cup {(5,6)}$. I.e. for element ${{\beta }_{\sigma }}(2i-1)=2\sigma (i)-1,\,\,{{\beta }_{\sigma }}(2i)=2\sigma (i)$, ${{\sigma }_{i}}\in \left\{ {{1,2,...,2}^{k-1}} \right\}$. Let us present new operation $\boxtimes $ (similar to that is in [@Dm]) as a even subdirect product of $Syl{{S}_{{{2}^{i}}}}$, $n = {{2}^{{{k}_{0}}}}+{{2}^{{{k}_{1}}}}+...+{{2}^{{{k}_{m}}}}$, $0\le {{k}_{0}}<{{k}_{1}}<...<{{k}_{m}}$ and $m\ge 0$, i.e. $Syl{{S}_{{{2}^{{{k}_{1}}}}}}\boxtimes Syl{{S}_{{{2}^{{{k}_{2}}}}}}\boxtimes ...\boxtimes Syl{{S}_{{{2}^{{{k}_{l}}}}}}=Par(Syl{{S}_{{{2}^{{{k}_{1}}}}}}\times Syl{{S}_{{{2}^{{{k}_{2}}}}}}\times ...\times Syl{{S}_{{{2}^{{{k}_{l}}}}}})$, where $Par(G)$ – set of all even permutations of $G$. Note, that $\boxtimes $ is not associated operation, for instance $ord({{G}_{1}}\boxtimes {{G}_{2}}\boxtimes {{G}_{3}})\,\,\,=\left| {{G}_{1}}\times {{G}_{2}}\times {{G}_{3}} \right|:2$ but $ord(({{G}_{1}}\boxtimes {{G}_{2}})\boxtimes {{G}_{3}})\,\,\,=\left| {{G}_{1}}\times {{G}_{2}}\times {{G}_{3}} \right|:4$. For cases $n=4k+1$, $n=4k+3$ it follows from formula of Legendre. It is well known that the $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{i}}}}}}\simeq \wr _{j=1}^{{{k}_{i}}}{{C}_{2}}$. Since Sylow $p$-subgroup of direct product is direct product of Sylow $p$-subgroups and fact that automorphism of rooted tree keeps an vertex-edge incidence relation then we have $Aut{{X}^{[{{k}_{0}}]}}\times Aut{{X}^{[{{k}_{1}}]}}\times ...\times Aut{{X}^{[{{k}_{m}}]}}\simeq Sy{{l}_{2}}{{S}_{n}}$, $n={{2}^{{{k}_{1}}}}+{{2}^{{{k}_{2}}}}+...+{{2}^{{{k}_{l}}}}$, ${{k}_{i}}\ge 0$, ${{k}_{i}}<{{k}_{i-1}}$. Let us denote a subgroup, that consists of all even substitutions from $Sy{{l}_{2}}{{S}_{n}}$ as $Aut{{X}^{[{{k}_{0}}]}}\boxtimes Aut{{X}^{[{{k}_{1}}]}}\boxtimes ...\boxtimes Aut{{X}^{[{{k}_{m}}]}}$, where a states of v.p. on ${{X}^{{{k}_{0}}-1}}\sqcup {{X}^{{{k}_{1}}-1}}\sqcup ...\sqcup {{X}^{{{k}_{m}}-1}}$ are related by congruence: $$\label{congruen} \sum\limits_{i=0}^{m}{\sum\limits_{j=1}^{{{2}^{{{k}_{i}}-1}}}{{{s}_{{{k}_{i}}-1,j}}({{\alpha }_{i}})\equiv 0\left( \bmod 2 \right)}}.$$ If number of active states on a last level of $Aut{{X}^{[{{k}_{i}}]}}$ from $Aut{{X}^{[{{k}_{0}}]}}\boxtimes \,\,...\,\,\boxtimes Aut{{X}^{[{{k}_{m}}]}}$ is odd, then it is subdirect product of groups $Aut{{X}^{[{{k}_{0}}]}},\,\,...\,\,,Aut{{X}^{[{{k}_{m}}]}}$. It is a quotient group which is a homomorphic image obtained by a mapping from $Aut{{X}^{[{{k}_{0}}]}}\times Aut{{X}^{[{{k}_{1}}]}}\times ...\times Aut{{X}^{[{{k}_{m}}]}} \simeq Sy{{l}_{2}}{{S}_{n}}$ to this quotient group. A kernel of $\varphi $ consists of all automorphisms which satisfy a congruence $\sum\limits_{i=0}^{m}{\sum\limits_{j=1}^{{{2}^{{{k}_{i}}-1}}}{{{s}_{{{k}_{i}}-1,j}}({{\alpha }_{i}})\equiv 1\left( \bmod 2 \right)}}$. At once from definition follows, that if number of states on last level of $Aut{{X}^{[{{k}_{i}}]}}$ is odd, then a subgroup from the condition is subdirect product of groups $Aut{{X}^{[{{k}_{0}}]}},\,\,...\,\,,Aut{{X}^{[{{k}_{m}}]}}$. Actually for every state of automorphism $\alpha $ from $Aut{{X}^{[{{k}_{i}}]}}$ on ${{X}^{l}}$, $l<{{k}_{i}}-1$ we have that $(e,...,e,{{\alpha }_{i}},e,...,e)$ belongs to $Aut{{X}^{[{{k}_{0}}]}}\times Aut{{X}^{[{{k}_{1}}]}}\times ...\times Aut{{X}^{[{{k}_{m}}]}}$. An arbitrary state from ${{X}^{{{k}_{i}}-1}}$ is included in $Aut{{X}^{[{{k}_{0}}]}}\boxtimes Aut{{X}^{[{{k}_{1}}]}}\boxtimes ...\boxtimes Aut{{X}^{[{{k}_{m}}]}}$ together with even number of states from last levels of ${{X}^{[{{k}_{0}}]}},\,...\,,{{X}^{[{{k}_{m}}]}}$. Analogous fact was proved in \[1\] for a direct sum of permutations groups and for their subgroups which consists of all even permutations. Our statement is a restiction on a $Sy{{l}_{2}}{{S}_{n}}$. The Sylow subgroup $Sy{{l}_{2}}({{A}_{n}})$ has index 2 in $Syl_{2}({{S}_{n}})$ and it’s structure: $Syl_2{S_{2^{{{k}_{1}}}}}\boxtimes Syl_2{{S}_{{{2}^{{{k}_{2}}}}}}\boxtimes ...\boxtimes Syl_2{S}_{{2}^{{k}_{l}}}$. \[isomorph\] If $n=4k+2$, then the subgroup $Syl_2A_n$ is isomorphic to $Syl_2 S_{4k}$, where $k\in \mathbb{N}$. Let us consider the subgroup $H_{4k+2} = \{(g,h_g)|g \in Syl_2(S_{4k}), h_g \in S_2\}$, where $$\label{HHHH} \begin{cases} h_g = e, \ \ if \ g|_{X^k} \in Syl_2(A_{4k+2}), \\ h_g = (4k+1,4k+2), \ if \, g|_{X^k} \in Syl_2(S_{4k+2}) \setminus Syl_2(A_{4k+2}). \end{cases}$$ For instance the structure of $Syl_2(A_{6})$ is the same as subgroup $H_6:$ $H_6 < Syl_2(S_4) \times (C_2)$, where $H_6= \{ (g, h) | g\in Syl_2(S_4), h \in AutX \}$. So last bijection determined by (\[HHHH\]) give us $Syl_2 A_{6} \simeq Syl_2 S_{4} $. As a corollary we have $Syl_2 A_{{2^k}+2} \simeq Syl_2 S_{2^k} $. Bijection correspondence between set of elements of $Syl_2(A_n)$ and $Syl_2(S_{4k})$ we have from (\[HHHH\]). Let’s consider a mapping $\phi: Syl_2 (S_{4k}) \rightarrow Syl_2 (A_{4k+2})$ if $\sigma \in Syl_2(S_{4k})$ then $\phi(\sigma)=\sigma \circ (4k+1, 4k+2)^{\chi(\sigma)}=(\sigma, (4k+1, 4k+2)^{\chi(\sigma)})$, where $\chi(\sigma)$ is number of transposition in $\sigma$ by module 2. So $\phi(\sigma) \in Syl_2(A_{4k+2})$. If $\phi(\sigma) \in A_{n}$ then ${\chi(\sigma)}=0$, so $\phi(\sigma) \in Syl_2(A_{n-1})$. Check that $\phi$ is homomorphism. Assume that ${{\sigma }_{1}}\in Sy{{l}_{2}}({{S}_{4k}}\backslash {{A}_{4k}}),\,\,{{\sigma }_{2}}\in Sy{{l}_{2}}({{A}_{4k}})$, then $\phi ({{\sigma }_{1}})\phi ({{\sigma }_{2}})=({{\sigma }_{1}},{{h}^{\chi({{\sigma }_{1}})}})({{\sigma }_{2}},e)=({{\sigma }_{1}}{{\sigma }_{2}},h)={{\sigma }_{1}}{{\sigma }_{2}}\circ (4k+1,4k+2)$, where $({{\sigma }_{i}},h)={{\sigma }_{i}}\circ {{h}^{\chi({{\sigma }_{i}})}}\in Sy{{l}_{2}}({{A}_{4k+2}})$. If ${\sigma _{1}},\,\,{\sigma_{2}}\in {{S}_{{{2}^{k}}}}\backslash {{A}_{{{2}^{k}}}}$, then $\phi ({{\sigma }_{1}})\phi ({{\sigma }_{2}})=({{\sigma }_{1}},{{h}^{\chi({{\sigma }_{1}})}})({{\sigma }_{2}},{{h}^{\chi({{\sigma }_{2}})}})=({{\sigma }_{1}}{{\sigma }_{2}},\,e)=(a,\,e)$, where ${{\sigma }_{1}}{{\sigma }_{2}}=a\in {{A}_{4k+2}}$. So it is isomorphism. Let ${{n}_{m}}={{2}^{{{k}_{0}}}}+{{2}^{{{k}_{1}}}}+...+{{2}^{{{k}_{m}}}}$, where $0\le {{k}_{0}}<{{k}_{1}}<...<{{k}_{m}}$ and $m\ge 0$. If ${{n}_{m}}=4k+2$, then the minimal set of generators for $Sy{{l}_{2}}{{A}_{n_m}}$ has $\sum\limits_{i=1}^{m}{{{k}_{i}}}$ elements. Actually, according to Lemma \[isomorph\], $Sy{{l}_{2}}{{A}_{4k+2}}$ is isomorphic to $Sy{{l}_{2}}{{S}_{4k}}$. In Statement 2 it was proved that $Sy{{l}_{2}}{{S}_{4k}}\simeq Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{1}}}}}}\times ...\times Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{m}}}}}}$, where $4k={{2}^{{{k}_{1}}}}+...+{{2}^{{{k}_{m}}}}$, ${{k}_{1}}<...<{{k}_{m}}$. On the other hand, $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{i}}}}}}\simeq Aut{{X}^{[{{k}_{i}}]}}$, so there exists the homomorphism $\varphi $ from every factor $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{i}}}}}}$ onto $C_{2}^{{{k}_{i}}}$. Such homomorphism was defined in Corollary \[qoutient\] and in [@Gr]. And what is more it is known that $Aut{{X}^{[{{k}_{i}}]}}$ has a minimal generating set of $k_i$ generators [@Gr]. Thus, there exists the homomorphism from $Aut{{X}^{[{{k}_{1}}]}}\times \,\,...\,\,\times Aut{{X}^{[{{k}_{m}}]}}$ onto $C_{2}^{{{k}_{1}}}\times \,...\,\times C_{2}^{{{k}_{m}}}$, so the rank of $Syl_2 A_{4k+2}$ is $\sum\limits_{i=1}^{m}{{{k}_{i}}}$, where $k_1=1$. This result was confirmed by the algebraic system GAP. Actually, it was founded by GAP that the minimal generating set for $Syl_2 A_{14}$, $Syl_2 A_{14} \simeq Sy{{l}_{2}}{{S}_{12}}\simeq Sy{{l}_{2}}{{S}_{{{2}^{{{2}}}}}} \times Syl_2 S_{2^3}$, of 5 elements: $(11,12)(13,14), (9,11)(10,12), (7,8)(9,10), (1,5)(2,6)(3,7)(4,8), (1,3)(2,4)$. \[Action\] If ${{n}_{m}}\equiv 1(\bmod 2)$, then there exists a point $n$ from tuple $M$ of ${{n}_{m}}$ points indexed by numbers from 1 to ${{n}_{m}}$, such that $S{{t}_{Sy{{l}_{2}}{{S}_{{{n}_{m}}}}}}(n)$ is isomorphic to $Sy{{l}_{2}}{{S}_{{{n}_{m}}}}$ acting on a tuple $M$. If ${{S}_{{{n}_{m}}}}$ acts on $M$, then one of a Sylow 2-subgroups $H<{{S}_{{{n}_{m}}}}$ is isomorphic to $Aut{{X}^{[{{k}_{0}}]}}\times Aut{{X}^{[{{k}_{1}}]}}\times ...\times Aut{{X}^{[{{k}_{m}}]}}$, that acts on tuple of ${{n}_{m}}$ points, where${{n}_{m}}={{2}^{{{k}_{0}}}}+{{2}^{{{k}_{1}}}}+...+{{2}^{{{k}_{m}}}}$, ${{k}_{0}}<{{k}_{1}}<...<{{k}_{m}}$. By virtue of the fact that ${{n}_{m}}\equiv 1(\bmod 2)$, then ${{k}_{0}}=0$. Thus point $n$, that is in ${{X}^{[{{k}_{0}}]}}$ has a stabilizer $S{{t}_{H}}(n)\simeq Sy{{l}_{2}}{{S}_{{{n}_{m}}}}$. It is so, because group of automorphism of such group keeps an vertex-edge incidence relation of ${{X}^{\left[ {{k}_{i}} \right]}},\,\,i\in \left\{ 0,...,m \right\}$. Thus, action of every Sylow 2-subgroup of ${{S}_{{{n}_{m}}}}$, where ${{n}_{m}}\equiv 1(\bmod 2)$, fix one element from $\{1,2,...,{{n}_{m}}\}$. According to the Sylow theorem all Sylows $p$-subgroups are conjugated so their actions are isomorphic. In particular, a Sylow 2-subgroup of $S_{2^r}$ is self-normalizing. The number of Sylow 2-subgroups of ${{S}_{{{2}^{r}}}}$ is ${{2}^{r}}!:{{2}^{e}}$ where $e=1+2+...+{{2}^{r-1}}$ [@Weisner]. The mentioned isomorphism may be extended to $Sy{{l}_{2}}{{A}_{4k+3}}\simeq Sy{{l}_{2}}{{A}_{4k+2}}\simeq Sy{{l}_{2}}{{S}_{4k+1}}\simeq Sy{{l}_{2}}{{S}_{4k}}$. Since in accordance with Lemma \[Action\] an action of $Sy{{l}_{2}}{{A}_{4k+3}}$ on the set of $4k+3$ elements fixes one point, then this group as group of action is isomorphic to $Syl_2 {{A}_{4k+2}}$. For a similar reason $Sy{{l}_{2}}{{A}_{4k+1}}\simeq Sy{{l}_{2}}{{A}_{4k}}$. As well as it was proved in Lemma \[isomorph\] that $Sy{{l}_{2}}{{A}_{4k+2}}\simeq Sy{{l}_{2}}{{S}_{4k}}$. The number of generating sets for $Aut X^{k_i}$ is not less then $N_{k_i}=1\cdot 2 \cdot 2^2 \cdot...\cdot 2^{k_i-1}=2^{\frac{k_i(k_i-1)}{2}}$. Hence, the number of generating sets for $Sy{{l}_{2}}{{A}_{4k+2}}$ is not less than $ 2^{\frac{k_1(k_1-1)}{2}} \cdot . . . \cdot 2^{\frac{k_m(k_m-1)}{2}} $. Thus, it can be applied in cryptography [@Myasn]. Relation between sizes of the Sylows subgroup for $n=4k-2$ and $n=4k$ is given by $\left| Sy{{l}_{2}}({{A}_{4k-2}}) \right|={{2}^{i}}\left| Sy{{l}_{2}}({{A}_{4k}}) \right|$, where value $i$ depends only of power of 2 in decomposition of prime number of $k$. Really $\left| {{A}_{4k-2}} \right|=\frac{(4k-2)!}{2}$, therefore $\left| {{A}_{4k}} \right|=\frac{(4k-2)!}{2}(4k-1)4k$, it means that $i$ determines only by $k$ and is not bounded. If $n=4k$, then index $Syl_2(A_{n+3})$ in $A_{n+3}$ is equal to $[S_{4k+1}: Syl_2 (A_{4k+1})](2k+1)(4k+3)$, index $Syl_2(A_{n+1})$ in $A_{n+1}$ as a subgroup of index $2^{m-1}$, where $m$ is the maximal natural number, for which $4k!$ is divisible by $2^m$. For $Syl_2(A_{n+3})$ its cardinality equal to maximal power of 2 which divide $(4k+3)!$ this power on 1 grater then correspondent power in $(4k+1)!$ because $(4k+3)!=(4k+1)!(4k+2)(4k+3)=(4k+1)!2(2k+1)(4k+3)$ so $\mid Syl_2 A_{n+3}\mid= 2^m \cdot 2 = 2^{m+1}$. As a result of it indexes of $A_{n+3}$ and $A_{n+1}$ are following: $ [S_{4k+1}: Syl_2 (A_{4k+1}) ] = \frac{(4k+1)!}{2^m} $ and $[S_{4k+3}: Syl_2 (A_{4k+3})] = [S_{4k+1}: Syl_2 (A_{4k+1}) ](2k+1)(4k+3) = \frac{(4k+1)!}{2^m}(2k+1)(4k+3) $. If $n=2k$ then $[Syl_2(A_n) : Syl_2(S_{n-1})] =2^{m-1}$, where $m$ is the maximal power of 2 in factorization of $n$. $|Syl_2(S_{n-1})|$ is equal to $t$ that is a maximal power of 2 in $(n-1)!$. $|Syl_2(A_{n})|$ is equal to maximal power of 2 in $(n!/2)$. Since $n=2k$ then $(n/2)!=(n-1)!\frac{n}{2}$ and $2^f$ is equal to product maximal power of 2 in $(n-1)!$ on maximal power of 2 in $\frac{n}{2}$. Therefore $\frac{|Syl_2(A_{n})|}{|Syl_2(S_{n-1})|}=\frac {2^{m-1}}{2^t} 2^t=2^{m-1}. $ Note that for odd $m=n-1$ the group $Syl_2(S_{m}) \simeq Syl_2(S_{m-1})$ i.e. $Syl_2(S_{n-1})\simeq Syl_2(S_{n-2})$. The group $Syl_2(S_{n-2})$ contains the automorphism of correspondent binary subtree with last level $X^{n-2}$ and this automorphism realizes the permutation $\sigma$ on $X^{n-2}$. For every $\sigma\in Syl_2(S_{n-2})$ let us set in correspondence a permutation $\sigma (n-1,n)^{\chi (\sigma)} \in Syl_2(A_{n})$, where $\chi (\sigma)$ – number of transposition in $\sigma$ by $mod\, 2$, so it is bijection $\phi(\sigma)\longmapsto \sigma (n-1, n){\chi (\sigma)}$ that has property of homomorphism, see Lemma \[isomorph\]. Thus, we prove that $Syl_2(S_{n-1}) \hookrightarrow Syl_2(A_{n})$ and its index is $2^{d-1}$. The ratio of $|Syl_2(A_{4k+3})|$ and $|Syl_2(A_{4k+1})|$ is equal to 2 and ratio of indexes $[A_{4k+3} : Syl_2(A_{4k+3})]$ and $[A_{4k+1} : Syl_2(A_{4k+1})]$ is equal $(2k+1)(4k+3)$. The ratio $|Syl_2(A_{4k+3})| : |Syl_2(A_{4k+1})|= 2$ holds because formula of Legendre gives us new one power of 2 in $(4k+3)!$ in compering with $(4k+1)!$. Second part of statement follows from theorem about $p$-subgroup of $H$, $[G:H] \neq kp $ then one of $p$-subgroups of $H$ is Sylow $p$-group of $G$. In this case $p=2$ but $|Syl_2(A_{4k+3})| : |Syl_2(A_{4k+1})|=2$ so we have to divide ratio of indexes on $2$. If $n=2k+1$ then $Syl_2(A_n) \cong Syl_2(A_{n-1})$ and $Syl_2(S_n) \cong Syl_2(S_{n-1})$. Sizes of these subgroups are equal to each other according to Legender’s formula which counts power of 2 in $(2k+1)!$ and $(2k)!$ we obtain that these powers are equal. So these maximal 2-subgroups are isomorphic. From Statment 1 can be obtained that vertex with number $2k+1$ will be fixed to hold even number of transpositions on $X^{k_1}$ from decomposition of $n$ which is in Statement 1. For instance $Syl_2(A_{7})\simeq Syl_2(A_{6})$ and by the way $Syl_2(A_{6})\simeq C_2 \wr C_2 \simeq D_4$, $Syl_2(A_{11})\simeq Syl_2(A_{10}) \simeq C_2 \wr C_2 \wr C_2 $. Sizes of these subgroups are equal to each other according to Legender’s formula which counts power of 2 in $(2k+1)!$ and $(2k)!$ we obtain that these powers are equal. So these maximal 2-subgroups are isomorphic. From Statment 1 can be obtained that vertex with number $2k+1$ will be fixed to hold even number of transpositions on $X^{k_1}$ from decomposition of $n$ which is in Statement 1. For instance $Syl_2(A_{7})\simeq Syl_2(A_{6})$ and by the way $Syl_2(A_{6})\simeq C_2 \wr C_2 \simeq D_4$, $Syl_2(A_{11})\simeq Syl_2(A_{10}) \simeq C_2 \wr C_2 \wr C_2 $. Let us denote by $S(n)$ and $S[n]$ a minimal generating system of $Sy{{l}_{2}}({{A}_{n}})$, $Sy{{l}_{2}}({{S}_{n}})$ correspondently. Let ${{n}_{m}}={{2}^{{{k}_{0}}}}+{{2}^{{{k}_{1}}}}+...+{{2}^{{{k}_{m}}}}$, where $0\le {{k}_{0}}<{{k}_{1}}<...<{{k}_{m}}$ and $m\ge 0$. We shall call the top of system of generators a portion of generators from subgroup with the largest degree ${{k}_{m}}$, which belong to tuples, that represent elements from subdirect product, where there are not other non-trivial elements. Any minimal set of generators for $Sy{{l}_{2}}{{A}_{n}}$ has $\sum\limits_{i=0}^{m}{{{k}_{i}}}-1$ generators, if $m>0$, and it has ${{k}_{0}}$ generators, if $m=0$. Construction of generating set $S({{n}_{m}})$ is such that it contains on the first coordinate all generators from $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{0}}}}}}$, it contains all generators of $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{1}}}}}}$ on the second one, analogously it contains on $i$-th coordinate all generators of $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{i}}}}}}$, $0<i\le m$. Hence we can generate on $i$-th coordinate an arbitrary element from $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{i}}}}}}$. States on ${{X}^{{{k}_{0}}-1}}\sqcup {{X}^{{{k}_{1}}-1}}\sqcup ...\sqcup {{X}^{{{k}_{m}}-1}}$ are related by congruence: $$\label{cong} \sum\limits_{i=0}^{m}{\sum\limits_{j=1}^{{{2}^{{{k}_{i}}-1}}}{{{s}_{{{k}_{i}}-1j}}(\alpha_i )\equiv 0\left( \bmod 2 \right)}}$$ This congruence holds due to structure of constructed by us system of generators. Let us consider a construction of a system of generators for $Sy{{l}_{2}}{{A}_{n}}$ and prove its minimality. Consider an induction base on the example of the group ${{A}_{28}}$, since ${{A}_{28}}=Sy{{l}_{2}}{{A}_{4}}\times Sy{{l}_{2}}{{A}_{8}}\times Sy{{l}_{2}}{{A}_{16}}$, according to the proved fact it is equivalent to a construction of minimal system of generators for $Sy{{l}_{2}}{{A}_{28}}=Sy{{l}_{2}}{{S}_{4}}\boxtimes Sy{{l}_{2}}{{S}_{8}}\boxtimes Sy{{l}_{2}}{{S}_{16}}$. All elements of this system define odd substitutions on the respective sets ${{X}^{2}}$ and ${{X}^{3}}$ by the generators of the subgroups $Aut{{X}^{[1]}}=Sy{{l}_{2}}{{S}_{4}}=\left\langle {{\alpha }_{0,1(1)}},{{\alpha }_{1}} \right\rangle $, and $Sy{{l}_{2}}{{S}_{8}}$ where $\left\langle {{\alpha }_{0,2(1)}},{{\alpha }_{1(2),2(1)}},{{\alpha }_{2}} \right\rangle =Aut{{X}^{[3]}}=Sy{{l}_{2}}{{S}_{8}}$, ${{\alpha }_{2}}={{\alpha }_{2(1)}}$. Last subgroup of the subdirect product has the next generating system $Sy{{l}_{2}}{{S}_{16}}=\left\langle {{\alpha }_{0}},\,\,{{\alpha }_{1(1)}},\,\,{{\alpha }_{2(1)}},\,\,{{\alpha }_{3(1)}} \right\rangle $ hence here only the last element determines odd permutation on ${{X}^{4}}$. Other generators for $Sy{{l}_{2}}{{S}_{8}}$ can be decomposed in a product of two special elements ${{\alpha }_{2}}{{\alpha }_{i}}={{\alpha }_{i(1);2(1)}}$. We call *such structure* of element as *odd structure*. Generators of $Syl_{2}{S_4}$, $0 \le i<2$ have the same structure ${{\alpha }_{1}}{{\alpha }_{0}}={{\alpha }_{0;1(1)}}$, where $0\le i<1$. Hence a product of two arbitrary elements from the set $\left\{ {{\alpha }_{0,2(1)}},{{\alpha }_{1(2),2(1)}},{{\alpha }_{2}} \right\}$ is automorphism that has even number of active states of v.p. on ${{X}^{2}}$. Analogously statement is true for $\left\{ {{\alpha }_{0,1(1)}},{{\alpha }_{1}} \right\}$ on ${{X}^{1}}$ and for ${{\alpha }_{1(1)}},\,\,{{\alpha }_{2(1)}},\,\,{{\alpha }_{3(1)}}$ on ${{X}^{3}}$. Now rename generators of group $Sy{{l}_{2}}{{S}_{4}}$ as ${{\beta }_{0,1(1)}}={{t}_{0}},\,\,{{\beta }_{1}}={{t}_{1}}$, for the second group as ${{\alpha }_{0,1(1),2(1)}}={{s}_{0}},\,\,{{\alpha }_{1,2(1)}}={{s}_{1}},\,\,{{\alpha }_{2}}={{s}_{2}}$, rename generators for $Sy{{l}_{2}}{{S}_{16}}$ as ${{f}_{0}}={{\alpha }_{0}},\,...\,,{{f}_{3}}={{\alpha }_{3(1)}}$. A characteristic feature of the built elements is, that those of them, which are generators of the same subgroup $Sy{{l}_{2}}{{S}_{{{2}^{i}}}}$ and implement odd substitutions, contains the unique common 2-cycle, that implements odd substitution on the set, where the correspondent group acts. Thus, for $Aut{{X}^{[2]}}$ it is the cycle $(1,2)$ on set with 4 elements. Therefore, according to the given by us structure of generators of groups $Sy{{l}_{2}}{{S}_{{{2}^{i}}}}$, those of them, that implement odd substitutions, their products and their squares, are already even substitutions. The system of generators for $Sy{{l}_{2}}{{A}_{28}}$, $Sy{{l}_{2}}{{A}_{28}}=Sy{{l}_{2}}{{S}_{4}}\boxtimes Sy{{l}_{2}}{{S}_{8}}\boxtimes Sy{{l}_{2}}{{S}_{16}}$ is $\left\langle({{s}_{0}},{{t}_{0}},e), (t_0, e, f_3), (e, s_{1}, f_3), ({{s}_{1}},{{t}_{0}},e),(e,{{t}_{1}},{{f}_{3}}),(e,{{t}_{2}},{{f}_{3}}),(e,e,\,{{f}_{2}}),(e,e,\,{{f}_{1}}),(e,e,\,{{f}_{0}}) \right\rangle $ that is why substitutions, that are defined by generators ${{f}_{2}},\,\,{{f}_{1}},\,\,{{f}_{0}}$ must be even (we shall call just this ***set as top***) and ${{f}_{3}}$ must has odd structure. Since for the group $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{i}}}}}}$ there exists a surjective homomorphism $({{\varphi }_{0}},\,{{\varphi }_{1}},\,{{\varphi }_{2}},...,\,{{\varphi }_{{{k}_{i}}-1}})$ onto $C_{2}^{k_{i}}$, where ${{\varphi }_{i}}$ is homomorphism from ${{G}_{{{k}_{i}}}}(l)$ onto ${{C}_{2}}$, so there exists a surjective homomorphism from $Sy{{l}_{2}}{{S}_{4}}\times Sy{{l}_{2}}{{S}_{8}}\times Sy{{l}_{2}}{{S}_{16}}$ onto $C_{2}^{9}$ [@Gr]. We apply it to every ${{G}_{{{k}_{i}}}}(l)$, which are the subgroups of $Sy{{l}_{2}}{{S}_{4}}\times Sy{{l}_{2}}{{S}_{8}}\times Sy{{l}_{2}}{{S}_{16}}$, where ${{k}_{i}}\in \left\{ 2,\,3,\,4 \right\}$ and $l$ changes from 0 to ${{2}^{{{k}_{i}}}}-1$ for each ${{k}_{i}}$. Show, that there is enough $\sum\limits_{i=0}^{m}{{{k}_{i}}}-1$ generators in general case, when $m>0$. Arbitrary element from can be generated on the first coordinate as follows $({{s}_{0}},{{\beta }_{1}},e)({{s}_{1}},{{\beta }_{1}},e)=({{\alpha }_{0}},e,e)$ and $({{\alpha }_{0}},e,e)({{s}_{0}},{{\beta }_{1}},e)=({{\alpha }_{1(2)}},{{\beta }_{1}},e)$ then $({{\alpha }_{1}},{{\beta }_{1}},e)({{\alpha }_{0,1(1)}},{{\beta }_{1}},e)=({{\alpha }_{1(12)}},e,e)$. Hence, we have derived on the first coordinate all non-trivial elements from $Sy{{l}_{2}}{{A}_{4}}$. All automorphisms, which has odd structure, provides an odd permutations on ${{X}^{{{k}_{i}}}}$, which contain in its cyclic decomposition the unique cycle of length 2 (in common case it is odd number of such cycles) that rearranges neighboring elements from $X^{{{k}_{i}}}$ and arises under the action of active state from ${{X}^{{{k}_{i}}-1}}$, all other cycles in substitutions are even, because they are defined by states from ${{X}^{k-j}},\,\,1<j<{{k}_{i}}$. Thus, in the multiplication of these automorphisms, reduction of such cycle is obtained or one more conjugated to it 2 cycle appears in the correspondent substitution. The considered by us group $Sy{{l}_{2}}{{S}_{4}}\boxtimes Sy{{l}_{2}}{{S}_{8}}\boxtimes Sy{{l}_{2}}{{S}_{16}}$ is a quotient group of $Sy{{l}_{2}}{{S}_{4}}\times Sy{{l}_{2}}{{S}_{8}}\times Sy{{l}_{2}}{{S}_{16}}$ by normal *closure* ${{R}^{G}}$, where $R$ is the relation that can be derived from the congruence \[cong\]. Thus from the group $Sy{{l}_{2}}{{S}_{4}}\boxtimes Sy{{l}_{2}}{{S}_{8}}\boxtimes Sy{{l}_{2}}{{S}_{16}}$ has on 1 generator less than $Sy{{l}_{2}}{{S}_{4}}\times Sy{{l}_{2}}{{S}_{8}}\times Sy{{l}_{2}}{{S}_{16}}$ because we have new relation from which this generator can be expressed from form ${{t}^{-1}}\omega ,\,\,\omega \in Sy{{l}_{2}}{{A}_{28}}$ and $t$ is one of generators for $Sy{{l}_{2}}{{S}_{4}}\times Sy{{l}_{2}}{{S}_{8}}\times Sy{{l}_{2}}{{S}_{16}}$. Hence now $t$ can be express due to a new relation similar as it was in transformation of Tietce. Thus we have on one generator less because $t$ is depends from rest of generators for $Sy{{l}_{2}}{{S}_{4}}\times Sy{{l}_{2}}{{S}_{8}}\times Sy{{l}_{2}}{{S}_{16}}$. For the group $Sy{{l}_{2}}{{A}_{28}}$ there exists a surjective homomorphism $({{\varphi }_{0}},\,{{\varphi }_{1}},\,{{\varphi }_{2}},\,{{\varphi }_{{{k}_{3}}}})$ onto $C_{2}^{8}$, which is built previously. We apply it to every ${{G}_{{{k}_{i}}}}(l)$, which are the subgroups of $Sy{{l}_{2}}{{S}_{4}}\boxtimes Sy{{l}_{2}}{{S}_{8}}\boxtimes Sy{{l}_{2}}{{S}_{16}}$, where ${{k}_{i}}\in \left\{ 2,\,3,\,4 \right\}$ and $l$ changes from $0$ to ${{2}^{{{k}_{i}}}}-1$ for each ${{k}_{i}}$. The main property of a homomorphism holds for all ${{G}_{{{k}_{i}}}}(l)$. Actually, a parity of index of $\alpha \cdot \beta $ on ${{X}^{l}}$ is equal to sum by $\bmod \,2$ of indexes of $\alpha $ and $\beta $ hence $\varphi ({{\alpha }_{l}}\cdot {{\beta }_{l}})=\left( \varphi ({{\alpha }_{l}})+\varphi ({{\beta }_{l}}) \right)$ because multiplication $\alpha \cdot \beta $ in ${{G}_{k}}$ does not change a parity of index of $\beta $, $\beta \in {{G}_{k}}$ on ${{X}^{l}}$. This parity depends only from parities of $\alpha $ and $\beta $ on ${{X}^{l}}$. That’s why the described mapping is a homomorphism onto ${{C}_{2}}$. The same is true for every $l\in \{0,...,3\}$ and for every ${{k}_{i}}\in \left\{ 2,\,3,\,4 \right\}$ from ${{G}_{{{k}_{i}}}}(l)$. A group ${{A}_{{{n}_{m}}}}$ has the similar structure $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{0}}}}}}\boxtimes \ldots \boxtimes Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{m-1}}}}}}\boxtimes Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{m}}}}}}$ and main property of constructed homomorphism $\varphi ({{\alpha }_{l}}\cdot {{\beta }_{l}})=\left( \varphi ({{\alpha }_{l}})+\varphi ({{\beta }_{l}}) \right)$ holds by the same reason for every, $l\in \left\{ 0,...,{{k}_{m}}-1 \right\}$. Let us separate out a subgroup which corresponds to the last level, at which the active states exist. The parity of the index of this subgroup, let it be ${{G}_{4}}(3)$, depends also on the parities of all other level subgroups of $Aut{{X}^{[2]}}\boxtimes Aut{{X}^{[3]}}\boxtimes Aut{{X}^{[4]}}$ because relation \[cong\] holds. Thus $Sy{{l}_{2}}{{A}_{28}}$ has minimal generating system from 8 elements. Let us present $S(28)$ in form useful for generalization and for counting rank of $S(28)$:\ $\left\langle ({{s}_{0}},e,{{f}_{3}}),({{s}_{1}},e,{{f}_{3}}),(e,\,{{t}_{0}},\,{{f}_{3}}),(e,{{t}_{1}},{{f}_{3}}),(e,{{t}_{2}},{{f}_{3}}),(e,e,\,{{f}_{2}}),(e,e,\,{{f}_{1}}),(e,e,\,{{f}_{0}}) \right\rangle $.\ So $rk(S(28))$=8. Obviously that generator ${{f}_{3}}$ from generating system of $Sy{{l}_{2}}{{S}_{16}}$ occurs only in combination with generators of other subgroups and any other of generators occurs only one time. Thus formula $\sum\limits_{i=0}^{m}{{{k}_{i}}}-1$ holds. This result was confirmed by algebraic system GAP by which it was founded the minimal generating set for $Syl_2 A_{28}$ from 8 elements: $(25,27)(26,28), (23,24)(25,26), (17,21)(18,22)(19,23)(20,24), (17,19)(18,20),\\ (15,16)(17,18), (1,9)(2,10)(3,11)(4,12)(5,13)(6,14)(7,15)(8,16), (1,5)(2,6)(3,7)(4,8),\\ (1,3)(2,4)$. Thus, the base of induction is checked. The construction of homomorphism and minimal generating set for case $n={{2}^{k}}$ was fully investigated in Corollary \[qoutient\]. Show, that there is enough $\sum\limits_{i=0}^{m}{{{k}_{i}}}-1$ generators in general case, when $m>0$. There are all ${{k}_{i}}$ generators of $Sy{{l}_{2}}{{A}_{{{2}^{{{k}_{i}}}}}}$ on $i$-th coordinate. Consider the example from the base of induction arbitrary element from $Sy{{l}_{2}}{{A}_{4}}$ can be generated on the first coordinate as follows $({{s}_{0}},{{\beta }_{1}},e)({{s}_{1}},{{\beta }_{1}},e)=({{\alpha }_{0}},e,e)$ and $({{\alpha }_{0}},e,e)({{s}_{0}},{{\beta }_{1}},e)=({{\alpha }_{1(2)}},{{\beta }_{1}},e)$ then $({{\alpha }_{1}},{{\beta }_{1}},e)({{\alpha }_{0,1(1)}},{{\beta }_{1}},e)=({{\alpha }_{1(12)}},e,e)$. Hence, we have derived on the first coordinate all non-trivial elements from $Sy{{l}_{2}}{{A}_{4}}$. All automorphisms, which has odd structure, provides an odd permutations on ${{X}^{{{k}_{i}}}}$, which contain in its cyclic decomposition the unique cycle of length 2 (in common case it is odd number of such cycles) that rearranges neighboring elements from $X^{{{k}_{i}}}$ and arises under the action of active state from ${{X}^{{{k}_{i}}-1}}$, all other cycles in substitutions are even, because they are defined by states from ${{X}^{k-j}},\,\,1<j<{{k}_{i}}$. Thus, in the multiplication of these automorphisms, reduction of such cycle is obtained or one more conjugated to it 2 cycle appears in the correspondent substitution. So, closure holds with respect to parity of obtained tuples of elements as a result of multiplication in the whole subdirect product, because in the product automorphism is reduced on the second coordinate, which is determined by the same state, that is determined by ${{\beta }_{3}}$. It is so because it has the order 2 therefore $e$ are on other coordinates, and automorphism on the first coordinate becomes an even substitution already after the first multiplication by $({{s}_{0}},{{\beta }_{1}},e)$. Indeed, system of generators for $\left\langle {{t}_{0}},\,{{t}_{1}} \right\rangle $ has such cyclic structure, that the same cycle implements the unique odd substitution - this is the cycle of length 2, therefore it is reduced as a result of multiplication of two such generators and even substitution is obtained, the generators ${{s}_{0}},\,{{s}_{1}},\,{{s}_{2}}$, ${{f}_{0}},{{f}_{1}},{{f}_{2}},{{f}_{3}}$ for $Sy{{l}_{2}}{{A}_{8}}$ and $Sy{{l}_{2}}{{A}_{16}}$ correspondently have the same structure, so this is the system of generators, that has only 7 elements, while the direct product of these groups should have 9 elements. Thus, for this example the statement is true. Let us show, that minimal system of generators can be constructed by the same method for ${A_{n_m}}$ that contains in subdirect product the factor $Sy{{l}_{2}}{{S}_{{2^{{k_m}}}}}$. All generator except those from the top of generating system for $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{m}}}}}}$ has odd structure. For $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{m}}}}}}$ we choose one generator that realize odd permutation on ${{X}^{{{k}_{m}}}}$ analogously as ${{\alpha }_{{{k}_{m}}-1}}$ and other generators ${{\alpha }_{{{k}_{0}}}},...,{{\alpha }_{{{k}_{m}}-2}}$ belong to top of system of generators of $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{m}}}}}}$, so they define even substitutions on ${{X}^{{{k}_{m}}}}$. Here every separately taken generator for the separated subgroup $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{i}}}}}},\,\,\,{{k}_{i}}<{{k}_{m}}$ defines odd substitution and has odd structure, only generators from *the top of system of generators* of the subgroup $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{m}}}}}}$ implement even substitutions. Only one generator from $S({{2}^{{{k}_{m}}}})$ has odd structure and only this generator stands in pair with generator from $S({{2}^{{{k}_{m}}-1}})$ in tuple of generators of subdirect product for $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{0}}}}}}\boxtimes Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{1}}}}}}\boxtimes \,\,...\,\,\boxtimes Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{m}}}}}}$. Thus rank of $S({{n}_{m}})$ in comparing to $S({{n}_{m-1}})$ will grows on ${{k}_{m}}$. Thus rank of $S({{n}_{m}})$ became to be $\sum\limits_{i=1}^{m}{{{k}_{i}}}-1$ as it states in formula from condition. For the subgroup $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{i}}}}}},\,\,i<m$, distinguish its subgroups of the form ${{G}_{{{k}_{i}}}}(l),\,\,l<{{k}_{i}}-1$ elements from this subgroup act on ${{X}^{{{k}_{i}}}}$ by even substitutions and therefore are not linked with elements ${{G}_{{{k}_{j}}}}(s),\,\,s<{{k}_{j}}-1$, where $i\ne j$, ${{k}_{i}}-s={{k}_{j}}-l$ by parity relation and they forming a direct product. States from last levels ${{X}^{{{k}_{0}}}},...,{{X}^{{{k}_{m}}}}$ are linked by parity relation and form the subdirect product, in which there is only even number of states in every tuple with $m+1$ coordinates. Construction of generating system $S({{n}_{m}})$ is such that it presents on the first coordinate all generators from $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{0}}}}}}$, it presents all generators of $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{1}}}}}}$ on the second one, analogously it presents on $i $-th coordinate all generators of $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{i}}}}}}$, $0<i\le m$. Hence we can generate on i-th coordinate an arbitrary element from $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{i}}}}}}$. A states on last level of ${{X}^{[{{k}_{0}}-1]}}\sqcup {{X}^{[{{k}_{1}}-1]}}\sqcup ...\sqcup {{X}^{[{{k}_{m}}-1]}}$ are related by congruence \[cong\]. And form the subgroup of $H=Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{0}}}}}}\boxtimes \,\,\,...\,\,\,\boxtimes Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{m}}}}}}$ that is a quotient group of $G=Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{0}}}}}}\times \,\,\,...\,\,\,\times Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{m}}}}}}$ by normal c closure ${{R}^{G}}$, where $R$ is the relation equivalent to relation \[cong\]. Since as in proof of induction base we have that the group $H$ has on 1 generator less than $G$ because we have new relation \[cong\] that does not belongs to normal closure of relations from the group $G$ and $R$ does not contains new generators than those for $G$, therefore one of old generators can be expressed. This congruence holds due to the structure of constructed by us system of generators. To prove that proposed by us system generates whole group $Sy{{l}_{2}}{{A}_{n}}$ let us prove that there is equality of orders of $\left\langle S(n) \right\rangle $ and $Sy{{l}_{2}}{{A}_{n}}$. Order of $Sy{{l}_{2}}{{S}_{n}}$ is ${{2}^{{{2}^{{{k}_{0}}}}+...+{{2}^{{{k}_{m}}}}-(m+1)}}={{2}^{n-m-1}}$ because $Syl{{S}_{{{2}^{k}}}}\simeq Aut{{X}^{[{{2}^{k}}]}}$ and its order is equal to ${{2}^{{{2}^{k}}-1}}$, where $n={{2}^{{{k}_{0}}}}+{{2}^{{{k}_{1}}}}+...+{{2}^{{{k}_{m}}}}$. Thus $\left| Sy{{l}_{2}}{{A}_{n}} \right|={{2}^{n-m-2}}$. But order of constructed by us quotient group $Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{0}}}}}}\boxtimes Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{1}}}}}}\boxtimes \,\,...\,\,\boxtimes Sy{{l}_{2}}{{S}_{{{2}^{{{k}_{m}}}}}}$ is also equal to ${{2}^{n-m-2}}$ because the relation \[cong\] restricts class of states from last level of ${{X}^{[{{k}_{0}}-1]}}\sqcup {{X}^{[{{k}_{1}}-1]}}\sqcup ...\sqcup {{X}^{[{{k}_{m}}-1]}}$ to the subclass that realize even permutations on ${{X}^{n}}$. The same order for $\frac{n!}{2}$ give us Legender’s formula for finding maximal power of 2 which divides $\frac{n!}{2}$: $\left[ \frac{n}{2} \right]+...+\left[ \frac{n}{{{2}^{{{k}_{m}}}}} \right]-1={{2}^{{{k}_{0}}}}+{{2}^{{{k}_{1}}}}+...+{{2}^{{{k}_{m}}}}-1=(n-m-1)-1=n-m-2$. Also from Statement \[comm\] and corollary from it about $(AutX^{[k]})'$ can be deduced that derived length of $Syl_2 A_2^k$ is not always equal to $k$ as it was said in Lemma 3 of [@Dm] because in case $A_{2^k}$ if $k=2$ its $Syl_2 A_4 \simeq K_4$ but $K_4$ is abelian group so its derived length is 1. Some applications of constructed generating systems =================================================== As it was calculated in previous paragraph a total cardinality of ${{S}_{d}}$ for $Sy{{l}_{2}}{{A}_{{{2}^{k}}}}$ is ${{2}^{{{2}^{k-1}}-k-2}}{{({{2}^{k-2}})}^{2}}$. Thus, if we associate generating set with alphabet and choice of generating set will be a private key, then it can be applied in cryptography [@Myasn]. More over our group $G_k$ has exponential grows of different generating sets and diagonal bases that can be used for extention of key space. Diagonal bases are useful for easy constructing of normal form [@Ushak] of an element $g\in {{G}_{k}}$. Let us consider a function of Morse [@Shar] $f:\,{D^2}\to \mathbb{R}$ that painted at pict. 2 and graph of Kronrod-Reeb [@Maks] that obtained by contraction every set’s component of level of ${{f}^{-1}}(c)$ in point. Group of automorphism of this graph is isomorphic to $Sy{{l}_{2}}{{S}_{{{2}^{k}}}}$, where $k=2$ in general case we have regular binary rooted tree for arbitrary $k\in \mathbb{N}$. According to investigations of [@Maks2] for $D^2$ we have that $Syl_{2}S_{2^k} > G_k \simeq Syl_{2} A_{2^k}$ is quotient group of diffeomorphism group that stabilize function and isotopic to identity. Analogously to investigations of [@Maks; @Maks2; @SkThes] there is short exact sequence $0\to {{\mathbb{Z}}^{m}}\to {{\pi }_{1}}{{O}_{f}}(f)\to G\to 0$, where $G$-group of automorphisms Reeb’s (Kronrod-Reeb) graph [@Maks] and $O_f(f)$ is orbit under action of diffeomorphism group, so it could be way to transfer it for a group $Sy{{l}_{2}}(S_{2^k})$, where $m$ in ${{\mathbb{Z}}^{m}}$ is number of inner edges or vertices in Reeb’s graph, in case for $Syl_{2}S_{4}$ we have $m=3$. Higher half of projection of manifold from pic. 2 can be determed by product of the quadratic forms $-({{(x+4)}^{2}}+{{y}^{2}})({{(x+3)}^{2}}+{{y}^{2}})({{(x-3)}^{2}}+{{y}^{2}})({{(x-4)}^{2}}+{{y}^{2}})=z$ in points $(-4,0) (-3,0) (3,0) (4,0) $ it reach a maximum value 0. Generally there is $-d_{1}^{2}d_{2}^{2}d_{3}^{2}d_{4}^{2}=z$. Conclusion ============ The proof of minimality of constructed generating sets was done, also the description of the structure $Syl_2 A_{2^k}$, $Syl_2 A_{n}$ and its property was founded. [9]{} *Alexey Muranov,* Finitely generated infinite simple groups of infinite commutator width. arXiv:math/0608688v4 \[math.GR\] 12 Sep 2009. *U. Dmitruk, V. Suschansky,* Structure of 2-sylow subgroup of symmetric and alternating group. UMJ. (1981), N. 3, pp. 304-312. *R. Skuratovskii* “Corepresentation of a Sylow p-subgroup of a group Sn”. Cybernetics and systems analysis, (2009), N. 1, pp. 27-41. The action of Sylow 2-subgroups of symmetric groups on the set of bases and the problem of isomorphism of their Cayley graphs. Algebra and Discrete Mathematics. (2016), Vol. 21, N. 2, pp. 264-281. *V. Ivanchenko,* System of generators for 2-sylow subgroup alternating group, Four ukraine conference of young scientists. Kiev: [*KPI*]{} (2015, (http://matan.kpi.ua/uk/ysmp4conf.html)), pp. 60. *R.V. Skuratovskii, Y.A. Drozd* Generators and and relations for wreath products of groups. Ukr Math J. (2008), vol. 60. Issue 7, pp. 1168–1171. *V. Nekrashevych,* Self-similar groups. International University Bremen. American Mathematical Society. 2005. Monographs, Vol. 117, 230 p. *R.I. Grigorchuk,* Solved and unsolved problems around one group. Infinite Groups: Geometric, Combinatorial and Dynamical Aspects. Âasel, (2005). Progress Math., vol 248. pp. 117-218. *V. Magnus, A. Karras, D. Soliter,* Combinatorial Group Theory: Presentations of Groups in Terms of Generators and Relations. – M.: Science, 1974, 453 p. *V.S. Sikora, V. I. Suschanskii,* Operations on groups of permutations. Chernivci: Ruta, 2003, 256 p. *M. I. Kargapolov, J. I. Merzljakov,* Fundamentals of the Theory of Groups Springer. Softcover reprint of the original 1st ed. 1979 edition. Springer. 1st ed. 1979. 312 p. *R. Skuratovskii*. Generators and relations for sylows ð-subgroup of group $S_n$. *Naukovi Visti KPI.* **4** (2013) *Louis Weisner,* On the Sylow Subgroups of the Symmetric and Alternating Groups. American Journal of Mathematics, (Apr., 1925), vol. 47, Issue 2, pp. 121-124. *Oleg Bogopolski,* An Introduction to the Theory of Groups. European Mathematical Society. 2008. — 189 p. *J. J. Rotman,* An introduction to the Theory of Groups. New Yourk. Springer, 1995, XV. 513 p. *V. V. Sharko* Smooth topological equivalence of functions of surfaces. Ukrainian Mathematical Journal, 5 (2003), 687-700 *A.G.Myasnikov, V. Shpilrain, A.Ushakov,* A practical attack on some braid group based cryptographic protocols, Crypto 2005, Springer Lect., Notes Comp. (2005), Sc. 3621, pp. 86-96. *V. V. Sharko,* Smooth topological equivalence of functions of surfaces. Ukrainian Mathematical Journal, May (2003), Volume 55, Issue 5, pp. 832–846. *J.D.P. Meldrum,* Wreath Products of Groups and Semigroups. Pitman Monographs and Surveys in Pure and Applied Mathematic English. 1st Edition. Jun (1995). 425 p. *S. I. Maksymenko,* Homotopy types of stabilizers and orbits of Morse functions on surfaces. Annals of Global Analysis and Geometry. (2006), May vol. 29, issue 3, pp. 241–285. *S. I. Maksymenko,* Path-components of Morse mappings space Some problems of contemporary mathematics. Proceedings of Intsitute of Mathematics of Ukrainian NAS. 1998, vol. 25, pp. 135-153. V. Shpilrain, A. Ushakov. A new key exchange protocol on the decomposition problem. Contemp. Math. Volume 418. (2006), pp. 161-167. *R. V. Skuratovskii,* Minimal generating systems and properties of $Syl_2A_{2^k}$ and $Syl_2A_{n}$. X International Algebraic Conference in Odessa dedicated to the 70th anniversary of Yu. A. Drozd. (2015), pp. 104. *R. V. Skuratovskii,* Minimal generating systems and structure of $Syl_2A_{2^k}$ and $Syl_2A_{n}$. International Conference and PhD-Master Summer School on Graphs and Groups, Spectra and Symmetries. (2016), source: http://math.nsc.ru/conference/g2/g2s2/exptext/Skuratovskii-abstract-G2S2+.pdf. *R. V. Skuratovskii,* Structure of commutant and centralizer of Sylow 2-subgroups of alternating and symmetric groups, minimal generating sets of $Syl_2A_{n}$, its applications in cryptography. CAIM 2017, ROMAI. Algebra, Logic and Geometry (2017), pp. 74-75, source: $http://www.romai.ro/documente poze/Conferinte/Caim17/CAIM 2017 \,\, 2.pdf$ *R. V. Skuratovskii,* Commutant of Sylow 2-subgroups of alternating group and minimal generating sets of $Syl_2 A_n$. Mal’tsev Meeting Novosibirsk, (2017), source: http://math.nsc.ru/conference/malmeet/17/Main.htm

--- abstract: 'We prove existence and regularity results for weak solutions of non linear elliptic systems with non variational structure satisfying $(p,q)$-growth conditions. In particular we are able to prove higher differentiability results under a dimension-free gap between $p$ and $q$.' address: - 'Charles University, Faculty of Mathematics and Physics, Mathematical Institute, Sokolovská 83, 186 75, Prague, Czech Republic' - 'Dipartimento di Matematica, Università di Bologna, Piazza di Porta S.Donato 5, 40126 Bologna, Italy ' - 'Dipartimento di Matematica e Applicazioni, Università di Napoli ”Federico II", via Cintia, 80126 Napoli, Italy' - 'Dipartimento di Matematica e Applicazioni, Università di Napoli “Federico II”, via Cintia, 80126 Napoli, Italy' author: - Miroslav Bulíček - Giovanni Cupini - Bianca Stroffolini - Anna Verde title: 'Existence and regularity results for weak solutions to $(p,q)$-elliptic systems in divergence form' --- Introduction {#S:1} ============ In this paper we focus on the existence and the regularity results for solutions $u$ to the Dirichlet problems associated with the following nonlinear system in divergence form (here $\alpha=1,\ldots,N$) $$\label{dirichlet} \left\{\begin{array}{ll} \displaystyle \sum_{i=1}^n \frac{\partial}{\partial x_i} A_{i}^{\alpha}(Du)=0 &\text{in $ \Omega$} \\ u=u_0 &\text{on $ \partial \Omega$,} \end{array}\right.$$ where the functions $A_i^{\alpha}(\xi)$ are locally Lipschitz continuous in $\mathbb{R}^{nN}$, $\Omega$ is an open bounded subset of $\mathbb{R}^n$ and $Du:\Omega \to \mathbb{R}^{nN}$ represents the gradient of a (vector-valued) function $u:\Omega \to \mathbb{R}^N$. We equip the problem with the general $(p,q)$-growth conditions, i.e., we assume that there are $1< p\le q<\infty$ and two positive constants $m, M$ such that for all $\xi, \lambda\in \mathbb{R}^{nN}$ and for all $i,j=1,\ldots,n$, and $\alpha,\beta=1,\ldots,N$ there holds $$\begin{aligned} \label{ellitticita1} m(1+|\xi|^2)^{\frac{p-2}{2}}|\lambda|^2&\le \sum_{i,j=1}^n\sum_{\alpha,\beta=1}^N \frac{\partial A_{i}^{\alpha}}{\partial \xi^{\beta}_j}(\xi)\lambda^{\alpha}_i\lambda_j^{\beta},\\ \label{crescita-q} \left|\frac{\partial A_{i}^{\alpha}}{\partial \xi_j^{\beta}}(\xi)\right|&\le M (1+|\xi|^2)^{\frac{q-2}{2}}.\end{aligned}$$ Notice that is the usual ellipticity condition and is the $q$-growth condition, from which the name of $(p,q)$-growth come from. Under these assumptions, one can easily observe (see Lemma \[stimaLp\]) that $|A^{\alpha}_i(\xi)|\le C(1+|\xi|)^{q-1}$ with some generic constant $C$ and therefore we can naturally define a notion of a weak solution to in the following way: Let $u_0\in W^{1,p}(\Omega; \mathbb{R}^N)\cap W^{1,q}_{\rm loc}(\Omega;\mathbb{R}^N)$. We say that $u$ is a weak solution to if $$\label{soldebole-Dirichlet1} u-u_0 \in W_0^{1,p} (\Omega;\mathbb{R}^N)\cap W_{\textrm{loc}}^{1,q} (\Omega;\mathbb{R}^N)$$ and for all open $\Omega'$ fulfilling $\overline{\Omega'}\subset \Omega$ and for all $\varphi \in W_0^{1,q}(\Omega';\mathbb{R}^N)$ there holds $$\label{soldebole-Dirichlet} \int_{\Omega}\sum_{i=1}^{n}\sum_{\alpha=1}^{N}A_{i}^{\alpha}(Du)\varphi^{\alpha}_{x_i}(x) \,dx=0.$$ Here, and also in what follows, we use the abbreviation $\varphi^{\alpha}_{x_i}:=\frac{\partial \varphi^{\alpha}}{\partial x_i}$ Our main task in the paper is to establish the existence of such a solution and further some regularity of arbitrary weak solutions. However, contrary to the classical result, we do not in general assume any symmetry condition on the derivative of $A^{\alpha}_i$ and so we do not assume that the system is in variational form. Nevertheless, as done in [@mar91] in the scalar framework, we will need to compensate this lack of symmetry by the following assumption on the asymptotic behavior of the skew-symmetric part, namely, for all $\xi, \lambda\in \mathbb{R}^{nN}$ and for all $i,j=1,\ldots,n$, and $\alpha,\beta=1,\ldots,N$ there holds $$\label{continuita} \left|\frac{\partial A_{i}^{\alpha}}{\partial \xi_j^{\beta}}(\xi)-\frac{\partial A_{j}^{\beta}}{\partial \xi_i^{\alpha}} (\xi)\right|\le M (1+|\xi|^2)^{\frac{q+p-4}{4}}.$$ If $p = q$, the existence of weak solutions to (\[dirichlet\]) can be established using the theory of coercive, monotone operators, see Leray–Lions [@lerlio65], Browder [@browder] and Hartman–Stampacchia . Also the regularity issue has been extensively studied, see the monographs [@gia2], [@giusti] and the surveys [@Mingione] and [@mingione2]. Notice also, that without any further additional structural assumptions, the best[^1] known regularity information about the solution is that $V(Du)\in W^{1,2}_{\rm loc}(\Omega; \mathbb{R}^{nN})$, where $$\label{defV} V(\xi):=(1+|\xi|^2)^{\frac{p-2}{4}}\xi.$$ On the other hand, if $p < q$ the above classical existence results cannot be applied due to the lack of coercivity in $W^{ 1,q}$ . Moreover, the request $u\in W^{1,q }_{\rm loc}(\Omega; \mathbb{R}^N)$ in the definition of weak solution, needed to have a well defined integral, is an additional difficulty. Notice that such a request is a priori assumed in some regularity results under the $p, q$-growth, see for example [@leonetti1], [@bilfuchsCalcVar] and [@cupmarmas4]. The first result of the paper is that any weak solution is in fact twice weakly differentiable. \[T:main-weak\] Let $1<p\le q<\infty$ be arbitrary and $A$ satisfy , and . Then any $u\in W^{1,\max\{q,2\}}_{\rm loc}(\Omega; \mathbb{R}^N)$ fulfilling satisfies for all $\eta\in C_{c}^{\infty}(\Omega)$ the following estimate $$\int_{\Omega}(1+|Du|^2)^{\frac{p-2}{2}} |D^2u|^2\eta^{2}\, dx \le c\int_{\Omega} (1+|Du|^2)^{\frac{q}{2}} |D\eta|^2 \,dx, \label{exdopoH3-1}$$ where the constant $c$ depends only on $m$ and $M$. In particular, we also have that $$\int_{\Omega}|D V(Du)|^2\eta^{2}\, dx \le c\int_{\Omega} (1+|Du|^2)^{\frac{q}{2}} |D\eta|^2 \,dx. \label{exdopoH3-V}$$ The above theorem provides the existence of the second derivatives for arbitrary $1<p\le q<\infty$ but the right hand side of or still depends on the $W^{1,q}$ norm of $u$. We shall improve this estimate provided that $p$ and $q$ are sufficiently close to each other. Thus, the second main theorem of the paper is the following. \[t:main\] Let $1<p\le q<\infty$ be arbitrary and $A$ satisfy , and and $u \in W^{1,\max\{q,2\}}_{\rm loc}(\Omega;\mathbb{R}^N)$ satisfy . Then for all open $\Omega' \subset \overline{\Omega'} \subset \Omega$ the following holds: - If $$\label{ipotesip-q} q<p\frac{n+2}{n}$$ then $$\int_{\Omega'}\left( |V(Du)|^{\frac{2q}{p}}+|D V(Du)|^2 + (1+|D u|^2)^{\frac{p-2}{2}}|D^2u|^2\right)\, dx \le C(\Omega', n, N, p, q, m, M, \|Du\|_{L^p(\Omega)}).$$ - If $u\in L^{\infty}(\Omega;\mathbb{R}^N)$ and $$\label{ipotesip-qb} q<p+2 \qquad \textrm{and} \qquad p<n$$ then $$\begin{split} \int_{\Omega'}&\left(|V(Du)|^{\frac{2q}{p}}+|D V(Du)|^2 + (1+|D u|^2)^{\frac{p-2}{2}}|D^2u|^2\right)\, dx \\ &\qquad \qquad\qquad \qquad\qquad \qquad\qquad \qquad \le C(\Omega', n, N, p, q, m, M, \|Du\|_{L^p(\Omega)},\|u\|_{L^{\infty}(\Omega)}). \end{split}$$ In particular, in both cases we have that $V(Du)\in W^{1,2}_{\rm loc}(\Omega;\mathbb{R}^{nN})$, which, due to the embedding theorem, leads to $Du\in L^{\frac{p2^*}{2}}_{\rm loc}(\Omega;\mathbb{R}^{nN})$. Finally, we state our last main result of the paper. It is an existence result for the Dirichlet problem . For this purpose, we need to consider a regularity assumption on the boundary datum. We shall require in what follows that $$\label{ipotesi-dato} u_0 \in W^{1,r} (\Omega;\mathbb{R}^N), \quad \textrm{with}\,\, r:=\max\left\{2,\frac{p(q-1)}{p-1}\right\}.$$ \[t:main2\] Let $1<p\le q<\infty$ be arbitrary and $A$ satisfy , and . Moreover, let $u_0$ fulfill . Then there exists a weak solution to the problem provided that at least one of the following conditions hold - $p$ and $q$ satisfy . - $p$ and $q$ satisfy , $u_0\in L^{\infty}(\partial \Omega;\mathbb{R}^N)$ [and $$\label{structure} \sum_{i=1}^{n}A_{i}^{\alpha}(\xi)\xi_i^{\alpha}\ge 0, \ \ \forall\xi\in \mathbb{R}^{nN},\ \ \forall \alpha\in \{1,\ldots,N\}.$$]{} As far as the regularity of solutions is concerned, the obstructions are essentially two: we are dealing with systems and under non-standard growth $(p < q)$. Indeed, in the vectorial case, even under the standard growth, the everywhere regularity of solutions for systems, or of minimizers of integrals, cannot be expected unless some structure conditions are assigned, and this holds also for the local boundedness, see e.g. the counterexamples by De Giorgi [@deg] and Šverák-Yan [@sverak]. Since the pioneering paper by Marcellini [@mar89], the theory of regularity in the framework of non- standard growth has been deeply investigated. The results and the contributions to regularity are so many, that it is a hard task to provide a comprehensive overview of the issue. For this, we refer to the survey of Mingione [@Mingione] for an accurate and interesting account on this subject. A common feature is that to get regularity results $p$ and $q$ must be not too far apart, as examples of irregular solutions by Giaquinta [@gia], Marcellini [@mar87] and Hong [@hong] show. On the other hand, many regularity results are available if the ratio $q/p$ is bounded above by suitable constant that in general depends on the dimension $n$, and converges to $1$ when $n$ tends to infinity ([@bil03], [@CarKriPas], [@espoleomin1], [@espoleomin2], [@espoleomin3]). Moreover, the condition on the distance between the exponents $p$ and $q$ can usually be relaxed if the solutions/minimizers are assumed locally bounded. Let us observe that the local higher differentiability results for bounded minimizers of integral functionals satisfying $p, q$-growth conditions is more studied than the analogous issue for systems of PDE’s. In particular, recently, the Authors, in [@CKP], considered integral functionals with convex integrand satisfying $p, q$-growth conditions. They proved local higher differentiability results for bounded minimizers under dimension-free conditions on the gap between the growth and the coercivity exponents; i.e., restricted to the case $p\geq 2$, using an improved Gagliardo-Nirenberg’s inequality. We also observe that an existence result in the $(p,q)$-framework was proved in [@CupLeoMas] for a Dirichlet problem with monotone operators possibly depending on the $x$-variable, but for $p\ge 2$ only. As a novel feature, the main results are achieved through uniform higher differentiability estimates for solutions to a class of auxiliary problems, constructed adding higher order perturbations to the integrand. Here we achieve the same result for systems with non variational structure with control on the skew-symmetric part (see (\[continuita\])). The plan of the paper is the following. In Section \[s:preliminaries\] we prove some preliminar algebraic inequalities. In Sections \[apriori\] and \[P-t:main\] we prove the higher differentiability results Theorem \[T:main-weak\] and Theorem \[t:main\], respectively. In the last section, we prove the existence result (Theorem \[t:main\]) for the problem . Auxiliary algebraic inequalities {#s:preliminaries} ================================ In this part, we recall several algebraic inequalities related to the mapping $A$. Although, their proof can be in some simplified setting found in many works, see e.g. [@mar91 Lemma 4.4, Lemma 2.4], [@Tolksdorf Lemma 1], [@cupmarmas4 Lemma 5.1] or [@mnrr96 Chapter 5], we provide for the sake of clarity a detailed proof here. We start with the first auxiliary result based on the assumptions –. \[stimaLp\] Let $A:\mathbb{R}^{nN}\to \mathbb{R}^{nN}$ be a continuous mapping fulfilling and . Then there exists a positive constant $K$ such that for all $\xi, \eta \in \mathbb{R}^{nN}$ there hold $$\begin{aligned} \label{dis-ellitticita} |\xi|^p &\leq K\left\lbrace (1+|\eta|^2)^{\frac{p(q-1)}{2(p-1)}} + \sum_{i=1}^{n} \sum_{\alpha=1}^NA_i^{\alpha}(\xi)(\xi^{\alpha}_i-\eta_i^{\alpha})\right\rbrace,\\ |A^{\alpha}_i(\xi)| &\leq K (1+|\xi|^2)^{\frac{q-1}{2}} \qquad \textrm{ for all } \alpha=1,\ldots, N \textrm{ and } i=1,\ldots,n. \label{crescitaai}\\ \label{casop>2} |\xi-\eta|^p&\le K\sum_{i=1}^{n} \sum_{\alpha=1}^{N}\left( A_i^{\alpha}(\xi)-A_i^{\alpha}(\eta)\right) (\xi_i^{\alpha}-\eta_i^{\alpha}) \quad \textrm{for }p\ge 2,\\ \label{casop<2} \left( 1+|\xi|^2+|\eta|^2\right)^{\frac{p-2}{2}}|\xi-\eta|^2 &\le K\sum_{i=1}^{n} \sum_{\alpha=1}^{N}\left(A_i^{\alpha}(\xi)-A_i^{\alpha}(\eta)\right) (\xi^{\alpha}_i-\eta^{\beta}_i) \quad \textrm{for } p \in (1,2).\end{aligned}$$ We start the proof with . Since $$\label{Nap1} A^{\alpha}_i(\xi)-A^{\alpha}_i(0)=\int_0^1 \sum_{j=1}^n \sum_{\beta=1}^N\frac{\partial A^{\alpha}_i(t\xi)}{\partial \xi^{\beta}_j}\xi^{\beta}_j\; dt,$$ we can use the assumption , to get $$\left|A^{\alpha}_i(\xi) \right|\le \left|A^{\alpha}_i(0)\right| + M \int_0^1 \sum_{j=1}^n \sum_{\beta=1}^N(1+t^2|\xi|^2)^{\frac{q-2}{2}} |\xi^{\beta}_j|\; dt\le \left|A^{\alpha}_i(0)\right| + M nN\int_0^1 (1+t^2|\xi|^2)^{\frac{q-2}{2}} |\xi|\; dt.$$ Thus, in case $q\ge 2$, the inequality immediately follows. If $q\in (1,2)$ we can continue with estimating the last integral in the following way $$\int_0^1 (1+t^2|\xi|^2)^{\frac{q-2}{2}} |\xi|\; dt = \int_0^{|\xi|} (1+t^2)^{\frac{q-2}{2}} \; dt \le 2^{\frac{2-q}{2}}\int_0^{|\xi|} (1+t)^{q-2} \; dt \le \frac{2^{\frac{2-q}{2}}}{q-1}(1+|\xi|)^{q-1}$$ and we again see that follows directly. To show –, we write $$\begin{aligned} &\sum_{i=1}^{n}\sum_{\alpha=1}^{N}\left( A_i^{\alpha}(\xi)-A_i^{\alpha}(\eta)\right) (\xi_i^{\alpha}-\eta_i^{\alpha}) =\int_0^1 \sum_{i,j=1}^{n}\sum_{\alpha=1}^{N} \frac{\partial A^{\alpha}_i(t\xi + (1-t)\eta)}{\partial \xi^{\beta}_j} ( \xi_j^{\beta}-\eta_j^{\beta}) (\xi_i^{\alpha}-\eta_i^{\alpha})\; dt\\ &\overset{\eqref{ellitticita1}}\ge m|\xi-\eta|^2\int_0^1 (1+|t\xi+(1-t)\eta|^2)^{\frac{p-2}{2}}\; dt. \end{aligned}$$ Then, following step by step proof of Lemma 1.19 in [@mnrr96 Chapter 5], we deduce –. To show , we first consider the case $p \geq 2$. Then by using and and also the Young’s inequality, we can observe that for all $\epsilon >0$ and all $\xi,\eta \in \mathbb{R}^{nN}$, we have $$\begin{aligned} &|\xi|^p \leq c(|\xi- \eta|^p +|\eta|^p) \leq c\left\{ \sum_{i=1}^{n}\sum_{\alpha=1}^{N} (A_i^{\alpha}(\xi)-A_i^{\alpha}(\eta)) (\xi_i^{\alpha}-\eta_i^{\alpha})+|\eta|^p\right\} \\ &\le c \left\lbrace |\eta|^p+ \sum_{i=1}^{n}\sum_{\alpha=1}^{N} A_i^{\alpha}(\xi) (\xi_i^{\alpha}-\eta_i^{\alpha})+ \bar{C}(1+|\eta|^2)^{\frac{q-1}{2}}(|\xi|+|\eta|) \right\rbrace \\& \le c \left\lbrace (1+|\eta|^2)^{\frac{p}{2}}+\sum_{i=1}^{n} \sum_{\alpha=1}^{N} A_i^{\alpha}(\xi) (\xi_i^{\alpha}-\eta_i^{\alpha})+ c_{\epsilon}(1+|\eta|^2)^{\frac{p(q-1)}{2(p-1)}}+\epsilon (|\xi|+|\eta|)^p \right\rbrace; \end{aligned}$$ thus if $\epsilon$ is small enough we get . In the case $1<p<2$, we proceed slightly differently. By using the Young ’s inequality with complementary exponents $\frac{2}{p}$ and $ \frac{2}{2-p}$ we get for $\epsilon >0$ $$\begin{aligned} |\xi|^p &\leq c\left(|\xi- \eta|^p +|\eta|^p\right) \le c\left( |\eta|^p+ (|\xi- \eta|^2)^{\frac{p}{2}} (1+|\xi|^2 +|\eta|^2)^{ \frac{p(p-2)}{4} + \frac{p(2-p)}{4}}\right) \\ &\le c \left\{(1+|\eta|^2)^{\frac{p}{2}}+ c_{\epsilon} (1+ |\xi|^2+|\eta|^2)^{\frac{p-2}{2}}|\xi-\eta|^2+\epsilon (1+ |\xi|^2+|\eta|^2)^{\frac{p}{2}}\right\}.\end{aligned}$$ Therefore, by , with a proper choice of (small) $\epsilon>0$, we get $$|\xi|^p \leq c \left\{ (1+|\eta|^2)^{\frac{p}{2}} +\sum_{i=1}^{n} \sum_{\alpha=1}^{N} (A_i^{\alpha}(\xi)-A_i^{\alpha}(\eta)) (\xi_i^{\alpha}-\eta_i^{\alpha}) \right\}$$ and we conclude by proceeding as above. The following estimate will play a crucial role for getting the information about the second derivatives of the weak solutions to . \[Tolsk\] Let $A$ be a continuous mapping fulfilling , and . Then there exists a positive constant $K$ such that for all $\xi,\eta,\zeta\in \mathbb{R}^{nN}$ we have $$\label{key-3} \begin{aligned} \frac{m}{2}(1+|\zeta|^2)^{\frac{p-2}{2}}|\xi|^2 & \le \sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N \frac{\partial A^{\alpha}_i(\zeta)}{\partial \zeta^{\beta}_j}(\xi^{\alpha}_i-\eta^{\alpha}_i)\xi^{\beta}_j+K(1+|\zeta|^2)^{\frac{q-2}{2}}|\eta|^2. \end{aligned}$$ For arbitrary $\zeta, \xi, \eta \in \mathbb{N}$, we define a bilinear form (for fixed $\zeta$) $$(\xi,\eta)_{\zeta}:=\frac12\sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N\left(\frac{\partial A^{\alpha}_i(\zeta)}{\partial \zeta^{\beta}_j}+\frac{\partial A^{\beta}_j(\zeta)}{\partial \zeta^{\alpha}_i}\right)\eta^{\alpha}_i\xi^{\beta}_j.$$ Trivially $(\xi,\eta)_{\zeta}=(\xi,\eta)_{\zeta}$. Moreover, using the assumption we get that $$(\xi,\xi)_{\zeta}=\sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N\frac{\partial A^{\alpha}_i(\zeta)}{\partial \zeta^{\beta}_j}\xi^{\alpha}_i\xi^{\beta}_j\ge m(1+|\zeta|^2)^{\frac{p-2}{2}}|\xi|^2,$$ and consequently, we see that for any fixed $\zeta$, the relation $(\xi,\eta)_{\zeta}$ is a scalar product on $\mathbb{R}^{nN}$ and therefore the Cauchy–Schwarz inequality holds, i.e., $$|(\xi,\eta)_{\zeta}|\le (\xi,\xi)^{\frac12}_{\zeta}(\eta,\eta)^{\frac12}_{\zeta}. \label{CSc}$$ Thus, by assumption and taking into account that $$\sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N \left(\frac{\partial A^{\alpha}_i(\zeta)}{\partial \zeta^{\beta}_j}- \frac{\partial A^{\beta}_j(\zeta)}{\partial \zeta^{\alpha}_i}\right)\xi^{\alpha}_i\xi^{\beta}_j=0,$$ we have $$\begin{aligned} m&(1+|\zeta|^2)^{\frac{p-2}{2}}|\xi|^2 \le (\xi,\xi)_{\zeta} = -(\xi-\eta,\xi-\eta)_{\zeta}+ 2(\xi,\xi-\eta)_{\zeta} + (\eta,\eta)_{\zeta}\le 2(\xi,\xi-\eta)_{\zeta} + (\eta,\eta)_{\zeta}\\ &=\sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N \left(\frac{\partial A^{\alpha}_i(\zeta)}{\partial \zeta^{\beta}_j}+ \frac{\partial A^{\beta}_j(\zeta)}{\partial \zeta^{\alpha}_i}\right)(\xi^{\alpha}_i-\eta^{\alpha}_i)\xi^{\beta}_j+ \sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N \frac{\partial A^{\alpha}_i(\zeta)}{\partial \zeta^{\beta}_j}\eta^{\alpha}_i\eta^{\beta}_j\\ &=2\sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N \frac{\partial A^{\alpha}_i(\zeta)}{\partial \zeta^{\beta}_j}(\xi^{\alpha}_i- \eta^{\alpha}_i)\xi^{\beta}_j-\sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N \left(\frac{\partial A^{\alpha}_i(\zeta)}{\partial \zeta^{\beta}_j}- \frac{\partial A^{\beta}_j(\zeta)}{\partial \zeta^{\alpha}_i}\right)(\xi^{\alpha}_i-\eta^{\alpha}_i)\xi^{\beta}_j\\ &\qquad +\sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N \frac{\partial A^{\alpha}_i(\zeta)}{\partial \zeta^{\beta}_j}\eta^{\alpha}_i \eta^{\beta}_j\\ &=2\sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N \frac{\partial A^{\alpha}_i(\zeta)}{\partial \zeta^{\beta}_j}(\xi^{\alpha}_i-\eta^{\alpha}_i)\xi^{\beta}_j + \sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N \left(\frac{\partial A^{\alpha}_i(\zeta)}{\partial \zeta^{\beta}_j}-\frac{\partial A^{\beta}_j(\zeta)}{\partial \zeta^{\alpha}_i} \right)\eta^{\alpha}_i\xi^{\beta}_j\\ &\qquad +\sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N \frac{\partial A^{\alpha}_i(\zeta)}{\partial \zeta^{\beta}_j}\eta^{\alpha}_i \eta^{\beta}_j\\ &\overset{\eqref{crescita-q},\eqref{continuita}}\le 2\sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N \frac{\partial A^{\alpha}_i(\zeta)}{\partial \zeta^{\beta}_j}(\xi^{\alpha}_i-\eta^{\alpha}_i)\xi^{\beta}_j +MnN(1+|\zeta|^2)^{\frac{q+p-4}{4}}|\eta||\xi|+MnN(1+|\zeta|^2)^{\frac{q-2}{2}}|\eta|^2. \end{aligned}$$ Taking into account that $$(1+|\zeta|^2)^{\frac{q+p-4}{4}}|\eta||\xi|=\left((1+|\zeta|^2)^{\frac{p-2}{4}}|\xi|\right) \left((1+|\zeta|^2)^{\frac{q-2}{4}}|\eta|\right)$$ and using the Young’s inequality, we get $$m(1+|\zeta|^2)^{\frac{p-2}{2}}|\xi|^2 \le 2\sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N \frac{\partial A^{\alpha}_i(\zeta)}{\partial \zeta^{\beta}_j} (\xi^{\alpha}_i-\eta^{\alpha}_i)\xi^{\beta}_j+\frac{m}{2}(1+|\zeta|^2)^{\frac{p-2}{2}}|\xi|^2+C(1+|\zeta|^2)^{\frac{q-2}{2}}|\eta|^2$$ for a suitable constant $C$. Then, easily follows. Proof of Theorem \[T:main-weak\] {#apriori} ================================ We proceed via difference quotients technique. Due to the assumed regularity of the solution $u$ and thanks to , it follows from that $$\int_{\Omega}\sum_{i=1}^n \sum_{\alpha=1}^N (A^{\alpha}_i(Du(x+he_k))-A^{\alpha}_i(Du(x)))\varphi^{\alpha}_{x_i}\; dx =0,$$ for all $\varphi\in W^{1,q}_0(\Omega_h;\mathbb{R}^N)$, all $h\in (0,1)$ and all $k=1,\ldots, n$, where $\Omega_h:=\{x\in \Omega: \; B_{2h}(x)\subset \Omega\}$ and $e_k$ is a unit vector in the $k$-th direction. Hence, setting $$\varphi(x):=(u(x+he_k)-u(x))\tau^2(x)$$ with $\tau \in \mathcal{C}^{\infty}_c(\Omega_{2h})$ (which is an admissible choice), we obtain the starting identity $$\label{start} \begin{split} 0&=\int_{\Omega}\sum_{i=1}^n \sum_{\alpha=1}^N (A^{\alpha}_i(Du(x+he_k))-A^{\alpha}_i(Du(x)))\tau(x)\cdot\\ &\qquad \cdot \left((u^{\alpha}_{x_i}(x+he_k)-u^{\alpha}_{x_i}(x))\tau(x)+2(u^{\alpha}(x+he_k)-u^{\alpha}(x))\tau_{x_i} \right)\; dx. \end{split}$$ Since $$\begin{aligned} &A^{\alpha}_i(Du(x+he_k))-A^{\alpha}_i(Du(x))\\ &=\int_0^t\sum_{j=1}^n \sum_{\beta=1}^N \int_0^1 \frac{\partial A^{\alpha}_i(tDu(x+he_k)+(1-t)Du(x))}{\partial \zeta^{\beta}_j}(u^{\beta}_{x_j}(x+he_k)-u^{\beta}_{x_j}(x))\; dt, \end{aligned}$$ the identity can be equivalently rewritten as $$\label{start2} \begin{split} 0&=\int_{\Omega}\sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N\int_0^1 \frac{\partial A^{\alpha}_i(tDu(x+he_k)+(1-t)Du(x))}{\partial \zeta^{\beta}_j}(u^{\beta}_{x_j}(x+he_k)-u^{\beta}_{x_j}(x))\tau(x)\cdot\\ &\qquad \cdot \left((u^{\alpha}_{x_i}(x+he_k)-u^{\alpha}_{x_i}(x))\tau(x)+2(u^{\alpha}(x+he_k)-u^{\alpha}(x))\tau_{x_i} \right)\; dt\; dx. \end{split}$$ Abbreviating for the moment $$\xi^{\alpha}_i:= \tau(x)(u^{\alpha}_{x_i}(x+he_k)-u^{\alpha}_{x_i}(x)), \quad \eta^{\alpha}_i:=-2(u^{\alpha}(x+he_k)-u^{\alpha}(x))\tau_{x_i}(x)$$ and $$\zeta:=tDu(x+he_k)+(1-t)Du(x),$$ we can formally rewrite as $$\begin{split} 0&=\int_{\Omega}\int_0^1\sum_{i,j=1}^n \sum_{\alpha,\beta=1}^N \frac{\partial A^{\alpha}_i(\zeta)}{\partial \zeta^{\beta}_j}\xi^{\beta}_j\left(\xi^{\alpha}_i-\eta^{\alpha}_i\right)\; dt\; dx. \end{split}$$ Thus, using , we obtain (here $C$ is some constant depending only on $m,M,n,N,p,q$) $$\int_{\Omega}\int_0^1 (1+|\zeta|^2)^{\frac{p-2}{2}}|\xi|^2\; dt\; dx \le C\int_{\Omega}\int_0^1(1+|\zeta|^2)^{\frac{q-2}{2}}|\eta|^2\; dt \; dx,$$ which in terms of original variables after division by $h^2$ means that $$\begin{split}\label{huz} &\int_{\Omega}\int_0^1 (1+|tDu(x+he_k)+(1-t)Du(x))|^2)^{\frac{p-2}{2}}\frac{|Du(x+he_k)-Du(x)|^2}{h^2}\tau^2(x)\; dt\; dx \\ &\le 4C\int_{\Omega}\int_0^1(1+|tDu(x+he_k)+(1-t)Du(x))|^2)^{\frac{q-2}{2}}\frac{|u(x+he_k)-u(x)|^2}{h^2}|D\tau(x)|^2\; dt \; dx. \end{split}$$ Finally, we let $h\to 0_+$. First, we focus on the limit in the term on the right hand side of . In case that $q\le 2$, we use the assumption that $u\in W^{1,2}_{\rm loc}(\Omega; \mathbb{R}^N)$ and therefore, we can use the Lebesgue dominated convergence theorem to conclude that $$\begin{aligned} \limsup_{h\to 0}&\int_{\Omega}\int_0^1(1+|tDu(x+he_k)+(1-t)Du(x))|^2)^{\frac{q-2}{2}}\frac{|u(x+he_k)-u(x)|^2}{h^2}|D\tau(x)|^2\; dt \; dx\\ &=\int_{\Omega}(1+|Du|^2)^{\frac{q-2}{2}}|u_{x_k}|^2|D\tau|^2 \; dx\le \int_{\Omega}(1+|Du|^2)^{\frac{q}{2}}|D\tau|^2 \; dx. \end{aligned}$$ Next, if $q>2$, we use the Hölder inequality, the assumption $u\in W^{1,q}_{\rm loc}(\Omega; \mathbb{R}^N)$ and the Lebesgue dominated convergence theorem to conclude $$\begin{aligned} \limsup_{h\to 0}&\int_{\Omega}\int_0^1(1+|tDu(x+he_k)+(1-t)Du(x))|^2)^{\frac{q-2}{2}}\frac{|u(x+he_k)-u(x)|^2}{h^2}|D\tau(x)|^2\; dt \; dx\\ &= \limsup_{h\to 0}\int_{\Omega}\int_0^1\left(((1+|tDu(x+he_k)+(1-t)Du(x))|^2)^{\frac{q-2}{2}}|D\tau(x)|^{2\frac{q-2}{q}}\right)\cdot \\ &\qquad \cdot\frac{|u(x+he_k)-u(x)|^2}{h^2}|D\tau(x)|^{\frac{4}{q}}\; dt \; dx\\ &\le \limsup_{h\to 0}\int_0^1 \left(\int_{\Omega}((1+|tDu(x+he_k)+(1-t)Du(x))|^2)^{\frac{q}{2}}|D\tau(x)|^{2}\; dx \right)^{\frac{q-2}{q}}\cdot \\ &\qquad \left(\int_{\Omega}\frac{|u(x+he_k)-u(x)|^q}{h^q}|D\tau(x)|^{2} \; dx\right)^{\frac{2}{q}}\; dt\\ &\le \int_{\Omega}(1+|Du|^2)^{\frac{q}{2}}|D\tau|^2 \; dx. \end{aligned}$$ Consequently, substituting these limits into , we have $$\begin{split}\label{huz2} &\limsup_{h\to 0}\int_{\Omega}\int_0^1 (1+|tDu(x+he_k)+(1-t)Du(x))|^2)^{\frac{p-2}{2}}\frac{|Du(x+he_k)-Du(x)|^2}{h^2}\tau^2(x)\; dt\; dx \\ &\le 4C\int_{\Omega}(1+|Du|^2)^{\frac{q}{2}}|D\tau|^2 \; dx. \end{split}$$ &gt;From this estimate it immediately follows that $u\in W^{2,\min\{2,p\}}_{\rm loc}(\Omega; \mathbb{R}^N)$, in particular we know that $D^2u$ exists and that for almost all $x$ $$\frac{Du(x+he_k)-Du(x)}{h}\to (D^2 u)_{x_k}(x)$$ where $(D^2 u)_{x_k}$ stands for $\frac{\partial Du}{\partial x_k}$. Therefore, we can use the Fatou lemma in to conclude $$\begin{split} &\int_{\Omega}(1+|Du|^2)^{\frac{p-2}{2}}| D^2 u_{x_k}(x)|^2\tau^2\; dx\le 4C\int_{\Omega}(1+|Du|^2)^{\frac{q}{2}}|D\tau|^2 \; dx. \end{split}$$ Since $k$ is arbitrary, the relation obviously follows. In addition, using the following algebraic inequality $$|DV(Du)|^2\le K(1+|Du|^2)^{\frac{p-2}{2}}|D^2u|^2,$$ we see that holds as well. Hence the proof is complete. Proof of Theorem \[t:main\] {#P-t:main} =========================== We shall start by recalling the definition of the Sobolev embedding exponent $$\label{2star} 2^{*}=\left\{\begin{array}{ll}\frac{2n}{n-2}&\text{if $n\ge 3$}\\ \text{arbitrary $>2$} & \text{if $n=2$}.\end{array}\right.$$ The value $2^*$ in dimension $n=2$ will be finally chosen sufficiently large. Since $u$ is assumed to be a weak solution belonging to $W^{1,\max\{q,2\}}_{\rm loc}(\Omega;\mathbb{R}^N)$, we can use Theorem \[T:main-weak\] and after summing and , we obtain the starting inequality valid for all $\tau \in \mathcal{C}^{\infty}_c (\Omega)$ $$\label{split} \int_{\Omega}\left((1+|Du|^2)^{\frac{p-2}{2}}|D^2u|^2\tau^2 + |DV(Du)|^2\tau^2\right) \; dx \le K\int_{\Omega}(1+|Du|^2)^{\frac{q}{2}}|D\tau|^2\; dx$$ Moreover, we remark that $$\label{e:Lq} \int_{\Omega}(1+|Du|^2)^{\frac{q}{2}}|D\tau|^2\; dx\le c\int_{\Omega}\left( (1+|Du|^2)^{\frac{p}{2}}+|V(Du)|^{\frac{2q}{p}}\right)|D\tau|^2\,dx.$$ Indeed, in $\{|Du|\le 1\}$ we have $$(1+|Du|^2)^{\frac{q}{2}}\le 2(1+|Du|^2)^{\frac{p}{2}}$$ and, in $\{|Du|> 1\}$, $$(1+|Du|^2)^{\frac{q}{2}}= \left\{(1+|Du|^2)^{\frac{p-2}{2}}(1+|Du|^2)\right\}^{\frac{q}{p}}\le 2^{\frac{q}{p}} |V(Du)|^{\frac{2q}{p}}.$$ Next, we split the proof for the case i) and ii). The case $q<p\frac{n+2}{n}$ --------------------------- In this case, we first use the Sobolev embedding to conclude that (with some $C$ depending on $2^*$) $$\begin{aligned} \|V(Du)\tau\|_{2^*}^2 &\le C\|D(V(Du)\tau)\|_2^2 \le 2C\int_{\Omega} \left(|DV(Du)|^2\tau^2 + |V(Du)|^2|D\tau|^2\right)\; dx \\ &\le 2C\int_{\Omega}\left(|DV(Du)|^2\tau^2 + (1+|Du|^2)^{\frac{p}{2}}|D\tau|^2\right)\; dx. \end{aligned}$$ Using this inequality in , and taking into account we get $$\label{split2} \begin{split} &\|V(Du)\tau\|_{2^*}^2+ \int_{\Omega}\left((1+|Du|^2)^{\frac{p-2}{2}}|D^2u|^2\tau^2 + |DV(Du)|^2\tau^2\right) \; dx \\ &\quad\le K_1\int_{\Omega}\left((1+|Du|^2)^{\frac{p}{2}}|D\tau|^2+(1+|Du|^2)^{\frac{q}{2}}|D\tau|^2\right)\; dx\\ &\quad\le K_2\int_{\Omega}\left((1+|Du|^2)^{\frac{p}{2}}|D\tau|^2+|V(Du)|^{\frac{2q}{p}}|D\tau|^2\right)\; dx. \end{split}$$ In particular, we have that $V(Du)\in L_{\rm loc}^{2^*}$. Let us now estimate the last integral on the right hand side. Since $q\in (p,p\frac{2^*}{2})$, which follows from the assumption that $q<p \frac{n+2}{n}$ (note here that the value of $2^*$ in dimension $n=2$ has to be chosen greater than $\frac{2q}{p}$), there exists a unique $\theta\in (0,1)$ such that $$\frac{q}{2}=\frac{p}{2}(1-\theta)+\frac{p2^*}{4}\theta, \qquad \theta:=\frac{q-p}{p(\frac{2^*}{2}-1)}.$$ As we will prove below, under our assumptions on the exponents $p$ and $q$ and, if $n=2$, with a suitable choice of $2^*$, we have $$2>2^*\theta. \label{asss}$$ Consider $\eta\in \mathcal{C}^{\infty}_c(\Omega)$ an arbitrary nonnegative cut-off function and set $$\tau:=\eta^{\gamma} \qquad \textrm{with } \gamma:=\frac {2}{ 2-2^*\theta}.$$ We have that $$\label{eta-tau} \frac{|D\tau|^2}{\tau^{2^*\theta}}= \gamma^2\eta^{\gamma(2-2^*\theta)-2}|D \eta|^2= \gamma^2|D\eta|^2.$$ Then by the Hölder inequality, we have $$\begin{aligned} \int_{\Omega} |V(Du)|^{\frac{2q}{p}}|D\tau|^2\,dx&=\int_{\Omega} |V(Du)|^{2(1-\theta)}(V(Du)\tau)^{2^*\theta} \frac{|D\tau|^2}{\tau^{2^*\theta}}\,dx\\ &\le \|V(Du)\|_2^{2(1-\theta)}\|V(Du)\tau\|_{2^*}^{2^*\theta} \left\|\frac{|D\tau|^2}{\tau^{2^*\theta}}\right\|_{\infty} \end{aligned}$$ and we can apply the Young ’s inequality to deduce that for arbitrary $\varepsilon>0$ we have $$\label{ssop} \begin{aligned} \int_{\Omega} |V(Du)|^{\frac{2q}{p}}|D\tau|^2\,dx&\le \varepsilon \|V(Du)\tau\|_{2^*}^2 + C(\varepsilon,\gamma) \|V(Du)\|_2^{2(1-\theta)\gamma} \left\|\frac{|D\tau|^2}{\tau^{2^*\theta}}\right\|^{\gamma}_{\infty}. \end{aligned}$$ Therefore, combining , , and taking into account that $V(Du)\le (1+|Du|^2)^{\frac{p}{4}}$, with a proper choice of $\varepsilon>0$, we obtain $$\begin{split} &\|V(Du)\tau\|_{2^*}^2+\int_{\Omega}\left((1+|Du|^2)^{\frac{p-2}{2}}|D^2u|^2\eta^{2\gamma} + |DV(Du)|^2\eta^{2\gamma}\right) \; dx \\ &\quad\le K(\gamma,\|D \eta\|_{\infty}) \left(\int_{\Omega}(1+|Du|^2)^{\frac{p}{2}}\; dx\right)^{\tilde{q}}, \end{split}$$ with some power $\tilde{q}$ whose value depends on $p,q$ and $\gamma$. &gt;From this inequality the statement i) of Theorem \[t:main\] follows directly. Now, we check the validity of , which, by using of definition of $\theta$, it can be written as $$q<2p\left(1-\frac{1}{2^*}\right).$$ If $n=2$, we can choose $2^*$ arbitrarily large, therefore in this case the condition reduces to $q<2p$, which is exactly the assumption for $n=2$. If $n\ge 3$ we have $2^*=2n/(n-2)$ and the above condition is then equivalent to $$q<p\frac{n+2}{n},$$ which is nothing else than the assumption . Hence the proof of the statement i) is finished. The case $q<p+2$ and $p<n$ -------------------------- We again start to estimate the integral on the right hand side of . Using a simple inequality and the integration by parts, we find that (here $K$ is again a generic constant depending only on $q$) $$\begin{aligned} \nonumber \int_{\Omega} &(1+|Du|^2)^{\frac{q}{2}} |D\tau|^2\; dx \le K+ \sum_{k=1}^n\sum_{\alpha=1}^N \int_{\Omega}(1+|Du|^2)^{\frac{q-2}{2}} u^{\alpha}_{x_k}u^{\alpha}_{x_k}|D\tau|^2\,dx\\ \nonumber &= K-\sum_{k=1}^n\sum_{\alpha=1}^N \int_{\Omega}\left((1+|Du|^2)^{\frac{q-2}{2}} u^{\alpha}_{x_k}|D\tau|^2 \right)_{x_k}u^{\alpha}\; dx\\ &\le K+K\|u\|_{\infty} \int_{\Omega}\left((1+|Du|^2)^{\frac{q-2}{2}} |D^2 u||D\tau|^2 +(1+|Du|^2)^{\frac{q-1}{2}}|D\tau||D^2\tau|\right)\; dx. \label{e:stima}\end{aligned}$$ Let us now set $\tau:=\eta^{\gamma}$, $\gamma\ge 2$ to be chosen later, where $\eta\in \mathcal{C}^{\infty}_c(\Omega)$ is an arbitrary nonnegative cut-off function. By the Young ’s inequality, $$\begin{aligned} \nonumber & K\|u\|_{\infty} \int_{\Omega}(1+|Du|^2)^{\frac{q-2}{2}} |D^2 u||D\tau|^2\,dx \\&= \int_{ \Omega} \left\{(1+|Du|^2)^{\frac{p-2}{4}}|D^2u|\tau \right\}\left\{\|u\|_{\infty} (1+|Du|^2)^{\frac{2q-p-2}{4}}\frac{|D\tau|^2}{\tau} \right\} \,dx \nonumber \\&\le \varepsilon \int_{\Omega}(1+|Du|^2)^{\frac{p-2}{2}}|D^2u|^2\tau^2\; dx + c_{\varepsilon,K} \|u\|^2_{\infty} \int_{\Omega}(1+|Du|^2)^{\frac{2q-p-2}{2}}\frac{|D\tau|^4}{\tau^2}\; dx. $$ Therefore, $$\label{inter} \begin{split} &\int_{\Omega} (1+|Du|^2)^{\frac{q}{2}} |D\tau|^2\; dx \le \varepsilon \int_{\Omega}(1+|Du|^2)^{\frac{p-2}{2}}|D^2u|^2\tau^2\; dx \\ &\quad + K+K\|u\|^2_{\infty} \int_{\Omega}(1+|Du|^2)^{\frac{2q-p-2}{2}}\frac{|D\tau|^4}{\tau^2}\; dx + K\|u\|_{\infty}\int_{\Omega}(1+|Du|^2)^{\frac{q-1}{2}}|D\tau||D^2\tau|\; dx, \end{split}$$ with a possibly different positive constant $K$ than before. Let us now discuss first the case $q\in [p,p+1]$. If $q$ belongs to this range, the above inequality immediately reduces to $$\begin{split} &\int_{\Omega} (1+|Du|^2)^{\frac{q}{2}} |D\tau|^2\; dx \le \varepsilon \int_{\Omega}(1+|Du|^2)^{\frac{p-2}{2}}|D^2u|^2\tau^2\; dx \\ &\quad + K+K\|u\|^2_{\infty} \int_{\Omega}(1+|Du|^2)^{\frac{p}{2}}\frac{|D\tau|^4}{\tau^2}\; dx +K\|u\|_{\infty}\int_{\Omega}(1+|Du|^2)^{\frac{p}{2}}|D\tau||D^2\tau|\; dx. \end{split}$$ Let us now choose $\gamma=2$, that is $\tau:=\eta^2$. Thus we get $$\label{intera} \begin{split} &\int_{\Omega} (1+|Du|^2)^{\frac{q}{2}} |D\tau|^2\; dx \le \varepsilon \int_{\Omega}(1+|Du|^2)^{\frac{p-2}{2}}|D^2u|^2\tau^2\; dx \\ &\quad + K+C(\|u\|_{\infty}, \|\eta\|_{2,\infty}) \int_{\Omega}(1+|Du|^2)^{\frac{p}{2}}\; dx. \end{split}$$ Hence by and taking a proper $\varepsilon>0$, so that we can absorb the first term on the right hand side in by the left hand side in , it is not difficult to arrive to the statement ii) of Theorem \[t:main\] for $q\in [p,p+1]$. Next, we focus on the case when $q\in (p+1,p+2)$. There exist $\theta_1,\theta_2\in (0,1)$ such that $$\begin{aligned} q-1&=p(1-\theta_1)+q\theta_1, && \theta_1:= \frac{q-1-p}{q-p},\\ 2q-p-2&=p(1-\theta_2)+ q\theta_2, && \theta_2:=\frac{2(q-1-p)}{q-p}.\end{aligned}$$ In addition, considering $\tau=\eta^\gamma$ with $$\label{special} \gamma= \frac{2-\theta_2}{1-\theta_2},$$ we have that $\frac{|D\tau|^{2-\theta_2}}{\tau}=\gamma |D\eta|^{2-\theta_2}$. With this setting, we can now estimate the remaining integrals on the right hand side of by means of the Hölder inequality as follows $$\begin{aligned} &\int_{\Omega}(1+|Du|^2)^{\frac{2q-p-2}{2}}\frac{|D\tau|^4}{\tau^2}\; dx = \int_{\Omega}\left((1+|Du|^2)^{\frac{q}{2}}|D\tau|^2\right)^{\theta_2}(1+|Du|^2)^{\frac{p(1-\theta_2)}{2}}\frac{|D\tau|^{4-2\theta_2}}{\tau^2}\; dx\\ &\quad\le C\left\|\frac{|D\tau|^{4-2\theta_2}}{\tau^2}\right\|_{\infty}\|(1+|Du|)\|_p^{p(1-\theta_2)} \left(\int_{\Omega}(1+|Du|^2)^{\frac{q}{2}}|D\tau|^2\; dx\right)^{\theta_2}. \end{aligned}$$ Then the above estimate reduces to $$\label{godot} \begin{aligned} &\int_{\Omega}(1+|Du|^2)^{\frac{2q-p-2}{2}}\frac{|D\tau|^4}{\tau^2}\; dx \\ &\qquad \le C(\theta_2,\|\eta\|_{1,\infty})\|(1+|Du|)\|_p^{p(1-\theta_2)} \left(\int_{\Omega}(1+|Du|^2)^{\frac{q}{2}}|D\tau|^2\; dx\right)^{\theta_2}. \end{aligned}$$ We proceed similarly also with the remaining integral in , i.e., using the Hölder inequality, we have $$\label{godot2} \begin{aligned} &\int_{\Omega}(1+|Du|^2)^{\frac{q-1}{2}}|D\tau||D^2\tau|\; dx\\ &\quad =\int_{\Omega}(1+|Du|^2)^{\frac{p(1-\theta_1)}{2}}\left((1+|Du|^2)^{\frac{q}{2}}|D\tau|^2\right)^{\theta_1}|D\tau|^{1-2\theta_1}|D^2\tau|\; dx\\ &\quad\le K\|(1+|Du|)\|_p^{p(1-\theta_1)} \||D\tau|^{1-2\theta_1}|D^2\tau|\|_{\infty} \left(\int_{\Omega}(1+|Du|^2)^{\frac{q}{2}}|D\tau|^2\; dx \right)^{\theta_1}\\ &\quad\le C(\|\tau\|_{2,\infty},\theta_2)\|(1+|Du|)\|_p^{p(1-\theta_1)} \left(\int_{\Omega}(1+|Du|^2)^{\frac{q}{2}}|D\tau|^2\; dx \right)^{\theta_1}, \end{aligned}$$ where the last inequality follows from the fact that $1-2\theta_1=1-\theta_2>0$. Finally, using and in , keeping in mind the special choice of $\tau$ in and applying the Young’s inequality (notice that $\theta_1,\theta_2<1$) we observe that $$\begin{split} \int_{\Omega} (1+|Du|^2)^{\frac{q}{2}} |D\tau|^2\; dx &\le \varepsilon \int_{\Omega}(1+|Du|^2)^{\frac{p-2}{2}}|D^2u|^2\tau^2\; dx \\ &\quad + C(\varepsilon,\theta_2,\|\eta\|_{2,\infty}, \|u\|_{\infty}, \|u\|_{1,p}). \end{split}$$ Thus, going back to , choosing $\varepsilon>0$ sufficiently small to absorb the term involving the second derivatives by the left hand side, we finally get the statement ii) of Theorem \[t:main\]. Proof of Theorem \[t:main2\] ============================ In this final section we establish the existence of a weak solution to the Dirichlet problem , under the assumption on the boundary datum $u_0$; i.e., $$u_0 \in W^{1,r} (\Omega;\mathbb{R}^N), \qquad r:=\max\left\{2,p\frac{q-1}{p-1}\right\}.$$ We use an approximation procedure. For arbitrary $\epsilon \in (0,1)$ we introduce the approximate problem ($\alpha=1,\ldots,N$) $$\label{systemepsilon} \left\{\begin{array}{ll}\sum_{i=1}^n\frac{\partial }{\partial x_i}\left(A_{\epsilon,i}^{\alpha}(Du_{\epsilon})\right)=0 &\textrm{in } \Omega,\\ u_{\epsilon}=u_0 &\textrm{on }\partial \Omega, \end{array}\right.$$ where $A_{\epsilon,i}^{\alpha}: \mathbb{R}^{nN}\to \mathbb{R}$ is defined as $$\label{Aepsilon}A_{\epsilon,i}^{\alpha}(\xi):=A_{i}^{\alpha}(\xi)+\epsilon (1+|\xi|^2)^{\frac{\max\{q,2\}-2}{2}}\xi_i^{\alpha}.$$ In addition, in case we deal with the statement ii) of the theorem, we shall require that $u_0\in L^{\infty}(\partial \Omega)$. Due to , and we have that $A_{\epsilon,i}^{\alpha}(\xi)$ satisfies the following properties: $$\sum_{i=1}^n\sum_{\alpha=1}^N A_{\epsilon,i}^{\alpha}(\xi)\xi_i^{\alpha}\ge \epsilon |\xi|^{\max\{q,2\}}-\lambda,$$ $$\label{e:crescita-eps} |A_{\epsilon,i}^{\alpha}(\xi)|\le M'(1+|\xi|)^{\max\{q,2\}-1}.$$ for some positive $\lambda$ and $M'$ independent on $\epsilon$. We can apply the theory of monotone operators (see e.g. [@lerlio65; @browder; @hart-stamp]) to prove the existence of a unique solution to , i.e., the existence of $u_{\epsilon}\in u_0+W^{1,\max\{q,2\}}(\Omega; \mathbb{R}^N)$ fulfilling $$\int_{\Omega}\sum_{\alpha=1}^N \sum_{i=1}^n{A_{\epsilon,i}^{\alpha}}(Du_{\epsilon})\varphi_{x_i}^\alpha\,dx=0\qquad \forall \varphi\in W_0^{1,\max\{q,2\}}(\Omega;\mathbb{R}^N). \label{firstvariation}$$ First a priori estimates ------------------------ We now derive estimates for $u_{\epsilon}$ independent of $\epsilon$. Using $\varphi:=u_{\epsilon}-u_0$ as a test function in , we get $$\label{e:begin} \begin{split} 0&=\int_{\Omega}\sum_{\alpha=1}^N \sum_{i=1}^n{A_{\epsilon,i}^{\alpha}}(Du_{\epsilon})((u_{\epsilon})_{x_i}^{\alpha}-(u_0)_{x_i}^{\alpha})\; dx\\ &=\int_{\Omega}\sum_{\alpha=1}^N \sum_{i=1}^n\left\{{A_{i}^{\alpha}}(Du_{\epsilon})((u_{\epsilon})_{x_i}^{\alpha}-(u_0)_{x_i}^{\alpha})+ \epsilon(1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}-2}{2}} (u_{\epsilon})^{\alpha}_{x_i}((u_{\epsilon})_{x_i}^{\alpha}-(u_0)_{x_i}^{\alpha})\right\}\; dx\\ &\overset{\eqref{dis-ellitticita}}\ge \int_{\Omega}\left(K^{-1}|Du_{\epsilon}|^p -(1+|Du_0|^2)^{\frac{p(q-1)}{2(p-1)}} + \epsilon(1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}-2}{2}}|Du_{\epsilon}|(|Du_{\epsilon}|-|Du_0|)\right)\; dx. \end{split}$$ Since $$\begin{split}(1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}-2}{2}}|Du_{\epsilon}|(|Du_{\epsilon}|-|Du_0|)=& (1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}-2}{2}}|Du_{\epsilon}|^2\\ &-(1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}-2}{2}}|Du_{\epsilon}||Du_0|,\end{split}$$ then implies $$\begin{aligned} \nonumber &\int_{\Omega}\left(|Du_{\epsilon}|^p +\epsilon (1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}-2}{2}}|Du_{\epsilon}|^2\right)\; dx\\ &\le c \int_{\Omega}\left((1+|Du_0|^2)^{\frac{p(q-1)}{2(p-1)}}+\epsilon(1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}-2}{2}}|Du_{\epsilon}||Du_0|\right) \,dx. \label{e:begin2}\end{aligned}$$ We claim that implies $$\int_{\Omega}\left(|Du_{\epsilon}|^p + \frac{\epsilon}{2} (1+|Du_{\epsilon}|^2)^{\frac{\max\{2,q\}-2}{2}}|Du_{\epsilon}|^2\right)\; dx\le c \int_{\Omega}(1+|Du_0|^2)^{\frac{r}{2}}\,dx \label{goal}$$ If $q\le 2$, we can conclude using Young’s inequality with exponent $\frac12$ on the last term in : $$\begin{aligned} &|Du_{\epsilon}||Du_0|\le \frac{1}{2c} |Du_{\epsilon}|^2+c' |Du_0|^2.\end{aligned}$$ Therefore, recalling that $r=\max\{2,\frac{p(q-1)}{p-1}\}$ the inequality follows. Otherwise, if $q>2$, the last term in can be estimate as follows: $$\epsilon(1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}-2}{2}}|Du_{\epsilon}||Du_0| \le \epsilon\left\{c(1+|Du_0|^2)^{\frac{r}{2}}+c(1+|Du_{\epsilon}|^2)^{\frac{q-2}{4}} |Du_{\epsilon}|^{\frac{q}{2}} |Du_0|\right\}.\label{e:second}$$ Indeed, in $\{|Du_{\epsilon}|\le 1\}$ we have $$(1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}-2}{2}}|Du_{\epsilon}||Du_0|\le 2^{\frac{q-2}{2}}|Du_0|\le c(1+|Du_0|^2)^{\frac{r}{2}}$$ and, in $\{|Du_{\epsilon}|> 1\}$, $$(1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}-2}{2}}|Du_{\epsilon}||Du_0|\le 2^{\frac{q-2}{4}}(1+|Du_{\epsilon}|^2)^{\frac{q-2}{4}}|Du_{\epsilon}|^{\frac{q}{2}}|Du_0|$$ and follows. To estimate the last term in , we use Young’s inequality with exponents $2, q ,\frac{2q}{q-2}$. Recalling that $\epsilon<1$, we have $$\begin{aligned} \nonumber & \epsilon c(1+|Du_{\epsilon}|^2)^{\frac{q-2}{4}} |Du_{\epsilon}|^{\frac{q}{2}} |Du_0|=\epsilon c\left\{(1+|Du_{\epsilon}|^2)^{\frac{q-2}{4}} |Du_{\epsilon}|\right\} |Du_{\epsilon}|^{\frac{q-2}{2}} |Du_0|\\ &\le\nonumber \frac{\epsilon}{8} (1+|Du_{\epsilon}|^2)^{\frac{q-2}{2}}|Du_{\epsilon}|^2+ \frac{\epsilon}{8}|Du_{\epsilon}|^{q} + \epsilon c|Du_0|^q \\ &\le \frac{\epsilon}{4} (1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}-2}{2}}|Du_{\epsilon}|^2 + c(1+|Du_0|^2)^{\frac{r}{2}} \label{e:fourth}\end{aligned}$$ with $c$ independent of $\epsilon$. Therefore, collecting , and , the inequality follows also in the case $q>2$. Thus, we can find a universal constant $C>0$ such that (using also the Poincaré inequality) $$\|u_{\epsilon}\|_{1,p} + \epsilon \|u_{\epsilon}\|^{\max\{q,2\}}_{1,\max\{q,2\}}\le C. \label{f-ap}$$ If the assumption holds, then for every $\alpha\in \{1,\ldots, N\}$ we have $$\label{BulHLP} \sum_{i=1}^nA_{\epsilon,i}^{\alpha}(\xi)\xi_i^{\alpha}\ge \epsilon (1+|\xi|^{2})^{\frac{\max\{2,q\}-2}{2}}|\xi^{\alpha}|^2\ge \epsilon |\xi^{\alpha}|^{\max\{q,2\}}$$ and $$|A_{\epsilon,i}^{\alpha}(\xi)|\le (K+1)(1+|\xi|^{2})^{\frac{\max\{2,q\}-1}{2}},$$ where $K$ is as in . Next we denote $\tilde{M}:=\|u_0\|_{L^{\infty}(\partial \Omega)}$ and define $$\varphi^{\alpha}:=\max \{u_{\epsilon}^{\alpha}-\tilde{M},0\} \qquad \alpha\in \{1,\ldots, N\}.$$ Evidently, $\varphi=(\varphi^1,\ldots,\varphi^N) \in W^{1,\max \{2,q\}}_0(\Omega;\mathbb{R}^N)$ and can be used as a test function in . Doing so, and using the definition of $\varphi$ we obtain (here $\chi_{u_{\epsilon}^{\alpha}\ge \tilde{M}}$ denotes the characteristic function of the set, where $u^{\alpha}_{\epsilon}\ge \tilde{M}$) $$0=\int_{\Omega}\sum_{\alpha=1}^N \sum_{i=1}^n{A_{\epsilon,i}^{\alpha}}(Du_{\epsilon})\varphi_{x_i}^\alpha\,dx=\int_{\Omega}\sum_{\alpha=1}^N \sum_{i=1}^n{A_{\epsilon,i}^{\alpha}}(Du_{\epsilon})Du_{\varepsilon}^{\alpha}\chi_{u_{\epsilon}^{\alpha}\ge \tilde{M}}\,dx \label{firstvariationM}$$ Using finally , we see that $$\begin{split} 0&=\int_{\Omega}\sum_{\alpha=1}^N \sum_{i=1}^n{A_{\epsilon,i}^{\alpha}}(Du_{\epsilon})Du_{\varepsilon}^{\alpha}\chi_{u_{\epsilon}^{\alpha}\ge \tilde{M}}\,dx\ge \epsilon \int_{\Omega}\sum_{\alpha=1}^N |Du_{\epsilon}^{\alpha}|^{\max\{q,2\}}\chi_{u_{\epsilon}^{\alpha}\ge \tilde{M}}\,dx\\ &=\epsilon\int_{\Omega}\sum_{\alpha=1}^N |D\varphi^{\alpha}|^{\max\{q,2\}}\,dx. \label{firstvariationMM} \end{split}$$ Consequently, $\varphi$ is a constant function. Since it has zero trace, it must be identically zero and it directly follows from its definition that $u_{\varepsilon}^{\alpha}\le \tilde{M}=\|u_0\|_{L^{\infty}(\partial \Omega)}$ for all $\alpha\in \{1,\ldots, N\}$. The minimum principle can be obtained by repeating step by step the above procedure for a test function defined as $$\varphi^{\alpha}:=\min \{u_{\epsilon}^{\alpha}+\tilde{M},0\} \qquad \alpha\in \{1,\ldots, N\}.$$ Therefore, we conclude that, for every $\epsilon\in (0,1)$, $$\label{apest3} \|u_{\epsilon}\|_{L^{\infty}(\Omega)}\le \|u_0\|_{L^{\infty}(\partial \Omega)}.$$ Uniform higher order estimates ------------------------------ Due to the proof of a priori estimates we can use Theorem \[T:main-weak\] to get the existence of the second order derivatives of $u_{\epsilon}$, but with their estimates depending on $\epsilon$. Nevertheless, we can repeat step by step the estimates in Theorem \[T:main-weak\] to get the following inequality $$\label{hot} \begin{aligned} &\int_{\Omega}(1+|Du_{\epsilon}|^2)^{\frac{p-2}{2}}|D^2u_{\epsilon}|^2\tau^2\; dx \le c\int_{\Omega}(1+|Du_{\epsilon}|^2)^{\frac{q}{2}}|D\tau^2|\; dx \\ &\qquad -c\epsilon\int_{\Omega}\sum_{i,k=1}^n\sum_{\alpha=1}^N \left((1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}-2}{2}}(u_{\epsilon})^{\alpha}_{x_i}\right)_{x_k} \left((u_{\epsilon})^{\alpha}_{x_k}\tau^2 \right)_{x_i}\; dx \end{aligned}$$ for every $\tau\in \mathcal{C}_c^{\infty}(\Omega)$. Thus, we need to bound uniformly the last integral. By a rather standard manipulation and using the Young inequality, it is not difficult to check that $$\begin{aligned} &\sum_{i,k=1}^n\sum_{\alpha=1}^N \left((1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}-2}{2}}(u_{\epsilon})^{\alpha}_{x_i}\right)_{x_k} \left((u_{\epsilon})^{\alpha}_{x_k}\tau^2 \right)_{x_i}\\ &\qquad \ge (1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}-2}{2}}|D^2u_{\epsilon}|^2\tau^2 - 2\max\{q,2\} (1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}-2}{2}}|D^2u_{\epsilon}|\tau |Du_{\epsilon}| |D\tau|\\ &\qquad \ge -C(1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}}{2}}|D\tau|^2 \end{aligned}$$ with $C$ independent of $\epsilon$. Substituting this into , we derive $$\label{hot2} \begin{aligned} &\int_{\Omega}(1+|Du_{\epsilon}|^2)^{\frac{p-2}{2}}|D^2u_{\epsilon}|^2\tau^2\; dx \le c\int_{\Omega}(1+|Du_{\epsilon}|^2)^{\frac{q}{2}}|D\tau|^2\; dx\\ &\qquad +c\epsilon \int_{\Omega}(1+|Du_{\epsilon}|^2)^{\frac{\max\{q,2\}}{2}}|D\tau|^2\; dx\\ &\qquad \overset{\eqref{f-ap}}\le c +c\int_{\Omega}(1+|Du_{\epsilon}|^2)^{\frac{q}{2}}|D\tau|^2\; dx. \end{aligned}$$ Hence, we are in the same starting position as in the proof of Theorem \[t:main\] and due to uniform ($\epsilon$-independent) uniform bounds and , we deduce that for arbitrary open $\Omega'\subset \overline{\Omega'} \subset \Omega$, $$\int_{\Omega'}\left(|Du_{\epsilon}|^{q}+|D V(Du_{\epsilon})|^2+ (1+|Du_{\epsilon}|^2)^{\frac{p-2}{2}}|D^2u_{\epsilon}|^2\right)\, dx \le C(\Omega', u_0). \label{exdopoH3-V2ns}$$ Further, it is then not difficult to observe with the help of the Hölder inequality that $$\int_{\Omega'}|D^2u_{\epsilon}|^{\min\{2,p\}}\, dx \le C(\Omega', u_0). \label{exdopoH3-V2ns3}$$ Limit $\epsilon \to 0$ ---------------------- Using the uniform bounds , and , the compact Sobolev embedding and the diagonal procedure, we can find a subsequence, that we do not relabel, and it exists $$u\in (u_0+W^{1,p}(\Omega; \mathbb{R}^N))\cap W_{\rm loc}^{1,q}(\Omega; \mathbb{R}^N)$$ such that for arbitrary open $\Omega'\subset \overline{\Omega'} \subset \Omega$, we have $$\begin{aligned} u^{\epsilon}&\rightharpoonup u &&\textrm{weakly in }W^{1,p}(\Omega; \mathbb{R}^N),\label{co1}\\ u^{\epsilon}&\rightharpoonup u &&\textrm{weakly in }W^{1,q}(\Omega'; \mathbb{R}^N),\label{co2}\\ Du^{\epsilon}&\to Du &&\textrm{strongly in }L^p(\Omega; \mathbb{R}^N),\label{co3}\\ Du^{\epsilon}&\to Du &&\textrm{almost everywhere in }\Omega,\label{co4}\\ \epsilon (1+|Du^{\epsilon}|^2)^{\frac{\max(2,q)-2}{2}}Du^{\epsilon}&\to 0 &&\textrm{strongly in }L^1(\Omega'; \mathbb{R}^{nN}).\label{co5}\end{aligned}$$ Having –, it is easy to let $\epsilon \to 0$ in with arbitrary $\varphi \in \mathcal{C}^{\infty}_c(\Omega; \mathbb{R}^N)$ to deduce for the same class of $\varphi$’s. The density result then leads to the validity of in the full generality. This finishes the proof. [10]{} Michael Bildhauer. , volume 1818 of [*Lecture Notes in Mathematics*]{}. Springer-Verlag, Berlin, 2003. Michael Bildhauer and Martin Fuchs. -solutions to non-autonomous anisotropic variational problems. , 24(3):309–340, 2005. Felix E. Browder. Nonlinear monotone operators and convex sets in [B]{}anach spaces. , 71:780–785, 1965. Menita Carozza, Jan Kristensen, and Antonia Passarelli di Napoli. Higher differentiability of minimizers of convex variational integrals. , 28(3):395–411, 2011. Menita Carozza, Jan Kristensen, and Antonia Passarelli di Napoli. Regularity of minimizers of autonomous convex variational integrals. , 13(4):1065–1089, 2014. Giovanni Cupini, Francesco Leonetti, and Elvira Mascolo. Existence of weak solutions for elliptic systems with [$p,q$]{}-growth. , 40(2):645–658, 2015. Giovanni Cupini, Paolo Marcellini, and Elvira Mascolo. Existence and regularity for elliptic equations under [$p,q$]{}-growth. , 19(7-8):693–724, 2014. Ennio De Giorgi. Un esempio di estremali discontinue per un problema variazionale di tipo ellittico. , 1:135–137, 1968. Luca Esposito, Francesco Leonetti, and Giuseppe Mingione. Higher integrability for minimizers of integral functionals with $(p,q)$ growth. , 157(2):414–438, 1999. Luca Esposito, Francesco Leonetti, and Giuseppe Mingione. Regularity results for minimizers of irregular integrals with $(p,q)$ growth. , 14(2):245––272, 2002. Luca Esposito, Francesco Leonetti, and Giuseppe Mingione. Sharp regularity for functionals with $(p,q)$ growth. , 204(1):5–55, 2004. Mariano Giaquinta. , volume 105 of [*Annals of Mathematics Studies*]{}. Princeton University Press, Princeton, NJ, 1983. Mariano Giaquinta. Growth conditions and regularity, a counterexample. , 59(2):245–248, 1987. Enrico Giusti. . World Scientific Publishing Co., Inc., River Edge, NJ, 2003. Philip Hartman and Guido Stampacchia. On some non-linear elliptic differential-functional equations. , 115:271–310, 1966. Min Chun Hong. Some remarks on the minimizers of variational integrals with nonstandard growth conditions. , 6(1):91–101, 1992. Francesco Leonetti. Weak differentiability for solutions to nonlinear elliptic systems with [$p,q$]{}-growth conditions. , 162:349–366, 1992. Jean Leray and Jacques-Louis Lions. Quelques résultats de [V]{}išik sur les problèmes elliptiques nonlinéaires par les méthodes de [M]{}inty-[B]{}rowder. , 93:97–107, 1965. J. Málek, J. Nečas, M. Rokyta, and M. Ržička. . Chapman & Hall, London, 1996. P Marcellini. Un example de solution discontinue d’un problème variationnel dans le cas scalaire. Preprint 11, Istituto Matematico “U.Dini”, Università di Firenze, 1987. Paolo Marcellini. Regularity of minimizers of integrals of the calculus of variations with nonstandard growth conditions. , 105(3):267–284, 1989. Paolo Marcellini. Regularity and existence of solutions of elliptic equations with [$p,q$]{}-growth conditions. , 90(1):1–30, 1991. Giuseppe Mingione. Regularity of minima: an invitation to the dark side of the calculus of variations. , 51(4):355–426, 2006. Giuseppe Mingione. Singularities of minima: a walk on the wild side of the calculus of variations. , 40(1-3):209–223, 2008. Vladim[í]{}r [Š]{}ver[á]{}k and Xiaodong Yan. A singular minimizer of a smooth strongly convex functional in three dimensions. , 10(3):213–221, 2000. Peter Tolksdorf. Regularity for a more general class of quasilinear elliptic equations. , 51(1):126–150, 1984. [^1]: This information can be as usual slightly improved by the Gehring lemma

--- abstract: | We consider, in the setting of stratified groups $G,$ two analogues of the Ornstein-Uhlenbeck semi-group, namely Markovian diffusion semi-groups acting on $L^{q}(p(\gamma )d\gamma ),$ whose invariant density $p$ is a heat kernel at time 1 on $G.$ The first one is symmetric on $L^{2}(pd\gamma ),$ its generator is $\sum_{i=1}^{n}X_{i}^{\ast }X_{i},$ where $(X_{i})_{i=1}^{n}$  is a basis of the first layer of the Lie algebra of $G.$ The second one, denoted by $T_{t}=e^{-tN},t>0,$ is non symmetric on $L^{2}(pd\gamma )$ and the formal real part of $N$ is $\sum_{i=1}^{n}X_{i}^{\ast }X_{i}.$ The operators $e^{-tN}$ are compact on $L^{q}(pd\gamma ),$ $1<q<\infty .$ The spectrum of $N$ on this space is the set of integers $\mathbb{N}$ if polynomials are dense in $L^{2}(p(\gamma )d\gamma ),$ i.e if $G$ has at most 4 layers; and we determine in this case its eigenspaces. When $G$ is step 2, we give another description of these eigenspaces, very similar to the classical definition of “Hermite polynomials” by their generating function. *Keywords:* stratified groups, sub Laplacian, heat kernel measure, Ornstein-Uhlenbeck semi-groups. *MSC classification: 43A80, 47D06, 47D07.* author: - 'Françoise Lust-Piquard' title: 'Ornstein-Uhlenbeck semi-groups on stratified groups' --- **Introduction and notation** ============================= Let $G$ be a stratified Lie group equipped with its (biinvariant) Haar measure $dg$ and dilations $(\delta _{t})_{t\geq 0}$. Let $Q$ be the homogeneous dimension of $G$. We denote by $\mathcal{D}(G)$ the space of $\mathcal{C}^{\infty }$ compactly supported functions on $G,$ by $\mathcal{S}(G)$ the space of Schwartz functions, by $\mathcal{S}^{\prime }(G)$ its dual, and $L^{q}(\varphi dg)=L^{q}(G,\varphi dg)$ for a measurable non negative function $\varphi $. As usual, elements $Z$ of the Lie algebra $\mathcal{G}$ are identified with left invariant vector fields by $$(Zf)(g)=\frac{d}{dt}\mid _{t=0}f(g\exp tZ).$$ Let $L$ be a subLaplacian on $G,$ i.e. an operator on $\mathcal{S(}G\mathcal{)}$ defined by $$L=-\sum_{1}^{n}X_{i}^{2} \label{0}$$ where $(X_{i})_{1\leq i\leq n}$ is a linear basis of the first layer of $\mathcal{G}.$ Obviously $L$ commutes with left translations and satisfies $$\delta _{t^{-1}}L\delta _{t}=t^{2}L,\;t>0. \label{1}$$ The following facts are well known, see e.g. [@FS propositions 1.68, 1.70, 1.74]: $-\frac{L}{2}$ generates a strongly continuous semi-group $e^{-\frac{t}{2}L}$ of convolution operators which are contractions on $L^{q}(dg),$ $1\leq q\leq \infty .$ The kernel $p_{t}$ of $e^{-\frac{t}{2}L}$ is a positive function such that $p_{t}(g)=p_{t}(g^{-1}),$ it lies in $\mathcal{S(}G\mathcal{)}$ and has norm one in $L^{1}(dg).$ Denoting $p_{1}=p,$ $$p_{t}(g)=t^{-\frac{Q}{2}}p\circ \delta _{\frac{1}{\sqrt{t}}}(g).$$ Equivalently, for $f\in L^{q}(dg),$ $$e^{-\frac{t}{2}L}(f)(\gamma )=f\ast p_{t}(\gamma )=\int_{G}f(\gamma g^{-1})p_{t}(g)dg=\int_{G}f(\gamma \delta _{\sqrt{t}}g^{-1})p(g)dg. \label{2}$$ The aim of this paper is to generalize the Ornstein-Uhlenbeck semi-group in the setting of stratified groups, namely to consider Markovian semi-groups acting on $L^{q}(p(\gamma )d\gamma ),1\leq q\leq \infty ,$ for which $p(\gamma )d\gamma $ is an invariant measure, whose generators are related to the first layer gradient $$\nabla =(X_{1},..,X_{n}).$$ The classical Ornstein-Uhlenbeck semi-group is defined on $\mathcal{S}(\mathbb{R}^{n})$ by Mehler formula $$e^{-tN_{0}}(f)(x)=\int_{\mathbb{R}^{n}}f(e^{-t}x+\sqrt{1-e^{-2t}}y)p(y)dy,\;t\geq 0,$$ where the gaussian density $p(y)=\frac{1}{(2\pi )^{\frac{n}{2}}}e^{-\frac{1}{2}\left\vert y\right\vert ^{2}}$ is the kernel of $e^{-\frac{\Delta }{2}},$ and $\Delta $ is the (positive) Laplacian on $\mathbb{R}^{n}$. The O-U semi-group is contracting on $L^{q}(\mathbb{R}^{n},pdx),$ $1\leq q\leq \infty ,$ compact if $1<q<\infty ,$ but not compact on $L^{1}(\mathbb{R},pdx)$ [@D theorem 4.3.5], and $p$ is an invariant measure. The generator $-N_{0}$ satisfies $$N_{0}=\sum_{j=1}^{n}(\frac{\partial }{\partial x_{j}})^{\ast }\frac{\partial }{\partial x_{j}}=\Delta -\sum_{j=1}^{n}\frac{\frac{\partial p}{\partial x_{j}}}{p}\frac{\partial }{\partial x_{j}}=\Delta +\sum_{j=1}^{n}x_{j}\frac{\partial }{\partial x_{j}}=\Delta +A$$ where $(\frac{\partial }{\partial x_{j}})^{\ast }$ denotes the adjoint on $L^{2}(\mathbb{R}^{n},pdx)$ and $A$ is the generator of dilations on $\mathbb{R}^{n}.$ On $L^{q}(\mathbb{R}^{n},pdx),1<q<\infty ,$ the spectrum of $N_{0}$ is $\mathbb{N},$ and the Hermite polynomials on $\mathbb{R}^{n}$ form an orthogonal basis of eigenvectors of $e^{-tN_{0}}$ in $L^{2}(\mathbb{R}^{n},pdx).$ The generator $N_{0}$ has a fruitful generalization in (commutative or non commutative) analysis on deformed or $q$-Fock spaces, namely the number operator $N,$ i.e. the second differential quantization of identity. A substitute of Mehler formula holds and $(e^{-tN})_{t>0}$ is the compression of a one parameter group of unitary dilations, see e.g. [@LP2]. Our motivation in this paper is to exploit Mehler formula in another direction: in the setting of stratified groups Mehler formula still defines a semi-group $(e^{-tN})_{t>0}$ and we study which properties of the classical O-U semi-group remain valid. We also hope that this semi-group might throw some light on properties of the heat density $p$. **Results and organization of the paper** In section 2 we recall some properties of the self-adjoint semi-group on $L^{2}(pd\gamma )$ whose generator is $-\nabla ^{\ast }\nabla =-\sum_{i=1}^{n}X_{i}^{\ast }X_{i},$ $X_{i}^{\ast }$ being the formal adjoint of $X_{i}$ with respect to $L^{2}(pd\gamma ).$ We give in passing a simple proof of the known Poincaré inequality in $L^{2}(pd\gamma )$. In the main section 3 we consider another generalization, the Mehler semi-group, which is defined for $t\geq 0$ by (theorem \[P2\]) $$T_{t}(f)(\gamma )=\int_{G}f(\delta _{e^{-t}}\gamma \delta _{\sqrt{1-e^{-2t}}}g)p(g)dg=e^{-tN}(f)(\gamma ).$$ Some properties are described in 3.2, in particular $pd\gamma $ is an invariant measure. This semi-group is not selfadjoint on $L^{2}(pd\gamma ),$ but formally the real part of its generator $-N$ is $-\nabla ^{\ast }\nabla $ and $N=L+A$ where $A$ is the generator of the group ($\delta _{e^{t}})_{t\in \mathbb{R}}$ of dilations$,$ studied in 3.3. We show in 3.4 that every $T_{t},t>0,$ is compact on $L^{q}(pd\gamma ),1<q<\infty ,$ (proposition \[P3\]), with common spectrum $e^{-t\mathbb{N}}$ on the closed subspace spanned by polynomials (theorem \[spectre\]), which coincides with the whole space only if the number of layers of $\mathcal{G}$ is $\leq 4$ (proposition \[dense\]). We describe the eigenspaces in this case. In 3.5 we give another description of these eigenspaces if $G$ is step two, similar to the usual definition of one variable Hermite polynomials by their generating function.       **More** **notation** We denote $\mathcal{G=}V_{1}\oplus ..\oplus V_{k}$, where $V_{1},..,V_{k}$ are the layers of the Lie algebra $\mathcal{G}$ of $G,$ $V_{k}=\mathcal{Z}$ being the central layer, so that [@FS p. 5] $$\lbrack V_{j},V_{h}]\subset V_{j+h},\;[V_{1},V_{h}]=V_{h+1},1\leq h<k$$ The homogeneous dimension of $G$ is $$Q=\sum_{j=1}^{k}j\dim V_{j}.$$ Generic elements of the layers are denoted respectively by $X,Y...,U,$ and respective basis of the layers are denoted by $(X_{1},..,X_{n})$, $(Y_{1},..,Y_{m})$, ..., $(U_{1},..,U_{k}).$ Such a basis is also denoted by $(Z_{j})_{1\leq j\leq N}.$ We denote accordingly $$\begin{aligned} g &=&\exp (\sum x_{i}X_{i}+\sum y_{i}Y_{i}+..+\sum u_{i}U_{i})=\exp (X+Y+..+U) \\ &=&(x,y,..,u)=\exp (\sum_{j=1}^{N}z_{j}Z_{j})=(z_{j})_{j=1}^{N},\end{aligned}$$ since the mapping $(z_{j})_{j=1}^{N}\rightarrow g$ is a diffeomorphism: $\mathbb{R}^{N}\rightarrow G.$ We denote by $\mathcal{P}$ the space of polynomials on $G,$ as defined in [@FS chapter I-C] for the fixed basis $(Z_{j})_{j=1}^{N}$: they are polynomials w.r. to the coordinates $z_{j},1\leq j\leq N.$ The dilation $\delta _{t},t\geq 0,$ are defined on $\mathcal{G}$ and $G$ by $$\delta _{t}(X+Y+..+U)=tX+t^{2}Y+..+t^{k}U,\;\;\delta _{t}(\exp Z)=\exp \delta _{t}(Z),\;Z\in \mathcal{G}.$$ For a function $f$ on $G,$ $$\delta _{t}(f)=f\circ \delta _{t}.$$ The generator $A$ of the one parameter group ($\delta _{e^{s}})_{s\in \mathbb{R}}$ of dilations on $G$ satisfies$:$ for $f\in \mathcal{S}(G)$ and $s>0$ $$\frac{d}{dt}\mid _{t=1}f\circ \delta _{t}=A(f)=-tt^{A}\frac{d}{dt}t^{-A}(f)=-t\delta _{t}\frac{d}{dt}(f\circ \delta _{\frac{1}{t}}). \label{dil}$$ **Acknowledgment:** We thank W. Hebisch who gave us the idea of the proof of proposition \[dense\]. The semi-group $e^{-t\protect\nabla ^{\ast }\protect\nabla }$ on $L^{2}(pdg)$ ============================================================================= This semi-group has already been introduced in [@BHT], under a probabilistic point of view, in connection with some Markov processes on Lie groups.  We use instead an analytic point of view as in [@O]. We consider this semi-group firstly because it is a natural generalization of the classical O-U semi-group, secondly because its generator $\nabla ^{\ast }\nabla $ is the real part of the generator $N$ we shall study in part 3, see theorem \[P2\]. Definition and some properties ------------------------------ We consider the (closed) accretive sesquilinear form $$a(f,h)=\int_{G}(\nabla f.\nabla h)pdg=\int_{G}\sum_{i=1}^{n}X_{i}f\overline{X_{i}h}pdg$$  whose (dense) domain in $L^{2}(pdg)$ is the Hilbert space $$H^{1}(p)=\{f\in L^{2}(pdg)\mid X_{i}f\in L^{2}(pdg),1\leq i\leq n\}$$ equipped with the norm $\left\Vert f\right\Vert _{H^{1}(p)}^{2}=\left\Vert f\right\Vert _{L^{2}(p)}^{2}+\left\Vert \left\vert \nabla f\right\vert \right\Vert _{L^{2}(p)}^{2}$; this form is continuous on $H^{1}(p)\times H^{1}(p).$ Hence [@O proposition 1.51, theorem 1.53] it defines an operator, which we denote by $\nabla ^{\ast }\nabla ,$ such that $-\nabla ^{\ast }\nabla $ is the generator of a strongly continuous semi-group of contractions on $L^{2}(pdg);$ moreover this semi-group is holomorphic on the sector $\Sigma _{\frac{\pi }{2}}=\{\left\vert \arg z\right\vert <\frac{\pi }{2},z\neq 0\},$ and $e^{-z\nabla ^{\ast }\nabla }$ is a contraction on $L^{2}(pdg)$ for $z\in \Sigma _{\frac{\pi }{2}}$. Obviously, on $\mathcal{S}(G),$ $$\nabla ^{\ast }\nabla =\sum_{i=1}^{n}X_{i}^{\ast }X_{i}=L-\sum_{i=1}^{n}\frac{X_{i}p}{p}X_{i}=L-B. \label{500}$$ Since $X_{i}$ is a derivation, the chain rule holds$,$ hence $X_{i}(f^{+})=(X_{i}f)1_{\{f>0\}}$ by the same proof as for usual derivations on $\mathbb{R}^{N}$ [@O proposition 4.4], and $a(f^{+},f^{-})=0$; since the form $a$ also preserves real valued functions, the semi-group $e^{-t\nabla ^{\ast }\nabla }$  is positivity preserving [O]{}. Since $e^{-t\nabla ^{\ast }\nabla }(1)=1,$ the semi-group is thus contracting on $L^{\infty }(pdg).$ Since moreover $\nabla ^{\ast }\nabla $ is self-adjoint, $e^{-t\nabla ^{\ast }\nabla }$ is measure preserving, i.e. $$\int_{G}e^{-t\nabla ^{\ast }\nabla }(f)pdg=\int_{G}fpdg,t>0,$$ so it extends as a contraction semi-group on $L^{1}(pdg)$ hence on $L^{q}(pdg),1<q<\infty $ by interpolation. Poincaré inequality in $L^{2}(pdg)$ ----------------------------------- Poincaré inequality [@DM theorem 4.2] means that the spectrum of $\nabla ^{\ast }\nabla $ on $L^{2}(pdg)$ lies in $\{0\}\cup \lbrack C^{-1},\infty \lbrack $: there exists $C>0$ such that, for $f\in \mathcal{S}(G),$ $$\left\Vert f-\int_{G}fpdg\right\Vert _{L^{2}(pdg)}^{2}\leq C\int_{G}\left\vert \nabla f\right\vert ^{2}pdg=C\int_{G}f(\nabla ^{\ast }\nabla f)pdg. \label{P}$$ (\[P\]) follows from the inequality (used for $q=2)$ [@DM theorem 4.1] $$\left\vert \nabla (e^{-tL}f)\right\vert ^{q}\leq C_{q}e^{-tL}(\left\vert \nabla f\right\vert ^{q}),\;1<q<\infty , \label{DM}$$ which B. Driver and T. Melcher proved, first for $\mathbb{H}_{1}, $ then for nilpotent groups $G$ (see T. Melcher’s thesis), using Malliavin calculus. See also [@BHT] for some extensions. We shall show in proposition \[Poincare\] that (\[DM\]) also follows easily from gaussian estimates of $p$ and $\nabla p$. Using the explicit formula for the Carnot-Caratheodory distance, H.Q. Li [@Li corollary 1.2] obtained (\[DM\]) for $q=1$, on the 3-dimensional Heisenberg group $G=\mathbb{H}_{1}.$ As well known [@A théorème 5.4.7], this implies Log-Sobolev inequality for the measure $pdg$ on $\mathbb{H}_{1}$ and (\[P\]). Another proof of this Log-Sobolev inequality for $\mathbb{H}_{1},$ hence for $\mathbb{H}_{k},$ is given in [@HZ theorem 7.3]$.$ \[Poincare\][@DM] Let $G$ be a stratified group. Then (\[DM\]) and Poincaré inequality (\[P\]) hold true. Proof: By [@DM theorem 4.2, proposition 2.6, lemma 2.3] it is enough to prove (\[DM\]) for $t=\frac{1}{2}$, at $\gamma =0$. Hence, it is enough to prove, for an element $X$ of the basis of $V_{1},$ and $f\in \mathcal{S}(G),$ $$\left\vert X(e^{-\frac{1}{2}L}f)(0)\right\vert =\left\vert X(f\ast p)(0)\right\vert =\left\vert \int_{G}(\widehat{X}f)(g)p(g)dg\right\vert \leq C_{q,X}\left\Vert \nabla f\right\Vert _{L^{q}(pdg)};$$ here [@FS p. 22 and proposition 1.29] $$(\widehat{X}f)(g)=\frac{d}{dt}\mid _{t=0}f((\exp tX)g),\;\;\widehat{X}=X+\sum_{j>n}Q_{X,j}Z_{j}$$ where $(Z_{j})_{j=1}^{N}$ is a basis of $\mathcal{G}$ respecting the layers and $Q_{X,j}$ is a polynomial (with homogeneous degree $h-1$ if $Z_{j}\in V_{h},2\leq h\leq k).$ Since $[V_{1},V_{h-1}]=V_{h},2\leq h\leq k,$ we may choose $Z_{j}\in V_{h}$ such that $Z_{j}=[Y,A],$ where $Y$ is an element of the basis of $V_{1}$ and $A\in V_{h-1}.$Then $$\left\vert \int_{G}Z_{j}f(g)Q_{X,j}(g)p(g)dg\right\vert \leq \left\vert \int_{G}Yf\;A(Q_{X,j}p)dg\right\vert +\left\vert \int_{G}Af\;Y(Q_{X,j}p)dg\right\vert .$$ Iterating for $A\in V_{1}+..+V_{k-1}$ and so on, $\left\vert \int_{G}(\widehat{X}f)(g)p(g)dg\right\vert $ is finally less than a finite number (which does not depend on $f)$ of terms $\left\vert \int_{G}Yf\;Z(Qp)dg\right\vert $ where $Y$ is an element of the basis of $V_{1},Z\in \mathcal{G},$ and $Q$ is a polynomial. Each of these terms can be estimated by $$\left\vert \int_{G}Yf\;Z(Qp)dg\right\vert \leq \left\Vert \left\vert \nabla f\right\vert \right\Vert _{L^{q}(pdg)}(\left\Vert ZQ\right\Vert _{L^{q^{\prime }}(pdg)}+\left\Vert Q\frac{Zp}{p}\right\Vert _{L^{q^{\prime }}(pdg)})$$ where $\frac{1}{q}+\frac{1}{q^{\prime }}=1.$ Then $\left\Vert ZQ\right\Vert _{L^{q^{\prime }}(pdg)}$ is finite since $ZQ$ is a polynomial and $p\in \mathcal{S}(G).$ The main point is that $\left\Vert Q\frac{Zp}{p}\right\Vert _{L^{q^{\prime }}(pdg)}$ is finite. Indeed, denoting $d(g)=d(0,g) $ where $d$ is the Carnot-Caratheodory distance on $G,$ one uses [@CSV theorem IV.4.2 and Comments on chapter IV]: for $0<$ $\varepsilon <1,$ $$C_{\varepsilon }e^{-\frac{1}{2-2\varepsilon }d^{2}(g)}\leq p(g)\leq K_{\varepsilon }e^{-\frac{1}{2+2\varepsilon }d^{2}(g)}. \label{11}$$ and, for $Z\in \mathcal{G},$ $$(Zp)(g)\leq K_{\varepsilon ,Z}\;e^{-\frac{1}{2+2\varepsilon }d^{2}(g)}. \label{111}$$  Hence $Q\frac{Zp}{p}$ lies in $L^{r}(pdg),1\leq r<\infty ,$ which ends the proof.$\blacksquare $ Definition and properties of the Mehler semi-group ================================================== Preliminaries ------------- The next proposition extends a classical property of independant gaussian variables and will imply the semi-group property of our family of operators. \[P1\]Let $\gamma ,g$ be independant $G-$valued random variables with law $pdg.$ Then the r.v.$$\delta _{\cos \theta }\gamma \delta _{\sin \theta }g,\;0\leq \theta \leq \frac{\pi }{2}$$ has the same law, i.e. for any bounded borelian function $f$ on $G, $ $$\int_{G^{2}}f(\delta _{\cos \theta }\gamma \delta _{\sin \theta }g)p(\gamma )p(g)d\gamma dg=\int_{G}f(g)p(g)dg.$$ More generally, if $g_{1},...,g_{n}$ are $G$-valued i.i.d r.v. with law $pdg$ and $\sum\limits_{1\leq j\leq n}a_{j}^{2}=1,$ $(a_{j}\geq 0),$ the law of $\prod\limits_{j=1}^{j=n}\delta _{a_{j}}g_{j}$ is $pdg.$ Proof: By two changes of variables, denoting $C=\sin \theta \cos \theta ,$ $$\begin{aligned} \int_{G^{2}}f(\delta _{\cos \theta }\gamma \delta _{\sin \theta }g)p(g)p(\gamma )d\gamma dg &=&\frac{1}{C^{Q}}\int_{G^{2}}f(\gamma ^{\prime }g^{\prime })p(\delta _{\frac{1}{\cos \theta }}\gamma ^{\prime })p(\delta _{\frac{1}{\sin \theta }}g^{\prime })d\gamma ^{\prime }dg^{\prime } \\ &=&\frac{1}{C^{Q}}\int_{G^{2}}f(g)p(\delta _{\frac{1}{\cos \theta }}\gamma ^{\prime })p(\delta _{\frac{1}{\sin \theta }}(\gamma ^{\prime -1}g))d\gamma ^{\prime }dg \\ &=&\int_{G}f(g)(p_{\cos ^{2}\theta }\ast p_{\sin ^{2}\theta })(g)dg \\ &=&\int_{G}f(g)p(g)dg.\end{aligned}$$ The second assertion follows by iteration.      *Remark 1*: A central limit theorem for i.i.d centered random variables with values in a stratified group $G$ and law $\mu $ with order 2 moments is proved in [@CR theorem 3.1]. The density $p$ of the limit law is the kernel at time 1 of a diffusion semi-group whose generator satisfies (\[1\]). *Remark 2:* If $X,Y$ are i.i.d standard gaussian vectors with values in $\mathbb{R}^{n}$, the couple $(X\cos \theta +Y\sin \theta ,$ $\frac{d}{d\theta }(X\cos \theta +Y\sin \theta ))$ has the same joint law as $(X,Y)$. This fact implies, in the O-U case,* *$\ $that $\cos ^{N_{0}}\theta $ is the compression of the isometry $R_{\theta }$ of $L^{2}(\mathbb{R}^{n}\times \mathbb{R}^{n},p(x)p(y)dxdy)$ defined by $$R_{\theta }(F)(x,y)=F(x\cos \theta +y\sin \theta ,-x\sin \theta +y\cos \theta )$$ and $(R_{\theta })_{\theta \in \mathbb{R}}$ is a one parameter group preserving the measure $p(x)p(y)dxdy.$ This point of view was exploited e.g. in [@Q theorem 2.2] in order to get a concentration inequality for the gaussian measure . In the stratified setting we were not able to exhibit explicit unitary dilations for the Mehler operators $T_{t}$ defined below. The Mehler semi-group --------------------- We now define the Mehler semi-group on $L^{q}(G,pdg).$ \[P2\]Let  $L,$ defined by (\[0\]), be a subLaplacian on a stratified group $G,$ and let $p$ be the kernel of $e^{-\frac{L}{2}}$. a\) The family of operators $(T_{t})_{t\geq 0}$ defined on $\mathcal{S(}G\mathcal{)}$ by $$T_{t}(f)(\gamma )=\int_{G}f(\delta _{e^{-t}}\gamma \delta _{\sqrt{1-e^{-2t}}}g)p(g)dg=e^{-\frac{L}{2}(1-e^{-2t})}(f)(\delta _{e^{-t}}\gamma ) \label{1bis}$$ is a semi-group whose generator $-N$ is defined on $\mathcal{S(}G\mathcal{)}$ by$$N=L+A. \label{7bis}$$ b)The probability measure $pd\gamma $ is invariant by $(T_{t})_{t\geq 0}$ i.e.$$\int_{G}T_{t}(f)(\gamma )p(\gamma )d\gamma =\int_{G}f(\gamma )p(\gamma )d\gamma \label{3bis}$$ and, for $f\in \mathcal{S(}G\mathcal{)},\int_{G}(Nf)pdg=0.$ c\) $(T_{t})_{t\geq 0}$ extends as a Markovian semi-group of contractions on $L^{q}(G,pd\gamma ),1\leq q\leq \infty $, strongly continuous if $q\neq \infty .$ d\) If $f\in L^{q}(pd\gamma ),1\leq q<\infty ,$ $$\left\Vert T_{t}(f)-\int_{G}fpdg\right\Vert _{L^{q}(pd\gamma )}\rightarrow _{t\rightarrow \infty }0.$$ e\) $(T_{t})_{t>0}$ is not self-adjoint on $L^{2}(G,pd\gamma )$ as soon as $G$ is not abelian. Formally  $\nabla ^{\ast }\nabla $ is the real part of $N,$ i.e., for $f,h\in \mathcal{S(}G\mathcal{)},$$$\left\langle Nf,h\right\rangle _{L^{2}(p)}=\left\langle (\nabla ^{\ast }\nabla +iC)f,h\right\rangle _{L^{2}(p)}$$ where $C$ is a non zero first order differential operator satisfying $\left\langle Cf,h\right\rangle =\left\langle f,Ch\right\rangle .$ In particular, for $f\in \mathcal{S(}G\mathcal{)},$ $$\Re \int_{G}(Nf)fpd\gamma =\int_{G}\left\vert \nabla f\right\vert ^{2}pd\gamma =\int_{G}(\nabla ^{\ast }\nabla f)fpd\gamma .$$ If  moreover $f$ is real valued, the left integral is real. By the change of notation $e^{-t}=\cos \theta $, $<\theta <\frac{\pi }{2}$, (\[1bis\]) can be rewritten as $$\cos ^{N}\theta (f)(\gamma )=\int_{G}f(\delta _{\cos \theta }\gamma \delta _{\sin \theta }g)p(g)dg=\delta _{\cos \theta }\circ e^{-\frac{1}{2}\sin ^{2}\theta L}(f)(\gamma ). \label{7}$$ Proof: a) Let $\varphi (g^{\prime })=$ $T_{t}(f)(g^{\prime });$ we compute $$\begin{aligned} T_{s}(\varphi )(\gamma ) &=&\int_{G}\varphi (\delta _{e^{-s}}\gamma \;\delta _{\sqrt{1-e^{-2s}}}h)p(h)dh \\ &=&\int_{G^{2}}f(\delta _{e^{-t}}[\delta _{e^{-s}}\gamma \delta _{\sqrt{1-e^{-2s}}}h]\;\delta _{\sqrt{1-e^{-2t}}}g)p(g)p(h)dgdh \\ &=&\int_{G}f(\delta _{e^{-(t+s)}}\gamma \delta _{\sqrt{1-e^{-2(s+t)}}}k)p(k)dk=T_{s+t}(f)(\gamma )\end{aligned}$$ where the third equality comes from proposition \[P1\] applied to $(h,g)$. By the chain rule applied to (\[1bis\]), $$Nf=-\frac{d}{dt}\mid _{t=0}T_{t}(f)=Lf+A(f).$$ b\) Proposition \[P1\] gives (\[3bis\]). Differentiating ([3bis]{}) at $t=0$ for $f\in \mathcal{S}(G)$ implies $$\int_{G}(Nf)pdg=0.$$ Another proof will be given in Remark 3. c\) $T_{t}$ is contracting both on $L^{1}(G,pd\gamma ),$ since it is positivity and measure preserving, and on $L^{\infty }(G,pd\gamma ),$ since it is positivity preserving and $T_{t}(1)=1.$ Hence $T_{t}$ is contracting on $L^{q}(G,pd\gamma ),1\leq q\leq \infty $ by interpolation.      Since $\mathcal{D}(G)$ is norm dense in $L^{q}(G),\mathcal{\ }$it is norm dense in $L^{q}(pd\gamma ),1\leq q<\infty :$ indeed, if $F\in L^{q^{\prime }}(pd\gamma )$ $(\frac{1}{q}+\frac{1}{q^{\prime }}=1)$ and $\int_{G}fFpd\gamma =0$ for every $f\in \mathcal{D}(G),$ then $Fp\in L^{q^{\prime }}(G)$ hence $Fp=0$ $d\gamma $ a.s.$.$ Writing $e^{-t}=\cos \theta ,$ one has, for $f\in \mathcal{D}(G),$ $$\begin{aligned} \left\Vert T_{t}(f)-f\right\Vert _{L^{q}(pd\gamma )}^{q} &=&\left\Vert \int_{G}[f(\delta _{\cos \theta }\gamma \delta _{\sin \theta }g)-f(\gamma )]p(g)dg\right\Vert _{L^{q}(pd\gamma )}^{q} \\ &\leq &\int_{G^{2}}\left\vert f(\delta _{\cos \theta }\gamma \delta _{\sin \theta }g)-f(\gamma )\right\vert ^{q}p(\gamma )p(g)d\gamma dg,\end{aligned}$$ which converges to 0 as $\theta \rightarrow 0$ by the dominated convergence theorem. Since $T_{t}$ is contracting, the strong continuity on $L^{q}(pd\gamma )$ follows by density. d\) Similarly, if $f$ is bounded and continuous on $G,$ $$f(\delta _{e^{-t}}\gamma \;\delta _{\sqrt{1-e^{-2t}}}g)\rightarrow _{t\rightarrow \infty }f(g);$$ by dominated convergence theorem $T_{t}(f)\rightarrow _{t\rightarrow \infty }\int_{G}f(g)p(g)dg$  pointwise and in the norm of $L^{q}(pd\gamma ).$ The claim follows by density. e\) By (\[7bis\]), (\[500\]) and lemma \[20\] below, for $f\in \mathcal{S}(G),$ $$(N-\nabla ^{\ast }\nabla )f=A(f)+\sum\limits_{1\leq j\leq n}\frac{X_{j}p}{p}X_{j}f=\sum\limits_{1\leq j\leq N}b_{j}Z_{j}f$$  where the functions $b_{j}$ are not all zero if $j>n=\dim V_{1}.$ Hence for $h\in \mathcal{S}(G),$ $$\int_{G}(N-\nabla ^{\ast }\nabla )(f)\overline{h}pdg=-\int_{G}f[\sum\limits_{1\leq j\leq N}b_{j}(g)(Z_{j}\overline{h})p+\overline{h}Z_{j}(b_{j}p)]dg.$$ By b), the left hand side is zero for $h=1,$ hence $\sum\limits_{1\leq j\leq N}Z_{j}(b_{j}p)=0.$ Since $T_{t}$ preserves real valued functions, so does $N,$ hence $$\int_{G}(N-\nabla ^{\ast }\nabla )(f)\overline{h}pdg=-\int_{G}f(N-\nabla ^{\ast }\nabla )(\overline{h})pdg=-\int_{G}f\overline{(N-\nabla ^{\ast }\nabla )(h)}pdg,$$ which proves ($iC)^{\ast }=-iC,$ where $iC=N-\nabla ^{\ast }\nabla =A+B$. The remaining assertions are obvious. *Remark 3*: We now give another instructive proof of $\int_{G}(Nf)pdg=0,f\in \mathcal{S}(G),$ hence of (\[3bis\]). We claim that, for $f,h\in \mathcal{S}(G)$, $$\int_{G}(Nf)hdg=\int_{G}f[L(h)-Qh+\frac{d}{ds}\mid _{s=1}h\circ \delta _{\frac{1}{s}}]dg=\int_{G}f(L-QId-A)(h)dg.$$ Indeed, $N=L+A,$ $L$ is formally selfadjoint on $L^{2}(dg)$ and the claim follows by differentiating at $s=1$ the right hand side of $$\int_{G}f(\delta _{s}\gamma )h(\gamma )d\gamma =s^{-Q}\int_{G}f(\gamma ^{\prime })h(\delta _{\frac{1}{s}}\gamma )d\gamma ^{\prime }.$$ By (\[dil\]) and [@LP lemma 2], $p$ may be precisely defined as the unique solution in $L^{1}(G),$ satisfying $\int_{G}p(g)dg=1,$ of $$(L-QId-A)(p)=Lp-Qp+s\delta _{s}\frac{d}{ds}(p\circ \delta _{\frac{1}{s}})=0.\blacksquare$$ *Remark 4:* As already mentioned in section 2.2, Log-Sobolev inequality for $pd\gamma $ is known for $\mathbb{H}_{k}.$ It is equivalent both to hypercontractivity of $e^{-tN}$ and to hypercontractivity of $e^{-t\nabla ^{\ast }\nabla }$ on $\mathbb{H}_{k},$ since $p$ is an invariant measure for these markovian semigroups and $N,\nabla ^{\ast }\nabla $ are diffusion operators [@A theorem 2.8.2]. The generator of dilations -------------------------- ** **We may identify $G$ with a group of finite matrices [@V theorem 3.6.6]. The derivation formula for an exponential of a matrix valued function, see e.g. [@H theorem 69], applied to a smooth function $Z(s)$: $\mathbb{R}\rightarrow \mathcal{G},$ where $\mathcal{G}$ has $k$ layers, gives $$\begin{aligned} \frac{d}{ds}\exp Z(s) &=&\lim_{h\rightarrow 0}\frac{\exp Z(s+h)-\exp Z(s)}{h} \nonumber \\ &=&\lim_{h\rightarrow 0}\frac{\exp (Z(s)+hZ^{\prime }(s))-\exp Z(s)}{h} \nonumber \\ &=&[\exp Z(s)]V(Z(s)), \label{100}\end{aligned}$$ where $$V(Z(s))=(d\exp )_{Z(s)}(Z^{\prime }(s))=Z^{\prime }(s)+\sum_{l=1}^{k-1}\frac{(-1)^{l}}{(l+1)!}(AdZ(s))^{l}(Z^{\prime }(s)). \label{501}$$ Hence $$\exp Z(s+h)=\exp Z(s)\exp h[V(Z(s))+o(1)],$$ which entails for $f\in \mathcal{C}^{\infty }(G)$ $$\frac{d}{ds}f(\exp Z(s))=V(Z(s))(f)(\exp Z(s)). \label{101}$$ \[20\] Let $A$ be the generator of the group of dilations ($\delta _{e^{t}})_{t\in \mathbb{R}}.$ Then$$A(f)(g)=\sum\limits_{1\leq j\leq N}a_{j}(g)Z_{j}f(g)$$ where the functions $a_{j}$ are polynomials w.r. to the coordinates of $g,$ and are not all zero for $j>n=\dim V_{1}$. Proof: Assume that $\mathcal{G}$ has $k$ layers, $k\geq 2$. Let $$\delta _{s}g=\exp (sX+s^{2}Y+..+s^{k}U)=\exp Z(s).$$ By (\[101\]) $A=V(Z(1)).$ Noting that $Z^{\prime }-Z\in V_{2}+..+V_{k},$ we get $(AdZ(1))^{l}(Z^{\prime }(1))\in V_{3}+..+V_{k},l\geq 1.$ So $V(Z(1))-(X+2Y)$ lies in $V_{3}+..+V_{k}$.$\blacksquare $ **Notation:** We denote by $\mathcal{P}_{n}$ the (finite dimensional) space of homogeneous polynomials on $G$ with homogeneous degree $n,n\in \mathbb{N},$ i.e. satisfying $$\delta _{s}(P)=s^{n}P,\;P\in \mathcal{P}_{n}; \label{vpA}$$ equivalently, $\mathcal{P}_{n}$ is the eigensubspace of $A$ on $\mathcal{P}$ associated to $n.$ The finite dimensional subspaces $B_{n}=\mathcal{P}_{0}+..+\mathcal{P}_{n}$ are stable under $L$ and dilations, hence under $e^{-\frac{tL}{2}}$ and $\cos ^{N}\theta $ by (\[1bis\]), these operators being naturally extended on $\mathcal{S}^{\prime }(G).$ In particular $e^{\frac{L}{2}}$ is well defined on $B_{n}$ and is the inverse of $e^{-\frac{L}{2}},$ which is thus one to one on every $B_{n}$ hence on $\mathcal{P=\cup }_{n\geq 0}B_{n}$. The next lemma is the key for the computation of the spectrum of $\cos ^{N}\theta .$ It will be exploited again in section 3.5. \[LA\] a) The generator $A$ of dilations on $G$ satisfies $[L,A]=2L$ on $\mathcal{C}^{\infty }(G).$ b\) $e^{-\frac{L}{2}}\circ \cos ^{N}\theta =\delta _{\cos \theta }e^{-\frac{L}{2}}$ on $\mathcal{S}^{\prime }(G).$ c\) The set of polynomials $e^{\frac{L}{2}}(\mathcal{P}_{n})$ is a space of eigenvectors of $\cos ^{N}\theta $ associated to the eigenvalue $\cos ^{n}\theta ,n\geq 0.$ Proof: a) We rewrite (\[1\]) as $$Le^{tA}=e^{2t}e^{tA}L,\;t\in \mathbb{R},$$ and a) follows by differentiating at $t=0.$ b\) By (\[2\]), on $\mathcal{S}(G),$ hence on $\mathcal{S}^{\prime }(G),$ for $t>0,$ $$e^{-\frac{t^{2}}{2}L}=\delta _{\frac{1}{t}}\circ e^{-\frac{L}{2}}\circ \delta _{t}. \label{800}$$ Hence, on $\mathcal{S}^{\prime }(G)),$ by (\[1bis\]) and ([800]{}) applied to $t=\cos \theta ,$ $$e^{-\frac{L}{2}}\circ \cos ^{N}\theta =e^{-\frac{L}{2}}\circ \delta _{\cos \theta }\circ e^{-\frac{\sin ^{2}\theta }{2}L}=\delta _{\cos \theta }\circ e^{-\frac{L}{2}}.$$ c\) Since $e^{-\frac{L}{2}}$ is invertible on $\mathcal{P},$ and $\mathcal{P}$ is stable under $cos^{N}\theta ,$ b) implies on $\mathcal{P}$ $$\cos ^{N}\theta \circ e^{\frac{L}{2}}=e^{\frac{L}{2}}\circ \delta _{\cos \theta }.$$ Applying this to $\mathcal{P}_{n}$ proves the result. Compacity and spectrum of $\cos ^{N}\protect\theta $ on $L^{q}(pd\protect\gamma )$ ---------------------------------------------------------------------------------- \[P3\] Let $\cos ^{N}\theta $ be defined by (\[7\]). Then a\) $\cos ^{N}\theta $ is a Hilbert-Schmidt operator on $L^{2}(pd\gamma ).$ b\) $\cos ^{N}\theta $ is compact on $L^{q}(pd\gamma ),1<q<\infty ;$ its non zero eigenvalues and corresponding eigenspaces are the same on $L^{2}(pd\gamma )$ and $L^{q}(pd\gamma ).$ In particular its spectrum $\sigma (\cos ^{N}\theta )$ does not depend on $q$ and$$\sigma (\cos ^{N}\theta )=(\cos \theta )^{\sigma (N)}\cup \{0\}.$$ Actually, $\cos ^{N}\theta $ is a trace class operator on $L^{2}(pd\gamma )$ by a) and the semi-group property of $(e^{-tN})_{t>0}.$ Proof: a) We must show that the kernel of $\cos ^{N}\theta $ lies in $L^{2}(G\times G,pd\gamma \otimes pdg).$ For fixed $\gamma $ and $\theta ,0<\theta <\frac{\pi }{2},$ $$\int_{G}f(\delta _{\cos \theta }\gamma \delta _{\sin \theta }g)p(g)dg=\frac{1}{\sin ^{Q}\theta }\int_{G}f(z)p(\delta _{\frac{\cos \theta }{\sin \theta }}\gamma ^{-1}\delta _{\frac{1}{\sin \theta }}z)dz,$$ so we must prove the convergence of the integral $$I(\theta )=\int_{G^{2}}p^{2}(\delta _{\frac{\cos \theta }{\sin \theta }}\gamma ^{-1}\delta _{\frac{1}{\sin \theta }}z)\frac{p(\gamma )}{p(z)}dzd\gamma .$$ By the gaussian estimates (\[11\]) $$\frac{C_{\varepsilon }}{K_{\varepsilon }^{3}}p^{2}(\delta _{\frac{\cos \theta }{\sin \theta }}\gamma ^{-1}\delta _{\frac{1}{\sin \theta }}z)\frac{p(\gamma )}{p(z)}\leq \exp (\frac{d^{2}(z)}{2-2\varepsilon }-\frac{d^{2}(\gamma )}{2+2\varepsilon }-\frac{d^{2}(\delta _{\frac{\cos \theta }{\sin \theta }}\gamma ^{-1}\delta _{\frac{1}{\sin \theta }}z)}{1+\varepsilon })=\exp \beta .$$ The Carnot distance $d$ satisfies $$d(g)\leq d(\gamma ^{-1}g)+d(\gamma )\;and\;d(\delta _{t}g)=td(g).$$ Hence $$\begin{aligned} (1+\varepsilon )\beta &\leq &\frac{d^{2}(z)}{2(1-\varepsilon )^{2}}-\frac{d^{2}(\gamma )}{2}-(\frac{1}{\sin \theta }d(z)-\frac{\cos \theta }{\sin \theta }d(\gamma ))^{2} \\ &\leq &d^{2}(z)(\frac{1}{2-4\varepsilon }-\frac{1-\cos \theta }{\sin ^{2}\theta })+d^{2}(\gamma )(\frac{\cos \theta -\cos ^{2}\theta }{\sin ^{2}\theta }-\frac{1}{2}).\end{aligned}$$ Since $\frac{1-\cos \theta }{\sin ^{2}\theta }>\frac{1}{2}$ on $]0,\frac{\pi }{2}],$ the coefficient of $d^{2}(\gamma )$ is strictly negative, and so is the coefficient of $d^{2}(z)$ for small enough $\varepsilon >0$. Hence, for some $c,C>0,$ $$I(\theta )\leq C\int \int_{G^{2}}e^{-c(d^{2}(z)+d^{2}(\gamma ))}dzd\gamma =C(\int_{G}e^{-cd^{2}(z)}dz)^{2}.$$ By the left hand side of (\[11\]), for small $\varepsilon ,$ $$C_{\varepsilon }\int_{G}e^{-cd^{2}(z)}dz\leq \int_{G}p^{2c(1-\varepsilon )}(z)dz,$$ and the last integral is finite since $p\in \mathcal{S}(G)$. This proves a). b\) By interpolation, since $cos^{N}\theta $ is compact on $L^{2}(p(g)dg)$ and bounded on $L^{\infty }(pdg)$ and $L^{1}(pdg),$ it is compact on $L^{q}(pdg),1<q<\infty ,$ with the same spectrum and the same eigenspaces associated to non zero eigenvalues [@D theorems 1.6.1 and 1.6.2]. By the compacity on $L^{q}(pdg),$ the set of these eigenvalues is $\{\cos ^{\lambda }\theta \mid \lambda \in \sigma _{q}(N)\}$ where $\sigma _{q}(N)$ denotes the spectrum of $N$ on $L^{q}(pdg)$ [@L chap. 34.5, theorem 13]. Hence $\sigma _{q}(N)=\sigma _{2}(N)$ is discrete and lies in {$\lambda \in \mathbb{C}\mid \Re \lambda \geq 0$} since $cos^{N}\theta $ is contracting on $L^{2}(pdg)$ (or since $\Re \left\langle Nf,f\right\rangle \geq 0).\blacksquare $ \[spectre\]Let $G$ be a step $k$ stratified group. 1\) If $k\leq 4$ a\) the spectrum of $\cos ^{N}\theta $ on $L^{2}(pdg)$ is $\sigma (\cos ^{N}\theta )=(\cos \theta )^{\mathbb{N}}\cup \{0\}$ and $\sigma (N)=\mathbb{N}.$ b\) the corresponding eigenspaces $E_{n},n\geq 0,$ (which are not pairwise orthogonal in $L^{2}(pdg)$) are $$E_{n}=e^{\frac{1}{2}L}(\mathcal{P}_{n}).$$ 2\) If $k>4,$ assertions a) b) remain true for the restriction of $\cos ^{\mathbb{N}}\theta $ to the closed subspace $L_{\mathcal{P}}^{2}(pdg)$ spanned by polynomials. If $k=1$ polynomials in $E_{n}$ are the Hermite polynomials with degree $n$. Proof: 1) follows from 2) and proposition \[dense\] below. 2\) We first define $E_{n}$ by $E_{n}=e^{\frac{L}{2}}(\mathcal{P}_{n}).$ By lemma \[LA\], $E_{n}$ lies in the eigenspace of $\cos ^{N}\theta $ associated to the eigenvalue $\cos ^{n}\theta .$ By proposition \[P3\], $\cos ^{N}\theta $ is compact on $L_{\mathcal{P}}^{2}(pdg).$ The claim then follows from the following facts: Let $T:E\rightarrow E$ be a compact operator on an infinite dimensional Banach space $E;$ let $\Lambda $ be a set of eigenvalues of $T$ and let $E_{\lambda },\lambda \in \Lambda ,$ be eigensubspaces whose union is total in $E.$ Then a\) the spectrum of $T$ is $\Lambda \cup \{0\}$ b\) for $\lambda \in \Lambda ,E_{\lambda }$ is the whole eigenspace associated to $\lambda .$ Indeed, assume that $T$ has an eigenvalue $\lambda _{0}\notin \Lambda .$ Then $T-\lambda _{0}I$  has a closed range with non zero finite codimension (see e.g. [@L chap. 21.1, theorems 3, 4]). But this range contains the linear span of the $E_{\lambda }$’s, $\lambda \in \Lambda ,$ hence is the whole of $E.$ This is a contradiction, which proves a). Let $\lambda _{0}\in \Lambda ;$ since $E_{\lambda _{0}}$ is stable under $T,$ $T$ acts on the quotient space $E/E_{\lambda _{0}}$ and is still compact. The $E_{\lambda }$’s, $\lambda \in \Lambda \backslash \{\lambda _{0}\}$ span a dense subspace of $E/E_{\lambda _{0}}.$ Applying a) to $E/E_{\lambda _{0}}$, $\lambda _{0}$ cannot belong to the spectrum of $T$ on the quotient space, which proves b).$\blacksquare $ The proof of the next proposition is essentially due to W. Hebisch (private communication). \[dense\]Let $G$ be a stratified group. Then the polynomials are dense in $L^{2}(pdg)$ if and only if $G$ is step $k$ with $k\leq 4$. Proof: 1) We recall that polynomials are dense in $L^{2}(\mathbb{R},e^{-c\left\vert x\right\vert ^{\alpha }}dx)$ if and only if $\alpha \geq \frac{1}{2}:$ obviously, this does not depend on $c$ and is equivalent to the density of polynomials in $L^{2}(\mathbb{R}^{+},e^{-x^{\alpha }}dx).$ If $0<\alpha <\frac{1}{2},$ [@PS Part III, problem 153] produces a non zero bounded function $g_{\alpha }$ which is orthogonal to polynomials in $L^{2}(\mathbb{R}^{+},e^{-\cos (\alpha \pi )x^{\alpha }}dx).$ If $\alpha \geq \frac{1}{2},$ the result follows from the trick of [@Ham p 197-198]. Indeed, if $\psi \in L^{2}(\mathbb{R}^{+},e^{-x^{\alpha }}dx)$ and $\alpha \geq \frac{1}{2}$, the function $$F(z)=\int_{\mathbb{R}^{+}}\psi (x)e^{\sqrt{x}z}e^{-x^{\alpha }}dx=\int_{\mathbb{R}^{+}}\psi (y^{2})e^{yz}e^{-y^{2\alpha }}ydy$$ is bounded and holomorphic on $\{\Re z<\beta \}$ for some $\beta >0,$ by Cauchy-Schwarz inequality. Expanding $z\rightarrow e^{\sqrt{x}z}$ in power series, one gets $F(-z)=-F(z)$ if $\psi $ is orthogonal to polynomials in $L^{2}(\mathbb{R}^{+},e^{-x^{\alpha }}dx)$. Thus $F$ extends as a bounded entire function, which must be zero by Liouville theorem since $F(0)=0$. Hence the Fourier transform of $y\rightarrow \psi (y^{2})e^{-y^{2\alpha }}y$ is zero, i.e. $\psi =0$ a.s.. 2\) We identify $g=\exp Z\in G$ with the coordinates $(x,y,..,w)$ of $Z$ w.r. to a basis respecting the layers and denote $$\eta (g)=\sum_{i\leq l}\left\vert x_{i}\right\vert ^{2}+\sum_{i\leq m}\left\vert y_{i}\right\vert ^{\frac{2}{2}}+...+\sum_{i\leq r}\left\vert w_{i}\right\vert ^{\frac{2}{k}}.$$ Obviously $\eta (\delta _{s}g)=s^{2}\eta (g),$ in particular $\eta (g)=d^{2}(g)\eta (\delta _{\frac{1}{d(g)}}g),$ $d$ denoting the Carnot distance. Since $\eta $ is strictly positive and bounded on the $d$-unit sphere of $G,$ there exist constants $c^{\prime },C^{\prime }>0$ such that $$c^{\prime }\eta (g)\leq d^{2}(g)\leq C^{\prime }\eta (g).$$ By (\[11\]) there exist constants $c,C>0$ such that the following embeddings $$L^{2}(e^{-C\eta (g)}dg)\rightarrow L^{2}(pdg)\rightarrow L^{2}(e^{-c\eta (g)}dg)$$ are continuous, with dense ranges since $\mathcal{D}(G)$ is dense in the three spaces. 3\) The algebraic tensor product $$\mathcal{E}=\otimes _{i\leq l}L^{2}(e^{-Cx_{i}^{2}}dx_{i})\otimes ...\otimes _{i\leq p}L^{2}(e^{-C\left\vert w_{i}\right\vert ^{\frac{2}{k}}}dw_{i}),$$is dense in $L^{2}(e^{-C\eta (g)}dg).$ For $k\leq 4,$ one variable polynomials are dense in every factor of $\mathcal{E}$ by step 1), hence polynomials are dense in $L^{2}(e^{-C\eta (g)}dg)$ and in $L^{2}(pdg).$ Let $k\geq 5.$ By 1) there exists a non zero function $g\in L^{2}(e^{-c\left\vert w_{r}\right\vert ^{\frac{2}{k}}}dw_{r})$ which is orthogonal to polynomials w.r. to $w_{r}$. Then $1\otimes ....\otimes 1\otimes g\in L^{2}(e^{-c\eta (g)}dg)$ is orthogonal to all polynomials, so polynomials are neither dense in $L^{2}(e^{-c\eta (g)}dg),$ nor in $L^{2}(pdg).\blacksquare $ Generating functions of polynomial eigenvectors of $N$ ------------------------------------------------------ The usual Hermite polynomials on $\mathbb{R},$ denoted by $H_{n},n\in \mathbb{N},$ are the eigenvectors of the Ornstein-Uhlenbeck operator $N_{0},$ and have the generating function $$e^{ixt+\frac{1}{2}t^{2}}=\sum_{n\geq 0}\frac{(it)^{n}}{n!}H_{n}(x)=e^{\frac{1}{2}\Delta }(e^{ixt})=e^{\frac{1}{2}\Delta }\circ \delta _{t}(e^{ix}),$$ noting that $x\rightarrow e^{ix}$ is a bounded eigenvector of $\Delta .$ In particular $$i^{n}H_{n}(x)=\frac{d^{n}}{dt^{n}}\mid _{t=0}e^{\frac{1}{2}\Delta }\circ \delta _{t}(e^{ix}).$$We shall verify (proposition \[P5\]) that a similar formula gives polynomial eigenvectors of $N.$ When $G$ is step two, these vectors are total in $L^{q}(pdg),1\leq q<\infty ,$ see theorem \[9\] below. More precisely we give in 3.5.1 a technical lemma producing eigenvectors of $N$ out of eigenvectors of $L.$ In 3.5.3 we use this lemma when $\varphi $ is both an eigenvector of $L$ and a coefficient function of a representation of $G$ (proposition \[P5\]). We shall first gather in 3.5.2 well known facts about these functions. ### Candidates for generating functions of eigenvectors of $N$ In the next lemma \[GF\] we state technical assumptions ensuring the validity of the computation of some eigenvectors of $N.$ Using lemma \[LA\] b), the point is to define $"e^{\frac{L}{2}}\varphi "$ for suitable functions $\varphi :$ in lemma \[LA\] c), we choose $\varphi \in \mathcal{P},$ here we choose eigenvectors of $L.$ \[GF\]Let $G$ be a stratified group and let $\varphi \in $ $\mathcal{S}^{\prime }(G)\cap \mathcal{C}^{\infty }(G)$ be an eigenvector of $L$ such that $L\varphi =\lambda \varphi .$ We assume that, for $n\geq 1,$ \(i) $\frac{d^{n}}{dt^{n}}\mid _{t=0}\int_{G}\delta _{t}(\varphi )(\gamma g^{-1})p(g)dg=\int_{G}\frac{d^{n}}{dt^{n}}\mid _{t=0}\delta _{t}(\varphi )(\gamma g^{-1})p(g)dg$ \(ii) $\frac{d^{n}}{dt^{n}}\mid _{t=0}\delta _{t}(\varphi )$ is a polynomial on $G.$ Let $$\ f_{t}=e^{\frac{t^{2}\lambda }{2}}\delta _{t}(\varphi ),\;t>0;\;h_{n}=\frac{d^{n}}{dt^{n}}\mid _{t=0}f_{t}.$$ Then $h_{n}$ is a polynomial on $G$ and $$\cos ^{N}\theta (h_{n})=\cos ^{n}\theta \;h_{n}.$$ Proof: Since $\varphi \in $ $\mathcal{C}^{\infty }(G)$, $t\rightarrow f_{t}$ is $\mathcal{C}^{\infty }$ on $\mathbb{R}^{+}.$ By (\[1\]) $L\circ \delta _{t}(\varphi )=t^{2}\lambda \delta _{t}(\varphi )$, so that $\delta _{t}(\varphi )=e^{-\frac{L}{2}}f_{t}.$ By lemma \[LA\] b) $$e^{-\frac{L}{2}}\cos ^{N}\theta (f_{t})=\delta _{\cos \theta }e^{-\frac{L}{2}}f_{t}=\delta _{\cos \theta }\delta _{t}(\varphi )=\delta _{t\cos \theta }(\varphi )=e^{-\frac{L}{2}}f_{t\cos \theta }. \label{213}$$ We claim that $$\frac{d^{n}}{dt^{n}}\mid _{t=0}e^{-\frac{L}{2}}\cos ^{N}\theta (f_{t})=e^{-\frac{L}{2}}\cos ^{N}\theta (\frac{d^{n}}{dt^{n}}\mid _{t=0}f_{t})=e^{-\frac{L}{2}}\cos ^{N}\theta (h_{n}). \label{214}$$ In particular, applying (\[214\]) with $\theta =0,\frac{d^{n}}{dt^{n}}\mid _{t=0}e^{-\frac{L}{2}}(f_{t})=e^{-\frac{L}{2}}(h_{n}).$ Hence, by (\[214\]) and (\[213\]), $$e^{-\frac{L}{2}}\cos ^{N}\theta \;(h_{n})=\frac{d^{n}}{dt^{n}}\mid _{t=0}e^{-\frac{L}{2}}f_{t\cos \theta }=e^{-\frac{L}{2}}\cos ^{n}\theta \;h_{n}. \label{216}$$ By Leibnitz rule, it is enough to prove the claim for $\delta _{t}(\varphi )$ instead of $f_{t}.$ By lemma \[LA\] b) we may replace $e^{-\frac{L}{2}}\cos ^{N}\theta $ in the claim by $\delta _{\cos \theta }e^{-\frac{L}{2}}.$ The claim now follows from assumption (i). By Leibnitz rule and assumption (ii), $h_{n}$ is a polynomial. So is $\cos ^{N}\theta (h_{n})$ and the result follows from (\[216\]) since $e^{-\frac{L}{2}}$ is one to one on $\mathcal{P}.$ *Remark 5:* $\varphi $ and $\varphi \circ \delta _{\beta },\beta >0,$ give colinear $h_{n}$’s, since $$\frac{d^{n}}{dt^{n}}\mid _{t=0}e^{\frac{1}{2}t^{2}\beta ^{2}\lambda }\delta _{t\beta }(\varphi )=\beta ^{n}\frac{d^{n}}{dt^{n}}\mid _{t=0}e^{\frac{1}{2}t^{2}\lambda }\delta _{t}(\varphi )=\beta ^{n}h_{n}.$$ ### A total set of eigenvectors of $L$ in $L^{q}(pdg),1\leq q<\infty .$ Let $\Pi :G\rightarrow B(L^{2}(\mathbb{R}^{k},d\xi ))$ be a non trivial unitary irreducible representation of $G.$ By definition, $F\in L^{2}(\mathbb{R}^{k})$ is a $\mathcal{C}^{\infty }$ vector for $\Pi $ if the vector valued function: $g\rightarrow \Pi (g)(F)$ is $\mathcal{C}^{\infty }$ on $G.$ We still denote by $\Pi $ the associated differential representation, defined for a $\mathcal{C}^{\infty }$ vector $F$ and $X\in \mathcal{G}$ by $$X\Pi (g)(F)=\frac{d}{dt}\mid _{t=0}\Pi (g\exp tX)(F)=\Pi (g)\Pi (X)(F),\;g\in G, \label{50}$$ and $\Pi (X^{m})=\Pi (X)^{m},$ see e.g. [@CG p.227]; by definition, $\Pi (X^{m})(F)$ still lies in $L^{2}(\mathbb{R}^{k})$ and is still a $\mathcal{C}^{\infty }$ vector for $\Pi .$ $\Pi $ extends as a representation of the convolution algebra $M(G) $ by $$\Pi (\mu )=\int\limits_{G}\Pi (g)d\mu (g).$$In particular ($\Pi (p_{t}dg))_{t\geq 0}$ is a semigroup of operators on $L^{2}(\mathbb{R}^{k}),$ whose generator is $-\Pi (L).$ Indeed, for a $\mathcal{C}^{\infty }$ vector $F$, by (\[50\]), $$\begin{aligned} -\frac{d}{dt}\int_{G}\Pi (g)(F)p_{t}(g)dg &=&\int_{G}\Pi (g)(F)(Lp_{t})(g)dg=\int_{G}L\circ \Pi (g)(F)p_{t}(g)dg \\ &=&\int_{G}\Pi (g)\circ \Pi (L)(F)p_{t}(g)dg\rightarrow _{t\rightarrow 0^{+}}\Pi (L)(F).\end{aligned}$$ Since $p\in \mathcal{S}(G)$, $\Pi (pdg)=e^{-\frac{1}{2}\Pi (L)}$ is a trace class operator [@CG theorem 4.2.1]; in particular its non zero eigenvalues are {$e^{-\frac{1}{2}\lambda },\lambda \in \sigma _{2}(\Pi (L))\},$ where $\lambda $ runs through the eigenvalues of $\Pi (L)$ on $L^{2}(\mathbb{R}^{k})$. Moreover, for $F\in L^{2}(\mathbb{R}^{k}),$ the function $\Pi (pdg)(F)$ is a $\mathcal{C}^{\infty }$ vector for $\Pi $ [@CG theorem A.2.7 p. 241]. Let $\mathcal{U}$ be a set of non trivial unitary irreducible representations of $G$ whose equivalence classes support the Plancherel measure for $G.$ By Kirillov theory, there exists an integer $k,$ which does not depend on $\Pi \in \mathcal{U}$ , such that $\Pi :G\rightarrow B(L^{2}(\mathbb{R}^{k}))$, see more details in 3.5.4 below. \[L5\]Let $G$ be a stratified group and let $\mathcal{F}$ be the set of coefficient functions $$\mathcal{F=}\{\varphi ^{\Pi ,\mu ,\mu ^{\prime }}=\left\langle \Pi (.)(F_{\mu }),F_{\mu ^{\prime }}\right\rangle \mid \Pi \in \mathcal{U},F_{\mu },F_{\mu ^{\prime }}\in \mathcal{B}_{\Pi }\}\subset L^{\infty }(dg)$$ where $\mathcal{B}_{\Pi }$ is an orthogonal basis of $L^{2}(\mathbb{R}^{k})$ chosen among eigenvectors of $e^{-\frac{1}{2}\Pi (L)}.$ Then $\mathcal{F},$ which lies in $\mathcal{C}^{\infty }(G),$ is a set of eigenvectors of $L$ which is total in $L^{q}(p(g)dg),1\leq q<\infty .$ For fixed $\ \Pi ,\mu $ the functions $\{\varphi ^{\Pi ,\mu ,\mu ^{\prime }}\mid F_{\mu ^{\prime }}\in \mathcal{B}_{\Pi }\}$ are independent and belong to the same eigenspace of $L.$ Proof: a) For every non trivial unitary irreducible representation $\Pi $ of $G,$ since $\Pi (pdg)(F_{\mu })=e^{-\frac{1}{2}\Pi (L)}(F_{\mu })=e^{-\frac{1}{2}\lambda _{\mu }}F_{\mu },$ $F_{\mu }$ is a $\mathcal{C}^{\infty }$ vector for $\Pi ,$ hence $\varphi ^{\Pi ,\mu ,\mu ^{\prime }}\in \mathcal{C}^{\infty }(G)$; $\varphi ^{\Pi ,\mu ,\mu ^{\prime }}$ is an eigenvector of $L$ with eigenvalue $\lambda _{\mu }$ by (\[50\]). Since $\Pi $ is irreducible, the closed invariant subspace $$\{F\in L^{2}(\mathbb{R}^{k})\mid \forall g\in G\;\left\langle \Pi (g)(F_{\mu }),F\right\rangle =0\}$$ is reduced to $\{0\},$ which implies the independence of the $\varphi ^{\Pi ,\mu ,\mu ^{\prime }}$’s. (In the Heisenberg case, see [@T p. 19, 51]). b\) Let $\psi \in L^{q^{\prime }}(pdg),\frac{1}{q}+\frac{1}{q^{\prime }}=1,$ be orthogonal to $\mathcal{F}$, i.e. for $\Pi \in $ $\mathcal{U},$ $$0=\int_{G}\left\langle \Pi (g)(F_{\mu }),F_{\mu ^{\prime }}\right\rangle \psi (g)p(g)dg=\left\langle (\int_{G}\Pi (g)\psi (g)p(g)dg)(F_{\mu }),F_{\mu ^{\prime }}\right\rangle .$$ Equivalently $\Pi (\psi p)=\widehat{\psi p}(\Pi )=0$ for $\Pi \in \mathcal{U}.$ Then Plancherel formula for $G$ (see e.g. [CG]{}) implies that $\psi p=0$ $dg$ a.s.. Indeed, this is clear if $\psi p\in L^{2}(dg),$ in particular if $q^{\prime }\geq 2.$ In general, $\psi p\in L^{1}(dg),\left\Vert (\psi p)\ast p_{t}-\psi p\right\Vert _{L^{1}(dg)}\rightarrow _{t\rightarrow 0}0$ and $(\psi p)\ast p_{t}\in L^{2}(dg);$ moreover $(\psi p)\ast p_{t}=0$ a.s. since, for every $\Pi \in \mathcal{U},$ $$\Pi ((\psi p)\ast p_{t})=\Pi (\psi p)\Pi (p_{t})=0.\blacksquare$$ ### Polynomial eigenvectors of N built from coefficients of representations We now consider the functions $e^{\frac{1}{2}t^{2}\lambda _{\mu }}\varphi ^{\Pi ,\mu ,\mu ^{\prime }}\circ \delta _{t}$ as generating functions of polynomial eigenvectors of $N.$ \[P5\] Let $\varphi ^{\Pi ,\mu ,\mu ^{\prime }}=\left\langle \Pi (.)(F_{\mu }),F_{\mu ^{\prime }}\right\rangle \in \mathcal{F}$ be as in proposition \[L5\]. For $n\geq 1,$ let$$h_{n}^{\Pi ,\mu ,\mu ^{\prime }}=\frac{d^{n}}{dt^{n}}\mid _{t=0}e^{\frac{1}{2}t^{2}\lambda _{\mu }}\varphi ^{\Pi ,\mu ,\mu ^{\prime }}\circ \delta _{t}.$$ Then $h_{n}^{\Pi ,\mu ,\mu ^{\prime }}$ is a polynomial eigenvector of $\cos ^{N}\theta $ with eigenvalue $\cos ^{n}\theta $. Proof: By proposition \[L5\] and lemma \[GF\], it is enough to prove assumptions (i) and (ii) in lemma \[GF\]. We claim the existence of a polynomial $\psi _{n},$ $n\geq 1,$ which does not depend on $t,$ such that, for $0\leq t\leq 1$ and $n\geq 0,$$$\;\;\left\vert \frac{d^{n}}{dt^{n}}\;\varphi ^{\Pi ,\mu ,\mu ^{\prime }}\circ \delta _{t}\right\vert \leq \psi _{n}.$$  Since $g\rightarrow \psi _{n}(\gamma g^{-1})$ is still a polynomial, it lies in $L^{1}(pdg)$ for every $\gamma \in G,$ and this will prove assumption (i). We now verify the claim. **Case 1:** The computation of derivatives being easier if $G$ is step two, we first consider this setting. By Schur lemma, the restriction of $\Pi $ to the center $\exp \mathcal{Z}$ of $G$ is given by a character $u\rightarrow e^{i\left\langle l,u\right\rangle }$ where $l$ is some linear form on $\mathcal{Z},$ see e.g. [@CG p. 184]. If $g=(x,u)$ and $X=\sum_{j=1}^{n}x_{j}X_{j}\in V_{1},$ $$\varphi ^{\Pi ,\mu ,\mu ^{\prime }}(\delta _{t}g)=e^{it^{2}\left\langle l,u\right\rangle }\left\langle \Pi (\exp tX)(F_{\mu }),F_{\mu ^{\prime }}\right\rangle =e^{it^{2}\left\langle l,u\right\rangle }\Phi _{t}^{\Pi ,\mu ,\mu ^{\prime }}(x)$$ and, by (\[50\]), $$\frac{d^{m}}{dt^{m}}\Phi _{t}^{\Pi ,\mu ,\mu ^{\prime }}(x)=\left\langle \Pi (\exp tX)\Pi (X)^{m}(F_{\mu }),F_{\mu ^{\prime }}\right\rangle . \label{215}$$ Since $\Pi (X)^{m}(F_{\mu })$ lies in $L^{2}(\mathbb{R}^{k},d\xi ),\left\langle \Pi (X)^{m}(F_{\mu }),F_{\mu ^{\prime }}\right\rangle $ and $\left\Vert \Pi (X)^{m}(F_{\mu })\right\Vert _{L^{2}(d\xi )}$ are polynomials w.r. to $x,$ $\frac{d^{m}}{dt^{m}}\mid _{\alpha =0}\delta _{t}(\varphi ^{\Pi ,\mu ,\mu ^{\prime }})$ is a polynomial w.r. to $x,u,$ and $\left\vert \frac{d^{m}}{dt^{m}}e^{it^{2}\left\langle l,u\right\rangle }\Phi _{t}^{\Pi ,\mu ,\mu ^{\prime }}(x)\right\vert $ is, for $0\leq t\leq 1,$ less than a polynomial $\psi _{n}$ which does not depend on $t.$ This proves (i) and (ii) in this case. **General case:** As in (\[100\]) and (\[501\]), for $g=\exp Z=\exp (X+Y+..+U)$ and $t>0,$ since $V(\Pi (\delta _{t}Z))=\Pi (V(\delta _{t}Z)),$ $$\frac{d}{dt}\varphi ^{\Pi ,\mu ,\mu ^{\prime }}(\delta _{t}g)=\frac{d}{dt}\left\langle \exp \Pi (\delta _{t}Z)(F_{\mu }),F_{\mu ^{\prime }}\right\rangle =\left\langle \Pi (V(\delta _{t}Z))(F_{\mu }),\exp -\Pi (\delta _{t}Z)(F_{\mu ^{\prime }})\right\rangle .$$ At $t=0$ this reduces to the polynomial $\left\langle \Pi (X)(F_{\mu }),F_{\mu ^{\prime }}\right\rangle $. Since $\Pi (V(\delta _{t}Z)$ has polynomial coefficients w.r. to $t$ and the coordinates of $g,$ so does $\left\Vert \Pi (V(\delta _{t}Z))(F_{\mu })\right\Vert _{L^{2}(d\xi )}.$ Hence there is a polynomial $\psi _{1}$ w.r. to the coordinates of $g$ such that $$sup_{0\leq t\leq 1}\left\Vert \Pi (V(\delta _{t}Z))(F_{\mu })\right\Vert _{L^{2}(d\xi )}\leq \psi _{1}.$$ This proves the claim for $n=1$. Clearly this can be iterated for upper derivatives, which proves (i) and (ii).$\blacksquare $ ### The step two setting: generalized Hermite polynomials In this case, the key facts are the extension of the explicit functions $\varphi ^{\Pi ,\mu ,\mu ^{\prime }}\in \mathcal{F}$ as entire functions on the complexification of $G$ and the explicit expression of $p.$ Theorem [9]{} gives another proof of theorem \[spectre\] a) in this setting, with another description of the eigenspaces of $N$ by generating functions. \[9\]Let $G$ be a step two stratified group. Then a\) every $\varphi ^{\Pi ,\mu ,\mu ^{\prime }}\in \mathcal{F}$ lies in the closed subspace of $L^{q}(pdg),1\leq q<\infty ,$ spanned by constants and the polynomials $\{h_{n}^{\Pi ,\mu ,\mu ^{\prime }},n\geq 1\}$ defined in proposition \[P5\]$.$ b\) The set of generalized Hermite polynomials$$\cup _{\varphi ^{\Pi ,\mu ,\mu ^{\prime }}\in \mathcal{F}}\{h_{n}^{\Pi ,\mu ,\mu ^{\prime }},n\geq 1\}$$ together with the constants is a set of eigenvectors of $N$ which is total in $L^{q}(pdg),1\leq q<\infty .$ c\) For $\ $fixed $n\geq 1,$ $\cup _{\varphi ^{\Pi ,\mu ,\mu ^{\prime }}\in \mathcal{F}}\{h_{n}^{\Pi ,\mu ,\mu ^{\prime }}\}$ spans the eigenspace of $N$ associated to $n$ in $L^{q}(pdg),1<q<\infty .$ In contrast, if $G$ has more than 4 layers, assertion b) is false by proposition \[dense\], hence a) is false for some $\varphi ^{\Pi ,\mu ,\mu ^{\prime }}\in \mathcal{F}$, by proposition \[L5\]$.$ If $G$ has 3 or 4 layers, we do not know if the conclusions of theorem \[9\] hold true. Proof of theorem \[9\]: a) implies b) by propositions \[L5\] and \[P5\]. b\) implies c) as recalled in the proof of theorem \[spectre\].       a) The proof is given in three steps. In step 1 we state two standard sufficient conditions ensuring statement a); in step 2 we verify these conditions when $G$ is a Heisenberg group; in step 3 we show how the general step 2 case mimicks the Heisenberg  case.      **Step 1:**  Let $\varphi ^{\Pi ,\mu ,\mu ^{\prime }}\in \mathcal{F}$ and assume that \(i) for every $g\in G,$ the function $t\rightarrow \varphi ^{\Pi ,\mu ,\mu ^{\prime }}(\delta _{t}g)$ extends as a holomorphic function $z\rightarrow \varphi _{z}^{\Pi ,\mu ,\mu ^{\prime }}(g)$ on $\mathbb{C}.$ \(ii) for some connected neighborhood $\Omega $ of the real axis, for every compact $K\subset \Omega ,\ $there exists $w_{K}\in L^{q}(pdg),$ $1\leq q<\infty ,$ such that $$\left\vert \varphi _{z}^{\Pi ,\mu ,\mu ^{\prime }}\right\vert \leq w_{K},\;z\in K.$$ We claim that $\varphi ^{\Pi ,\mu ,\mu ^{\prime }}=\varphi $ then lies in the closed subspace of $L^{q}(pdg)$ spanned by $h_{n}^{\Pi ,\mu ,\mu ^{\prime }},n\geq 1.$ Indeed, let $\psi \in L^{q^{\prime }}(pdg),\frac{1}{q}+\frac{1}{q^{\prime }}=1,$ and let $$m(t)=\int_{G}\varphi (\delta _{t}g)\psi (g)p(g)dg.$$ By the assumptions, $m$ extends as a holomorphic function on $\Omega $ and $$\frac{d^{n}}{dz^{n}}m=\int_{G}(\frac{d^{m}}{dz^{m}}\varphi _{z})\psi pdg,\;m\geq 0.$$ By proposition \[L5\], $L(\varphi )=\lambda \varphi $ for some $\lambda =\lambda _{\mu }.$ Hence $t\rightarrow e^{\frac{1}{2}t^{2}\lambda }m(t)$ also extends as a holomorphic function on $\Omega $  and $$\frac{d^{n}}{dz^{n}}\mid _{z=0}e^{\frac{1}{2}z^{2}\lambda }m=\int_{G}[\frac{d^{n}}{dz^{n}}\mid _{z=0}e^{\frac{1}{2}z^{2}\lambda }\varphi _{z}]\psi pdg=\int_{G}h_{n}^{\Pi ,\mu ,\mu ^{\prime }}\psi pdg,\;n\geq 0.$$ If $\psi $ is orthogonal to $\{h_{n}^{\Pi ,\mu ,\mu ^{\prime }},n\geq 0\},$ these derivatives are zero, hence $e^{\frac{1}{2}z^{2}\lambda }m$  is zero on $\Omega .$ In particular $m(1)=0,$ i.e. $\psi $ is orthogonal to $\varphi ,$ which proves the claim. **Step 2: The Heisenberg groups** $\mathbb{H}_{k}$ A basis of the first layer of the Lie algebra is $X_{1},Y_{1},..,X_{k},Y_{k}$ where $[X_{j},Y_{j}]=-4U,$ $U$ spans the center, and the other commutators are zero. By the Campbell-Hausdorff formula, $$g=\exp (\sum_{j=1}^{k}x_{j}X_{j}+y_{j}Y_{j}+uU)=\exp uU\ \prod\limits_{j=1}^{k}\exp (-2x_{j}y_{j}U)\exp y_{j}Y_{j}\exp x_{j}X_{j}.$$ We first consider the Schrödinger (unitary irreducible) representation $\Pi _{S}:\mathbb{H}_{k}\rightarrow B(L^{2}(\mathbb{R}^{k})),$ defined on the Lie algebra by $$\Pi _{S}(X_{j})=\frac{\partial }{\partial \xi _{j}},\;\Pi _{S}(Y_{j})=i\xi _{j},\;\Pi _{S}(U)=-\frac{1}{4}[\frac{\partial }{\partial \xi _{j}},i\xi _{j}]=-\frac{i}{4}I.$$For $F\in L^{2}(\mathbb{R}^{k}),$ this implies $$\Pi _{S}(g)(F)(\xi )=e^{-i\frac{u}{4}}e^{\frac{i}{2}\sum_{j=1}^{k}x_{j}y_{j}}e^{i\sum_{j=1}^{k}y_{j}\xi _{j}}F(\xi +x), \label{19}$$  and $$\Pi _{S}(L)=H=\sum_{j=1}^{k}(-\frac{\partial ^{2}}{\partial \xi _{j}^{2}}+\xi _{j}^{2})$$ is the harmonic oscillator. If $k=1,$ an o.n. basis of eigenvectors of $H$ in $L^{2}(\mathbb{R})$ is the sequence of Hermite functions $F_{\mu },\mu \in \mathbb{N}.$ The so called special Hermite functions [@T p. 18-19] are, for $\mu ,\mu ^{\prime }\in \mathbb{N}$ and $\varepsilon _{\mu ,\mu ^{\prime }}=sgn(\mu ^{\prime }-\mu ),$ $$\begin{aligned} \left\langle \Pi _{S}(x,y,0)(F_{\mu }),F_{\mu ^{\prime }}\right\rangle &=&\Phi _{\mu ,\mu ^{\prime }}(x,y)=\int_{\mathbb{R}}e^{iy\xi }F_{\mu }(\xi +\frac{x}{2})F_{\mu ^{\prime }}(\xi -\frac{x}{2})d\xi \nonumber \\ &=&r_{\mu ,\mu ^{\prime }}(x^{2}+y^{2})e^{-\frac{1}{2}(x^{2}+y^{2})}(x+i\varepsilon _{\mu ,\mu ^{\prime }}y)^{\left\vert \mu -\mu ^{\prime }\right\vert }, \label{32}\end{aligned}$$ where $r_{\mu ,\mu ^{\prime }}=r_{\mu ^{\prime },\mu }$ is a one variable polynomial with real coefficients. An o.n basis of eigenvectors of $H$ in $L^{2}(\mathbb{R}^{k})$ is the sequence $\left( \prod\limits_{j=1}^{k}F_{\mu _{j}}(\xi _{j})\right) _{\mu \in \mathbb{N}^{k}}$, which gives, for $\mu ,\mu ^{\prime }\in \mathbb{N}^{k}$ and $g=(x,y,u),$ $$\varphi ^{\Pi _{S},\mu ,\mu ^{\prime }}(g)=\left\langle \Pi (g)(F_{\mu }),F_{\mu ^{\prime }}\right\rangle =e^{-i\frac{u}{4}}\prod\limits_{j=1}^{k}\Phi _{\mu _{j},\mu _{j}^{\prime }}(x_{j},y_{j}).$$ By (\[32\]) the function $z\rightarrow \varphi ^{\Pi _{S},\mu ,\mu ^{\prime }}(zx,zy,z^{2}u)$ is holomorphic on $\mathbb{C}$. Let $$R_{a,\delta }=\{\alpha +i\beta \mid \left\vert \alpha \right\vert <a,\left\vert \beta \right\vert <\delta \}\subset \mathbb{C}.$$ For some constant $C_{a,\delta },$ and $z\in $ $\overline{R_{a,\delta }},$ $$\left\vert \varphi ^{\Pi _{S},\mu ,\mu ^{\prime }}(zx,zy,z^{2}u)\right\vert \leq C_{a,\delta }e^{\frac{1}{2}a\delta \left\vert u\right\vert }\prod\limits_{j=1}^{k}e^{\delta ^{2}(x_{j}^{2}+y_{j}^{2})}.$$ We now look for conditions on $a,\delta $ ensuring that the right hand side lies in $L^{q}(pdg).$ We recall [@Hu] that $$p(x,y,u)=\int_{\mathbb{R}}e^{i\lambda u}Q(x,y,\lambda )d\lambda =c_{k}\int_{\mathbb{R}}e^{i\lambda u}\prod\limits_{j=1}^{k}\frac{2\lambda }{sh2\lambda }e^{-\frac{\lambda }{th2\lambda }(x_{j}^{2}+y_{j}^{2})}d\lambda .$$ Noting that $Q(x,y,\lambda )=\prod\limits_{j=1}^{k}$ $Q_{1}(x_{j},y_{j},\lambda )$ is even w.r. to $\lambda ,$ we get, for $q\geq 1,$ $$\frac{1}{2}\int_{\mathbb{R}}e^{\frac{q}{2}a\delta \left\vert u\right\vert }p(x,y,u)du\leq \int_{\mathbb{R}}ch(\frac{q}{2}a\delta u)p(x,y,u)du=Q(x,y,ia\delta \frac{q}{2}).$$ We need the convergence of $$\int_{\mathbb{R}^{2k}}\prod\limits_{j=1}^{k}e^{q\delta ^{2}(x_{j}^{2}+y_{j}^{2})}Q_{1}(x_{j},y_{j},i\frac{q}{2}a\delta )dxdy=c\prod\limits_{j=1}^{k}\int_{\mathbb{R}^{2}}e^{(q\delta ^{2}-\frac{1}{2}\frac{qa\delta }{tgqa\delta })(x_{j}^{2}+y_{j}^{2})}dx_{j}dy_{j},$$ which holds for $qa\delta \leq \frac{\pi }{4}$ and $a>2\delta .$ Thus, taking $a=N\in \mathbb{N},$ $\varphi ^{\Pi _{S},\mu ,\mu ^{\prime }}$ satisfies the assumptions of step 1 on $$\Omega =\cup _{N\geq 2}R_{N,\frac{\pi }{4qN}}.$$ Plancherel formula for $\mathbb{H}_{k}$ (see e.g. [@T Theorem 1.3.1] or [@CG p.154]) involves the representations $$\rho _{h}(x,y,u)=e^{-\frac{i}{4}hu}\Pi _{S}(x,hy,0).$$ By the Stone-Von Neumann theorem [@T theorem 1.2.1] every irreducible unitary representation $\Pi $ of $\mathbb{H}_{k}$ satisfying $\Pi (0,0,u)=e^{-\frac{i}{4}hu}$ for a real $h\neq 0$ is unitarily equivalent to $\rho _{h}.$ Hence $\rho _{\beta ^{2}}$ (resp. $\rho _{-\beta ^{2}})$ is unitarily equivalent to $\Pi _{S}\circ \delta _{\beta }$, (resp. $\Pi _{S}\circ \sigma \circ \delta _{\beta })$, $\beta >0,$ where $\sigma $ is the automorphism of $\mathbb{H}_{k}$ defined by $\sigma (x,y,u)=(x,-y,-u)).$  Since $\Pi _{S}(L)=\Pi _{S}\circ \sigma (L),$ we get $\varphi ^{\Pi _{S}\circ \sigma ,\mu ,\mu ^{\prime }}=\varphi ^{\Pi _{S},\mu ,\mu ^{\prime }}\circ \sigma =\overline{\varphi ^{\Pi ,\mu ,\mu ^{\prime }}},$ hence $$\mathcal{F}=\{\varphi ^{\Pi _{S},\mu ,\mu ^{\prime }}\circ \delta _{\beta },\overline{\varphi ^{\Pi _{S},\mu ,\mu ^{\prime }}}\circ \delta _{\beta },\beta >0,\mu ,\mu ^{\prime }\in \mathbb{N}^{k}\}.$$ The conditions of step 1 are satisfied by $\varphi ^{\Pi ,\mu ,\mu ^{\prime }}\circ \delta _{\beta },$ replacing $R_{a,\delta }$ by $R_{\beta a,\beta \delta },$ which ends the proof of theorem \[9\] for $\mathbb{H}_{k}$. Taking remark 5 into account, the set $\cup _{\mu ,\mu ^{\prime },n}\{h_{n}^{\Pi _{S},\mu ,\mu ^{\prime }},\overline{h_{n}^{\Pi _{S},\mu ,\mu ^{\prime }}}\}$ is total in $L^{2}(\mathbb{H}_{k},pdg).$ **Step 3.** We first recall some more facts on representations and compute the set $\mathcal{F}$ for step 2 stratified groups. We shall follow Cygan’s scheme [@Cy].      Let $l\in \mathcal{G}^{\ast }.$ Among the Lie subalgebras $\mathcal{M\subset G}$ satisfying $\left\langle l,[X,Y]\right\rangle =0$ for every $X,Y\in \mathcal{M},$ some have minimal codimension $m_{l}$ and are denoted by $\mathcal{M}_{l}$. Then the map $$Z\in \mathcal{M}_{l}\rightarrow e^{i\left\langle l,Z\right\rangle } \label{88}$$ is a representation of the subgroup $\exp \mathcal{M}_{l}$ and induces an irreducible unitary representation of $G$ as follows [@CG theorems 1.3.3, 2.2.1 and p 41] : One chooses independent vectors $(X_{j})_{i=1}^{m_{l}}$ such that $\mathcal{G}=\mathcal{M}_{l}+span\{(X_{j})_{i=1}^{m_{l}}\}.$ For ($g,\xi )\in G$ $\times \mathbb{R}^{m_{l}}$ there exist ($\xi ^{\prime },M)\in \mathbb{R}^{m_{l}}\times \mathcal{M}_{l}$ such that $$\exp (\sum\limits_{i=1}^{m_{l}}\xi _{i}X_{i}).\;g=\exp M\;.\exp (\sum\limits_{i=1}^{m}{}\xi _{i}^{\prime }X_{i}).$$ Then, for $F\in L^{2}(\mathbb{R}^{m_{l}}),$ $$\Pi _{l}(g)(F)(\xi )=e^{i\left\langle l,M\right\rangle }F(\xi ^{\prime }). \label{29}$$ The set of $\mathcal{C}^{\infty }$ vectors for $\Pi _{l}$ is $\mathcal{S(\mathbb{R}}^{k}\mathcal{)}$ [@CG corollary 4.1.2]. Every irreducible unitary representation of $G$ is equivalent to a representation constructed in this way; different $\mathcal{M}_{l},\mathcal{M}_{l}^{\prime } $ and different $l,l^{\prime }$  in the same coadjoint orbit induce equivalent representations [@CG theorems 2.2.2, 2.2.3, 2.2.4]. By Kirillov theory there is an integer $k$ and a set $\mathcal{U}_{0}\mathcal{\subset G}^{\ast }$  of “generic” orbits with maximal dimension $2k,$ such that $m_{l}=k$ for $l\in \mathcal{U}_{0}.$ The Plancherel measure is supported by $\mathcal{U}_{0}$ [@CG theorem 4.3.10]. We now compute such a $\Pi _{l}$ when $G$** **is** **step 2**.** Let $U_{1},..,U_{d}$ be a basis of the central layer $\mathcal{Z}$ and let $\chi _{1},.,\chi _{n}$ be a basis of the first layer $V_{1}$ of $\mathcal{G}$. Let $l\mathcal{\in G}^{\ast }$ and let $\lambda =\sum_{j=1}^{d}\lambda _{j}U_{j}^{\ast }$  be its central part, identified with a vector $\lambda \in $ $\mathbb{R}^{d}.$ Let $A_{\lambda }$ be the $n\times n$ matrix with coefficients $\left\langle \lambda ,[\chi _{j},\chi _{h}]\right\rangle .$ By Campbell-Hausdorff formula, for $Y\in \mathcal{G},X\in V_{1},U\in \mathcal{Z},g=\exp (X+U),$ $$\exp Adg(Y)=g\;\exp Y\;g^{-1}=e^{[X,Y]}\exp Y=\exp (Y+[X,Y]),$$ hence the coadjoint orbit of $l,$  i.e. $\{l\circ Adg,\;g\in \mathcal{G}\}\subset \mathcal{G}^{\ast },$ is $l+range\;A_{\lambda }.$ We now assume that $l$ lies in $\mathcal{U}_{0},$ so that the range of $A_{\lambda }$ has dimension $2k.$There exists an orthogonal matrix $\Omega _{\lambda }$ such that $$A_{\lambda }=\Omega _{\lambda }A_{\lambda }^{\prime }\Omega _{\lambda }^{\ast }$$ where $A_{\lambda }^{\prime }$ is block diagonal, the non zero blocks having the form $$\nu _{j}(\lambda )\left( \begin{array}{cc} 0 & 1 \\ -1 & 0\end{array}\right) ,\;\nu _{j}(\lambda )>0. \label{258}$$ The new basis of $V_{1}$ (defined by the columns of $\Omega _{\lambda })$ is denoted by $X_{1},Y_{1},..,$   $X_{k},Y_{k},S_{1},...,S_{n-2k},$ so that $$\left\langle \lambda ,[X_{j},X_{h}]\right\rangle =0=\left\langle \lambda ,[Y_{j},Y_{h}]\right\rangle ,\;\left\langle \lambda ,[X_{j},Y_{h}]\right\rangle =\nu _{j}(\lambda )\delta _{jh},\;1\leq j,h\leq k. \label{256}$$ We denote $t=\Omega _{\lambda }(x,y,s)\in \mathbb{R}^{n},$ where $$\sum_{j=1}^{n}t_{j}\chi _{j}=\sum_{j=1}^{k}x_{j}X_{j}+y_{j}Y_{j}+\sum_{h=1}^{n-2k}s_{h}S_{h}=X+Y+S\in V_{1}.$$ Choosing $\mathcal{M}_{l}=\mathcal{Z+}span\{Y_{j},S_{h}\}_{1\leq j\leq k,1\leq h\leq n-2k}$, let us compute $\Pi _{l}.$ By definition $\Pi _{l}(\exp u_{j}U_{j})=e^{iu_{j}\lambda _{j}}$. For $g=\exp (X+Z)$ and $Z=Y+S, $ $$\begin{aligned} \exp (\sum_{j=1}^{k}\xi _{j}X_{j})g &=&\exp [\sum_{j=1}^{k}\xi _{j}X_{j},X+Z]\;g\;\exp (\sum_{j=1}^{k}\xi _{j}X_{j}) \\ &=&\exp ([\sum_{j=1}^{k}\xi _{j}X_{j},X+Z]+\frac{1}{2}[X,Z])\exp Z\exp X\exp (\sum_{j=1}^{k}\xi _{j}X_{j}) \\ &=&\exp M\exp (\sum_{j=1}^{k}(\xi _{j}+x_{j})X_{j}).\end{aligned}$$ Hence, by (\[29\]) and (\[256\]), for $F\in L^{2}(\mathbb{R}^{k}),$ $$\Pi _{l}(g)(F)(\xi )=e^{i\left\langle l,M\right\rangle }F(\xi +x)=e^{i\sum_{j=1}^{k}\nu _{j}y_{j}(\xi _{j}+\frac{1}{2}x_{j})}e^{i\left\langle l,Y+S\right\rangle }F(\xi +x). \label{31}$$ Since we may replace $l$ by $l^{\prime }$ in the orbit of $l,$ we may suppose $\left\langle l,Y_{j}\right\rangle =0,$ $1\leq j\leq k.$ In particular, by (\[31\]), $$\Pi _{l}(X_{j})=\frac{\partial }{\partial \xi _{j}},\Pi _{l}(Y_{j})=i\nu _{j}\xi _{j},\;1\leq j\leq k,\Pi _{l}(S_{h})=i\left\langle l,S_{h}\right\rangle I,\;1\leq h\leq n-2k.$$ Since $\Omega _{\lambda }$ is orthogonal, $-L=\sum_{j=1}^{k}(X_{j}^{2}+Y_{j}^{2})+\sum_{h=1}^{n-2k}S_{h}^{2},$ which entails $$\Pi _{l}(L)=\sum_{j=1}^{k}-\frac{\partial ^{2}}{\partial \xi _{j}^{2}}+\nu _{j}^{2}\xi _{j}^{2}+\sum_{h=1}^{n-2k}\left\langle l,S_{h}\right\rangle ^{2}I.$$ A basis of eigenvectors of $\Pi _{l}(L)$ is thus $\left( \prod\limits_{j=1}^{k}F_{\mu _{j}}(\sqrt{\nu _{j}}\xi _{j})\right) _{\mu \in \mathbb{N}^{k}}$. By (\[31\]) and (\[32\]), for $g=(x,y,s,u),$ $$\varphi ^{\Pi _{l},\mu ,\mu ^{\prime }}(g)=e^{i\left\langle \lambda ,u\right\rangle }e^{i\sum_{h=1}^{n-2k}s_{h}\left\langle l,S_{h}\right\rangle }\prod\limits_{j=1}^{k}\frac{1}{\sqrt{\nu _{j}}}\Phi _{\mu _{j,}\mu _{j}^{\prime }}(\sqrt{\nu _{j}}x_{j},\sqrt{\nu _{j}}y_{j}).$$ Hence, for $z\in R_{a,\delta }$ and some constant $C_{a,\delta },$ with $t=$ $\Omega _{\lambda }(x,y,s),$ $$\begin{aligned} \left\vert \varphi ^{\Pi _{l},\mu ,\mu ^{\prime }}(zt,z^{2}u)\right\vert &\leq &C_{a,\delta }e^{2a\delta \left\vert \left\langle \lambda ,u\right\rangle \right\vert }e^{\delta \sum_{h=1}^{n-2k}\left\vert s_{h}\left\langle l,S_{h}\right\rangle \right\vert }\prod\limits_{j=1}^{k}\frac{1}{\sqrt{\nu _{j}}}e^{\delta ^{2}\nu _{j}(x_{j}^{2}+y_{j}^{2})} \\ &=&e^{2a\delta \left\vert \left\langle \lambda ,u\right\rangle \right\vert }w_{a,\delta ,l}(x,y,s).\end{aligned}$$ By [@Cy corollary 5.5] the heat kernel $p(t,u)$ is the Fourier transform of $CQ(t,\lambda )$ w.r. to the central variables, where $$Q(t,\lambda )=\prod\limits_{h=1}^{n-2k}e^{-\frac{1}{2}s_{h}^{2}}\prod\limits_{j=1}^{k}Q_{1}(x_{j},y_{j},\frac{\nu _{j}}{4})=Q(t,-\lambda ).$$ Again, we need the convergence  of $$\int_{\mathbb{R}^{n}}w_{a,\delta ,l}^{q}(x,y,s)\prod\limits_{h=1}^{n-2k}e^{-\frac{1}{2}s_{h}^{2}}\prod\limits_{j=1}^{k}Q_{1}(x_{j},y_{j},\frac{iqa\delta \nu _{j}}{2})dxdyds,$$ which holds if $qa\delta \max \nu _{j}\leq \frac{\pi }{4}$ and $a>2\delta .$ This ends the proof of theorem \[9\].$\blacksquare $ [LP$_{2}$]{} C. Ané et altri: Sur les inégalités de Sobolev logarithmiques, Panoramas et synthèse 10, SMF (2000). F. Baudoin, M. Hairer, J. Teichmann: Ornstein-Uhlenbeck processes on Lie groups, J. Funct. Analysis 255 (2008) 877-890. L. Corwin, F.P. Greenleaf: Representations of nilpotent Lie groups and their applications Part 1: Basic theory and examples, Cambridge Studies in Advanced Math. 18 (2004). T. Coulhon, L. Saloff-Coste, N.TH. Varopoulos: Analysis and geometry on groups, Cambridge UP (1992). P. Crepel, A. Raugi: Théorème central limite sur les groupes nilpotents, Annales de l’IHP, section B, 14-2 (1978) 145-164. J. Cygan: Heat kernels for class 2 nilpotent groups, Studia Math. LXIV (1979) 227-238. E.B Davies: Heat kernels and spectral theory, Cambridge University Press (1989). B. Driver, T. Melcher: Hypoelliptic heat kernel inequalities on the Heisenberg group, J. Funct. Analysis 221 (2005) 340-365. G.B. Folland, E. Stein: Hardy spaces on homogeneous groups, Princeton UP (1982). B.C. Hall: Lie groups, Lie algebras and representations: an elementary introduction. Graduate Text in Maths 222, Springer Verlag (2003). H.Hamburger: Zur Konvergenztheorie der Stieltjesschen Kettenbrüche, Math. Z. 4 (1919) 209-211. A. Hulanicki: The distribution of energy in the Brownian motion in the Gaussian field and analytic hypoellipticity of certain subelliptic operators on the Heisenberg group**, Studia Math. 56 (1976) 165-173. W. Hebisch, B. Zegarlinski: Coercive inequalities on metric measure spaces, preprint. P. Lax: Functional analysis, Wiley (2002). H-Q Li: Estimation optimale du gradient du semi-groupe de la chaleur sur le groupe de Heisenberg, J. Funct. Anal. 236 (2006) 369-394. F. Lust-Piquard: A simple minded computation of heat kernels on Heisenberg groups, Colloquium Math. 97 (2003) 233-249. F. Lust-Piquard: Riesz transforms on deformed Fock spaces, Comm. Math. Phy. 205 (1999) 519-549. E.M. Ouhabaz: Analysis of heat kernels on domains, Princeton UP (2006), London Math. Soc. monograph series 31. G. Pisier: Probabilistic methods in the geometry of Banach spaces. Probability and Analysis, Varenna (Italy) (1985), Lecture Notes in Maths 1206 (1986), 167-241, Springer Verlag. G.Polya, G.Szegö: Problems and theorems in analysis, Springer Verlag. S. Thangavelu: Harmonic analysis on the Heisenberg group, Progress in Maths 159, Birkhauser (1998). V.S. Varadarajan: Lie groups, Lie algebras and their representations, Springer Verlag (1984). Affiliation: équipe AGM-UMR 8088 Address: Dept Maths, PST, Université de Cergy, 2 Av. A. Chauvin, 95803, Cergy, France E-mail: francoise.piquard@math.u-cergy.fr

\ \ [**Turbulent behaviour in magnetic hydrodynamics is not universal** ]{} \ \ [ D. Wolchenkov]{} *Department of Theoretical Physics, State University of St. Petersburg,* Ul’yanovskaya 1, Petrodvoretc, St. Petersburg, Russia \ \ [**Abstract**]{} A short distance expansion method (SDE) that is well known in the quantum field theory for analysis of turbulent behaviour of stochastic magnetic\ hydrodynamics of incompressible conductive fluid is applied. As a result is shown that in an inertial range the turbulent spectra of magnetic hydrodynamics depend on a scale of arising of curls. \ \ Introduction ==============      The methods of quantum field theory are successfully applied for description of the critical phenomena and developed turbulence recently. This approach has an important advantage before the classical one. For example it enables one to renormalize all the correlation functions of a model as well as to define their various asymptotics. A stochastic problem of magnetic hydrodynamics (MHD) of incompressible\ conductive fluid with external random force $ f^\varphi$ and rotor of random current $f^\theta$ $$\label{1} {\cal D}_t v_i = \nu \Delta v_i - \partial_i p +(\theta \partial) \theta_i + f^{v}_i,\quad {\cal D}_t = \partial_t + v \partial,$$ $${\cal D}_t \theta_i = \nu' \Delta \theta_i +(\theta \partial) v_i + f^{\theta}_i$$ is equivalent to a theory of four fields with generating functional of renormalized correlation functions (Green functions) $$G(A_{\phi})=\int {\cal D}\Phi \det M \exp (S_{R}(\Phi)+\Phi A_{\phi}), \quad\Phi=\{\varphi, \varphi',\theta ,\theta'\}$$ in which the renormalized functional of action is $$\label{2} S _R= \frac12 g_1 \nu^3 M^{2 \varepsilon} \varphi' D ^{\varphi\varphi} \varphi' + \frac12 g_2 \nu^3 M^{2a \varepsilon} \theta' D ^{\theta\theta} \theta' + g_3 \nu^3 M^{(1+a) \varepsilon} \varphi' D ^{\varphi \theta} \theta' +$$ $$+ \varphi' \bigl [ -\partial _t \varphi + Z _1\nu \Delta \varphi - ( \varphi \partial) \varphi + Z _3 (\theta \partial)\theta \bigr] + \theta' \bigl [ -\partial _t \theta + Z _2\nu u\Delta \theta - ( \varphi \partial) \theta + (\theta \partial)\varphi \bigr]$$ (necessary integration on $ { \bf x } $ and $t$ as well as summation on repeated badges are implied). Here the fields ${\bf \varphi}({\bf x}, t) $ and $ { \bf \theta } ( { \bf x }, t) $ are both transversal (the vector field of velocity $ {\bf \varphi } $ due to incompressibility of fluid ($\partial _i\varphi _i =0$) and the pseudovector field $ { \bf\theta } $ as connected with transversal field of a magnetic induction $ { \bf B }:$ ${ \bf \theta } = { \bf B } / \sqrt { 4\pi\varrho }$ ($\varrho$ is a density of fluid, $\nu$ is a viscosity, $p$ is a pressure). We shall use the dimensionless constant $\lambda$ (it means inverted number of Prandtl) instead of $\nu'=c^2 / 4\pi\sigma$ ( here $\sigma$ is a conductivity, $c$ is the velocity of light) by this way $\nu'=\lambda\nu$. Under construction of the model existence of an inertial range is implied therein a real external energy pumping (correlators of the random forces) can be simulated by exponential model of $\delta$-function. $$\label{3} D^{vv}_{is} = {g_1}_0 \nu^3_0 P_{is}d_{vv} , \quad D^{\theta\theta}_{is} = {g_2}_0 \nu^3_0 P_{is}d_{\theta\theta} , \quad D^{v\theta}_{is} = {g_3}_0 \nu^3_0 \varepsilon_{ism} k_m d_{v\theta} ,$$ $$d_{vv}= k^{4-d-2\varepsilon}, \quad d_{\theta\theta} =k^{4-d-2a\varepsilon},\quad d_{v\theta}= k^{3-d-(1+ a)\varepsilon}.$$ In momentum-frequency representation (from frequency the correlators $D_{is}$ do not depend) $P_{is}$ is the transversal projector, $d$ means the dimension of space (completely antisymmetric pseudotensor $\varepsilon _ { ism } $ is determined only at $d=3$). The factors $g_0$ in correlators are played a role of charges; the positive constant $a$ is an arbitrary parameter. The parameter $\varepsilon$ serves for construction of\ decomposition of correlation functions, as the physical value of $\varepsilon$ the $\varepsilon_p = 2$ is considered that simulates energy pumping from the large-scale movements of fluid. The model (\[1\] -\[3\]) was investigated in [@1] in the first order on $\varepsilon$ with the help of recursive renormalization group method. It is shown that in the system two different regimes of developed turbulence can be realized: “kinetic” and “magnetic” that are connected with existence of two infrared (IR) stable fixed points of RG transformation. In [@1] the critical asymptotics of spectral density of energy $<{\bf\varphi}({\bf k}) {\bf\varphi}({\bf -k})>$ and $<{\bf\theta}({\bf k}) {\bf\theta}({\bf -k})>$ were determined. In [@2] more general formulation of the problem with inclusion of the cross correlator of random force $D^ { \varphi\theta } $ was considered by the quantum field RG method. The renormalization constants $Z_{i}$ were calculated there in the first order of $ g_{\alpha} $: $$Z _1 = 1-\frac{g_1d(d-1)}{4B\varepsilon}- \frac{g_2(d^2+d-4)}{4Ba\lambda^2\varepsilon}, \quad Z _3 = 1+\frac {g_1}{B\lambda\varepsilon}-\frac {g_2}{Ba\lambda^2\varepsilon},$$ $$Z _2 = 1-\frac{g_1(d +2)(d-1)}{2B\lambda(\lambda+1)\varepsilon}- \frac{g_2(d+2)(d-3)}{2Ba\lambda^2(\lambda+1)\varepsilon};$$ here $B=d(d+2){(4 \pi)}^{d/2}\Gamma (d/2)$ and $\Gamma (x)$ is the gamma-function. Renormalizability of the model was proven, and situation in the charging space of IR-stable fixed points corresponding to kinetic and magnetic regimes were found: $$\label{4} {g'_1}_{*}\equiv\frac{{g _1}_{*}}{B\lambda_{*}} =\frac{\varepsilon (1+\lambda_{*})}{15},\quad {g _2}_{*} =0,\quad \lambda_{*}= \frac{\sqrt{43/3}-1}{2},$$ ( the region of stability of the point is $a < 1.16.$ ) $$\label{5} {g _1}_{*} =\lambda_{*}=0 ,\quad {g'_2}_{*}\equiv\frac{{g _2}_{*}}{B{\lambda_{*}}^2} =a\varepsilon$$ ( this one is stable at $a\geq 0.25$). In paper [@3] features of a scaling behaviour in the model were investigated, and critical dimensions of fields and parameters of the theory are found out in the both critical regimes. We shall be interested in dependence of characteristics of developed turbulence (the correlation functions) in the inertial range from conditions of arising of large-scale curls. In the theory of developed turbulence it is supposed, that the energy pumping into the inertial range is executed by the vortices of a large size $\Lambda$. We shall take into account this scale having supplied the model (\[1\] -\[3\] ) by an infrared mass parameter $m\equiv 1 / \Lambda$ to consider the relative corrections to developed turbulent spectra with the help of SDE method. According to this method following the operator decomposition is fair: $$\label{6} { \bf \phi } ( { \bf x _1 }, t ){ \bf \phi } ( { \bf x _2 }, t) \simeq \sum _ {i } c _i ( { \bf r } ) F _i ( { \bf x }, t);$$ where $x\equiv ( x _1 + x _2 )/ 2$, $r\equiv x_1 -x _2$, $F _i$ are various local averages (composite operators). The averaging of (\[6\]) yields the asymptotics for pair correlation functions at $mr\to 0$ $$\label{7} < { \bf \phi } ( { \bf x _1 }, t ){ \bf \phi } ( { \bf x _2 }, t) >\simeq\sum _ { i } c _i ( { \bf r } ) a _i m^ { \Delta _ { F _i }};$$ here $\Delta _ { F _i } $ are the critical dimensions of the composite operators; $a_i$ means some constants. On decomposition (\[7\]) we conclude, that from the point of view of an opportunity of transition to a massless theory ( $m\to 0$ ) in the inertial range the operators with a negative critical dimension are dangerous. Critical dimensions of the composite operators in the model of magnetic hydrodynamics ===================================================================================== The renormalized operators are defined by the formula $F_i=Z_ { ik } F^R_k ( \Phi_R ).$ On the known matrix $Z_ { ik } $ a matrix of anomalous dimensions $\gamma_ { ik } = ( Z^ { -1 } ) _ { ij } { \cal D } _MZ_ { jk }$ is calculated and then the matrix of critical dimensions $\Delta_ { ik } = ( d^k_F ) _ { ik } + \Delta_\omega ( d^\omega_F ) _ { ik } + \gamma_ { ik }$ where $\Delta _\omega$ designates critical dimension of frequency, and $d^k_F$, $d^\omega_F$ are the momentum and frequency canonical dimensions of $F$; they are being determined from requirement of frequency and momentum dimensionless of terms of the action functional. The particular critical dimensions are eigenvalues of the matrix $\Delta_ { ik }$. They correspond to linear combinations of composite operators $L_i (F^R ) =U_ { ik } F^R_k$ which diagonalize the matrix $\Delta_ { ik }$. We shall consider the dimensions of elementary composite operators of the MHD model: $\phi_i\phi_k$, $\phi'\phi$, as well as vector operators $ ( \partial \phi\phi ) _i$ with various transpositions of badges. The tensor $\phi_i\phi_k$ is a sum of two independent tensors $$\label{8} \frac 1d \phi^2 \delta_{ik}, \quad\quad\quad \phi_i\phi_k-\frac 1d\phi^2\delta_{ik}.$$ Convolution of the first operator on badges yields family of scalar operators $\phi\phi$ $$F _1=\frac 12 v _iv _i;\quad F _2=\frac 12 \theta _i\theta _i;\quad F _3= \theta _iv _i.$$ Trace of the second expression in (\[8\]) is equal to zero. Essential property of the theory (\[1\] - \[3\]) is Galileian invariancy; it was used in [@4] for investigation of the composite operators in a problem of usual stochastic hydrodynamics. A non-stationary Galileian transformations of the fields $$\varphi_a ({\bf x},t)= \varphi ({\bf x +u}(t),t)-{\bf a}(t);\quad \varphi' _a ({\bf x},t)= \varphi' ({\bf x +u}(t),t);$$ $$\theta _a ({\bf x},t)= \theta ({\bf x +u}(t),t); \quad \theta' _a ({\bf x},t)= \theta' ({\bf x +u}(t),t);$$ (a parameter of transformation - $ { \bf v } (t) $ is the vector function dependent only from a time well decreasing at $ | t | \to \infty$ and $ { \bf u } (t) = \int _ { -\infty } ^ { t } dt { \bf v } ( t' ) $ ) realized in the Ward identity leads [@5] to $$\label{9} \int dx\{a _{0\alpha}[\frac {\partial F _\alpha}{\partial \varphi _s}-{ \frac {\partial F _\alpha}{\partial (\partial _t \varphi _k)}} \partial _s \varphi _k] + \partial _t[a _{0\alpha} \frac {\partial F _\alpha}{\partial (\partial _s \varphi _k)}]\} <\infty.$$ Here $a _ { 0\alpha } $ are the nonrenormalized functions of sources of family of operators $F _\alpha$, and symbol ”$ < \infty$ ” means that the considered functional is finite. We shall assign $G_1= v_iv_k / 2$ and $G_2=\theta_i\theta_k / 2$. The operator $G_3=v_i\theta_k$ is pseudotensor so it doesn’t mix with the first two operators due to renormalization. $G_1$ is noninvariant of the Galileian transformations so the formula (\[9\]) permits to prove that $G _1 $ is finite ($Z_ { 11 } =1 $), and that it doesn’t mix to the operator $G _2$ in proceeding of renormalization, hence $\quad Z_ { 21 } =0$ (one can say, that the operator $G _1$ aren’t being renormalized and doesn’t mix to the Galileian invariant operator $G _2$ since $G _1$ is not Galileian invariant). Besides, from (\[9\]) the absence of divergences in diagrams with the composite operator $G_3$ (it is noninvariant of Galileian transformations also) is followed. Then from the definition of renormalized composite operators it is easy to get the expression: $Z_ { 33 } =Z_ { v_i\theta_k } =Z_\theta^ { -1 }.$ Thus, the matrix of renormalization constants ${\bf Z}_{\alpha\delta}$ of the operators $G_i$ is $${\bf Z} _{\alpha\beta}= \left(\begin{array}{ccc} 1&Z _{12}&0\\ 0&Z _{22}&0\\ 0&0&Z _\theta ^{-1} \end{array}\right).$$ Unknown elements of the matrix one can calculate with the help of standard diagrams with the following propagators $$\begin{aligned} \label{10} <\varphi _i({\bf k},t)\varphi' _j({\bf-k},0)> _0= P _{ij}e^{-\nu k^2 t}\theta (t); \, \nonumber\\ <\varphi _i ({\bf k},t)\varphi _j({\bf-k},0)> _0= \frac12 P _{ij}g \nu^2k^{2-d-2\varepsilon} M^{2\varepsilon}e^{-\nu k^2 |t|}; \, \nonumber\\ <\theta _i({\bf k},t)\theta' _j({\bf-k},0)> _0= P _{ij}\theta (t)e ^{-\nu \lambda k^2 t}; \, \nonumber\\ <\theta _i({\bf k},t)\theta _j({\bf-k},0)> _0= \frac12 P_{ij}\frac{g'\nu^2k^{2-d-2a\varepsilon}}{\lambda} M^{2a\varepsilon}e^{-\nu\lambda k^2|t|};\end{aligned}$$ $$<\theta _i({\bf k},t)\varphi _j({\bf-k},0)> _0= \varepsilon _{isj}\frac{g'' \nu^2k^{1-d-(1+a)\varepsilon}}{(1+\lambda)} M^{(1+a) \varepsilon}k _s [e^{-\lambda\nu k^2t}\theta (t)+e^{\nu k^2 t}\theta (-t)];$$ Account of the renormalization constants for the scalar operator we execute having curtailed the badges $\delta _ { ik }.$ The calculations in one-loop approximation under diagrams specified on a fig. 1 of the appropriate constants $Z _ { \alpha\delta } $ in the theory (\[2\]) yelds a renormalized action for the generating functional of correlation functions with the composite operators $F_\alpha$: $\bar S _R= S _R + a_{0\alpha}Z_{\alpha\delta}F_\delta^R (Z_\phi\phi)$; here the renormalization constants are $$\begin{aligned} \label{11} {Z^{\varphi\theta}} _{12} = -\frac{\lambda(d +2)(d-1)}{2(\lambda+1)}(\frac {g'_1}{\varepsilon}-\frac{g' _2}{a\varepsilon}); \quad {Z^{\theta^2}} _{22}Z _\theta ^2 = 1+ \frac {(d-1)(d+2)}{2(\lambda+1)}(\frac {g'_1}{\varepsilon}-\frac{g' _2}{a\varepsilon});\, \nonumber \\ {Z^{\varphi_{i}\theta_{j}}}_ {12} = -\frac \lambda{2(\lambda+1)}(\frac {g'_1}{\varepsilon}- \frac{g' _2}{a\varepsilon}), \quad {Z^{\theta_{i}\theta_{j}}}_{22} Z _\theta ^2 = 1+\frac 1{2(\lambda+1)}(\frac {g'_1}{\varepsilon}-\frac{g' _2}{a\varepsilon}).\end{aligned}$$ As far as in the fixed points of RG transformation some values of the charges approach to zero in avoidance of trivialization of asymptotics in all orders of the $\varepsilon$-decomposition it is useful to redefine the fields as follows: $$\label{12} \theta \rightarrow\sqrt{g_2\nu ^3}M ^{a\varepsilon}\theta; \quad\quad\quad v\rightarrow\sqrt{g_1\nu^3} M^\varepsilon v;$$ $$\theta' \rightarrow\theta'/\sqrt{g_2\nu ^3}M ^{a\varepsilon}; \quad\quad\quad\varphi'\rightarrow\varphi'/\sqrt{g_1\nu^3} M^\varepsilon.$$ The transformations don’t change positions of the fixed points and values of the renormalization constants (the canonical dimensions of the operators $F _\alpha$ vary only). The critical dimensions of the operators $G_i$ calculated on the constants $Z _ { \alpha\alpha } $ in the first order of $\varepsilon$ and at any $d$ are listed in table I. It should be noticed that the value of $\Delta _1$ is exact. The constant $Z_{12}$ defines an admixture to $G _1$ of the operator $G _2$. Considering (\[2\]) in a condition of the kinetic regime with the fields of (\[12\]) we have $G_2\to g_2 G_2$ and taking into account that in the kinetic fixed point (\[4\]) $g _2^ { * } =0$ it’s easy to show that just the operators $G _1,\quad G _2$ (instead of their mixture) have the dimension $\Delta _i$ in this point. The constant $Z_{12}\sim O(\lambda)$ which is responsible for mixing of the operators disappears in the magnetic point. We shall consider elements of a matrix $Z_{ik}$ correspond to renormalization of the set of scalar and vector composite operators $\phi'\phi$ and $ ( \partial \phi\phi ) _i$. At this set there are the operators reducing to a total differential $\partial ( \phi\phi )$ with various transpositions of badges. Renormalization of them is equivalent to renormalization of the operators ${\bf\phi}{\bf\phi}$ have considered above. The critical dimensions is being appropriated to these operators surpass dimensions located in table I per unit of. The operators of a $\phi'\phi$-type don’t mix to any other operators and aren’t being renormalized because of 1-irreducible diagrams that is responsible for mixing of these operators with the other are equal to zero as far as they contain cycles of advancing lines. A structure of interactions in the (\[2\]) provides removal from each diagram a one derivative on each external line of a $\phi'$-type that effectively lowers an index of divergence of the diagrams. Therefore, in the minimal subtraction (MS) scheme of renormalization that is appropriate to these operators the diagonal elements are $Z_ { \alpha\alpha } =1$, and all nondiagonal ones are equal to zero. Thus, $\Delta_{\phi'\phi}=3.$ The remaining vector operators $F_1=v_iv^2$ and $F_2=v\theta ^2$ are true tensors, and $F_3=\theta_i\theta^2$, $\quad F_4=\theta v^2$ are pseudotensors; these pairs of the operators are being renormalized independently from each other. The operators $F_{1-2}$ are noninvariant to Galileian transformations, thereof, it is easy to approve [@5] the finiteness of the operator $v^3$ as well as that it doesn’t mix to $v{\theta^2}$ due to renormalization. Thus, $Z_ { 11 } =1,\quad Z_ { 21 } =0.$ Similarly, the Galileian invariant operator $F_3$ can’t mix to the operator $F_4$ that means $Z_{34}=0.$ For definition of remaining elements of the matrix $Z_ { ik } $ in a one-loop\ approximation it is necessary to consider the diagrams that is shown on a fig. 2. It gives $$Z_{12}=-3 \frac{\lambda (d+2)(d-1)} {(\lambda+1)}(\frac{g'_1}{\varepsilon}-\frac{g'_2}{a\varepsilon}) ,\quad Z_{33}Z^3_\theta=1+ 3 \frac{(d+2)(d-1)} {(\lambda+1)}(\frac{g'_1}{\varepsilon}-\frac{g'_2}{a\varepsilon}),$$ $$\label{13} Z_{22}Z^2_\theta= 1+\left( \frac{(d-1)(d+2)}{(\lambda+1)}-1\right) (\frac{g'_1}{\varepsilon}-\frac{g'_2}{a\varepsilon}), \label{16}$$ $$\quad Z_{43}Z_\theta=- \frac{\lambda (d+2)(d-1)} {(\lambda+1)}(\frac{g'_1}{\varepsilon}-\frac{g'_2}{a\varepsilon}),\quad Z_{44}Z_\theta=1- \frac{g'_1}{\varepsilon}+\frac{g'_2}{a\varepsilon},$$ The matrix of renormalization constants of this family of the operators has a block triangular form, so the critical dimensions are determined by the diagonal elements $Z_{\alpha\alpha}.$ The values of critical dimensions calculated on (\[13\]) are listed in table 2. The nondiagonal elements $Z_ { 12 } $ and $Z_ { 43 } $ define an admixture of the operators $F_1$ and $F_4$ to $F_2$ and $F_3.$ Taking into account that in the kinetic mode $F_1= { g_1 } ^ { 3 / 2 } v^3$, $F_2=g_1^ { 1 / 2 } g_2 v\theta^2$ $F_3= { g_2 } ^ { 3 / 2 } \theta ^3$, and $F_4= g_1 { g_2 } ^ { 1 / 2 } \theta v^2$ and that $ {g_2 } _ { * } =0$ in the fixed point (\[5\]) it is possible to assert the certain critical dimensions are belonged to the operators, instead of their linear combinations. In the magnetic point (\[6\]) $Z_ { 12 } =Z_ { 34 } =0.$ Discussion of the results ========================== It is important to note that any of the operators considered can’t participate as amendments for phenomenological equations of MHD. The operators are being possessed of the essential critical dimensions don’t contain auxiliary fields $\phi'$, but the operators of a $\phi'\phi$-type (a function of response) are inessential and don’t satisfy the requirements of Galileian invariancy also. The operators of canonical dimension $d=3$ define new nonanalytic corrections to the spectra of developed turbulence that was found in [@3]. According to SDE method such corrections for a pair correlation function $ < \phi_1\phi_2>$ can be represented as follows: $$<\phi_1({\bf k},t)\phi_2({\bf -k},t)>= Ak^{-d -\Delta _{\phi_1}-\Delta _{\phi_2}}\left(1+\sum_i b_i \left(\frac {m}{k}\right)^{ \Delta_{F_i}}\right); \label{14}$$ (here $\Delta_\phi$ are the critical dimensions of the fields). As far as in the inertial range the following estimation is correct $ m / k \ll 1$, the formula (\[14\]) results to nonanalyticities in a case of $ \Delta_ { F_i } < 0.$ At the real value of $\varepsilon =2$ the critical dimensions of $G_2 \quad ( a>1 / 2 $, $F_1,\quad F_2 \quad ( a>0.243 ), \quad F_3,\quad F_4 \quad ( a>0.82 ) $ in the kinetic point become negative. The dimension of $F _4$ at $a>3 / 4$ is negative in the magnetic point also. Thanking to Galileian invariancy of the theory we can refer to the results of [@4] where the terms of the sum of (\[14\]) for static correlation functions connected with the Galileian noninvariant composite operators were proven to yield not a contribution. Those in our case are $F_1,\quad F_2$, and $F_4.$ They can participate in the decomposition (\[16\]) only for dynamic correlation functions. The Galileian invariant operator $F_3$ gives the contribution in the sum in all the cases. Comparing of the dimensions of the composite operators of families $\phi\phi$ and $( \partial \phi\phi ) _i$ at the real value $\varepsilon=2$ one can see that the set $ ( \partial \phi\phi ) _i$ appears more essential in the scaling range (so in the kinetic regime $\Delta_ { \theta^2 } > \Delta_ { v\theta^2 },\Delta_ { \theta^3 },\quad \Delta_ { v^2 } >\Delta_ { v^3 } $ ). One can assume that the tendency of growth of essentiality of operators is being demonstrated by the elementary operators in the model of magnetic hydrodynamics will be saved for more complex operators. The results received by us testify that in the inertial range in MHD the correlation functions depend on the external integral turbulent scale $ m.$ Thus, the behaviour in the model is not universal. This important result can be checked experimentally besides the particular values of critical dimensions of the composite operators calculated here can be measured on an experiment too that will be served certainly to check of the offered theory. Acknowledgments ================ The author is grateful to M. Yu. Nalimov and L. Ts. Adzhemyan for useful discussion at preparation of the paper as well as to mayorate of St.-Petersburg (Russia ) for rendered support. [99]{} Fournier J.-P., Frish U. [*Phys.Rev.*]{}, 1983. Vol. A28,$N^0$ 2. P.1000. L. Ts. Adzhemyan, A.N. Vasil’ev, M. Gnatich, [*Theor. and math. physics*]{}, 1985, Vol. 64, $N^0$ 2, p. 196. L. Ts. Adzhemyan, D.Yu. Wolchenkov, M. Yu. Nalimov, [*Theor. and math. physics*]{}, Vol. 107, $N^0 1$, pp. 142-154. L. Ts. Adzhemyan, N.V. Antonov, A.N. Vasil’ev, [*Journal of experimental and theoretical physics*]{}, 1989, Vol. [ 95]{}, $N^{0} 4$, pp. 1272, (on Russian). L. Ts. Adzhemyan, A.N. Vasil’ev, M. Gnatich, [*Theor. and math. physics*]{}, 1988, Vol. [ 74]{}, N2, p. 180.       [||c|c|c||]{}\ & &\ *$(G_1)_{ij}$&$2-4/3\varepsilon$&$2$\ *$(G_2)_{ij}$&$2-2(a-3/10)\varepsilon$&$2+3a\varepsilon$\ *$(G_3)_{ij}$ & $2-(a+1/3)\varepsilon$ & $2+a\varepsilon$\ *$(G_1)_{ij}\delta_{ij} $ & $2-4/3\varepsilon$& $2$\ *$(G_2)_{ij}\delta_{ij} $ & $2-2a\varepsilon$& $2+12a\varepsilon$\ *$(G_3)_{ij}\delta_{ij} $ & $2-(a+1/3)\varepsilon$& $2+a\varepsilon$\ ******             [||c|c|c||]{}\ & &\ *$F_1$&$3-2\varepsilon$&$3$\ *$F_2$&$3-2(a+0.507)\varepsilon$&$3+18a\varepsilon$\ *$F_3$ & $3-3(a+1)\varepsilon$ & $3+60a\varepsilon$\ *$F_4$ & $3-(a+0.68)\varepsilon$& $3-2a\varepsilon$\ ****       =1.00mm (154.00,37.00) (154.00,15.00)[(0,0)\[cc\][$,$]{}]{} (19.00,19.00) ------------------------------------------------------------------------ (55.00,19.00) ------------------------------------------------------------------------ (82.00,19.00) ------------------------------------------------------------------------ (116.00,19.00) ------------------------------------------------------------------------ (20.00,24.00)[(0,0)\[cc\][$\theta$]{}]{} (25.00,29.00)[(0,0)\[cc\][$\theta'$]{}]{} (32.00,35.00)[(0,0)\[cc\][$\theta$]{}]{} (33.00,26.00)[(0,0)\[cc\][$\varphi$]{}]{} (33.00,14.00)[(0,0)\[cc\][$\varphi$]{}]{} (20.00,14.00)[(0,0)\[cc\][$\theta$]{}]{} (25.00,9.00)[(0,0)\[cc\][$\theta'$]{}]{} (32.00,3.00)[(0,0)\[cc\][$\theta$]{}]{} (44.00,19.00)[(0,0)\[cc\][$,$]{}]{} (56.00,24.00)[(0,0)\[cc\][$\theta$]{}]{} (62.00,29.00)[(0,0)\[cc\][$\theta$]{}]{} (68.00,35.00)[(0,0)\[cc\][$\theta$]{}]{} (69.00,26.00)[(0,0)\[cc\][$\varphi'$]{}]{} (69.00,14.00)[(0,0)\[cc\][$\varphi$]{}]{} (56.00,14.00)[(0,0)\[cc\][$\theta$]{}]{} (62.00,9.00)[(0,0)\[cc\][$\theta'$]{}]{} (68.00,3.00)[(0,0)\[cc\][$\theta$]{}]{} (83.00,24.00)[(0,0)\[cc\][$\varphi$]{}]{} (89.00,29.00)[(0,0)\[cc\][$\varphi'$]{}]{} (94.00,35.00)[(0,0)\[cc\][$\theta$]{}]{} (96.00,26.00)[(0,0)\[cc\][$\theta$]{}]{} (96.00,14.00)[(0,0)\[cc\][$\theta$]{}]{} (83.00,14.00)[(0,0)\[cc\][$\varphi$]{}]{} (89.00,9.00)[(0,0)\[cc\][$\varphi'$]{}]{} (94.00,3.00)[(0,0)\[cc\][$\theta$]{}]{} (75.00,19.00)[(0,0)\[cc\][$,$]{}]{} (105.00,19.00)[(0,0)\[cc\][$,$]{}]{} (118.00,24.00)[(0,0)\[cc\][$\varphi$]{}]{} (124.00,29.00)[(0,0)\[cc\][$\varphi$]{}]{} (129.00,35.00)[(0,0)\[cc\][$\theta$]{}]{} (130.00,27.00)[(0,0)\[cc\][$\theta'$]{}]{} (130.00,14.00)[(0,0)\[cc\][$\theta$]{}]{} (118.00,14.00)[(0,0)\[cc\][$\varphi$]{}]{} (124.00,9.00)[(0,0)\[cc\][$\varphi'$]{}]{} (129.00,3.00)[(0,0)\[cc\][$\theta$]{}]{} [**fig. 1**]{} The set of one-loop 1-irreducible diagrams are responsible for the nontrival renormalization constances of the composite operators of a $ \phi_{i}\phi_{j}$ type. =0.50mm (288.00,298.00) (22.00,279.00)[(0,0)\[cc\][$Z _{12}$:]{}]{} (55.00,278.00) ------------------------------------------------------------------------ (110.00,278.00) ------------------------------------------------------------------------ (172.00,279.00)[(0,0)\[cc\][$Z _{33}$:]{}]{} (205.00,278.00) ------------------------------------------------------------------------ (260.00,278.00) ------------------------------------------------------------------------ (102.00,230.00)[(0,0)\[cc\][$Z _{43}$:]{}]{} (135.00,229.00) ------------------------------------------------------------------------ (190.00,229.00) ------------------------------------------------------------------------ (44.00,272.00)[(0,0)\[cc\][$v$]{}]{} (56.00,284.00)[(0,0)\[cc\][$v$]{}]{} (63.00,292.00)[(0,0)\[cc\][$v'$]{}]{} (78.00,293.00)[(0,0)\[cc\][$\theta$]{}]{} (74.00,287.00)[(0,0)\[cc\][$\theta$]{}]{} (76.00,273.00)[(0,0)\[cc\][$\theta'$]{}]{} (74.00,265.00)[(0,0)\[cc\][$\theta$]{}]{} (62.00,264.00)[(0,0)\[cc\][$v$]{}]{} (54.00,271.00)[(0,0)\[cc\][$v$]{}]{} (97.00,272.00)[(0,0)\[cc\][$v$]{}]{} (108.00,271.00)[(0,0)\[cc\][$v$]{}]{} (117.00,265.00)[(0,0)\[cc\][$v'$]{}]{} (130.00,265.00)[(0,0)\[cc\][$\theta$]{}]{} (129.00,273.00)[(0,0)\[cc\][$\theta$]{}]{} (129.00,286.00)[(0,0)\[cc\][$\theta$]{}]{} (133.00,294.00)[(0,0)\[cc\][$\theta$]{}]{} (112.00,285.00)[(0,0)\[cc\][$v$]{}]{} (120.00,294.00)[(0,0)\[cc\][$v'$]{}]{} (196.00,273.00)[(0,0)\[cc\][$\theta$]{}]{} (206.00,285.00)[(0,0)\[cc\][$\theta$]{}]{} (207.00,272.00)[(0,0)\[cc\][$\theta$]{}]{} (212.00,265.00)[(0,0)\[cc\][$\theta$]{}]{} (213.00,292.00)[(0,0)\[cc\][$\theta'$]{}]{} (229.00,294.00)[(0,0)\[cc\][$\theta$]{}]{} (224.00,285.00)[(0,0)\[cc\][$v$]{}]{} (225.00,273.00)[(0,0)\[cc\][$v'$]{}]{} (224.00,265.00)[(0,0)\[cc\][$\theta$]{}]{} (251.00,272.00)[(0,0)\[cc\][$\theta$]{}]{} (262.00,284.00)[(0,0)\[cc\][$\theta$]{}]{} (260.00,272.00)[(0,0)\[cc\][$\theta$]{}]{} (268.00,292.00)[(0,0)\[cc\][$\theta'$]{}]{} (267.00,266.00)[(0,0)\[cc\][$\theta'$]{}]{} (281.00,272.00)[(0,0)\[cc\][$v$]{}]{} (281.00,285.00)[(0,0)\[cc\][$v$]{}]{} (283.00,295.00)[(0,0)\[cc\][$\theta$]{}]{} (280.00,264.00)[(0,0)\[cc\][$\theta$]{}]{} (211.00,237.00)[(0,0)\[cc\][$\theta$]{}]{} (210.00,223.00)[(0,0)\[cc\][$\theta$]{}]{} (214.00,246.00)[(0,0)\[cc\][$\theta$]{}]{} (210.00,215.00)[(0,0)\[cc\][$\theta$]{}]{} (179.00,221.00)[(0,0)\[cc\][$\theta$]{}]{} (191.00,221.00)[(0,0)\[cc\][$v$]{}]{} (192.00,236.00)[(0,0)\[cc\][$v$]{}]{} (201.00,245.00)[(0,0)\[cc\][$v'$]{}]{} (200.00,216.00)[(0,0)\[cc\][$v'$]{}]{} (156.00,216.00)[(0,0)\[cc\][$\theta$]{}]{} (157.00,244.00)[(0,0)\[cc\][$\theta$]{}]{} (156.00,237.00)[(0,0)\[cc\][$\theta$]{}]{} (154.00,225.00)[(0,0)\[cc\][$\theta'$]{}]{} (125.00,223.00)[(0,0)\[cc\][$\theta$]{}]{} (135.00,222.00)[(0,0)\[cc\][$v$]{}]{} (137.00,236.00)[(0,0)\[cc\][$v$]{}]{} (144.00,243.00)[(0,0)\[cc\][$v'$]{}]{} (140.00,216.00)[(0,0)\[cc\][$v$]{}]{} (10.00,182.00)[(0,0)\[cc\][$Z _{22}$:]{}]{} (43.00,181.00) ------------------------------------------------------------------------ (98.00,181.00) ------------------------------------------------------------------------ (116.00,190.00)[(0,0)\[cc\][$\theta$]{}]{} (116.00,173.00)[(0,0)\[cc\][$\theta$]{}]{} (118.00,197.00)[(0,0)\[cc\][$v$]{}]{} (118.00,167.00)[(0,0)\[cc\][$\theta$]{}]{} (87.00,178.00)[(0,0)\[cc\][$\theta$]{}]{} (100.00,177.00)[(0,0)\[cc\][$v$]{}]{} (100.00,188.00)[(0,0)\[cc\][$\theta$]{}]{} (109.00,197.00)[(0,0)\[cc\][$\theta'$]{}]{} (106.00,167.00)[(0,0)\[cc\][$v'$]{}]{} (64.00,168.00)[(0,0)\[cc\][$\theta$]{}]{} (64.00,197.00)[(0,0)\[cc\][$v$]{}]{} (60.00,190.00)[(0,0)\[cc\][$v$]{}]{} (61.00,177.00)[(0,0)\[cc\][$v'$]{}]{} (31.00,174.00)[(0,0)\[cc\][$\theta$]{}]{} (45.00,176.00)[(0,0)\[cc\][$\theta$]{}]{} (45.00,188.00)[(0,0)\[cc\][$v$]{}]{} (52.00,195.00)[(0,0)\[cc\][$v'$]{}]{} (50.00,168.00)[(0,0)\[cc\][$\theta$]{}]{} (168.00,181.00) ------------------------------------------------------------------------ (186.00,190.00)[(0,0)\[cc\][$v$]{}]{} (186.00,173.00)[(0,0)\[cc\][$v$]{}]{} (188.00,197.00)[(0,0)\[cc\][$\theta$]{}]{} (188.00,167.00)[(0,0)\[cc\][$\theta$]{}]{} (157.00,178.00)[(0,0)\[cc\][$v$]{}]{} (169.00,175.00)[(0,0)\[cc\][$\theta$]{}]{} (170.00,188.00)[(0,0)\[cc\][$\theta$]{}]{} (179.00,197.00)[(0,0)\[cc\][$\theta'$]{}]{} (176.00,166.00)[(0,0)\[cc\][$\theta'$]{}]{} (237.00,181.00) ------------------------------------------------------------------------ (255.00,190.00)[(0,0)\[cc\][$v$]{}]{} (256.00,175.00)[(0,0)\[cc\][$v$]{}]{} (257.00,197.00)[(0,0)\[cc\][$v$]{}]{} (257.00,167.00)[(0,0)\[cc\][$\theta$]{}]{} (226.00,178.00)[(0,0)\[rc\][$\theta$]{}]{} (236.00,175.00)[(0,0)\[cc\][$\theta$]{}]{} (239.00,188.00)[(0,0)\[cc\][$v$]{}]{} (248.00,197.00)[(0,0)\[cc\][$v'$]{}]{} (243.00,165.00)[(0,0)\[cc\][$\theta'$]{}]{} (43.00,127.00) ------------------------------------------------------------------------ (64.00,114.00)[(0,0)\[rc\][$\theta$]{}]{} (66.00,142.00)[(0,0)\[cc\][$v$]{}]{} (61.00,134.00)[(0,0)\[cc\][$\theta$]{}]{} (62.00,122.00)[(0,0)\[cc\][$\theta'$]{}]{} (33.00,122.00)[(0,0)\[cc\][$\theta$]{}]{} (45.00,122.00)[(0,0)\[lc\][$v$]{}]{} (45.00,134.00)[(0,0)\[cc\][$\theta$]{}]{} (52.00,141.00)[(0,0)\[cc\][$\theta'$]{}]{} (49.00,114.00)[(0,0)\[cb\][$v$]{}]{} (96.00,127.00) ------------------------------------------------------------------------ (117.00,114.00)[(0,0)\[cc\][$v$]{}]{} (120.00,143.00)[(0,0)\[cc\][$\theta$]{}]{} (113.00,136.00)[(0,0)\[cc\][$\theta$]{}]{} (115.00,121.00)[(0,0)\[cc\][$\theta'$]{}]{} (87.00,121.00)[(0,0)\[cc\][$\theta$]{}]{} (98.00,122.00)[(0,0)\[cc\][$\theta$]{}]{} (98.00,134.00)[(0,0)\[cc\][$v$]{}]{} (105.00,141.00)[(0,0)\[cc\][$v'$]{}]{} (103.00,113.00)[(0,0)\[cc\][$\theta$]{}]{} (166.00,127.00) ------------------------------------------------------------------------ (187.00,114.00)[(0,0)\[cc\][$v$]{}]{} (190.00,143.00)[(0,0)\[cc\][$\theta$]{}]{} (183.00,136.00)[(0,0)\[cc\][$v$]{}]{} (185.00,121.00)[(0,0)\[cc\][$v'$]{}]{} (157.00,121.00)[(0,0)\[cc\][$\theta$]{}]{} (168.00,122.00)[(0,0)\[cc\][$v$]{}]{} (168.00,134.00)[(0,0)\[cc\][$\theta$]{}]{} (175.00,141.00)[(0,0)\[cc\][$\theta'$]{}]{} (173.00,114.00)[(0,0)\[cc\][$v$]{}]{} (233.00,127.00) ------------------------------------------------------------------------ (254.00,114.00)[(0,0)\[cc\][$\theta$]{}]{} (257.00,143.00)[(0,0)\[cc\][$\theta$]{}]{} (253.00,136.00)[(0,0)\[cc\][$v$]{}]{} (252.00,123.00)[(0,0)\[cc\][$v'$]{}]{} (223.00,122.00)[(0,0)\[cc\][$v$]{}]{} (235.00,122.00)[(0,0)\[cc\][$\theta$]{}]{} (235.00,134.00)[(0,0)\[cc\][$\theta$]{}]{} (242.00,141.00)[(0,0)\[cc\][$\theta'$]{}]{} (240.00,114.00)[(0,0)\[cc\][$\theta$]{}]{} (10.00,77.00)[(0,0)\[cc\][$Z _{44}$:]{}]{} (43.00,76.00) ------------------------------------------------------------------------ (98.00,76.00) ------------------------------------------------------------------------ (118.00,85.00)[(0,0)\[cc\][$v$]{}]{} (118.00,69.00)[(0,0)\[cc\][$v$]{}]{} (122.00,93.00)[(0,0)\[cc\][$v$]{}]{} (118.00,62.00)[(0,0)\[cc\][$v$]{}]{} (87.00,71.00)[(0,0)\[cc\][$\theta$]{}]{} (98.00,70.00)[(0,0)\[cc\][$v$]{}]{} (100.00,83.00)[(0,0)\[cc\][$v$]{}]{} (109.00,92.00)[(0,0)\[cc\][$v'$]{}]{} (107.00,60.00)[(0,0)\[cc\][$v'$]{}]{} (64.00,63.00)[(0,0)\[cc\][$v$]{}]{} (67.00,93.00)[(0,0)\[cc\][$v$]{}]{} (62.00,85.00)[(0,0)\[rb\][$v$]{}]{} (62.00,72.00)[(0,0)\[cc\][$v'$]{}]{} (33.00,70.00)[(0,0)\[cc\][$\theta$]{}]{} (43.00,70.00)[(0,0)\[cc\][$v$]{}]{} (45.00,83.00)[(0,0)\[cc\][$v$]{}]{} (52.00,90.00)[(0,0)\[cc\][$v'$]{}]{} (50.00,63.00)[(0,0)\[cc\][$v$]{}]{} (168.00,76.00) ------------------------------------------------------------------------ (189.00,85.00)[(0,0)\[cc\][$v$]{}]{} (189.00,71.00)[(0,0)\[cc\][$v$]{}]{} (192.00,93.00)[(0,0)\[cc\][$v$]{}]{} (188.00,62.00)[(0,0)\[cc\][$\theta$]{}]{} (157.00,71.00)[(0,0)\[cc\][$v$]{}]{} (169.00,70.00)[(0,0)\[cc\][$\theta$]{}]{} (170.00,83.00)[(0,0)\[cc\][$v$]{}]{} (179.00,92.00)[(0,0)\[cc\][$v'$]{}]{} (176.00,61.00)[(0,0)\[cc\][$\theta'$]{}]{} (237.00,76.00) ------------------------------------------------------------------------ (258.00,85.00)[(0,0)\[cc\][$\theta$]{}]{} (258.00,71.00)[(0,0)\[cc\][$\theta$]{}]{} (259.00,94.00)[(0,0)\[cc\][$v$]{}]{} (257.00,62.00)[(0,0)\[cc\][$\theta$]{}]{} (227.00,69.00)[(0,0)\[cc\][$v$]{}]{} (238.00,70.00)[(0,0)\[cc\][$\theta$]{}]{} (239.00,83.00)[(0,0)\[cc\][$v$]{}]{} (248.00,92.00)[(0,0)\[cc\][$v'$]{}]{} (245.00,63.00)[(0,0)\[cc\][$\theta'$]{}]{} (43.00,22.00) ------------------------------------------------------------------------ (64.00,9.00)[(0,0)\[cc\][$v$]{}]{} (70.00,39.00)[(0,0)\[cc\][$\theta$]{}]{} (62.00,31.00)[(0,0)\[cc\][$\theta$]{}]{} (62.00,18.00)[(0,0)\[cc\][$\theta'$]{}]{} (34.00,18.00)[(0,0)\[cc\][$v$]{}]{} (43.00,15.00)[(0,0)\[cc\][$\theta$]{}]{} (45.00,29.00)[(0,0)\[cc\][$v$]{}]{} (52.00,36.00)[(0,0)\[cc\][$v'$]{}]{} (48.00,9.00)[(0,0)\[cc\][$\theta$]{}]{} (96.00,22.00) ------------------------------------------------------------------------ (117.00,9.00)[(0,0)\[cc\][$v$]{}]{} (119.00,38.00)[(0,0)\[cc\][$\theta$]{}]{} (116.00,31.00)[(0,0)\[cc\][$v$]{}]{} (115.00,18.00)[(0,0)\[cc\][$v'$]{}]{} (87.00,18.00)[(0,0)\[cc\][$v$]{}]{} (96.00,15.00)[(0,0)\[cc\][$v$]{}]{} (98.00,29.00)[(0,0)\[cc\][$\theta$]{}]{} (104.00,38.00)[(0,0)\[cc\][$\theta'$]{}]{} (103.00,9.00)[(0,0)\[cc\][$v$]{}]{} (166.00,22.00) ------------------------------------------------------------------------ (187.00,9.00)[(0,0)\[ct\][$\theta$]{}]{} (190.00,39.00)[(0,0)\[cc\][$v$]{}]{} (185.00,31.00)[(0,0)\[lb\][$\theta$]{}]{} (185.00,18.00)[(0,0)\[cc\][$\theta'$]{}]{} (156.00,17.00)[(0,0)\[cc\][$v$]{}]{} (165.00,14.00)[(0,0)\[cc\][$v$]{}]{} (168.00,29.00)[(0,0)\[cc\][$\theta$]{}]{} (172.00,37.00)[(0,0)\[cc\][$\theta'$]{}]{} (173.00,8.00)[(0,0)\[cc\][$v$]{}]{} (233.00,22.00) ------------------------------------------------------------------------ (254.00,9.00)[(0,0)\[rc\][$\theta$]{}]{} (256.00,38.00)[(0,0)\[cc\][$v$]{}]{} (251.00,31.00)[(0,0)\[cc\][$v$]{}]{} (252.00,18.00)[(0,0)\[cc\][$v'$]{}]{} (222.00,17.00)[(0,0)\[cc\][$v$]{}]{} (233.00,15.00)[(0,0)\[cc\][$\theta$]{}]{} (235.00,29.00)[(0,0)\[cc\][$v$]{}]{} (242.00,38.00)[(0,0)\[cc\][$v'$]{}]{} (240.00,8.00)[(0,0)\[cc\][$\theta$]{}]{} (144.00,278.00)[(0,0)\[cc\][;]{}]{} (288.00,278.00)[(0,0)\[cc\][;]{}]{} (223.00,228.00)[(0,0)\[cc\][;]{}]{} (270.00,125.00)[(0,0)\[cc\][;]{}]{} (270.00,19.00)[(0,0)\[cc\][;]{}]{} \ \ [**fig. 2**]{} The set of one-loop 1-irreducible diagrams are responsible for the nontrival renormalization constances of the composite operators of a $ (\partial\phi\phi)_{i}$ type.

--- abstract: 'By mapping the strong interaction between Rydberg excitations in ultra-cold atomic ensembles onto single photons via electromagnetically induced transparency, it is now possible to realize a nonlinear optical medium which exhibits a strong optical nonlinearity at the level of individual photons. We review the theoretical concepts and the experimental state-of-the-art of this exciting new field, and discuss first applications in the field of all-optical quantum information processing.' address: - '$^1$Department of Physics of Complex Systems, Weizmann Institute of Science, Rehovot 76100, Israel' - '$^2$Joint Quantum Centre (JQC) Durham-Newcastle, Dept. of Physics, Durham University, Durham DH1 3LE, UK.' - '$^3$5. Physikalisches Institut, Universität Stuttgart, Pfaffenwaldring 57, 70569 Stuttgart, Germany' author: - 'O. Firstenberg$^1$' - 'C. S. Adams$^2$' - 'S. Hofferberth$^3$' bibliography: - 'biblio\_EITReview.bib' title: Nonlinear quantum optics mediated by Rydberg interactions --- Introduction ============ One remarkable success of advances in ultra-cold Rydberg physics is the realisation of a medium with a large optical nonlinearity at the single photon level [@Kuzmich2012b; @Vuletic2012; @Adams2013]. Highly-excited Rydberg atoms bring something new to the history of optics as they enable quantum nonlinear media where photons are strongly interacting! This is significant for a number of reasons. For example, previously, it was generally accepted that the prospects for nonlinear all-optical quantum computing were bleak due to the weakness of optical nonlinearities. Consequently the main focus turned towards linear optics quantum computing (LOQC), which exploits measurement to implement gates [@Milburn2001]. However, as this is a probabilistic protocol, scaling is a problem. But now Rydberg quantum optics brings the nonlinear approach back into the frame. A fundamental question remains, even if there is a sufficiently large nonlinearity, is this sufficient to build an optical quantum computer [@Shapiro2006]? This question can be partially addressed. As we show here, the Rydberg nonlinearity is not only large but different because of the nature of Rydberg blockade [@Lukin2001c]. As the interactions between highly-excited Rydberg atoms are long-range, unlike conventional nonlinear optics such as the optical Kerr effect, the Rydberg nonlinearity is also long-range and so standard no-go theorems do not apply. Also interesting on a more fundamental level, is that the realisation of strongly-interacting photons allows us to study exotic quantum many-body states of light such as photon liquids or photon crystals [@Vuletic2013b; @Fleischhauer2013; @Buechler2014; @Fleischhauer2015]. The principle of Rydberg nonlinear optics [@Pritchard2013] is simple. The idea is to take the long-range dipolar interaction between highly-excited Rydberg atoms [@Saffman2010; @Pillet2010] and map it onto a large interaction between photons. The difficulty is to localise a photon to the characteristic length scale of the dipole-dipole interaction, typically a few microns. There is more than one way to achieve this localisation: For a single emitter one can reverse the emission process. However in free space, mode matching between the input field and the dipolar emission pattern is challenging and the efficiency is limited to $\sim 10\%$ [@Kurtsiefer2008; @Agio2008; @Sandoghdar2009; @Kurtsiefer2013]. Alternatively, one can use a cavity or waveguide to solve the mode-matching problem. This works well and cavity QED is a well established and extremely successful field where large single photon nonlinearities are possible albeit at the cost of additional complexity of a hybrid system [@Kimble2005; @Rempe2007; @Kimble2008; @Vuletic2013]. Third, in an ensemble of atoms the photon localisation or compression occurs naturally (to some extent) due to the phenomenon of slow light [@Harris1991; @Fleischhauer2005]. A light pulse inside a medium is a mixture of electromagnetic wave and a dipolar excitation, which at the level of single photons we call a polariton [@Fleischhauer2000]. The speed of the polariton and the compression ratio are determined by the group index, $n_{\rm g}$, and hence the dispersive response of the medium. To localise a photon, we would like the group index to be as large as possible. The nonlinear response of the Rydberg medium is proportional to the group index and to the strength of the dipole-dipole interactions and, as we show below, both can be extremely large. A brief history =============== The idea of using Rydberg blockade to generate nonclassical states of light appears in the original blockade paper in 2001 [@Lukin2001c]. Lukin [*et al*]{}. write that the “collective spin states generated by means of dipole blockade $\ldots$, can be transferred from the spin degrees of freedom to the optical field” allowing the creation of interesting quantum states of light “without the use of high-$Q$ cavities”. But at the time, the experimental techniques were not sufficiently advanced to make this work, for example, nearly all experiments involving highly-excited Rydberg states used ionisation for detection and no one had observed a coherent atom-light interaction where the presence of the Rydberg state is read-out directly by an optical field. The key turned out to be the technique of electromagnetically induced transparency (EIT), where an additional control field coupling to a third level renders a medium transparent to resonant light. For more details on this phenomenon, which is now routinely exploited in a wide variety of quantum optics experiments, see the excellent review by Fleischhauer [*et al*]{}. [@Fleischhauer2005]. Most work on EIT has focused on $\Lambda$-type systems, where a strong control laser couples the excitated state to another ground state. But in 2005, Friedler [*et al*]{}. for the first time discussed the idea that one could instead couple the excited state to a highly-excited Rydberg state which are conveniently metastable. They showed that one could transfer the strong interactions between Rydberg atoms onto the optical transition and thereby realise a photonic phase gate [@Kurizki2005]. The first experiments on EIT to highly-excited Rydberg states with principal quantum numbers up to $n=124$ were reported by Mohapatra [*et al.*]{} in 2007 [@Adams2007]. Although this was a classical linear optics experiment, the significant result was that the resonances were narrow, and the combined dephasing and decoherence rates did not exceed $\sim 300~$kHz. This was a breakthrough as it showed that potential problems such as ionisation of the Rydberg atoms were not a ‘show stopper’. The effect of Rydberg blockade on the optical transmission through an ensemble of ultra-cold atoms was first demonstrated in 2010 [@Adams2010]. The first experiments demonstrating manipulation of light at the level of single quantum followed in 2012 by Dudin and Kuzmich [@Kuzmich2012b], Peyronel [et al.]{} [@Vuletic2012], and Maxwell [et al.]{} [@Adams2013]. In this review we focus on the underlying mechanism of quantum nonlinear optics using interacting Rydberg atoms, on progress since 2012, and on the challenges ahead. But before looking at the quantum nonlinearity, we present a simple classical argument of why Rydberg EIT offers the largest optical nonlinearities ever demonstrated. Rydberg nonlinear optics ======================== At the level of a few photons, optical nonlinearities arise when the response of the medium to a second photon is different to the first. This can occur either because the medium cannot absorb or scatter a second photon at the same time — as in the case of a single emitter — or because the resonance condition for the second photon is different. This second case is true for both cavity QED and Rydberg ensembles. For Rydberg-mediated nonlinearities, the first photon creates a Rydberg excitation or Rydberg polariton (where the excitation is spread over many atoms), which both result in a shift of the energy of Rydberg states of nearby atoms. If this shift is signifcantly larger than the excitation linewdith, then a second excitation becomes impossible. This process is known as Rydberg blockade [@Lukin2001c]. To understand the nonlinear optical response of a Rydberg ensemble it is convenient to start with the case of a single photon or less than one photon such that there are no dipole-dipole interactions, and see how we can map the exaggerated electric field sensitivity of a highly-excited Rydberg atom into a strong optical response. From here it is a small step to imagine that this external field arises due to the proximity of another Rydberg atom and hence another photon. Single-photon Rydberg Kerr effect --------------------------------- If an optical nonlinearity arises due to a field-dependent shift in the atomic resonance, then to first order the nonlinear response is proportional to the product of the shift and the gradient of frequency dependence of the refractive index, *i.e.*, the dispersion [@Boyd1992]. It is convenient to parameterise the gradient in the refractive index in terms of the group refractive index which is defined as $$\begin{aligned} n_{\rm g}=1+\omega\frac{\partial n}{\partial \omega}~,\end{aligned}$$ where $\omega$ is the angular frequency of the light. In a dilute medium where the refractive index is close to unity, we can write $n=1+\textstyle{1\over 2}\chi_{\rm r}$, where $\chi_{\rm r}$ is the real part of the electric susceptibility. Hence for a large group index, $n_{\rm g}=\textstyle{1\over 2}\omega \partial \chi_{\rm r}/\partial \omega$. Next we consider the shift of the Rydberg level, which for an external electric field has the form of a dc or ac electric Stark shift, $$\begin{aligned} \Delta_{\rm Ryd}=-\frac{1}{2}\frac{\alpha {\cal E}^2}{\hbar}~,\end{aligned}$$ where $\alpha$ is the atomic polarizabilty at the frequency of the external field, which can be different to the frequency of the optical field. This latter point is the key to origin of large nonlinearities in Rydberg ensembles. The polarizability of Rydberg atoms at optical frequencies is small but the polarizability from dc to microwave frequencies can be enormous [@Gallagher1994]. This fact has been exploited in microwave cavity QED experiments for decades, where individual Rydberg atoms are used as ultra-sensitive probes able to monitor few-photon intra-cavity fields [@Haroche2001]. These low frequency susceptibilities scale as the principal quantum number $n^7$. To qualitatively understand the Rydberg-Rydberg interaction, consider that another Rydberg atom produces a low frequency field ${\cal E}$ proportional to the induced Rydberg dipole which scales as $n^2$. Consequently, when considering only the dipole-dipole term of the interaction Hamiltonian, we obtain a van-der-Waals type interaction scaling as $\alpha{\cal E}^2\sim n^{11}$. Writing the nonlinear optical response as slope $\partial \chi_{\rm r}/\partial \omega$ times shift $\Delta_{\rm Ryd}$, we get a term that scales quadratically with the field, *i.e.*, a Kerr-like effect, $$\begin{aligned} \chi^{(3)}{\cal E}^2&=&\frac{\partial \chi_r}{\partial \omega}\Delta_{\rm Ryd}=-\frac{n_{\rm g}\alpha}{ \omega}{\cal E}^2~.\end{aligned}$$ So the Kerr nonlinearity $\chi^{(3)}$ is proportional to the product of the group index and the polarizability. To get a large group index we would like to work close to resonance or even on resonance, but this has the disadvantage that the imaginary part of the susceptibility is also large, giving rise to off-axis scattering and hence loss. The solution is electromagnetically induced transparency (EIT), where an additional control field renders the medium transparent on resonance due to destructive interference between excitation pathways [@Fleischhauer2005]. Group indices as large as $10^6$ are possible using EIT in atomic ensembles as first demonstrated in 1999 [@Harris1999]. A large group index gives rise to the phenomenon of [*slow light*]{}, enabling compression of the light pulse inside the medium. In addition, the ability to control the group index enables ‘storing’ and retrieving light pulses, which is the basis of [*quantum memory*]{} (See *e.g.*, K. Hammerer [*et al.*]{} for a recent review [@Polzik2010]). By combining EIT and Rydberg states, we get the best of both worlds, *i.e.*, both the largest possible group index and the exaggerated sensitivity of Rydberg state to low frequency fields either applied externally or induced by other nearby Rydberg atoms. Linear EIT susceptibility {#sec:3.2} ------------------------- To see how the large group index arises in an EIT medium, we present a brief derivation of the EIT susceptibility following Gea-Banacloche [*et al*]{}. [@Xiao1995] and Fleischhauer [*et al.*]{} [@Fleischhauer2005]. Consider the level structure in Fig. \[fig:chi\](a) of an atom with states $\vert {\rm g}\rangle$ and $\vert {\rm e}\rangle$ that is excited by a probe laser with Rabi frequency $\Omega_{\rm p}$ and detuning $\Delta_{\rm p}$. The excited state $\vert {\rm e}\rangle$ is coupled to a highly-excited Rydberg state $\vert {\rm r}\rangle$ by a coupling laser with detuning $\Delta_{\rm c}$ and Rabi frequency $\Omega_{\rm c}$. The equations for the coherences of this 3-level ladder system are $$\begin{aligned} \dot{\tilde{\rho}}_{\rm eg} &=&-{\rm i}\frac{\Omega_{\rm p}}{2}(\tilde{\rho}_{\rm gg}-\tilde{\rho}_{\rm ee}) +{\rm i}\Delta_{\rm p}\tilde{\rho}_{\rm eg} -{\rm i}\frac{\Omega_{\rm c}}{2}\tilde{\rho}_{\rm rg}-\gamma\tilde{\rho}_{\rm eg}~\nonumber\\ \dot{\tilde{\rho}}_{\rm rg} &=&{\rm i}\frac{\Omega_{\rm p}}{2}\tilde{\rho}_{\rm re} +{\rm i}(\Delta_{\rm p}+\Delta_{\rm c})\tilde{\rho}_{\rm rg} -{\rm i}\frac{\Omega_{\rm c}}{2}\tilde{\rho}_{\rm eg} -{\gamma_{\rm r}}\tilde{\rho}_{\rm rg}~. \nonumber\\ \dot{\tilde{\rho}}_{\rm re} &=&-{\rm i}\frac{\Omega_{\rm c}}{2}(\tilde{\rho}_{\rm ee}-\tilde{\rho}_{\rm rr}) +{\rm i}\Delta_{\rm c}\tilde{\rho}_{\rm re} +{\rm i}\frac{\Omega_{\rm p}}{2}\tilde{\rho}_{\rm rg} -\gamma'\tilde{\rho}_{\rm re}~,\nonumber\end{aligned}$$ where $\gamma$ and $\gamma'$ are decoherence rates of the driven transitions ($g\leftrightarrow e$ and $e\leftrightarrow r$), which are often much larger than the decoherence rate ${\gamma_{\rm r}}$ of the two-photon transition $g\leftrightarrow r$. If spontaneous emission is the only decay mechanism then $\gamma=\Gamma/2$, where $\Gamma$ is the spontaneous decay rate of state $\vert {\rm e}\rangle$. For the steady-state solution in the weak-probe limit (where the populations $\rho_{\rm gg}=1$, $\rho_{\rm ee}=0$, $\rho_{\rm rr}=0$), we find that $\tilde{\rho}_{\rm re}=0$ and then from the second equation $$\begin{aligned} \tilde{\rho}_{\rm rg} = -\frac{{\rm i}\Omega_{\rm c}/2}{{\gamma_{\rm r}}-{\rm i}(\Delta_{\rm p}+\Delta_{\rm c})}\tilde{\rho}_{\rm eg}~,\nonumber\end{aligned}$$ and if we substitute this into the steady-state solution of the first equation, we find an expression for the coherence on the probe transition $$\begin{aligned} \tilde{\rho}_{\rm eg} = -\frac{{\rm i}\Omega_{\rm p}/2}{\gamma-{\rm i}\Delta_{\rm p}+\Omega_{\rm c}^2/[{\gamma_{\rm r}}-{\rm i}(\Delta_{\rm p}+\Delta_{\rm c})]/4}~,\nonumber\end{aligned}$$ which determines the induced dipole on the probe transition. The resulting electrical susceptibility is $$\begin{aligned} \chi_{\rm 3-level}=\frac{2\gamma}{k {l_{\rm a}}}\frac{\tilde{\rho}_{\rm eg}}{\Omega_{\rm p}}=\chi_{\rm 2-level}\left[1-\frac{\Omega_{\rm c}^2}{4({\gamma_{\rm r}}-{\rm i}{\Delta_{\rm p}}-{\rm i}{\Delta_{\rm c}})(\gamma-{\rm i}{\Delta_{\rm p}})+\Omega_{\rm c}^2}\right]~, \label{eq:chi_eit}\end{aligned}$$ with the susceptibility of the bare two-level system ($\Omega_{\rm c}=0$) given by $$\begin{aligned} \chi_{\rm 2-level}=\frac{1}{k {l_{\rm a}}} \frac{{\rm i}\gamma}{\gamma-{\rm i}\Delta_{\rm p}}~. \label{eq:chi_2}\end{aligned}$$ Here ${l_{\rm a}}= (N\sigma)^{-1}$ is the resonant attenuation length, with $N$ the number density of atoms and $\sigma$ the optical cross-section ($\sigma=3\lambda^2/2\pi$ for a closed two-level transition). These susceptibilities are plotted in Figs. \[fig:chi\](b,c). ![Electromagnetically-induced transparency (EIT) with Rydberg states. (a) Level scheme of an unperturbed atom (left) and with a shifted Rydberg level (right). (b+c) The real part of the susceptibility $\chi_{\rm 3-level}$ of an EIT medium from Eq. (\[eq:chi\_eit\]). The insets show the corresponding imaginary part, over the same frequency range. The dashed lines are the unperturbed EIT susceptibilities, with the control field either (b) on resonance or (c) red detuned (${\Delta_{\rm c}}\approx-\Gamma$) from resonance. The solid lines show the effect of a shift of the Rydberg state: In (b) we demonstrate the effect of a small shift, $\Delta_{\rm Ryd}=-0.1\Gamma$, where the approximation that the nonlinear response equals slope$\times$shift holds. In (c) we demonstrate a large shift, $|\Delta_{\rm Ryd}|>\Gamma$, where the susceptibility reverts to the two-level response, $\chi_{\rm 2-level}$. This corresponds to the case of *Rydberg blockade*. []{data-label="fig:chi"}](figures/fig_1_proof.pdf){width="8.6"} To simplify the presentation, we now focus on the resonance case ${\Delta_{\rm c}}=0$. Whereas in the two-level system we have maximum scattering on resonance, in EIT the scattering is suppressed. Minimum scattering is obtained at two-photon resonance ${\Delta_{\rm p}}=-{\Delta_{\rm c}}=0$, where $$\chi_{\rm 3-level}({\Delta_{\rm p}}=-{\Delta_{\rm c}}=0)=\chi_{\rm 2-level}\left[1-\frac{\Omega_{\rm c}^2}{4{\gamma_{\rm r}}\gamma+\Omega_{\rm c}^2}\right]~,$$ so we require $\Omega_{\rm c}^2 \gg 4{\gamma_{\rm r}}\gamma$ to induce significant transparency. In this regime, the EIT linewidth ${\Gamma_{\rm EIT}}=\Omega_{\rm c}^2/(4\gamma)$ is dominated by power broadening (${\Gamma_{\rm EIT}}\gg{\gamma_{\rm r}}$). Assuming an EIT transparency window much narrower than the one-photon absorption line (${\Gamma_{\rm EIT}}\ll\gamma$), we find from Eq. (\[eq:chi\_eit\]) $$\chi_{\rm 3-level}\approx\chi_{\rm 2-level}\left(1-\frac{{\Gamma_{\rm EIT}}}{{\Gamma_{\rm EIT}}-{\rm i}{\Delta_{\rm p}}}\right)~$$ for probe tuned within the EIT linewidth ($|{\Delta_{\rm p}}|\ll{\Gamma_{\rm EIT}}$ and ${\Delta_{\rm c}}=0$). If we now linearize the susceptibility around $\Delta_{\rm p}=0$, we find a real part $\chi_{\rm r}\approx(k {l_{\rm a}})^{-1} {\Delta_{\rm p}}/ {\Gamma_{\rm EIT}}$ and a corresponding group index $$\begin{aligned} n_{{\rm g}}\approx\frac{\omega}{2}\frac{\partial \chi_{\rm r}}{\partial {\Delta_{\rm p}}} &\approx& \frac{1}{k {l_{\rm a}}} \frac{\omega}{2{\Gamma_{\rm EIT}}}~.\label{eq:n_g_EIT}\end{aligned}$$ Thus narrow EIT resonances enable a large group index and hence large optical nonlinearity. To obtain a narrow resonance (small ${\Gamma_{\rm EIT}}$) while satisfying the above requirement ${\Gamma_{\rm EIT}}\gg{\gamma_{\rm r}}$ to guarantee significant transparency, we require a long-lived state $\vert {\rm r}\rangle$ (small ${\gamma_{\rm r}}$). This is indeed the case if $\vert {\rm r}\rangle$ is another ground state ($\Lambda$–EIT) or a Rydberg state. The nonlinear response arises from a level shift which changes $\Delta_{\rm c}$, as illustrated in Fig. \[fig:chi\]. The success of Rydberg nonlinear optics relies on the ability to map the large shifts arising from low frequency fields onto an optical field using EIT. The first experiment demonstrating a large Kerr effect due to an external electric field using Rydberg EIT was reported in 2008 [@Adams2008]. Optical nonlinearity due to Rydberg blockade -------------------------------------------- The probe field propagates in the Rydberg-EIT medium as a so-called Rydberg polariton with a group velocity $$v_{{\rm g}}=\frac{c}{n_{{\rm g}}}.$$ A large group index thus implies a small photonic component, on order $v_{{\rm g}}/c$, and correspondingly a large Rydberg component. The interaction between the Rydberg atoms shifts the Rydberg level, effectively altering the control-field detuning ${\Delta_{\rm c}}$. The term *Rydberg blockade* refers to the case where the interaction-induced shift is much larger than the EIT linewidth. In this case, the nonlinearity can be considered as a switch from the 3-level EIT susceptibility $\chi_{\rm 3-level}$ to the 2-level susceptibility $\chi_{\rm 2-level}$ [@Pohl2011b; @Lukin2011; @Fleischhauer2011], as illustrated in Fig. \[fig:chi\](c). The volume around a Rydberg polariton in which EIT is suppressed is known as the blockade sphere. Its radius is found from the requirement $V({r_{\rm b}})=2\hbar{\Gamma_{\rm EIT}}$, where $V(r)$ is the Rydberg-Rydberg interaction potential. For a van der Waals interaction $V(r)=C_6/r^6$, we find $${r_{\rm b}}=\sqrt[6]{C_6/(2\hbar{\Gamma_{\rm EIT}})}.$$ If there are enough atoms contributing to the 2-level susceptibility within the blockade sphere, the effect of a single Rydberg polariton on the transmission of nearby photons can be dramatic. This is the quantum nonlinear regime that will be the focus of the rest of this article and requires that the medium has a high optical depth per blockade sphere. However, even in the ‘classical’ – or partially blockaded – regime, the non-linearities can be enormous! For a weak probe, $\Omega_{\rm p}\ll \Omega_{\rm c}$, as the probe intensity increases there is a gradual transition from the 3–level to the 2–level response as each successive Rydberg excitation converts a fraction $(\Omega_{\rm p}/\Omega_{\rm c})^2$ of nearby atoms to 2–level scatterers. It follows that the classical Rydberg non-linearity scales as the 2–level response times the fraction of blockading excitations [@Pohl2011b]: $$\begin{aligned} \chi^{(3)}&=&N\frac{4\pi}{3}r_{\rm b}^3\frac{\Omega_{\rm p}^2}{\Omega_{\rm c}^2}\chi_{\rm 2-level}~.\end{aligned}$$ Substituting $\Omega_{\rm p}=d{\cal E}_{\rm p}/\hbar$, where $d$ is the dipole matrix element for the 2–level transition, this equation gives a [*Kerr-like optical non-linearity*]{}. For ${r_{\rm b}}=5~\mu$m (typical for principal quantum numbers $50-100$ and $\Omega_c$ on order of a few MHz, satisfying $\Omega_{\rm c}^2 \gg 4{\gamma_{\rm r}}\gamma$) and $N=3\times 10^{12}~{\rm cm}^{-3}$ in Rb, one obtains an estimate of the Rydberg nonlinearity of $\chi_{\rm 3-level}\sim 5\times 10^{-2}~{\rm V}^{-2}{\rm m}^2$, which is 5 orders of magnitude larger than conventional EIT media [@Fleischhauer2005]. This classical nonlinearity was first demonstrated experimentally in 2010 using rubidium atoms prepared at densities of $N=2\times 10^{10}~{\rm cm}^{-3}$ using a magneto-optical trap [@Adams2010]. Having reviewed the classical nonlinearity arising from Rydberg EIT we now move on to the quantum limit. Quantum nonlinearity ==================== Quantum nonlinear optics [@Lukin2014b], that is, the extreme limit where the optical nonlinearity becomes significant on the level of single photons, calls for a quantum description of the probe field. In this limit, the nonlinearity is characterized by comparing the transmission amplitude of single photons to that of a photon pair. If single photons are transmitted much better than photon pairs, or conversely absorbed (scattered) much more, we denote the nonlinearity as *dissipative*. This type of nonlinearity can be quantified, almost by definition, by measuring the normalized second-order correlation function $g^{(2)}(t_1,t_2)=\langle n(t_1)n(t_2) \rangle / [ \langle n(t_1)\rangle \langle n(t_2) \rangle ]$ of the outgoing field for a weak classical input. Here, $t_1$ and $t_2$ are the photon detection times, and $n(t)$ the detection rate. For example, anti-bunching $g^{(2)}(t_1=t_2) \ll 1$ would indicate strong scattering of pairs (the numerator of $g^{(2)}$) relatively to singles (approximately the denominator). The other type of nonlinearity is the so-called *dispersive*, when singles and pairs are equally transmitted (*e.g.*, in a lossless medium) but acquire different optical phases. Strong nonlinearity is then reached when the phase difference, sometimes refereed to as the *conditional* or *nonlinear phase* $\phi$, is on the order of $\pi$. We discuss below how one could generalize the standard $g^{(2)}$ measurement to characterize the conditional phase. In a Rydberg-EIT setup, the type of the nonlinearity is determined by the detunings ${\Delta_{\rm p}}$ of the probe from the intermediate level $\vert e\rangle$ [@Lukin2011]. Each polariton, carrying a single Rydberg excitation, suppresses EIT within the blockade sphere around it. The transmission amplitude for other photons crossing this blockade sphere is thus given by the bare two-level response $$t_{\rm 2-level}({\Delta_{\rm p}})=\exp\left (- \frac{\gamma}{\gamma-{\rm i}{\Delta_{\rm p}}} \frac{{{\rm {OD_b}}}}{2} \right). \label{eq:t2}$$ Here the optical depth of the blockade sphere ${{\rm {OD_b}}}=2{r_{\rm b}}/{l_{\rm a}}$ depends on the sphere diameter $2{r_{\rm b}}$ and the attenuation length ${l_{\rm a}}$. The quantum nonlinearity follows from the ratio $t_{\rm 2-level}/t_{\rm 3-level}$, which can be estimated from the classical 2- and 3-level susceptibilities of the form plotted in Figs. \[fig:chi\](b,c).Since $t_{\rm 3-level}=1$ at the EIT resonance (assuming ${\Gamma_{\rm EIT}}\gg {\gamma_{\rm r}}$), we only need to examine $t_{\rm 2-level}$: For a resonant probe, $t_{\rm 2-level}({\Delta_{\rm p}}=0)=\exp(-{{\rm {OD_b}}}/2)$ leads to scattering of blocked photons and thus to dissipative nonlinearity. For an off-resonant probe, $t_{\rm 2-level}(|{\Delta_{\rm p}}|\gg \gamma)\approx\exp({\rm i}\phi)$ yields no loss and a conditional phase of $\phi=-({{\rm {OD_b}}}/2)(\gamma/{\Delta_{\rm p}})$, thus rendering a dispersive nonlinearity. Dissipative nonlinearities {#subsec:dissipative_nonlinearities} -------------------------- It follows from the above expressions that the strength of the nonlinearity is governed by ${{\rm {OD_b}}}$, and we require ${{\rm {OD_b}}}\ge1$ for the *quantum* nonlinear limit, in both the dissipative and the dispersive regimes. A medium with ${{\rm {OD_b}}}\approx 5$ was realized in 2012 by Peyronel [*et al.*]{} [@Vuletic2012]. In this experiment, an elongated optical trap compressed a cloud of ultracold rubidium to an atomic density of $N= 2\times 10^{12}~{\rm cm}^{-3}$. The Rydberg-EIT had a (half) linewidth of ${\Gamma_{\rm EIT}}/(2\pi)\approx 10~{\rm MHz}$ when tuned to the Rydberg level 100S$_{1/2}$, yielding a blockade radius on order ${r_{\rm b}}\approx 10~\mu{\rm m}$ and approximately 10,000 atoms within a single blockade sphere. The probe field was focused throughout the $100~\mu{\rm m}$-long cloud to a diameter $2w_0$ on the order of ${r_{\rm b}}$. By that, the condition $w_0\ll{r_{\rm b}}$, for keeping the dynamics one-dimensional and blocking two polaritons from propagating side-by-side, was nearly fulfilled. ![(a) Transmission spectra of EIT using the Rydberg state 100S$_{1/2}$ for various incoming photon rates. The weak field transmission is determined by the 3-level susceptibility, Eq. (\[eq:chi\_eit\]), plotted in Fig. \[fig:chi\](b) inset. However, the transmission on resonance begins to be substantially reduced at a level of a few photons per $\mu$s. With the group delay in the medium being $\tau_{\rm d}\approx 0.25~\mu$s, this rate corresponds to (on average) less than two photons inside the medium. (b) Outgoing photon rate (rescaled to compensate for the $50\%$ linear transmission and for the finite detection efficiency) versus incoming photon rate. The transmission is saturated at about one photon per $\mu$s. The dashed curves are expected rates assuming that multi-photon events are either (black) blocked or (green) converted into a one-photon state. (c) Normalized second-order correlation function $g^{(2)}$ of the outgoing photons versus their time separation $\tau$. The anti-bunching feature, a result of the dissipative nonlinearity, has a temporal width $\tau_{\rm c}$ on the order of the group delay $\tau_{\rm d}$. Inset: a reference experiment with the Rydberg state 46S$_{1/2}$, showing a negligible effect. The solid lines are results of full numerical simulations of the 2-photon wavefunction evolution in the medium. (d) The anti-bunching temporal width $\tau_{\rm c}$ versus the inverse bandwidth $B={\Gamma_{\rm EIT}}/\sqrt{8{\rm OD}}$ for various experimental parameters (different symbols). The finite transmission bandwidth $B$ broadens the anti-bunching feature in the two-photon wavefunction during propagation. Figure adapted from Peyronel [*et al.*]{} [@Vuletic2012].[]{data-label="fig:dissipativeQNLO"}](figures/fig_dissipative.pdf){width="8.6"} In these conditions, nonlinear transmission was measured at probe powers as low as 0.25 pW, corresponding to less than two photons in the medium — see Fig. \[fig:dissipativeQNLO\](a) and caption. For incoming photon rates above $2~\mu{\rm s}^{-1}$, the outgoing photon rate became constant \[Fig. \[fig:dissipativeQNLO\](b)\], realizing a photonic version of an hourglass which allows the transmission of only about one photon per $\mu$s. The quantum nature of the nonlinearity is evidenced by a strong photon anti-bunching of the outgoing light \[Fig. \[fig:dissipativeQNLO\](c)\], measured with single-photon detectors. In the regime of the experiment, the (linear) transmission bandwidth of the medium, which sets a lower limit on the duration of probe pulses, sets a similar lower limit on the temporal extent of the anti-bunching feature \[Fig. \[fig:dissipativeQNLO\](d)\]. Dispersive nonlinearities {#subsec:dispersive_nonlinearities} ------------------------- Changing the nonlinearity in a Rydberg-EIT experiment from dissipative to dispersive is straight forward: one simply detunes the probe and control fields from the intermediate state. Firstenberg [*et al.*]{} realized this in 2013 with detunings ${\Delta_{\rm p}}\approx-{\Delta_{\rm c}}\approx 5\gamma$ [@Vuletic2013b]. The measured transmission and phase-shift spectra, corresponding respectively to the imaginary and real parts of the susceptibility, are shown in Fig. \[fig:dispersiveQNLO1\]. At the chosen detuning, the 3-level and 2-level responses differ only in their phase, yielding purely dispersive nonlinearity. Similarly to the case of dissipative nonlinearity, the measured spectra alone do not suffice to characterize the quantum nonlinearity, and 2-photon correlation measurements are needed. Since now it is a phase, rather than transmission, that is of interest, an interferometric setup is required. To this end, one introduces a reference photonic mode, *e.g.,* with a different polarization or frequency, for which the transmission is linear. By measuring the correlations between the probe and the reference photons when interfered in different bases, one can utilize quantum tomography techniques to reconstruct the full two-photon density matrix $\rho(t_1,t_2)$ (with $t_1$ and $t_2$ being the detection times of the two photons). Normalizing this density matrix by the one-photon density matrices $\rho(t_1)\otimes\rho(t_2)$ renders an *interaction matrix* $\tilde{\rho}(t_1,t_2)$ that generalizes the standard $g^{(2)}(t_1,t_2)$. The interaction matrix, $\tilde{\rho}(t_1,t_2)$, yields not only the conditional phase $\phi$, but also information on the decoherence and entanglement generation during the process. Experimental results extracted from $\tilde{\rho}(t_1,t_2)$ in the 2013 experiment [@Vuletic2013b] are shown in Figs. \[fig:dispersiveQNLO2\](a-b). A conditional phase shift as high as $|\phi|=\pi/4$ was observed in this experiment at a linear transmission of $50\%$. Photon-photon interaction ------------------------- The optical nonlinearity we observe originates from the strong dipolar interaction between Rydberg atoms. A reciprocal and complimentary notion is the *effective interaction between photons* that can be used to describe the nonlinearity at the quantum level. In fact, this effective photon-photon interaction inherits from the dipolar atomic interaction, but is regulated by the optical response of the medium. Dissipative and dispersive nonlinearities thus result from effective dissipative or dispersive photon-photon interactions. To illustrate this point, we define the 2-photon wavefunction $\psi(z_1,z_2)$ [@Vuletic2013b], with $z_1$ and $z_2$ the photon coordinates inside the medium. $|\psi|^2$ and ${\rm arg}(\psi)$ are respectively the probability to find the two photons and their phase-shift relatively to the non-interacting (Poissonian) case. In the absence of nonlinearity, $\psi=1$. Moving to the center-of-mass $R=(z_1+z_2)/2$ and relative $r=z_1-z_2$ coordinates, we can approximately relate $\psi(R,r)$ to the outgoing light by $$\psi(R=L,r=v_{{\rm g}}\tau)=\sqrt{g^{(2)}(\tau)}{\rm e}^{{\rm i}\phi(\tau)}.\nonumber$$ It can then be shown that the evolution of a stationary Poissonian input $\psi(R=0)=1$ in a dispersive nonlinear medium, assuming $\Omega_{\rm c}\ll{\Delta_{\rm p}}$, is given approximately by a Schrödinger-like equation [@Vuletic2013b] $${\rm i}\frac{\partial \psi}{\partial R} = \frac{4{l_{\rm a}}{\Delta_{\rm p}}}{\gamma} \frac{\partial ^{2}\psi }{\partial r^{2}}+ \frac{\gamma}{{l_{\rm a}}|{\Delta_{\rm p}}|} U(r)\psi. \label{eq:RydSch}$$ The center-of-mass coordinate, varying from $R=0$ to $R=L$, plays the role of time in this Schrödinger evolution. The first term on the right-hand side accounts for an effective photon mass. It stems from the finite bandwidth of the linear EIT transmission, rendering a quadratic dispersion of the individual polaritons. The second term describes an effective potential $$\begin{aligned} &U(r)=\frac{1}{(r/{r_{\rm b}})^6 + {\rm sign}({\Delta_{\rm p}})}=\left\{ \begin{array}{cc} {\rm sign}({\Delta_{\rm p}}) & r\ll {r_{\rm b}}\\ 0 & r\gg {r_{\rm b}}\end{array}\right.,\label{eq:effpot}\end{aligned}$$ where ${\rm sign}({\Delta_{\rm p}})={\Delta_{\rm p}}/|{\Delta_{\rm p}}|$ and ${r_{\rm b}}=\sqrt[6]{2C_6|{\Delta_{\rm p}}|/(\hbar\Omega_{\rm c}^2)}$ (in an off-resonance EIT, where ${\Gamma_{\rm EIT}}=\Omega_{\rm c}^2/|4{\Delta_{\rm p}}|$). Equation (\[eq:effpot\]) assumes a repulsive van-der-Waals interaction between the Rydberg atoms, $C_6/r^6$ with $C_6>0$. We thus find that the behaviour of a photon pair in a Rydberg-EIT medium resembles that of a pair of interacting massive particles. Since ${\Delta_{\rm p}}$ determines the sign of both the mass and the potential, the overall behaviour is that of an *attractive potential well*, or an attractive force, irrespective of the sign of ${\Delta_{\rm p}}$. However, the well $U(r)$ is well-behaved only for ${\Delta_{\rm p}}>0$; for ${\Delta_{\rm p}}<0$, the denominator in Eq. (\[eq:effpot\]) vanishes at the boundaries ($|r|\approx {r_{\rm b}}$) due to a resonant Raman absorption, creating local resonance-like features in $U(r)$. For ${\Delta_{\rm p}}>0$, bound-state solutions of Eq. (\[eq:RydSch\]) may be termed ‘molecules’ of two photons. In experiments so far [@Vuletic2013b], the well was shallow and supported only a single bound state $\psi_{\rm bound}$. The “finite-time” evolution (from $R=0$ to $R=L$) following Eq. (\[eq:RydSch\]) is then governed by $\psi_{\rm bound}$. The initial state describing the lack of photon-photon correlations at the entrance to the medium is $\psi(R=0,r)=1$ ([*i.e.*]{}, a uniform distribution in the relative coordinate $r$), which is a superposition of a bound-state component $\psi_{\rm bound}$ and a scattering component $1-\psi_{\rm bound}$. The difference in the accumulated phase between these two components leads to constructive interference between them at $r=0$, and hence to photon bunching [@Vuletic2013b]. This bunching and a reminiscence of the bound state are shown in Figs. \[fig:dispersiveQNLO2\](a,c). The case ${\Delta_{\rm p}}<0$ was theoretically analyzed by Maghrebi, Gullans, [*et al.*]{} [@Gorshkov2015b]. The resonance-like features in $U(r)$ resemble a Coulomb potential at $|r|\lesssim{r_{\rm b}}$. They support a continuum of metastable bound states with an hydrogen-like energy spectrum, with the two polaritons separated by a finite distance $\sim{r_{\rm b}}$. In the latter sense, this type of photonic ’molecule’ perhaps resembles a real molecule more than the one at ${\Delta_{\rm p}}>0$ (described above), which peaks at zero separation. However, the metastable states propagate in the medium with negative group velocity while decaying to a pair of Rydberg atoms at a rate $|\Omega_{\rm c}^2/{\Delta_{\rm p}}|/\phi^2$ \[where $\phi=-({{\rm {OD_b}}}/2)(\gamma/{\Delta_{\rm p}})$ as defined above\]. Correlated states: from two to many photons ------------------------------------------- The successful realization of quantum nonlinearity with Rydberg-EIT prompted great theoretical efforts to better describe and understand the system. A full description of a uniform EIT system involves one dark and two bright polariton branches, obtained by diagonalizing the non-interacting Hamiltonian in momentum space. In this basis, the Rydberg interaction appears as a non-trivial local scatterer, coupling between the different branches. The scattering properties for the case of two photons in one dimension were derived by Bienias [*et al.*]{} [@Buechler2014] using quantum scattering theory. This approach requires very little assumptions and is thus applicable for a wide range of parameters, including for large blockade radii and strong control fields $\Omega_{\rm c}$, on the order of or even much larger than ${\Delta_{\rm p}}$. In particular, Bienias [*et al.*]{} show that increasing $\Omega_{\rm c}$ can modify the 1D scattering length from attractive, as in Eq. (\[eq:RydSch\]), to repulsive. Such transition occurs at a scattering resonance similar to a Feshbach resonance. Furthermore, this approach provides a generalization of Eq. (\[eq:RydSch\]) to the non-stationary regime: For fields slowly-varying in time, one only needs to replace $\partial/\partial R\rightarrow\partial/\partial R+\partial/(v_{{\rm g}}\partial t)$, rendering a spatio-temporal Schrödinger-like dynamics. Finally, an effective *many-body* Hamiltonian can be formulated in terms of the 1D scattering length for low energies. An alternative approach to scattering theory is the input-output formalism. In this formalism, the system is described in the Heisenberg picture by defining a set of operators for the optical modes, including special operators for the input and output modes. For many quantum optics systems, this approach is more natural and proves extremely useful [@Collett1985; @Fan2010]. Recently, Caneva [*et al.*]{} introduced a generalized input-output formalism for describing the Rydberg-EIT system [@Chang2015]. They effectively model the system as a one-dimensional chain of interacting three-level atoms (a spin model) that is tailored to reproduce the mean-field parameters of the real continuous medium, such as OD and $v_{{\rm g}}$. Calculating high-order correlations of the outgoing photonic state is then done in a relatively straight-forward procedure by solving for higher moments of the Heisenberg operators. An exact many-body formulation of the continuous system was recently presented by Moos [*et al.*]{} [@Fleischhauer2015]. The model includes photon loss from the bright polariton branches, arising from dark-bright coupling due to the Rydberg-Rydberg interaction. The effect of finite beam size (paraxial propagation) is also considered. Even so, an effective one-dimensional, many-body, model for only the dark polaritons is still valid under certain conditions, as verified by exact numerical simulations [@Fleischhauer2015]. An exciting prospect for quantum nonlinear optical systems is their potential to realize strongly-correlated states and dynamical many-body phases with photons. While experiments are being setup to pursue this regime, there have been a few theoretical predictions for a many-body behavior. Honer [*et al.*]{} proposed a single-photon absorber based on an effective two level system (large ${\Delta_{\rm p}}$), with the Rydberg state sufficiently interacting such that a single excitation blockades the complete optical medium [@Buechler2011]. In this case, after the first absorption event, the blockaded medium becomes completely transparent for all subsequent signal photons. If the optical depth is sufficient to absorb the first photon with large probability, this removes with high fidelity exactly one photon from an arbitrary input state. In the opposite limit of dissipative nonlinearity (${\Delta_{\rm p}}=0$), the system transmits single photons while scattering the multi-photon components. The back-action of this scattering on the properties of the transmitted photon was investigated by Gorshkov [*et al.*]{} [@Pohl2013]. This work implies that when the medium is smaller than one blockade sphere, it will transform an intense coherent input into a stream of single photons with a well-defined separation. Such a photonic state, a so-called one-dimensional crystal of photons, promises to be a valuable resource for metrology and quantum computation. In the absence of dissipation, with only dispersive (conservative) interaction between the photons, crystallization of an incoming coherent field is akin to Wigner crystallization of electrons. This many-body process, which requires dynamic control of the group-velocity, was studied by Otterbach [*et al.*]{} using Luttinger liquid theory in the dilute, low-energy regime [@Fleischhauer2013] and afterwards validated by Moos [*et al.*]{} using numerical simulations of their many-body model [@Fleischhauer2015]. Applications ============ ![Schematic of different schemes for making individual photons interact inside a Rydberg medium. Photons can interact while (a) counter-propagating or (b) co-propagating through the medium. (c) Alternatively, one (or more) photons can first be completely stopped and stored in the medium as a Rydberg atom and then interact with subsequent photons sent into the medium.[]{data-label="fig:photonphoton_schematic"}](figures/collisions.png){width="7"} The effective interaction between individual photons discussed in the previous section enables a variety of optical quantum information applications. Since the first proposal for a photonic phase gate using Rydberg-EIT [@Kurizki2005], a range of different ideas have been suggested for photonic quantum gates or non-classical light sources [@Kurizki2005; @Lukin2011; @Buechler2011]. The immense experimental progress in the last years has lead to a number of demonstrations of applications making use of the Rydberg interaction in an atomic medium. The dissipative interaction (section \[subsec:dissipative\_nonlinearities\]) has been used to demonstrate highly efficient single-photon generation [@Kuzmich2012b; @Adams2013], atom-photon entanglement [@Kuzmich2013], as well as single-photon all-optical switches [@Duerr2014] and transistors [@Hofferberth2014; @Rempe2014b]. In turn, the dispersive nonlinearity (section \[subsec:dispersive\_nonlinearities\]) has been exploited to imprint large conditional phase shifts on weak target pulses. Very recently, a record phase-shift exceeding $\pi$ conditioned on the storage of a single gate photon has been reported by Tiarks [*et al*]{}. [@Duerr2016]. All these application make use of the already discussed Rydberg EIT ladder scheme \[Fig. \[fig:chi\](a)\]. Single photons or weak coherent ‘target’ pulses on or near resonance with the $\vert {\rm g}\rangle\leftrightarrow\vert {\rm e}\rangle$ transition are sent into the optically thick medium, while a strong control field couples $\vert {\rm e}\rangle$ to the target Rydberg state $\vert {\rm r}\rangle$. For conditional operations, a second weak ‘gate’ pulse is coupled either to the same or via a second control laser to a different Rydberg state $\vert {\rm r'}\rangle$ [@Duerr2014; @Hofferberth2014; @Rempe2014b]. Employing different Rydberg states simultaneously greatly enhances the flexibility of the implemented schemes as target and gate photons can for example be individually slowed and stored in the medium. Various different schemes for achieving interaction either between different photons in the target pulse or between target and gate photons have been proposed and implemented (Fig. \[fig:photonphoton\_schematic\]). Friedler [*et al*]{}. initially considered two single photon pulses counter-propagating through the medium. For dispersive interaction this results in a phase-imprint during the “collision” of the two slowly-propagating polaritons, which maps into a phase-shift of the optical field outside the medium [@Kurizki2005]. Alternatively, signal and gate photons can co-propagate through the medium, in which case they will interact during their travel time through the medium [@Vuletic2012; @Vuletic2013b]. While this can maximize the interaction time, it may be more challenging to separate signal and gate photons and thus to perform a controlled operation on a target pulse. Finally, the ability to completely stop a photon and convert it into a stored spin-wave inside the medium enables the configuration shown in Fig. \[fig:photonphoton\_schematic\](c): one photon is first stored and subsequently interacts with multiple photons propagating through the medium. This scheme is particularly suited for applications where a single gate photon should interact with many signal photons, as in the optical transistor applications [@Hofferberth2014; @Rempe2014b]. One has to keep in mind though, that the retrieval of the stored gate photon, required for the realization of a full quantum gate, can be strongly affected by the interaction with the signal photons [@Lesanovsky2015; @Hofferberth2015b]. In the following, we review three specific applications, which are currently being explored both experimentally and theoretically, namely storage-based generation of non-classical light, photonic two-qubit phase gates, and all-optical switches and transistors. Generation of non-classical light from collective Rydberg excitations {#subsec:generation_of_nonclassical_light} --------------------------------------------------------------------- ![Manipulation of light stored as a Rydberg excitation using an external microwave field [@Adams2013; @Adams2014b]. The microwave field drives Rabi oscillations between the polariton Rydberg state and another Rydberg level as depicted schematically in the inset (upper left). This induces resonant dipole-dipole interactions which modify the photon statistics suppressing $g^{(2)}$ at the first revival in the retrieved pulse, inset (upper right). Figure adapted from Maxwell [*et al.*]{} [@Adams2013; @Adams2014b].[]{data-label="fig:Adams_singlephoton_microwave"}](figures/fig_Adams_singlephotonsMW.pdf){width="0.7\columnwidth"} The concept of a Rydberg-mediated non-classical light source is to convert a weak coherent input pulse into non-classical output via the effective interaction inside the medium. In other words, the Rydberg medium acts as a filter for the input light, changing its photon statistics based on how the Rydberg interaction in the medium is coupled to the signal photons. The Rydberg blockade mechanism enables such operation with the aid of EIT and slow-light, but also outside the EIT regime. The first quantum light source originating from a Rydberg excitation was demonstrated by Dudin and Kuzmich [@Kuzmich2012b] in 2012. A coherent signal pulse with large detuning ${\Delta_{\rm p}}$ creates a collective Rydberg excitation inside a medium that is larger than a single blockade. The long-range interaction between different excitations results in strong dephasing of the initial many-atom state [@Kuzmich2012b; @Grangier2012b]. As a consequence, when a resonant (${\Delta_{\rm p}}=0$) readout field is turned on after some time, the retrieved signal photons are scattered in random directions. On the other hand, if only a single excitation is created by the input beam, no dephasing occurs and a single photon is retrieved in the phase-matched direction. The single-photon character of the signal output was characterized in the experiment by measuring the second order correlation function. The reported value of $g^{(2)}(0) = 0.040(14)$ shows the extreme suppression of readout containing more than one photon. Furthermore, for an optical medium smaller than a single blockade volume, Dudin [*et al.*]{} showed that this optical readout can be used to probe the dynamics of the collective Rydberg excitation [@Kuzmich2012c]. This approach to filtering a coherent input field to achieve a non-classical output was further explored by Maxwell [*et al*]{}. [@Adams2013; @Adams2014b]. In their experiment, the signal photons were stored as a collective excitations of a Rydberg $S$-state by turing off the control field during the probe pulse. The Rydberg excitation was subsequently coupled to a neighbouring $P$-state using a resonant microwave field before retrieving the stored light by turing the control field back on (Fig. \[fig:Adams\_singlephoton\_microwave\]). Due to the microwave field, the interaction between excitations was tuned from van-der-Waals to dipolar, which changes both the strength and the angular dependence [@Cote2005; @Walker2008; @Hofferberth2015]. The resulting dynamics of the interacting Rydberg excitations [@Ates2013] became visible in the photon statistics of the retrieved signal light, as is shown in Fig. \[fig:Adams\_singlephoton\_microwave\]. From this first demonstration it becomes apparent that the extreme tunability of the Rydberg interaction opens further possibilities for tailoring the stored light and for realizing more complex non-classical light states. Photonic phase gates {#subsec:photonic_phase_gates} -------------------- To implement a two-photon phase gate, the dispersive interaction (section \[subsec:dispersive\_nonlinearities\]) between polaritons is used. The case of two counter-propagating polaritons was first discussed by Friedler [*et al*]{}. [@Kurizki2005] and further explored and extended to co-propagating and stored pulses by Gorshkov [*et al.*]{} [@Lukin2011]. The basic idea is the same in all configurations: the interaction between two individual polaritons results in a $\pi$-phase shift imprinted on the transmitted light, while the large detuning from the intermediate state results in (ideally) zero scattering of the photons. To turn such a conditional phase shift into a quantum gate, the mechanism needs to be state-dependent with regard to whatever basis is used to encode the quantum information in individual photons. A straightforward example is the common polarization encoding, in which case selection rules can be used to couple only one specific combination of signal and gate photon polarization to the Rydberg state [@Vuletic2013b]. A Rydberg-interaction mediated phase shift of an optical pulse was first reported by Parigi [*et al.*]{} in 2012 employing an atomic ensemble inside an optical resonator [@Grangier2012]. Tiarks [*et al.*]{} have very recently reported a conditional phase shift for single photons exceeding $\pi$, by making use of an optical medium with large optical density ${\rm OD}=25$ and storage and retrieval of the gate photon [@Duerr2016]. As explained in section \[subsec:dispersive\_nonlinearities\], the characterization of the conditional phase-shift requires an interferometric measurement. Tiarks [*et al.*]{} employ quantum state tomography in the polarization basis, similarly to that used by Firstenberg [*et al.*]{} in the first demonstration of Rydberg-mediated dispersive interactions [@Vuletic2013b]. Besides achieving record single-photon phase shifts, the Rydberg-mediated approach may overcome fundamental limitations of single-photon $\pi$ phase gates in conventional nonlinear media [@Shapiro2006; @GeaBanacloche2010]. The key point here is that due to the long-range interaction the nonlinearity is no longer local, and the phase-shift can be uniform over the full size of the stored spin wave [@Adams2014], which should allow a high-fidelity phase gate without the unavoidable pulse distortion for conventional nonlinearities. Optical switches and transistors {#subsec:optical_switches} -------------------------------- ![(a) Switch contrast (red), i.e. the relative attenuation of the source beam, of a Rydberg mediated single photon transistor as function of mean photon number $\overline{N}_{g,in}$ in a coherent gate pulse. The experimentally observed attenuation of the source light is mainly determined by the Poissonian statistics of the gate photon number, with the dashed line giving the fundamental limit $C_{coh} = 1- e^{-\overline{N}_{g,in}}$ for a coherent input pulse. Knowing the gate photon statistics, the achievable switch contrast for one-, two- and three-photon Fock input states (black data points) can be predicted. (b) With the measured storage efficiency for individual gate photons, the data in (a) can be rescaled to show how much contrast is provided by gate excitations with Poissonian number distribution (red). This is again compared to the fundamental limit set by the gate excitation statistics. Block dots show the achievable contrast for deterministic single and two stored gate excitations. Figure adapted from Gorniaczyk [*et al.*]{} [@Hofferberth2014].[]{data-label="fig:Stuttgart_transistor"}](figures/fig_Transistor_contrast.pdf){width="0.7\columnwidth"} In a single photon switch the transmission of a target photon is controlled by a single gate photon. The first Rydberg-based switch was demonstrated by Baur [*et al*]{}. in 2014 [@Duerr2014]. The scheme is similar to the phase gate described in the previous section: the gate photon is converted into a stationary Rydberg excitation in the medium, either via storing of a Rydberg polariton or via direct excitation. The target photon is subsequently sent into the medium on resonant EIT coupled to a different Rydberg state. The gate excitation changes the optical response inside its blockade volume to that of an effective two-level system, see Eq. (\[eq:t2\]), resulting in scattering of the target photon from ground state atoms inside this volume. Since the blockaded optical depth ${{\rm {OD_b}}}$ can be (much) larger than one, the Rydberg-based switch can achieve a very high on/off contrast $C>0.9$ for a single gate excitation (Fig. \[fig:Stuttgart\_transistor\]). . \[fig:Rempe\_transistor\] This concept was then used to demonstrate for the first time single photon transistors by Gorniaczyk [*et al.*]{} [@Hofferberth2014] and Tiarks [*et al.*]{} [@Rempe2014b], where a single gate photon attenuates a stronger source input beam. This transistor performance is quantified by the optical gain $G$, which shows how many photons are removed from the source input by a single gate photon [@Vuletic2013]. Both experiments achieved a gain of $G \sim 15...20$. Key to this achievement is the immense flexibility of the Rydberg interaction [@Cote2005; @Shaffer2006; @Walker2008]. For the transistor experiments in particular, employing Rydberg states with different principal quantum numbers coupled to gate and source photons turned out to be an essential step. Practically, this avoids cross-talk between gate and source photons and excitations, but more importantly it enables tuning of the different interaction strengths involved in the scheme. For an optimal transistor, the interaction between the gate and source Rydberg state should be maximal, while the interaction among source photons is ideally small. One particular interesting feature for tuning the Rydberg interaction are Förster resonances [@Martin2004; @Pillet2006; @Raithel2008; @Entin2010; @Pfau2012; @Weidemueller2013b; @Browaeys2015b], which occur when two dipole-coupled pairs of Rydberg states are resonant with each other. This results in the transition from the van-der-Waals interaction regime $V_{\mathrm{vdW}} \sim C_6/r^6$ to dipolar interaction $V_{\mathrm{dd}} \sim C_3/r^3$. Tiarks [*et al*]{}. showed that the transistor performance improves when choosing gate and source Rydberg states which are close to such a resonance even in zero field (Fig. \[fig:Rempe\_transistor\]). Tuning the Förster defect to exact zero by applying external fields has recently enabled further improvement of the transistor, demonstrating an optical gain $G \sim 200$ [@Hofferberth2015b]. In these experiments, the transistor is operated classically, meaning that the gate photon is lost in the process. Similar to the electronic transistor, this classical device enables the amplification of signals, one example application being the high-fidelity all-optical detection of single Rydberg excitations [@Weidemueller2012; @Lesanovsky2011]. With the optical gain now demonstrated, spatially resolved single-shot imaging of Rydberg excitations embedded in a background gas is feasible. In contrast, retrieving the gate photon after the transistor operation constitutes the first step towards a quantum transistor. Such a quantum device with $G >2$ would enable quantum circuits with gain and feedback or creation of entangled multi-photon states. The finite coherence time of the stored gate spin-wave reduces the possible transistor operation time, but with the recent improvements, a gain $G>2$ could be demonstrated even if the gate photon is retrieved afterwards [@Hofferberth2015b]. At the moment, the fidelity of this coherent transistor is limited because the scattering of source photons results in projection of the stored gate spin wave [@Lesanovsky2015]. Similarly to the phase gates discussed in the last section, the long-range character of the interaction can in principle avoid this problem. For this, the blockade volume of the single gate excitation must exceed the total system size (and ${{\rm {OD_b}}}\gg 1$), an experimentally challenging task. Challenges and outlook ====================== In less than a decade, Rydberg nonlinear optics has created new capabilities that were only dreamed of before. Previously, there were no optical media with a large nonlinearity at the single photon level and now there are. Ultra-cold Rydberg ensembles have been used to create bound states of photons, single photon phase shifts of order $\pi$, and all-optical transistors with gain larger than 100. These are remarkable successes but there remain considerable challenges both in terms of our theoretical understanding and practical applications. Current challenges and open questions ------------------------------------- A Rydberg EIT medium is a complex quantum many-body system where atoms couple collectively to a continuum of photonic modes. Light propagation, interactions and nonlinearity are all coupled. Currently, exact theoretical treatments are only able to describe interactions between two photons. Theoretical developments tend to advance hand in hand with experimental progress which provides a direct validation of approximations. With experiments starting to explore beyond pair-wise interaction, theory also must evolve from effective mean-field descriptions to true many-body treatment of the system [@Buechler2014; @Fleischhauer2015; @Chang2015]. Much of the theoretical challenge in quantum nonlinear optics, as opposed, for example, to cold atoms and condensed-matter systems, stems from the optical nature of the system. It is naturally a nonequilibrium system, which is constantly driven and constantly monitored. It is also naturally dissipative, with the particle number (here, the photon number) not necessarily conserved. Therefore, to describe the evolution of highly-correlated states and nonequilibrium phases in a quantum nonlinear medium, modern methods of quantum field theory are required. These include scattering and input-output formalisms for quantised electromagnetic fields with multiple photons and multiple scatterers, dynamical many-body simulations, and other tools adopted from strongly-correlated condensed-matter theory. In turn, Rydberg-EIT systems of growing size could provide ideal testing grounds offering flexibility encountered in few other systems. On the experimental side, probably the most fundamental challenge encountered so far is the limitation of the atomic density in which Rydberg excitations can be embedded without affecting their coherence or even lifetime [@Pfau2014c; @Duerr2014]. Due to the large size of each Rydberg atom, at high atomic densities ($\gtrsim 10^{13}/$cc) surrounding ground state atoms are not only found inside the blockade volume, but even inside the orbit of the Rydberg electron. While this opens up access to highly interesting new physics such as low-energy electron-atom collisions and Rydberg molecule formation [@Sadeghpour2000; @Pfau2009], in the context of quantum nonlinear optics this imposes a major obstacle for increasing the strength of the effective photon-photon interaction. The increasing rate of Rydberg electron-atom collisions at higher atomic densities has been found to result in shorter coherence times [@Duerr2014] and even reduced Rydberg lifetimes [@Ott2015b]. The key parameter encountered in virtually any application of Rydberg nonlinearities is ${{\rm {OD_b}}}$. Many figures of merit depend strongly on ${{\rm {OD_b}}}$ and only weakly on the total optical depth OD, such as the blockade probability for co-propagating photons $1-{\rm e}^{-{{\rm {OD_b}}}}/\sqrt{\rm{OD}}$ [@Vuletic2012] or their conditional phase shift $\propto\sqrt{\rm{OD}}\cdot{{\rm {OD_b}}}$ [@Vuletic2013b]. Other merits depend solely on ${{\rm {OD_b}}}$, such as the attenuation of ‘signal’ polaritons by a stored ‘gate’ polariton in the photon transistor which equals $1-{\rm e}^{-{{\rm {OD_b}}}}$ [@Hofferberth2014]. The parameters dominating the optical depth per blockade sphere ${{\rm {OD_b}}}$ are the atomic density and the principle quantum number. Since the atomic density is fundamentally limited by the electron-atom collisions, the remaining knob for experimentalists is the principal quantum number. But this raises the same problem of electron-atom collisions, and there are practical issues such as available laser power and ability to compensate stray electric fields that have limited Rydberg-EIT experiments to principle quantum numbers $n \approx 100$. While this number will be pushed further upwards in future experiments, the practically achievable optical depth per blockade region will not grow much beyond ${{\rm {OD_b}}}\sim 20$. Future directions ----------------- To circumvent the limitations on ${{\rm {OD_b}}}$, various attempts are currently in progress to implement more intricate schemes: Including, for example, the idea of enhancing the optical cross section using magic monolayers [@Adams2016]. Another option to further increase the photon-photon interaction is the introduction of an optical cavity around the atomic ensemble. A ground-breaking experiment by Parigi [*et al*]{}. provided a proof of concept for such a system in 2012 [@Grangier2012], recently followed up by the demonstration of long-lived cavity-Rydberg polaritons by Ningyuan [*et al*]{}. [@Simon2015] and detailed study of Rydberg-induced nonlinearities inside a resonator by Boddeda [*et al*]{}. [@Grangier2016]. In parallel, theoretical analyses of this system suggest promising predictions for controlling the quantum statistics of light [@Grangier2014b; @Grangier2015; @Shaffer2016], high-fidelity conditional-phase gates [@Sorensen2016], and the realization of quantum crystals and fractional quantum Hall states of light [@Gorshkov2015; @Simon2015b]. The latter make use of the rich 2D transverse-mode spectra of multimode cavities. While the introduction of an optical resonator makes experimental setups somewhat more complicated again, the new physics offered by these combined systems easily justifies this addition. It seems more than likely that fundamental new steps will emerge from these systems. Another exciting research direction is to transfer the concepts discussed in this review and explored in ultra-cold atomic samples to room-temperature vapour cells. This could result in a great simplification of the experimental setup and pave the way for employing Rydberg-based nonlinearities in future devices. Rydberg interaction effects have already been observed in room-temperature experiments [@Carr2012; @Pfau2013b; @Loew2015], suggesting that the concepts of quantum nonlinear optics with Rydberg atoms can be transferred from the ultra-cold. So far, experimental demonstration of light manipulation on the quantum level is still outstanding, but it seems safe to expect further progress in the future. Based on these open challenges and future steps, the field of Rydberg-based quantum nonlinear optics will keep both theorists and experimentalists busy for quite some time. Certainly we can expect both fundamental breakthroughs, as well as paradigm shifting applications in photonic quantum information processing in the upcoming years. *Acknowledgments*.— We are particularly grateful to Hans Peter Büchler, Stephan Dürr, Michael Fleischhauer, Tom Gallagher, Alexey Gorshkov, Philippe Grangier, Igor Lesanovsky, Misha Lukin, Klaus Molmer, Tilman Pfau, Pierre Pillet, Thomas Pohl, Gerhard Rempe, Vladan Vuletic, and many others for their insightful contributions that helped to shape this field over the last decade. O. F. acknowledges financial support from the Israel Science Foundation, the Laboratory in Memory of Leon and Blacky Broder, the Sir Charles Clore research prize and the MINERVA Stiftung with the funds from the BMBF of the Federal Republic of Germany. C. S. A. acknowledges financial support from EPSRC Grant Ref. No. EP/M014398/1, the EU RIA project ‘RYSQ’ project and DSTL. S. H. is supported by the German Research Foundation through Emmy-Noether-grant HO 4787/1-1 and within the SFB/TRR21, and by RiSC grant 33-7533.-30-10/37/1 from the Ministry of Science, Research and the Arts of Baden-Württemberg. References ==========

--- abstract: 'We investigate two-body Cabibbo-angle enhanced weak decays of charmed baryons into octet baryon and pseudoscalar meson in the current algebra framework with inclusion of the factorization terms which are evaluated using the HQET guided baryonic form factors. We obtain the branching ratios and asymmetry parameters for various Cabibbo-enhanced decays of $\Lambda_c^+$, $\Xi_c^+$ and $\Xi_c^0$ baryons. Sensitivity of the flavor dependence of the spatial wavefunction overlap on the branching ratios and asymmetry parameters is also investigated.' author: - | K.K. Sharma and R. C. Verma$^*$\ Centre for Advanced Study in Physics, Department of Physics,\ Panjab University, Chandigarh -160014, [**India**]{}. title: '**A Study of Weak Mesonic Decays of $\Lambda_c$ and $\Xi_c$ Baryons on the Basis of HQET Results**' --- —————————————————————————— -1 cm -2.5 cm 0.5 cm PACS number(s): 13.30.Eg, , 11.30.Hv, 11.40.Ha, 14.20.Kp\ 1.0 cm $^{*}$ Present Address: Department of Physics, Punjabi University, Patiala - 147 002, India. Introduction ============ A significant progress in the experimental determination of masses, lifetimes of charmed baryons and their decays has taken place during the last few years. Masses of the charm unity baryons have been measured within accuracy of a few percent. Charmed baryons can decay through numerous channels. However, data on their exclusive weak decays are available mainly for $\Lambda_c^+$ baryon \[1\], though, a few decay modes of $\Xi_c^+$ baryon have also been observed \[2\]. Recently, the asymmetry parameters of $\Lambda_c^+ \rightarrow \Lambda \pi^+ $ and $ \Lambda_c^+ \rightarrow \Sigma^+\pi^0$ decays have been measured by the CLEO collaboration \[3\]. In the near future a large quantity of new and more accurate data on the exclusive nonleptonic decays of heavy baryon can be expected which calls for a comprehensive analysis of these decays. Even the meager data available for the charm baryon decays have already started to distinguish between various theoretical models. These models have been developed employing the flavor symmetries \[4\], factorization \[5\], pole model \[6-8\], current algebra \[9,10\] frameworks. So far none of these attempts has been able to explain the available data on the nonleptonic decays of the charmed baryons. The analysis of weak hadronic decays of baryons gets complicated by their being the three quark systems. Further, it not straightforward to estimate the strong interaction effects on their decays. Initially, it was hoped that like meson decays the spectator quark processes would dominate charm baryon decays also. However, this scheme does not seem to be supported by the experiment as the observed branching ratios of $\Lambda_c^+ \rightarrow \Sigma^+ \pi^0$, $\Lambda_c^+ \rightarrow \Sigma^0 \pi^+$, $\Lambda_c^+ \rightarrow \Sigma^+ \eta$ and $\Lambda_c^+ \rightarrow \Xi^0 K^+$ decays, forbidden by the spectator quark process, are significantly large thereby indicating the need of the W-exchange contributions. Unlike the mesons, W-exchange seems to play a dramatic role in the charmed baryon decays, as this mechanism is neither helicity nor color suppressed in baryon decays due to the presence of of a scalar diquark system inside the baryons. Theoretically the contribution from this process has been expected to be proportional to $|\psi(0)|^2$, which renders it quite significant for these decays. For two-body baryon decays, W-emission process leads to the factorization which expresses decay amplitude as coupling of weak baryon transition with the meson current. The matrix elements of the weak transition between baryon states in general involve six form factors which control the factorization contributions \[5\]. Fortunately, in the past few years the discovery of new flavor and spin symmetries has simplified the heavy flavor physics \[11\]. In the framework of Heavy Quark Effective Theory (HQET), the form factors get mutually related, though $1/M$ corrections are certainly needed \[12,13\]. At present one does not know how to carry out these corrections from first principles particularly for heavy to light baryon transitions and one takes the help of phenomenological models. Recently, Cheng and Tseng \[14\] have determined such corrections to the baryonic form factors in the nonrelativistic quark model, which gives excellent agreement with the experimental value for the only measured semileptonic decay $\Lambda_c^+ \rightarrow \Lambda e^+\nu_e$. Similar result has also been obtained by Ivanov et al. in a relativistic three quark model \[15\]. It is worth remarkable here that the agreement has been achieved due to the flavor-suppression factor, resulting from the HQET considerations, for the factorizable contribution. The full implications of this feature for the nonleptonic decays of the charmed baryons is yet to be considered. In the present work, we study Cabibbo-enhanced two-body weak decays of $\Lambda_c^+$, $\Xi_c^+$ and $\Xi_c^0$ into octet baryons $(J^P=1/2^+)$ and a pseudoscalar meson $(J^P=0^-)$. We include the factorization terms using the HQET guided form factors and the nonspectator contributions. Since current algebra is the most common approach used before for the study of the nonleptonic decays, we employ it to obtain the nonspectator contributions. Section-II describes the methodology of the calculations. Section-III deals with the numerical results for branching ratios and asymmetries of the charmed baryon decays and their comparison with the available data. Here, we also study the effect of flavor dependence of the $|\psi(0)|^2$ on these decays. In our analysis, we find that all factorization, pole and equal time current commutator (ETC) terms are equally important in the charm baryon decays, though, one may dominate over other depending upon the decay channel. Methodology =========== Weak Hamiltonian ---------------- The general weak current $\otimes$ current weak Hamiltonian for Cabibbo enhanced ($\Delta C= \Delta S = -1$) decays in terms of the quark fields is $$H_{W}= {G_{F}\over {\sqrt 2}} V_{ud}V_{cs}^{*}[ c_{1} (\bar u d) (\bar s c) + c_{2} (\bar s d) (\bar u c)], \label {(1a)}$$ where $\bar q_{1}q_{2} \equiv \bar q_{1}\gamma_{\mu}(1-\gamma_{5})q_{2}$ represents the color-singlet combination. $V_{ud}$ and $V_{cs}$ are the Cabibbo-Kobayashi-Maskawa (CKM) weak mixing matrix elements. The perturbative QCD coefficients for the charm sector, $c_{1} = {1 \over 2}(c_{+} + c_{-}) = 1.26 \pm 0.04$ and $c_{2} = {1 \over 2}(c_{+} - c_{-}) = -0.51 \pm 0.05$, are usually taken at the charm mass scale \[16\]. Decay Width and Asymmetry Formulas ---------------------------------- The matrix element for the baryon ${1 \over 2}^{+} \rightarrow {1\over 2}^{+} + 0^{-}$decay process is written as $$M = - \langle B_{f} P \vert H_{W}\vert B_{i} \rangle = i \bar u_{B_{f}}( A - \gamma_{5}B) u_{B_{i}} \phi_{P} ,\label {(5)}$$ where A and B represent the parity violating (PV) and the parity conserving (PC) amplitudes respectively. The decay width is computed from $$\Gamma = C_{1} [ \vert A \vert^{2} + C_{2} \vert B \vert ^{2} ], \label {(6)}$$ where $$C_{1} = \frac { \vert {\bf q} \vert} {8 \pi} \frac { (m_{i} + m_{f})^{2} - m_{P}^{2}} {m_{i}^{2}},$$ $$C_{2} = \frac { (m_{i} - m_{f})^{2} - m_{P}^{2}}{(m_{i} + m_{f})^{2} + m_{P}^{2}},$$ and $$\vert {\bf q} \vert = \frac {1}{2m_{i}} {\sqrt {[m_{i}^{2}- (m_{f} - m_{P})^{2}][m_{i}^{2} - (m_{f} + m_{P})^{2}]}},$$ is the magnitude of centre of mass three-momentum in the rest frame of the parent particle. $m_{i}$ and $m_{f}$ are the masses of the initial and final baryons and $m_{P}$ is the mass of the emitted meson. Asymmetry parameter is given by $$\alpha = \frac { 2 Re ( A \bar B^{*})}{ ( |A|^{2} + |\bar B |^{2})}, \label {(7)}$$ where $\bar B = \sqrt {C_{2}} B$. Decay Amplitudes ----------------- Generalising the current algebra (CA) framework of the hyperon decays \[9,10\] to the charm sector, the charmed baryon decay amplitudes receive contributions from the pole diagrams involving the W-exchange process and the ETC term. Including the factorization contributions, the nonleptonic decay amplitude becomes $$< B_f P | H_W | B_i > = M_{fac} + M_{ETC} + M_{pole}. \label {(fep)}$$ We discuss the contribution of each of these terms in the context of PC and PV amplitudes. ### Factorization Contributions The first term $M_{fac}$ in eq(8), corresponding to the factorization contribution, can be obtained by inserting vacuum intermediate states, which express it as a product of two current matrix elements \[5\]; $$<P | A_{\mu} |0> <B_{f}(P_f )| V^{\mu} - A^{\mu} | B_{i}(P_i) > \label {(f1)}$$ where $$<P | A_{\mu} |0> = i f_P q_{\mu} \label {(qm)}$$ with $q_{\mu}$ being the meson four momenta, and $f_P$ is the decay constant of the meson emitted. The matrix element for the baryonic transition $B_i \rightarrow B_f$ can be expanded as $$< B_{f}(P_f )| V_{\mu} | B_{i}(P_i) > = \bar {u}_{f} (P_f) [ f_{1}(q^{2}) \gamma_{\mu} + i f_{2}(q^{2}) \sigma_{\mu \nu} q^{\nu} + f_{3}(q^{2}) q_{\mu} ] u_{i} (P_{i}), \label {(f2)}$$ $$< B_{f}(P_f )| A_{\mu} | B_{i}(P_i) > = \bar {u}_{f}(P_f) [ g_{1}(q^{2}) \gamma_{\mu} + i g_{2}(q^{2}) \sigma_{\mu \nu} q^{\nu} + g_{3}(q^{2}) q_{\mu} ] \gamma_{5} u_{i} (P_{i}), \label {(f3)}$$ where $f_i$’s and $g_i$’s are the vector and axial vector form factors. In the HQET framework, the matrix elements can be parameterised \[11\] in terms of the baryon velocities $v$ and $v'$, $$< B_{f}(v' )| V_{\mu} | B_{i}(v) > = \bar {u}_f [ F_{1}(\omega) \gamma_{\mu} + F_{2}(\omega) v_{\mu} + F_{3}(\omega) v'_{\mu} ] u_{i} , \label {(f4)}$$ $$< B_{f}(v' )| A_{\mu} | B_{i}(v) > = \bar {u}_f [ G_{1}(\omega) \gamma_{\mu} + G_{2}(\omega) v_{\mu} + G_{3}(\omega) v'_{\mu} ] \gamma_{5}u_{i}, \label {(f5)}$$ with $ \omega = v \cdot v'$. The form factors $f_i$’s and $g_i$’s are related to $F_i$’s and $G_i$’s via $$f_1 = F_1 + {1 \over 2 } (m_i + m_f) ( { F_{2} \over m_{i} } + { F_{3} \over m_{f} } ),$$ $$f_2 = {1 \over 2 } ( { F_{2} \over m_{i} } + { F_{3} \over m_{f} } ),$$ $$f_3 = {1 \over 2 } ( { F_{2} \over m_{i} } - { F_{3} \over m_{f} } );$$ $$g_1 = G_1 - {1 \over 2 } ( \Delta m) ( { G_{2} \over m_{i} } + { G_{3} \over m_{f} } ),$$ $$g_2 = {1 \over 2 } ( { G_{2} \over m_{i} } + { G_{3} \over m_{f} } ),$$ $$g_3 = {1 \over 2 } ( { G_{2} \over m_{i} } - { G_{3} \over m_{f} } ), \label {(f6)}$$ where $\Delta m = m_i - m_f $. Employing the nonrelativistic quark model framework, Cheng and Tseng \[14\] have calculated these form factors at maximum $q^2$, $${ f_{1}(q_{m}^{2}) \over N_{fi} } = 1 - { {\Delta m} \over { 2 m_i } } + { { \Delta m } \over { 4 m_i m _f } } (1- { {\bar \Lambda } \over { 2 m_f } }) (m_i + m_f - \eta \Delta m)$$ $$- { { \Delta m } \over { 8 m_i m _f } } { {\bar \Lambda } \over { m_Q } } (m_i + m_f + \eta \Delta m), \label {(f7)}$$ $${ g_{1}(q_{m}^{2}) \over N'_{fi} } = 1 + { {\Delta m {\bar \Lambda}} \over { 4 } } ( { { 1 } \over { m_i m_q } } - { {1} \over { m_f m_Q }} ), \label {(f8)}$$ where $ \eta~=~N'_{fi}/N_{fi}$, ${\bar \Lambda} = m_f - m_q $ and $ q_{m}^{2} = (\Delta m)^2$ denotes the maximum $q^2$ transfer. $N_{fi}$ and $N'_{fi}$ are the flavor factors, $$N_{fi} = _{flavor-spin} <B_f| b_q^{\dagger} b_Q | B_i >_{flavour-spin}, \label {(nfi)}$$ $$N'_{fi} = _{flavour-spin} <B_f| b_q^{\dagger} \sigma_{z}^Q b_Q | B_i >_{flavour-spin}, \label {(nfid)}$$ for the heavy quark $Q$ in the parent baryon $B_i$ transiting into the light quark $q$ in the daughter baryon $B_f$. $m_Q$ and $m_q$ denote masses of these heavy and light quarks respectively. The light diquark present in the parent baryon behaves as spectator. In the absence of a direct evaluation, $q^2$ dependence of the baryonic form factors can be realized by assuming a pole dominance of the form, $$f(q^{2}) = { {f(0)} \over { (1 - { { q^{2}} \over {m_{V}^{2}} } )^{n}} }, \label {(d1)}$$ $$g(q^{2}) = { {g(0)} \over { (1 - { { q^{2}} \over {m_{A}^{2}} } )^{n}} },\label {(d2)}$$ where $m_V$ and $m_A$ denote, respectively, pole masses of the vector meson and axial-vector meson having the quantum numbers of the current involved. Generally for the baryons, one takes $ n = 2$. Upto the first order of parameterization, the factorization amplitudes are given by $$A_{fac} = - { { G_F} \over {\sqrt{2}} } F_C f_P a_k (m_i - m_f) f_{1}^{B_i B_f } (m_{P}^{2}),\label {(fpv)}$$ $$B_{fac} = { { G_F} \over {\sqrt{2}} } F_C f_P a_k (m_i + m_f) g_{1}^{B_i B_f } (m_{P}^{2}). \label {(fpc)}$$ $F_{C}$ is the CKM factor. $a's$ are the two undetermined coefficients assigned to the effective charged current, $a_1,$ and the effective neutral current, $a_2$, parts of the weak Hamiltonian given in eq.(1). Values of these parameters can be related to the QCD coefficients as $$a_{1,2}=c_{1,2}+\zeta c_{2,1},$$ where $\zeta = 1/N_{color}$. The values $$a_1~~=~~1.26,~~~a_2~~=~~-0.51,$$ give the best fit to the experimental data on charm meson decays corresponding to $\zeta \rightarrow 0$ \[16\]. In this approach, the quark currents of weak Hamiltonian are considered as interpolating meson fields generating a $q \bar q$ state. The factorization contributions, being proportional to the meson momenta, can be considered as the correction to the decay amplitudes obtained in the CA framework which employs the soft meson limit. ### ETC and Pole Contributions The second term $M_{ETC}$ in eq(8), corresponding to the equal time current commutator (ETC), is given by the matrix element of $H_W$ between the initial and the final state baryons, $$\langle B_{f} \vert H_{W}\vert B_{i} \rangle = \bar u_{f}(P_f) ( a_{if} - b_{if} \gamma_{5} ) u_{i} (P_i). \label {(me)}$$ It is well known that the PV matrix elements $b_{if}$ vanish for the hyperons due to C-parity null theorem \[9\] in the flavor symmetry limit. In the case of the charm baryon decays, in analogy with hyperons, it has been shown that $b_{if} << a_{if} $. Hence, the ETC term enters only in the s-wave (PV) amplitudes; $$A^{ETC} = { {1} \over {f_{k} }} < B_f | [ Q_{k}^{5} , H_{w}^{PV} ] | B_i > = { {1} \over {f_{k} }} < B_f | [ Q_{k}, H_{w}^{PC} ] | B_i >, \label {(ETC)}$$ where $Q_k$ and $Q_{k}^5 $ denote the vector and axial vector charges respectively. The p - wave (PC) decay amplitudes are then described by the $J^P = 1/2^{+} $ pole terms $(M_{pole})$. The baryon pole terms, arising from s- and u- channels contributions to PC decay amplitude, are given by $$B_{pole} = { { g_{ \ell fk} a_{i \ell} } \over { m_i - m_{\ell} } } { { m_i + m_f } \over { m_{\ell }+ m_{f} } } + { { g_{i \ell 'k} a_{\ell' f} } \over { m_f - m_{\ell '} } } { { m_i + m_f } \over { m_{i}+ m_{\ell '} } }, \label {(pole)}$$ where $g_{ijk}$ are the strong baryon-baryon coupling constants, $\ell$ and $\ell '$ are the intermediate states - corresponding to the respective s- and u-channels. This pole contribution differs from the simple pole model calculations due to the appearance of extra mass factors. This term is actually a modified pole term and contains the contributions from the surface term, the soft-meson Born-term contraction and the baryon-pole term \[9\], combined in a well-defined way. It has been pointed out by Karlsen and Scadron \[10\] that in this way this term accounts for the large momentum dependence away from the soft pion limit as occurs in the charmed baryon decays. The weak matrix elements $a_{ij}$ for baryonic transition $B_i \rightarrow B_f $ are evaluated in the constituent quark model following the work of Riazuddin and Fayyazuddin \[17\]. For the strong baryon-meson coupling constants $ g_{ijk}$, we introduce SU(4) breaking effects \[18\] through $$g_{B B' P} = { { M_B + M_{B'} }\over { 2 M_N } } g_{B B' P}^{sym} \label {(gbb)},$$ where $ g_{B B' P}^{sym} $ denotes the SU(4) symmetric coupling. Numerical Calculations and Discussion of Results ================================================ We first determine the factorizable contributions to the Cabibbo-angle enhanced decays of $\Lambda_c^+$ baryon using the HQET guided form factors, which have been calculated earlier in the nonrelativistic quark model framework \[14\]; $$f_{1}^{ \Lambda_c \Lambda } = 0.50 N_{ \Lambda_c \Lambda }, ~~~g_{1}^{ \Lambda_c \Lambda } = 0.65 N_{ \Lambda_c \Lambda };$$ $$f_{1}^{ \Lambda_c p } = 0.34 N_{ \Lambda_c p }, ~~~g_{1}^{ \Lambda_c p } = 0.53 N_{ \Lambda_c p },$$ where the flavor-spin factors are $$N_{ \Lambda_c \Lambda}' ~=~ N_{ \Lambda_c \Lambda}~~ = ~~{{1} \over {\sqrt 3}};~~~ N_{ \Lambda_c p }' ~~=~~N_{ \Lambda_c p}~~ = ~~{{1} \over {\sqrt 2}}.$$ Reliability of these form factors has been well tested by computing decay width of the semileptonic mode $ \Lambda_c^+ \rightarrow \Lambda e^+ \nu_e $, $$\Gamma (\Lambda_c^+ \rightarrow \Lambda e^+ \nu_e) = 7.1 \times 10^{10} s^{-1}$$ which is consistent with the experimental value \[1\]. It is worth pointing that the flavor factors $N_{\Lambda_c \Lambda } $ plays a crucial role for the agreement. Earlier theoretical models \[19\] have given quite large values for this semileptonic decay rate. The weak Hamiltonian eq(1) allows only $ \Lambda_c^+ \rightarrow p\bar {K^0}/\Lambda\pi^+$ decays to receive the factorization contributions. For these decays, the form factors given in eqs. (35) and (36) yield the following branching ratios and asymmetries: $$Br(\Lambda_c^+\rightarrow p\bar {K^0}) = 0.48\% ~~ ((2.2\pm0.4)\% ~Expt.),$$ $$\alpha(\Lambda_c^+\rightarrow p\bar {K^0}) = - 0.94,$$ $$Br(\Lambda_c^+\rightarrow \Lambda \pi^+) = 1.29\%~~ ((0.79 \pm 0.18)\% ~~Expt.),$$ $$\alpha(\Lambda_c^+\rightarrow \Lambda \pi^+) = - 0.97~~ (-0.94 \pm 0.29 ~~Expt.).$$ Though, the asymmetry of $\Lambda_c^+\rightarrow \Lambda \pi^+$ is in good agreement with experiment, its branching ratio is rather large. In contrast, branching ratio of $\Lambda_c^+\rightarrow p\bar {K^0}$, is much less than the experimental value. Thus, the spectator contributions alone cannot explain even these decays. The branching ratio of $\pi^+$ emitting mode is greater than that of the $\bar {K^0}$ emitting mode due to the color enhancement factor $(a_1/a_2)^2$. However, in a typical $\pi^+$ emitting decay $\Lambda_c^+\rightarrow \Sigma^0 \pi^+,$ the factorization term vanishes due to the Clebsch-Gordon coefficient. It proceeds only through the nonspectator processes, which are also responsible for the remaining $\Lambda_c^+$ decays where the factorization terms do not appear. Accurate experimental measurements of these decays can clearly determine the relative strength of the nonspectator terms in the charmed baryon decays. In our framework, nonspectator ETC and pole terms involve the matrix elements of the kind $\langle B_{f} \vert H_{W}^{PC} \vert B_{i} \rangle$. We evaluate such matrix elements following the scheme of Riazuddin and Fayyazuddin \[17\], which gives the nonrelativistic reduction of the PC-Hamiltonian, $$H_W^{PC} = c_-(m_c)(s^{\dag}c~u^{\dag}d - s^{\dag}{\bf \sigma} c \cdot u^{\dag} {\bf \sigma} d)\delta^3({\bf r}).$$ Note that only $c_-$ appears in this limit, because the part of Hamiltonian corresponding to $c_+$ is symmetric in the color indices. We take the QCD enhancement at the charm mass scale $c_-(m_c) = 1.75,$ which is lower than $c_-(m_s) = 2.23$ used in the hyperon sector. To reduce the number of free parameters, we determine the scale for the ETC and pole terms using $$\langle \psi_{\Lambda} \vert \delta^3(r)\vert \psi_{\Lambda_c^+} \rangle \approx \langle \psi_{p} \vert \delta^3(r)\vert \psi_{\Sigma^+} \rangle.$$ Combining all the ingredients of PV and PC decay amplitudes, we compute the branching ratios and asymmetries for various decays. These are given in the Table 1. Experimentally measured \[1\] masses, lifetimes, and decay constants have been used in the present analysis. Comparing the theoretical values with those obtained in the pure factorization case, we find that inclusion of the nonspectator terms modifies the branching ratios in the desired direction without affecting the asymmetry parameters. We make the following observations: [**1.**]{} The branching fraction for $\Lambda_c^+\rightarrow p\bar{K^0}$ increases from $0.48\%$ to $1.23\%$ bringing it closer to the experiment. The increase in the branching ratio occurs due to constructive interference between the ETC and factorization terms, comparable in magnitude, in the PV mode. Similarly for the PC mode also, the pole and factorization terms interfere constructively, though the pole contribution is around $30\%$ only. We predict its asymmetry $\alpha(\Lambda_c^+\rightarrow p\bar{K^0})$  = - 0.99. [**2.**]{} For the decay $\Lambda_c^+\rightarrow \Lambda \pi^+$, the branching ratio decreases from $1.29\%$ to $1.17\%$ in the right direction. For this decay, the ETC contribution vanishes, so its PV amplitude is given only by the factorization term. For its PC amplitude, there exists a destructive interference between the pole and factorization contributions for the choice of the form factors given in eq. (35). We wish to remark that even if the pole and factorization terms interfere constructively, its branching ratio would hardly be raised to 1.44%. This is due to the reason that the pole terms in s- and u-channels tend to cancel each other thereby reducing the pole strength to around $10\%$ of the factorization. We obtain its asymmetry $\alpha (\Lambda_c^+ \rightarrow \Lambda \pi^+) ~=~~- 0.99$ in nice agreement with the experimental value recently measured by the CLEO collaboration. The CLEO measurement \[3\] has determined the following sets of PV and PC amplitudes (in the units of $~G_F V_{ud} V_{cs}^* \times 10^{-2}~ GeV^2)$: $$Set I:~~~ A(\Lambda_c^+\rightarrow \Lambda \pi^+)~=~ -3.0_{-1.2}^{+0.8}, ~~~ B(\Lambda_c^+\rightarrow \Lambda \pi^+)~=~+12.7_{- 2.5}^{+2.7}~;$$ $$Set II: ~~~ A(\Lambda_c^+\rightarrow \Lambda \pi^+)~=~ -4.3_{-0.9}^{+0.8}, ~~~ B(\Lambda_c^+\rightarrow \Lambda \pi^+)~=~ +8.9_{-2.4}^{+3.4}.$$ Our analysis gives $$A(\Lambda_c^+\rightarrow \Lambda \pi^+)~=~-4.6,~~ B(\Lambda_c^+\rightarrow \Lambda \pi^+)~=~+15.8,~~$$ which seem to favor the first set. As this decay occurs largely through the spectator quark process, the present data seems to demand lower values of the form factors involved, or more accurate measurement is desired to clarify the situation. [**3.**]{} The same CLEO experiment \[3\] has measured the asymmetry of $\Lambda_c^+\rightarrow \Sigma^+ \pi^0$ decay, $$\alpha(\Lambda_c^+\rightarrow \Sigma^+ \pi^0)~=~ -0.45\pm 0.31,$$ which is in good agreement with our prediction, $$\alpha(\Lambda_c^+\rightarrow \Sigma^+ \pi^0)~=~ -0.31.$$ In contrast, earlier theoretical efforts \[6-8\] have given large positive value, ranging from 0.78 to 0.92, for this asymmetry parameter. The calculated branching ratio in our analysis, $$Br(\Lambda_c^+\rightarrow \Sigma^+ \pi^0) = 0.69\% ~~((0.88 \pm 0.22)\% ~~Expt.),$$ also matches well the experimental value. Considering the PV and PC amplitudes explicitly, the measured values are (in the units of $~G_F V_{ud} V_{cs}^* \times 10^{-2}~ GeV^2)$; $$Set I:~~~ A(\Lambda_c^+\rightarrow \Sigma^+ \pi^0)~=~+1.3_{- 1.1}^{+0.9} , ~~~ B(\Lambda_c^+\rightarrow \Sigma^+ \pi^0)~=~-17.3_{- 2.9}^{+2.3};$$ $$Set II:~~~ A(\Lambda_c^+\rightarrow \Sigma^+ \pi^0)~=~ +5.4_{-0.7}^{+0.9}, ~~~ B(\Lambda_c^+\rightarrow \Sigma^+ \pi^0)~=~~ -4.1_{-3.0}^{+3.4}.$$ For these decay amplitudes, we obtain, $$A(\Lambda_c^+\rightarrow \Sigma^+ \pi^0)~=~+5.4;~~ B(\Lambda_c^+\rightarrow \Sigma^+ \pi^0)~=~-2.7,~~$$ consistent with the second set. [**4.**]{} For $\Lambda_c^+\rightarrow \Sigma^0 \pi^+ $ decay, our analysis yields, $$Br(\Lambda_c^+\rightarrow \Sigma^0 \pi^+) = 0.69\%~~ ((0.88 \pm 0.20)\% ~~Expt.),$$ agreeing well with the experiment, and the asymmetry $\alpha(\Lambda_c^+\rightarrow \Sigma^0 \pi^+) = -0.31$, i.e. the same as that of the $\Lambda_c^+\rightarrow \Sigma^+ \pi^0$, as expected from the isospin symmetry arguments. [**5.**]{} For $\eta-\eta'$ emitting decays, we calculate: $$Br(\Lambda_c^+\rightarrow \Sigma^+ \eta) ~=~0.26\% ~~ ((0.48\pm 0.17)\% ~~Expt.),$$ $$\alpha(\Lambda_c^+\rightarrow \Sigma^+ \eta) ~=~- 0.99;$$ $$Br(\Lambda_c^+\rightarrow \Sigma^+ \eta') ~=~0.08\%,$$ $$\alpha(\Lambda_c^+\rightarrow \Sigma^+ \eta') ~=~+0.49;$$ for the $\eta-\eta'$ physical mixing angle $-10^o$. Here, the branching ratio of $\Lambda_c^+\rightarrow \Sigma^+ \eta $ decay is consistent with the observed value. We find that this branching ratio comes closer ($0.29\%$) to the experiment for the physical mixing angle $(-23^o)$ given by the linear mass formulae \[1\]. [**6.**]{} The decay $\Lambda_c^+\rightarrow \Xi^0K^+,$ is theoretically the cleanest of all the $\Lambda_c^+$ decays as it acquires only p-wave contribution to its decay amplitude and has null asymmetry. For this mode, we obtain $$Br(\Lambda_c^+\rightarrow \Xi^0 K^+)~~=~~0.07\%~~((0.34\pm 0.09)\%~~ Expt.),$$ which is smaller than the experimental value. [**7.**]{} Among the $\Xi_c^+$ decays, there are only two possible modes. Recently, the branching ratio of $\Xi^+_c\rightarrow \Xi^0 \pi^+$ decay has also been measured in a CLEO experiment \[2\], for which our analysis yields, $$Br(\Xi_c^+\rightarrow \Xi^0 \pi^+)~~=~~1.08\% ~~((1.2\pm 0.5\pm 0.3)\% ~~Expt.),$$ in excellent agreement with experiment. It may be remarked that though, both the $\Xi_c^+$ modes get contributions from the factorization, pole, and ETC terms, yet the decay $(\Xi^+_c\rightarrow \Xi^0 \pi^+)$ dominates over $(\Xi^+_c\rightarrow \Sigma^+ \bar {K^0}) $ by an order of magnitude $$\frac {Br(\Xi^+_c\rightarrow \Xi^0 \pi^+)} {Br(\Xi^+_c\rightarrow \Sigma^+ \bar {K^0})} ~=~13.2.$$ [**8.**]{} Among $\Xi_c^0$ decays, we find that the dominant mode is $\Xi^0_c\rightarrow \Xi^- \pi^+~$ which has branching ratio around 2% in our model. Variation of $|\psi(0)|^2$ -------------------------- So far, we have taken the scale $|\psi(0)|^2$ for the nonspectator terms same as that of the hyperon sector. However, this being a dimensional quantity, it may be incorrect to ignore its variation with flavor. Unfortunately, evaluation of $|\psi(0)|^2$ is as yet uncertain for the baryons and more complicated, because unlike mesons these are three body systems. However, a naive estimate for the scale may be obtained using the hyperfine splitting, $$\Delta E_{HFS} = { {4\pi\alpha_s } \over { 9m_1m_2 } } \vert \psi(0)\vert^2 \langle{\bf \sigma_1} \cdot {\bf \sigma_2} \rangle,$$ leading to $${{\Sigma_c - \Lambda_c} \over {\Sigma - \Lambda}} ={ {\vert \psi(0)\vert^2_c} \over {\vert \psi(0)\vert^2_s}} {{\alpha_s(m_c)} \over {\alpha_s(m_s)}} {{m_s(m_c-m_u)} \over {m_c(m_s-m_u)}}.$$ For the choice $\alpha_s(m_c)/\alpha_s(m_s) \approx 0.53,$ we obtain, $r ~\equiv ~ \vert \psi(0)\vert^2_c /\vert \psi(0)\vert^2_s \approx 2.1$. However, we do not expect this ratio to hold for the weak decays considered in the present work, as the weak baryon transitions occurring in the charmed baryon decays involve $~ _s< \psi~ | \delta^3({\bf r}) |~ \psi >_c ~$ which should lie between 1 and 2. We have investigated the implications of this scale ratio, varying from 1 to 2, on the branching ratios and asymmetry parameters. We make the following observations: [**1.**]{} Asymmetry of all the decays, except those of $\Xi_c \rightarrow \Sigma + {\bar K}^0$, remain almost unaffected and stay in good agreement with the experiment. Asymmetry of the $\Xi_c \rightarrow \Sigma + {\bar K}^0$ decays show change in sign for scale parameter is increased. [**2.**]{} Branching ratios of $\Lambda_c^+\rightarrow p \bar {K^0}/\Lambda \pi^+/\Xi^0 K^+$ decays are found to require this ratio on the higher side (1.5 to 2.1) for better agreement with the experiment, whereas the $\Lambda_c^+ \rightarrow \Sigma \pi / \eta $ decays prefer a small enhancement ratio (1.1 to 1.3). [**3.**]{} Ratio of the decay rates, $${{ Br(\Lambda_c^+\rightarrow \Lambda\pi^+) } \over { Br(\Lambda_c^+\rightarrow p\bar {K^0}) } } ~= ~(0.92 ~to ~0.40),$$ for the chosen range $(~ r~ = ~1~ to~ 2~)$ approaching the experimental value ${0.36\pm 0.10}$. It may be noticed that this ratio has been theoretically estimated to be as high as 13 in some of the earlier models due to the expected color enhancement. [**4.**]{} The scale ratio certainly increases the branching ratio of $\Lambda_c^+\rightarrow \Xi^0 ~K^+$ as desired by the experiment. Since all the decays $\Lambda_c^+\rightarrow \Sigma^+\pi^0/\Sigma^0\pi^+$, $\Lambda_c^+ \rightarrow \Sigma^+\eta/\Sigma^+\eta'$ and $\Lambda_c^+ \rightarrow \Xi^0K^+$ occur only through the nonspectator terms, their relative ratios remain independent of the scale factor in our analysis. [**5.**]{} We expect the ratio $r$ to lie close to 1.4 using the following ansatz: $$(~ _s< \psi | \delta^3({\bf r}) | \psi >_c~)^2 ~~\approx~~ _s< \psi | \delta^3({\bf r}) | \psi >_s ~ \times ~ _c< \psi | \delta^3({\bf r}) | \psi >_c .$$ To show the trends of the results, we give corresponding values of the branching ratios and asymmetry parameters of the charmed baryon decays in the Table 2. Summary and Conclusions ======================= We have studied the two-body Cabibbo-angle favored decays of the charmed baryons $\Lambda_c^+$, $\Xi_c^+$, and $\Xi_c^0$ into the octet baryons and pseudoscalar mesons. It is now established that factorization alone cannot explain the available data as the branching ratio of $\Lambda_c^+\rightarrow \Sigma \pi/\eta$ and $\Lambda_c^+\rightarrow \Xi^0 K^+$ decays, forbidden in the factorization scheme, have been measured to be comparable to that of $\Lambda _c^+ \rightarrow \Lambda \pi^+$. Hence the nonspectator processes, like W-exchange diagram, seem to play a significant role in understanding these decays. Further, $\Lambda_c^+\rightarrow p {\bar K}^0$ decay which receives color- suppressed factorization has branching ratio greater than that of the color-favored decay $\Lambda _c^+ \rightarrow \Lambda \pi^+$ by a factor of 2.5. In the absence of a direct method for calculating the nonspectator terms, we have employed the standard current algebra framework to estimate their strength. The factorization contributions, being proportional to the meson momentum, provide corrections to this framework. We have evaluated the factorization terms using the HQET guided baryonic form factors. We have obtained branching ratios and asymmetry parameters for these decays, which are found to be consistent with the experimental data. We observe that inclusion of the nonspectator contributions increases $Br(\Lambda_c^+\rightarrow p\bar{K^0})$ from $0.48\%$ to $1.23\%$ and decreases $Br(\Lambda_c^+\rightarrow \Lambda \pi^+)$ from $1.29\%$ to $1.17\%$ in the desired directions. Further, branching ratios of $\Lambda_c^+\rightarrow \Sigma^+ \pi^0/\Sigma^0\pi^+/ \Sigma^+\eta $ and the only measured $\Xi^+_c\rightarrow \Xi^0\pi^+$ decay, obtained in the present analysis, are in good agreement with the experiment. The experimentally available asymmetries of $\Lambda^+_c$ decays are also found in nice agreement with our results. However, branching ratio of $\Lambda_c^+ \rightarrow \Xi^0 K^+$ decay is found to be much less than the observed value. Theoretically, one expect $\Lambda^+_c\rightarrow \Xi^0 K^+$ to be the cleanest of all the modes as neither factorization nor ETC term contributes to this process, so it should provide a measure of the pole terms. We have also investigated the effects, flavor dependence of $|\psi(0)|^2$ as is evident by $\Sigma_c$ and $\Lambda_c$ mass splitting, on these decays. It can result into the desired enhancement of $\Lambda^+_c\rightarrow \Xi^0K^+$ but simultaneously it would increase the decay rates of $\Lambda_c^+\rightarrow \Sigma \pi/ \eta $, as these bear fixed ratios in our model. Though we find that a small scale enhancement is acceptable to the present level of data on charmed baryon decays, it needs some new physics. Small branching ratio for $\Lambda \rightarrow \Xi^0 K^+$ decay in fact results due to near cancellation of the pole terms in the s- and u- channels, which involve antitriplet (C = 1) to octet baryon and sextet (C = 1) to octet baryon transition respectively. We expect that the HQET considerations may differentiate between the two type of the transitions. Further, final state interactions (FSI), well known to substantially alter the decay rates of the charm mesons, may also affect the charm baryon decays by feeding one decay mode into the other. 1.0 cm [**Acknowledgements**]{}\ One of the authors (RCV) gratefully acknowledges the financial support from the Department of Science and Technology, New Delhi, India. [99]{} Particle Data Group, L. Montanet, [*et al.*]{}, Phys. Rev. D [**54**]{}, 55 (1996). CLEO Collaboration: K. W. Edwards, [*et al.*]{}, Phys. Lett. [**373B**]{}, 261 (1996); M. S. Alam, [*et al.*]{}, ‘Measurement of Branching Fractions of $\Lambda_c^+\rightarrow p\bar K n(\pi)$’ CLNS 97/1502, CLEO 97-18, Sept. 10, 1997. CLEO Collaboration: M. Bishai, [*et al.*]{}, Phys. Lett. [**350B**]{}, 256 (1995); R. Ammar, [*et al.*]{}, Phys. Rev. Lett. [**74**]{}, 3534 (1995). M. J. Savage and R. P. Singer, Phys. Rev. D [**42**]{}, 1527 (1990); S. M. Sheikholeslami, M. P. Khanna, and R. C. Verma, Phys. Rev. D [**43**]{}, 170 (1991); Y. Kohara, Phys. Rev. D [**44**]{}, 2799 (1991); M. P. Khanna and R. C. Verma, Phys. Rev. D [**43**]{}, 170 (1991); J. G. Körner and M. Krämmer, Z. Phys. C [**55**]{}, 659 (1992); M. P. Khanna, Phys. Rev. D [**49**]{}, 5921 (1994); R. C. Verma and M. P. Khanna, [*ibid*]{} Phys. Rev. D [**53**]{}, 3723 (1996); K. K. Sharma and R. C. Verma, [*ibid*]{} [**55**]{}, 7067 (1997). J. D. Bjorken, Phys. Rev. D [**40**]{}, 1513 (1989); A. Garcia and R. Huerta, Phys. Rev. D [**45** ]{}, 3266 (1992). H. Y. Cheng and B. Tseng, Phys. Rev. D [**46**]{}, 1042 (1992). H. Y. Cheng and B. Tseng, Phys. Rev. D [**48**]{}, 4188 (1993). P. Zenczykowski, Phys. Rev. D [**50** ]{}, 402 (1994); Q. P. Xu and A. N. Kamal, Phys. Rev. D [**46**]{}, 3836 (1992). R. E. Marshak, Riazuddin, and C. P. Ryan, [*Theory of Weak Interactions in Particle Physics* ]{} (Wiley, New York, 1969); M.D. Scadron, Rep. Prog. Phys. [**44**]{}, 213 (1981) and references therein. S. Pakvasa, S. F. Tuan, and S. P. Rosen, Phys. Rev. D [**42**]{}, 3746 (1990); G. Kaur and M. P. Khanna, Phys. Rev. D [**44**]{}, 182 (1991); R. E. Karlsen and M. D. Scadron, Euro Phys. Lett. [**14** ]{}, 319 (1991); Q. P. Xu and A. N. Kamal, Phys. Rev. D [**46**]{}, 270 (1992); T. Uppal, R. C. Verma, and M. P. Khanna, Phys. Rev. D [**49**]{}, 3417 (1994). N. Isgur and M. B. Wise, Phys. Rev. Letts. [**66**]{}, 1130 (1991); N. Isgur and M. B. Wise, Nucl. Phys. B [**348**]{}, 276 (1991); T. Mannel, [*et al.,*]{} Nucl. Phys. B[**355**]{}, 38 (1991); Tung-Mow Yan, [*et al.,*]{} Phys. Rev. [**46** ]{}, 1148 (1992); N. Isgur and M. B. Wise, “Heavy Quark Symmetry” in "[**B DECAYS**]{}” [*edited by*]{} Sheldon Stone, [**World Scientific**]{} Singapore, Second edition (1994); M. Neubert, Phys. Rep. [**245**]{}, 259 (1994); H. Y. Cheng [*et al.,*]{} Phys. Rev. D [**51**]{}, 1199 (1995); X.-H. Guo and T. Muta, Phys. Rev. D [**54**]{}, 4629 (1996). Robert Singleton, Jr., Phys. Rev. D [**43**]{}, 2939 (1991). X.-H. Guo, P. Kroll, Z. Phys. C [**59**]{}, 567 (1993); Yuan-Ben Dai, [*et al.,*]{} Phys. Rev. D [**51**]{}, 3532 (1995); Xin-Heng Guo, Tao Huang and Zuo-Hong Li, Phys. Rev. D [**53**]{}, 4946 (1996). H. Y. Cheng and B. Tseng, Phys. Rev. D [**53**]{}, 1457(1996). M. A. Ivanov, [*et al.,*]{} Phys. Rev. D [**56**]{}, 348 (1997). M. Wirbel, Prog. Part. Nucl. Phys. [**21**]{}, 33 (1988). Riazuddin and Fayyazuddin, Phys. Rev. D [**18**]{}, 1578 (1978); [*ibid*]{} [**55**]{}, 255 (1997). M. P. Khanna and R. C. Verma, Z. Phys. C [**47**]{}, 275 (1990). R. Perez-Marcial, [*et al.,*]{} Phys. Rev. D [**40**]{}, 2955 (1989); [*ibid*]{} [**44**]{}, 2303(E) (1991). 0.3 cm ------------------------------------------------------------------------------------------------------------------------------------- Decay Br.(%) Expt. Br. (%) $\alpha$ Expt. $\alpha$ -------------------------------------------------- -------------------- ------------------- --------------------- ------------------- $ $1.23$ $2.2\pm 0.4$ $-0.99$ \Lambda_{c}^{+} \rightarrow p \bar K^{0}$ $ \Lambda_{c}^{+}\rightarrow \Lambda \pi^{+}$ $ 1.17$ $0.79 \pm 0.18$ $-0.99$ $ -0.94 \pm 0.24$ $ $0.69$ $0.88 \pm $-0.31 $ $-0.45 \pm 0.31$ \Lambda_{c}^{+}\rightarrow \Sigma^{+} \pi^{0} $ 0.22$ $ \Lambda_{c}^{+}\rightarrow $ 0.26^a~(0.29^b)$ $0.48 \pm 0.17$ $-0.99^a~(-0.91^b)$ \Sigma^{+} \eta $ $ $ 0.08^a~(0.05^b)$ $+0.49^a~(+0.78^b)$ \Lambda_{c}^{+}\rightarrow \Sigma^{+} \eta' $ $\Lambda_{c}^{+}\rightarrow \Sigma^{0} \pi^{+} $ $ 0.69$ $0.88\pm 0.20$ $-0.31$ $ \Lambda_{c}^{+} \rightarrow \Xi^{0} $ 0.07$ $0.34 \pm 0.09$ $ 0.00$ K^{+}$ $ $ 1.08 $ $ 1.2\pm0.5\pm0.3 $-0.74$ \Xi_{c}^{+} \rightarrow \Xi^{0} \pi^{+}$ $ $ \Xi_{c}^{+} \rightarrow \Sigma^{+} \bar K^{0}$ $ 0.08$ $-0.38$ $ \Xi_{c}^{0} \rightarrow $ 0.44$ $-0.80$ \Xi^{0} \pi^{0}$ $ \Xi_{c}^{0} $0.08^a~(0.11^b)$ $+0.01^a~(+0.21^b)$ \rightarrow \Xi^{0} \eta$ $ \Xi_{c}^{0} $0.05^a~(0.03^b)$ $+0.68^a~(+0.80^b)$ \rightarrow \Xi^{0} \eta'$ $ $ 1.99$ $-0.99$ \Xi_{c}^{0} \rightarrow \Xi^{-} \pi^{+}$ $ \Xi_{c}^{0} \rightarrow \Sigma^{+} K^{-}$ $ 0.06$ $0.00 $ $ \Xi_{c}^{0} \rightarrow \Sigma^{0} \bar K^{0}$ $ 0.08$ $-0.15 $ $ \Xi_{c}^{0} \rightarrow \Lambda \bar K^{0}$ $0.34$ $-0.85$ ------------------------------------------------------------------------------------------------------------------------------------- : Branching Ratios and Asymmetries of Charmed Baryon Decays ($~r = 1$) $a~ =~\phi_{\eta-\eta'}~ = ~-10^o,~~~b~ =~\phi_{\eta-\eta'}~ = ~-23^o$ 0.3 cm ------------------------------------------------------------------------------------------------------------------------------------- Decay Br. (%) Expt. Br. (%) $\alpha$ Expt. $\alpha$ -------------------------------------------------- -------------------- ------------------- --------------------- ------------------- $ $1.64$ $2.2\pm 0.4$ $-0.98$ \Lambda_{c}^{+} \rightarrow p \bar K^{0}$ $ \Lambda_{c}^{+}\rightarrow \Lambda \pi^{+}$ $ 1.12$ $0.79 \pm 0.18$ $-0.99$ $ -0.94 \pm 0.24$ $ $1.34$ $0.88 \pm $-0.31 $ $-0.45 \pm 0.31$ \Lambda_{c}^{+}\rightarrow \Sigma^{+} \pi^{0} $ 0.22$ $ \Lambda_{c}^{+}\rightarrow $ 0.50^a~(0.57^b)$ $0.48 \pm 0.17$ $-0.99^a~(-0.91^b)$ \Sigma^{+} \eta $ $ $ 0.15^a~(0.10^b)$ $+0.49^a~(+0.78^b)$ \Lambda_{c}^{+}\rightarrow \Sigma^{+} \eta' $ $\Lambda_{c}^{+}\rightarrow \Sigma^{0} \pi^{+} $ $ 1.34$ $0.88\pm 0.20$ $-0.31$ $ \Lambda_{c}^{+} \rightarrow \Xi^{0} $ 0.13$ $0.34 \pm 0.09$ $ 0.00$ K^{+}$ $ $ 0.53 $ $ 1.2\pm0.5\pm0.3 $-0.27$ \Xi_{c}^{+} \rightarrow \Xi^{0} \pi^{+}$ $ $ \Xi_{c}^{+} \rightarrow \Sigma^{+} \bar K^{0}$ $ 0.04$ $+0.54$ $ \Xi_{c}^{0} \rightarrow $ 0.87$ $-0.80$ \Xi^{0} \pi^{0}$ $ \Xi_{c}^{0} $0.16^a~(0.22^b)$ $+0.01^a~(+0.21^b)$ \rightarrow \Xi^{0} \eta$ $ \Xi_{c}^{0} $0.10^a~(0.06^b)$ $+0.68^a~(+0.80^b)$ \rightarrow \Xi^{0} \eta'$ $ $ 2.46$ $-0.97$ \Xi_{c}^{0} \rightarrow \Xi^{-} \pi^{+}$ $ \Xi_{c}^{0} \rightarrow \Sigma^{+} K^{-}$ $ 0.12$ $0.00 $ $ \Xi_{c}^{0} \rightarrow \Sigma^{0} \bar K^{0}$ $ 0.07$ $+0.48 $ $ \Xi_{c}^{0} \rightarrow \Lambda \bar K^{0}$ $0.54$ $-0.79$ ------------------------------------------------------------------------------------------------------------------------------------- : Branching Ratios and Asymmetries of Charmed Baryons ($~r = 1.4$) $a~ =~\phi_{\eta-\eta'}~ = ~-10^o,~~~b~ =~\phi_{\eta-\eta'}~ = ~-23^o$

--- author: - 'U. Heiter, F. Kupka, C. van ’t Veer-Menneret, C. Barban, W.W. Weiss, M.-J. Goupil, W. Schmidt, D. Katz' - 'R. Garrido' bibliography: - 'models.bib' date: 'Received ; accepted ' title: 'New grids of ATLAS9 atmospheres I: Influence of convection treatments on model structure and on observable quantities' --- Introduction ============ Convective transport of energy in a stellar atmosphere is one of the most complex astrophysical problems. Many of the approximations usually admitted for the stellar interior, such as diffusive radiative transfer, are no longer valid. Moreover, throughout most of a convective stellar atmosphere radiative losses are large enough to make convection less efficient in transporting energy than radiation. Only stars which have a surface convection zone (CZ) extending deep into the stellar envelope can maintain efficient convective energy transfer near the bottom of their atmosphere. On the other hand, inefficient convection appears in all stars near the boundary of a convection zone close to locally stable regions. The modelling of inefficient convection requires a detailed knowledge about the effect of radiative gains and losses on the fluid flow. The situation is particularly complex for stars which are cool enough to develop a granulation pattern, such as the sun. In this case, at identical geometrical depths, vastly different physical conditions may be encountered depending on whether upflow in a granule or downflow in an intergranular lane is considered. The former may be optically thick while the latter is already optically thin, a consequence of the extreme temperature sensitivity of the dominant opacity source in the solar photosphere, the H$^{-}$ ion [cf. also @Stein98]. Currently, only very simple convection models are available for routine computation of extended grids of model atmospheres, while detailed numerical simulations are still unaffordable for applications that require the calculation of many thousands of individual model atmospheres over the HR diagram. Our intention here is first to review the convection models which are available for use together with the popular ATLAS9 model atmosphere code by @kuruczI [@kuruczII] [see also @cast97]. We provide an overview on what is known about the effects of the different convection treatments on model atmosphere structure and consequently on observable quantities. The second purpose of the present paper is to determine to what extent the precision of fundamental parameters derived from the observed stellar spectrum, i.e. , gravity and metallicity depend on the model atmosphere. Another objective is to obtain very accurate colors and more importantly very accurate derivatives of colors, color indices and limb darkening coefficients. These quantities are needed in the procedure of pulsation mode identification which is the first and a crucial step in any seismological study. Indeed probing the stellar interior of a pulsating star requires the knowledge of the resonant cavity within which each mode propagates, i.e. the physical nature of the pulsation mode associated with each observed oscillation frequency. One such procedure is based on the computation of oscillation amplitude ratios and phase differences which in turn depend on the variation of the colors with effective temperature and gravity. The results of this application of the model atmosphere grids will be presented in the next papers of this series [@Bar2002; @Gar2002]. Finally, due to their enhanced resolution the new model grids are also useful to improve the outer boundary conditions of stellar structure calculations [@Mont2001; @DAntona2002]. These goals are part of a program performed in the framework of preparing the COROT space mission (see @COROT). To achieve these purposes, we have used the ATLAS9 code in several versions modified for the convection zone treatment to compute new grids of model atmospheres, corresponding fluxes, surface intensities, $uvby$ colors, synthetic spectra for some representative lines, and compared them with relevant observations. We have three versions of the ATLAS9 code at our disposal: 1. The original version from CDROM13 of Kurucz [@kuruczI] in which the convection zone is treated using mixing length theory (MLT). While ATLAS versions from 5 to 8 remained basically close to the formulations given in @BV58 and in @CG68, some improvements were added in ATLAS9 [cf. @cast97]. In Sect. \[previous\] we discuss the reasons for our specific selection among these improvements. 2. The other two versions were provided by one of the authors (FK) who modified the code to include turbulent convection models from @cm [ CM], and from @cgm [ CGM]. Each convection model has been extensively used in the model atmosphere grid computations which we describe below. All the convection models are of local type and thus require the prescription of a characteristic length scale. Formally, it is possible to interchange the different length scales associated with the convection models. The motivation for doing so and a particular example will be discussed in the next paper of this series [@Kup2002]. This paper is organized as follows. In Sect. \[previous\] we review previous works about the effect of the model structure on theoretical photometric colors and justify the need for new grids of model atmospheres. In Sect. \[convection\] we describe the specific different convection treatments used and discuss their physical content. In Sect. \[grids\] we give details of the grid computations. In Sect. \[effects\] we set out and comment the role of the convection treatments and convection parameters on the model structure, as well as its dependence on effective temperature, surface gravity, and metallicity. Finally, we discuss the consequences on observable quantities such as Balmer line profiles, flux distributions, and colors. A need for new grids {#previous} ==================== The original grids of model atmospheres and colors based on the ATLAS9 code were published by @kuruczI. They were computed using the classical mixing length theory. Kurucz chose and fixed the mixing length parameter $\alpha$, i.e. the ratio $l/H_p$ of convective scale length $l$ and local pressure scale height $H_p$, to be 1.25. He also used a prescription for overshooting at the top of the convection zone (cf. also Sect. \[convection\]) to achieve a better match between computed and observed solar fluxes for the range of $\alpha$ considered. The parameters obtained from the comparison with solar data were used for the entire grids published in @kuruczI. These grids have now been superseded by a new set with a slightly modified prescription of the overshooting treatment [for details see @cast97]. More recently, they have also become available in electronic form [@kuruczII]. @cast97 compared Johnson colors and the $(b-y)$ and $c$ indices from the Strömgren system with colors from grids of model atmospheres based on MLT with and without the overshooting prescription, and with an identical choice for the mixing length. Considering different methods of determining  they concluded that models without the overshooting treatment yield more consistent results, while for the solar case a model with overshooting was favored. As a consequence of this study, new grids of models, fluxes and colors were computed by Castelli without any overshooting for several metallicities and different microturbulent velocities. They are available at the Kurucz website (“NOVER” grids). @cast99 analysed synthetic Johnson UBV colors from these model atmosphere grids, all based on MLT with $\alpha$ = 1.25. She analysed the effect of metallicity and microturbulent velocity and concluded that the indices are affected by both the convection treatment and the amount of line blanketing. This has to be considered in parameter determinations for stars with unknown metallicity. @kunz97 have used the revised version of model atmosphere grids of @kuruczII to provide a new calibration of Geneva photometry for B to G type stars. Comparing their photometrically determined  and  with evolutionary tracks for the Hyades they noticed a systematic trend in  below 7000 K and a rather pronounced “bump” in  located in the same region. Both results were considered to indicate shortcomings in the model atmospheres used for the computations of the synthetic color indices. @smk [ SK] were the first to study the role of different convection treatments implemented in the ATLAS9 code of @kuruczI [@kuruczII] for the synthetic $uvby$ colors. They compared observed color indices with synthetic ones computed using two versions of ATLAS9: the original version of @kuruczI based on MLT treatment of convection, with and without the overshooting option, and another version modified to employ the convection model of @cm [@CM92], known as the CM model and described in Sect. \[convection\]. For the MLT they prove that models built with overshooting at the top of the convection zone, as illustrated in @cast97, are discrepant with the observed color indices. This confirmed similar conclusions drawn by @cvm [ hereafter VM] for the case of Balmer line profiles. SK also showed that the CM models give results generally superior to those obtained with MLT using $\alpha = 1.25$, because they are in better overall agreement with the observed indices $(b-y)_0$ and $c_0$. The metallicity index $m_0$ was found to be the most discrepant one with observations, the CM models remaining in good agreement only for stars with  larger than 7000 K, but clearly discrepant for solar type stars. A peculiar feature in the gravity sensitive $c_0$-index for  around 7000 K was found to be present in colors predicted using any of the convection models investigated, similar to the results found by @kunz97 for MLT model atmospheres for the Geneva photometric system. A similar investigation to the one of SK for the Strömgren photometric system was done later by @schmw, but for the Geneva system. Moreover, he extended it to the CGM convection model which had meanwhile been implemented into the ATLAS9 code (see Sect. \[convection\]). His main conclusion, similar to the one of SK, can be summarized as follows: [*synthetic color indices are more sensitive to the scale length used than to the particular convection model*]{}. For instance, a value of $\alpha = 1.25$ yields differences in the colors in comparison with models where $\alpha = 0.5$ which are much larger than the difference among CM and CGM models as well as MLT models with $\alpha = 0.5$. He concluded that a value of $\alpha = 1.25$ does not allow reproducing the observed photometric colors of late A and F stars. However, discrepancies were also found for the other convection treatments he had studied, in agreement with the results of SK on the $uvby$ colors. @heiter investigated the temperature structure and observed quantities calculated with different convection models for two $\lambda$ Bootis stars with (, ) values of ($-$1, 6800 K) and ($-$2, 7800 K). They found a smaller difference between the synthetic colors and fluxes and the observations when using the CM model or MLT without overshooting compared to MLT with overshooting ($\alpha$ was set to 1.25 for the MLT models). For the cooler one among the two stars, the inclusion of overshooting changed the C, Ti, Cr, and Fe abundances derived from high resolution spectra by +0.1 dex. They also compared the UV fluxes of these stars with IUE and TD1 measurements and found the CM convection model to yield results in best overall agreement while the discrepancies were largest for MLT models with $\alpha = 1.25$ with overshooting. Recently, @Gard99 extended the comparison of SK to the CGM model for the case of Balmer line profiles. It was found that differences between model atmospheres based on the CM or CGM convection treatment, and models based on MLT without overshooting yield rather similar results, while MLT models with overshooting are clearly different. A recommendation for a particular model was found to be possible only for distinct, limited regions in . Their results indicated that a more thorough study of the hydrogen line broadening mechanisms is necessary to draw more reliable conclusions on the convection model, as well as a larger number of standard stars with more accurately known fundamental parameters. For cool dwarf stars such as the sun, one source of problems in matching observed Balmer line profiles with synthetic ones has been to neglect the self-broadening (line broadening due to collisions with neutral hydrogen) in the hydrogen line profile calculations [@Bark2000]. However, this effect is too weak in A and F stars to explain the extent of the discrepancies found in matching the Balmer lines H$_{\alpha}$ and H$_{\beta}$ with some of the model atmospheres for the stars in the above mentioned works. From these previous works we have thus drawn the following considerations for our grid computations. First, the overshooting prescription of ATLAS9 was generally found to be less successful in reproducing observations for A to G type stars, even though for solar observations the case is less settled. Thus, we have decided not to include models computed with this treatment in our grids. However, for comparison we computed individual models with overshooting (always using the correction by @cast96) for our case studies (Figs. \[Ttau\_single\], \[BLP\_single\] and \[irradiance\]). Second, it has been found that model atmospheres which predict temperature gradients closer to the radiative one, i.e. where convection is less efficient than predicted by MLT models with $\alpha > 1$, are in better overall agreement with observations. This was first noticed by @fag93 and, quite independently, for the case of ATLAS9 models by where in order to reproduce the sequence of Balmer line profiles of the sun with the same solar model they had to reduce the value of $\alpha$ of their MLT model atmospheres down to $0.5$. Similar results were found by @fag93, and @Veer:98 for a large range of metallicities and stars of spectral types between A5 and G5 where Balmer lines are both strong and primarily sensitive to the temperature stratification. As shown above, this overall conclusion can also be drawn from other types of measurements such as photometry and is found to hold in particular for A type stars with  larger than 7000 K, while results for stars with lower  were generally more discrepant. Consequently, we have decided to base the majority of our model grid computations on convection treatments which predict less efficient convection than the previous model grids published by @kuruczI [@kuruczII] and @cast99. As far as oscillation mode identification procedures are concerned, it has been demonstrated that the dependency of the colors on  and  is not captured smoothly enough by the standard ATLAS9 models. The effects of the non smooth behavior of the color and limb darkening coefficient derivatives are larger than the expected effect used for identifying the modes [@Gar2000]. In order to obtain smooth variations of these quantities, we have found that it is necessary to compute our model atmospheres with a higher resolution in temperature distribution with depth and built finer grids in  and . Convection Treatment: MLT and FST versions of ATLAS9 {#convection} ==================================================== Mixing length theory (MLT) -------------------------- Model atmospheres computed with ATLAS9 are based on the classical assumptions of stationarity and horizontal homogeneity. With these restrictions only some of the properties of stellar convection can be taken into account. ATLAS9 permits to include: 1. the thermal contributions of convection to the energy flux through the atmosphere; 2. the effect of convective motions on the line opacity due to the additional Doppler broadening of spectral lines caused by turbulent velocity fluctuations on length scales smaller than the mean free optical path. This is achieved by specifying a microturbulent velocity $v_{\rm micro}$ [cf. @gray]. 3. Optionally, ATLAS9 permits to account for changes in pressure stratification due to a turbulent pressure $p_{\rm turb}$. The convective energy flux $F_{\rm conv}$ in the different versions of ATLAS has been computed traditionally with the classical mixing length theory [cf. @Bier48; @Oepik50; @BV58; @cast97]. Classical MLT includes radiative cooling of the fluid which is particularly important where convection is most inefficient, near the boundary of stably stratified layers. It requires the specification of a characteristic scale length $l$ which is prescribed to be a fraction $\alpha$ of the local pressure scale height, $$\label{Hp_formula} H_p = \frac{P}{\rho g} = \frac{l}{\alpha}.$$ $l$ is used to describe the distance which fluid elements can travel before they dissolve. It also specifies the geometrical size of the fluid elements (“bubbles”) together with a second parameter, the ratio of the fluid element volume $V$ over its surface area $A$. The quantity $V/(A l)$ has been changed during upgrades of the ATLAS code [cf. @cast96]. The present choice results in the same convective efficiency as the original one of @BV58 if slightly smaller values of $\alpha$ are used, i.e. the usual choice of $\alpha = 1.25$ in the grids of @kuruczI [@kuruczII] corresponds to an “$\alpha_{\rm BV}$” of about 1.4 for A to G type main sequence stars. A detailed summary of the modifications of MLT as used in the ATLAS code can be found in @cast96, together with various numerical coefficients which we have kept unaltered. One strong motivation to apply a more complete description of stellar turbulent convection stems from the result that low values of the scale length parameter $\alpha$, e.g. 0.5, are required to fit Balmer line profiles for the sun and other cool dwarfs [@fag93 VM], while much larger values (between 1 and 2) are necessary to reproduce their observed radii [@Mor94]. Likewise, the scale length ratio has to be varied over an even larger domain ($1 < \alpha < 3$) to reproduce the red giant branch in HR diagrams of galactic open clusters and associations for stars with masses ranging from 1 $M_{\sun}$ to 20 $M_{\sun}$ [@SC95; @SC97]. Full spectrum turbulence (FST) convection models ------------------------------------------------ An alternative to MLT which can address these problems was introduced by @cm [@CM92] and is referred to as the CM convection model. An improved version was proposed by @cgm which is known as the CGM formulation. A main intention behind both models was to improve the physical description of convection while keeping computational expenses as low as for MLT. Both models achieve this goal by providing a gradient (diffusion) approximation for the convective (enthalpy) flux: $$\label{Fc_formula} F_{\rm conv} = K_{\rm t} \beta = K_{\rm rad} T H_p^{-1} (\nabla-\nabla_{\rm ad})\Phi(S),$$ where $K_{\rm rad} = 4 a c T^3 / (3 \kappa \rho)$ is the radiative conductivity, $\Phi = K_{\rm t} / K_{\rm rad}$ is the ratio of turbulent to radiative conductivity, and $$\beta = -\left(\frac{dT}{dz}-\left(\frac{dT}{dz}\right)_{\rm ad}\right) = T H_p^{-1} \left(\nabla-\nabla_{\rm ad}\right)$$ is the superadiabatic gradient. The convective efficiency $S$ is given by $$\label{S_def} S = {\rm Ra}\cdot{\rm Pr} = \frac{g\alpha_{\rm v}\beta l^4}{\nu\chi}\cdot\frac{\nu}{\chi},$$ ${\rm Ra}$ and ${\rm Pr}$ are Rayleigh and Prandtl numbers of the convective flow, $\alpha_{\rm v}$ is the volume expansion coefficient, and the meaning of the other symbols is standard. We recall here that the thermometric conductivity $\chi$ is related to the radiative conductivity through $K_{\rm rad} = c_p \rho \chi$ and that $\nu$ is the kinematic viscosity. The quantity $S$ is a useful measure of efficiency for flows which feature a very low ${\rm Pr}$ number, as occurs in stellar convection, and for which hence the detailed dependence on $\nu$ can be neglected in parameterisations. This is possible because viscous processes act on much longer timescales than radiation ($t_{\chi} = l^2/\chi$) and buoyancy ($t_{\rm b} = (g\alpha_{\rm v}\beta)^{-1/2}$) which in turn are responsible for the energy balance in stellar atmospheres and envelopes. Thus, the convective efficiency in a star can be characterized using only $(t_{\chi} / t_{\rm b})^2 = S$. The latter can easily be related to an efficiency definition more common in astrophysics [@CG68] that uses the quantity $$\Gamma = \frac{1}{2}\left((1 + \Sigma)^{1/2}-1\right),$$ where $$\Sigma = 4 A^2 (\nabla-\nabla_{\rm ad}) = \frac{2}{81} S,\quad A = \frac{Q^{1/2} c_p \rho^2 \kappa l^2}{12 a c T^3} \sqrt{\frac{g}{2 H_p}},$$ and in which $Q = T V^{-1}(\partial V/\partial T)_P = 1-(\partial \ln \mu/ \partial \ln T)_P$ is the variable average molecular weight. Using this notation the MLT of @BV58 can be viewed as a phenomenologically derived prescription to compute $\Phi$ which reads $$\begin{aligned} \label{PhiMLT} \Phi^{\rm MLT} & = & \frac{9}{8} \Sigma^{-1} \left((1+\Sigma)^{1/2}-1\right)^3 \nonumber\\ & = & \frac{729}{16} S^{-1} \left((1 + \frac{2}{81} S)^{1/2}-1\right)^3,\end{aligned}$$ as mentioned by @cm who pointed out that alternatively the MLT can be understood as a one-eddy approximation made for the spectrum $E(k)$ of turbulent kinetic energy [see also @Lesi:90]. The latter describes how the kinetic energy of the velocity field generated by convection is distributed among different spatial scales $k^{-1}$. @Can96 has shown how MLT underestimates the convective flux in the high efficiency regime ($S \gg 1$) while it overestimates $F_{\rm conv}$ in the low efficiency regime ($S \ll 1$). Both the CM and CGM convection models attempt to overcome the one-eddy approximation by using a turbulence model to compute the full spectrum $E(k)$ of a turbulent convective flow for a given $S$, but keep the assumption of horizontal homogeneity and the Boussinesq approximation used in MLT. Hence, they are also referred to as [*full spectrum turbulence (FST)*]{} convection models. In the case of the CM convection model, the so-called eddy damped quasi-normal Markovian (EDQNM) model [@Ors77] of turbulence is used to compute $\Phi(S)$. This model provides a rather detailed treatment of the nonlinear interactions in a turbulent flow, but requires the specification of a growth rate. The latter was computed from the linear unstable convective modes. To avoid the solution of the equations of the turbulence model each time in a stellar code, the results for $F_{\rm conv}$ were tabulated in a dimensionless form. This was achieved by computing the quantity $\Phi({\rm Ra}, {\rm Pr})$ for a large range of ${\rm Ra}$ and ${\rm Pr}$ numbers. For ${\rm Pr} < 10^{-3}$ the function $\Phi$ was found to saturate. This agrees with the previous remark that $S$ is a useful measure of convective efficiency in a star, where ${\rm Pr}$ is even orders of magnitudes lower, and it was hence sufficient to consider only the results for the lowest ${\rm Pr}$ number for a tabulation of $\Phi(S)$, or actually $\Phi(\Sigma)$, given by the EDQNM model. @cm found that $\Phi(\Sigma)$ can be represented by the following analytical fit formula to an accuracy of better than 3%: $$\label{Phi} \Phi^{\rm CM}(\Sigma) = a_1\Sigma^k \left((1 + a_2 \Sigma)^m - 1\right)^n, \quad\mbox{where}$$ $$\begin{aligned} \label{coeffCM} & & a_1 = 24.868,\, a_2 = 0.097666, \nonumber\\ & & k = 0.14972,\, m = 0.18931,\, n = 1.8503.\end{aligned}$$ The comparison with the CGM model published later is simplified if one considers a change of variable from $\Sigma$ to $S$. In that case $$\label{Phi_CM_S} \Phi^{\rm CM}(S) = b_1 S^k \left(( 1 + b_2 S)^m - 1\right)^n, \quad\mbox{where}$$ $$\begin{aligned} \label{coeffCM_S} & & b_1 = 14.288,\, b_2 = 0.0024115, \nonumber\\ & & \mbox{and $k, m, n$ are the same as above}.\end{aligned}$$ While the asymptotic behavior of both MLT and CM models are equal, i.e. they fulfill the limiting relations $k+m n \simeq 1/2$ and thus $$\label{phi11} \Phi(S) \sim S^{1/2} ~~~~~~~{\rm for} ~~~~~~~S ~~\gg ~~1$$ as well as $k+n \simeq 2$ and hence $$\label{phi12} \Phi(S) \sim S^2 ~~~~~~{\rm for} ~~~~~~ S ~~\ll~~ 1,$$ a distinguishing feature of $\Phi^{\rm CM}(S)$ is to yield about 10 times more flux than (\[PhiMLT\]) for $S \gg 1$, i.e.$$\label{cm_mlt_1} \Phi^{\rm CM}(S) \sim 10 \Phi^{\rm MLT}(S) ~~~~~~~{\rm for} ~~~~~~~S ~~\gg ~~1$$ while $$\label{cm_mlt_2} \Phi^{\rm CM}(S) \sim 0.1 \Phi^{\rm MLT}(S) ~~~~~~~{\rm for} ~~~~~~~S ~~\ll ~~1.$$ The function $\Phi^{\rm CM}$ defined by (\[Phi\])–(\[coeffCM\]) (or (\[Phi\_CM\_S\])–(\[coeffCM\_S\])) is only the first ingredient of the “CM model”. Because $\Phi$ is computed as a function of local variables (\[S\_def\]), it depends on a characteristic length scale which cannot be provided by the formalism itself. Following the physical argument that the Boussinesq approximation leaves no natural unit of length other than the distance to a boundary and that eddies near the boundary of the convection zone are smaller than in the middle of the same (stacking), @cm proposed to take $$\label{CM_length} l = z$$ where $z$ is the distance to the nearest stable layer. The combination of (\[Phi\])–(\[coeffCM\]) and (\[CM\_length\]) has subsequently been called the “CM model”. In this form it was implemented by @Kup96 into ATLAS9 and used for the model grid computations presented here, although other prescriptions of $l$ had been implemented and experimented with as well. In a subsequent paper, @cgm proposed a different FST convection model which avoided the usage of a growth rate. Rather, it was taken into account that the rate of energy input which feeds the velocity fluctuations and thus keeps convection from decaying is controlled by both the source of instability (buoyancy) and by the turbulence it generates. However, the treatment of the nonlinear interactions had to be more simplified to keep the analytical model manageable. The equations of the turbulence model were solved in the limit for low ${\rm Pr}$ numbers. The new self-consistently computed input rate results in an increase of the convective flux for a given efficiency $S$ which is largest at intermediate values of $S \sim 300$. For that reason a more complicated analytical fit formula had to be used to represent the predictions of the turbulence model to an accuracy better than 3% for all values of $S$. The CGM expression for $\Phi$ reads $$\label{PhiCGM} \Phi^{\rm CGM} = F_1(S) F_2(S)$$ where $F_1(S)$ has the same structural form as in the CM model, $$\label{Phi_CGM_F1} F_1(S) = ({\rm Ko}/1.5)^3 a S^k \left(( 1 + b S)^m - 1\right)^n, \quad \mbox{with}$$ $$\begin{aligned} \label{coeff_CGM_F1} & & a = 10.8654,\, b = 0.00489073, \nonumber\\ & & k = 0.149888,\, m = 0.189238,\, n = 1.85011,\end{aligned}$$ while $F_2(S)$ is given by $$\label{Phi_CGM_F2} F_2(S) = 1 + \frac{c S^p}{1+d S^q} + \frac{e S^r}{1+f S^t},\quad\mbox{with}$$ $$\begin{aligned} \label{coeff_CGM_F2} & & c = 0.0108071,\, d = 0.00301208, \nonumber\\ & & e = 0.000334441,\, f = 0.000125, \nonumber\\ & & p = 0.72,\, q = 0.92,\, r = 1.2,\, t=1.5.\end{aligned}$$ Here, ${\rm Ko}$ is the Kolmogorov constant which has been taken 1.7 in all our calculations, a value well inside of the experimental range [@Pras94]. Note that $\Phi^{\rm CGM}$ shows the same asymptotic behavior as $\Phi^{\rm MLT}$ and $\Phi^{\rm CM}$ in the limits of $S \ll 1$ and $S \gg 1$. Moreover, the CM and CGM functions $\Phi$ approach these limits in a very similar manner, because $F_2(S)\rightarrow 1$ for both very large and very small $S$, and the power exponents $k,m,n$ of (\[coeffCM\]), (\[coeffCM\_S\]), and (\[coeff\_CGM\_F1\]) are almost identical. However, while $$\label{cgm_cm} \Phi^{\rm CGM}(S) \sim \Phi^{\rm CM}(S) ~~~~~~~{\rm for} ~~~S ~~\gg ~~1,$$ the low efficiency results differ, as $$\label{cgm_mlt} \Phi^{\rm CGM}(S) \sim 0.3 \Phi^{\rm MLT}(S) ~~~~~~~{\rm for} ~~~S ~~\ll ~~1$$ (cf. (\[cm\_mlt\_2\])). On a logarithmic scale, the low efficiency limit of (\[PhiCGM\]) is almost exactly the average of the fluxes of (\[PhiMLT\]) and (\[Phi\_CM\_S\])–(\[coeffCM\_S\]). The second difference between the two FST convection models is the choice of the scale length $l$ which @cgm have proposed to be $$\label{CGM_length} l = z + \alpha^* H_{p,{\rm top}}.$$ This accounts for the observed fact that convection penetrates into neighboring stable regions and thus the scale length cannot decay to zero right at the layer where the stratification becomes stable according to the Schwarzschild criterion. The additional term in (\[CGM\_length\]) is thus supposed to account for overshooting and provides a possibility for small adjustments, if exact stellar radii are needed, e.g. in helioseismology. However, the meaning of overshooting in this context must not be confused with the overshooting option offered by the ATLAS9 code. This point deserves special attention to which we turn in the following. Length scale parameters and overshooting ---------------------------------------- The term $\alpha^* H_{p,{\rm top}}$ in (\[CGM\_length\]) accounts for the increase of the efficiency of convection due to convective penetration at the boundary between a stably and an unstably stratified region compared to a rigid boundary, for instance a fixed plate. The stellar scenario thus implies to increase the scale length $l$ which can no longer be forced to zero as in (\[CM\_length\]). The total flux within convectively stable layers is still taken equal to the radiative flux. On the other hand, the overshooting prescription included in @kuruczI [@kuruczII] as illustrated in @cast97 was invented to take into account that overshooting directly changes the temperature gradient also in a stable region next to a convection zone. The procedure suggested is to simply smooth out the convective flux over as much as 0.5 $H_p$ in each direction around the last point where $\nabla = \nabla_{\rm rad}$. This mimics the well-known property found in many numerical simulations [e.g. @Hurl86; @Hurl94] and in solutions of the nonlocal Reynolds stress equations [@Kup99; @KM2002] where $F_{\rm conv} > 0$ even though $\nabla-\nabla_{\rm ad} < 0$ in layers right next to a neighboring convection zone. A steeply decaying $F_{\rm conv}$ cannot be modeled this way while the adjacent region where $F_{\rm conv}< 0$ has to be neglected by taking $\nabla = \nabla_{\rm rad}$. The effect of this flux smoothing procedure of ATLAS9 on the emergent flux is large enough to provide an additional degree of freedom to improve the match of solar observations by adjusting the smoothing width. In the CGM model, the parameter $\alpha^*$ of (\[CGM\_length\]) is typically of order 0.1 and may be slightly changed to compensate for uncertainties in opacities and in the treatment of convection. Values of 0.08 and 0.09, similar to @cgm, were used for the different grids presented in Sect. \[grids\]. However, the effect of such small changes is minute. No inconsistencies were found in a recent work by @Mont2001 when model atmospheres computed with $\alpha^*$=0.09 were matched on top of stellar envelopes at different $\tau_{\rm Ross}$, despite a slightly larger value was used in the stellar structure computations to obtain the correct solar radius when using the most recent opacity data. On the other hand, using $\alpha^*$ to compensate for the Boussinesq approximation and various homogeneity assumptions in ATLAS9 by a match to, say, the entropy jump near the stellar surface as found from numerical simulations (cf. @Ludwig99 who used a combination of the CM fluxes (\[Phi\])–(\[coeffCM\]) and the scale length (\[CGM\_length\])) may require larger variations for models very different from the sun. However, such a procedure cannot bring the temperature gradient of ATLAS9 model atmospheres into agreement with the simulations. The latter avoid horizontal homogeneity assumptions but cannot be afforded together with a treatment of frequency dependent radiative transfer which is comparably sophisticated as that one used in ATLAS9. Hence, emergent fluxes, spectra, and photometric colors will be different as well. As long as such a matching procedure is not shown to allow an improved match of fundamental star data over extended parts of the HR diagram (and thus improving over present models, cf. various publications discussed in Sect. \[previous\]), its practical advantages appear more limited. For that reason, we have preferred to use the CGM model as intended by its authors and studied grids with a constant $\alpha^*$ which makes them suitable to be matched with stellar structure calculations using the same treatment of convection [@Mont2001]. Implementation of FST models into ATLAS9 ---------------------------------------- In the ATLAS9 implementation of the CGM convection model the quantities (\[PhiCGM\])–(\[coeff\_CGM\_F2\]) are actually computed as functions of $\Sigma$. Thus, only minimal changes were necessary in the subroutine TCORR, which performs the temperature correction, and in CONVEC, which computes the convective flux, for replacing the CM with the CGM model. TCORR and CONVEC were also the only subroutines that had to be changed for implementing the CM model into ATLAS9. The scale length of the CGM model is evaluated in the following way: $$l = \min (z_{\rm top} + \alpha^* H_{p,{\rm top}}, z_{\rm bottom} + \alpha^* H_{p,{\rm bottom}})$$ This choice makes convection slightly more efficient in comparison with (\[CGM\_length\]) and more consistent with the idea of accounting for overshooting, as the latter is also expected to occur below convection zones. For most model atmospheres we found that the differences between these alternative prescriptions are either zero or negligibly small, because the temperature gradient for convection zones which are entirely contained within the atmosphere is practically radiative while for convection zones extending below the atmosphere the evaluation of $l$ in a pure model atmosphere code necessarily has to occur at the top of the convection zone. We note here that in principle (\[Phi\])–(\[coeffCM\]) and (\[PhiCGM\])–(\[coeff\_CGM\_F2\]) could also be used together with the common scale length $l = \alpha H_p$ with $\alpha < 1$, or other scale lengths. Results on such calculations will be reported in @Kup2002. Turbulent pressure and the optically thin limit ----------------------------------------------- For the CM model, a prescription for the turbulent pressure was published as well, although the results were given only for $S \gg 1$ and in tabular form. In stellar atmospheres, $S \gg 1$ is usually attained only in cool stars and close to the bottom where the Rosseland mean optical depth $\tau_{\rm Ross} > 10$. Hence, the ATLAS9 implementation of the CM model does not account for turbulent pressure. On the other hand, for the CGM convection model analytical fit formulae for $v_{\rm turb}$ and $p_{\rm turb}$ were published by @cgm which can be used even for $S \ll 1$ and were implemented into ATLAS9 as well. A number of model atmospheres for A to early M type dwarfs and for giants were computed with the CGM model with and without the prescription of $p_{\rm turb}$. Differences were found only for stars with deep envelope convection zones, although in most cases both $T$ and $P$ changed by less than 0.1% for $\tau_{\rm Ross} < 5$, and by no more than 0.5% to 1% for $10 < \tau_{\rm Ross} < 100$. As the inclusion of $p_{\rm turb}$ slowed down the convergence of models while spectra and colors remained indistinguishable from the case $p_{\rm turb}=0$, all the CM and CGM model atmospheres grids presented here are computed without a $p_{\rm turb}$, just as their MLT counterparts. We note that for stellar structure calculations the change in temperature structure due to $p_{\rm turb}$ may be more important than for flux predictions derived from ATLAS9 model atmospheres. To avoid discrepancies with the CGM treatment as used in the model grids a reasonable compromise is to match model atmospheres and stellar envelopes at a $\tau_{\rm Ross} \sim 10$. Following a suggestion by Canuto (private communication) the correction of @Spieg57 for radiative losses in optically thin media was implemented for the case of the CM model. However, except for late K and early M dwarfs, where ATLAS9 models are not reliable any more due to the dominance of molecular lines, the effects were found to be negligible. The primary reason for this are the very low values of $F_{\rm conv}$ predicted by the CM model for $\tau_{\rm Ross} < 2$ for stars with  $>4000$ K. For the CGM model, convection is slightly more efficient, but still the effects of such a correction are expected to be very small. Therefore, no further experiments with radiative loss rates were made with FST convection models. The case is different for MLT where the results are more sensitive to the different cooling rates of “optically thin bubbles”, as $$\label{mlt_cgm_cm} \Phi^{\rm MLT}(S) > \Phi^{\rm CGM}(S) > \Phi^{\rm CM}(S) ~~{\rm for} ~~S ~\lesssim ~10$$ because of (\[cm\_mlt\_2\]) and (\[cgm\_mlt\]) and due to the much larger $l$ of MLT for $\alpha > 1$ if $z < H_p$. A correction of $F_{\rm conv}$ for the optically thin limit is always included in the MLT implementation of ATLAS9 [@kuruczI; @cast97]. Model grid computation {#grids} ====================== Two model grids have been computed independently at the Paris and Vienna observatories. At the Paris Observatory an automatic procedure was created by one of the authors (DK). The procedure is interactive, and allows the computation of grids of model atmospheres based on the ATLAS9 code, of Balmer line profiles, surface fluxes and intensities, colors and synthetic spectra, all in one run. The flux and temperature computations are iterated until the following convergence criteria are satisfied: the maximum of the flux and flux derivative errors have to be equal to or less than one and ten percent, respectively. In addition, the maximum of the temperature correction has to be equal to or less than one K. In the MLT case, we started from the original Kurucz grids [@kuruczI; @kuruczII] and recomputed the models by the scaling procedure of the ATLAS9 code. The thickness of the layers of the model atmospheres was divided by 2 or 4 in comparison with the original @kuruczI [@kuruczII] models, in order to solve numerical instabilities in the iteration procedure for the flux computation, and to provide more accurate photometric colors (see Sect. \[flux\_col\] and next paper in this series). Models with higher resolution converged faster and smaller flux errors were achieved. [lllll]{} &\ & Min & Max & Step &\ & 6000 & 8500 & 250 &\ & 2.0 & 4.5 & 0.1 &\ &\ & & & &\  \[\] &\ Convection & MLT & CGM & CM &\ Parameter & 1.25, 0.5 & 0.08 & &\ $\Delta$log $^{\rm a}$ &\ Number of layers &\ & & & &\ \ [lllll]{}\ & Min & Max & Step\ &4000 & 10000 & 200\ & 2.0 & 5.0 & 0.2\ \ \ \ MLT & CGM & CGM & CM\ 0.5 & &\ 0.125 & 0.125 & 0.03125 & 0.03125\ 72 & 72 & 288 & 288\ & & & &\ \ The parameters used for these model grids are given in Table \[grid\_tab\]. We recall that the metallicity is given in terms of the logarithmic ratio between the total number of atoms of each species, except for hydrogen and helium, over the number of hydrogen atoms, with respect to the solar metallicity defined in the same way. For instance, \[M/H\]=0.0 and $-$1.0 means that the opacities entering the model calculations are computed using either solar element abundances or solar element abundances divided by 10 for all elements other than hydrogen and helium. The MLT models were computed for two values of $\alpha$, the original value used by Kurucz $\alpha = 1.25$, and the lower value $\alpha = 0.5$, chosen for reasons given in Sects. \[convection\] and \[effects\]. In the CM and CGM cases, we started from our MLT models with $\alpha = 0.5$, and computed grids with the same set of parameters. For the CGM convection a value of $\alpha^* = 0.08$ was chosen (see Sect. \[convection\] for a discussion). At the Vienna Observatory, model grids with several combinations of convection treatment and vertical resolution were computed for slightly smaller step sizes in , larger step sizes in  and more  values. For MLT models a value of 0.5 has been chosen for $\alpha$. Convection has been turned off for models with $\ge$8600 K, because the convective flux can be neglected for higher temperatures, as can be seen from Fig. \[Fconv\_max\]. As in the Paris grid the uppermost layer is located at log = $-$6.875. The difference of consecutive layers in log is 0.125 and 0.03125 for models with 72 and 288 layers, respectively. In addition to the model atmospheres, fluxes and colors in 12 systems have been computed. Furthermore, information on the convergence extracted from the ATLAS9 output is provided for each model. The atmospheric and computational parameters are summarized in Table \[grid\_tab\]. The grid computations were performed with the perl package SMGT (Stellar Model Grid Tool), described in @schmw[^1]. This non-interactive program runs ATLAS9 repeatedly until the convergence criteria are satisfied for each model. The output of ATLAS9 is evaluated directly and selected information is provided for each model, such as the root mean square (RMS) and maximum values of the flux and flux derivative errors, the maximum of the convective to total flux ratio, the extension of the convection zone, and the optical depth where the temperature equals . The grids defined in Table \[grid\_tab\] are available on CDROM on request from the authors. We note here that two different, but overlapping grids of model atmospheres were computed as there were different applications in mind. The main motivation for the computation of the Paris grids was to calculate photometric colors and their derivatives with respect to  and , which will be used in view of seismic applications [@Wats:88; @Garri90; @Balo99]. This required rather small steps in , but a restricted range for  and few metallicity values. The results of this specific application will be discussed in a subsequent paper of this series. The Vienna grids, on the other hand, are intended for general use, which is the reason for choosing intermediate values for the parameter step sizes and covering as much of the HR diagram as possible. Examples for already published applications of these grids can be found in @Mont2001 [ see below] and in @DAntona2002. Resolution ---------- To show that for specific applications it is necessary to use the models with 288 layers, we examined the quantity $\Delta z = z($$=3.162) - z(F_{\rm conv}=0)$, where $z$ is the depth (distance from top layer) in the atmosphere in km and $z(F_{\rm conv}=0)$ is the depth of the upper limit of the convection zone. This quantity has been used by @Mont2001 for the calculation of the convective scale length in stellar interior models which use convective atmospheres computed with ATLAS9 as a boundary condition. It turned out that for a particular region in the HR diagram, calculating this quantity from atmospheric models with 72 layers results in unphysical oscillations in solar evolutionary tracks which disappear for higher resolutions (J. Montalbán, private communication). Fig. \[length\] shows the values of $\Delta z$ for a small grid of CGM model atmospheres with \[M/H\]=0, =4200…4800 and =3.0…3.6 for four different resolutions, with the stepsize $\Delta$ log  divided by two for each successive resolution value. For 144 layers, the results are rather different from the 72 layer ones (note the peak at (, ) = (4400, 3.2)). There is a small change when increasing to 288 layers, whereas the change when using 576 layers is negligible. This shows that 288 layers are sufficient in low to moderate temperature atmospheric models, in particular as all structural quantities (e.g. the temperature gradient $\nabla$) are resolved. For models with $\ge$ 10000 K, on the other hand, we verified that 72 layers are sufficient. The effects of Convection Treatment {#effects} =================================== Effects on model atmosphere structure ------------------------------------- We first examine changes of temperature and convective flux distribution when using different convection models. Figs. \[Ttau\_single\], \[Ttau\_general\], and \[Fconv\_max\] show the intricate dependence of the effect of convection treatment on , , and  of the model. Fig. \[Ttau\_single\] displays the temperature and the convective flux as a function of Rosseland optical depth (log ) corresponding to the models used for three specific main sequence solar metallicity stars which have been chosen so as to cover the temperature range of interest: the Sun, Procyon – a well studied reference star, and $\beta$ Ari – a well observed hot star. For each star, several models are computed which differ only for the convection treatment. , , metallicity, and microturbulent velocity of the models are taken from previous detailed analyses (by CV for $\beta$ Ari, @Veer:98 for Procyon and @kuruczII for the Sun). The slope of the $T-\tau$ relation within the CZ indicates the efficiency of the convection transport. It is steeper for a less efficient convection, i.e. a temperature gradient closer to the radiative one. For instance, in Fig. \[Ttau\_single\]a, it can be seen that the CM model predicts the least efficient convection, followed with increasing convective efficiency by the MLT ($\alpha=0.5$), the CGM and the MLT ($\alpha=1.25$) models. The same trend is observed for the convective flux in Fig. \[Ttau\_single\]b. This is a consequence of the fact that radiative losses of the convective fluid are always large within the stellar atmosphere where the gas is optically transparent. Hence, the inequality chain (\[mlt\_cgm\_cm\]) always holds at least at lower optical depths ($ \tau \lesssim 1$). The scale lengths (\[Hp\_formula\]), (\[CGM\_length\]), (\[CM\_length\]) of MLT, CGM and CM respectively also obey such an inequality chain for distances $z$ closer to stably stratified layers than $\alpha H_p$ and for the ranges of $\alpha$ and $\alpha^*$ considered in our work. Therefore, the amount of convective flux and the associated $T-\tau$ relations shown in Fig. \[Ttau\_single\] are an immediate consequence of these inequality chains. For hotter stars ($\simeq 8000 $ K), as the convective efficiency decreases, all convection models predict a temperature gradient close to the radiative one. The effect of convection treatment on the atmosphere structure depends on metallicity, gravity and  in a complex way, as illustrated in Fig. \[Ttau\_general\]. Evidently, a metal rich atmosphere reduces the efficiency of the convection transport, as does a low gravity, or a high . The influence of  on the convective efficiency depends strongly on , , and the convection model. For instance, a model at  = 6500 K,  = 2.5, $\alpha_{\rm MLT}\ = 0.5$, and ten times solar metallicity is completely radiative (see Fig. \[Ttau\_general\]a, thin dashed lower line), while for  = 4.5 and identical parameters otherwise a small deviation from radiative stratification is found (thick dashed lower line). This deviation grows significantly when decreasing the metallicity to one tenth of the solar one (Fig. \[Ttau\_general\]b). The variation of the maxima of the convective flux with , , and  is shown in Fig. \[Fconv\_max\] for the CGM models. The decrease of convective flux with increasing metallicity is a consequence of the lower mass density ($\rho$) found in metal rich atmospheres. The latter is a result of the increased opacity, which requires a smaller column density for a given optical depth. Due to increased line blanketing in metal rich atmospheres, the requirements of flux constancy and hydrostatic equilibrium then result in both lower temperature and lower pressure in the outermost layers. As a consequence, lower densities are also found near the boundary of the CZ. This makes convection less efficient, although this effect is partially compensated by a higher convective velocity found for metal rich atmospheres. Fig. \[Fconv\_max\] also shows the influence of using a four times higher resolution in optical depth, resulting in a much smoother run of the curves and in a small shift towards lower temperatures. Consequences on observable quantities ------------------------------------- ### Balmer line profiles The effect of changing the physical parameters entering the models on the Balmer line profiles (BLPs) is very complex. This is illustrated in Fig. \[BLP\_single\], where the synthetic profiles for several different convection models are compared to the observed ones for the same three stars as in Fig. \[Ttau\_single\]. The spectra shown in Fig. \[BLP\_single\] were obtained at the Haute-Provence Observatory, with the spectrograph Aurèlie attached to the 152cm reflector, equipped with a CCD receptor. The resolution is about 25000. The Aurèlie spectra are observed in the first or second order, depending on the wavelength. The wavelength range is 200 Å, and the continuum tracing is local, using the most suitable windows. With a signal to noise ratio of at least 400 we can expect an accuracy for the continuum location of 0.3 %, i.e. 0.5 % for the ratio of line to continuum fluxes in the line wings. This corresponds to a 30 to 60 K change in effective temperature for F to G stars. In the case of H$_{\alpha}$ the effects are never larger than 0.5%. Fig. \[BLP\_single\] shows that the H$_{\alpha}$ profile is insensitive to the choice of any of the scale lengths or convection models discussed in Sect. \[convection\], while in the case of H$_{\beta}$ the profiles computed with MLT and $\alpha = 1.25$ are too narrow. As an example, in the case of Procyon this H$_{\beta}$ profile must be computed with   around 300 K higher to represent the observed profile. The insensitivity of H$_{\alpha}$ to any convection treatment is one of the reasons why it is a very good temperature indicator. However, it is formed close to the boundary of convectively instable layers and therefore can be modified by inclusion of overshooting. Fig. \[BLP\_single\] is a convincing illustration that by the use of the CM or CGM convection treatment, the observed H$_{\alpha}$ and H$_{\beta}$ profiles can be represented by the same atmosphere model, and this constraint can be achieved in the MLT case provided a value for $\alpha$ of about 0.5 is adopted. Thus, we want to emphasize that for the three stars investigated here, [*less efficient convection within ATLAS9 type model atmospheres allows the best fit of H$_{\alpha}$ and H$_{\beta}$ using the same atmosphere model.*]{} In the MLT case, the consequences of these effects on the BLPs have been extensively described by and @fag93, who demonstrated that the BLPs are sensitive probes of the atmosphere structure and of effective temperature for late A, F, and G dwarf stars. Fig. \[BLP\_alpha\]a,b illustrates the effect of changing the MLT parameter $\alpha$ on the BLPs, and to what extent this modification depends on the model parameters. This sensitivity strongly depends on the selected combinations of the model parameters. For instance, the largest differences are seen for models with  between 7000K and 8000K,  = 4.5 and high metallicity. In contrast, the differences are insignificant at low gravity, high metallicity and high temperature. The shape of the profiles is also affected, the most for low temperature and low metallicity models. Fig. \[BLP\_alpha\]c,d shows the difference between two H$_{\beta}$ line profiles, computed with MLT and CM. The difference between the CGM and CM H$_{\beta}$ profiles was not plotted, because it is similar to MLT($\alpha = 0.5$) $-$ CM. The differences with MLT($\alpha = 1.25$) are the largest and strongly depend on temperature and gravity, but less on metallicity. These statements mean that the most efficient convection treatment is MLT with $\alpha = 1.25$, in agreement with the $T-\tau$ laws shown in Figs. \[Ttau\_single\] and \[Ttau\_general\]. From the observer’s point of view, Fig. \[BLP\_alpha\]a-d also reveals that Balmer line profiles have to be measured and normalized to an accuracy of at least 0.5% to draw a clear distinction between convection models with different efficiency. Insufficiently determined profiles may thus easily introduce erroneous trends or large scatter when analyzing their dependence on a particular convection treatment. We stress here that the sensitivity to MLT’s parameter $\alpha$ strongly depends on gravity. For instance, for  $\le$ 7500 K all convection models with low gravity yield inefficient convection. This is due to the fact that low gravity implies lower densities, and the convective efficiency is related to the density as explained in subsection 5.1 above. Moreover, we suggest to consider the [*commonly assumed insensitivity of BLPs to gravity change for  below 8000 K with real caution*]{}. Indeed, Fig. \[BLP\_logg\] shows clearly that the sensitivity of BLPs to gravity changes depends more than usually expected on metallicity, , and finally on all parameters playing a role in the efficiency of the convective transport. It depends also on the gravity itself, the second derivative is not zero. This effect is most important for the highest metallicity and largest . The main reason is that an increase of metallicity leads to a lowering of the density on the one hand while a higher effective temperature favors radiative transfer on the other. ### Fluxes and colors {#flux_col} We have computed fluxes for solar models with different convection treatments as follows: CGM, CM, MLT ($\alpha$=0.5), MLT ($\alpha$=1.25), MLT ($\alpha$=1.25 with overshooting). The same parameters have been chosen for all models:  = 5777 K,  = 4.4377,  = 1.5 , element abundances from @Ande:89, except for Fe, for which the current value of $\log(N_{\rm Fe}/N_{\rm tot}) = -4.51$ was used [@kuruczII]. To compare the calculated solar fluxes to observations, solar irradiance data have been taken from @Neck:84, and in addition from two more recent sources: @Lock:92 [ Lowell Observatory, 1985] and @Thui:98 [ SOLSPEC spectrometer on ATLAS I mission, 1992]. The irradiances in the region of maximum emitted radiation, i.e.  410 to 510 nm, are displayed in Fig. \[irradiance\]. As can be seen, the three observational data sets are different from each other by up to 15 % (upper panel), although @Neck:84 estimate 0.5 % as an upper limit for the local systematic error of their measurements. But they point out that intrinsic intensity variations depending on solar activity can occur when comparing measurements made at different times. These amount to 2 % in certain spectral regions (e.g. the CaII K line) in their data, which are derived from observations made over a 20 yr period [see also @Livi:91]. For comparison, @Lock:92 give an upper limit for the errors of their measurements of 2 % (their observations were made at a phase of low solar activity), and @Thui:98 quote a mean uncertainty of 2-3 % (data obtained at high solar activity) [^2]. Detailed discussions of the error sources and comparisons with previous observations are given in each of the three references. However, the mean of the maximum relative difference between the irradiances from the three sources is 5 % in the region of 450 to 480 nm, which is much larger than the differences between the fluxes calculated with different convection models (2 %, cf. lower panel of Fig. \[irradiance\]). The CM and MLT ($\alpha$=0.5) fluxes are almost identical to the CGM flux. Therefore, the solar irradiance measurements cannot be used to decide between the various models. A similar conclusion would result if measurements of solar central intensity would be used, because the error estimates by @Neck:84 for these measurements (the other two sources did not include this kind of measurements) are equal to that for the irradiance spectra. Thus, we regard tests of central intensity calculations against observations [e.g. @cast97] as having limited significance until accuracy and absolute calibration of these data will have been established with the necessary reliability. The general dependence of the calculated flux on the convection model can be seen in Fig. \[flux\_fig\], where the ratios of MLT and CGM to CM fluxes are displayed for two different values of  and and the extreme case of = $-$1. For all cases, the CGM flux is closest to the CM flux, followed by MLT ($\alpha$=0.5) and with a larger discrepancy by MLT ($\alpha$=1.25). The differences between the models are very small for the highest  and lowest  values. Otherwise they depend strongly on the wavelength range, and no general trend is visible. This is illustrated by Table \[flux\_tab\], which lists the (, ) combinations for three metallicities in order of increasing flux differences from top to bottom. Three different wavelength ranges have been regarded: blue, UV and red. It can be seen that in the latter two, the convective efficiency effect is inversed compared to the one in the blue part. Thus, one can only guess that the calculated emitted flux depends in a complex way on the combination of , , and the convection model. [llllllll]{} blue & & & & UV & red & &\ \ 0.000 & 7500 & 2.5 & & 0.000 & 0.000 & 7500 & 2.5\ 0.024 & 6500 & 4.5 & & 0.011 & 0.005 & 6500 & 4.5\ 0.038 & 6500 & 2.5 & & 0.012 & 0.007 & 6500 & 2.5\ 0.046 & 7500 & 4.5 & & 0.022 & 0.011 & 7500 & 4.5\ \ 0.004 & 7500 & 2.5 & & 0.003 & 0.001 & 7500 & 2.5\ 0.030 & 6500 & 4.5 & & 0.014 & 0.007 & 6500 & 4.5\ 0.043 & 7500 & 4.5 & & 0.022 & 0.008 & 6500 & 2.5\ 0.048 & 6500 & 2.5 & & 0.027 & 0.013 & 7500 & 4.5\ \ 0.009 & 7500 & 2.5 & & 0.006 & 0.003 & 7500 & 2.5\ 0.026 & 6500 & 4.5 & & 0.023 & 0.007 & 6500 & 2.5\ 0.034 & 7500 & 4.5 & & 0.030 & 0.011 & 6500 & 4.5\ 0.048 & 6500 & 2.5 & & 0.034 & 0.015 & 7500 & 4.5\ We use our grids of computed fluxes to derive colors and color indices in the $uvby$ photometric system. The role of convection on this photometric system has already been studied by SK (see Sect. \[previous\]), also using the ATLAS9 code in the MLT and CM cases. Here, we extend this study to the CGM case, and investigate how the variations of color indices due to temperature and gravity variations are affected by the convection formulation. We find that: - there are no measurable differences between colors or indices computed with CM and CGM models and both are very close to those computed with MLT($\alpha$=0.50) models. - differences become much more important if colors or indices computed with MLT($\alpha$=1.25) models are compared to those computed with CM, CGM, or MLT($\alpha$=0.50) models. Thus, we can extend the results of SK (see Sect. \[previous\]) to the CGM model, and conclude that color indices computed with CGM models are generally in better agreement with observations than those computed with MLT($\alpha$=1.25) models. The $(b-y)$ index is the most sensitive one with respect to temperature changes, and this sensitivity is also strongly influenced by the convection model considered. We have investigated the variation of $(b-y)$ indices computed using models differing only by the convection treatment. Fig. \[col\_fig\] shows that the sensitivity of $(b-y)$ to convection change is  and gravity dependent, and the temperature changes associated with the ones of $(b-y)$ are written along the curves. The results are very similar, and in the same order of magnitude, for metallicities ten times or one tenth of the solar one. The same conclusion is reached when the CM model is replaced by the CGM or by MLT($\alpha$=0.50) formulation. From this result we can establish that the “error" (or “change") on temperature variation estimations can be important, i.e. as large as 200 K, when using MLT($\alpha$=1.25) instead of CM convection treatment (or CGM or MLT($\alpha$=0.50)). It can reach up to 400 K, if the overshooting option of ATLAS9 is not removed. Conclusions =========== One of the main conclusions to be raised from this study is that as long as one considers inefficient convection, whatever is the choice of the formulation, either MLT with low $\alpha$, or FST, the interpretation of spectroscopic or photometric observations is equivalent: observed BLPs and Strömgren color indices of dwarf and subgiant stars between A5 and G5 spectral types, and in a large range of metallicity are best represented by the use of less efficient convection transport, i.e. MLT with $\alpha = 0.5$, or with FST formulation. This confirms results already obtained by @fag93, and @Veer:98 for the Sun, Procyon, and other cool metal-poor stars using MLT models. @Gard99 reported a few opposite cases (see Sect. \[previous\]), but for parts of their sample of stars fundamentally known values were not available. An analysis of a larger sample of stars in binary systems with revised fundamental parameters for [*both*]{}  and  [@Smal:02] did not confirm the discrepancies previously found. Furthermore, for the case of F stars @Smal:02 noticed a larger systematic difference between fundamental effective temperatures and those obtained from H$_{\beta}$ lines for MLT($\alpha$=1.25) than for less efficient convection models, although this discrepancy remains within the overall uncertainties. Nevertheless, we have to emphasize that in models with deep convection zones (e.g. for Sun, Procyon) MLT($\alpha = 0.5$) and FST treatments have comparable effects on calculated fluxes, but not on atmosphere structure. They produce different temperature gradients in the deep layers, as can be seen in Figs. \[Ttau\_single\] and \[Ttau\_general\], but those cannot be distinguished by the computed BLPs. In other words, the BLPs allow to discriminate among different values for the MLT parameter $\alpha$, but not among MLT($\alpha = 0.5$), CM, and CGM models. In any case, the sensitivity of BLPs to convection parameters depends significantly on the other physical parameters. This holds especially for the sensitivity to gravity change, which can be more important than usually expected. In case of weakly efficient convection, fluxes and colors depend only weakly on the selected convection treatment. On the other hand, when the convection is highly efficient, then fluxes and colors become strongly dependent on the convection modelling, as the differences among the models show up more clearly within the photosphere. Thus, significant uncertainties on stellar global parameters arise from the convective treatment in model atmospheres. Ignoring these uncertainties can lead to systematic differences affecting subsequent interpretations. The calculations of color and limb darkening partial derivatives are significantly improved when using the present model atmosphere grids which are finer spaced in  and  and have a higher resolution in the temperature distribution with depth [@Bar2002]. Smoothness of these derivatives is of crucial importance in the mode discrimination problem for non-radially pulsating stars, which is basically due to the dependence of the color amplitude ratios on these derivatives. Details of the required precision of these partial derivatives in order to be useful for mode identification will be given in @Gar2002. There we will show that the next space asteroseismology missions – COROT, MONS/Rømer and Eddington – will supply light curves with high enough precision to permit a direct comparison up to the second order to partial derivatives with respect to temperature and gravity as calculated with the present model atmospheres. The improved resolution of the new model grids also avoids unphysical oscillations in evolutionary track calculations when using ATLAS9 model atmospheres as boundary conditions (see Sect. \[resolution\]). Moreover, the possibility to choose among different convection models allows a self-consistent match between model atmospheres and model envelopes [@Mont2001]. However, we must stress here that the different relations $T$ and $F_{\rm conv}/F_{\rm tot}$ vs. depth represent stars which are different in their radii and luminosities. The broad effects of the convective treatment can only be assessed by studying a complete stellar model, i.e. a model with an atmosphere and an internal structure which are consistently built with the same convection formulation. We will address this topic in follow up work [@Kup2002]. The authors would like to thank Gerard Thuillier for providing the SOLSPEC solar irradiance data. Many thanks go to Robert Kurucz for allowing us to use his model atmosphere code and opacity data. We would like to thank the referee, F. Castelli, for helpful comments and the rapid evaluation of the manuscript. This research was carried out within the working group [*Asteroseismology–AMS*]{}, supported by the Fonds zur Förderung der wissenschaftlichen Forschung (project [*P13936-TEC*]{}). Appendix: Description of SMGT {#appendix-description-of-smgt .unnumbered} ============================= The program can be run in either of two modes depending on the temperature structure used for initialization: - In the [*static*]{} mode, an existing model file or a gray atmosphere is used. - In the [*dynamic*]{} mode, an existing model file or a weighted average of all existing models within one grid step of each parameter is used. The weights are calculated in the following way: For each atmospheric parameter $p$, the quantity $$\exp\left(-\frac{|p^i-p^m|}{p_{\rm max}-p_{\rm min}}\right)$$ is computed, where $i$ denotes the initialization model, $m$ the model to be computed, and $p_{\rm max/min}$ the maximum and minimum of the parameter as given in the grid definition. The results for all parameters (at most four) are multiplied to yield the weight. The state of convergence of a particular model is measured by calculating the RMS and the maximum of the flux errors ($\Delta F_{l}$), the flux derivative errors ($\Delta F'_{l}$) and the temperature correction ($\Delta T_{l}$) in all layers $l$. A model is considered as fully converged if these values satisfy certain criteria, which are given in Table \[conver\_crit\] (labeled “primary”). Models for which these criteria cannot be achieved after a reasonable number of iterations are also stored, if they satisfy the criteria labeled “secondary” in Table \[conver\_crit\], but the corresponding files are marked with “$\sim$”. RMS Maximum ------------------------------ ----------- ----------------------------- --------------------- primary $\Delta F_{l}$ $\le$ 1% $\le$ 5% $\Delta F'_{l}$ $\le$ 2% $\le$ 10% or $\le$ 10% ($N/10\le l \le N$) and $\Delta T_{l}$ $\le$ 1K ($0\le l \le N/10$) secondary $\Delta F_{l},\Delta F'_{l}$ $\le$ 10% $\le$ 100% : Convergence criteria used for the Vienna grid computations.[]{data-label="conver_crit"} In order to achieve the convergence criteria without wasting time when no further improvements can be expected from further iterations, the required number of iterations is calculated and checked dynamically, after an initial sequence of 12 iterations. From a sequence of $n$ iterations the ATLAS9 output is processed every $n$/4 iteration (but at least every 15$^{th}$ and at most every 3$^{rd}$ iteration) and the speed of convergence is characterized in the following way. The ratios between the rms errors of two subsequent iterations ($r_F^i, r_{F'}^i$) are calculated. If they are found smaller than a threshold value (0.95), damping exponents are computed iteratively for the flux errors: $$\gamma_F^i = \left[\gamma_F^{i-1} n^{i-1} - a \ln(1-r_F^i)\right]/n^i,$$ where $i$ goes from 2 to the number of processed outputs, $\gamma_F^1$ is set to zero or the value determined from the previous iteration sequence, $a = 1$ for $i = 2$ and $a = (n^i-n^{i-1})/n^i$ for $i > 2$, and an analogous formula is used for the flux derivative errors. The total number of iterations is then given by $$\begin{aligned} n_{\rm tot} & = & {\rm int}\left(\frac{\ln(r_{\rm crit})}{\ln(1 - \exp(-\gamma_F n))}\right) \nonumber\\ {\rm or} \quad n_{\rm tot} & = & {\rm int}\left(\frac{\ln(r_{\rm crit})}{\ln(r_F)}\right)\end{aligned}$$ with $$r_{\rm crit} = \frac{{\rm RMS_{prim}}(\Delta F_{l})}{{\rm RMS}(\Delta F_{l})}$$ where the largest of the values resulting from $\Delta F_{l}$ or $\Delta F'_{l}$ is taken. Apart from the first iteration sequence, the actual number of further iterations is a fraction depending on the ratio between the current predicted number of total iterations and the previous prediction. The computations are terminated if the errors increase (after a few further trials) or if the total number of iterations would be too large (we use a limit of 480). [^1]: A summary is given in the Appendix and directions for the use of this program can be found at http://ams.astro.univie.ac.at/ heiter/smgt\_usage\_1.html. [^2]: Information on the solar activity level has been obtained from the National Solar Observatory Digital Library (http://www.nso.noao.edu/diglib/ftp.html).

--- abstract: 'We apply the general formalism of nilpotent polynomials \[Mandilara [*et al*]{}, [*Phys. Rev. A*]{} [**74**]{}, 022331 (2006)\] to the problem of pure-state multipartite entanglement classification in four qubits. In addition to establishing contact with existing results, we explicitly show how the nilpotent formalism naturally suggests constructions of entanglement measures invariant under the required unitary or invertible class of local operations. A candidate measure of fourpartite entanglement is also suggested, and its behavior numerically tested on random pure states.' author: - Aikaterini Mandilara - Lorenza Viola title: 'Nilpotent polynomials approach to four-qubit entanglement' --- Introduction ============ Characterizing and quantifying multipartite entanglement is a problem whose complexity rapidly increases with the number of particles, and a major challenge within current quantum information science. In spite of intensive effort, a complete understanding of entanglement properties remains limited to date to few-body small-dimensional composite quantum systems: in particular, such understanding has been achieved for pure states of three two-level systems (qubits) [@Dur; @Gingrich], mixed-state entanglement having also been investigated for this system in [@Acin]. Thus, the analysis of pure-state entanglement in an ensemble of four qubits is a critical test for any entanglement theory, as it is provides the first highly non-trivial case whose complexity remains tractable. Different approaches have been attempted so far for unraveling the classification of multipartite entanglement, ranging from so-called hyper-determinants [@Akimasa], to normal forms [@Verstraete], and invariants [@Grassl; @MLV; @Levay2] and covariants [@Briand] of the relevant group of local transformations. Even if such methods offer equivalent answers for ensembles of two or three qubits, a complete description of four-qubit entanglement has only been obtained by Verstraete *et al* [@Verstraete2] based on the method of normal forms, which simplifies considerably in this case thanks to the fact that the group $SO(4,\mathbb{C})$ is isomorphic to $ SL(2,\mathbb{C})\times SL(2,\mathbb{C})$. The resulting classification has been partially independently verified in [@Akimasa]. Closely related with the problem of classification is, in turn, the problem of quantifying entanglement through appropriate measures, as the identification of proper classes should provide the physical boundaries for prospective measures. In addition, the invariants which are often utilized to discriminate among different entanglement classes satisfy themselves the minimum set of requirements that measures are expected to fulfill [@Verstraete2; @Vidal2]. In this work, we tackle the problem of pure-state four-qubit entanglement via a recently introduced approach based on [*nilpotent polynomials*]{} [@nilpotent]. In addition to providing a simple entanglement criterion for any bipartition of a multipartite ensemble, the nilpotent method has the advantage of offering, in principle, a physically motivated procedure for entanglement classification, based on the idea of reducing the nilpotent polynomials to suitable [*canonic*]{} forms, which are [*extensive*]{} with respect to the number of subsystems and [*invariant*]{} under the desired groups of transformations. Such a reduction procedure is considerably facilitated if the dynamical equations of the polynomials are derived and employed. The coefficients of the resulting invariant forms have the same values as polynomial invariants, and may then be used for constructing measures of entanglement. The content of the paper is organized as follows. After recalling in Sec. \[secI\] the basic ingredients of the general nilpotent formalism, we specialize it in Sec. \[secII\] to the four-qubit setting, and derive both general and special entanglement classes for this ensemble. Note that we obtain more entanglement classes than in [@Verstraete2], as a consequence of the fact that we consider at each stage of our reduction procedure transformations that preserve the canonic form of the nilpotent polynomials. In Sec. \[secIII\], the problem of entanglement quantification is discussed in terms of the invariant coefficients of the nilpotent polynomials. Measures for comparing entanglement within classes are proposed, as well as a measure of genuine fourpartite entanglement. Sec. \[secIV\] concludes with a summary of the results, and a discussion of the main advantages and limitations of our approach. Nilpotent polynomials for entanglement description {#secI} ================================================== Consider a pure state $|\Psi\rangle$ describing an ensemble of $n$ qubits. With respect to the computational basis in ${\cal H}\simeq (\mathbb{C}^{2})^{\otimes n}$, $|\Psi\rangle$ may be expressed in the form$$\begin{aligned} \hspace*{-1cm}\left\vert \Psi \right\rangle &=&{\sum }_{\{k_{i}\}=0,1}\psi _{k_{n}k_{n-1}\cdots k_{1}}\left\vert k_{n}k_{n-1}\cdots k_{1}\right\rangle \nonumber \\ \hspace*{-1cm} &=&\psi _{00\cdots 0}\left\vert 00\cdots 0\right\rangle +\psi _{10\cdots 0}\left\vert 10\cdots 0\right\rangle +\ldots +\psi _{11\cdots 1}\left\vert 11\cdots 1\right\rangle , \label{EQ14}\end{aligned}$$where $\psi_{k_{n}k_{n-1}\cdots k_{1}}\in \mathbb{C}$. By introducing pseudospin creation operators $\sigma_{i}^{+}$, the above expression may be rewritten as $$\begin{aligned} \hspace*{-1.5cm}\left\vert \Psi \right\rangle &=&\psi _{00\cdots 0}\left\vert 00\cdots 0\right\rangle +\psi _{10\cdots 0}\sigma _{n}^{+}\left\vert 0\cdots 0\right\rangle +\ldots +\psi _{11\cdots 1}\sigma _{n}^{+}\sigma _{n-1}^{+}\cdots \sigma _{1}^{+}\left\vert 00\cdots 0\right\rangle \\ \hspace*{-1.5cm} &=&\bigg(\sum_{\{k_{i}\}=0,1}\psi _{k_{n}k_{n-1}\cdots k_{1}}\prod_{i=1}^{n}\sigma _{i}^{+}\bigg)\left\vert 00\cdots 0 \right\rangle \:, \end{aligned}$$that is, a polynomial in the nilpotent operator $\left\{ \sigma _{i}^{+}\right\}$ (recall that $(\sigma_{i}^+)^{2}=0$ due to Pauli algebra), acting on the vacuum (or reference) state $\left\vert \mathrm{O}\right\rangle =\left\vert 00\cdots 0\right\rangle $. By setting the population of the latter to be maximal (equal to one), we construct, equivalently, the nilpotent polynomial $F$, $$\hspace*{-1.5cm}F(\{\sigma _{i}^{+}\})=\sum_{\left\{ k_{i}\right\} ={0,1}}\alpha _{k_{n}k_{n-1}\ldots k_{1}}\prod_{i=1}^{n}\left( \sigma _{i}^{+}\right) ^{k_{i}}=\sum_{\{k_{i}\}={0,1}}\frac{\psi _{k_{n}k_{n-1}\ldots k_{1}}}{\psi _{00\ldots 0}}\prod_{i=1}^{n}\left( \sigma _{i}^{+}\right) ^{k_{i}}\ .$$Furthermore, by taking the logarithm of $F$, and by Taylor-expanding around the unit value of the vacuum state population, we obtain the *nilpotential* $f$, $$\hspace*{-1cm}f(\{\sigma _{i}^{+}\})=\ln \left[ F(\{\sigma _{i}^{+}\})\right] =\sum_{\{k_{i}\}={0,1}}\beta _{k_{n}k_{n-1}\ldots k_{1}}\prod_{i=1}^{n}\left( \sigma _{i}^{+}\right) ^{k_{i}}\ .$$The nilpotential makes it possible to readily check whether two subsets $A$ and $B$ of qubits are entangled or not. The following criterion holds [@nilpotent]: **The entanglement criterion:** *The subsets* $A$*and* $B$ * of a binary partition of an assembly of* $n$* qubits are unentangled iff* $$\frac{\partial ^{2}f(\{x_{i}\})}{\partial x_{k}\partial x_{m}}=0\,,\;\;\;\forall k\in A,\:\forall m\in B\:.$$Thus, $A$ and$~B$ are disentangled iff $ f_{A\cup B}=f_{A}(\left\{ x_{\in A}\right\} )+f_{B}(\left\{ x_{\in B}\right\} )$. In spite of the fact that the nilpotential $f$ gives the possibility of applying the above entanglement criterion, $f$ may not yet be regarded as a satisfactory description of entanglement present in the overall composite system, as the latter should naturally be invariant under operations which act [*locally*]{} on individual subsystems only (see [@ge] for a generalization of entanglement [*beyond*]{} the distinguishable subsystem framework we focus on here). The local transformations on each qubit may either be considered to be restricted to unitary transformations in $SU(2)$ – in which case, we talk about $su$-entanglement – or they may be more generally allowed to be any invertible transformation in $SL(2,C)$ – in which case, we talk about $sl$-entanglement. Physically, the latter correspond to the family of [*stochastic local operations assisted by classical communication operations*]{} ([SLOCC]{}) [@Bennett; @Verstraete]. Under the action of local transformations (unitary or merely invertible), the state vector undergoes changes but still remains within a subset $\mathcal{O}$, which coincides with a $su$-orbit (or, respectively, $sl$-orbit) within the overall Hilbert space $\mathcal{H}$. Thus, the nilpotential $f$ should retain the same form for all states belonging to a given orbit, and a [*canonic*]{} form of the resulting nilpotential may accordingly be taken as an “orbit marker”. Canonic forms may be used as an alternative to the method of invariants [@Linden] for identifying different orbits, thereby entanglement classes. The number of independent (real) parameters in a given canonic form should equal the number of independent invariants identifying the orbit, or else equal the dimension of the coset $\mathcal{H}/\mathcal{O}$. According to the general arguments given in [@Carteret2; @nilpotent], the [*$su$-canonic nilpotential*]{} is defined as the nilpotential of the state in the orbit with the maximum reference state population. Under this condition, the orbit-marker is the canonic nilpotential, which we also term the *su-tanglemeter*, $f_{c}$, $$f_{c}(\{\sigma _{i}^{+}\})=\beta _{ij}\sigma _{i}^{+}\sigma _{j}^{+}+\ldots ,\;\;\; \label{EQ13a}$$where the $n$ linear terms are absent and the number of parameters involved equals the dimension of the coset, $D_{su}=2^{n+1}-3n-2$, $n \geq 3$. In order to construct the [*$sl$-canonic nilpotential*]{}, or $sl$-tanglemeter, we begin with the tanglemeter $f_{c}$, and we further reduce the number of parameters down to $D_{sl}\ =\ 2^{n+1}-6n-2$, $n \geq 4$. To achieve this we impose the following conditions: in addition to the requirement for $f_{c}$ that all $n$ terms linear in $\sigma_i^{+}$ be equal to zero, we require that *all $n$ terms of $(n-1)$-th order vanish as well*. Thus, the $sl$-tanglemeter, $f_C$, takes the form $$f_{C}(\{\sigma _{i}^{+}\})=\sum_{\quad \Sigma _{i}k_{i}\not \in \{1,n-1\}}\beta _{k_{n}k_{n-1}\ldots k_{1}}\prod_{i=1}^{n}\left( \sigma _{i}^{+}\right) ^{k_{i}}. \label{EQ14a}$$ Since $D_{sl}<D_{su}$, different $su$-orbits may become equivalent under local $SL$-transformations. For this reason, the classification given by $SL$ is more general than the one given by $SU$, thus usually the term entanglement classes is taken to refer to different $sl$-orbits. Given an arbitrary pure state $|\Psi\rangle$, the task of determining the tanglemeter by applying local operations is, in general, not trivial. The difficulty is substantially reduced if one is able to take advantage of explicit dynamical equation obeyed by the nilpotential of the state, subject to appropriate consistency (or “feedback”) conditions. For qubit systems, the dynamic equation reads $$\mathrm{i}\frac{\partial f}{\partial t}=\mathrm{e}^{-f}H\mathrm{e}^{f}\ , \label{EQ21af}$$where the generators of the local operations$$H=\sum_{i}P_{i}^{-}(t)\sigma _{i}^{+}+P_{i}^{+}(t)\sigma _{i}^{-}+P_{i}^{z}(t)\sigma _{i}^{z}\ , \label{EQ48}$$should be formally substituted as $$\begin{aligned} \sigma _{i}^{+}f &=&\sigma _{i}^{+}f, \nonumber \\ \sigma _{i}^{-}f &=&\frac{\partial f}{\partial \sigma _{i}^{+}}\,, \label{EQ21aa} \\ \sigma _{i}^{z}f &=&-f+2\sigma _{i}^{+}\frac{\partial f}{\partial \sigma _{i}^{+}}\ . \nonumber\end{aligned}$$ For the special case of local [*unitary*]{} operations, $\left( P_{i}^{+}\right) ^{\ast }=\left( P_{i}^{-}\right)$ in Eq. (\[EQ48\]), and the feedback conditions for obtaining $f_{c}$ are $$P_{i}^{-}=\left( P_{i}^{+}\right) ^{\ast }=-\mathrm{i}\beta _{i}\,, \label{EQ60}$$where $$\beta _{i}=\left. \frac{\partial f}{\partial \sigma _{i}^{+}}\right\vert _{\sigma \rightarrow 0}\, \label{EQ58}$$are the coefficients of the linear terms in the nilpotential at a given time. A similar procedure for reducing the nilpotential to the canonic form $f_{C}$ may be carried out also for $SL$-transformations. We begin in this case by reducing $f$ to the tanglemeter $f_{c}$, so that the terms linear in $\sigma _{i}^{+}$ vanish. Next, we apply $SL$ operations as in Eq. (\[EQ48\]), where however $P_{i}^{-}$ and $P_{i}^{+}$ are no longer constrained to be complex conjugates, and choose such operations in such a way that the terms in the nilpotential involving the monomials of order one and of order $n-1$ in $\sigma_i^+$ decrease exponentially with time. The two feedback conditions to be imposed in this case are: ([i]{}) the condition $$P_{j}^{-}=-\sum_{i=1}^{n}\left. P_{i}^{+}\frac{\partial ^{2}f}{\partial \sigma _{i}^{+}\partial \sigma _{j}^{+}}\right\vert _{\sigma \rightarrow 0}=-\sum_{i=1}^{n}P_{i}^{+}\beta _{i,j}\:, \label{EQ62a}$$expressing $P_{i}^{-}$ via $P_{i}^{+}$, which ensures that the nilpotential remains in the form of a tanglemeter at each stage; and ([ii]{}) the condition $$\hspace*{-2cm}\left. \mathrm{i}\frac{\partial ^{n-1}f}{{\textstyle\prod_{i\neq j}}\partial \sigma _{i}^{+}}\right\vert _{\sigma \rightarrow 0}=-P_{j}^{+}\frac{\partial ^{n}f}{{\textstyle\prod_{i}}\partial \sigma _{i}^{+}} +\sum_{m=1}^{n}P_{k}^{+}\left. \frac{\partial ^{n-1}}{{\textstyle \prod_{i\neq m}}\partial \sigma _{i}^{+}}\left[ \sigma _{i}^{+}\left( \frac{\partial f}{\partial \sigma _{i}^{+}}\right) ^{2}\right] \right\vert _{\sigma \rightarrow 0}, \label{EQ62b}$$which ensures the exponential decrease of all $n$ coefficients in front of the second-highest order terms. Unfortunately, no immediate physical meaning seems to be attributable in general to the requirements of vanishing of the $sl$-tanglemeter coefficients of $\left( n-1\right) $-th order – in contrast to the case of $SU$ transformations, where vanishing of the first-order terms reflects maximum ground state population. Mathematically, however, such a requirement is suggested by symmetry considerations: $n$ complex conditions are imposed on $n$ complex coefficients of the same type. After having eliminated the monomials of orders $1$ and $(n-1)$, it is possible to specify the scaling parameters $P_{i}^{z}$ so that $n$ additional conditions are imposed on the tanglemeter coefficients. For example, we can set to unity the coefficients in front of the highest order term, and adjust $(n-1)$ coefficients in front of certain monomials to be equal to $(n-1)$ coefficients of other monomials. The condition in Eq. (\[EQ62b\]) for $P_{j}^{+}$ is written implicitly as a set of $n$ linear equations that can be solved for [*generic*]{} states. However, no solution exists for those $P_{j}^{+}$ parameters corresponding to a zero determinant. Such singularities may correspond to [*special*]{} classes of entangled states which require separate consideration – as we are going to see explicitly in the four-qubit example. $sl$-tanglemeters for four qubits {#secII} ================================== A generic normalized pure state of four qubits may be described by $2\cdot 2^{4}-2=30$ real parameters. The $su$-entanglement of this state requires less parameters to be characterized, $D_{su}=30-3\cdot 4=18$ and, according to the discussion in Sec. \[secI\], for four qubits the $su$-tanglemeter defined in Eq. (\[EQ13a\]) reads $$\begin{aligned} \hspace{-1.8cm}f_{c} &=&\beta _{3}\sigma _{2}^{+}\sigma _{1}^{+}+\beta _{5}\sigma _{3}^{+}\sigma _{1}^{+}+\beta _{9}\sigma _{4}^{+}\sigma _{1}^{+}+\beta _{6}\sigma _{3}^{+}\sigma _{2}^{+}+\beta _{10}\sigma _{4}^{+}\sigma _{2}^{+}+\beta _{12}\sigma _{4}^{+}\sigma _{3}^{+} \label{EQ14.4} \\ \hspace{-1.8cm}&+&\beta _{7}\sigma _{3}^{+}\sigma _{2}^{+}\sigma _{1}^{+}+\beta _{13}\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{1}^{+}+\beta _{11}\sigma _{4}^{+}\sigma _{2}^{+}\sigma _{1}^{+}+\beta _{14}\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{2}^{+}+\beta _{15}\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{2}^{+}\sigma _{1}^{+}\,. \nonumber\end{aligned}$$In the above expression, we have used the local phase operations that did not contribute to the elimination of the linear coefficients to make the trilinear coefficients $\beta_{7},\beta_{13},\beta_{11},\beta_{14}$ real numbers. In addition, a compact notation has been introduced by considering the indexes of $\beta $ as a binary representation of decimal numbers, e.g., $0011\mapsto 3$, *etc*. Allowing for more general local transformations on each qubit, such as indirect measurements with stochastic outcomes, the number of the parameters necessary to describe a state may be further reduced. The $sl$-tanglemeter (\[EQ14a\]) of a generic state of four qubits contains $D_{sl}=30-6\cdot 4=6$ real parameters, and may be cast in the following form: $$\begin{aligned} \hspace{-1.5cm}f_{C} &=&\beta _{3}\left( \sigma _{1}^{+}\sigma _{2}^{+}+\sigma _{3}^{+}\sigma _{4}^{+}\right) +\beta _{5}\left( \sigma _{1}^{+}\sigma _{3}^{+}+\sigma _{2}^{+}\sigma _{4}^{+}\right) +\beta _{6}\left( \sigma _{1}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{3}^{+}\right) \nonumber \\ \hspace{-1.5cm}&&+(1-\beta _{3}^{2}-\beta _{5}^{2}-\beta _{6}^{2})\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}, \label{EQ15}\end{aligned}$$where the scaling factors (that is, the parameter in front of $\sigma_{i}^{z}$ in Eq. (\[EQ48\])), have been chosen so that the $f_{C}$ becomes equivalent to the expression $G_{abcd}$ in *Theorem 2* of [@Verstraete]. We proceed to explicitly illustrate the procedure for evaluating the $sl$-tanglemeter in Eq. (\[EQ15\]) by means of the dynamic equations (\[EQ21af\])-(\[EQ48\]), starting from the $su$-tanglemeter given in Eq. (\[EQ14.4\]). First, one may notice that in the system of eleven first-order nonlinear differential equations for the coefficients $\beta _{i}$, the coupling of the second-order terms $\beta _{ij}\sigma _{i}^{+}\sigma _{j}^{+}$ to the fourth-order term $ \beta _{15}\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{2}^{+}\sigma _{1}^{+}$ occurs via the third-order terms $\beta _{7}\sigma _{3}^{+}\sigma _{2}^{+}\sigma _{1}^{+}$, $\beta _{13}\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{1}^{+}$, $\beta _{11}\sigma _{4}^{+}\sigma _{2}^{+}\sigma _{2}^{+}$, $ \beta _{14}\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}$. Thus, the time evolution of all $\beta _{i}$ stops when these third-order coefficients $\beta _{7}$, $\beta _{13}$, $\beta $, and $\beta _{14}$ vanish – indicating that for four qubits the $sl$-tanglemeter is a stationary solution for the dynamic equations. If the coefficients $P_{i}^{-}$ satisfy the requirement of Eq. (\[EQ62a\]), which ensures that the [*nilpotential always remains in the form of a valid $su$-tanglemeter $f_{c}$ during such evolution*]{}, what it is left is to adjust the time dependence of the parameters $P_{1}^{+}$, $P_{2}^{+}$, $P_{3}^{+}$ and $P_{4}^{+}$ so that they drive all four third-order coefficients to zero. From the differential equations of the third-order coefficients,$$\begin{aligned} \mathrm{i}\dot{{\beta }}_{14} &=&\!-P_{1}^{+}\beta _{15}+2P_{2}^{+}\beta _{6}\beta _{10}+2P_{3}^{+}\beta _{6}\beta _{12}+2P_{4}^{+}\beta _{10}\beta _{12}, \nonumber \\ \mathrm{i}\dot{{\beta }}_{13} &=&2P_{1}^{+}\beta _{5}\beta _{9}-P_{2}^{+}\beta _{15}+2P_{3}^{+}\beta _{5}\beta _{12}+2P_{4}^{+}\beta _{9}\beta _{12}, \nonumber \\ \mathrm{i}\dot{{\beta }}_{11} &=&2P_{1}^{+}\beta _{3}\beta _{9}+2P_{2}^{+}\beta _{3}\beta _{10}-P_{3}^{+}\beta _{15}+2P_{4}^{+}\beta _{9}\beta _{10}, \nonumber \\ \mathrm{i}\dot{{\beta }}_{7} &=&2P_{1}^{+}\beta _{3}\beta _{5}+2P_{2}^{+}\beta _{3}\beta _{6}+2P_{3}^{+}\beta _{5}\beta _{6}-P_{4}^{+}\beta _{15}\,, \label{EQ62d}\end{aligned}$$we see that, in the general case, feedback conditions may be imposed by a proper choice of the parameters $P_{i}^{+}$, in such a way that these equations take the form $$\dot{{\beta }}_{7}=-\beta _{7};\quad \dot{{\beta }}_{11}=-\beta _{11};\quad \dot{{\beta }}_{13}=-\beta _{13};\quad \dot{{\beta }}_{14}=-\beta _{14}\,. \label{EQ62da}$$The evolution implied by these equations brings, in turn, the nilpotential to the following form: $$\begin{aligned} \hspace{-1mm}f&=&\beta _{3}\sigma _{2}^{+}\sigma _{1}^{+}+\beta _{5}\sigma _{3}^{+}\sigma _{1}^{+}+\beta _{9}\sigma _{4}^{+}\sigma _{1}^{+}+\beta _{6}\sigma _{3}^{+}\sigma _{2}^{+} \nonumber \\ \hspace{-1mm}& + &\beta _{10}\sigma _{4}^{+}\sigma _{2}^{+}+\beta _{12}\sigma _{4}^{+}\sigma _{3}^{+}+\beta _{15}\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{2}^{+}\sigma _{1}^{+}\:. \label{EQ14b}\end{aligned}$$We can invoke the four scaling operators $e^{B_{i}\sigma _{i}^{z}}$, and further reduce Eq. (\[EQ14b\]) to the $sl$-canonic form $f_{C}$ of Eq. (\[EQ15\]), unless one or more of the above $\beta $ coefficients vanish. Such cases correspond to zero-measure manifolds – in other words, to special classes of entanglement. For example, when $\beta _{3}=0$ in  (\[EQ14b\]), the tanglemeter may be cast, by scaling, in the form $$\begin{aligned} f_{C}^{(2)} &=&\sigma _{3}^{+}\sigma _{4}^{+}+\beta _{5}(\sigma _{1}^{+}\sigma _{3}^{+}+\sigma _{2}^{+}\sigma _{4}^{+}) \label{EQ16a} \\ &&+\beta _{6}(\sigma _{1}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{3}^{+})+(1-\beta _{5}^{2}-\beta _{6}^{2})\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}~, \nonumber\end{aligned}$$characterized by only two parameters. If $\beta _{3}=\beta _{10}=0$, the $sl$-tanglemeter reads $$\hspace*{-1cm}f_{C}^{(1)}=\sigma _{3}^{+}\sigma _{4}^{+}+\sigma _{1}^{+}\sigma _{3}^{+}+\beta _{6}(\sigma _{1}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{3}^{+})+(1-\beta _{6}^{2})\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}~, \label{EQ16b}$$which only involves a single parameter. Lastly, if $\beta_{3}=\beta _{10}=\beta _{9}=0$, $$f_{C}^{(0)}=\sigma _{3}^{+}\sigma _{4}^{+}+\sigma _{1}^{+}\sigma _{3}^{+}+\sigma _{2}^{+}\sigma _{3}^{+}+\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}~. \label{EQ16c}$$Note that the tanglemeters of Eqs. (\[EQ16a\]), (\[EQ16b\]), and (\[EQ16c\]) correspond to the special families $L_{abc_{2}}$, $L_{a_{2}b_{2}}$ and $ L_{a_{2}0_{3\oplus 1}}$ of the classification given in *Theorem 2* of [@Verstraete]. However, it is important to bear in mind that the latter classification applies to [*un*]{}-normalized states, whereas our tanglemeter corresponds to states of unit population in the reference state. When the fourth-order coefficient $\beta _{15}=0$ and, additionally, one or more of the quadratic coefficients are also zero, singular classes of states without genuine fourpartite entanglement emerge: for instance, the $sl$-tanglemeter of a four-qubit $W$ state, $$f_{C}=\sigma _{3}^{+}\sigma _{4}^{+}+\sigma _{1}^{+}\sigma _{3}^{+}+\sigma _{2}^{+}\sigma _{3}^{+},$$belongs to one of such classes, and separable states with tanglemeters of the type $$f_{C}=\sigma _{3}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{3}^{+}+\sigma _{2}^{+}\sigma _{4}^{+}~,$$and similar, belong to other. On the other hand, reducing $\ f_{c}$ to the canonic $sl$-form $f_{C}$ cannot be achieved when the determinant $${\cal D}_4=\left\vert \begin{array}{cccc} -\beta _{15} & 2\beta _{6}\beta _{10} & 2\beta _{6}\beta _{12} & 2\beta _{10}\beta _{12} \\ 2\beta _{5}\beta _{9} & -\beta _{15} & 2\beta _{5}\beta _{12} & 2\beta _{9}\beta _{12} \\ 2\beta _{3}\beta _{9} & 2\beta _{3}\beta _{10} & -\beta _{15} & 2\beta _{9}\beta _{10} \\ 2\beta _{3}\beta _{5} & 2\beta _{3}\beta _{6} & 2\beta _{5}\beta _{6} & -\beta _{15}\end{array}\right\vert \label{EQ62e}$$of the system of differential equations (\[EQ62d\]) vanishes – which makes it impossible to impose any required feedback conditions. In such a situation, we loose the functional independence of the right hand sides of (\[EQ62d\]), which ensures complete controllability of the dynamics of $\beta _{7}$, $\beta _{13}$, $\beta _{11}$, and $\beta _{14}$ in the generic case. In turn, this means that some linear combinations of these coefficients, determined by the system’s eigenvectors, cannot be set to zero by any choice of $P_{i}^{+}$, and a tanglemeter $f_{C}$ of a special form should be defined in such instances. In [@nilpotent], four special families of tanglemeters are derived,$$\begin{aligned} f_{C}^{(s1)} &=&\beta _{3}\left( \sigma _{2}^{+}\sigma _{1}^{+}+\sigma _{4}^{+}\sigma _{3}^{+}\right) +\beta _{5}\left( \sigma _{3}^{+}\sigma _{1}^{+}+\sigma _{4}^{+}\sigma _{2}^{+}\right) \nonumber \\ &&+\beta _{6}\left( \sigma _{4}^{+}\sigma _{1}^{+}+\sigma _{3}^{+}\sigma _{2}^{+}\right) \nonumber \\ &&+\sigma _{3}^{+}\sigma _{2}^{+}\sigma _{1}^{+}-\sigma _{4}^{+}\sigma _{2}^{+}\sigma _{1}^{+}+\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{1}^{+}-\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{2}^{+} \nonumber \\ &&+2\left( \beta _{5}\beta _{6}-\beta _{3}\beta _{6}+\beta _{3}\beta _{5}\right) \sigma _{4}^{+}\sigma _{3}^{+}\sigma _{2}^{+}\sigma _{1}^{+}\,, \label{EQ14.4a}\end{aligned}$$ $$\begin{aligned} f_{C}^{(s2)} &=&\sigma _{2}^{+}\sigma _{1}^{+}+\sigma _{3}^{+}\sigma _{1}^{+}+\sigma _{4}^{+}\sigma _{2}^{+}+\sigma _{4}^{+}\sigma _{3}^{+} \nonumber \\ &&+\beta _{6}\left( \sigma _{4}^{+}\sigma _{1}^{+}+\sigma _{3}^{+}\sigma _{2}^{+}\right) +\beta _{7}\left( \sigma _{3}^{+}\sigma _{2}^{+}\sigma _{1}^{+}-\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{2}^{+}\right) \nonumber \\ &&+\beta _{11}\left( \sigma _{4}^{+}\sigma _{2}^{+}\sigma _{1}^{+}-\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{1}^{+}\right) +2\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{2}^{+}\sigma _{1}^{+}\,. \label{EQ14.4b}\end{aligned}$$$$\begin{aligned} f_{C}^{(s3)} &=&\sigma _{2}^{+}\sigma _{1}^{+}+\sigma _{3}^{+}\sigma _{1}^{+}+\sigma _{4}^{+}\sigma _{2}^{+}+\sigma _{4}^{+}\sigma _{3}^{+}+\sigma _{4}^{+}\sigma _{1}^{+}+\sigma _{3}^{+}\sigma _{2}^{+} \nonumber \\ &&+\beta _{14}\left( \sigma _{3}^{+}\sigma _{2}^{+}\sigma _{1}^{+}-\sigma _{4}^{+}\sigma _{2}^{+}\sigma _{1}^{+}+\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{1}^{+}-\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{2}^{+}\right) \nonumber \\ &&+\beta _{13}\left( \sigma _{3}^{+}\sigma _{2}^{+}\sigma _{1}^{+}+\sigma _{4}^{+}\sigma _{2}^{+}\sigma _{1}^{+}-\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{1}^{+}-\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{2}^{+}\right) \nonumber \\ &&+\beta _{11}\left( \sigma _{3}^{+}\sigma _{2}^{+}\sigma _{1}^{+}-\sigma _{4}^{+}\sigma _{2}^{+}\sigma _{1}^{+}-\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{1}^{+}+\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{2}^{+}\right) \nonumber \\ &&+2\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{2}^{+}\sigma _{1}^{+}\,. \label{EQ14.44c}\end{aligned}$$$$\begin{aligned} f_{C}^{(s4)} &=&\sigma _{3}^{+}\sigma _{2}^{+}\sigma _{1}^{+}+\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{1}^{+}+\sigma _{4}^{+}\sigma _{2}^{+}\sigma _{1}^{+}+\sigma _{4}^{+}\sigma _{3}^{+}\sigma _{2}^{+} \nonumber \\ &&+\beta _{3}\sigma _{2}^{+}\sigma _{1}^{+}+\beta _{5}\sigma _{3}^{+}\sigma _{1}^{+}+\beta _{6}\sigma _{3}^{+}\sigma _{2}^{+}, \label{EQ14.44d}\end{aligned}$$corresponding, respectively, to one, two, three, or four of the eigenvalues $\gamma_j$ of ${\cal D}_4$ vanishing, where explicitly $$\begin{aligned} \gamma_{1} &=&\beta _{15}-2\sqrt{\beta _{5}\beta _{6}\beta _{9}\beta _{10}}+2\sqrt{\beta _{3}\beta _{6}\beta _{9}\beta _{12}}-2\sqrt{\beta _{3}\beta _{5}\beta _{10}\beta _{12}}\,, \nonumber \\ \gamma_{2} &=&\beta _{15}+2\sqrt{\beta _{5}\beta _{6}\beta _{9}\beta _{10}}-2\sqrt{\beta _{3}\beta _{6}\beta _{9}\beta _{12}}-2\sqrt{\beta _{3}\beta _{5}\beta _{10}\beta _{12}}\,, \nonumber \\ \gamma_{3} &=&\beta _{15}-2\sqrt{\beta _{5}\beta _{6}\beta _{9}\beta _{10}}-2\sqrt{\beta _{3}\beta _{6}\beta _{9}\beta _{12}}+2\sqrt{\beta _{3}\beta _{5}\beta _{10}\beta _{12}}\,, \nonumber \\ \gamma_{4} &=&\beta _{15}+2\sqrt{\beta _{5}\beta _{6}\beta _{9}\beta _{10}}+2\sqrt{\beta _{3}\beta _{6}\beta _{9}\beta _{12}}+2\sqrt{\beta _{3}\beta _{5}\beta _{10}\beta _{12}}\,. \label{EQ62.1}\end{aligned}$$Observe that the number of parameters in such special tanglemeters is $3$ complex numbers, the same as in the general case given in Eq. (\[EQ15\]). At present, it remains to be proved whether all four special tanglemeters (\[EQ14.4a\])-(\[EQ14.44d\]) correspond to distinct special entanglement classes, since they result from considering a dynamic evolution based on a series of sequential [*infinitesimal*]{} local operations which preserve the $su$-canonic form of the nilpotential. Thus, a situation where some of the obtained tanglemeters turn out to be equivalent under a [*finite*]{} local $sl$-transformation, cannot be ruled out in principle by our current approach. In [@Verstraete], such special classes are not explicitly identified, although the last three classes of *Theorem 2* – that is, $L_{0_{7\oplus \bar{1}}}$ , $L_{0_{3\oplus \bar{1}}0_{3\oplus \bar{1}}}$, and $L_{0_{5\oplus \bar{3}}}$, may be easily identified as special cases of Eq. (\[EQ14.44d\]) when one or more terms vanish. The class $L_{ab_{3}}$ in [@Verstraete]  is not identified by our method. We summarize in Table 1 the entanglement classes for pure states of four qubits we have thus obtained. \[tables\] --------------------------------------------------------------------------------------------------------- *General class* 3 complex parameters -------------------------- ------------------------------------------------------------------------------ $G_{a}$ $f=\beta _{3}(\sigma _{1}^{+}\sigma _{2}^{+}+\sigma _{3}^{+}\sigma _{4}^{+})+\beta _{5}(\sigma _{1}^{+}\sigma _{3}^{+}+\sigma _{2}^{+}\sigma _{4}^{+})$ $+\beta _{6}(\sigma _{1}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{3}^{+})+(1-\beta _{3}^{2}-\beta _{5}^{2}-\beta _{6}^{2})\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}$ *Singular $3D$ classes* 3 complex parameters $G_{b}$ $f=\beta _{3}\left( \sigma _{1}^{+}\sigma _{2}^{+}+\sigma _{3}^{+}\sigma _{4}^{+}\right) +\beta _{5}\left( \sigma _{1}^{+}\sigma _{3}^{+}+\sigma _{2}^{+}\sigma _{4}^{+}\right) +\beta _{6}\left( \sigma _{1}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{3}^{+}\right) $ $+\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}-\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{4}^{+}+\sigma _{1}^{+}\sigma _{3}^{+}\sigma _{4}^{+}-\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}$ $+2\left( \beta _{5}\beta _{6}-\beta _{3}\beta _{6}+\beta _{3}\beta _{5}\right) \sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}$ $G_{c}$ $f=\sigma _{1}^{+}\sigma _{2}^{+}+\sigma _{1}^{+}\sigma _{3}^{+}+\sigma _{2}^{+}\sigma _{4}^{+}+\sigma _{3}^{+}\sigma _{4}^{+}+\beta _{6}\left( \sigma _{1}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{3}^{+}\right) $ $+\beta _{7}\left( \sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}-\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}\right) +\beta _{11}\left( \sigma _{1}^{+}\sigma _{2}^{+}\sigma _{4}^{+}-\sigma _{1}^{+}\sigma _{3}^{+}\sigma _{4}^{+}\right) $ $+2\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}$ $G_{d}$ $f=\sigma _{1}^{+}\sigma _{2}^{+}+\sigma _{1}^{+}\sigma _{3}^{+}+\sigma _{2}^{+}\sigma _{4}^{+}+\sigma _{3}^{+}\sigma _{4}^{+}+\sigma _{1}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{3}^{+}$ $+\beta _{14}\left( \sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}-\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{4}^{+}+\sigma _{1}^{+}\sigma _{3}^{+}\sigma _{4}^{+}-\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}\right) $ $+\beta _{13}\left( \sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}+\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{4}^{+}-\sigma _{1}^{+}\sigma _{3}^{+}\sigma _{4}^{+}-\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}\right) $ $+\beta _{11}\left( \sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}-\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{4}^{+}-\sigma _{1}^{+}\sigma _{3}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}\right) $ $+2\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}$ $G_{e}$ $f=\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}+\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{4}^{+}+\sigma _{1}^{+}\sigma _{3}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}+$ $\beta _{3}\sigma _{1}^{+}\sigma _{2}^{+}+\beta _{6}\sigma _{2}^{+}\sigma _{3}^{+}+\beta _{5}\sigma _{2}^{+}\sigma _{3}^{+}$ *Singular $2D$ classes* 2 complex parameters $LG2_{a}$ $f=\sigma _{3}^{+}\sigma _{4}^{+}+\beta _{5}(\sigma _{1}^{+}\sigma _{3}^{+}+\sigma _{2}^{+}\sigma _{4}^{+})+\beta _{6}(\sigma _{1}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{3}^{+})+(1-\beta _{5}^{2}-\beta _{6}^{2})\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}$ $LG2_{b}$ $f=\sigma _{1}^{+}\sigma _{2}^{+}+\sigma _{3}^{+}\sigma _{4}^{+}+\beta _{5}(\sigma _{1}^{+}\sigma _{3}^{+}+\sigma _{2}^{+}\sigma _{4}^{+})+\beta _{6}(\sigma _{1}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{3}^{+})$ $LG2_{c}$ $f=\sigma _{1}^{+}\sigma _{3}^{+}\sigma _{4}^{+}+\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}+\sigma _{1}^{+}\sigma _{2}^{+}+\beta _{5}\sigma _{1}^{+}\sigma _{3}^{+}+\beta _{6}\sigma _{2}^{+}\sigma _{3}^{+}$ *Singular $1D$ classes* 1 complex parameters $LG1_{a} $ $f=\sigma _{1}^{+}\sigma _{2}^{+}+\sigma _{1}^{+}\sigma _{3}^{+}+\beta _{6}(\sigma _{1}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{3}^{+})+(1-\beta _{6}^{2})\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}$ $LG1_{b}$ $f=\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}+\sigma _{1}^{+}\sigma _{2}^{+}+\sigma _{1}^{+}\sigma _{3}^{+}+\beta _{6}\sigma _{2}^{+}\sigma _{3}^{+}$ *Singular point classes* no parameters $S_{a}$ $f=\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}+\sigma _{1}^{+}\sigma _{3}^{+}\sigma _{4}^{+}+\sigma _{2}^{+}\sigma _{4}^{+}$ $S_{b}$ $f=\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}+\sigma _{1}^{+}\sigma _{3}^{+}\sigma _{4}^{+}+\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{4}^{+}$ $S_{c}$ $f=\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}+\sigma _{1}^{+}\sigma _{3}^{+}\sigma _{4}^{+}$ $S_{d}$ $f=\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}$ $S_{e}$ $f=\sigma _{3}^{+}\sigma _{4}^{+}+\sigma _{1}^{+}\sigma _{3}^{+}+\sigma _{2}^{+}\sigma _{3}^{+}+\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}$ $S_{f}$ $f=\sigma _{1}^{+}\sigma _{2}^{+}+\sigma _{2}^{+}\sigma _{3}^{+}+\sigma _{3}^{+}\sigma _{1}^{+}+\sigma _{1}^{+}\sigma _{2}^{+}\sigma _{3}^{+}\sigma _{4}^{+}$ $\dots $ $\dots $ --------------------------------------------------------------------------------------------------------- : Classification of four-qubit entanglement classes following from $ SL(2,{\mathbb C})$ transformation properties of the canonic form, see Sec. 3. Entanglement measures for four qubits {#secIII} ===================================== Measures for $sl$- and $su$-entanglement ---------------------------------------- From an information-theoretic standpoint, the construction of well-defined entanglement measures typically relies on the concept of [*entanglement monotone*]{}, that is, of a quantity that is required to be invariant under local unitary transformations and non-increasing on average under LOCC transformations [@Bennett]. For instance, the most widely utilized measures for two and three qubits, the *concurrence*, $C$, and the *residual entanglement* (or 3-tangle), $\tau$, [@Wootters] are entanglement monotones. For a four-qubit system, we have seen in Sec. \[secII\] that the classification is much richer than in the case of three qubits. In the context of such a classification, we would like to first revisit the role of entanglement monotones, and then argue that another class of measures may also be meaningful. In particular, we show how a measure for fourpartite entanglement should be also more precisely defined by imposing additional requirements beside the ones mentioned above. A standard way to construct entanglement monotones is based on exploiting polynomial (algebraic) invariants. Polynomial invariants are polynomial functions of the state coefficients, and a linearly independent finite set of them may be used to distinguish different orbits in the same way the set of invariant tanglemeter’s coefficients does. For example, for a three-qubit system, five (as many as the tanglemeter’s parameters) independent invariants under local unitary transformations exist [@Gingrich], namely the three real numbers $$\begin{aligned} I_{1}& =\psi _{kij}\psi ^{\ast pij}\psi _{pmn}\psi ^{\ast kmn}\ , \nonumber \\ I_{2}& =\psi _{ikj}\psi ^{\ast ipj}\psi _{mpn}\psi ^{\ast mkn\ }, \nonumber \\ I_{3}& =\psi _{ijk}\psi ^{\ast ijp}\psi _{mnp}\psi ^{\ast mnk\ }, \label{EQ3}\end{aligned}$$and the real and the imaginary part of a complex number, $$I_{4}+\mathrm{i}I_{5}=\psi _{ijk}\psi ^{ijp}\psi _{mnp}\psi ^{mnk}\,. \label{EQ3a}$$In the above equations, $\psi ^{ijk}=\epsilon ^{ii^{\prime }}\epsilon ^{jj^{\prime }}\epsilon ^{kk^{\prime }}\psi _{i^{\prime }j^{\prime }k^{\prime }}$, with the convention that summation over repeated indexes ranging over $\{0,1\}$ is left implicit, $\psi ^{\ast }{}^{ijk}$ denotes the complex conjugate of $\psi _{ijk}$, and $\epsilon ^{ii^{\prime }}$ is the antisymmetric tensor of rank $2$. The $su$-invariant quantity $\tau=2|I_{4}+iI_{5}|$ is exactly the 3-tangle, which also remains invariant under the class of local transformations $\otimes _{i=1}^{3}SL_{i}(2,\mathbb{C})$. It was proved in [@Verstraete] that $sl$-invariants behave as entanglement monotones for normalized pure states, since the (square) vector length $\sum_i \psi _{i}\psi _{i}^{\ast}$ is non-increasing under $sl$-transformations and thus may be employed as a measure of $su$-entanglement within a given $sl$ orbit. The main reason for choosing such a vector length as a measure is based on the relation between the determinant $D_{et}\leq 1$ of the physical transformation corresponding to the chosen $sl$-transformation and the probability $\left(\sum_i \psi _{i}\psi _{i}^{\ast}\right)^{-1}$ of the desired outcome of the indirect measurement implementing this transformation: The $\left( n-1\right)$-th power of the probability upper-bounds the determinant, $D_{et}\leq \left( \sum_i \psi _{i}\psi _{i}^{\ast }\right) ^{-n}$. Note that, by definition, an entanglement monotone is an object able to quantify $su$-entanglement by distinguishing different $su$-orbits that belong to the same $sl$-orbit. However, in the case of four (or more) qubits, there exists an infinite number of general $sl$-orbits (see Sec. \[secII\]). This suggests that measures able to compare the [*$sl$-entanglement between such general orbits*]{} should be considered in addition to the $su$-measures. A reasonable suggestion for $sl$-entanglement measures is provided by $sl$-invariants that are also scaling invariants and, therefore, are independent of the specific normalization of the state. One may construct $sl$-invariants for a four-qubit ensemble in a way similar to how the invariant $I_{4}+iI_{5}$ of Eq. (\[EQ3a\]) is constructed; that is, by taking products of several factors $\sim \psi $ (but not factors $\sim \psi ^{\ast }$) and by considering contractions over $SU(2)$-indexes with invariant antisymmetric tensors $\epsilon ^{ii^{\prime }}$. The simplest combination one finds in this way, $$I^{(2)}\ =\ \psi _{ijkl}\psi ^{ijkl}\,, \label{J2}$$is a $sl$-invariant of second order. There also exist three different $sl$-invariants of fourth order, $$\begin{aligned} I_{12}^{(4)}& =I_{34}^{(4)}\ =\ \psi _{ijkl}\psi ^{ijmn}\psi _{opmn}\psi ^{opkl}\,, \nonumber \\ I_{13}^{(4)}& =I_{24}^{(4)}\ =\ \psi _{ikjl}\psi ^{imjn}\psi _{ompn}\psi ^{okpl}\,, \nonumber \\ I_{14}^{(4)}& =I_{23}^{(4)}\ =\ \psi _{iklj}\psi ^{imnj}\psi _{omnp}\psi ^{oklp}\,. \label{4-th}\end{aligned}$$The ratios $I_{12}^{(4)}/(I^{(2)})^{2}$, $I_{13}^{(4)}/(I^{(2)})^{2}$, and $I_{14}^{(4)}/(I^{(2)})^{2}$ are, in addition, invariant with respect to multiplication of the state vector by an arbitrary complex constant. Were these ratio linearly independent, they would suffice for a complete characterization of four-qubit entanglement. However, they are not. The following identity, $$I_{12}^{(4)}+I_{13}^{(4)}+I_{14}^{(4)}\ =\ \frac{3}{2}\left( I^{(2)}\right)^{2}\,, \label{identJ}$$makes such quantities inconvenient for entanglement characterization. Thus, it is necessary to turn to the sixth-order invariants. We consider the following three independent combinations, $$\begin{aligned} \hspace*{-2cm}I_{12}^{(6)}\ & =\frac{1}{6}\ \left( \psi _{ingd}\psi _{mrko}\psi _{sjph}-\psi _{ingo}\psi _{mrkh}\psi _{sjpd}\right) \psi ^{mrgd}\psi ^{inph}\psi ^{sjko}\,, \nonumber \\ \hspace*{-2cm}I_{23}^{(6)}\ & =\frac{1}{6}\ \left( \psi _{ijpo}\psi _{mngh}\psi _{srkd}-\psi _{ijpd}\psi _{mngo}\psi _{srkh}\right) \psi ^{mrgd}\psi ^{inph}\psi ^{sjko}\,, \nonumber \\ \hspace*{-2cm}I_{13}^{(6)}\ & =\frac{1}{6}\ \left( \psi _{ijkh}\psi _{mnpd}\psi _{srgo}-\psi _{ijgh}\psi _{mnkd}\psi _{srpo}\right) \psi ^{mrgd}\psi ^{inph}\psi ^{sjko}\,, \label{6-th}\end{aligned}$$ whose differences give the invariants of Eq. (\[4-th\]) multiplied by $I^{(2)}$. The explicit form of these invariants for a generic state is awkward. However, they take a simple form for the canonic state under $sl$-transformations, which allows us to explicitly relate them to the canonic amplitudes. One finds $$\begin{aligned} \hspace*{-2cm}\psi _{0000}& =\frac{\sqrt{\sqrt{I_{13}^{(6)}+Q}+\sqrt{I_{23}^{(6)}+Q}+\sqrt{I_{12}^{(6)}+Q}-\left( I^{(2)}\right) ^{3/2}}}{\sqrt{2}\left( I^{(2)}\right) ^{1/4}}, \nonumber \\ \hspace*{-2cm}\psi _{1100}& =\psi _{0011}=\frac{\sqrt{\sqrt{I_{13}^{(6)}+Q}-\sqrt{I_{23}^{(6)}+Q}-\sqrt{I_{12}^{(6)}+Q}+\left( I^{(2)}\right) ^{3/2}}}{2\left( I^{(2)}\right) ^{1/4}}, \nonumber \\ \hspace*{-2cm}\psi _{1001}& =\psi _{0110}=\frac{\sqrt{\sqrt{I_{23}^{(6)}+Q}-\sqrt{I_{13}^{(6)}+Q}-\sqrt{I_{12}^{(6)}+Q}+\left( I^{(2)}\right) ^{3/2}}}{2\left( I^{(2)}\right) ^{1/4}}, \nonumber \\ \hspace*{-2cm}\psi _{0101}& =\psi _{1010}=\frac{\sqrt{\sqrt{I_{12}^{(6)}+Q}-\sqrt{I_{23}^{(6)}+Q}-\sqrt{I_{13}^{(6)}+Q}+\left( I^{(2)}\right) ^{3/2}}}{2\left( I^{(2)}\right) ^{1/4}}, \nonumber \\ \hspace*{-2cm}\psi _{1111}& =\frac{\sqrt{I_{13}^{(6)}+Q}+\sqrt{I_{23}^{(6)}+Q}+\sqrt{I_{12}^{(6)}+Q}+\left( I^{(2)}\right) ^{3/2}}{2\sqrt{2}\left( I^{(2)}\right) ^{1/4}\sqrt{\sqrt{I_{13}^{(6)}+Q}+\sqrt{I_{23}^{(6)}+Q}+\sqrt{I_{12}^{(6)}+Q}-\left( I^{(2)}\right) ^{3/2}}}, \label{AMPL}\end{aligned}$$where $Q$ is a root of the following cubic equation: $$(I_{13}^{(6)}+Q)(I_{23}^{(6)}+Q)(I_{12}^{(6)}+Q)=( I^{(2)} )^{3}Q^{2}\,. \label{QE}$$The set of Eqs. (\[AMPL\]) determines the canonic state vector form with respect to pure $SL$-transformations. Upon dividing Eqs. (\[AMPL\]) by $\psi _{0000},$ the ratios $\psi _{1100}/\psi _{0000}$, $\psi _{1001}/\psi _{0000}$, and $\psi _{0101}/\psi _{0000}$ respectively yield the $sl$-tanglemeter coefficients $\beta _{3}$, $\beta _{5}$, and $\beta _{6}$, which are also scaling-invariant. Different roots of the cubic equation (\[QE\]) yield different $sl$-canonic states related by $SL$ transformations. We can choose one particular root by minimizing the difference between the normalization of the canonic state and the initial normalization. Thus, as conjectured in Sec. \[secII\], the $sl$-entanglement in the four-qubit assembly may be completely characterized by [*three independent scale-invariant complex ratios*]{}, $$\begin{aligned} \hspace*{-1cm}\beta _{3}& =\frac{\sqrt{\sqrt{I_{13}^{(6)}+Q}-\sqrt{I_{23}^{(6)}+Q}-\sqrt{I_{12}^{(6)}+Q}+\left( I^{(2)}\right) ^{3/2}}}{\sqrt{2}\sqrt{\sqrt{I_{13}^{(6)}+Q}+\sqrt{I_{23}^{(6)}+P}+\sqrt{I_{12}^{(6)}+Q}-\left( I^{(2)}\right) ^{3/2}}}, \nonumber \\ \hspace*{-1cm} \beta _{5}& =\frac{\sqrt{\sqrt{I_{23}^{(6)}+Q}-\sqrt{I_{13}^{(6)}+Q}-\sqrt{I_{12}^{(6)}+Q}+\left( I^{(2)}\right) ^{3/2}}}{\sqrt{2}\sqrt{\sqrt{I_{13}^{(6)}+Q}+\sqrt{I_{23}^{(6)}+Q}+\sqrt{I_{12}^{(6)}+Q}-\left( I^{(2)}\right) ^{3/2}}}, \nonumber \\ \hspace*{-1cm}\beta _{6}& =\frac{\sqrt{\sqrt{I_{12}^{(6)}+Q}-\sqrt{I_{23}^{(6)}+Q}-\sqrt{I_{13}^{(6)}+Q}+\left( I^{(2)}\right) ^{3/2}}}{\sqrt{2}\sqrt{\sqrt{I_{13}^{(6)}+Q}+\sqrt{I_{23}^{(6)}+Q}+\sqrt{I_{12}^{(6)}+Q}-\left( I^{(2)}\right) ^{3/2}}}, \label{betas}\end{aligned}$$emerging from the invariants of Eqs. (\[6-th\])-(\[J2\]). In view of this, a natural measure of $sl$-entanglement is provided by the sum of squared moduli of the $sl$-tanglemeter coefficients $\beta$, $\mathcal{S}_{2}=\sum \left\vert \beta \right\vert ^{2}$. This yields $\mathcal{S}_{2}=$ $0$ for the GHZ canonic state, whereas $\mathcal{S}_{2}\not=$ $0$ for all other states, thereby exhibiting a similar behavior to the hyper-determinant [@Akimasa]. Accordingly, this measure quantifies how close the orbit is to the [GHZ]{}-orbit. The quantity $\sum \left\vert \beta -\beta ^{\prime }\right\vert ^{2}$ may likewise serve as a measure characterizing the distance between two different $sl$-orbits. As a next question, we wish to suggest a simple measure for characterizing $su$-entanglement in four qubits. A natural candidate is the sum $\mathcal{S}_{1}=\sum \left\vert \psi \right\vert ^{2}$ over the probabilities in Eq. (\[AMPL\]), which gives the standard normalization of the canonic-like state. Once the invariants of Eqs. (\[J2\])-(\[6-th\]) are calculated for a state with unit normalization, this sum quantifies the extent by which the $SL$ transformation required for setting the state to the canonic form differs from a unitary transformation. Thus, $\left\vert \ln \sum \left\vert \psi \right\vert ^{2}\right\vert$ provides us with a suitable measure of such a non-unitarity. By construction, the latter quantity is able to discriminate between different $su$-orbits that belong to the same $sl$-orbit. Interestingly, as found in [@nilpotent], this measure exhibits a strong correlation with the quadratic $sl$-invariant $I^{(2)}$. Measures for fourpartite entanglement ------------------------------------- Having suggested $sl$- and $su$-measures for four qubits in terms of the tanglemeter’s coefficients, we finally proceed to address the more delicate issue of constructing a measure of genuine fourpartite entanglement [@Emary]. In addition to behaving as an entanglement monotone, such a measure should satisfy the requirement of being zero in the $sl$-orbits that do not bear genuine fourpartite entanglement. Within constructions based on $sl$-invariants (different approaches have also been suggested, see e.g. [@Emary; @Osterloh; @Levay1]), the combination of invariants able to satisfy the last requirement is not known to date. For example, $I^{(2)}$ in Eq. (\[J2\]) is a low-order entanglement monotone, but it cannot serve as a good fourpartite measure since it attains its maximum value $1$ for both the four-qubit GHZ state and for a product of two Bell pairs, that is a four-qubit state which manifestly contains no genuine fourpartite correlations [@Verstraete2]. The so-called $4$-concurrence introduced in [@Wong], that is just $\left|I^{(2)}\right|$, exhibits a similar unfavorable behavior. On the other hand, the hyper-determinant $\Delta $ [@Akimasa] is nonzero in the general family of orbits $G_{abcd}$, and zero in all others as well as in the GHZ orbit. According to our results (Table I), recall that the families of orbits $L_{abc_{2}} $, $L_{a_{2}b_{2}}$, and $L_{a_{2}0_{3\oplus 1}}$ are derived as special cases of the general family, and also contain genuine fourpartite entanglement in general. Observing that the determinant ${\cal D}_4$ of the infinitesimal transformations given in Eq. (\[EQ62e\]) is precisely equal to zero in the orbits $G_{abcd}$, $L_{abc_{2}}$, $L_{a_{2}b_{2}}$, and $L_{a_{2}0_{3\oplus 1}}$, we express it in terms of the canonic state amplitudes, $$\kappa_4 =\left\vert \begin{array}{cccc} \Psi _{15} & 2\psi _{6}\psi _{10} & 2\psi _{6}\psi _{12} & 2\psi _{10}\psi _{12} \\ 2\psi _{5}\psi _{9} & \Psi _{15} & 2\psi _{5}\psi _{12} & 2\psi _{9}\psi _{12} \\ 2\psi _{3}\psi _{9} & 2\psi _{3}\psi _{10} & \Psi _{15} & 2\psi _{9}\psi _{10} \\ 2\psi _{3}\psi _{5} & 2\psi _{3}\psi _{6} & 2\psi _{5}\psi _{6} & \Psi _{15}\end{array}\right\vert/\psi_0^8 ,$$where $\Psi _{15}=$ $-\psi _{15}\psi _{0}+\psi _{6}\psi _{9}+\psi _{3}\psi _{12}+\psi _{5}\psi _{10}$. Our proposal is to consider the quantity $$\mathcal{K}_4 =4 \sqrt{\left\vert \kappa_4 \right\vert/\mathcal{A}^4}\,$$ as a measure of proper fourpartite entanglement. Here, the normalization $\mathcal{A}=\sum_i\left\vert\psi _{i}/\psi _{0}\right\vert^2$ ensures that $\mathcal{K}_4$ consistently ranges between $0,1$. Note that $\mathcal{K}_4$ is constructed as a function of $su$-canonic amplitudes, thus it remains invariant under local unitary transformations, while it is not a $sl$-invariant. In order to gain further insight into the properties of $\mathcal{K}_4$, its behavior on a typical four-qubit state is numerically investigated. The distribution of $\mathcal{K}_4$ for pure states sampled uniformly over the Haar measure is depicted in Fig. \[Fig0\]. A non-monotonic behavior is clearly seen, peaked around the value $0.1$, qualitatively resembling the distribution of the $3$-tangle for three-qubit pure states [@Kendon]. Unlike the $4$-concurrence, $\mathcal{K}_4$ attains its maximum value $1$ for the GHZ state, and gives zero for all the states which are separable in some way. For a four qubit cluster state [@Briegel], for instance, it gives $\mathcal{K}_4= 0.3265$, which is well above the average value. While the above features make $\mathcal{K}_4$ an attractive candidate for quantifying fourpartite entanglement, a main disadvantage of $\mathcal{K}_4$ is that it inherits the redundancy of our classification, vanishing whenever the general class of orbits cannot be reached by infinitesimal transformations – irrespective of whether it might be reached by finite transformations. Furthermore, the calculation of $\mathcal{K}_4$ for a given pure state requires in general that the latter is first reduced to its $su$-canonic form. On the other hand, extending the construction of this measure to $n>4$ qubits is relatively straightforward in principle. For $n=4$, the fact that the determinant $\kappa_4$ does not contain the second-highest order terms is a sign that this measure is approximate for arbitrary states. However, this effect may expected to become less pronounced (hence the accuracy of such approximation improves) with increasing $n$. Finally, it remains an open question to ascertain whether this measure is an entanglement monotone. Discussion {#secIV} ========== In summary, we have demonstrated how the approach based on nilpotent polynomials may be employed to identify entanglement classes for the illustrative yet highly nontrivial situation of four qubits in a pure state. Even if the approach is somewhat redundant compared to mathematically more sophisticated methods, we believe it has the advantage of offering a clearer physical interpretation, and it may also be extended straightforwardly to larger multipartite ensembles and higher-dimensional subsystems. In the context of the resulting classification, we have suggested additional class of measures beside the existing ones, which remain invariant under either local unitary ($su$) or arbitrary local invertible ($sl$) transformations. We employ the nilpotent invariant coefficients for the construction of such measures as an alternative route to invoking polynomial algebraic invariants. Finally, we have also suggested a measure of genuine fourpartite entanglement. Our prospective measure is both, by construction, a $su$-invariant and it vanishes on the special orbits where no genuine fourpartite entanglement exists. We also note that while it would be very illustrative to apply the nilpotent method in the well-explored three qubit case, the procedure implemented here is not viable, since in this case the $SL$ coset dimension $D_{sl}<0$ and, consequently, the dynamical set of equations employed in the analysis becomes degenerate. It is our hope that the results presented here may serve as a stimulus to prompt further investigations and applications of the nilpotent polynomial formalism as a tool exploring entanglement. Acknowledgments {#acknowledgments .unnumbered} =============== It is a pleasure to thank Vladimir Akulin and Andrei Smilga for insightful discussions and collaboration on the nilpotent entanglement program. The authors are also indebted to Winton G. Brown for valuable suggestions and a critical reading of the manuscript. AM gratefully acknowledges Ile de France for financial support. References {#references .unnumbered} ========== [99]{} Dür W, Vidal G, and Cirac J I *Three qubits can be entangled in two inequivalent ways* 2000 *Phys. Rev. A.* **62** 062314 Gingrich R M *Properties of entanglement monotones for three-qubit pure states* 2002 *Phys. Rev. A* **65** 052302 Acin A, Bruss D, Lewenstein M, and Sanpera A *Classification of Mixed Three-Qubit States* 2001 *Phys. Rev. Lett.* **87** 040401 Miyake A *Classification of multipartite entangled states by multidimensional determinants* 2003 *Phys. Rev. A* **67** 012108; Miyake A and Verstraete F *Multipartite entanglement in 2 x 2 x n quantum systems* 2004 *Phys. Rev. A* **69** 012101 Verstraete F, Dehaene J, DeMoor B, and Verschelde H *Four qubits can be entangled in nine different ways* 2002 *Phys. Rev. A* **65** 052112 Grassl M, Rötteler M, and Beth T *Computing Local Invariants of Quantum-Bit System* 1998 *Phys. Rev. A* **58**, 1833 Luque J G and Thibon J Y *Polynomial Invariants of Four Qubits* 2003 *Phys. Rev. A* **67** 042303 Leváy P *On the geometry of four-qubit invariants* 2006 *J. Phys. A: Math. Gen.* **39** 9533 Briand E, Luque J G and Thibon J Y *A complete set of covariants of the four qubit system* 2003 *J. Phys. A: Math. Gen.* **36** 9915-9927 Verstraete F, Dehaene J, and DeMoor B *Normal forms and entanglement measures for multipartite quantum states* 2003 *Phys. Rev. A* **68** 012103 Vidal G *Entanglement monotones* 2000 *J. Mod. Opt.* **47** 355 Mandilara A, Akulin V M, Smilga A V and Viola L *Quantum entanglement via nilpotent polynomials* 2006 *Phys. Rev. A* **74** 022331 Bennett C H, Popescu S, Rohrlich D, Smolin J A, and Thapliyal A V *Exact and asymptotic measures of multipartite pure-state entanglement* 2001 *Phys. Rev. A* **63** 012307 Barnum H, Knill E, Ortiz G, and Viola L *Generalizations of entanglement based on coherent states and convex sets* 2003 *Phys. Rev. A* [**68**]{} 032308; Somma R, Barnum H, Knill E, Ortiz G, and Viola L *A subsystem-independent generalization of entanglement* 2004 *Phys. Rev. Lett.* [**92**]{} 107902 Linden N and Popescu S *On Multi-Particle Entanglement* 1998 *Fortschr. Phys.* **46** 567 Carteret H A, Higuchi A, and Sudbery A *Multipartite generalization of the Schmidt decomposition* 2000 *J. Math. Phys.* **41** 7932 Wootters W K *Entanglement of Formation of an Arbitrary State of Two Qubits* 1998 *Phys. Rev. Lett.* **80** 2245; Coffman V, Kundu J, and Wootters W K *Distributed entanglement* 2000 *Phys. Rev. A* **61** 052306 Emary C *A bipartite class of entanglement monotones for N-qubit pure states* 2004 *J. Phys. A: Math. Gen*. **37** 8293-8302 Osterloh A and Siewert J *Constructing N-qubit entanglement monotones from antilinear operators* 2005 *Phys. Rev. A* **72** 012337 Leváy P *On the geometry of a class of N-qubit entanglement monotones* 2005 *J. Phys. A: Math. Gen.* **38** 9075 Wong A and Christensen N *Potential multiparticle entanglement measure* 2001 *Phys. Rev. A* **63** 044301 Kendon V M, Nemoto K, and Munro W J *Typical entanglement in multi-qubit systems* 2002 *J. Mod. Optics* **49** 1709 Briegel H J and Raussendorf R *Persistent entanglement in arrays of interacting particles* 2001 *Phys. Rev. Lett.***86** 910

--- author: - | Maxime Bucher, Stéphane Herbin\ ONERA - The French Aerospace Lab\ Palaiseau, France\ [maxime.bucher@onera.fr,]{}\ [stephane.herbin@onera.fr]{} - | Frédéric Jurie\ Normandie Univ, UNICAEN, ENSICAEN, CNRS\ Caen, France\ [frederic.jurie@unicaen.fr]{} bibliography: - '0-bibfile.bib' title: 'Generating Visual Representations for Zero-Shot Classification' ---

--- abstract: 'Sequences based on the Distant Dipolar Field (DDF) have shown great promise for novel spectroscopy and imaging. Unless spatial variation in the longitudinal magnetization, $M_{z}(s)$, is eliminated by relaxation, diffusion, or spoiling techniques by the end of a single repetition, unexpected results can be obtained due to spatial harmonics in the steady state $M_{z}^{SS}(s)$ profile. This is true even in a homogeneous single-component sample. We have developed an analytical expression for the $M_{z}^{SS}(s)$ profile that occurs in DDF sequences when smearing by diffusion is negligible in the $TR$ period. The expression has been verified by directly imaging the $M_{z}^{SS}(s)$ profile after establishing the steady state.' address: 'AHSC, P.O. Box 245067, Tucson, AZ 85724-5067 USA' author: - 'C. A. Corum' - 'A. F. Gmitro' bibliography: - 'mz.bib' date: 1st April 2004 title: 'Spatially Varying Steady State Longitudinal Magnetization in Distant Dipolar Field-based Sequences' --- distant dipolar field ,DDF,intermolecular multiple quantum coherence ,iMQC ,steady state longitudinal magnetization 82.56.Jn ,82.56.Na ,87.61.Cd Introduction ============ NMR and MRI sequences utilizing the Distant Dipolar Field (DDF) have the relatively unique property of preparing, utilizing, and leaving spatially-modulated longitudinal magnetization, $M_{z}(s)$, where $\hat{s}$ is in the direction of an applied gradient. In fact this is fundamental to producing the novel “multiple spin echo”[@DBD79; @BBG90] or “non-linear stimulated echo” [@ASDK97] of the classical picture and making the “intermolecular multiple quantum coherence (iMQC)” [@QH93] observable in the quantum picture. Existing analytical signal equations for DDF/iMQC sequences depend on $M_{z}(s)$ being sinusoidal during the signal build period[@alw98b; @CG04b]. Experiments that probe sample structure also require a well-defined “correlation distance” which is defined as the repetition distance of $M_{z}(s)$ [@RB96; @WAM+98; @ACM02]. If the repetition time $TR$ of the DDF sequence is such that full relaxation is not allowed to proceed $TR<5T_{1}$, or diffusion does not average out the modulation, spatially-modulated longitudinal magnetization will be left at the end of one iteration of the sequence. The next repetition of the sequence will begin to establish “harmonics” in what is desired to be a purely sinusoidal modulation pattern. Eventually a steady state is established, potentially departing significantly from a pure sinusoid. Experimental Methods\[sec:Experimental-Methods\] ================================================ In order to study the behavior of the steady state $M_{z}^{SS}(s)$ profile we have implemented a looped DDF preparation subsequence followed by a standard multiple-phase encode imaging sub-sequence. (Figure \[fig:Pulse-Sequence\].) The $\alpha$ pulse excites the system, the gradient $G_{q}$ twists the transverse magnetization into a helix. $\beta$ rotates one component of the helix back into the longitudinal direction. For simplicity we have omitted the $180^{\circ}$ pulses used to create a spin echo during TM and/or TB sometimes present in DDF sequences. Also, we are only interested in $M_{z}(s)$ in this experiment, not the actual DDF-generated transverse signal. Looping the “preparation” sub-sequence thus creates the periodic $M_{z}(s)$ profile, spoils remaining transverse magnetization, and establishes $M_{z}^{SS}(s)$. The $\varepsilon$ pulse converts $M_{z}^{SS}(s)$ into transverse magnetization, allowing it to be imaged via the subsequent spin echo “image” sub-sequence. $M_{z}^{SS}(s)$ must be re-established by the “preparation” sub-sequence for each phase encode. After a suitably long full relaxation delay “relax,” the sequence is repeated to acquire the next k-space line. This is clearly a slow acquisition method because many $TR$ periods are required to reach steady state in the preparation before each k-space line is acquired. The sequence is intended as a tool to directly image the $M_{z}^{SS}(s)$ profile, verifying the $M_{z}^{SS}(s)$ that would occur in a steady state DDF sequence, not as a new imaging modality. Theory ====== The effect of the ”preparation” pulse sequence was first determined for a single iteration. The progress along the sequence is denoted by the the superscript. Starting with fully relaxed equilibrium magnetization before the $\alpha$ pulse:$$M_{z}^{Eq}(s)=M_{0}\label{eq:initial}$$ after the $\alpha$ pulse, the mix delay $TM$ and the $\beta$ pulse we have:$$\begin{aligned} M_{z}^{\beta}(s) & =[A^{\beta}cos(q\, s)+B^{\beta}]\, M_{z}^{Eq}+C^{\beta}M_{0}\label{eq:beta}\end{aligned}$$ $$A^{\beta}=-sin(\alpha)\, e^{-\frac{TM}{T_{2}}}sin(\beta)$$ $$B^{\beta}=cos(\alpha)\, e^{-\frac{TM}{T_{1}}}cos(\beta)$$ $$C^{\beta}=(1-e^{-\frac{TM}{T_{1}}})\, cos(\beta)$$ The parameter $q=\frac{2\pi}{\lambda}$, where $\lambda$ is the helix pitch resulting from the applied gradient. Diffusion has been assumed to be negligible at the scale of $\lambda$. Note that $T_{2}$ is used in $A$ rather than $T_{2}^{*}$ when $G_{q}$ is larger than background inhomogeneity and susceptibility gradients. After the build delay $TB$ we have: $$M_{z}^{TB}(s)=[A^{TB}cos(q\, s)+B^{TB}]\, M_{z}^{Eq}(s)+C^{TB}M_{0}\label{eq:TE}$$ $$A^{TB}=-sin(\alpha)\, e^{-\frac{TM}{T_{2}}}sin(\beta)\, e^{-\frac{TB}{T_{1}}}$$ $$B^{TB}=cos(\alpha)\, e^{-\frac{TB}{T_{1}}}cos(\beta)\, e^{-\frac{TB}{T_{1}}}$$ $$C^{TB}=[(1-e^{-\frac{TB}{T_{1}}})\, cos(\beta)-1]\, e^{-\frac{TB}{T_{1}}}+1$$ At the start of the next repetition, after a $TR$ period inclusive of $TM$ and $TB$ we have$$M_{z}^{TR}(s)=[A^{TR}cos(q\, s)+B^{TR}]M_{z}^{Eq}(s)+C^{TR}\, M_{0}\label{eq:one}$$ $$A^{TR}=-sin(\alpha)\, e^{-\frac{TM}{T_{2}}}sin(\beta)\, e^{-\frac{TR-TM}{T_{1}}}$$ $$B^{TR}=cos(\alpha)\, cos(\beta)\, e^{-\frac{TR}{T_{1}}}$$ $$C^{TR}=[(1-e^{-\frac{TM}{T_{1}}})\, cos(\beta)-1]\, e^{-\frac{TR-TM}{T_{1}}}+1$$ If we apply the sequence $N$ times and re-arrange the terms we get the series:$$M_{z}^{NxTR}(s)=M_{0}+M_{0\,}[A^{TR}cos(q\, s)+B^{TR}+C^{TR}-1]\sum\limits _{n=1}^{N}[A^{TR}cos(q\, s)+B^{TR}]^{n-1}\label{eq:sum}$$ for the starting magnetization state after $N$ repetitions of the sequence. Summing an infinite number of terms results in the expression for the steady state $M_{z}^{SS}(s)$ after a large number of TR periods:$$M_{z}^{SS}(s)=M_{0}-M_{0\,}[\frac{A^{TR}cos(q\, s)+B^{TR}+C^{TR}-1}{A^{TR}cos(q\, s)+B^{TR}-1}]\label{eq:steady_state}$$ One can then calculate the magnetization state after the $\beta$ pulse in the steady state:$$M_{z}^{SS,\,\beta}(s)=[A^{\beta}cos(q\, s)+B^{\beta}]\, M_{z}^{SS}(s)+C^{\beta}M_{0}\label{eq:ss_beta}$$ and after $TB$: $$M_{z}^{SS,\, TB}(s)=[A^{TB}cos(q\, s)+B^{TB}]\, M_{z}^{SS}(s)+C^{TB}M_{0}\label{eq:ss_tb}$$ We show graphs of Eq. [\[]{}\[eq:steady\_state\][\]]{}, [\[]{}\[eq:ss\_beta\][\]]{}, and [\[]{}\[eq:ss\_tb\][\]]{} in Figure \[fig:Mz\_theory\_image\] for $TR=2s$. Results ======= We now show in Figure \[fig:Mz\_images\] representative $M_{z}^{SS}(s)$ magnitude images obtained with the sequence described in section \[sec:Experimental-Methods\] for four different values of $TR=5s,\ 2s,\ 1s,\ 500ms$. Figure \[fig:Mz\_fit\] shows several cross sections through row \#128 of Figure \[fig:Mz\_images\]. The object is an 18mm glass sphere filled with silicone oil. Data points are superimposed with the corresponding magnitude of the theoretical curve. The $T_{1}$ of the silicone oil (at 400MHz) was measured by spectroscopic inversion recovery to be 1.4s. A Bruker DRX400 Micro 2.5 system was used with a custom 27mm diameter 31P/1H birdcage coil. 10 $TR$ periods were used to establish steady state. A 10s “relax” delay was used between phase encodes to establish full relaxation. $G_{q}$ was 3ms and 2.5mT/mm, with $G_{spoil1}$ of 5ms and 100mT/mm. No attempt was made to account for $B_{1}$ inhomogeneity. A single scaling parameter was used for all theoretical curves. We achieved good agreement with the theoretical predictions. In the sequence as used $TM=TB=~7ms$. A variety of other $G_{q}$ directions and strengths show similar agreement with theory. Better agreement in the fit between experiment and theory can be obtained with $\alpha=\beta=75^{\circ}$than with the nominal $90^{\circ}$. A $B_{1}$ map needs to be determined to see if this corresponds more closely to the actual experimental conditions. Conclusions =========== The expressions developed and verified above should be useful to those wishing to understand or utilize harmonics in the $M_{z}^{SS}(s)$ profile in DDF based sequences in the situation where the diffusion distance during $TR$ compared with $\lambda$ in negligible. This is especially true for those carrying out structural measurements which depend on a well defined correlation distance. The theory should also hold for spatially varying magnetization density $M_{0}=M_{0}(\vec{r})$, and longitudinal relaxation $T_{1}=T_{1}(\vec{r})$. Acknowledgements ================ This work and preparation leading to it was carried out under the support of the Flinn Foundation, a State of Arizona Prop. 301 Imaging Fellowship, and NIH 5R24CA083148-05.

--- abstract: 'In natural language the intended meaning of a word or phrase is often implicit and depends on the context. In this work, we propose a simple yet effective method for sentiment analysis using contextual embeddings and a self-attention mechanism. The experimental results for three languages, including morphologically rich Polish and German, show that our model is comparable to or even outperforms state-of-the-art models. In all cases the superiority of models leveraging contextual embeddings is demonstrated. Finally, this work is intended as a step towards introducing a universal, multilingual sentiment classifier.' author: - 'Katarzyna Biesialska [^1]' - | \ Magdalena Biesialska[^^]{} - | \ Henryk Rybinski bibliography: - 'bibliography.bib' title: | Sentiment Analysis with\ Contextual Embeddings and Self-Attention --- Introduction ============ All areas of human life are affected by people’s views. With the sheer amount of reviews and other opinions over the Internet, there is a need for automating the process of extracting relevant information. For machines, however, measuring sentiment is not an easy task, because natural language is highly ambiguous at all levels, and thus difficult to process. For instance, a single word can hardly convey the whole meaning of a statement. Moreover, computers often do not distinguish literal from figurative meaning or incorrectly handle complex linguistic phenomena, such as: sarcasm, humor, negation etc. In this paper, we take a closer look at two factors that make automatic opinion mining difficult – the problem of representing text information, and sentiment analysis (SA). In particular, we leverage contextual embeddings, which enable to convey a word meaning depending on the context it occurs in. Furthermore, we build a hierarchical multi-layer classifier model, based on an architecture of the Transformer encoder [@vaswani2017attention], primarily relying on a self-attention mechanism and bi-attention. The proposed sentiment classification model is language independent, which is especially useful for low-resource languages (e.g. Polish). We evaluate our methods on various standard datasets, which allows us to compare our approach against current state-of-the-art models for three languages: English, Polish and German. We show that our approach is comparable to the best performing sentiment classification models; and, importantly, in two cases yields significant improvements over the state of the art. The paper is organized as follows: Section 2 presents the background and related work. Section 3 describes our proposed method. Section 4 discusses datasets, experimental setup, and results. Section 5 concludes this paper and outlines the future work. Related Work ============ Sentiment classification has been one of the most active research areas in natural language processing (NLP) and has become one of the most popular downstream tasks to evaluate performance of neural network (NN) based models. The task itself encompasses several different opinion related tasks, hence it tackles many challenging NLP problems, see e.g. [@Liu2012SentimentAA; @Mohammad2016SentimentAD]. Sentiment Analysis Approaches ----------------------------- The first fully-formed techniques for SA emerged around two decades ago, and continued to be prevalent for several years, until deep learning methods entered the stage. The most straight-forward method, developed in [@turney-2002-thumbs], is based on the number of positive and negative words in a piece of text. Concretely, the text is assumed to have positive polarity if it contains more positive than negative terms, and vice versa. Of course, the term-counting method is often insufficient; therefore, an improved method was proposed in [@Kennedy2006SentimentCO], which combines counting positive and negative terms with a machine learning (ML) approach (i.e. Support Vector Machine). Various studies (e.g. [@Turney2010FromFT]) have shown that one can determine the polarity of an unknown word by calculating co-occurrence statistics of it. Moreover, classical solutions to the SA problem are often based on lexicons. Traditional lexicon-based SA leverages word-lists, that are pre-annotated with positive and negative sentiment. Therefore, for many years lexicon-based approaches have been utilized when there was insufficient amount of labeled data to train a classifier in a fully supervised way. In general, ML algorithms are popular methods for determining sentiment polarity. A first ML model applied to SA has been implemented in [@pang2002thumbs]. Moreover, throughout the years, different variants of NN architectures have been introduced in the field of SA. Especially recursive neural networks [@Paulus2014GlobalBR], such as recurrent neural networks (RNN) [@Socher2013RecursiveDM; @tai-etal-2015-improved; @Kumar2015AskMA], or convolutional neural networks (CNN) [@Kalchbrenner2014ACN; @Kim2014ConvolutionalNN] have become the most prevalent choices. Vector Representations of Words ------------------------------- One of the principal concepts in linguistics states that related words can be used in similar ways [@firth1957synopsis]. Importantly, words may have different meaning in different contexts. Nevertheless, until recently it has been a dominant approach (e.g. word2vec [@Mikolov2013DistributedRO], GloVe [@Pennington2014GloveGV]) to learn representations such that each and every word has to capture all its possible meanings. However, lately a new set of methods to learn dynamic representations of words has emerged [@mccann2017learned; @howard2018universal; @Peters2018; @radfordimproving; @devlin2019bert]. These approaches allow each word representation to capture what a word means in a particular context. While every word token has its own vector, the vector can depend on a variable-length sequence of nearby words (i.e. context). Consequently, a context vector is obtained by feeding a neural network with these context word vectors and subsequently encoding them into a single fixed-length vector. [R]{}[0.5]{} {width="50.00000%"} ULMFiT [@howard2018universal] was the very first method to induce contextual word representations by harnessing the power of language modeling. The authors proposed to learn contextual embeddings by pre-training a language model (LM), and then performing task-specific fine-tuning. ULMFiT architecture is based on a vanilla 3-layer Long Short-Term Memory (LSTM) NN without any attention mechanism. The other contextual embedding model introduced recently is called ELMo (Embeddings from Language Models) [@Peters2018]. Similarly to ULMFiT, this model uses tokens at the word-level. ELMo contextual embeddings are “deep” as they are a function of all hidden states. Concretely, context-sensitive features are extracted from a left-to-right and a right-to-left 2-layer bidirectional LSTM language models. Thus, the contextual representation of each word is the concatenation of the left-to-right and right-to-left representations as well as the initial embedding (see Fig. \[fig:elmo\]). The most recent model – BERT [@devlin2019bert] – is more sophisticated architecturally-wise, as it is a multi-layer masked LM based on the Transformer NN utilizing sub-word tokens. However, as we are bound to use word-level tokens in our sentiment classifier, we leverage the ELMo model for obtaining contextual embeddings. More specifically, by means of ELMo we are able to feed our classifier model with context-aware embeddings of an input sequence. Hence, in this setting we do not perform any fine-tuning of ELMo on a downstream task. Self-Attention Deep Neural Networks ----------------------------------- The attention mechanism was introduced in [@bahdanau2014neural] in 2014 and since then it has been applied successfully to different computer vision (e.g. visual explanation) and NLP (e.g. machine translation) tasks. The mechanism is often used as an extra source of information added on top of the CNN or LSTM model to enhance the extraction of sentence embedding [@Santos2016AttentivePN; @Lin2017ASS]. However, this scenario is not applicable to sentiment classification, since the model only receives a single sentence on input, hence there is no such extra information [@Lin2017ASS]. Self-attention (or intra-attention) is an attention mechanism that computes a representation of a sequence by relating different positions of a single sequence. Previous work on sentiment classification has not covered extensively attention-based neural network models for SA (especially using the Transformer architecture [@vaswani2017attention]), although some papers have appeared recently [@ambartsoumian2018self; @letarte2018importance]. The Proposed Approach ===================== Our proposed model, called Transformer-based Sentiment Analysis (TSA) (see Fig. \[fig:tsa\]), is based on the recently introduced Transformer architecture [@vaswani2017attention], which has provided significant improvements for the neural machine translation task. Unlike RNN or CNN based models, the Transformer is able to learn dependencies between distant positions. Therefore, in this paper we show that attention-based models are suitable for other NLP tasks, such as learning distributed representations and sentiment analysis, and thus are able to improve the overall accuracy. [R]{}[0.35]{} {width="27.00000%"} The architecture of the TSA model and steps to train it can be summarized as follows: 1. At the very beginning there is a simple text pre-processing method that performs text clean-up and splits text into tokens. 2. We use contextual word representations to represent text as real-valued vectors. 3. After embedding the text into real-valued vectors, the Transformer network maps the input sequence into hidden states using self-attention. 4. Next a bi-attention mechanism is utilized to estimate the interdependency between representations. 5. A single layer LSTM together with self-attentive pooling compute the pooled representations. 6. A joint representation for the inputs is later passed to a fully-connected neural network. 7. Finally, a softmax layer is used to determine sentiment of the text. Embeddings and Encoded Positional Information --------------------------------------------- Non-recurrent models, such as deep self-attention NN, do not necessarily process the input sequence in a sequential manner. Hence, there is no way they can record the position of each word in a sequence, which is an inherent limitation of every such model. Therefore, in the case of the Transformer, the need has been addressed in the following manner – the Transformer takes into account the order of the words in the input sequence by encoding their position information in extra vectors (so called positional encoding vectors) and adding them to input embeddings. There are many different approaches to embed position information, such as learned or fixed positional encodings (PE), or recently introduced relative position representations (RPR) [@shaw2018]. The original Transformer used sine and cosine functions of different frequencies. In this work, we explore the effectiveness of applying a modified approach to incorporate positional information into the model, namely using RPR instead of PE. Furthermore, we use global average pooling in order to average the output of the last self-attention layer and prepare the model for the final classification layer. The Transformer Encoder ----------------------- The input sequence is combined with word and positional embeddings, which provide time signal, and together are fed into an encoder block. Matrices for a query *Q*, a key *K* and a value *V* are calculated and passed to a self-attention layer. Next, a normalization is applied and residual connections provide additional context. Further, a final dense layer with vocabulary size generates the output of the encoder. A fully-connected feed-forward network within the model is a single hidden layer network with a ReLU activation function in between: $$FFN(x)=\max\left(0, x W_{1}+b_{1}\right) W_{2}+b_{2}$$ Self-Attention Layer -------------------- The self-attention block in the encoder is called multi-head self-attention. A self-attention layer allows each position in the encoder to access all positions in the previous layer of the encoder immediately, and in the first layer all positions in the input sequence. The multi-head self-attention layer employs *h* parallel self-attention layers, called heads, with different *Q*, *K*, *V* matrices obtained for each head. In a nutshell, the attention mechanism in the Transformer architecture relies on a scaled dot-product attention, which is a function of *Q* and a set of *K*-*V* pairs. The computation of attention is performed in the following order. First, a multiplication of a query and transposed key is scaled through the scaling factor of $1/\sqrt{d_{z}}$ (Eq. 2) $$m_{i j}=\frac{Q K^{T}}{\sqrt{d_{z}}}$$ Next, the attention is produced using the softmax function over their scaled inner product: $$\alpha_{i j}=\frac{e^{m_{i j}}}{\sum_{k=1}^{n} e^{m_{i k}}}$$ Finally, the weighted sum of each attention head and a value is calculated as follows: $$z_{i}=\sum_{j=1}^{n} \alpha_{i j} V$$ Masking and Pooling ------------------- Similar to other sources of data, the datasets used for training and evaluation of our models contain sequences of different length. The most common approach in the literature involves finding a maximal sequence length existing in the dataset/batch and padding sentences that are shorter than the longest one with trailing zeroes. In the proposed TSA model, we deal with the problem of variable-length sequences by using masking and self-attentive pooling. The inspiration for our approach comes from the BCN model proposed in [@mccann2017learned]. Thanks to this mechanism, we are able to fit sequences of different length into the final fixed-size vector, which is required for the computation of the sentiment score. The self-attentive pooling layer is applied just after the encoder block. Experiments =========== Datasets -------- In this work, we compare sentiment analysis results considering four benchmark datasets in three languages. All datasets are originally split into training, dev and test sets. Below we describe these datasets in more detail. **Dataset** **\# Classes** **Train** **Dev** **Test** **Domain** **Language** --------------- ---------------- ----------- --------- ---------- ------------------- -------------- SST-2 2 6,920 872 1,821 movies English SST-5 5 8,544 1,101 2,210 movies English PolEmo 2.0-IN 5 5,783 723 722 medical, hotels Polish GermEval 3 19,432 2,369 2,566 travel, transport German : Sentiment analysis datasets with number of classes and train/dev/test split.[]{data-label="tab1"} #### **Stanford Sentiment Treebank (SST)** This collection of movie reviews [@Socher2013RecursiveDM] from the `rottentomatoes.com` is annotated for the binary (SST-2) and fine-grained (SST-5) sentiment classification. SST-2 divides reviews into two groups: *positive* and *negative*, while SST-5 distinguishes 5 different review types: *very positive*, *positive*, *neutral*, *negative*, *very negative*. The dataset consists of 11,855 single sentences and is widely used in the NLP community. #### **PolEmo 2.0** The dataset [@kocon-etal-2019-multi] comprises online reviews from education, medicine and hotel domains. There are two separate test sets, to allow for in-domain (medicine and hotels) and out-of-domain (products and university) evaluation. The dataset comes with the following sentiment labels: *strong positive*, *weak positive*, *neutral*, *weak negative*, *strong negative*, and *ambiguous*. #### **GermEval** This dataset [@germevaltask2017] contains customer reviews of the railway operator (Deutsche Bahn) published on social media and various web pages. Customers expressed their feedback regarding the service of the railway company (e.g. travel experience, timetables, etc.) by rating it as *positive*, *negative*, or *neutral*. Experimental Setup ------------------ Pre-processing of input datasets is kept to a minimum as we perform only tokenization when required. Furthermore, even though some datasets, such as SST or GermEval, provide additional information (i.e. phrase, word or aspect-level annotations), for each review we only extract text of the review and its corresponding rating. The model is implemented in the Python programming language, PyTorch[^2] and AllenNLP[^3]. Moreover, we use pre-trained word-embeddings, such as ELMo [@Peters2018], GloVe [@Pennington2014GloveGV]. Specifically, we use the following ELMo models: Original[^4], Polish [@janz-elmo-2019] and German [@GerElmo]. In the ELMO+GloVe+BCN model we use the following 300-dimension GloVe embeddings: English[^5], Polish [@dadas2019] and German[^6]. In order to simplify our approach when training the sentiment classifier model, we establish a very similar setting to the vanilla Transformer. We use the same optimizer - Adam with $\beta_{1}=0.9, \beta_{2}=0.98$, and $\epsilon=10^{-9}$. We incorporate four types of regularization during training: dropout probability $P_{drop}=0.1$, embedding dropout probability $P_{emb}=0.5$, residual dropout probability $P_{res}=0.2$, and attention dropout probability $P_{attn}=0.1$. We use 2 encoder layers. In addition, we employ label smoothing of value $\epsilon_{ls}=0.1$. In terms of RPR parameters, we set clipping distance to $k=10$. Results and Discussion ---------------------- In Table \[tab:results\], we summarize experimental results achieved by our model and other state-of-the-art systems reported in the literature by their respective authors. We observe that our models, baseline and ELMo+TSA, outperform state-of-the-art systems for all three languages. More importantly, the presented accuracy scores indicate that the TSA model is competitive and for two languages (Polish and German) achieves the best results. Also noteworthy, in Table \[tab:results\], there are two models that use some variant of the Transformer: SSAN+RPR [@ambartsoumian2018self] uses the Transformer encoder for the classifier, while Polish BERT [@KLEJ] employs Transformer-based language model introduced in [@devlin2019bert]. One of the reasons why we achieve higher score for the SST dataset might be that the authors of SSAN+RPR used word2vec embeddings [@Mikolov2013DistributedRO], whereas we employ ELMo contextual embeddings [@Peters2018]. Moreover, in our TSA model we use not only self-attention (as in SSAN+RPR) but also a bi-attention mechanism, hence this also should provide performance gains over standard architectures. In conclusion, comparing the results of the models leveraging contextual embeddings (CoVe+BCN, Polish BERT, ELMo+GloVe+BCN and ELMo+TSA) with the rest of the reported models, which use traditional distributional word vectors, we note that the former category of sentiment classification systems demonstrates remarkably better results. Conclusion and Future Work ========================== We have presented a novel architecture, based on the Transformer encoder with relative position representations. Unlike existing models, this work proposes a model relying solely on a self-attention mechanism and bi-attention. We show that our sentiment classifier model achieves very good results, comparable to the state of the art, even though it is language-agnostic. Hence, this work is a step towards building a universal, multi-lingual sentiment classifier. In the future, we plan to evaluate our model using benchmarks also for other languages. It is particularly interesting to analyze the behavior of our model with respect to low-resource languages. Finally, other promising research avenues worth exploring are related to unsupervised cross-lingual sentiment analysis. [^1]: Both authors contributed equally to this work, which was mostly done at the Warsaw University of Technology. [^2]: https://pytorch.org [^3]: https://allennlp.org [^4]: https://allennlp.org/elmo [^5]: http:$//$nlp.stanford.edu$/$data$/$glove.840B.300d.zip [^6]: https://wikipedia2vec.github.io/wikipedia2vec/pretrained

--- abstract: 'We study the dynamics of asymptotically regular discrete isometry groups $\Gamma\subset G$ on the ideal boundary of higher rank symmetric spaces of noncompact type $X=G/K$. We identify geometrically domains of discontinuity in all (partial) flag manifolds $G/P$ associated to $G$. Under the additional assumption of chamber conicality - a natural generalization of conicality of the limit set for Kleinian groups in rank one - and antipodality, we establish cocompactness of the $\Ga$-action on various domains of discontinuity, in particular for domains in the full flag manifold $G/B$. The latter are nonempty at least for all irreducible Weyl groups $\neq B_2,G_2$. We also compare in higher rank several notions related to classical geometric finiteness.' author: - 'Michael Kapovich, Bernhard Leeb, Joan Porti' date: 'June 16, 2013' title: Dynamics at infinity of regular discrete subgroups of isometries of higher rank symmetric spaces --- Introduction ============ In this paper we study discrete isometry groups of symmetric spaces $X$ of noncompact type and higher rank, $\Gamma\subset\Isom(X)=G$. Our goal is two-fold: 1. We study the dynamics of the group $\Gamma$ on the ideal boundary (sphere at infinity) $\geo X$ of $X$, more precisely, on the (partial) flag manifolds $G/P$ associated to the group $G$. The main purpose is to identify geometrically domains of discontinuity in these flag manifolds and establish, under certain additional (RCA) assumptions on $\Ga$, the cocompactness of the $\Ga$-action on various domains. 2. We compare in higher rank several notions related to classical geometric finiteness. The [*conicality*]{} of limit chambers emerges as a key notion connecting (1) and (2). We now discuss the two major themes and sets of results in detail. Along the way, we compare and contrast the higher rank theory with the discrete isometry groups of rank one symmetric spaces. For rank one symmetric spaces, i.e. when $X$ has strictly negative sectional curvature, there is a precise dynamical decomposition of $\geo X$: The action $G\acts\geo X$ on the sphere at infinity is transitive and has [*convergence*]{} dynamics, which means that every divergent sequence of isometries $g_n\to\infty$ in $G$ subconverges on the complement of one point in $\geo X$ locally uniformly to a constant map. With respect to the $\Ga$-action, there is a decomposition of $\geo X$ into a (canonically defined) domain of discontinuity $\Om$ and the limit set $\La$. The action of $\Ga$ on $\Om$ is not always cocompact. It is cocompact if (but not only if) its extension to $X\cup\Om=\bar X-\La$ is cocompact, where $\bar X=X\cup\geo X$ denotes the visual compactification of $X$. The latter is, in rank one, equivalent to a series of other conditions (see section \[sec:rank1\]), including: \(i) $\Ga$ is convex cocompact. \(ii) All limit points of $\Ga$ are conical. \(iii) The action of $\Ga$ is expanding at $\La$. \(iv) $\Ga$ is word hyperbolic and $\geo\Ga$ is equivariantly homeomorphic to $\La$. \(v) $\Ga$ is undistorted (i.e. its orbits in $X$ are quasi-isometrically embedded). In higher rank, the situation is more complicated. First of all, the action $G\acts\geo X$ is not transitive any more. There is a partition of $\geo X$ into subsets each of which is ($G$-equivariantly) homeomorphic to the product of an open disk (open face of the Weyl chamber) with a (partial) flag variety, i.e. a quotient of $G/P$ of $G$ by a parabolic subgroup. Individual $G$-orbits $G\xi\subset\geo X$ are equivariantly diffeomorphic to flag manifolds $G/P$ and the action $G\acts G/P$ has no convergence dynamics any more; the convergence property is replaced by the weaker property that every sequence $g_n\to\infty$ in $G$ subconverges on the complement of a proper subvariety locally uniformly to a map into a proper subvariety. The $\Ga$-action on $\geo X$ still defines the dynamical decomposition of $\geo X$ into an open wandering set and a compact nonwandering set, but its relation to the limit set $\La\subset \geo X$ of $\Gamma$ is now more complicated. In general, there are nonwandering points outside the limit set, e.g. fixed points of isometries $\ga\in\Ga$ which are neither attractive nor repulsive. Moreover, the $\Ga$-action on the wandering set is in general not properly discontinuous and there is no unique maximal domain of discontinuity. Nevertheless, it remains true that “the further away from the limit set, the more discrete is the action”. The distance from the limit set can be measured, for instance, in terms of the Tits metric. We prove by a brief direct argument for any discrete subgroup: \[Theorem \[thm:pdgen\]\] \[thm:pdgenintro\] A discrete subgroup $\Ga\subset G$ acts properly discontinuously on the complement $\geo X-\bar N_{\pihalf}(\La)$ of the closed $\pihalf$-neighborhood of $\La$. However, this quick argument, which applies to any locally compact CAT(0) space, does not give an optimal result, as it says nothing about points within distance $\pihalf$ from $\La$; in particular, $\geo X-\bar N_{\pihalf}(\La)$ is in general not a maximal domain of discontinuity. Our next goal is to refine this theorem by taking into account the building structure of $\geo X$. We study the dynamical decomposition of the $\Ga$-action on all partial flag manifolds $G/P$ and the cocompactness on domains of discontinuity in them. In order to keep the discussion simpler, we restrict ourselves to the case when $\Ga$ is a [*regular*]{} discrete subgroup. More general cases with weaker regularity assumptions will be treated in a later paper. Regularity is an [*asymptotic*]{} property. It is a relaxation of [*uniform regularity*]{} (see definitions in section \[sec:reg\]); the latter requires that all limit points of $\Gamma$ are regular, i.e. interior points of Weyl chambers. Regular subgroups have a well-defined [*chamber limit set*]{} in the Fürstenberg boundary of $X$, $\La_{ch}\subset\D_F X\cong G/B$, where $B\subset G$ denotes the Borel subgroup; in the uniformly regular case, $\La_{ch}$ consists of the chambers at infinity intersecting the limit set. We first discuss subsets of $G/P$ with [*chaotic*]{} dynamics. Two points in distinct $\Ga$-orbits are called dynamically related if their orbits cannot be separated by invariant open sets; points in the same orbit are dynamically related if they are nonwandering. An invariant open subset $\Om\subset G/P$ is [*a domain of discontinuity*]{} if and only if there are no dynamical relations between points in $\Om$. A basic result of this paper (Theorem \[thm:dynrel\]) is a precise description of the dynamical relations on $\geo X$. Whether two points in a $G$-orbit are dynamically related depends on their [*position relative to the chamber limit set*]{}. The [*relative position*]{} $\pos(\xi,\si)$ of an ideal point $\xi\in G\xi\cong G/P$ with respect to a Weyl chamber $\si$, is a refinement of the Tits distance, taking into account the building structure of $\geo X$. In order to state our results, we have to first discuss the stratifications of the flag manifolds $G/P$ into Schubert cells determined by the function $\pos(\cdot, \cdot)$. The function $\pos(\xi,\si)$ takes values in the (finite) quotient $W^P=W/W_P$, where $W_P$ is the Weyl group of the Levi factor of $P$ and $W$ is the Weyl group of $G$. For fixed $\si$, the level sets of $\pos(\cdot ,\si)$ in $G/P$ are the [*Schubert cells*]{}. The set $W^P$ has a natural partial order which corresponds to the inclusion order on the closures of Schubert cells, the [*Schubert cycles*]{}. Every coset $wW_P\in W^P$ determines a sublevel in this order, corresponding to a Schubert cycle. A union of sublevels is called a [*thickening*]{} $\Th$. Such a thickening, in turn, determines a [*thickening*]{} $\Th(\La_{ch})\subset G/P$ of the chamber limit set, which the reader can regard as a “tubular neighborhood” of $\La_{ch}$: $$\Th(\La_{ch})= \{\zeta\in G/P: \exists \si\in \La_{ch} \hbox{~so that~} \pos(\zeta, \si)\in \Th\}.$$ The larger the relative position of $\xi$ with respect to $\La_{ch}$, the more “generic” $\xi$ is with respect to $\La_{ch}$. The longest element $w_0\in W$ induces an order reversing involution on relative positions. We call a thickening $\Th$ [*fat*]{} if $\Th\cup w_0\Th=W^P$, [*slim*]{} if $\Th\cap w_0\Th=\emptyset$, and [*balanced*]{} if it is both fat and slim. The trivial example of a slim thickening is given by the singleton $\Th=W_P$. Not every partial flag manifold admits a balanced thickening, but many of them do, including all full flag manifolds: \[See Lemmata \[lem:balthreg\] and \[lem:-1\].\] \[prop:exbal\] A balanced thickening always exists on the full flag manifold $\D_FX\cong G/B$. It also exists on all partial flag manifolds provided that the Weyl group $W$ contains $-1$, i.e. when all irreducible components of the symmetric space are of the type $A_1, B_\ell, C_\ell$, $D_{2k}, E_{7,8}$, $F_4$ or $G_2$. Furthermore, in all these cases the balanced thickening can be chosen to be a [*metric thickening*]{}, namely the $\frac{\pi}{2}$-metric thickening associated with a regular (rational) point in the positive chamber.[^1] Our first main result concerns the proper discontinuity of actions: For a regular discrete subgroup $\Ga\subset G$, the complement of a fat thickening of the chamber limit set is a domain of discontinuity for $\Ga$ on $G/P$. To obtain cocompactness of $\Ga$-actions at infinity, we need to impose an asymptotic condition on $\Gamma$-orbits in $X$; this asymptotic condition is a higher rank generalization of the notion of [*conicality*]{} of limit points in the rank one case and applies to regular discrete subgroups $\Ga$. \[dfn:chconicalintro\] We call a limit [*chamber*]{} $\si\in\La_{ch}$ [*conical*]{} if some (any) $\Ga$-orbit in $X$ has infinite intersection with any sufficiently large tubular neighborhood of a euclidean Weyl chamber asymptotic to $\si$; and we call the subgroup $\Ga$ [*chamber conical*]{} if all its limit chambers are conical. We note that it would be too restrictive to require limit [*points*]{} in $\La$ to be conical by working with tubular neighborhoods of rays, as in rank one. A quick argument using Dirichlet domains at infinity yields the following cocompactness result for actions of uniformly regular subgroups $\Ga$ on $G$-orbits of root type (see Definition \[def:rt\]), which already covers many interesting cases: \[Theorem \[thm:coco\].\] \[thm:cocointro\] If $\Ga\subset G$ is a uniformly regular discrete subgroup with conical chamber limit set and if the $G$-orbit $G\xi\cong G/P$ is of root type, then $$G\xi - N_{\pihalf}(\La)= G\xi - \bar{N}_{\pihalf}(\La)$$ is a cocompact domain for $\Ga$ on $G\xi$, and a domain of discontinuity by Theorem \[thm:pdgenintro\]. One shows that the region covered by the compact Dirichlet domains at infinity lies in between the sets $G\xi - \bar{N}_{\pihalf}(\La)$ and $G\xi - N_{\pihalf}(\La)$. Since $\La$ is regular due to the uniform regularity of $\Ga$, and since $G\xi$ has root type, there are no points in $G\xi$ at distance precisely $\pihalf$ from a limit point, and the assertion follows. In section \[sec:expand\] we obtain general cocompactness results for regular discrete subgroups $\Ga$ on arbitrary partial flag manifolds. The following class of discrete subgroups appears naturally: We call the subgroup $\Ga\subset G$ [*antipodal*]{} if any two distinct limit chambers are opposite; we call it [*RCA*]{} if in addition it is chamber conical. Examples of RCA subgroups are given by Schottky groups (see [@Benoist] and Remark \[rem:schott\]) and surface group representations in the Hitchin component [@Labourie]. We do not know examples of chamber conical regular subgroups which fail to be antipodal. Note that regularity and antipodality are automatic in rank one. Our second main result concerns the cocompactness of actions: \[Theorem \[thm:cocoslim\]\] For an RCA subgroup $\Ga\subset G$, the complement of a slim thickening of the chamber limit set is a cocompact domain for $\Ga$ on $G/P$. Our argument is based on the observation that chamber conicality implies that the action of $\Ga$ on the Fürstenberg boundary $\D_FX$ is expanding at $\La_{ch}$ in the sense of Sullivan; this is our generalization of Sullivan’s interpretation of convex cocompactness of Kleinian groups in terms of their dynamics near the limit set in [@Sullivan]. Expansion yields the cocompactness of the $\Gamma$-action on the complement of $\La_{ch}$ in $\D_FX$ and, more generally, on the complements of its projection to all partial flag manifolds. To verify cocompactness on the complement of any slim thickening in a partial flag manifold, we use a relative version of expansion, [*transverse*]{} to the natural fibrations of these thickenings over $\La_{ch}$. Such fibrations exist under the assumptions of the theorem, because slimness and antipodality imply that the thickenings of individual limit chambers are disjoint. Note that we cannot expect (absolut) expansion of the $\Ga$-action at thickenings of $\La_{ch}$ because these contain in general saddle fixed points of elements in $\Ga$. Combining both results we obtain: \[Theorem \[thm:cocodiscont\].\] \[thm:cocodiscontintro\] For an RCA subgroup $\Ga\subset G$, the complement of a balanced thickening of the chamber limit set is a cocompact domain of discontinuity for $\Ga$ on $G/P$. In particular, this yields cocompact domains of discontinuity in $\D_FX\cong G/B$, compare Proposition \[prop:exbal\]. On the other hand, it can happen that some $G$-orbits contain no properly discontinuous and cocompact domain at all, even though $\Ga$ is uniformly regular, preserves a rank one totally geodesic subspace and is convex cocompact. (Which implies chamber conicality and antipodality.) This also means that in higher rank it is too much to ask for cocompact domains of discontinuity of the form $X\cup\Om\subset\bar X$. The nonemptyness of domains is an issue. If for some subgroup $\Ga$ all domains given by Theorem \[thm:cocodiscont\] were empty, it would follow that the Tits boundary of $X$ admits a packing by $\pihalf$-balls (Proposition \[prop:neccdd\]). However, in the irreducible higher rank case the existence of such packings can be easily ruled out for most Weyl groups (Theorem \[thm:nopackmost\]), and one concludes: \[Theorem \[thm:nedoman\]\] If $X$ is irreducible of higher rank with Weyl group $\neq B_2,G_2$ and if $\Ga\subset G$ is an RCA subgroup, then some of the cocompact domains of discontinuity for $\Ga$ in $\D_FX$ provided by Theorem \[thm:cocodiscont\] are nonempty. There is overlap of our results with the work of Guichard and Wienhard [@GuichardWienhard_dd]. They found cocompact domains of discontinuity in $G$-homogeneous spaces closely related to $\D_FX\cong G/B$ (bundles with compact fiber) under the assumption that $\Ga$ is Anosov [@GuichardWienhard_dd Thm. 1.9]. The general definition of Anosov groups only requires a weaker version of regularity. On the other hand, regular Anosov groups are RCA. (One can also show the converse.) \[GIT\] The domain of properness and cocompactness for the action $\Ga \acts G/P$ in Theorem \[thm:cocodiscontintro\] depends on the choice of an auxiliary datum, namely, a thickening $\Th\subset W^P$. This brings us to one of the sources of inspiration for this paper, Geometric Invariant Theory (GIT), which was developed by Mumford in the context of actions $H\acts V$ of reductive algebraic groups on projective varieties (see [@Mumford]). The central issue of GIT is to identify the quotient projective variety $V//H$ (initially defined in terms of its ring of regular functions) as a [*geometric quotient*]{} of a subset $V_{ss}$ of semistable points in $V$ by the [*extended orbit equivalence*]{} relation: Two orbits are equivalent if their closures in $V_{ss}$ have nonempty intersection. From the viewpoint of classical topology, the resulting [*Mumford quotient*]{} accomplishes two, seemingly contradictory, tasks: It is both Hausdorff and compact. The apparent contradiction comes from the fact that if we remove too little from $V$ to define $V_{ss}$ then the quotient will not be Hausdorff, while if we remove too much then the quotient will not be compact. The set of semistable points $V_{ss}$ contains an open subset $V_{st}$ of stable points on which the $H$-action is proper. The Mumford quotient coincides with the quotient defined in terms of the usual orbit equivalence provided that $V_{st}=V_{ss}$: Every semistable point is also stable. The [*stability*]{} is not an intrinsic notion for the dynamics of the group $H$, it also depends on a certain auxiliary datum, which appears in the form of a positive line bundle over $V$. Our notion of [*thickening*]{} is a generalization of such line bundles. The [*balanced thickening*]{} case corresponds to the “stable=semistable” situation in GIT, as does the choice of root type $G$-orbit in Theorem \[thm:cocointro\]. In fact, there is not only an analogy with GIT, but also some overlap: If $\Ga$ preserves a totally geodesic subspace $Y\subset X$ and acts cocompactly on it, then $\Ga$ is a uniform lattice in a semisimple subgroup $H\subset G$ and the dynamical properties of the action $\Ga\acts G\xi$ are closely related to those of the action $H\acts G\xi$. The latter action is within the algebro-geometric GIT framework, at least if the groups $G$ and $H$ are complex. In this case, the type of stability of an ideal point can be expressed geometrically in terms of its Tits distance from the limit set, cf. [@KLM] and section \[sec:stab\]: It is stable if the distance is $>\pihalf$, semistable if it is $\geq\pihalf$ and unstable if it is $<\pihalf$. This is closely related to Theorems \[thm:pdgenintro\] and \[thm:cocointro\]. We now turn to the second objective and set of results of our paper, namely, comparing properties related to [*geometric finiteness*]{}. The strength of the notion of geometric finiteness for discrete subgroups of rank one Lie groups is that it can be reformulated in a variety of ways: Geometric, topological and dynamical (we review these formulations in section \[sec:rank1\]). Furthermore, such groups enjoy a number of nice properties which general discrete groups are lacking. Some of these reformulations and properties remain meaningful in the higher rank case. However, not all of them remain equivalent; some become stronger than others. Even though, currently, we do not have a theory of geometric finiteness as comprehensive as it exists in rank one, we are able to compare several properties and establish the equivalence of some of them. For rank one symmetric spaces, geometric finiteness is closely related to [*convex cocompactness*]{}; they are even equivalent in the absence of parabolic elements. It was shown by Kleiner and Leeb [@convcoco] that convex cocompactness is way too restrictive in higher rank, as it is satisfied only by few actions. On the other hand, convex cocompact actions on rank one symmetric spaces are characterized by the property that the orbit map defines a quasi-isometric embedding of the discrete group into the symmetric space. In the context of higher rank symmetric spaces, this [*undistortion*]{} assumption on the action of $\Ga$ on $X$ is way too weak; for instance, undistorted subgroups can fail to be finitely presented. (Note, however, that the only such examples that we currently have are not regular, see Example \[ex:stallings\].) Thus, one is forced to look for alternative generalizations of geometric finiteness in between. In this paper, we will consider, in the context of regular discrete subgroups $\Ga\subset G$, four notions generalizing geometric finiteness to higher rank, see the definitions given below: 1\. RCA subgroups 2\. Antipodal subgroups expanding at $\La_{ch}(\Ga)$ 3\. Asymptotically embedded subgroups 4\. Subgroups with nice asymptotics The notions 1–4 in the case of discrete subgroups of rank one Lie groups are well-known to be equivalent. We will see that the four notions are also equivalent to each other in the higher rank case, and it can be shown that they are in turn equivalent (under the regularity assumption) to the notion of [*Anosov representations*]{} introduced by Labourie [@Labourie] and developed further by Guichard and Wienhard [@GuichardWienhard_dd]. \(i) An advantage of the notion of conicality over the notions of [*asymptotic embeddedness*]{}, [*nice asymptotics*]{} or [*Anosov*]{} is that it does not a priori assume that the subgroup is word hyperbolic as an abstract group. For example, uniform lattices in semisimple Lie groups satisfy this property. Thus, the conicality condition could serve as a basis for a general theory of geometric finiteness in higher rank which includes groups which are not (relatively) hyperbolic. \(ii) The notion of [*expanding actions*]{} appears to be a variation on the dynamical part of the notion of Anosov representation, which avoids the language of expansion/contraction for flows on bundles appearing in [@Labourie; @GuichardWienhard_dd]. We define expansion directly in terms of the dynamics of $\Ga$ on the sphere at infinity. Furthermore, also the notion of expansion makes no a priori assumption on the hyperbolicity of $\Ga$. \[defn:hyperbolic\] We call an antipodal regular discrete subgroup $\Ga\subset G$ [*asymptotically embedded*]{} if $\Gamma$ is word hyperbolic and there exists an equivariant homeomorphism $$\label{eq:mapalpha} \alpha: \geo \Gamma \to \La_{ch} \subset \partial_F X$$ between the Gromov boundary of $\Gamma$ and the [*chamber limit set*]{} of $\Gamma$ in the Fürstenberg boundary $\partial_F X=G/B$ of $X$. \[dfn:nicas\] We say that a regular discrete subgroup $\Ga\subset G$ has [*nice asymptotics*]{} if $\Gamma$ is word hyperbolic and there exists an equivariant continuous embedding $$\label{eq:mapbeta} \beta: \geo \Gamma \to \partial_F X.$$ which maps different points to opposite chambers. Note that “nice asymptotics” is a priori weaker than “asymptotically embedded”. \[thm:equivalence\] For non-elementary regular discrete subgroups $\Ga\subset G$ the following properties are equivalent: 1\. $\Ga$ is RCA. 2\. $\Ga$ is antipodal and the action $\Ga\acts\D_FX$ is expanding at $\La_{ch}$. 3\. $\Ga$ is word hyperbolic and asymptotically embedded. 4\. $\Ga$ is word hyperbolic and has nice asymptotics in $G$. They imply that $\Ga$ is uniformly regular. The equivalence of expansion and asymptotic embeddedness might shed some light on Gromov’s cryptic remarks in [@Gromov p. 255] on expansion properties of discrete groups actions on ideal boundaries of higher rank symmetric spaces. Note that the maps $\al$ and $\beta$ in (\[eq:mapalpha\]) and (\[eq:mapbeta\]) can in general be different as there may be several choices for $\beta$, see Example \[ex:alnebe\]. However, we can show that they must agree if the subgroup $\Ga$ is Zariski dense in $G$. Part of the proof of Theorem \[thm:equivalence\] is an investigation of the coarse extrinsic geometry of RCA subgroups (see section \[sec:reghypnas\]). One calls a finitely generated subgroup $\Ga\subset G$ [*undistorted*]{} if the orbit maps $\Ga\to X$ are quasi-isometric embeddings. We show: \[Theorem \[thm:qiemb\].\] RCA subgroups are undistorted. Furthermore, the orbit map $\Ga\to X$ extends continuously, with respect to the visual topology on the Fürstenberg bordification $X\cup \D_F X$, to a homeomorphism $\alpha: \geo \Ga\to \La_{ch}(\Ga)$ (as in Definition \[defn:hyperbolic\]). More precisely, this extension is continuous even with respect to the topology of radial convergence on $\Ga \cup \geo \Ga$ and the topology of [*chamber-radial convergence*]{} on $X\cup \D_F X$. (This follows from the property that quasirays in $\Ga$ are mapped uniformly close to Weyl sectors in $X$.) Note that it was proven in [@GuichardWienhard_dd Thm. 5.3] that Anosov actions (under weaker regularity assumptions) are undistorted. Our argument avoids using the geodesic flow of $\Ga$. In section \[sec:dyn prelim\] we discuss several standard concepts of topological dynamics, namely, dynamical relation, wandering sets, convergence actions, conical limit points. We also show that [*transverse expansion*]{} implies cocompactness, which is one of the key technical results allowing us to establish cocompactness (at infinity) of RCA actions. In section \[sec:prelim\] we review basic definitions from the theory of symmetric spaces and spherical buildings, including Coxeter complexes, chamber-valued distance functions and Busemann functions. In section \[sec:geomsymmsp\] we introduce various geometric and combinatorial concepts dealing with spherical buildings, most importantly, the notion of [*relative position at infinity*]{}, various types of [*thickenings*]{} and establish sufficient conditions for the existence of balanced thickenings on $G$-orbits. In section \[sec:rank1\] we review notions of geometric finiteness for discrete isometry groups of rank one symmetric spaces. Section \[sec:domdisc\] and the following section \[sec:dynrelation\] are motivated by the problem of finding sufficient conditions for proper discontinuity of discrete group actions on subsets of the sphere at infinity. The key difference between the sections is that section \[sec:domdisc\] uses the ordinary limit set and the Tits distance, while section \[sec:dynrelation\] uses instead the chamber limit set and its thickenings, yielding better results. In section \[sec:domdisc\] we introduce limit sets and domains of discontinuity and establish first results on the (Tits) distance between non-wandering points and the limit set. In section \[sec:dynrelation\] we define regularity and uniform regularity for sequences of isometries of a symmetric space. We then sharpen the results of section \[sec:domdisc\] and establish a connection between the $\Ga$-dynamics at infinity and relative positions of ideal points with respect to the chamber limit set. In particular, we show that the $\Ga$-action outside of a fat thickening of the chamber limit set is properly discontinuous. Section \[sec:cocompact\] deals with the cocompactness problem of $\Ga$-actions on domains in partial flag manifolds. We introduce Dirichlet domains at infinity, and prove Theorems \[thm:coco\] and \[thm:cocodiscont\] which establish cocompactness of RCA actions on domains of discontinuity in some flag manifolds. In the process, we show that RCA actions are relatively expanding, which is a key tool in proving cocompactness in Theorem \[thm:cocodiscont\]. We also define the concepts of [*stability, semistability*]{} and [*unstability*]{} for $\Ga$-orbits in flag manifolds, imitating the notions coming from Geometric Invariant Theory, and see how these relate to fundamental domains at infinity. In section \[sec:ccddne\], we address the nonemptyness of cocompact domains of discontinuity. In section \[sec:hyp\] we prove Theorems \[thm:equivalence\] and \[thm:qiemb\] establishing the equivalence of four notions related to geometric finiteness in higher rank, and discuss the coarse extrinsic geometry of RCA subgroups. The first author was supported by NSF grants DMS-09-05802 and DMS-12-05312. The last author was supported by grants Mineco MTM2012-34834, AGAUR SGR2009-1207, and Icrea Acadèmia Award 2008. Topological dynamics preliminaries {#sec:dyn prelim} ================================== In this section we collect various definitions and results from topological dynamics; most of them are rather standard but some are new. Throughout this section, we let $\Ga$ be a discrete group, i.e. a group equipped with the discrete topology. We say that a sequence $(\ga_n)$ of elements of the discrete group $\Ga$ [*diverges to infinity*]{}, $\ga_n\to\infty$, if the map $\N \to \Ga, n\mapsto \ga_n$ is proper. We consider continuous actions $\Ga \acts Z$ on compact metric spaces $(Z,d)$. Dynamical relation, wandering set and domains of discontinuity {#sec:topdyn} -------------------------------------------------------------- We call two points $\xi,\xi'\in Z$ [*dynamically related*]{} with respect to the $\Ga$-action if there exist sequences $\xi_n\to\xi$ and $\ga_n\to\infty$ such that $\ga_n\xi_n\to\xi'$. We say then more specifically that $\xi$ is dynamically related to $\xi'$ with respect to the diverging sequence $(\ga_n)$ in $\Ga$. One then verifies that: 1\. Dynamical relation is a closed relation in $Z\times Z$. 2\. Points in different $\Ga$-orbits are dynamically related if and only if their orbits cannot be separated by disjoint $\Ga$-invariant open subsets. A point $\xi\in Z$ is called [*wandering*]{} with respect to the $\Ga$-action if it has a neighborhood $U$ such that $\{\ga\in\Ga|U\cap\ga U\neq\emptyset\}$ is finite. Thus, a point is nonwandering if and only if it is dynamically related to itself. Nonwandering points are called [*recurrent*]{}. Again, we call a point $\xi$, more specifically, recurrent with respect to a diverging sequence $(\ga_n)$ in $\Ga$ if for every neighborhood $U$ of $\xi$ we have $U\cap\ga_nU\neq\emptyset$ for sufficiently large $n$. The wandering set $\Om_{wand}(\Ga)\subset Z$ (i.e. the set of wandering points) is open and the $\Ga$-orbits in $\Om_{wand}(\Ga)$ are closed. Let a subset $\Om\subset Z$ be open and $\Ga$-invariant. The action $\Ga\acts\Om$ is called [*properly discontinuous*]{} if and only if $\{\ga\in\Ga|K\cap\ga K\neq\emptyset\}$ is finite for every compact subset $K\subset\Om$. We then call $\Om$ [*a domain of discontinuity*]{} for $\Ga$. This is equivalent to the property that no two points in $\Om$ are dynamically related. In particular, $\Om\subseteq\Om_{wand}(\Ga)$. The orbit space $\Om/\Ga$ is then Hausdorff. In general, there is no unique maximal domain of discontinuity. \[ex:nonundd\] Consider the infinite cyclic group $\Ga\cong\Z$ acting projectively on $Z=\R P^2$, so that the generator $\ga$ of $\Ga$ acts as the projectivization of a diagonal matrix with distinct positive eigenvalues $\la_1>\la_2>\la_3$. Let $e_1, e_2, e_3\in Z$ be the three fixed points of $\ga$ (eigenspaces for $\la_1, \la_2, \la_3$ respectively). Let $E_{ij}\subset Z$ denote the projective lines spanned by $e_i$ and $e_j$, $i< j$. Then both sets $U_1=Z-(E_{23} \cup e_1)$ and $U_3=Z-(E_{12}\cup e_3)$ are maximal domains of discontinuity for $\Gamma$. (The maximality follows from the fact that the points on $E_{12}$ are dynamically related to the points on $E_{23}$.) Observe also that in this example both $U_1/\Ga$ and $U_3/\Ga$ are compact. It is clear, that in order to find a domain of discontinuity for the action of $\Ga$, we need to remove all recurrent points from $Z$. The key issue is what else has to be removed in order to find “nice” domains of discontinuity. The main challenge for us (as in the case of Geometric Invariant Theory) is to find conditions on $\Ga$-actions under which we can remove “just the right” subset of $Z$ (containing the recurrent set), so that the quotient of the rest by $\Ga$ is compact and Hausdorff. Expanding and transversely expanding actions {#sec:trexpand} -------------------------------------------- The following notion due to Sullivan [@Sullivan §9] will be of basic importance to us: We say that the action $\Ga\acts Z$ is [*expanding*]{} at the [*point*]{} $z\in Z$ if there exists an element $\ga\in\Ga$ which is [*uniformly expanding*]{} on a neighborhood $U$ of $z$, i.e. for some constant $c>1$ and all points $z_1,z_2\in U$ we have $$d(\ga z_1,\ga z_2)\geq c\cdot d(z_1,z_2) .$$ We say that the action of $\Ga$ is [*expanding*]{} at a compact $\Ga$-invariant [*subset*]{} $E\subset Z$ if it is expanding at all points $z\in E$. In this paper we will also need the following more general notion of [*partial expansion*]{}. We suppose that the action $\Ga\acts Z$ has the following structure: There is a $\Ga$-invariant compact subset $E\subset Z$ and a continuous map $\pi: E\to\La$ onto a compact topological space $\La$ (e.g. a fiber bundle), such that the restricted action $\Ga\acts E$ is fiber preserving, i.e. it descends to a continuous action $\Ga\acts\La$. We set $E_{\la}:=\pi^{-1}(\la)$. In what follows, given a metric space $X$, a point $x\in X$ and a subset $Y\subset X$, we set $$d(x, Y):= \inf_{y\in Y} d(x,y).$$ We say that the action $\Ga\acts Z$ is [*expanding transversely to $\pi$ at the fiber $E_{\la}$*]{} if there exist an element $\ga\in\Ga$ and a neighborhood $U\subset Z$ of $E_{\la}$ such that for some constant $c>1$ we have $$\label{ineq:trexpan} d(\ga z,E_{\ga\la'})\geq c\cdot d(z,E_{\la'})$$ for all points $z\in U$ and fibers $E_{\la'}\subset U$. We say that the action $\Ga\acts Z$ is [*expanding at $E$ transversely to $\pi$*]{} if it is expanding at all fibers $E_{\la}$. The concept of expansion is important to us due to its connection to the dynamical characterization of hyperbolic groups (see section \[sec:conv actions\]) as well as to the following observation: \[prop:trexpcoco\] If the action $\Ga\acts Z$ is expanding at $E$ transversely to $\pi$, then the action $\Ga\acts Z-E$ is cocompact. We claim that for some constant $c>1$, $$\label{eq:maxprinctr} \sup d(\cdot,E)|_{\Ga z} > c\cdot d(z,E)$$ for all $z\in Z-E$ sufficiently close to $E$. Otherwise, there would exist a sequence $(z_n)$ in $Z-E$ accumulating at $E$ and a sequence of constants $c_n\to1$ such that $$d(\ga z_n,E)\leq c_n\cdot d(z_n,E)$$ for all $n\in\N$ and $\ga\in\Ga$. Since $E$ is compact, we may assume, after passing to a subsequence, that $(z_n)$ accumulates at a fiber, $z_n\to E_{\la}$. Due to expansion, there exists an element $\ga_{\la}\in\Ga$ which satisfies the expansion property (\[ineq:trexpan\]) on a neighborhood $U_{\la}\subset Z$ of $E_{\la}$ with some expansion factor $c_{\la}>1$. Let $E_{\ga_{\la}\la_n}$ be the fiber closest to $\ga_{\la}z_n$, $d(\ga_{\la}z_n,E_{\ga_{\la}\la_n})=d(\ga_{\la}z_n,E)$. Then $\la_n\to\la$. Since $z_n\in U_{\la}$ and $E_{\la_n}\subset U_{\la}$ for large $n$, it follows that $$c_{\la}\cdot d(z_n,E) \leq c_{\la}\cdot d(z_n,E_{\la_n}) \leq d(\ga_{\la}z_n,E_{\ga_{\la}\la_n}) = d(\ga_{\la}z_n,E)\leq c_n\cdot d(z_n,E) ,$$ a contradiction confirming our claim. Let $U\subset Z$ be an open tubular neighborhood of $E$ where (\[eq:maxprinctr\]) holds. Thus, no $\Ga$-orbit is entirely contained in $U-E$ and, therefore, every $\Ga$-orbit in $Z-E$ meets the compact subset $Z-U\subset Z-E$. Convergence actions {#sec:conv actions} ------------------- Let $Z$ be a compact metric space. We define the space $\Trip(Z)$ to be the subset of $Z^3$ consisting of triples of pairwise distinct points in $Z$. Every topological action $\Ga\acts Z$ induces a topological action $\Ga \acts\Trip(Z)$. The action $\Ga\acts Z$ is called a [*convergence action*]{} and the image of $\Ga$ in $Homeo(Z)$ is said to be a [*convergence group*]{} if one of the following equivalent conditions holds: \(i) The action $\Ga \acts\Trip(Z)$ is properly discontinuous. \(ii) For every sequence $\ga_n\to\infty$ in $\Ga$ there exist points $z_{\pm}\in Z$ and a subsequence of $(\ga_n)$ which converges to the constant map $\equiv z_+$ uniformly on compacts in $Z-\{z_-\}$. A convergence action $\Ga\acts Z$ is said to be [*uniform*]{} if the action $\Ga\acts\Trip(Z)$ is cocompact. A proof for the equivalence of both definitions can be found in [@Bowditch_config]. The main example of convergence actions comes from the following fact: Every discrete group $\Ga$ of isometries of a proper Gromov hyperbolic geodesic metric space $X$ acts as a convergence group on the Gromov boundary $\geo X$ of $X$. Furthermore, every word hyperbolic group $\Ga$ acts on its Gromov boundary as a uniform convergence group. Bowditch proved that, vice versa, this dynamical behavior characterizes the natural actions of word hyperbolic groups on their boundaries: \[thm:charhypbow\] Let $\Ga \acts Z$ be a uniform convergence action on a non-empty perfect compact metric space. Then $\Ga$ is word hyperbolic and $Z$ is equivariantly homeomorphic to $\geo\Ga$. The uniformity of a convergence action is in turn equivalent to all points being conical. The notion of conical limit point (e.g. of a Kleinian group) can be expressed purely in terms of the dynamics at infinity and, therefore, extends to the more general context considered here: \[dfn:intrcon\] Let $\Ga\acts Z$ be a convergence action. A point $z\in Z$ is called [*intrinsically conical*]{} if there exists a sequence $\ga_n\to\infty$ in $\Ga$ such that the sequence of points $\gamma_n^{-1}z$ converges and the sequence of maps $\gamma_n^{-1}|_{Z-\{z\}}$ converges (uniformly on compacta) to a constant map with value $\neq\lim_{n\to\infty}\gamma_n^{-1}z$. \[thm:bowditch-conical\] A convergence action $\Ga \acts Z$ on a perfect compact metric space $Z$ is uniform if and only if every point in $Z$ is intrinsically conical. The easy direction is that uniformity implies conicality. This can be seen as follows: Let $z'\neq z$ and $z''_n\to z$ be a sequence of points different from $z$. By uniformity, there exist elements $\ga_n\to\infty$ in $\Ga$ such that we have convergence $\ga_n^{-1}z\to z_{\infty}$, $\ga_n^{-1}z'\to z'_{\infty}$ and $\ga_n^{-1}z''_n\to z''_{\infty}$ with pairwise different limits. Since $\Ga\acts Z$ is a convergence action, we have convergence of $\ga_n^{-1}$ to a constant map on $Z-\{z\}$ or on $Z-\{z'\}$. The latter is impossible because the convergence is locally uniform and $\ga_n^{-1}z''_n\to z''_{\infty}\neq z_{\infty}$. Thus, the point $z$ is intrinsically conical. The following result connects expanding actions with Bowditch’s theorem. Note that if we equip the boundary of a word hyperbolic group $\Ga$ with a visual metric $d$, then the natural action $\Ga\acts(\geo\Ga,d)$ is expanding, see e.g. [@CP]. \[lem:conical\] If $\Gamma\acts Z$ is an expanding convergence action on a perfect compact metric space $Z$, then all points in $Z$ are intrinsically conical. We start with a general remark concerning expanding actions. For every point $z\in Z$ there exist an element $\ga\in\Ga$ and constants $r>0$ and $c>1$ such that $\ga$ is a $c$-expansion on the ball $B(z,r)$ and $\ga(B(z,r'))\supset B(\ga z,cr')$ for all radii $r'\leq r$. To see this, suppose that $c$ is a local expansion factor for $\ga$ at $z$ and, by contradiction, that there exist sequences of radii $r_n\to0$ and points $z_n\not\in B(z,r_n)$ such that $\ga z_n\in B(\ga z,cr_n)$. Then $z_n\to z$ due to the continuity of $\ga^{-1}$ and, for large $n$, we obtain a contradiction to the local $c$-expansion of $\ga$. Since $Z$ is compact, the constants $r$ and $c$ can be chosen uniformly. It follows by iterating expanding maps that for every point $z$ and every neighborhood $V$ of $z$ there exists $\ga\in\Ga$ such that $\ga(V)\supset B(\ga z,r)$, equivalently, $\ga(Z-V)\subset Z-B(\ga z,r)$. To verify that a point $z$ is intrinsically conical, let $V_n$ be a shrinking sequence of neighborhoods of $z$, $$\bigcap_{n} V_n= \{z\},$$ and let $\ga_n\in\Ga$ be elements such that $\ga_n^{-1}(Z-V_n)\subset Z-B(\ga_n^{-1}z,r)$. Since $V_n$ is shrinking and $\ga_n^{-1}(V_n)\supset B(\ga_n^{-1}z,r)$ contains balls of uniform radius $r$, it follows that the $\ga_n^{-1}$ do not subconverge uniformly on any neighborhood of $z$; here we use that $Z$ is perfect. In particular, $\ga_n\to\infty$. The convergence action property implies that, after passing to a subsequence, the $\ga_n^{-1}$ must converge locally uniformly on $Z-\{z\}$. Moreover, we can assume that the sequence of points $\ga_n^{-1}z$ converges. By construction, its limit will be different (by distance $\geq r$) from the limit of the sequence of maps $\ga_n^{-1}|_{Z-\{z\}}$. Hence the point $z$ is intrinsically conical. Geometric preliminaries {#sec:prelim} ======================= In this section we collect some standard material on Coxeter complexes, the geometry of nonpositively curved symmetric spaces and associated spherical Tits buildings; we refer the reader to [@qirigid] and [@habil] for more details. Coxeter complexes {#sec:cox} ----------------- A [*spherical Coxeter complex*]{} is a pair $(S,W)$ consisting of a unit sphere $S$ in a Euclidean space and a finite group $W$ which acts isometrically on $S$ and is generated by reflections at hyperplanes. We will use the notation $\angle$ for the angular metric on $S$. Throughout the paper, we assume that $W$ does not fix a point in $S$ and is associated with a root system $R$. Spherical Coxeter complexes will occur as [*model apartments*]{} of spherical buildings, mostly of Tits boundaries of symmetric spaces, and will in this context usually be denoted by $a_{mod}$. A [*wall*]{} $m_{\rho}$ in $S$ is the fixed point set of a hyperplane reflection $\rho$ in $W$. A point $\xi\in S$ is called [*singular*]{} if it belongs to a wall and [*regular*]{} otherwise. The action $W\acts S$ determines on $S$ a structure as a simplicial complex whose facets, called [*chambers*]{}, are the closures of the connected components of $$S-\bigcup_{\rho} m_{\rho}$$ where the union is taken over all reflections $\rho$ in $W$. We will refer to the simplices in this complex as [*faces*]{}. (If one allows fixed points for $W$ on $S$, then $S$ carries only a structure as a cell complex.) Codimension one faces of this complex are called [*panels*]{}. The chambers are fundamental domains for the action $W\acts S$. We define the [*spherical model Weyl chamber*]{} as the quotient $\si_{mod}=S/W$. The natural projection $\theta:S\to\si_{mod}$ restricts to an isometry on every chamber. An identification of $\si_{mod}$ with a chamber $\si\subset S$ determines a generating set of $W$, namely the reflections at the walls bounding $\si_{mod}$, and hence a word metric on $W$; the longest element with respect to this metric is denoted $w_0$. This element sends $\si_{mod}$ to the opposite chamber in $S$. We say that two points $\xi, \hat\xi\in S$ are [*Weyl antipodes*]{} if $\hat\xi=w_0 \xi$. We define the [*standard involution*]{} $\iota=\iota_S: S\to S$ as the composition $-w_0$. It preserves $\si_{mod}$ and equals the identity if and only if $-\id_S\in W$ because then $w_0=-\id_S$. A point $\xi$ in $S$ is called a [*root*]{} if the hemisphere centered at $\xi$ is simplicial, equivalently, bounded by a wall. Note that a root is almost never regular: Only the Coxeter complexes of type $A_1$ and $A_2$ (where $W\cong \Z_2$ and $S_3$) contain regular roots. This is because except in these two cases, every root is orthogonal to another root. We also note that for every irreducible root system of type $\neq A_n$ there exists a root, which is a vertex of the Coxeter complex. (E.g. the highest root with respect to a positive chamber has this property, see [@Bourbaki]). Suppose that $S$ is identified with the sphere at infinity of a Euclidean space $F$, $S\cong\geo F$, where $\geo F$ is equipped with the angular (Tits) metric. For a chamber $\si\subset S$ and a point $x\in F$ we define the [*Weyl sector*]{} $V(x, \si)\subset F$ as the cone with the tip $x$, that is, as the union of rays emanating from $x$ and asymptotic to $\si$. After fixing an origin $o\in F$ the group $W$ lifts to a group of isometries of $F$ fixing $o$. The sectors $V(o,\si)$ are then fundamental domains for the action of $W\acts F$. We define the [*euclidean model Weyl chamber*]{} as the quotient $V_{mod}=F/W$; we will also denote it by $\Delta$ or $\De_{euc}$. It is canonically isometric to the complete euclidean cone over $\si_{mod}$. The natural projection $F\to V_{mod}$ restricts to an isometry on the sector $V(o,\si)$ for every chamber $\si\subset S$. We define the [*$\Delta$-valued distance function*]{} or [*$\Delta$-distance*]{} $d_{\De}$ on $F$ by $$d_\Delta(x,y)= proj(y-x)\in\De$$ where $proj: F\to F/W\cong\Delta$ is the quotient map. Note the symmetry property $$\label{eq:symprop} d_{\De}(x,y)= \iota_{\De} d_{\De}(y,x)$$ where $\iota_{\De}$ induces the isometric involution of $\De$ induced by $\iota_S$. Symmetric spaces of noncompact type {#sec:symm} ----------------------------------- A symmetric space, denoted by $X$ throughout this paper, is said to be of [*noncompact type*]{} if it is nonpositively curved, simply-connected and has no Euclidean de Rham factor. We will identify $X$ with the quotient $G/K$ where $G$ is a semisimple Lie group acting isometrically and transitively on $X$, and $K$ is a maximal compact subgroup of $G$. We will assume that $G$ is commensurable with the isometry group $\Isom(X)$ in the sense that we allow finite kernel and cokernel for the natural map $G\to\Isom(X)$. In particular, the image of $G$ in $\Isom(X)$ contains the identity component $\Isom(X)_o$. The Lie group $G$ carries a natural structure as a real algebraic group. A [*point reflection*]{} at a point $x\in X$ is an isometry $\si_x$ which fixes $x$ and has differential $-\id_{T_xX}$ in $x$. In a symmetric space, point reflections exist in all points (by definition). A [*transvection*]{} of $X$ is an isometry which is the product $\si_x\si_{x'}$ of two point reflections; it preserves the oriented geodesic through $x$ and $x'$ and translates along it in the sense that it preserves parallel vector fields. A nontrivial isometry $\phi$ of $X$ is called [*axial*]{} if it preserves a geodesic $l$ and translates along it. The geodesic $l$ is called an [*axis*]{} of $\phi$. Axes are in general not unique, and they are parallel to each other. Any two points in $X$ are connected by a unique geodesic segment since $X$ is a Hadamard manifold. We will use the notation $xy$ for the oriented geodesic segment connecting $x$ to $y$. A [*flat*]{} in $X$ is a totally geodesic flat submanifold, equivalently, a convex subset isometric to a Euclidean space. A maximal flat in $X$ is a flat which is not contained in any larger flat; we will use the notation $F$ for maximal flats. The group $\Isom(X)_o$ acts transitively on the set of maximal flats; the common dimension of maximal flats is called the [*rank*]{} of $X$. The space $X$ has rank one if and only if it has strictly negative sectional curvature. A maximal flat $F$ is preserved by all transvections along geodesic lines contained in it. In general, there exist nontrivial isometries of $X$ fixing $F$ pointwise. The subgroup of isometries of $F$ which are induced by elements of $G$ is isomorphic to a semidirect product $W\ltimes \R^r$, where $r$ is the rank of $X$. The subgroup $\R^r$ acts simply transitively on $F$ by translations. The linear part $W$ is a finite reflection group, called the [*Weyl group*]{} of $G$ and $X$. Since maximal flats are equivalent modulo $G$, the action $W\acts F$ is well-defined up to isometric conjugacy. We will think of the Weyl group as acting on a [*model flat*]{} $F_{mod}\cong \R^r$ and on its ideal boundary sphere at infinity, the [*model apartment*]{} $a_{mod}=\tits F_{mod}\cong S^{r-1}$. The pair $(a_{mod},W)$ is the [*spherical Coxeter complex*]{} associated with $X$. The Weyl group is precisely the isometry group for the $\Delta$-valued distance on $F_{mod}$. We identify the spherical model Weyl chamber $\si_{mod}$ with a (fundamental) chamber in the model apartment, $\si_{mod}\subset a_{mod}$. Accordingly, we identify the [*euclidean model Weyl chamber*]{} $V_{mod}$ with the sector in $F_{mod}$ with tip in the origin and ideal boundary $\si_{mod}$, $V_{mod}\subset F_{mod}$. The $\Delta$-valued distance naturally extends from $F_{mod}$ to $X$ because every pair of points lies in a maximal flat. In order to define the distance $d_\Delta(x,y)$ of two points $x,y\in X$ one chooses a maximal flat $F$ containing $x,y$ and identifies it isometrically with $F_{mod}$ while preserving the types of points at infinity. The resulting quantity $d_\Delta(x,y)$ is independent of the choices. We refer the reader to [@KLM] for the detailed discussion of [*metric properties*]{} of $d_\Delta$. We will use the notation $\angle$ for the Riemannian angle on the tangent spaces $T_xX$ of $X$. Similarly, we will denote by $\angle_x(y,z)$ the angle between the geodesic segments $xy$ and $xz$ at the point $x$. The [*ideal*]{} or [*visual boundary*]{} of $X$, denoted $\geo X$, is the set of asymptote classes of geodesic rays in $X$, where two rays are [*asymptotic*]{} if and only if they have finite Hausdorff distance. Points in $\geo X$ are called [*ideal points*]{}. For $\xi\in \geo X$ and $x\in X$ we let $x\xi$ denote the geodesic ray emanating from $x$ and asymptotic to $\xi$, i.e. representing the ideal point $\xi$. For $x\in X$ we let $\Si_xX$ denote the [*space of directions*]{} of $X$ at $x$, i.e. the unit sphere in the tangent space $T_xX$. The ideal boundary $\geo X$ carries the natural [*Tits (angle) metric*]{} $\tangle$, defined as $$\tangle(\xi,\eta)= \sup_{x\in X} \angle_{x}(\xi,\eta)$$ where $\angle_{x}(\xi,\eta)$ is the angle between the geodesic rays $x\xi$ and $x\eta$. The supremum is attained on any flat in $X$ whose ideal boundary contains $\xi$ and $\eta$. We denote by $\tits X$ the metric space $(\geo X,\tangle)$. There are two natural topologies on $\geo X$: One is the [*visual topology*]{} (also known as the [*cone topology*]{}); with respect to it $\geo X$ is homeomorphic to the sphere of dimension $\dim(X)-1$. The second topology is the [*Tits topology*]{} induced by the Tits metric. The Tits metric is lower semicontinuous with respect to the visual topology, and the Tits topology is finer than the visual topology. For every maximal flat $F\subset X$ there is an identification $\geo F\cong a_{mod}$ of its ideal boundary with the model apartment, unique up to composition with elements in $W$. The Coxeter complex structure on $a_{mod}$ induces a simplicial structure on $\geo F$. The ideal boundaries of maximal flats cover $\geo X$ because every geodesic ray in $X$ is contained in a maximal flat. Moreover, their intersections are simplicial. One thus obtains a $G$-invariant piecewise spherical [*simplicial structure*]{} on $\geo X$ which makes $\geo X$ into a [*spherical building*]{} and, also taking into account the visual topology, into a topological spherical building. It is called the [*spherical*]{} or [*Tits building*]{} associated to $X$. The Tits metric is the path metric with respect to the piecewise spherical structure. We will refer to the simplices as [*faces*]{}. A point $\xi\in \geo X$ is called [*regular*]{}, respectively, [*singular*]{} if it is a regular, respectively, singular point of an apartment $a\subset \geo X$ containing $\xi$. We will use the notation $\geo^{reg}X$ and $\tits^{reg} X$ to denote the subsets of the ideal boundary of $X$ consisting of regular points. One topologizes $\bar{X}=X\cup \geo X$ as follows: Fix a base-point $x\in X$. A sequence $x_n$ converges to an ideal point $\xi\in \geo X$ if the sequence of geodesic segments $xx_n$ with base point $x$ converges to the ray $x\xi$ in the pointed Hausdorff topology. This topology on $\bar X$ is independent of $x$ and makes $\bar{X}$ into a compact space homeomorphic to a closed ball. We define the visual boundary of a subset $A\subset X$ as the set $\geo A=\bar A\cap\geo X$ of its accumulation points at infinity. The ideal boundaries $\geo F\subset\geo X$ of the maximal flats $F\subset X$ are precisely the [*apartments*]{} with respect to the spherical building structure at infinity, which in turn are precisely the convex subsets isometric to the unit $(r-1)$-sphere with respect to the Tits metric. Any two points in $\geo X$ lie in a common apartment. Two ideal points $\xi,\eta\in\geo X$ are called [*antipodal*]{} if $\tangle(\xi, \eta)=\pi$; equivalently, these points are antipodal in every apartment $a$ containing $\xi$ and $\eta$. Similarly, two chambers $\si_+,\si_-\subset\geo X$ are called [*opposite*]{} if they are opposite in an apartment containing them. This apartment is then unique and will be denoted by $a(\si_+,\si_-)$, and the unique maximal flat with ideal boundary $a(\si_+,\si_-)$ by $F(\si_+, \si_-)$. If $\si\subset\geo X$ is a chamber and $x\in X$ is a point, then there exists a unique (maximal) flat $F\subset X$ containing $x$ so that $\si\subset\geo F$. This defines a Weyl sector $V=V(x,\si)\subset X$ contained in $F$. Sometimes, we will call the space of directions $\Si_xV\subset \Si_xX$ a [*spherical chamber*]{} in $\Si_xX$, although we will not regard the unit sphere as a spherical building. \[def:wh\] We define the [*Weyl hull*]{} of a regular geodesic segment $x_1x_2$ in $X$ as the intersection $$WH(x_1,x_2)=V(x_1,\si_1)\cap V(x_2,\si_2)$$ of the Weyl sectors containing $x_1x_2$. Note that the sectors $V(x_i,\si_i)$ lie in the unique maximal flat containing $x_1x_2$. Alternatively, one can define the Weyl hull of $x_1x_2$ as the intersection of all closed halfspaces in this flat which contain $x_1x_2$ and are bounded by a singular hyperplane. (The latter definition generalizes to subsets of $X$ which are contained in some flat.) The stabilizers $B_{\si}\subset G$ of the chambers $\sigma\subset \geo X$ are the [*Borel subgroups*]{} of $G$. After identifying the model chamber with a chamber in $\geo X$, $\si_{mod}\subset\geo X$, we call $B=B_{\si_{mod}}$ the [*positive*]{} Borel subgroup. The group $G$ acts transitively on the set of chambers in $\geo X$, which we will then identify with $G/B$, the [*(generalized) full flag manifold*]{} of $G$. The Borel subgroups are algebraic subgroups of $G$, and $G/B$ is a real projective variety. The set $\D_FX\cong G/B$ of chambers in $\geo X$ is called the [*Fürstenberg boundary*]{} of $X$; we will equip it with the visual topology (as opposed to the Zariski topology coming from $G/B$) which coincides with its manifold topology as a compact homogeneous $G$-space. After picking an interior point $\xi \in\interior(\si_{mod})$, we can identify the $G$-orbit $G\xi\subset\geo X$ $G$-equivariantly and homeomorphically with $\D_FX$ by assigning to the (regular) point $g\xi$ the (unique) chamber $g\si_{mod}$ containing it. Similarly to the visual compactification $\bar{X}$ of $X$ above, one defines the [*(visual) Fürstenberg bordification*]{} $X\cup \D_F X$ as follows: A sequence of points $x_n\in X$ converges to a chamber $\si\in\D_FX$ whenever $x_n$ accumulates (at a compact subset) in the interior of $\si$ with respect to the visual topology on $\bar{X}$. The Fürstenberg bordification is no longer compact (unless $X$ has rank one). There is another useful topology on $X\cup \D_F X$: We say that a sequence of points $x_n\in X$ [*converges*]{} to a chamber $\si\in \D_F X$ in the [*chamber radial topology*]{} if for some (any) base point $x\in X$ the sequence of points $x_n$ is regular (see Definition \[def:regseq\]) and is contained in a tubular neighborhood of the Weyl sector $V(x,\si)$. In the context of rank one symmetric spaces (as well as hyperbolic groups) this is known as the [*topology of radial convergence*]{}, where a sequence of points $x_n\in X$ converges to an ideal point $\xi\in \geo X$ [*radially*]{} if the convergence takes place in a certain tubular neighborhood of a geodesic ray. Whereas the group $G$ acts transitively on chambers at infinity, the action $G\acts\geo X$ on ideal points is not transitive in higher rank. The quotient can be identified with the spherical model chamber, $\geo X/G\cong\si_{mod}$. We call the projection $$\theta:\geo X\to\geo X/G \cong\si_{mod}$$ the [*type*]{} map. It restricts to an isometry on every chamber $\si\subset\geo X$. We call the inverse $\kappa_{\si}=(\theta|_{\si})^{-1}: \si_{mod}\to\si$ the [*chamber chart*]{} for $\si$. We define the [*type*]{} of an ideal point $\xi\in\geo X$ as its image $\theta(\xi)\in\si_{mod}$. Thus, a point is regular if its type is an interior point of $\si_{mod}$. The [*type*]{} of a face $\tau\subset\geo X$ is defined as the face $\theta(\tau)\subset\si_{mod}$. \[def:rt\] An ideal point $\xi\in \geo X$ is said to be of [*root type*]{} if its projection $\theta(\xi)\in\si_{mod}\subset a_{mod}$ is a point of root type in the model apartment. We will call a $G$-orbit in $\geo X$ of [*root type*]{} if it consists if root type points. The stabilizers $P_{\tau}\subset G$ of faces $\tau\subset\geo X$ are the [*parabolic subgroups*]{} of $G$. The group $G$ acts transitively on faces of the same type. The set $\flag(\tau_{mod})\cong G/P_{\tau_{mod}}$ of the faces $\tau$ of type $\theta(\tau)=\tau_{mod}\subset\si_{mod}$ is called the [*(generalized partial) flag manifold*]{}. In particular, $\flag(\si_{mod})=\D_FX$. Again, we equip the flag manifolds with the visual topology; it agrees with their topology as compact homogeneous $G$-spaces. Busemann functions and horoballs {#sec:bus} -------------------------------- Let $\xi\in \geo X$ be an ideal point. For a unit speed geodesic ray $\rho:[0,+\infty)\to X$ asymptotic to $\xi$ one defines the [*Busemann function*]{} $b_\xi$ on $X$ as the uniform monotonic limit $$b_\xi(x)=\lim_{t\to+\infty} (d(x, \rho(t))-t).$$ Altering the ray $\rho$ changes $b_\xi$ by an additive constant. Along the ray holds $$b_\xi(\rho(t))=-t.$$ The Busemann function $b_\xi$ is convex 1-Lipschitz and measures the relative distance from $\xi$. The sublevel sets $$H=\{x\in X: b_\xi(x)\le c\}$$ are called (closed) [*horoballs centered at $\xi$*]{}. Horoballs are convex. Ideal boundaries of horoballs are closed $\pihalf$-balls at infinity, $$\geo H= \{\tangle(\xi, \cdot)\le \pi/2\}\subset\geo X .$$ Busemann functions are [*asymptotically linear*]{} along rays; if $\gamma:[0,+\infty)\to X$ is a unit speed geodesic ray asymptotic to $\eta\in\geo X$, then $$\lim_{t\to\infty} \frac{b_\xi(\gamma(t))}{t}=-\cos \tangle(\xi, \eta).$$ This limit is called the [*asymptotic slope*]{} of $b_\xi$ at $\eta$. In particular, rays asymptotic to $\eta$ enter horoballs centered at ideal points $\xi$ with $\tangle(\xi,\eta)<\pihalf$. Geometry of ideal boundaries {#sec:geomsymmsp} ============================ In this section we introduce definitions and prove some properties of symmetric spaces of noncompact type and their ideal boundaries of more specific nature which are needed for our study of discrete group actions at infinity. Stars at infinity {#sec:star} ----------------- \[dfn:star\] Let $A\subset\geo X$ be a subset. We define its [*closed star*]{} $\st(A)\subset\geo X$ as the union of all closed chambers intersecting $A$ nontrivially, and its [*open star*]{} $\ost(A)\subset\geo X$ as the union of all open faces whose closure intersects $A$ nontrivially. Note that if $A$ is closed, then $\st(A)$ is the closure of $\ost(A)$. For $G$-orbits at infinity we have $\st(G\xi)=\geo X$. The open star $\ost(G\xi)$ decomposes as the disjoint union of the open stars $\ost(g\xi)$ around the orbit points $g\xi$. These are, in fact, the connected components of $\ost(G\xi)$ with respect to the Tits topology. Note that $\ost(g\xi)$ consists of the points in $\geo X$ which are strictly closer (with respect to the Tits metric) to $g\xi$ than to any other point in the orbit $G\xi$, i.e. the nearest point projection $$\label{eq:fibost} \ost(G\xi)\to G\xi$$ is well-defined and $G$-equivariant. The inverse image of any subset of $A\subset G\xi$ under this projection is the open star of $A$. For a type $\bar\xi\in \si_{mod}$ we define $\ost(\bar\xi)\subset\si_{mod}$ analogously as the union of all open faces of $\si_{mod}$ whose closure contains $\bar\xi$. Then $\ost(\bar\xi)$ contains the open model chamber $\interior(\si_{mod})$, and for a regular type $\ost(\bar\xi)=\interior(\si_{mod})$. We observe that $$\ost(G\xi)=\theta^{-1}(\ost(\theta(\xi)))$$ In particular, $\ost(G\xi)$ contains the regular $G$-orbits. The projection (\[eq:fibost\]) restricts on every $G$-orbit $G\xi'$ contained in $\ost(G\xi)$ to a $G$-equivariant fibration with compact fibers $\ost(g\xi)\cap G\xi'=\st(g\xi)\cap G\xi'$. This projection is a diffeomorphism if and only if the types $\theta(\xi')$ and $\theta(\xi)$ are contained in the same open face of $\si_{mod}$. The folding order {#sec:foldord} ----------------- In this section we discuss a natural partial order on Weyl orbits in the model apartment and give different geometric definitions for this order. By a [*folding map*]{} $a_{mod}\to a_{mod}$ we mean a type preserving continuous map which sends chambers isometrically onto chambers. For distinct points $\bar\xi_1,\bar\xi_2\in a_{mod}$ we define $\bar\xi_1\prec\bar\xi_2$ if and only if there exists a folding map $a_{mod}\to a_{mod}$ fixing $\si_{mod}$ and mapping $\bar\xi_2\mapsto\bar\xi_1$. An isometry of $a_{mod}$ preserving $\si_{mod}$ preserves the folding order. A wall $m$ splits $a_{mod}$ into two hemispheres, the inner hemisphere $h^+$ containing $\si_{mod}$ and the outer hemisphere $h^-$. This decomposition gives rise to the folding map which fixes $h^+$ and reflects $h^-$ onto it. We call a composition of such folding maps at walls $m_i$ a [*special folding*]{}. The intersection $\cap_ih^+_i$ of inner hemispheres is fixed by the special folding. In particular, special foldings fix the model chamber $\si_{mod}$. In general, there are folding maps fixing $\si_{mod}$ which are not special. However, this makes no difference for the folding order: If for points $\bar\xi_1,\bar\xi_2\in a_{mod}$ there exists a folding map fixing $\si_{mod}$ and mapping $\bar\xi_2\mapsto\bar\xi_1$ then there exists a special folding with this property. We may assume that $\bar\xi_1$ and $\bar\xi_2$ are regular and different. We connect a point $\bar\eta$ in the interior of $\si_{mod}$ to $\bar\xi_2$ by a geodesic segment $\bar\ga$ which avoids faces of codimension at least two. Let $f$ be a folding map fixing $\si_{mod}$ with $f(\bar\xi_2)=\bar\xi_1$. Then $f\circ\bar\ga$ is a broken geodesic path which connects $\bar\eta$ to $\bar\xi_1$ and has the same length and initial direction as $\bar\ga$. Its bending points are interior points of panels and $\bar\ga$ is locally “reflected” at the walls containing these panels. The claim follows if we can replace $f\circ\bar\ga$ by a broken geodesic path from $\bar\eta$ to $\bar\xi_1$ of the same sort which is the image of $\bar\ga$ under a [*special*]{} folding. Let $\bar\eta_1$ denote the first bending point of $f\circ\bar\ga$ starting from $\bar\eta$. It lies in a wall $m_1$. If $f\circ\bar\ga$ crosses $m_1$ again in some point $\bar\eta_2$ then we replace the subpath $\bar\eta_1\bar\eta_2$ by its reflection at $m_1$. The modified broken path $\bar\ga'$ has again reflection folds, the same initial direction and the same endpoint. Moreover, its initial segment is strictly longer. After finitely many such modifications, we may assume that $\bar\ga'$ stays inside $h_1^+$. (The wall $m_1$ has changed in the process.) We then can obtain $\bar\ga'$ as the image of another broken path $\bar\ga''$ under the special folding $s_1$ at $m_1$, i.e. $\bar\ga'=s_1\circ\bar\ga''$, such that $\bar\ga''$ has a strictly longer initial segment than $\bar\ga'$. By induction it follows that $f\circ\bar\ga$ can be replaced by another broken path with the same endpoint and which is the image of $\bar\ga$ under a special folding. \[cor:altdeffold\] $\bar\xi_1\preceq\bar\xi_2$ if and only if there exists a special folding which maps $\bar\xi_2\mapsto\bar\xi_1$. We can regard the regular $W$-orbits as copies of $W$ by identifying the regular orbit point in $w\si_{mod}$ with the element $w\in W$. The corollary shows that under this identification the folding order coincides with the [*Bruhat order*]{}, see [@Humphreys ch. 5.9] or [@BjoernerBrenti ch. 2] for a definition. To verify this, one observes that if the chambers $w\si_{mod}$ and $w'\si_{mod}$ are symmetric with respect to a wall and if $w\si_{mod}$ lies in the inner hemisphere, then we have the inequality $l(w)<l(w')$ for word lengths. Here the word length on $W$ is defined using as generators the reflections at the walls of $\si_{mod}$. Recall that $w_0\in W$ denotes the involution sending $\si_{mod}$ to the opposite chamber $\hat\si_{mod}$ in $a_{mod}$, i.e. the longest element of $W$. \[lem:compposrev\] $w_0$ reverses the folding order. Suppose that the special folding $s_m$ at the wall $m$ maps $\bar\xi_2$ to $\bar\xi_1$, i.e. $s_m\bar\xi_2=\bar\xi_1$. When applying $w_0$, the inner hemisphere bounded by $m$ becomes the outer hemisphere bounded by $w_0m$ and vice versa, $w_0h^{\pm}_m=h^{\mp}_{w_0m}$. Hence $s_{w_0m}w_0\bar\xi_1=w_0\bar\xi_2$. The assertion follows by applying Corollary \[cor:altdeffold\] and induction. We provide now metric estimates for the folding order. For a type $\vartheta\in\si_{mod}$ the (angular) distance from $\vartheta$ defines partial orders $\prec_{\vartheta}$ on the Weyl orbits in $a_{mod}$; for different points $\bar\xi_1,\bar\xi_2$ in the same Weyl orbit, $W\bar\xi_1=W\bar\xi_2$, we put: $$\bar\xi_1\prec_{\vartheta}\bar\xi_2 \quad\Leftrightarrow\quad \tangle(\bar\xi_1,\vartheta)<\tangle(\bar\xi_2,\vartheta)$$ This partial order is a total order on a Weyl orbit if no two points in the orbit are equidistant from $\vartheta$. If $\vartheta$ is regular, then the orbit point in $\si_{mod}$ is a total minimum and the orbit point in $\hat\si_{mod}$ is a total maximum of the order $\prec_{\vartheta}$. Since folding maps are 1-Lipschitz, we have the following relation of partial orders: $$\label{eq:foldimplmet} \bar\xi_1\prec\bar\xi_2 \quad\Rightarrow\quad \bar\xi_1\prec_{\vartheta}\bar\xi_2$$ Note in particular, that for distinct points $\bar\xi_1$ and $\bar\xi_2$ we can have at most one of the inequalities $\bar\xi_1\prec\bar\xi_2$ and $\bar\xi_1\succ\bar\xi_2$. Relative position at infinity {#sec:relpos} ----------------------------- For an ideal point $\xi\in\geo X$ and a chamber $\si\subset\geo X$ there exists an apartment $a$ containing both. The apartment can be uniquely parametrized by an apartment chart $\al:a_{mod}\to a$ which restricts to the chamber chart of $\si$, i.e. $\al|_{\si_{mod}}=\kappa_{\si}$. We define the [*relative position*]{} of $\xi$ with respect to $\si$ as $$\pos(\xi,\si)=\xi'=\al^{-1}(\xi)\in a_{mod} .$$ The relative position is determined by the restricted distance function $\tangle(\xi,\cdot)|_{\si}$ and in particular does not depend on the choice of the apartment $a$. The projection $\pos(\cdot,\si):\geo X\to a_{mod}$ restricts to an isometry on every apartment containing $\si$. Two ideal points $\xi_1$ and $\xi_2$ have the same position relative $\si$ if and only if they lie in the same orbit of the Borel subgroup $B_{\si}=Stab_G(\si)$ fixing $\si$, $$\pos(\xi_1,\si)=\pos(\xi_2,\si) \Leftrightarrow B_{\si}\xi_1=B_{\si}\xi_2 .$$ The positions relative $\si$ thus correspond to the orbits of $B_{\si}$ and we have the identification $\geo X/B_{\si}\cong a_{mod}$. When restricting the position relative $\si$ to the $G$-orbit $G\xi$ with type $\theta(\xi)=\bar\xi\in\si_{mod}$ it takes values in the Weyl orbit $W\bar\xi\subset a_{mod}$. The Weyl orbit can be identified with the coset space $W/W_{\bar\xi}$ where $W_{\bar\xi}$ is the stabilizer of $\bar\xi$ in $W$. Thus, one may regard the relative position restricted to $G\xi$ as a function $\pos(\cdot,\si)|_{G\xi}\to W/W_{\bar\xi}$. The $G$-orbit $G\xi$ decomposes into finitely many $B_{\si}$-orbits called [*Schubert cells*]{}. The closure of a Schubert cell is called a [*Schubert cycle*]{} and is a union of Schubert cells. There is a unique dense open Schubert cell relative $\si$ in $G\xi$. When identifying the Fürstenberg boundary $\D_FX$ with a regular orbit, the [*open Schubert cell $C(\si)\subset\D_FX$ relative $\si$*]{} consists of the chambers opposite to $\si$. In section \[sec:nearbythickenings\], we will also need the following generalization of the notion of relative position. For an ideal point $\xi$ and a face $\tau\subset\geo X$ of type $\tau_{mod}\subset\si_{mod}$, the [*relative position*]{} $\pos(\xi,\tau)$ of $\xi$ with respect to $\tau$ is defined as an element in the quotient $a_{mod}/W_{\tau_{mod}}$ of the model apartment by the stabilizer $W_{\tau_{mod}}$ of $\tau_{mod}$ in $W$. Accordingly, the position $\pos(\tau',\tau)$ of a face $\tau'$ relative to another face $\tau$ (not necessarily of the same type) is defined as an element in the double coset $W_{\tau_{mod}}\backslash W/W_{\tau'_{mod}}$. The next result shows that the folding order on the set of relative positions defined in section \[sec:foldord\] coincides with the inclusion order on Schubert cycles. \[prop:relposint\] $\pos(\xi_1,\si)\preceq\pos(\xi_2,\si) \Leftrightarrow B_{\si}\xi_1\subseteq\ol{B_{\si}\xi_2}$ We denote $\bar\xi_i:=\pos(\xi_i,\si)$. Suppose first that $\xi_1\in\ol{B_{\si}\xi_2}$. Then there exist a sequence $b_n\in B_{\si}$ such that $b_n\xi_2\to\xi_1$. Let $a_n$ be apartments containing $\si$ and $b_n\xi_2$, and let $\al_n:a_{mod}\to a_n$ be the apartment charts which restrict to the chamber chart of $\si$, $\al_n|_{\si_{mod}}=\kappa_{\si}:\si_{mod}\to\si$. Then $\al_n(\bar\xi_2)=b_n\xi_2$. The isometric embeddings (with respect to the Tits metric) $a_n$ subconverge (with respect to the visual topology) to a limit map $\al:a_{mod}\to\geo X$. The map $\al$ is, in general, not an isometric embedding anymore, but only a folding map extending $\kappa_{\si}$. Here, by a [*folding map*]{} $a_{mod}\to\geo X$ we again mean a type preserving map which sends chambers isometrically onto chambers, compare section \[sec:foldord\]. The folding map $\al$ satisfies $\al(\bar\xi_2)=\xi_1$, and its composition with the natural projection $\geo X\to\geo X/B_{\si}\cong a_{mod}$ is a folding map $a_{mod}\to a_{mod}$ which maps $\bar\xi_2\mapsto\bar\xi_1$. Vice versa, suppose now that $\bar\xi_1\preceq\bar\xi_2$. By definition of the partial order there exists a folding map of $a_{mod}$ fixing $\si_{mod}$ and carrying $\bar\xi_2\mapsto\bar\xi_1$. Furthermore there is an isometric embedding $a_{mod}\to\geo X$ which extends the chamber chart $\kappa_{\si}$ and maps $\bar\xi_1\mapsto\xi_1$. By composition we obtain a folding map $\al:a_{mod}\to\geo X$ which extends $\kappa_{\si}$ and maps $\al(\bar\xi_2)=\xi_1$. We want to find a sequence of isometric embeddings $\al_n:a_{mod}\to\geo X$ extending $\kappa_{\si}$ such that $\al_n(\bar\xi_2)\to\al(\bar\xi_2)=\xi_1$. This will then imply that $\xi_1\in\ol{B_{\si}\xi_2}$. (Note that in general folding maps are not limits of isometric embeddings.) We may assume that the relative positions $\bar\xi_i$ are regular. (Otherwise, we may perturb them keeping the inequality $\bar\xi_1\preceq\bar\xi_2$ and perturb the $\xi_i$ accordingly.) We choose in $a_{mod}$ a geodesic $\bar\ga$ of length $\pi$ starting in an interior point $\bar\eta_0$ of $\si_{mod}$ and passing through $\bar\xi_2$ while avoiding simplices of codimension $\geq2$. It crosses successively a sequence (gallery) of chambers $\bar\si_0=\si_{mod},\bar\si_1,\dots,\bar\si_k=\hat\si_{mod}$ and intersects the intermediate panels $\bar\tau_i=\bar\si_i\cap\bar\si_{i-1}$ transversally in interior points $\bar\eta_i$. When applying the folding map $\al$, it may happen that successive chambers of the folded gallery coincide, i.e. that $\al(\bar\si_i)=\al(\bar\si_{i-1})$ for some $i$. (This happens if and only if $\al$ is not an isometric embedding.) One can arbitrarily well approximate (in the visual topology) the folded gallery by an embedded gallery with the same initial chamber $\si$. To obtain such approximations it is convenient to use the $G$-action as follows. If $\al(\bar\si_i)=\al(\bar\si_{i-1})$ then one may pick an element $g\in G$ close to the identity, which fixes $\al(\bar\tau_{i-1})$ and moves $\al(\bar\si_i)=\al(\bar\si_{i-1})$ away from itself, and apply it to the “tail” $\al(\bar\si_i),\dots,\al(\bar\si_k)$ of the gallery. Doing this inductively along the gallery, one obtains an arbitrarily good approximation of the folded gallery $\al(\bar\si_0)=\si,\dots,\al(\bar\si_k)$ by an embedded gallery $\si_0=\si,\dots,\si_k$, that is a sequence of chambers such that $\si_i\cap\si_{i-1}$ is precisely a panel for all $i$. This yields at the same time an approximation of the broken geodesic $\al(\bar\ga)$ in $\geo X$ by a true geodesic $\ga$ such that $\ga\cap\si_i$ and $\bar\ga\cap\bar\si_i$ are corresponding subsegments of the same type. Now we use the path $\bar\ga$ as a “guiding line” to extend the correspondence $\bar\si_i\mapsto\si_i$ of galleries to an isometric embedding $\al':a_{mod}\to\geo X$ extending $\kappa_{\si}$: Since $\bar\ga$ connects two antipodal regular points there exists a unique such $\al'$ extending the isometry $\bar\ga\to\ga$ and hence mapping $\bar\si_i$ to $\si_i$. By construction, $\al'(\bar\xi_2)$ approximates $\al(\bar\xi_2)=\xi_1$ arbitrarily well. So we find a sequence of apartment charts $\al_n$ with the desired properties. We call two relative positions $\bar\xi_1,\bar\xi_2\in a_{mod}$ [*complementary*]{} if $\bar\xi_2=w_0\bar\xi_1$, where $w_0\in W$ is the longest element. In other words, the position complementary to $\bar\xi_1$ is $\pos(\bar\xi_1,\hat\si_{mod})$. We denote the relative position complementary to $\pos(\xi,\si)$ by $\cpos(\xi,\si)$, i.e. $\cpos=w_0\pos$. The relation of “complementarity” is clearly symmetric, $\ccpos=\pos$. Passing to complementary relative position reverses the partial order, $$\label{ineq:revordpos} \pos(\xi_1,\si)\preceq\pos(\xi_2,\si) \quad\Leftrightarrow\quad \cpos(\xi_1,\si)\succeq\cpos(\xi_2,\si) ,$$ compare Lemma \[lem:compposrev\]. In order to get a better geometric intuition for relative position, it is useful to compare it with Tits distance. Points with smaller position relative to a chamber are closer to it in a metric sense. Namely, (\[eq:foldimplmet\]) yields the inequality of Tits distances $$\label{ineq:ordrelpos1} \pos(\xi_1,\si)\preceq\pos(\xi_2,\si) \quad\Ra\quad \tangle(\xi_1,\cdot)|_{\si}\leq\tangle(\xi_2,\cdot)|_{\si} ,$$ respectively, $$\label{ineq:ordrelpos2} \pos(\xi_1,\si_1)\preceq\pos(\xi_2,\si_2) \quad\Ra\quad \tangle(\xi_1,\cdot)\circ\kappa_{\si_1} \leq\tangle(\xi_2,\cdot)\circ\kappa_{\si_2} .$$ The relative positions $\pos(\xi_1,\si_1)$ and $\pos(\xi_2,\si_2)$ are complementary if and only if $$\label{eq:complementrelpos} \tangle(\xi_1,\cdot)\circ\kappa_{\si_1}+ \tangle(\xi_2,\cdot)\circ\kappa_{\si_2}\circ\iota\equiv\pi$$ where $\iota=-w_0$ is the canonical involution of $\si_{mod}$. To see this, note that the formula reduces to the case when the chambers $\si_1$ and $\si_2$ are opposite to each other and the points $\xi_1$ and $\xi_2$ lie in the apartment spanned by them; it becomes then equivalent to the equality $\xi_1=\xi_2$. Later on we will use the following fact. \[lem:semcontrelpos\] If $\xi_n\to\xi$ and $\si_n\to\si$ (in the visual topology) are such that $\pos(\xi_n,\si_n)=\bar\xi$ for all $n$, then $\pos(\xi,\si)\preceq\bar\xi$. In other words, the set $\{\pos(\cdot,\si)\preceq\bar\xi\}$ is closed in $G\xi$. There exist apartment charts $\al_n:a_{mod}\to\geo X$ with $\al_n|_{\si_{mod}}=\kappa_{\si_n}$ and $\al_n(\bar\xi)=\xi_n$. The charts subconverge to a folding map $\al$ with $\al|_{\si_{mod}}=\kappa_{\si}$ and $\al(\bar\xi)=\xi$. It follows that subsets of the form $\{\pos(\cdot,\si)\succeq\bar\xi\}$ are open in $G\xi$ because their complements are finite unions of subsets of the form $\{\pos(\cdot,\si)\preceq\bar\xi'\}$. Chamber thickenings {#sec:chthick} ------------------- In this section we describe a construction of combinatorial “tubular neighborhoods” of families of chambers in $\geo X$. More generally, given a set of chambers $A\subset\D_FX$ we describe “tubular neighborhoods” of the projections of $A$ to single $G$-orbits $G\xi\subset\geo X$. \[dfn:thick\] A [*thickening*]{} of $\si_{mod}$ in $a_{mod}$ is a subset $\Th(\si_{mod})\subset a_{mod}$ which contains with every point $\bar\xi$ also every point $\bar\xi'$ satisfying $\pos(\bar\xi',\si_{mod})\preceq\pos(\bar\xi,\si_{mod})$. A thickening $\Th(\si_{mod})\subset a_{mod}$ induces thickenings $\Th(A)\subset\geo X$ of families of chambers $A\subset\D_FX$ as follows. One first defines the thickening $\Th(\si)\subset\geo X$ of a chamber $\si\subset\geo X$ as the set of all ideal points $\xi\in\geo X$ with $\pos(\xi,\si)\in\Th(\si_{mod})$. Then one puts $$\Th(A)=\cup_{\si\in A}\Th(\si)\subset\geo X .$$ If the thickening of $\si_{mod}$ is contained in a single Weyl orbit $W\bar\xi\subset a_{mod}$ for some $\bar\xi\in\si_{mod}$, i.e. $\Th(\si_{mod})\subset W\bar\xi$, then the thickening $\Th(A)$ is contained in the corresponding $G$-orbit $G\xi$ of type $\theta(\xi)=\bar\xi$. If $\Th(\si_{mod})\subset a_{mod}$ is closed and if $A\subset\D_FX$ is closed (with respect to the visual topology on $\D_FX$, i.e. its manifold topology as a homogeneous $G$-space), then $\Th(A)\subset\geo X$ is closed (with respect to the visual topology). Moreover, if $\Th(\si_{mod})\subset a_{mod}$ is an arbitrary thickening and if $A\subset\D_FX$ is closed, then the intersection $\Th(A)\cap G\xi$ with any $G$-orbit $G\xi$ is closed. Note that unions and intersections of thickenings of $\si_{mod}$ in $a_{mod}$ are again thickenings; in particular, intersections of thickenings with Weyl orbits are again thickenings. If a thickening has nonempty intersection with a Weyl orbit, $\Th(\si_{mod})\cap W\bar\xi\neq\emptyset$, then it contains its minimal element $\si_{mod}\cap W\bar\xi$. For $\bar\xi\in a_{mod}$ we define the [*sublevel thickening*]{} of $\si_{mod}$ inside the Weyl orbit $W\bar\xi$ as $$\Th_{\bar\xi}(\si_{mod}):=\{\bar\xi'\in W\bar\xi:\bar\xi'\preceq\bar\xi\}.$$ Note that the sublevel thickenings $\Th_{\bar\xi}(\si)\subset G\xi$ of chambers $\si\in\D_FX$ are precisely the Schubert cycles relative $\si$. By definition, general thickenings are unions of sublevel thickenings. The intersection $\Th(\si)\cap G\xi$ of a chamber thickening with a $G$-orbit is a finite union of Schubert cycles; in particular, it is compact. For a thickening $\Th(\si_{mod})\subset a_{mod}$ we define its [*complementary thickening*]{} as $$\left(\Th(\si_{mod})\right)^c:=w_0(a_{mod}-\Th(\si_{mod}))\subset a_{mod}.$$ That the complementary thickening is again a thickening of $\si_{mod}$ in $a_{mod}$, follows from the fact that $w_0$ reverses the folding order, cf. Lemma \[lem:compposrev\]. Furthermore, passing to the complementary thickening is an involutive operation: $$(\Th(\si_{mod})^c)^c =\Th(\si_{mod})$$ \[dfn:fatslimth\] We call a thickening $\Th(\si_{mod})\subset W\bar\xi$ [*fat*]{}, respectively, [*slim*]{} if for every pair of Weyl antipodes $w\bar\xi$ and $w_0w\bar\xi$ in $W\bar\xi$, at least one of them does, respectively, does not belong to $\Th(\si_{mod})$. We call a thickening [*balanced*]{} if it is both fat and slim. Note that a thickening is fat (slim, balanced) if and only if it contains (is contained in, coincides with) its complementary thickening. A Weyl orbit can contain a balanced thickening only if $w_0$ has no fixed point in the orbit. If any element in the Weyl orbit $W\bar\xi$ can be [*folding–compared*]{} with its $w_0$-image, i.e. if every pair of Weyl antipodes in $W\bar\xi$ is ordered, then there is a unique minimal fat thickening of $\si_{mod}$ in $W\bar\xi$. It consists of all elements which are $\preceq$ than their $w_0$-image. This happens in particular, if the Weyl orbit $W\bar\xi$ is totally ordered. We will use the following consequence of fatness: \[lem:posineqthick\] If $\Th(\si_{mod})\subset W\bar\xi$ is a fat thickening and if the points $\bar\xi$ and $\bar\xi'\in W\bar\xi$ satisfy $\bar\xi'\preceq w_0\bar\xi$, then $\Th(\si_{mod})$ contains at least one of them. If $\bar\xi\not\in\Th(\si_{mod})$, then $w_0\bar\xi\in\Th(\si_{mod})$ by fatness, and it follows that $\bar\xi'\in\Th(\si_{mod})$. \[ex:metthick\] For a type $\vartheta\in\si_{mod}$ and a radius $r\in(0,\pi)$ we define the [*metric $(\vartheta,r)$-thickening*]{} of $\si_{mod}$ in $a_{mod}$ as $$\Th^{met}_{\vartheta,r}(\si_{mod}) =\bar B(\vartheta,r)\subset a_{mod} .$$ It is a thickening by inequality (\[ineq:ordrelpos1\]). It has fat intersection with every Weyl orbit, if (but not only if) the type $\vartheta$ is $\iota$-invariant and $r\geq\pihalf$. To verify the latter, note that $\tangle(\bar\xi,\vartheta)+\tangle(\bar\xi,-\vartheta)=\pi$ and $\tangle(\bar\xi,-\vartheta)=\tangle(w_0\bar\xi,\vartheta)$ because $-w_0\vartheta=\iota\vartheta=\vartheta$. Given a thickening $\Th(\si_{mod})\subset a_{mod}$ and a type $\bar\eta\subset\si_{mod}$ then $$\Tr_{\bar\eta}\Th(\si_{mod}):= \st(\Th(\si_{mod})\cap W\bar\eta)\subset a_{mod}$$ is also a thickening. To verify this, suppose that $\bar\xi$ and $\bar\xi'$ are as in Definition \[dfn:thick\] and that $\bar\xi\in\st(w\bar\eta)$ with $w\bar\eta\in\Th(\si_{mod})$. The relative position inequality implies that there exists a folding map $f:a_{mod}\to a_{mod}$ restricting to the identity on $\si_{mod}$ such that $f(\bar\xi)=\bar\xi'$. Since $f(w\bar\eta)\in\Th(\si_{mod})$, it follows that $\bar\xi'\in\st(f(w\bar\eta))\subset\Tr_{\bar\eta}\Th(\si_{mod})$. If $\Th(\si_{mod})\cap W\bar\eta$ is fat, then $\Tr_{\bar\eta}\Th(\si_{mod})$ has fat intersection with every Weyl orbit, because for any pair of Weyl antipodal chambers at least one of them is contained in $\Tr_{\bar\eta}\Th(\si_{mod})$. We point out that the transfer thickening is a subcomplex. \[ex:rooth\] Let $\bar\eta\in\si_{mod}$ be a root type. We observe that the intersection of the metric thickening $\Th^{met}_{\vartheta,\pihalf}(\si_{mod})$ with the Weyl orbit $W\bar\eta$ is the same for every [*regular*]{} type $\vartheta\in\si_{mod}$, because $\tangle(w\bar\eta,\cdot)\neq\pihalf$ on $\interior(\si_{mod})$. The intersection with $W\bar\eta$ is fat. We define the [*$\bar\eta$-root thickening*]{} of $\si_{mod}$ in $a_{mod}$ as $$\Th^{root}_{\bar\eta}(\si_{mod}):= \Tr_{\bar\eta}\Th^{met}_{\vartheta,\pihalf}(\si_{mod})= \st\bigl(\bar B(\vartheta,\pihalf)\cap W\bar\eta\bigr)\subset a_{mod} .$$ If there is only one root type, we denote the root thickening of $\si_{mod}$ simply by $\Th^{root}(\si_{mod})$. We now discuss the existence of balanced thickenings depending on the Weyl group and the Weyl orbit. \[lem:balthreg\] The regular Weyl orbits $W\bar\xi\subset a_{mod}$ always admit a balanced thickening. Let $\vartheta\in \si_{mod}$ be a $\iota$-invariant type, and let $W\bar\xi$ be a regular orbit disjoint from the equatorial sphere $\partial B(\vartheta, \pi/2)$. Then $W\bar\xi\cap\Th^{met}_{\vartheta,\pihalf}(\si_{mod})$ is a balanced thickening inside $W\bar\xi$, compare Example \[ex:metthick\]. This implies the existence of balanced thickenings in all regular Weyl orbits because they are equivariantly identified with each other by natural order preserving bijections. \[lem:-1\] If $w_0=-\id$, equivalently, if the natural involution $\iota$ of $\si_{mod}$ is trivial, then every Weyl orbit $W\bar\xi\subset a_{mod}$ admits a balanced thickening. For a generic type $\vartheta \in \si_{mod}$, the distance function $\tangle(\vartheta, \cdot)$ does not take the value $\pihalf$ on $W\bar\xi$. Hence $W\bar\xi\cap\Th^{met}_{\vartheta,\pihalf}(\si_{mod})$ is a balanced thickening inside $W\bar\xi$. Indeed, for every $\bar\xi'\in W\bar\xi$, exactly one of the points $\bar\xi'$ and $w_0\bar\xi'=-\bar\xi'$ belongs to the closed ball $\bar B(\vartheta,\pihalf)$. If $W$ is irreducible, then $w_0=-\id$ if and only if $W$ is of type $A_1$, $B_{n\geq2}$, $D_{2k\geq4}$, $E_{7,8}$, $F_4$ or $G_2$, see [@Bourbaki]. Thus, if all irreducible components of the root system of $G$ are of type $A_1$, $B_{n\geq2}$, $C_{n\geq2}$, $D_{2k\geq4}$, $E_{7,8}$, $F_4$ or $G_2$, then every $G$-orbit $G\xi\subset\geo X$ at infinity admits a balanced thickening. We note that there are balanced thickenings in some singular Weyl orbits even if $w_0\neq-\id$. Convex cocompact groups of isometries of rank one symmetric spaces {#sec:rank1} ================================================================== In this section we review equivalent definitions and properties of convex cocompact groups of isometries of negatively curved symmetric spaces. Most of this discussion remains valid in the case of isometry groups of proper $CAT(-1)$ spaces. The main reference for this material is the paper of Bowditch [@Bowditch1995]. We also refer the reader to [@Kapovich2007] for a survey of discrete isometry groups of rank one symmetric spaces (primarily focused on higher-dimensional real-hyperbolic spaces). Let $\Gamma\subset G=Isom(X)$ be a discrete subgroup of the group of isometries of a negatively curved symmetric space $X$. We let $\Lambda=\Lambda(\Gamma)\subset \geo X$ denote the [*limit set*]{} of $\Gamma$, i.e. the accumulation set of a $\Gamma$-orbit in $X$. Note that $\La$ is necessarily closed in $\geo X$. Then $\Omega=\Omega(\Gamma)= \geo X-\Lambda$ is the [*domain of discontinuity*]{} of $\Gamma$, which is also the wandering set for the action $\Ga\acts \geo X$, and hence is the largest open subset of $\geo X$ where $\Gamma$ acts properly discontinuously. The [*Nielsen hull*]{} $N(\La)$ of $\La$ is defined as the smallest closed convex subset in $X$ whose ideal boundary contains $\La$. The set $N(\La)$ exists provided that $\La$ contains at least two points; in this case, $\geo N(\La)=\La$. The following definition explains the terminology [*convex cocompact*]{}. \[C1\] A discrete subgroup $\Gamma\subset G$ is called [*convex cocompact*]{} if $N(\La(\Ga))/\Ga$ is compact. A limit point $\la\in \La$ is called [*a conical limit point*]{} of $\Gamma$ if for some (every) geodesic ray $\rho$ in $X$ asymptotic to $\la$ there exists a sequence $\gamma_i x\subset X$ converging to $\la$ in an $R$-neighborhood of $\rho$ for some $R<\infty$. \[C2\] A discrete subgroup $\Gamma\subset G$ is called [*convex cocompact*]{} if every limit point of $\Ga$ is conical. Recall (see section \[sec:conv actions\]) that for a set $Z$, $\Trip(Z)$ denotes the set of triples of pairwise distinct points in $Z$. \[C3\] A discrete subgroup $\Gamma\subset G$ is called [*convex cocompact*]{} if $\La=\La(\Ga)$ contains at least two points and the action $\Gamma\acts \Trip(\La)$ is cocompact. Every discrete group $\Gamma\subset G$ acts properly discontinuously on $X\cup \Om$, which we equip with the subset topology induced from $\bar{X}=X\cup \geo X$. \[C4\] A discrete subgroup $\Gamma\subset G$ is called [*convex cocompact*]{} if the action $\Gamma\acts X\cup \Om$ is cocompact. Since $X\cup \Om$ (as $N(\La(\Ga))$ is contractible and, in particular, simply connected, one deduces: Every convex cocompact group is finitely presented. Definition C4 implies that for every convex cocompact subgroup $\Gamma\subset G$, the quotient $\Om/\Ga$ is compact. The converse is false, as the following examples show. 1\. Consider a cyclic group $\Gamma$ of parabolic isometries of the hyperbolic plane $\H^2=X$. Then $\La(\Ga)$ is a singleton, $\Om/\Ga$ is homeomorphic to $S^1$, while $$(X\cup \Omega)/\Ga \cong [0,\infty)\times S^1$$ is noncompact. Thus, $\Gamma$ is not convex cocompact. In this case, of course, $\Gamma$ contains unipotent (parabolic) elements. The next three examples contain only loxodromic elements. 2\. Let $S$ denote a closed hyperbolic surface, $\pi:=\pi_1(S)$. Then $\pi$ admits a discrete and faithful representation $\rho: \pi\to G=Isom(\H^3)$, so that its image $\Ga=\rho(\pi)$ is a [*totally-degenerate purely loxodromic subgroup*]{} of $G$: $\Om(\Ga)$ is simply-connected and nonempty, $\Gamma$ contains no parabolic elements and $$(X\cup \Om)/\Ga$$ is homeomorphic to $S\times [0,\infty)$, where $S\times \{0\}$ corresponds to $\Om/\Gamma$, see [@Bers]. Thus, $\Gamma$ is not convex cocompact. 3\. Let $M$ be a closed oriented hyperbolic $m$-manifold with nonseparating oriented compact totally-geodesic hypersurface $N$. Such manifolds exist for all $m$ (see [@Millson(1976)]). Let $\tilde{M}\to M$ denote the infinite cyclic cover determined by the homomorphism $\pi_1(M)\to \Z$ corresponding to the element of $H^1(M,\Z)$ Poincaré dual to the homology class $[N]$. Then $N$ lifts to a submanifold $N_0\subset \tilde{M}$ which is isometric to $N$ and which separates $\tilde{M}$ in two components $\tilde{M}_-, \tilde{M}_+$. Let $M'$ denote the metric completion of the Riemannian manifold $\tilde{M}_+$. Then $M'$ is a complete hyperbolic manifold with single geodesic boundary component isometric to $N_0$ and injectivity radius bounded below. The fundamental group $\Gamma$ of $M'$ is not finitely-generated. The hyperbolic structure on $M'$ determines a discrete isometric action $\Gamma \acts \H^m$, so that $\Gamma$ contains no parabolic elements. Then $$(\H^m\cup\Om)/\Gamma$$ is homeomorphic to $M'$. In particular, $\Om/\Gamma$ is compact and nonempty, while $\Gamma$ is not even finitely-generated; in particular, $\Gamma$ is not convex cocompact. 4\. Similarly, for every $G$, there are discrete subgroups $\Ga\subset G$ whose limit set is the entire sphere $\geo X$, but $\Ga$ is not finitely-generated. For instance, one can start with a uniform lattice $\widehat{\Ga}\subset G$; being a non-elementary word hyperbolic group, $\widehat{\Ga}$ admits a normal subgroup $\Ga$ which is isomorphic to the free groups of countably infinite rank. The limit set of $\Ga$ is necessarily the entire sphere $\geo X$. Furthermore, when $X$ is a real-hyperbolic 3-space or complex-hyperbolic plane or complex-hyperbolic 3-space, there are examples of finitely-generated subgroups $\Ga\subset G$ whose limit set is $\geo X$, but $\Ga$ is not a lattice in $G$. In the case $X=\H^3$, such examples can be constructed, for instance, using normal surface subgroups in fundamental groups of closed hyperbolic 3-manifolds fibering over the circle. For examples in $\C\H^2$ and $\C\H^3$ see e.g. [@Kapovich2013]: These are normal subgroups in complex-hyperbolic manifolds which admit (singular) holomorphic fibrations over hyperbolic Riemann surfaces. On the other hand, this phenomenon can essentially only occur in the real-hyperbolic case: \[See [@Ramachandran]\] Let $X$ be a negatively curved rank one symmetric space which is not real-hyperbolic. Suppose that $\Gamma\subset G=Isom(X)$ is a discrete torsion-free subgroup without unipotent elements so that $\Om(\Ga)/\Ga$ is compact and nonempty. Then $\Gamma$ is convex cocompact provided that $X$ is not isometric to $\C\H^2$. In the case $X=\C \H^2$, the same result holds provided that the Riemannian manifold $\Gamma\backslash X$ has injectivity radius bounded below. Let $\Ga\subset G$ be a discrete subgroup. Pick a point $x\in X$ which is not fixed by any nontrivial element of $\Ga$, and define the [*Dirichlet fundamental domain*]{} $D_x$ of $\Ga$ as $$D_x=\{y\in X: d(x,y)\le d(\ga x, y), \forall \ga\in \Ga\}.$$ Note that $D_x$ is convex if $X$ is real-hyperbolic, but is not convex otherwise. In general, $D_x$ is starlike with the center $x$; since $X$ is Gromov hyperbolic, this implies that $D_x$ is [*quasiconvex*]{} in $X$. Subsets of $D_x$ of the form $$D_x\cap \ga D_x, \quad\ga\in \Ga,$$ are called [*faces*]{} of $D_x$. Let $\bar{D}_x$ denote the union $$D_x\cup (\geo D_x \cap \Om(\Ga)),$$ which is a certain partial compactification of $D_x$. It follows (almost) immediately from C4 that $\Gamma$ is convex cocompact if and only if $\bar{D}_x$ is compact. The following definition is a more elaborate version of this observation: \[C5\] A discrete subgroup $\Gamma\subset G$ containing no parabolic elements is called [*convex cocompact*]{} if one (every) Dirichlet fundamental domain $D_x$ has finitely many faces. Note that a cyclic parabolic subgroup of $Isom(\H^4)$ can have a Dirichlet domain with infinitely many faces. \[C6\] A discrete subgroup $\Gamma\subset G$ is convex cocompact whenever $\Gamma$ is word hyperbolic as an abstract group and there exists an equivariant homeomorphism $$\beta: \geo \Gamma \to \La(\Ga),$$ where $\geo \Ga$ is the Gromov boundary of $\Gamma$. Note that the injectivity of $\beta$ is critical here: \[Mj, [@Mj]\] Suppose that $\Ga\subset\Isom(\H^3)$ is a word hyperbolic subgroup (not necessarily convex cocompact). Then there always exists a continuous equivariant map $\beta: \geo \Gamma\to \La(\Ga)$; the map $\beta$ is called the [*Cannon-Thurston map*]{} for the subgroup $\Ga\subset G$. Let $\Ga\subset G$ be a finitely-generated subgroup; we will equip $\Ga$ with a word metric. A point $x\in X$ defines the [*orbit map*]{} $\Ga \to \Ga x\subset X$. The subgroup $\Ga$ is called [*undistorted*]{} in $G$, if some (any) orbit map $\Ga \to X$ is a quasi-isometric embedding, equivalently, if the inclusion $\Ga\to G$ is a quasi-isometric embedding, where $G$ is equipped with a left invariant Riemannian metric. \[C7\] A discrete subgroup $\Gamma\subset G$ is convex cocompact if it is undistorted. Note that, in view of the hyperbolicity of $X$, undistortion of $\Ga$ implies that the quasi-isometrically embedded $\Ga$-orbits are quasi-convex subsets of $X$. In particular: 1\. $\Ga$ is word hyperbolic, and hence the orbit maps $\Ga\to\Ga x$ continuously extend at infinity to an equivariant homeomorphism $\beta: \geo \Ga \to \La(\Ga)$. 2\. The $\Ga$-equivariant relation in $X\times \Ga$ given by the nearest-point projection to an orbit $\Ga x$ is a coarse Lipschitz retraction $X\to\Ga x$. The converse to this is also easy: \[C8\] A discrete subgroup $\Gamma\subset G$ is convex cocompact if for some (every) $\Ga$-orbit $\Ga x\subset X$ there exists a $\Ga$-equivariant coarse Lipschitz retraction $X\to \Ga x$. The equivariance condition for the retraction can be omitted: \[C9\] A discrete subgroup $\Gamma\subset G$ is convex cocompact if for some (every) $\Ga$-orbit $ \Ga x\subset X$ there exists a coarse Lipschitz retraction $X\to \Ga x$. Our last characterization of convex cocompactness is in terms of expanding actions. We fix a visual metric $d$ on $S=\geo X$. \[C10\] A discrete subgroup $\Gamma\subset G$ is convex cocompact if its action is expanding at every point of $\La(\Ga)$: There exists a finite collection of open subsets $U_i\subset S, i=1,\ldots,k$ covering $\La(\Ga)$ and elements $\ga_i\in \Ga, i=1,\ldots,k$, so that the restrictions $\ga_i|U_i$ are expanding for all $i=1,\ldots k$. This interpretation of convex cocompactness appears in Sullivan’s paper [@Sullivan]. \[thm:rk1\] The definitions C1–C10 are equivalent. The equivalence of Definitions C1–C9 can be found for instance in [@Bowditch1995]. The implication C10 $\Rightarrow$ C2 is a corollary of Lemma \[lem:conical\], alternatively, it also follows from our Proposition \[prop:trexpcoco\]. In the case of real-hyperbolic space, the implication C5 $\Rightarrow$ C10 is immediate by taking a Ford fundamental domain (bounded by finitely many isometric spheres $I(\gamma_i), I(\gamma_i^{-1})$, $i=1,\ldots,k$) and observing that $\gamma_i$ is a strict expansion on every compact contained in the open round ball bounded by $I(\gamma_i)$). For the remaining rank one symmetric spaces the implication C2 $\Rightarrow$ C10 is a corollary of our Proposition \[prop:conicimplexpatr\]. Some domains of discontinuity at infinity {#sec:domdisc} ========================================= Let $X=G/K$ be a symmetric space of noncompact type, and let $\Ga\subset G$ be an infinite discrete subgroup. The [*limit set*]{} $\La=\La(\Ga)\subset\geo X$ of $\Ga$ is defined as the accumulation set of an orbit $\Ga x\subset X$. It is independent of the point $x\in X$. In this section, we establish preliminary results on the relation between domains of discontinuity of $\Ga$ and its limit set. Roughly speaking, we show that the $\Ga$-action is properly discontinuous sufficiently far away from the limit set in terms of the Tits distance. Domains of discontinuity for general discrete subgroups ------------------------------------------------------- Suppose that $K\subset\geo X$ is a compact subset such that there exist isometries $\ga_n\to\infty$ in $\Ga$ for which $K\cap\ga_n K\neq\emptyset$. \[lem:nonwandcomp\] $\tangle(K,\La(\Ga))\leq\pihalf$, i.e. there exist $\xi\in K$ and $\la\in\La(\Ga)$ with $\tangle(\xi,\la)\leq\pihalf$. Let $\xi_n\in K\cap\ga_n K$. We may assume that $\ga_n^{\pm1}x\to\la_{\pm}\in\La(\Ga)$ for some (any) base point $x\in X$ and, because $K$ is compact, that $\xi_n\to\xi\in K$ and $\ga_n^{-1}\xi_n\to\xi'\in K$. In this situation we have the inequality: $$\label{ineq:keyest} \tangle(\la_+,\xi)+\tangle(\la_-,\xi')\leq\pi$$ Let $x_n^{\pm}(t)$ be the point at distance $t$ from $x$ on the segment connecting $x$ to $\ga_n^{\pm1}x$ (it is defined for sufficiently large $n$ depending on $t$) and $x_{\pm}(t)$ the point at distance $t$ from $x$ on the ray $x\la_{\pm}$. Then $\angle_{x_n^+(t)}(\ga_nx,\xi_n) \buildrel{n\to\infty}\over\lra \angle_{x_+(t)}(\la_+,\xi) \buildrel{t\to\infty}\over\lra \tangle(\la_+,\xi)$, that is $$\lim_{t\to\infty}\lim_{n\to\infty}\angle_{x_n^+(t)}(\ga_nx,\xi_n)=\tangle(\la_+,\xi),$$ and, similarly, $$\lim_{t\to\infty}\lim_{n\to\infty}\angle_{x_n^-(t)}(\ga_n^{-1}x,\ga_n^{-1}\xi_n)=\tangle(\la_-,\xi').$$ Note that (again for sufficiently large $n$ depending on $t$) the points $x_n^+(t)$ and $\ga_nx_n^-(t)$ lie in this order on the segment from $x$ to $\ga_nx$. The ideal triangles $\De(x_n^+(t),\ga_nx_n^-(t),\xi_n)$ have angle sum $\leq\pi$, and we get $$\angle_{x_n^+(t)}(\ga_nx,\xi_n)+ \underbrace{\angle_{\ga_nx_n^-(t)}(x,\xi_n)}_{ \angle_{x_n^-(t)}(\ga_n^{-1}x,\ga_n^{-1}\xi_n)} =\angle_{x_n^+(t)}(\ga_nx_n^-(t),\xi_n)+ \angle_{\ga_nx_n^-(t)}(x_n^+(t),\xi_n) \leq\pi .$$ Passing to the double limits yields the assertion. To conclude the proof of the lemma, we observe that at least one of the summands on the left hand side of (\[ineq:keyest\]) must be $\leq\pihalf$. In the proof of (\[ineq:keyest\]) we used only that $X$ has nonpositive curvature. If one takes into account the higher rank geometry of $X$, one can refine the inequality (\[ineq:keyest\]) for Tits angles to an inequality for relative positions, compare e.g. (\[eq:dynrelposcond\]). As an application of Lemma \[lem:nonwandcomp\] we find domains of discontinuity for the $\Ga$-action at infinity: We call the closed $\pihalf$-neighborhood $$\bar N_{\pihalf}(\La(\Ga)) =\{\xi\in \geo X: \exists \la\in \La(\Ga), \tangle(\xi, \la)\le \pihalf\}$$ the [*$\pihalf$-thickening*]{} of the limit set $\La(\Ga)$ of $\Ga$. \[thm:pdgen\] $\geo X-\bar N_{\pihalf}(\La(\Ga))$ is a domain of discontinuity for $\Ga$. According to Lemma \[lem:nonwandcomp\], every compact subset of $\geo X-\bar N_{\pihalf}(\La(\Ga))$ has the property that it intersects only finitely many of its $\Ga$-translates. This means that the $\Ga$-action on $\geo X-\bar N_{\pihalf}(\La(\Ga))$ is properly discontinuous. Transferring domains of discontinuity {#sec:trans} ------------------------------------- Given a domain of discontinuity for $\Ga$ in a $G$-orbit at infinity $G\eta\subset\geo X$, one can construct from it a domain of discontinuity in $\geo X$ and in particular in other $G$-orbits. We refer to section \[sec:star\] for the definition of the closed star of a subset at infinity. \[prop:transf\] If $\Om\subset G\eta$ is a domain of discontinuity for $\Ga$ in $G\eta$, then $\geo X-\st(G\eta-\Om)$ is a domain of discontinuity for $\Ga$ in $\geo X$. In particular, the intersection of $\geo X-\st(G\eta-\Om)$ with any $G$-orbit $G\xi$ is a domain of discontinuity for $\Ga$ in $G\xi$. Suppose that $K\subset\geo X-\st(G\eta-\Om)$ is compact. Then $K_{\eta}=G\eta\cap\st(K)$ is compact and $K_{\eta}\subset\Om$. Note that for $\ga\in\Ga$ we have $K\cap\ga K=\emptyset$ if $K_{\eta}\cap\ga K_{\eta}=\emptyset$. By assumption, $K_{\eta}\cap\ga K_{\eta}=\emptyset$ for all but finitely many $\ga\in\Ga$. Hence $K\cap\ga K=\emptyset$ for all but finitely many $\ga\in\Ga$. Thus the $\Ga$-action on $\geo X-\st(G\eta-\Om)$ is properly discontinuous. If the $G$-orbit $G\xi$ is equivariantly diffeomorphic to $G\eta$, equivalently, if the types $\theta(\xi)$ and $\theta(\eta)$ are contained in the same open face of $\si_{mod}$, then the transferred domain of discontinuity $G\xi-\st(G\eta-\Om)$ corresponds to $\Om$ under the diffeomorphism of orbits. More generally, if there is an equivariant fibration $G\xi\to G\eta$ with compact fibers, equivalently, if $\theta(\xi)\in\ost(\theta(\eta))$, then the transferred domain of discontinuity $G\xi-\st(G\eta-\Om)$ is the inverse image of $\Om$. Otherwise, i.e. if $\theta(\xi)\not\in\ost(\theta(\eta))$, the transferred domain of discontinuity may be empty even if $\Om\neq\emptyset$. Compare the discussion in section \[sec:star\]. Applying Proposition \[prop:transf\] to the $\pihalf$-thickening, we deduce from Theorem \[thm:pdgen\]: \[cor:pdtrans\] For any $G$-orbit $G\eta\subset\geo X$ the set $\geo X-\st(G\eta\cap\bar N_{\pihalf}(\La(\Ga)))$ is a domain of discontinuity for $\Ga$ in $\geo X$. In particular, for any other $G$-orbit $G\xi$ the set $$G\xi-\st(G\eta\cap\bar N_{\pihalf}(\La(\Ga)))$$ is a domain of discontinuity for $\Ga$ in $G\xi$. \[rem:ostdec\] The open star $\ost(G\eta)$ decomposes as the [*disjoint*]{} union $$\ost(G\eta)= \ost(G\eta-\bar N_{\pihalf}(\La(\Ga))) \sqcup\ost(G\eta\cap\bar N_{\pihalf}(\La(\Ga)))$$ Hence, if $G\xi\subset\ost(G\eta)$, then $$G\xi-\st(G\eta\cap\bar N_{\pihalf}(\La(\Ga)))= G\xi\cap\st(G\eta-\bar N_{\pihalf}(\La(\Ga))).$$ Dynamical relation at infinity for regular subgroups {#sec:dynrelation} ==================================================== In this section, we will prove our main result regarding domains of discontinuity for $\Ga$-actions at infinity, see Theorem \[thm:pdth\]. The domains of discontinuity given in section \[sec:domdisc\] are in general not maximal, and to obtain them we only use that $X$ is a CAT(0)-space. We now prove more precise results by taking into account the building structure on $\geo X$ and describing dynamical properties of $\Ga$-orbits in terms of their position relative to the set of chambers which intersect the limit set. We will do this under the assumption that the discrete subgroup $\Ga\subset G$ satisfies an asymptotic [*regularity*]{} property. Regularity {#sec:reg} ---------- Regularity is an [*asymptotic*]{} property for unbounded subsets of isometries in $G$ or points in $X$. We start by defining it for sequences diverging to infinity. \[Regular sequences\] \[def:regseq\] (i) We call a sequence $\de_n\to\infty$ in $\De=\De_{euc}$ [*regular*]{} if $$d(\de_n,\D\De)\to+\infty,$$ and [*uniformly*]{} regular if $$\liminf_{n\to\infty}d(\de_n,\D\De)/\|\de_n\|>0.$$ \(ii) We call a sequence $x_n\to\infty$ in $X$ [*regular*]{}, respectively [*uniformly regular*]{}, if for some (any) base point $x$ the sequence of $\De$-valued lengths $d_{\De}(x,x_n)$ in $\De$ has this property. \(iii) We call a sequence $g_n\to\infty$ in $G$ [*regular*]{}, respectively [*uniformly regular*]{}, if some (any) orbit $(g_nx)$ in $X$ has this property. The independence of the base point and the orbit in parts (ii) and (iii) is due to the triangle inequality $|d_{\De}(x,y)-d_{\De}(x',y')|\leq d(x,x')+d(y,y')$. Note that a diverging sequence of vectors in $\De$ (of points in $X$) is uniformly regular if and only if its accumulation set in $\geo\De\cong\si_{mod}$ (in the visual boundary $\geo X$) is contained in the interior of $\si_{mod}$ (in the regular part $\geo^{reg} X$). In the general regular case, there may be singular accumulation points. Furthermore, a sequence $g_n\to\infty$ in $G$ is (uniformly) regular if and only if the sequence of inverses $(g_n^{-1})$ is, cf. the symmetry property (\[eq:symprop\]) for $\De$-lengths. Now we transfer the notions from sequences to subsets. \[Regular subgroups\] \[def:regsubs\] We call a subgroup of $G$ [*(uniformly) regular*]{} if all divergent sequences in it have this property. Note again that a subgroup $\Ga\subset G$ is uniformly regular if and only if its limit set $\La(\Ga):=\ol{\Ga x}\cap\geo X$ consists only of regular ideal points, $\La(\Ga)\subset\geo^{reg}X$. A discrete subgroup $\Ga\subset G$ need not be regular even if all its nontrivial elements are regular (transvections). This can happen e.g. for free abelian subgroups of rank $\geq2$. Chamber convergence and chamber limit set {#sec:chconv} ----------------------------------------- For regular sequences in $X$ and $G$ there is a natural weaker notion of convergence at infinity “transversely to Weyl chambers”. \(i) We say that a regular sequence $x_n\to\infty$ in $X$ [*chamber converges*]{} to the chamber $\si_{\infty}\in\D_FX$, if for some (any) base point $x$ the chambers $\si_n$ with $x_n\in V(x,\si_n)$ converge in $\D_FX$ to $\si_{\infty}$, $\si_n\to\si_{\infty}$. We write $x_n\to\si_{\infty}$. \(ii) We say that a regular sequence $g_n\to\infty$ in $G$ [*chamber converges*]{} to the chamber $\si_{\infty}\in\D_FX$, if some (any) orbit $(g_nx)$ in $X$ has this property, $g_nx\to\si_{\infty}$, and write $g_n\to\si_{\infty}$. The limit chamber $\si_{\infty}$ contains the accumulation points at infinity of the sequence $(x_n)$, respectively the orbits $(g_nx)$. In the uniform case, these accumulation points are regular and determine the limit chamber, which is hence independent of the base point $x$. In the general regular case the independence of the base point requires an argument. This is a consequence of the following result. \[lem:chconvwd\] Let $x_n,x'_n\to\infty$ be regular sequences in $X$ such that $\sup_nd(x_n,x'_n)<+\infty$, and let $(p_n),(p'_n)$ be bounded sequences in $X$. Moreover, let $\si_n,\si'_n\in\D_FX$ be chambers such that $x_n\in V(p_n,\si_n)$ and $x'_n\in V(p'_n,\si'_n)$. If $\si_n\to\si_{\infty}$, then $\si'_n\to\si_{\infty}$. After passing to a subsequence, we may assume that we have convergence $\si'_n\to\si'_{\infty}$. We must show that $\si'_{\infty}=\si_{\infty}$. We choose a reference chamber (“base point at infinity”) $\hat\si$ opposite to $\si_{\infty}$ and $\si'_{\infty}$. Then $\hat\si$ is also opposite to $\si_n$ and $\si'_n$ for large $n$. The maximal flats $F_n=F(\hat\si,\si_n)$ and $F'_n=F(\hat\si,\si'_n)$ are well-defined for large $n$ and form bounded families. To see this, note that $G$ acts transitively on the space $(\D_FX\times\D_FX)^{opp}$ of pairs of opposite chambers, and therefore the family of flats $U\cdot F(\hat\si,\si_{\infty})$ for a compact neighborhood $U\subset G$ of the identity contains $F_n$ for large $n$. Hence $d(p_n,F_n)<C$ and $d(p'_n,F'_n)<C$. Since $d(\cdot,F_n)|_{V(p_n,\si_n)}\leq d(p_n,F_n)$ by convexity of the distance function, we have $d(x_n,F_n)\leq d(p_n,F_n)$. Thus $d(x_n,F_n)$ and equally $d(x'_n,F'_n)$ are also uniformly bounded. Let $\bar p_n,\bar x_n\in F_n$ be the nearest point projections of $p_n,x_n$ to $F_n$, and $\bar p'_n,\bar x'_n\in F'_n$ the projections of $p'_n,x'_n$ to $F'_n$. By regularity, we have $d(x_n,\D V(p_n,\si_n))\to+\infty$. Since $p_n$ and $x_n$ lie at uniformly bounded distance from $F_n$, it follows that $\bar x_n\in V(\bar p_n,\si_n)$ for large $n$ and $d(\bar x_n,\D V(\bar p_n,\si_n))\to+\infty$, and hence $\bar p_n\in V(\bar x_n,\hat\si)$ for large $n$ and $d(\bar p_n,\D V(\bar x_n,\hat\si))\to+\infty$. Since $d(\bar x_n,F'_n)$ is uniformly bounded as a consequence of the triangle inequality), this implies that $d(\bar p_n,F'_n)\to0$. To see this, we choose a regular ideal point $\hat\zeta\in\hat\si$. There exist points $z_n$ on the rays $\bar x_n\hat\zeta$ such that $d(z_n,\bar x_n)\to+\infty$ while still $\bar p_n\in V(z_n,\hat\si)\subset V(\bar x_n,\hat\si)$. Then $d(\bar x_n,F'_n)\geq d(z_n,F'_n)\geq d(\bar p_n,F'_n)$ because $\geo V(\bar x_n,\hat\si)=\hat\si\subset\geo F'_n$. Since the rays $\bar x_n\hat\zeta$ approach the flats $F'_n$ at a uniform (exponential) rate, controlled by the bound on $d(\bar x_n,F'_n)$ and the type of $\hat\zeta$, it follows that $d(\bar p_n,F'_n)\to0$. Since the points $\bar p_n\in F_n$ form a bounded subset of $X$, and since the flats $F_n$ converge, $F_n\to F(\hat\si,\si_{\infty})$, it follows that the flats $F'_n$ converge to the same limit flat, $F'_n\to F(\hat\si,\si_{\infty})$, and so $\si'_n\to\si_{\infty}$. A regular sequence of points in $X$ or of isometries in $G$ has chamber convergent subsequences, due to the compactness of $\D_FX$. We define the [*chamber limit set*]{} $\La_{ch}(\Ga)\subset\D_FX$ of a regular discrete subgroup $\Ga\subset G$ as the set of all chambers which occur as the limit of some chamber convergent sequence $\ga_n\to\infty$ in $\Ga$. Clearly, $\La_{ch}(\Ga)$ is compact and $\Ga$-invariant. If $\Ga$ is uniformly regular, then $\La_{ch}(\Ga)$ is the image of $\La(\Ga)$ under the natural projection $\geo^{reg}X\to\D_FX$ to the Fürstenberg boundary. Basic properties of the limit set have been studied by Benoist [@Benoist]. He proves (under more general assumptions) that there exists a closed convex subset $\Theta(\Ga)\subset \si_{mod}$ so that the type map $\theta:\geo X\to\si_{mod}$ restricts to a product fibration $\theta: \La(\Ga) \to \Theta(\Ga)$ with fiber $\La_{ch}(\Ga)$. We call a regular discrete subgroup $\Ga\subset G$ [*non-elementary*]{} if it has at least three limit chambers, $|\La_{ch}(\Ga)|\geq3$. Dynamical relation {#sec:dynrel} ------------------ In order to determine the dynamical relations between points in $\geo X$ with respect to actions of regular discrete subgroups $\Ga\subset G$, we need to determine the dynamical relations with respect to diverging sequences of isometries in $G$. By passing to subsequences, we may restrict ourselves to chamber convergent sequences of isometries. We therefore fix a regular sequence $g_n\to\infty$ in $G$ such that both sequences $(g_n^{\pm1})$ chamber converge, $g_n^{\pm1}\to\si_{\pm}$. Note that the limit chambers $\si_{\pm}$ are in general not opposite to each other (and may even agree). As a first step, we study the dynamics of $(g_n)$ on chambers, that is on $\D_FX$, and show that $\si_+$ attracts the open Schubert cell relative $\si_-$, i.e. the dense open subset of chambers opposite to $\si_-$, in a locally uniform way. This is our generalization of the convergence property for group actions on ideal boundaries of rank one symmetric spaces. \[lem:unifattr\] On the open Schubert cell $C(\si_-)$ relative $\si_-$, the maps $g_n$ converge locally uniformly to the constant map $\equiv\si_+$. By symmetry, on the open Schubert cell $C(\si_+)$ relative $\si_+$, the maps $g_n^{-1}$ converge locally uniformly to the constant map $\equiv\si_-$. Let $\si$ be a chamber opposite to $\si_-$. Let $\si_n^{\pm}:=\geo V(x,g_n^{\pm1}x)\to\si_{\pm}$ and denote by $\zeta_n^{\pm},\zeta_{\pm},\zeta$ the regular points in $\si_n^{\pm},\si_{\pm},\si$ of a fixed $\iota$-invariant type $\bar\zeta\in\si_{mod}$. We want to verify that $$\label{eq:centang} \angle_{g_n^{-1}x}(\zeta_n^-,\zeta)\to\pi .$$ From $g_n^{-1}x\in V(x,\si_n^-)$ it follows that $x\in V(g_nx,g_n\si_n^-)$. Therefore the chambers $g_n\si_n^-$ and $\si_n^+$ are opposite to each other, and (\[eq:centang\]) yields that $\angle_x(\zeta_n^+,g_n\zeta)=\pi-\angle_x(g_n\zeta_n^-,g_n\zeta)\to0$. Consequently, $g_n\zeta\to\zeta_+$ and $g_n\si\to\si_+$. To verify (\[eq:centang\]), we consider the maximal flats $F_n=F(\si_n^-,\si)$. They are well-defined for large $n$ because then $\si_n^-$ is opposite to $\si$. Moreover, the flats $F_n$ form a bounded family, i.e. $d(x,F_n)<C$, cf. the proof of Lemma \[lem:chconvwd\]. Property (\[eq:centang\]) is implied by $$\label{eq:orbappr} d(g_n^{-1}x,F_n)\to0 .$$ The latter property is obtained as in the proof of Lemma \[lem:chconvwd\] from the following facts: $g_n^{-1}x\in V(x,\si_n^-)$, $\si_n^-\subset\geo F_n$, $d(x,F_n)<C$, and the sequence $(g_n^{-1})$ is regular. This shows pointwise convergence $g_n\si\to\si_+$ for all chambers $\si$ opposite to $\si_-$. The same argument yields locally uniform convergence of $g_n$ to the constant map $\equiv\si_+$ on the open Schubert cell opposite to $\si_-$. Indeed, if $C$ is a compact set of chambers opposite to $\si_-$, then for large $n$ these chambers are simultaneously opposite to $\si_n^-$ and the family of maximal flats $F(\si_n^-,\si)$ with $\si\in C$ and $n$ sufficiently large is bounded. The same estimates as before work uniformly on $C$. In terms of dynamical relations, Lemma \[lem:unifattr\] tells that all points in the open Schubert cell $C(\si_-)$ are dynamically related to $\si_+$ with respect to the sequence $(g_n)$. From this one can deduce the dynamical relations between arbitrary ideal points. \[Dynamical relation on $\geo X$ with respect to regular sequences of isometries\] \[prop:dynrel\] $\xi$ is dynamically related to $\xi'$ with respect to $(g_n)$ if and only if $$\label{eq:dynrelposcond} \pos(\xi',\si_+)\preceq\cpos(\xi,\si_-) .$$ Suppose first that the relative positions $\pos(\xi,\si_-)$ and $\pos(\xi',\si_+)$ are complementary, $\pos(\xi',\si_+)=\cpos(\xi,\si_-)$. There exist chambers $\hat\si_{\pm}$ opposite to $\si_{\pm}$ such that $\xi\in a(\si_-,\hat\si_-)$ and $\xi'\in a(\hat\si_+,\si_+)$. By Lemma \[lem:unifattr\], we know that $g_n\hat\si_-\to\si_+$ and $g_n^{-1}\hat\si_+\to\si_-$. Hence $g_n^{\pm1}\hat\si_{\mp}$ is opposite to $\hat\si_{\pm}$ for large $n$, and we have that $$\label{eq:aptconv} a(g_n^{-1}\hat\si_+,\hat\si_-)\to a(\si_-,\hat\si_-) \hbox{ and } g_n\cdot a(g_n^{-1}\hat\si_+,\hat\si_-)=a(\hat\si_+,g_n\hat\si_-)\to a(\hat\si_+,\si_+) .$$ More precisely, this is to be understood as the convergence $\al_n\to\al_-$, respectively, $g_n\al_n\to\al_+$ for the corresponding apartment charts $\al_n:a_{mod}\to a(g_n^{-1}\hat\si_+,\hat\si_-)$, $\al_-:a_{mod}\to a(\si_-,\hat\si_-)$ and $\al_+:a_{mod}\to a(\hat\si_+,\si_+)$, which restrict to the chamber charts ${\al_n}|_{\si_{mod}}=\kappa_{g_n^{-1}\hat\si_+}$, ${\al_-}|_{\si_{mod}}=\kappa_{\si_-}$ and ${\al_+}|_{\si_{mod}}=\kappa_{\hat\si_+}$. The formula (\[eq:aptconv\]) says that the apartment $a(\si_-,\hat\si_-)$ is dynamically related to $a(\hat\si_+,\si_+)$. It follows that $\xi$ is dynamically related to $\xi'$ because for $\xi_n\in a(g_n^{-1}\hat\si_+,\hat\si_-)$ with $\pos(\xi_n,\hat\si_-)=\pos(\xi,\hat\si_-)$ we have that $\xi_n\to\xi$ and $g_n\xi_n\to\xi'$. Indeed, the convergence follows from the convergence of apartment charts (\[eq:aptconv\]) since the points $\xi_n,\xi,g_n\xi_n$ and $\xi'$ correspond to the same point $\bar\xi\in a_{mod}$ under the charts $\al_n,\al_-,g_n\al_n$ and $\al_+$, because $\pos(\xi,\hat\si_-)=\cpos(\xi,\si_-)=\pos(\xi',\si_+)$. If only the inequality (\[eq:dynrelposcond\]) holds, then we can approximate $\xi'$ by points $\tilde\xi'$ with $\pos(\tilde\xi',\si_+)=\cpos(\xi,\si_-)$. By the above, $\xi$ is dynamically related to these points $\tilde\xi'$, and since dynamical relation is a closed relation it follows that $\xi$ is dynamically related also to $\xi'$. Suppose now vice versa that $\xi$ is dynamically related to $\xi'$. Then they lie in the same $G$-orbit on $\geo X$, $G\xi=G\xi'$, and there are points $\xi_n$ in this $G$-orbit such that $\xi_n\to\xi$ and $g_n\xi_n\to\xi'$. Let $a\subset\geo X$ be an apartment containing $\si_-$ and $\xi$. Nearby apartments $a_n$ containing $\xi_n$ can be obtained by using isometries $h_n\to e$ in $G$ with $\xi_n=h_n\xi$ and putting $a_n=h_n a$. (Note, however, that we cannot expect the $a_n$ to also contain $\si_-$; we have only the inequality $\pos(\xi_n,\si_-)\succeq\pos(\xi,\si_-)$ for large $n$.) Let $\hat\si_-\subset a$ be the chamber opposite to $\si_-$, and let $\si_n=h_n\hat\si_-\subset a_n$. Then $\si_n\to\hat\si_-$. The uniform convergence of $g_n$ near $\hat\si_-$ to the constant map $\equiv\si_+$, compare Lemma \[lem:unifattr\], implies that $g_n\si_n\to\si_+$. We observe that $$\pos(\xi',\si_+)\preceq\pos(g_n\xi_n,g_n\si_n) =\pos(\xi_n,\si_n) =\pos(h_n\xi,h_n\hat\si_-) =\pos(\xi,\hat\si_-)=\cpos(\xi,\si_-)$$ where the first inequality follows from Lemma \[lem:semcontrelpos\]. Hence (\[eq:dynrelposcond\]) holds. Let now $\Ga\subset G$ be a regular discrete subgroup. We define $\La^{(2)}_{ch}(\Ga)\subseteq\La_{ch}(\Ga)\times\La_{ch}(\Ga)$ as the set of pairs of chambers $(\si_+,\si_-)$ such that there exists a sequence $\ga_n\to\infty$ in $\Ga$ with $\ga_n^{\pm1}\to\si_{\pm}$. It is compact and invariant under switching the components of the pairs. Proposition \[prop:dynrel\] yields a complete description of the dynamical relations on the ideal boundary with respect to the $\Ga$-action: \[thm:dynrel\] Let $\Ga\subset G$ be a regular discrete subgroup. Two ideal points $\xi,\xi'\in\geo X$ are dynamically related with respect to the action $\Ga\acts\geo X$ if and only if there exists $(\si_+,\si_-)\in\La^{(2)}_{ch}(\Ga)$ such that $$\label{eq:dynrelposcondcop} \pos(\xi',\si_+)\preceq\cpos(\xi,\si_-)$$ holds. The points $\xi$ and $\xi'$ are dynamically related with respect to the $\Ga$-action if and only if they are dynamically related with respect to some sequence $\ga_n\to\infty$ in $\Ga$. After passing to a subsequence, we may assume that $(\ga_n^{\pm1})$ chamber converges, $\ga_n^{\pm1}\to\si_{\pm}$. The relative position inequality (\[eq:dynrelposcondcop\]) implies metric inequalities which are easier to apply, namely using (\[ineq:ordrelpos2\]) and (\[eq:complementrelpos\]) we obtain: $$\label{ineq:necdynrel2} \tangle(\xi,\cdot)\circ\kappa_{\si_-} +\tangle(\xi',\cdot)\circ\kappa_{\si_+}\circ\iota \leq\pi .$$ Note that this implies the estimate (\[ineq:keyest\]) and Lemma \[lem:nonwandcomp\] for regular groups. Theorem \[thm:dynrel\] yields: \[cor:dynrelmet\] Let $\Ga\subset G$ be a regular discrete subgroup. If two ideal points $\xi,\xi'\in\geo X$ are dynamically related with respect to the action $\Ga\acts\geo X$ then there exists $(\si_+,\si_-)\in\La^{(2)}_{ch}(\Ga)$ such that (\[ineq:necdynrel2\]) holds. Wandering set and domains of discontinuity {#sec:wanddomdisc} ------------------------------------------ We now look for new invariant open subsets, besides those already found in section \[sec:domdisc\], of $\geo X$ and of individual $G$-orbits $G\eta\subset\geo X$ on which $\Ga$ acts properly discontinuously. Such subsets are necessarily contained in the [*wandering set*]{} $\Om_{wand}(\Ga)$ which is the open subset of $\geo X$ consisting of the points which are not dynamically related to themselves. Theorem \[thm:dynrel\] yields a precise description for it: \[cor:nwdescr\] $\Om_{wand}(\Ga)=\geo X- \bigcup_{(\si_+,\si_-)\in\La^{(2)}_{ch}(\Ga)} \{\xi|\pos(\xi,\si_+)\preceq\cpos(\xi,\si_-)\}$ In general, the action on the wandering set is not properly discontinuous, see e.g. Example \[ex:nonundd\]. We will now provide different choices for domains of discontinuity inside $\Om_{wand}(\Ga)$. The description in Theorem \[thm:dynrel\] asserts that there are no dynamical relations between points which are in some sense far from the limit set. For instance, there are no dynamical relations between points outside the [*$\pihalf$-thickening*]{} $\bar N_{\pihalf}(\La(\Ga))$ of the limit set, i.e. $\geo X-\bar N_{\pihalf}(\La(\Ga))$ is a domain of discontinuity for $\Ga$. We proved this already in Theorem \[thm:pdgen\] for not necessarily regular discrete subgroups $\Ga$. By choosing a $G$-orbit $G\eta$ and transferring its domain of discontinuity, one obtains furthermore the domains of discontinuity $\geo X-\st(G\eta\cap\bar N_{\pihalf}(\La(\Ga)))$, see Corollary \[cor:pdtrans\]. Returning to regular discrete subgroups $\Ga$, we observe that for a $G$-orbit $G\eta$ of [*root*]{} type the domain of discontinuity $\geo X-\st(G\eta\cap\bar N_{\pihalf}(\La(\Ga)))$ depends only on the [*chamber*]{} limit set $\La_{ch}(\Ga)\subset\D_FX$. This is so because the $\pihalf$-balls centered at root type points $g\eta$ are subcomplexes, compare Example \[ex:rooth\]. We describe now other domains of discontinuity which depend only on the chamber limit set using the chamber thickenings defined in section \[sec:chthick\]. The following is the main result of this section. \[thm:pdth\] Let $\Ga\subset G$ be a regular discrete subgroup. For any fat thickening $\Th(\si_{mod})\subset W\bar\xi$ in a Weyl orbit $W\bar\xi\subset a_{mod}$, and for $\xi\in\geo X$ with $\theta(\xi)=\bar\xi$, the subset $G\xi-\Th(\La_{ch}(\Ga))$ is a domain of discontinuity for $\Ga$ in $G\xi$. Suppose that $\xi$ and $\xi'$ are dynamically related points in $G\xi$. Applying Theorem \[thm:dynrel\] and Lemma \[lem:posineqthick\] we see that at least one of the points $\xi$ and $\xi'$ must belong to $\Th(\La_{ch}(\Ga))$. Let $\Ga\subset G$ be a regular discrete subgroup. If $\Th(\si_{mod})\subset a_{mod}$ is a closed thickening whose intersection with any Weyl orbit is fat, then $\geo X-\Th(\La_{ch}(\Ga))$ is a domain of discontinuity for $\Ga$ in $\geo X$. $\Th(\La_{ch}(\Ga))$ is closed, because $\Th(\si_{mod})$ is closed in $a_{mod}$, cf. section \[sec:chthick\]. The assertion follows from Theorem \[thm:pdth\], since dynamical relations occur in $G$-orbits. Cocompact actions at infinity {#sec:cocompact} ============================= Let $\Ga\subset G$ be an infinite discrete subgroup. In this section, we look for domains in $G$-orbits $G\xi\subset\geo X$ at infinity where the $\Ga$-action is cocompact and, if possible, also proper. To obtain cocompactness on certain domains, we need to impose an asymptotic condition on $\Ga$. Following the definition of conicality in the rank one case, cf. section \[sec:rank1\], one may call a limit point $\xi\in\La$ [*ray conical*]{} if it is the limit of a sequence of orbit points $\ga_nx$ which are contained in a tubular neighborhood of some geodesic ray asymptotic to $\xi$. However, ray conicality is a too strong condition in the higher rank case. It is satisfied by convex-cocompact subgroups, which are rare [@convcoco], but one can construct many RCA Schottky subgroups which are not ray conical. For regular subgroups we relax the ray conicality condition as follows. \[dfn:chconical\] Let $\Ga\subset G$ be a regular discrete subgroup. We call a limit chamber $\si\in\La_{ch}(\Ga)$ [*conical*]{} if there exists a sequence $\ga_n\to\infty$ in $\Ga$ such that for some (any) base point $x\in X$ the sequence of orbit points $\ga_nx$ is contained in a tubular neighborhood of the Weyl sector $V(x,\si)$. We say that $\Ga$ has [*conical chamber limit set*]{} or is [*chamber conical*]{} if all limit chambers in $\La_{ch}(\Ga)$ are conical. In other words, $\si$ is a conical limit chamber of $\Ga$ if there exists a sequence $(\ga_n x)$ in the orbit $\Ga x$ which converges to $\si$ in the chamber radial topology. Chamber conicality is satisfied by a large class of discrete subgroups including regular Anosov subgroups and RCA Schottky subgroups, cf. Remarks \[rem:anos\] and \[rem:schott\]. We present two methods to prove cocompactness of actions at infinity for groups with conical chamber limit set. In section \[sec:dd\] we give a quick argument using Dirichlet domains at infinity which however yields only special results, cf. Theorem \[thm:coco\] and Corollary \[cor:coco\]. In section \[sec:expand\] we give an independent argument by proving an expansion property for the action on the limit set which implies a general cocompactness result, cf. Theorem \[thm:cocoslim\]. Dirichlet domains {#sec:dd} ----------------- We use Dirichlet domains at infinity as a tool to verify the cocompactness of $\Ga$-actions. For this purpose, we must estimate the region covered by the Dirichlet domains. This turns out to be closely related to a notion of stability motivated by geometric invariant theory. In some cases we can show that the Dirichlet domains precisely fill out a domain of discontinuity on which the $\Ga$-action is consequently cocompact. ### Stability {#sec:stab} The following notion of stability for points at infinity is motivated by Mumford stability in geometric invariant theory, as interpreted in the geometric language of Busemann functions in [@KLM]. Recall that $b_\xi$ denotes a Busemann function on $X$ associated with an ideal point $\xi\in \geo X$, cf. section \[sec:bus\]. We pick a base point $x\in X$ and call an ideal point $\xi\in\geo X$ \(i) [*stable*]{} with respect to the $\Ga$-action, if $b_{\xi}|_{\Ga x}$ has finite sublevels, i.e. if the horoballs centered at $\xi$ have finite intersection with the orbit $\Ga x$. \(ii) [*semistable*]{}, if $b_{\xi}|_{\Ga x}$ is bounded below, $\inf b_{\xi}|_{\Ga x}>-\infty$, i.e. if $\Ga x$ does not enter every horoball centered at $\xi$, and \(iii) [*unstable*]{}, if $\inf b_{\xi}|_{\Ga x}=-\infty$, i.e. if $\Ga x$ enters all horoballs centered at $\xi$. These stability properties are independent of the chosen orbit $\Ga x$. We are interested in the relation between the stability of an ideal point and its position relative to the limit set $\La=\La(\Ga):=\ol{\Ga x}\cap\geo X$ of $\Ga$. Points sufficiently far from the limit set are always stable: $$\label{eq:farawaystable} \geo X-\bar N_{\pihalf}(\La) \subseteq \geo X^{stable}$$ If $\xi$ is not stable, then some horoball $H$ centered at $\xi$ has infinite intersection with $\Ga x$. Hence $\La$ intersects $\geo H=\bar B(\xi,\pihalf)$ and $\xi\in\bar N_{\pihalf}(\La)$. ### Conicality and unstability {#sec:con} The link between stability and position relative to the limit set becomes stronger if one imposes suitable asymptotic conditions on $\Ga$: \[lem:rayconlimunst\] If $\Ga$ has ray conical limit set, then complementary to (\[eq:farawaystable\]) we also have: $$\label{eq:nearbyunstable} N_{\pihalf}(\La) \subseteq \geo X^{unstable}$$ Let $\la\in\La$ and $\xi\in\geo X$ with $\tangle(\xi,\la)<\pihalf$. According to the ray conical condition, there exists a ray $\rho$ asymptotic to $\la$ and a sequence $\ga_nx\to\la$ in $N_R(\rho)$ for some $R>0$. The Busemann function $b_{\xi}$ is asymptotically linear along $x\la$ with strictly negative slope because $\tangle(\xi,\la)<\pihalf$. It follows that $b_{\xi}(\ga_nx)\to-\infty$, i.e. $\xi$ is unstable. Hence, in this situation, the points at distance $>\pihalf$ from the limit set are stable, the points at distance $<\pihalf$ from the limit set are unstable, and the points at distance exactly $\pihalf$ from the limit set may have any type of stability. For regular subgroups, the weaker condition of chamber conicality implies the following version of Lemma \[lem:rayconlimunst\]: \[lem:conunst\] Suppose that $\Ga$ is regular. If $\si\in\La_{ch}(\Ga)$ is conical and if $\xi\in\geo X$ with $\si\subset\bar B(\xi,\pihalf)$, then $\xi$ is unstable. Let $F$ be a maximal flat in $X$ asymptotic to $\si$ and $\xi$, $\geo F\supset\si\cup\{\xi\}$, and let $x\in F$. According to the chamber conical condition there exists $R>0$ and a sequence of points $\ga_nx\to\infty$ in $N_R(V(x,\si))$. We denote by $x'_n$ the projection of $\ga_nx$ to $F$. Due to the regularity of $\Ga$, we have that $x'_n\in V(x,\si)$ for large $n$ and $d(x'_n,\D V(x,\si))\to+\infty$ as $n\to+\infty$. Note that $b_{\xi}|_F$ is linear because $\xi\in\geo F$. Since $\si\subset\bar B(\xi,\pihalf)$, the sublevels of $b_{\xi}|_F$ are half-spaces containing $\si$ in their ideal boundary hemisphere. It follows that $b_{\xi}(x'_n)\to-\infty$, and hence also $b_{\xi}(\ga_nx)\to-\infty$, i.e. $\xi$ is unstable. If $\Ga$ is uniformly regular with conical chamber limit set, then when restricting to certain $G$-orbits at infinity we get a sharper partition into stable and unstable points: Recall that an ideal point $\eta\in\geo X$ has root type if and only if the closed $\pihalf$-ball $\bar B(\eta,\pihalf)$ is a subcomplex with respect to the simplicial structure on $\geo X$. Equivalently, the $\pihalf$-sphere $S(\eta,\pihalf)=\bar B(\eta,\pihalf)-B(\eta,\pihalf)$ consists of singular points. Since the limit set contains only regular points due to uniform regularity, no root type point can have distance precisely $\pihalf$ from the limit set. Thus, for an ideal point $\eta$ of root type we have $$\label{eq:nopihalf} (\bar N_{\pihalf}(\La)-N_{\pihalf}(\La))\cap G\eta=\emptyset.$$ Furthermore, if a root type point $\eta$ has distance $<\pihalf$ from a limit point $\la$, then $\tangle(\eta, \xi)<\pihalf$ for all interior points $\xi$ of the limit chamber containing $\la$. Hence, Lemma \[lem:conunst\] implies that (\[eq:nearbyunstable\]) holds along root type $G$-orbits: $$\label{eq:nearbyunstableroot} N_{\pihalf}(\La)\cap G\eta\subseteq\geo X^{unstable}$$ The properties (\[eq:nopihalf\]) and (\[eq:nearbyunstableroot\]) allow together with (\[eq:farawaystable\]) to pin down on the $G$-orbit $G\eta$ the stable, semistable and unstable regions in terms of the distance from the limit set. \[prop:decoorbstunst\] If $\Ga$ is uniformly regular with conical chamber limit set and if $\eta\in\geo X$ has root type, then $G\eta$ decomposes into the unstable part $N_{\pihalf}(\La)\cap G\eta=\bar N_{\pihalf}(\La)\cap G\eta$ and the stable part $G\eta-N_{\pihalf}(\La)$, and all semistable points in $G\eta$ are stable. Since $\La$ is closed, also $\bar N_{\pihalf}(\La)$ is closed with respect to the visual topology, as a consequence of the lower semicontinuity of the Tits metric. Thus the unstable part is closed and the stable part is open in the manifold $G\eta$ (with respect to its visual topology which coincides with its manifold topology as a homogeneous $G$-space). ### Dirichlet domains at infinity {#sec:diri} The [*Dirichlet domain*]{} of a point $x\in X$ with respect to the $\Ga$-action on $X$, is the subset $$D_x=\{d(\cdot,x)\leq d(\cdot,\ga x)\;\forall\ga\in\Ga\}\subset X.$$ The Dirichlet domain is closed and, as a consequence of the triangle inequality, [*star-shaped*]{} with respect to $x$. Moreover, $\ga D_x=D_{\ga x}$ for all $\ga\in \Ga$, and $D_x$ is a fundamental domain for the action $\Ga\acts X$ if the stabilizer of $x$ in $\Ga$ is trivial. We call the ideal boundary $\geo D_x$ of $D_x$ the [*Dirichlet domain at infinity*]{} of the point $x\in X$ with respect to the $\Ga$-action. Being the ideal boundary of a subset of $X$, it is compact. Again, $\ga\cdot\geo D_x=\geo D_{\ga x}$. $\geo D_x$ consists of the centers of the horoballs touching $\Ga x$ in $x$, $$\geo D_x= \{\xi\in\geo X:b_{\xi}(x)\leq b_{\xi}(\ga x)\;\forall\ga\in\Ga\} .$$ The star-shape of $D_x$ with respect to $x$ implies that an ideal point $\xi$ belongs to $\geo D_x$ if and only if the ray $x\xi$ lies in $D_x$. The latter is equivalent to the property that the balls with center on $x\xi$ and with $x$ in their boundary contain no points of $\Ga x$ in their interiors, equivalently, if the horoball centered at $\xi$ bounded by the horosphere through $x$ contains no points of $\Ga x$ in its interior, i.e. if $b_{\xi}|_{\Ga x}$ attains a minimum in $x$. As a consequence, the region covered by the Dirichlet domains at infinity can be described as $$\Ga\cdot\geo D_x= \cup_{\ga\in\Ga}\geo D_{\ga x}= \{\xi\in\geo X:{b_{\xi}}|_{\Ga x}\hbox{ attains a minimum}\} .$$ This shows that the location of this region is related to stability; it is enclosed between the stable and the semistable parts: $$\label{eq:diriencl} \geo X^{stable}\subseteq \Ga\cdot\geo D_x \subseteq\geo X^{semistable}$$ We combine this with our previous discussion of the relation between stability and the limit set. If $\Ga$ has ray conical limit set, then (\[eq:farawaystable\]) and (\[eq:nearbyunstable\]) yield: $$\label{eq:coverstable} \geo X-\bar N_{\pihalf}(\La) \subseteq \Ga\cdot\geo D_x \subseteq \geo X-N_{\pihalf}(\La)$$ If $\Ga$ is uniformly regular with conical chamber limit set and if $\eta\in\geo X$ has root type, then according to Proposition \[prop:decoorbstunst\] semistable points in $G\eta$ are stable and (\[eq:diriencl\]) yields that the Dirichlet domains at infinity cover precisely the stable part of $G\eta$, $$\label{eq:covstab} \Ga\cdot\geo D_x\cap G\eta =G\eta-N_{\pihalf}(\La)=G\eta-\bar N_{\pihalf}(\La) .$$ ### Some cocompact domains {#sec:coco} Since the Dirichlet domains at infinity are compact, the $\Ga$-action on their union $\Ga\cdot\geo D_x$ is cocompact (but possibly not proper and the union might not be open!) and, in particular, its restriction to the intersections $\Ga\cdot\geo D_x\cap G\xi$ with $G$-orbits is also cocompact. Under suitable assumptions, these intersections turn out to be open subsets of $G\xi$ and domains of discontinuity for the $\Ga$-action. A [*cocompact domain*]{} for $\Ga$ is an open $\Ga$-invariant subset on which $\Ga$ acts cocompactly. \[thm:coco\] If $\Ga$ is uniformly regular with conical chamber limit set and if $\eta\in\geo X$ is of root type, then $G\eta-N_{\pihalf}(\La)=G\eta-\bar N_{\pihalf}(\La)$ is a cocompact domain for $\Ga$ in $G\eta$. This is a consequence of (\[eq:covstab\]). We know from Theorem \[thm:pdgen\] that the $\Ga$-action on $\geo X-\bar N_{\pihalf}(\La)$ is properly discontinuous, and hence also its restriction to $G\eta-\bar N_{\pihalf}(\La)$. We deduce from Theorem \[thm:coco\] the cocompactness of the $\Ga$-action on related domains in other $G$-orbits at infinity and in $\geo X$. Recall the definition of stars from section \[sec:star\]. \[prop:transfcoco\] If $\Om\subset G\xi$ is a cocompact domain for $\Ga$ in $G\xi$, then $\st(\Om)\subset\geo X$ is a cocompact domain for $\Ga$ in $\geo X$. It is immediate that the action of $\Ga$ on $\st(\Om)$ is cocompact. To see that $\st(\Om)$ is open, we observe that nearby closed chambers intersect $G\xi$ in nearby points. Note that this transfer domain is larger than the transfer domain constructed in Proposition \[prop:transf\], $\geo X-\st(G\xi-\Om)\subset\st(\Om)$. Both transfer domains agree on $\ost(G\xi)$. Indeed, since the inverse images of subsets of $G\xi$ under the natural equivariant fibration $\ost(G\xi)\to G\xi$, cf. (\[eq:fibost\]), are their open stars, we have that $\ost(G\xi)-\st(G\xi-\Om)=\ost(\Om)=\ost(G\xi)\cap\st(\Om)$. However, the transfer domains differ on the $G$-orbits outside $\ost(G\xi)$. Compare the discussion in section \[sec:star\] and the comment after Proposition \[prop:transf\]. Theorem \[thm:coco\] and Proposition \[prop:transfcoco\] imply: \[cor:coco\] If $\Ga$ is uniformly regular with conical chamber limit set and if $\eta\in\geo X$ is of root type, then $\st(G\eta-N_{\pihalf}(\La))=\st(G\eta-\bar N_{\pihalf}(\La))$ is a cocompact domain for $\Ga$ in $\geo X$. In particular, the intersection of $\st(G\eta-N_{\pihalf}(\La))$ with any $G$-orbit $G\xi$ is a cocompact domain for $\Ga$ in $G\xi$. If $G\xi\subset\ost(G\eta)$, then we know from Corollary \[cor:pdtrans\] and Remark \[rem:ostdec\] that the cocompact domain $$G\xi\cap\st(G\eta-\bar N_{\pihalf}(\La(\Ga)))= G\xi-\st(G\eta\cap\bar N_{\pihalf}(\La(\Ga)))$$ is also a domain of discontinuity for $\Ga$. In rank one, our discussion recovers well-known results on Kleinian groups: The Tits metric is discrete (it takes only the values $0$ and $\pi$) and $\bar N_{\pihalf}(\La)=N_{\pihalf}(\La)=\La$. By (\[eq:farawaystable\]), the points in $\geo X-\La$ are always stable. If the limit set $\La$ of $\Ga$ is conical, then it consists of unstable points, cf. (\[eq:nearbyunstable\]). Moreover, the Dirichlet domains at infinity cover precisely $\Ga\cdot\geo D_x=\geo X-\La$, see (\[eq:coverstable\]). It follows that the action of $\Ga$ on $\geo X-\La$ is cocompact. It is also properly discontinuous, cf. Theorem \[thm:pdgen\]. Expansion {#sec:expand} --------- In this section, we prove our main result concerning the cocompactness of $\Ga$-actions at infinity, see Theorem \[thm:cocoslim\]. In section \[sec:conical-&gt;expanding\] we first verify that, under suitable assumptions, the $\Ga$-action on the Fürstenberg boundary $\D_FX$ is expanding at the chamber limit set in the sense of Sullivan [@Sullivan], compare section \[sec:trexpand\]. We then proceed to showing that the $\Ga$-actions on all $G$-orbits $G\xi\subset\geo X$ are expanding, transversally to certain natural fibrations, at all slim thickenings of the chamber limit set $\La_{ch}$. This implies the cocompactness of the $\Ga$-actions on their complements. In section \[sec:nearbythickenings\], we establish certain preliminary geometric results needed for the proofs in section \[sec:conical-&gt;expanding\]. ### Nearby chamber thickenings {#sec:nearbythickenings} We consider now the (generalized) partial flag manifolds $\flag(\tau_{mod})$ consisting of the faces $\tau\subset\geo X$ with type $\theta(\tau)=\tau_{mod}\subset\si_{mod}$, cf. section \[sec:symm\]. In particular, $\flag(\si_{mod})=\D_FX$. For any two incident face types $\ups_{mod}\subset\tau_{mod}\subset\si_{mod}$ there is the natural forgetful map $$\pi_{\ups_{mod}\tau_{mod}}:\flag(\tau_{mod})\to\flag(\ups_{mod})$$ assigning to a face $\tau$ of type $\tau_{mod}$ its face $\ups$ of type $\ups_{mod}$. It is an $G$-equivariant smooth fibration with compact base and fiber. We fix auxiliary Riemannian metrics on all flag manifolds $\flag(\tau_{mod})$. Thereby, also the $G$-orbits $G\xi\subset\geo X$ at infinity are equipped with Riemannian metrics by equivariantly identifying them with the appropriate flag manifolds. The fibrations $\pi_{\ups_{mod}\tau_{mod}}$ are Lipschitz continuous by compactness. Vice versa, we have the following simple observation regarding controlled lifts: \[lem:liftsimpnear\] Let $\tau$ and $\ups'$ be faces of types $\tau_{mod}$ and $\ups_{mod}$, $\ups_{mod}\subset\tau_{mod}$, and let $\ups\subset\tau$ be the face of type $\ups_{mod}$. Then there exists a face $\tau'\supset\ups'$ of type $\tau_{mod}$ such that $$d(\tau',\tau)\leq C_0\cdot d(\ups',\ups)$$ with a uniform constant $C_0\geq1$ only depending on the chosen Riemannian metrics. The Riemannian metrics on $\flag(\tau_{mod})$ and $\flag(\ups_{mod})$ can be chosen so that $\pi_{\ups_{mod}\tau_{mod}}$ becomes a Riemannian submersion. In this case, there exists $\tau'$ so that $d(\tau,\tau')=d(\ups,\ups')$. For other choices of the metrics, a multiplicative factor enters. This lemma generalizes (by induction along galleries) to: \[lem:liftgallnear\] Let $\tau,\tau'$ be faces of type $\tau_{mod}$ and let $\tilde\tau$ be a face of type $\tilde\tau_{mod}$. Then there exists another face $\tilde\tau'$ of type $\tilde\tau_{mod}$ with relative position $\pos(\tilde\tau',\tau')=\pos(\tilde\tau,\tau)$ such that $$d(\tilde\tau',\tilde\tau) \leq C_1\cdot d(\tau',\tau)$$ with a uniform constant $C_1\geq1$ only depending on the chosen Riemannian metrics. Now we consider a $G$-orbit $G\xi\subset\geo X$ of type $\bar\xi=\theta(\xi)$ and a thickening $\Th(\si_{mod})\subset W\bar\xi$ of $\si_{mod}$ inside the corresponding Weyl orbit $W\bar\xi$. \[lem:findsi\] The following assertions hold with a uniform constant $C\geq1$ only depending on the chosen Riemannian metrics: \(i) For a point $\xi'\in G\xi$ and a chamber $\si\in\D_FX$ exists a chamber $\si'\in\D_FX$ such that $\xi'\in\Th(\si')$ and $$d(\si',\si)\leq C\cdot d(\xi',\Th(\si)) .$$ (ii) The Hausdorff distance of the thickenings of any two chambers $\si',\si\in\D_FX$ is controlled by $$d_H(\Th(\si'),\Th(\si))\leq C\cdot d(\si',\si) .$$ \(i) Suppose that $\xi\in\Th(\si)$ is the point closest to $\xi'$, i.e. $d(\xi',\xi)=d(\xi',\Th(\si))$. Lemma \[lem:liftgallnear\] yields a chamber $\si'$ with $\pos(\xi',\si')=\pos(\xi,\si)\in\Th(\si_{mod})$ and controlled distance $d(\si',\si)\leq C\cdot d(\xi',\xi)$, whence the first inequality. \(ii) If $\eta\in\Th(\si)$ is arbitrary, then applying Lemma \[lem:liftgallnear\] again yields a point $\eta'$ with $\pos(\eta',\si')=\pos(\eta,\si)\in\Th(\si_{mod})$ and controlled distance $d(\eta',\eta)\leq C\cdot d(\si,\si')$. ### Chamber conical implies transversely expanding {#sec:conical->expanding} Let $\Ga\subset G$ be a regular discrete subgroup. \[prop:conicimplexpa\] If the limit chamber $\si\in\La_{ch}(\Ga)$ is conical, then the action $\Ga\acts\D_FX$ is expanding at $\si$. In particular, if $\Ga$ has conical chamber limit set $\La_{ch}(\Ga)\subset\D_FX$, then the action $\Ga\acts\D_FX$ is expanding at $\La_{ch}(\Ga)$. By chamber conicality, there exists a sequence $\ga_n\to\infty$ in $\Ga$ such that for a(ny) point $x\in X$ the sequence of orbit points $\ga_nx$ is contained in a tubular neighborhood of the Weyl sector $V(x,\si)$. Let $F\subset X$ be a maximal flat asymptotic to $\si$, $\si\subset\geo F$. One can write $\ga_n=\tau_n\beta_n$ with transvections $\tau_n\in G$ along $F$ and a bounded sequence of isometries $\beta_n\in G$. Since $\Ga$ is regular by assumption, the sequence $(\tau_n)$ in $G$ is regular and we have chamber convergence $\tau_n\to\si$, cf. section \[sec:chconv\]. It follows e.g. from Lemma \[lem:unifattr\] that the differentials of $\tau_n$ on $\D_FX$ contract at their fixed point $\si$ arbitrarily strongly as $n\to+\infty$. This in turn implies that the differentials of the $\ga_n^{-1}=\beta_n^{-1}\tau_n^{-1}$ expand arbitrarily strongly at $\si$ as $n\to+\infty$ (with respect to the chosen auxiliary metric on $\D_FX$). Thus for $n$ sufficiently large, the element $\ga_n^{-1}$ is uniformly expanding on a neighborhood of $\si$. The same argument shows more generally an expansion property for the $\Ga$-actions on all $G$-orbits. Let $G\xi\subset\geo X$ be the $G$-orbit of type $\bar\xi=\theta(\xi)$. Let $\La_{\bar\xi}(\Ga)\subset G\xi$ be the image of $\La_{ch}(\Ga)$ under the natural projection $\pi_{\bar\xi}:\D_FX\to G\xi$. Then for any conical limit chamber $\si\in\La_{ch}(\Ga)$ the action $\Ga\acts G\xi$ is expanding at the point $\pi_{\bar\xi}\si$. In particular, if $\Ga$ has conical chamber limit set, then the action $\Ga\acts G\xi$ is expanding at $\La_{\bar\xi}(\Ga)$. Together with Proposition \[prop:trexpcoco\] (its special case for expanding actions) this implies that the action $\Ga\acts G\xi-\La_{\bar\xi}(\Ga)$ is cocompact. However, this action is in general not properly discontinuous. Now we generalize Proposition \[prop:conicimplexpa\] to thickenings of the chamber limit set in arbitrary $G$-orbits at infinity. Note that at thickenings we cannot expect expansion any more because they contain in general saddle fixed points of elements of $\Ga$. We consider a $G$-orbit $G\xi\subset\geo X$ of type $\bar\xi=\theta(\xi)\in\si_{mod}$ and a thickening $\Th(\si_{mod})\subset W\bar\xi$ of $\si_{mod}$ inside the corresponding Weyl orbit $W\bar\xi\subset a_{mod}$, cf. Definition \[dfn:thick\]. The thickenings $\Th(\si)\subset G\xi$ of the limit chambers $\si\in\La_{ch}(\Ga)$ are compact subvarieties (finite unions of Schubert cycles), and the thickening $\Th(\La_{ch}(\Ga))=\cup_{\si\in\La_{ch}(\Ga)}\Th(\si)$ of the chamber limit set is a compact subset. \[dfn:rca\] We call a regular discrete subgroup $\Ga\subset G$ [*antipodal*]{}, if any two distinct limit chambers in $\La_{ch}(\Ga)$ are opposite to each other. If in addition $\Ga$ is chamber conical, we call it [*RCA*]{}. If $\Ga\subset G$ is antipodal and if the thickening $\Th(\si_{mod})\subset W\bar\xi$ is slim in the sense of Definition \[dfn:fatslimth\], then the thickenings $\Th(\si)$ of the limit chambers $\si\in\La_{ch}(\Ga)$ are pairwise disjoint and we therefore have the natural map $$\pi:\Th(\La_{ch}(\Ga))\to\La_{ch}(\Ga)$$ projecting $\Th(\si)$ to $\si$. It is a fibration; we only need: $\pi$ is continuous. If $\xi_n\to\xi$ and $\si_n\to\si$ with $\xi_n\in\Th(\si_n)$, then $\xi\in\Th(\si)$ by semicontinuity of relative position, cf. Lemma \[lem:semcontrelpos\]. \[prop:conicimplexpatr\] Suppose that the regular discrete subgroup $\Ga\subset G$ is antipodal and that $\Th(\si_{mod})\subset W\bar\xi$ is a slim thickening. If the limit chamber $\si\in\La_{ch}(\Ga)$ is conical, then the action $\Ga\acts G\xi$ is expanding transversely to $\pi$ at $\Th(\si)$. In particular, if $\Ga$ is RCA, then the action $\Ga\acts G\xi$ is expanding at $\Th(\La_{ch}(\Ga))$ transversely to $\pi$. We need to verify the expansion inequality (\[ineq:trexpan\]). Let $V\subset G\xi$ be a neighborhood of $\Th(\si)$. Let $\xi\in V$ be a point and let $\si'\in\La_{ch}(\Ga)$ be a limit chamber such that $\Th(\si')\subset V$. According to the first part of Lemma \[lem:findsi\], there exists a chamber $\si''$ such that $\xi\in\Th(\si'')$ and $$d(\si'',\si')\leq C\cdot d(\xi,\Th(\si'))$$ Now let $\ga_n=\tau_n\beta_n$ be a sequence for the chamber $\sigma$ as in the proof of Proposition \[prop:conicimplexpa\]. Furthermore, let $W_n\subset\D_FX$ be an open neighborhood of $\si$, as provided by Proposition \[prop:conicimplexpa\], where $\tau_n$ contracts uniformly with contraction factor $\la_n\to0$. Fix a large $n$. If $V$ is small enough (depending on $n$), then $\si'',\si'\in W_n$ and $$d(\tau_n\si'',\tau_n\si')\leq\la_n\cdot d(\si'',\si') .$$ Applying the second part of Lemma \[lem:findsi\], we obtain $$d(\tau_n\xi,\Th(\tau_n\si')) \leq d_H(\Th(\tau_n\si''),\Th(\tau_n\si')) \leq C\cdot d(\tau_n\si'',\tau_n\si')$$ and hence $$d(\tau_n\xi,\Th(\tau_n\si'))\leq C^2\la_n\cdot d(\xi,\Th(\si'))$$ This implies that for $n$ sufficiently large, the element $\ga_n^{-1}$ satisfies the expansion property (\[ineq:trexpan\]) on a sufficiently small neighborhood (namely $\tau_nV$) of $\Th(\si)$. The result holds more generally without assuming antipodality as soon as the thickenings $\Th(\si)$ of individual limit chambers $\si\in\La_{ch}(\Ga)$ are disjoint and we hence have a natural fibration $\Th(\La_{ch}(\Ga))\to\La_{ch}(\Ga)$. Propositions \[prop:trexpcoco\] and \[prop:conicimplexpatr\] imply our main cocompactness result: \[thm:cocoslim\] Suppose that $\Ga\subset G$ is an RCA subgroup, and let $\bar\xi=\theta(\xi)$. Then for every slim thickening $\Th(\si_{mod})\subset W\bar\xi$ the action $\Ga\acts G\xi-\Th(\La_{ch}(\Ga))$ is cocompact. A special case of this result was Corollary \[cor:coco\]. Combining our main results on proper discontinuity and cocompactness, Theorems \[thm:pdth\] and \[thm:cocoslim\], we obtain: \[thm:cocodiscont\] If $\Ga\subset G$ is an RCA subgroup and if $\Th(\si_{mod})\subset W\bar\xi$ is a balanced thickening, then for $\xi\in\geo X$ with $\theta(\xi)=\bar\xi$ the open subset $G\xi-\Th(\La_{ch}(\Ga))$ is a cocompact domain of discontinuity for $\Ga$ in $G\xi$. \[rem:ddreg\] By Lemma \[lem:balthreg\], the regular Weyl orbits in $a_{mod}$ admit balanced thickenings, and Theorem \[thm:cocodiscont\] therefore provides cocompact domains of discontinuity for $\Ga$ in $\D_FX$. Nonemptyness of cocompact domains of discontinuity {#sec:ccddne} -------------------------------------------------- Let $\Ga\subset G$ be an RCA subgroup. If some $G$-orbit $G\xi\subset\geo X$ contains a nonempty cocompact domain of discontinuity for $\Ga$, then so does $\D_FX$ because it $G$-equivariantly fibers over $G\xi$ with compact fiber. Cocompact domains of discontinuity in $\D_FX$ are provided by Theorem \[thm:cocodiscont\], cf. Remark \[rem:ddreg\]. The following observation explains why it is hard for these domains to be empty. \[prop:neccdd\] If the cocompact domains of discontinuity for $\Ga$ in $\D_FX$ provided by Theorem \[thm:cocodiscont\] are all empty, then for every $\iota$-invariant type $\vartheta\in\si_{mod}$ exists a packing of $\geo X$ by a compact family of $\pihalf$-balls with centers of type $\vartheta$. The compactness of the family of balls refers to the visual topology; the balls are parametrized by the compact chamber limit set $\La_{ch}(\Ga)$. That they form a [*packing*]{} of $\geo X$ means that, with respect to the Tits metric, the open $\pihalf$-balls are disjoint and the closed $\pihalf$-balls cover. We consider the metric thickening $\Th^{met}_{\vartheta,\pihalf}(\si_{mod}) =\bar B(\vartheta,\pihalf)\subset a_{mod}$ of type $\vartheta$. Its intersection $W\bar\xi\cap B(\vartheta,\pihalf)$ with every Weyl orbit $W\bar\xi\subset a_{mod}$, which avoids the great sphere $\D B(\vartheta,\pihalf)$, is a balanced thickening inside $W\bar\xi$, compare Example \[ex:metthick\] and the proof of Lemma \[lem:balthreg\]. The corresponding $G$-orbits $G\xi$ fill out a dense open subset of $\geo X$. If the cocompact domains of discontinuity $G\xi-\Th^{met}_{\vartheta,\pihalf}(\La_{ch}(\Ga))$ in these $G$-orbits provided by Theorem \[thm:cocodiscont\] are all empty, $G\xi\subset\Th^{met}_{\vartheta,\pihalf}(\La_{ch}(\Ga))$, then $\Th^{met}_{\vartheta,\pihalf}(\La_{ch}(\Ga))=\geo X$ because $\Th^{met}_{\vartheta,\pihalf}(\La_{ch}(\Ga))$ is closed. The thickenings $\Th^{met}_{\vartheta,\pihalf}(\si)$ of the individual limit chambers $\si\in\La_{ch}(\Ga)$ are the $\pihalf$-balls around the points of type $\vartheta$ in the limit chambers. Their centers are pairwise antipodal because $\Ga$ is antipodal and $\vartheta$ is $\iota$-invariant. Therefore these balls have pairwise disjoint interiors and form a packing of $\geo X$. They are parametrized by the compact set $\La_{ch}(\Ga)$. In the irreducible higher rank case, the existence of such packings can be easily ruled out for most Weyl groups. We start with the $A_2$ case. \[lem:nopacka2\] If $W=A_2$, then $\geo X$ admits no packing by a compact family of $\pihalf$-balls. It is convenient to use the interpretation of the type $A_2$ spherical building structure on $\geo X$ as an abstract [*topological projective plane*]{} ${\mathcal P}$: There are two singular $G$-orbits in $\geo X$. The ideal points in one of them can be regarded as the points in ${\mathcal P}$, and the ideal points in the other as the lines; the Weyl chambers (arcs) then correspond to flags, i.e. pointed lines in ${\mathcal P}$. The topology on ${\mathcal P}$ is induced by the visual topology on $\geo X$. Note that the plane ${\mathcal P}$ is connected, because it is a homogeneous space for the identity component $\Isom(X)_o$ of $G$. Suppose now that there exists a packing of $\geo X$. The centers of the packing balls are pairwise antipodal and must therefore be midpoints of Weyl chambers (arcs); they correspond to flags in this projective plane. We call these flags and their lines and points [*special*]{}. We observe that, since the special flags are pairwise in general relative position, every special line contains exactly one special point. Furthermore, nonspecial points correspond to boundary vertices of packing balls, and hence all lines through nonspecial points are special. Since the packing balls form a compact family, the condition “special” is closed and “nonspecial” is open. On the other hand, take a special line and on it a nonspecial point. A small perturbation of this flag leaves the point nonspecial. It follows that the perturbed line remains special. Hence the condition “special” is also open. This contradicts the connectedness of the plane ${\mathcal P}$ because both conditions “special” and “nonspecial” are nonempty. For most irreducible higher rank Weyl groups, the question can be reduced to the $A_2$ case. \[thm:nopackmost\] If $X$ is irreducible of higher rank with Weyl group $\neq B_2,G_2$ then for some root type $\vartheta$ there exists no packing of $\geo X$ by a compact family of $\pihalf$-balls with centers of type $\vartheta$. We make use of the spherical building geometry of $\tits X$, see [@qirigid] for a detailed discussion. The question of nonexistence can be reduced to lower rank by observing that $\pihalf$-ball packings of spherical buildings induce such packings of their links. The links $\Si_{\xi}\tits X$ of points $\xi\in\geo X$ carry again natural spherical building structures. Furthermore, the links of $\pihalf$-balls $B(\zeta,\pihalf)$ at boundary points $\xi\in\D B(\zeta,\pihalf)$ are again $\pihalf$-balls, $\Si_{\xi}B(\zeta,\pihalf)=B(\oa{\xi\zeta},\pihalf)$; this follows from the first variation formula in $S^2$, because for any point $\eta$ sufficiently close to $\xi$ the three points $\xi,\eta,\zeta$ are the vertices of an embedded spherical triangle. If two balls $B(\zeta_i,\pihalf)$ have disjoint interiors and if $\xi$ is a point in the intersection of their boundaries, then the links $\Si_{\xi}B(\zeta_i,\pihalf)$ have disjoint interiors. Suppose now that we are given a packing of $\geo X$ by a compact family ${\mathcal B}$ of $\pihalf$-balls. Our discussion shows that it induces packings of the links $\Si_{\xi}\tits X$ of all boundary points $\xi$ of the packing balls by families ${\mathcal B}_{\xi}$ of $\pihalf$-balls in the links. To see that the closedness of ${\mathcal B}$ implies the closedness of the induced families ${\mathcal B}_{\xi}$, consider a convergent sequence $\zeta_n\to\zeta$ of centers of packing balls in ${\mathcal B}$ such that $\tangle(\xi,\zeta_n)=\pihalf$ for all $n$. Then $\tangle(\xi,\zeta)\leq\pihalf$ by the semicontinuity of Tits distance. However, strict inequality is impossible, because then $\xi$ would be an interior point of the packing ball $B(\zeta,\pihalf)$, which is absurd. Thus also $\tangle(\xi,\zeta)=\pihalf$ and it follows that $\oa{\xi\zeta_n}\to\oa{\xi\zeta}$. Hence the centers $\oa{\xi\zeta_n}$ of packing balls in ${\mathcal B}_{\xi}$ converge to the center of such a ball. This shows that also the families ${\mathcal B}_{\xi}$ are closed. Let us now focus on simplicial packings. The simplicial $\pihalf$-balls in $\geo X$ are precisely the $\pihalf$-balls centered at points of root type. Hence, if there exist packings of $\geo X$ for all (of the at most two) root types, then there exist packings of links $\Si_{\xi}\tits X$ for vertices $\xi$ of all possible vertex types. The induced packings have again root type because they are simplicial. Moreover, if in a link $\Si_{\xi}\tits X$ exist several (two) root types, then both can be realized by induced packings. The type of a vertex $\xi$ corresponds to a wall of the fundamental Weyl chamber, and the Dynkin diagram for the link $\Si_{\xi}\tits X$ is obtained from the Dynkin diagram for $\tits X$ by removing the corresponding node. Based on Lemma \[lem:nopacka2\], we arrive at a contradition for all Dynkin diagrams which can be reduced to $A_2$ inductively by removing nodes, i.e. which contain a simple edge. This is the case for the diagrams $A_{n\geq2}$, $B_{n\geq3}=C_{n\geq3}$, $D_{n\geq4}$, $E_{6,7,8}$ and $F_4$. Combining Proposition \[prop:neccdd\] and Theorem \[thm:nopackmost\] we conclude: \[thm:nedoman\] If $X$ is irreducible of higher rank with Weyl group $\neq B_2,G_2$ and if $\Ga\subset G$ is an RCA subgroup, then some of the cocompact domains of discontinuity for $\Ga$ in $\D_FX$ provided by Theorem \[thm:cocodiscont\] are nonempty. Chamber conicality versus hyperbolicity {#sec:hyp} ======================================= We compare various asymptotic properties for non-elementary regular subgroups. In particular, we show that RCA subgroups are word hyperbolic as abstract groups and have good boundary behavior. Antipodal regular subgroups --------------------------- We assume that $\Ga\subset G$ is a non-elementary antipodal regular discrete subgroup. Because of antipodality, Lemma \[lem:unifattr\] implies that the action $\Ga\acts\La_{ch}(\Ga)\subset\D_FX$ is a convergence action. In order to apply Bowditch’s dynamical characterization of hyperbolic groups, we compare intrinsic and extrinsic notions of conicality. \[Chamber conical equivalent to intrinsically conical\] \[lem:concon\] A limit chamber $\si\in\La_{ch}(\Ga)$ is conical if and only if it is intrinsically conical for the action $\Ga\acts\La_{ch}(\Ga)$. Let us first assume that $\si$ is conical. Since $\Ga$ is non-elementary, there exists another limit chamber $\hat\si\neq\si$. By antipodality, $\hat\si$ is opposite to $\si$ and there exists a unique maximal flat $F(\si,\hat\si)$ asymptotic to $\si$ and $\hat\si$. Since $\si$ is conical, there exists a sequence $\ga_n\to\infty$ in $\Ga$ such that for a point $x\in F(\si,\hat\si)$ the points $\ga_nx$ are contained in a tubular neighborhood of the Weyl sector $V(x,\si)$. It follows that we can write the isometries $\ga_n$ in the form $\ga_n=\tau_n\beta_n$ with transvections $\tau_n\to\infty$ along $F(\si,\hat\si)$ and uniformly bounded isometries $\beta_n\in G$. Clearly, $\tau_n^{-1}\si=\si$. Moreover, since the sequence $(\ga_n)$ is regular and $(\beta_n)$ is bounded, the $(\tau_n^{\pm})$ are regular sequences of transvections along $F$, and hence $\tau_n^{-1}|_{C(\si)}\to\hat\si$ locally uniformly. In particular, $\tau_n^{-1}|_{\La_{ch}(\Ga)-\{\si\}}\to\hat\si$ locally uniformly, again by antipodality. Since $\ga_n^{-1}=\beta_n^{-1}\tau_n^{-1}$, after passing to a subsequence, we have that $(\ga_n^{-1}\si)$ converges and $\ga_n^{-1}|_{\La_{ch}(\Ga)-\{\si\}}$ converges locally uniformly to a constant map with value $\neq\lim_{n\to\infty}\ga_n^{-1}\si$. Thus $\si$ is intrinsically conical. Let us now assume that $\si$ is intrinsically conical. By the definition of intrinsic conicality, there exists a sequence $\ga_n\to\infty$ in $\Ga$ such that the sequence of chambers $\gamma_n^{-1}\si$ converges and the sequence of maps $\gamma_n^{-1}|_{\La_{ch}(\Ga)-\{\si\}}$ converges (uniformly on compacta) to a constant map with value $\neq\lim_{n\to\infty}\ga_n^{-1}\si$. Let $F(\si,\hat\si)\subset X$ be the maximal flat asymptotic to $\si$ and to another limit chamber $\hat\si\in\La_{ch}(\Ga)$. (By antipodality, $\hat\si$ must be opposite to $\si$.) Since $\gamma_n^{-1}\si$ and $\gamma_n^{-1}\hat\si$ converge to opposite (limit) chambers, it follows that the flats $\ga_n^{-1}F(\si,\hat\si)$ converge (to a maximal flat). This implies that the elements $\ga_n$ have the form $\ga_n=\tau_n\beta_n$ with transvections $\tau_n\to\infty$ along $F(\si,\hat\si)$ and uniformly bounded isometries $\beta_n\in G$. Hence a sequence $(\ga_nx)$ stays (for every limit chamber $\hat\si\neq\si$) in a tubular neighborhood of the maximal flat $F(\si,\hat\si)$. After passing to a subsequence, we may assume that the regular sequence $(\ga_nx)$ chamber converges, $\ga_nx\to\si_{\infty}$. Since $(\ga_nx)$ stays Hausdorff close to the maximal flat $F(\si,\hat\si)$, we have that $\si_{\infty}\subset\geo F(\si,\hat\si)$, cf. Lemma \[lem:chconvwd\], and that $(\ga_nx)$ is contained in a tubular neighborhood of the sector $V(x,\si_{\infty})$, i.e. $\si_{\infty}$ is a conical limit chamber. Now, antipodality becomes important. It implies that $\geo F(\si,\hat\si)$ contains no limit chambers besides $\si$ and $\hat\si$, i.e. $\si_{\infty}$ must equal one of them. Since by assumption there are at least three limit chambers, we may repeat the same argument after replacing $\hat\si$ by another limit chamber opposite to $\si$. It follows that $\si_{\infty}=\si$. Thus, $\si$ is conical. \[lem:perfect\] The chamber limit set of a non-elementary antipodal regular discrete subgroup $\Ga\subset G$ is perfect. This is a consequence of the attraction-repulsion dynamics, as described by Lemma \[lem:unifattr\]. Indeed, let $\si_+$ be an arbitrary limit chamber. There exists a sequence $\ga_n\to\infty$ in $\Ga$ such that we have chamber convergence $\ga_n^{\pm1}\to\si_{\pm}$. Then according to Lemma \[lem:unifattr\], $\ga_n^{\pm1}\to\si_{\pm}$ locally uniformly as maps on the open Schubert cell $C(\si_{\mp})$ and, due to antipodality, in particular on $\La_{ch}(\Ga)-\{\si_{\mp}\}$. Since there exists a third limit chamber, its $\Ga$-orbit accumulates at $\si_{\pm}$. Thus, $\La_{ch}(\Ga)$ is perfect. \[thm:regopphyp\] Let $\Ga\subset G$ be a non-elementary antipodal regular discrete subgroup. Then the following are equivalent: \(i) $\Ga$ is chamber conical, and hence RCA. \(ii) $\Ga$ is word hyperbolic and $\La_{ch}(\Ga)$ is $\Ga$-equivariantly homeomorphic to $\geo\Ga$. We call subgroups $\Ga$ as in part(ii) [*asymptotically embedded*]{}, compare Definition \[defn:hyperbolic\]. By Lemma \[lem:concon\], property (i) is equivalent to $\La_{ch}(\Ga)$ being intrinsically conical. Suppose that $\La_{ch}(\Ga)$ is intrinsically conical. According to Lemma \[lem:perfect\], $\La_{ch}(\Ga)$ is perfect. Therefore, we can apply Bowditch’s Theorems \[thm:bowditch-conical\] and \[thm:charhypbow\]. ($\La_{ch}(\Ga)$ is metrizable as a subset of the manifold $\D_FX$.) Theorem \[thm:bowditch-conical\] implies that the convergence action $\Ga\acts\La_{ch}(\Ga)$ is uniform, and property (ii) follows from Theorem \[thm:charhypbow\]. The converse implication holds because the action of a word hyperbolic group on its Gromov boundary is intrinsically conical. Alternatively, one can prove that the conicality of the chamber limit set implies its intrinsic conicality as follows. If $\Ga$ has conical chamber limit set, then Proposition \[prop:conicimplexpa\] implies that the action $\Ga \acts \La_{ch}(\Ga)$ is expanding (with respect to any Riemannian metric on $\D_FX$). As in the proof of Theorem \[thm:regopphyp\] one verifies that $\La_{ch}(\Ga)$ is perfect. Lemma \[lem:conical\] then yields that all limit chambers are intrinsically conical. Coarse extrinsic geometry of regular hyperbolic subgroups with nice asymptotics {#sec:reghypnas} ------------------------------------------------------------------------------- Let $\Ga\subset G$ be a non-elementary regular discrete subgroup which, as an abstract group, is word hyperbolic. We assume in addition that $\Ga$ has [*nice asymptotics*]{} in the sense that there is a $\Ga$-equivariant embedding $$\beta:\geo\Ga\to\D_FX$$ such that the chambers in its image are pairwise opposite, compare Definition \[dfn:nicas\]. Note that “nice asymptotics” is a weaker condition than “aymptotically embedded”. In particular, we do not a priori assume that $\Ga$ is antipodal or that $\beta(\geo\Ga)\subseteq\La_{ch}(\Ga)$. We fix a word metric on $\Ga$ and study the coarse geometry of the inclusion embedding $\Ga\to G$, respectively, of the orbit maps $\Ga\to\Ga x\subset X$ for $x\in X$. We define a [*directed maximal flat*]{} in $X$ as a pair $(F,\si_+)$ consisting of a maximal flat and a chamber $\si_+\subset\geo F$, equivalently, as a pair $(\si_-,\si_+)$ of opposite chambers. The space ${\cal F}_{dir}(X)$ of directed maximal flats is thus naturally identified with the open dense $G$-orbit $$(\D_FX\times\D_FX)^{opp}\subset\D_FX\times\D_FX$$ consisting of the pairs of opposite chambers, $${\cal F}_{dir}(X)\cong(\D_FX\times\D_FX)^{opp}.$$ The boundary map $\beta$ induces a map $(\beta,\beta)$ of pairs of boundary points. By assumption, $(\beta,\beta)$ maps pairs of distinct boundary points into $(\D_FX\times\D_FX)^{opp}$ and we obtain a continuous $\Ga$-equivariant embedding $$\label{eq:leafmapmov} (\geo\Ga\times\geo\Ga)-\De\buildrel(\beta,\beta)\over\lra (\D_FX\times\D_FX)^{opp}$$ Here and in what follows, $\Delta$ denotes (also) the diagonal. We obtain from the boundary map some preliminary control on the images of quasigeodesics in $\Ga$ under the orbit maps. \[lem:qgeoflat\] An $(L,A)$-quasigeodesic $q:\Z\to\Ga$ with ideal endpoints $\zeta_{\pm}\in\geo\Ga$ is mapped by the orbit map $\Ga\to\Ga x\subset X$ into a tubular neighborhood of uniform radius $r=r(\Ga,L,A,x)$ of the maximal flat $F(\beta(\zeta_-),\beta(\zeta_+))$. Fix $\ga\in\Ga$ and consider the set of all $(L,A)$-quasigeodesics $q:\Z\to\Ga$ with $q(0)=\ga$. This set of quasigeodesics is compact, and therefore also the set of pairs of their ideal endpoints $(\zeta_-,\zeta_+)$ in $(\geo\Ga\times\geo\Ga)-\De$ and, consequently, its $(\beta,\beta)$-image in $(\D_FX\times\D_FX)^{opp}$, that is, the family of corresponding (directed) maximal flats $F(\beta(\zeta_-),\beta(\zeta_+))$. Hence, these flats have bounded distance from the orbit point $\ga x$. The distance bound depends, besides on the subgroup $\Ga\subset G$, on the constants $L,A$ and the point $x$, but not on $\ga$, due to equivariance. This means that for any $(L,A)$-quasigeodesics $q:\Z\to\Ga$ and any point $\ga\in\Ga$ on it, the orbit point $\ga x$ has uniformly bounded distance from the corresponding maximal flat. The geometry of the images of quasigeodesics can be further restricted by taking into account that $\Ga$ is regular. The regularity of $\Ga$, cf. section \[sec:reg\] for the definition, implies that $$\label{ineq:asyreg} d(d_{\De}(x,\ga x),\D\De)\geq d$$ for all $\ga\in\Ga$ with $d_{\Ga}(\ga,e)\geq R=R(\Ga,d_{\Ga},x,d)$. \[lem:qgeoconvinf\] There exists a pair of opposite chambers $\beta'(\zeta_{\pm})\subset\geo F(\beta(\zeta_-),\beta(\zeta_+))$ such that $q(m\pm n)x$ is contained in a tubular neighborhood of uniform radius $r'=r'(\Ga,L,A,x)$ of the Weyl sector $V(q(m)x,\beta'(\zeta_{\pm}))$ for $m,n\in\N$. By Lemma \[lem:qgeoflat\], $q(m\pm n)x$ is contained in the $r(\Ga,L,A,x)$-neighborhood of the maximal flat $F(\beta(\zeta_-),\beta(\zeta_+))$ and hence (by the triangle inequality) in the $2r(\Ga,L,A,x)$-neighborhood of the sector $V(q(m)x,\si(m,\pm n))$ for one of the chambers $\si(m,\pm n)\subset\geo F(\beta(\zeta_-),\beta(\zeta_+))$. It follows from the regularity of $\Ga$ that $\si(m,\pm n)$ stabilizes as $n\to+\infty$, independently of $m$ and the quasigeodesic $q$ asymptotic to $\zeta_{\pm}$, i.e.  $\si(m,\pm n)=\beta'(\zeta_{\pm})$ for $n\geq n(\Ga,L,A,x)$. The chambers $\beta'(\zeta_{\pm})$ are opposite to each other, because for the nearest point projections $\bar q_{\pm n}$ of $q(\pm n)x$ to $F(\beta(\zeta_-),\beta(\zeta_+))$ holds $\bar q_{\pm n}\in V(\bar q_{\mp n},\beta'(\zeta_{\pm}))$ for large $n$. The control obtained so far suffices to show that the orbit map sends quasigeodesics to quasigeodesics and therefore is a quasi-isometric embedding. \[lem:qgeotoqgeo\] $qx$ is a quasigeodesic with quasi-isometry constants depending on $\Ga,L,A,x$. Since $\Ga$ is a discrete subgroup of $G$, the distance between orbit points can be bounded from below in terms of the word metric, i.e. there is an estimate of the form $$d(\ga x,\ga' x)\geq f_x(d_{\Ga}(\ga,\ga'))$$ with $f_x(t)\to+\infty$ as $t\to+\infty$. As a consequence, for $\rho>0$ we have $d(q(m)x,q(m+n)x)\geq\rho$ for $n\geq n(f_x,L,A,\rho)$. Consider now the nearest point projection $\bar q$ of $qx$ to $F(\beta(\zeta_{-}), \beta(\zeta_{+}))$. Choosing $\rho>> r(\Ga,L,A,x),r'(\Ga,L,A,x)$ and invoking again the regularity of $\Ga$, cf. (\[ineq:asyreg\]), we obtain as in the end of the proof of Lemma \[lem:qgeoconvinf\] that $$\label{eq:projinwhu} \bar q_{m\pm n}\in V(\bar q_m,\beta'(\zeta_{\pm}))$$ for $n\geq n'(\Ga,L,A,x)$. It follows that along the coarsening $\bar q|_{n'\Z}$ of $\bar q$, the $\De$-distances between its points are additive in the sense that $$\label{eq:distadd} d_{\De}(\bar q_{m_1},\bar q_{m_2})+d_{\De}(\bar q_{m_2},\bar q_{m_3}) =d_{\De}(\bar q_{m_1},\bar q_{m_3})$$ if $m_3-m_2,m_2-m_1\geq n'$. In particular, $\bar q$ and hence $qx$ is a uniform quasigeodesic. \[thm:qiemb\] Let $\Ga\subset G$ be a non-elementary regular discrete subgroup which is word hyperbolic and has nice asymptotics. Then the orbit maps $\Ga\to\Ga x\subset X$ are quasi-isometric embeddings, with quasi-isometry constants depending on $x$. Since $\Ga$ is word hyperbolic, through any two elements $\ga,\ga'\in\Ga$ exists a quasigeodesic $q$ with uniform quasi-isometry constants $L,A$. The assertion then follows from Lemma \[lem:qgeotoqgeo\]. Our discussion of quasigeodesics yields the following additional information. \[prop:unifreg\] $\Ga$ is uniformly regular. We continue the argument in the proof of Lemma \[lem:qgeotoqgeo\] and consider the coarsened quasigeodesic $qx|_{n'\Z}$. The $\De$-distances between pairs of successive points of it take finitely many (large) values. The types of the segments between successive points of $\bar q|_{n'\Z}$ approximate the types of the segments between successive points of $qx|_{n'\Z}$, and hence are contained in a compact subset of $\interior(\si_{mod})$ which does not depend on the $(L,A)$-quasigeodesic $q$. The additivity of $\De$-distances (\[eq:distadd\]) yields that the curves $\bar q|_{n'\Z}$ are uniformly regular, and therefore also the coarsened quasigeodesics $qx|_{n'\Z}$, and the subgroup $\Ga$. As a consequence of our discussion of the coarse geometry of the embedding of $\Ga$ into $G$, we can sharpen the information on its asymptotics. The boundary map $$\beta':\geo\Ga\to\D_FX$$ provided by Lemma \[lem:qgeoconvinf\] is also a $\Ga$-equivariant embedding with pairwise opposite image chambers. It has the further property that it continuously extends the orbit map $\Ga\to\Ga x\subset X$ at infinity in the following sense: If $r:\N_0\to\Ga$ is a quasiray asymptotic to $\zeta\in\geo\Ga$, then its image quasiray $rx:\N_0\to X$ must enter within uniform time a uniform tubular neighborhood of the Weyl sector $V(r(0)x,\beta'(\zeta))$ and cannot leave it any more. (In other words, the quasiray $rx$ converges to the chamber $\beta'(\zeta)$ in the chamber radial topology.) We can now show that the image of the boundary map $\beta'$ is the chamber limit set. \[lem:homeochlim\] $\La_{ch}(\Ga)=\beta'(\geo\Ga)$. In particular, $\Ga$ is antipodal. It follows from Lemmata \[lem:qgeoconvinf\] and \[lem:chconvwd\] that $\beta'(\geo\Ga)\subseteq\La_{ch}(\Ga)$. In order to show that $\La_{ch}(\Ga)$ cannot be strictly larger, let $\si\in\La_{ch}(\Ga)$ and let $\ga_n\to\infty$ be a sequence in $\Ga$ chamber converging to $\si$, $\ga_n\to\si$. There exists a sequence of uniform quasigeodesics $q_n:\Z\to\Ga$ such that $q_n(0)=e$ and $\ga_n$ lies on the quasiray ${q_n}|_{\N}$. Let $\zeta_n\in\geo\Ga$ denote their forward ideal endpoints. By Lemma \[lem:qgeoconvinf\], the distance from $\ga_nx$ to the Weyl sector $V(x,\beta'(\zeta_n))$ is uniformly bounded. Lemma \[lem:chconvwd\] yields that $\beta'(\zeta_n)\to\si$. Thus $\si\in\beta'(\geo\Ga)$. \[rem:bdmp\] (i) Since both embeddings $\beta$ and $\beta'$ are continuous and $\Ga$-equivariant, the relative position $\pos(\beta',\beta):\geo\Ga\to W$ is continuous (locally constant) and $\Ga$-periodic. It must be constant because the action of a word hyperbolic group on its Gromov boundary is minimal. \(ii) One can show that if $\Ga$ is Zariski dense in $G$ then $\beta=\beta'$. \(iii) On the other hand, in general, $\beta$ and $\beta'$ can be different as the following example shows. \[ex:alnebe\] Consider the embedding of Weyl groups $W_{A_1\circ A_1}\subset W_{B_2}$ and the corresponding refinement of Coxeter complexes where an $A_1\circ A_1$-Weyl arc of length $\pihalf$ breaks up into two $B_2$-Weyl arcs of length $\piforth$. Let $Y\embed X$ be an isometric embedding of symmetric spaces inducing this embedding of Weyl groups, and $H\subset G$ a corresponding embedding of semisimple Lie groups, for instance, $SO(2,1)\times SO(2,1)\subset SO(4,2)$. The symmetric space $Y$ is reducible and decomposes as a product of rank one spaces, $Y\cong Y_1\times Y_2$. Accordingly, its Tits boundary decomposes as a spherical join, $\tits Y\cong\tits Y_1\circ\tits Y_2$. The Weyl chambers of $Y$ are arcs $\xi_1\circ\xi_2$ of length $\pihalf$ with endpoints $\xi_i\in\geo Y_i$. The embedding $\geo Y\embed\geo X$ of visual boundaries sends the Weyl chamber $\xi_1\circ\xi_2$ to an arc denoted by $\xi_1\xi_2$. With respect to the spherical building structure on $\geo X$ it decomposes as the union of two Weyl chambers $\xi_i\mu$ of length $\piforth$, where $\mu$ is the midpoint of $\xi_1\xi_2$. We see that there are [*two*]{} $H$-equivariant continuous embeddings of Fürstenberg boundaries $\iota_i:\D_FY\embed\D_FX$ obtained by assigning to each Weyl chamber $\xi_1\circ\xi_2$ of $Y$ its half $\xi_i\mu$ of type $i$. The embeddings send opposite chambers to opposite chambers. It is easy to construct regular Schottky subgroups $\Ga\subset H$ which remain regular in $G$, and by composing the embeddings $\geo\Ga\to\D_FY$ with $\iota_i$, one obtains two $\Ga$-equivariant embeddings $\geo\Ga\to\D_FX$ mapping distinct boundary points to opposite chambers. Lemma \[lem:homeochlim\] implies that, although we imposed weaker assumptions in the beginning of this section, $\Ga$ nevertheless satisfies the hypotheses of Theorem \[thm:regopphyp\] (ii). We therefore obtain the following strengthening of Theorem \[thm:regopphyp\]: \[thm:regopphyp2\] Let $\Ga\subset G$ be a non-elementary regular discrete subgroup. Then the following are equivalent: \(i) $\Ga$ is RCA. \(ii) $\Ga$ is word hyperbolic and has nice asymptotics. They imply that $\La_{ch}(\Ga)$ is $\Ga$-equivariantly homeomorphic to $\geo\Ga$. \[rem:anos\] Let $\Ga\subset G$ be a non-elementary regular discrete subgroup and suppose that it is [*Anosov*]{} in the sense of [@Labourie], see also [@GuichardWienhard_dd]. It is part of the definition of Anosov that $\Ga$ is word hyperbolic as an abstract group and has nice asymptotics. Theorem \[thm:regopphyp2\] thus implies that regular Anosov subgroups are RCA. One can also show the converse. \[rem:schott\] Benoist constructs in [@Benoist] Zariski dense free Schottky subgroups with prescribed set $\theta(\La(\Ga))\subset\si_{mod}$ of limit point types. In particular, such subgroups can be made uniformly regular. By a ping-pong argument based on the contraction behavior of regular elements on the Fürstenberg boundary $\D_FX$ (and analogous to the one in [@Tits]) one verifies that there are subgroups $\Ga$ of this kind with an equivariant embedding $\geo\Ga\to\D_FX$. Zariski density implies antipodality (this is related to Remark \[rem:bdmp\] (ii)) and, therefore, nice asymptotics for $\Ga$. Thus, one obtains RCA Schottky subgroups. \[Infinitely presented undistorted subgroups\] \[ex:stallings\] Consider the group $F_2\times F_2$ where $F_2$ is the free group of rank $2$. Let $\phi: F_2\to \Z$ be the homomorphism which sends both free generators of $F_2$ to the generator of $\Z$. Let $\Ga\triangleleft F_2\times F_2$ denote the normal subgroup $$\Ga=\{(h_1, h_2): \phi(h_1)=\phi(h_2)\}.$$ Then $\Ga$ is finitely-generated but $\Ga$ is not finitely-presentable (see [@Baumslag-Roseblade]). We claim that the subgroup $\Ga$ is undistorted in $F_2\times F_2$. Indeed, let $w$ be a path in the Cayley graph of $F_2\times F_2$ connecting the unit element $e$ to an element $\ga\in \Ga$. We will equip $\Z$ with the presentation which is the quotient of the presentation of $F_2\times F_2$. Then $\phi(w)$ is a loop in $\Z$, which, therefore, bounds a van Kampen diagram $D$ in the Cayley complex of $\Z$. Since $\Z$ has linear Dehn function, the diagram $D$ has combinatorial area at most $C\ell(w)$, where $\ell(w)$ is the length of $w$ (which is the same as the length of $\phi(w)$). Lifting $D$ to the Cayley complex of $F_2\times F_2$ results in a van Kampen diagram $\tilde{D}$ bounding a bigon one of whose sides is $w$ and the other is a path $u$ in the Cayley graph of $\Ga$. Since $\tilde{D}$ has the same combinatorial area as $D$, we conclude that $$\ell(u)\le (C+1)\ell(w).$$ Thus, $\Ga$ is indeed undistorted in $F_2\times F_2$. Realizing $F_2$ as a convex cocompact subgroup of $Isom(\H^2)$, we obtain a discrete quasi-isometric embedding $F_2\times F_2\to Isom(X), X=\H^2\times \H^2$. Then the subgroup $\Ga\subset\Isom(X)$ is undistorted and not finitely-presentable. On the other hand, since $\Ga$ is not finitely-presented, there is no coarse Lipschitz retraction $X\to\Ga x$. Note that the group $\Ga$ in this example is not regular. Comparing asymptotic properties related to geometric finiteness {#sec:compgfpr} --------------------------------------------------------------- We now prove the equivalence of the properties stated in Theorem \[thm:equivalence\] in the introduction. Properties (1), (3) and (4) are equivalent by Theorems \[thm:regopphyp\] and \[thm:regopphyp2\]. The implication (1)$\Rightarrow$(2) is proven in Proposition \[prop:conicimplexpa\]. We are left with proving the converse implication (2)$\Rightarrow$(1). Since $\Ga$ is antipodal, the action $\Ga\acts\La_{ch}(\Ga)$ is a convergence action by Lemma \[lem:unifattr\]. By assumption, it is expanding at $\La_{ch}(\Ga)$. The latter is perfect due to Lemma \[lem:perfect\]. Now Lemma \[lem:conical\] applies and yields that all limit chambers are intrinsically conical, and hence conical by Lemma \[lem:concon\]. Thus $\Ga$ is RCA. We conclude with several questions; the main theme of these questions can be described as: [*To which extent does Theorem \[thm:rk1\] still hold in the context of higher rank symmetric spaces?*]{} 1\. Suppose that $\Ga\acts X$ is an RCA action. Is it true that there exists a ($\Ga$-equivariant) coarse Lipschitz retraction from $X$ to a $\Ga$-orbit? 2\. Suppose that $\Ga\acts X$ is a regular action satisfying the chamber-conicality condition. Does it follow that this action also satisfies the antipodality condition, i.e. is RCA? 3\. Suppose that $\Ga\acts X$ is an Anosov action so that every infinite order element of $\Ga$ is regular. Does it follow that the action $\Ga\acts X$ is regular? 4\. Suppose that $\Ga\acts X$ is an RCA action. Is the quotient orbifold $X/\Ga$ [*topologically tame*]{}, i.e.  homeomorphic to the interior of a compact orbifold with boundary? 5\. Suppose that $\Ga\acts X$ is an isometric action of a hyperbolic group, which admits an equivariant injective continuous map $$\beta: \geo \Ga \to \D_F X$$ sending distinct points to antipodal chambers. Assume, in addition, that every nontrivial element of $\Ga$ is regular. Is it true that $\Ga$ is a regular subgroup of $Isom(X)$ (i.e. that $\Ga$ has nice asymptotics)? [BLP05]{} G. Baumslag, J. E. Roseblade, [*Subgroups of direct products of free groups*]{}, J. London Math. Soc. (2) Vol. [**30**]{} (1984), p. 44–52. Y. Benoist, [*Propriétés asymptotiques des groupes linéaires*]{}, Geom. Funct. Anal. Vol. [**7**]{} (1997), no. 1, p. 1–47. L. Bers, [*On boundaries of Teichmüller spaces and on Kleinian groups I*]{}, Ann. Math. Vol. [**91**]{} no. 3 (1970), p. 570–600. A. Björner, F. Brenti, “Combinatorics of Coxeter groups,” Springer, GTM 231, 2005. N. Bourbaki, “Lie Groups and Lie Algebras,” ch. 4-6, tables I-IX. B. Bowditch, [*Geometrical finiteness with variable negative curvature*]{}, Duke Math. J. Vol. [**77**]{} (1995) p. 229–274. B. Bowditch, [*A topological characterisation of hyperbolic groups*]{}, J. Amer. Math. Soc. Vol. [**11**]{} (1998), no. 3, p. 643–667. B. Bowditch, [*Convergence groups and configuration spaces*]{}, in “Geometric group theory down under” (Canberra, 1996), p. 23–54, de Gruyter, Berlin, 1999. M. Coornaert, A. Papadopoulos, “Symbolic dynamics and hyperbolic groups,” Lecture Notes in Mathematics, Vol. 1539, Berlin, 1993. L. Ford, “Automorphic Functions” Chelsea, New York, 1972. M. Gromov, [*Hyperbolic groups*]{}. In: “Essays in group theory,” p. 75–263, Math. Sci. Res. Inst. Publ., Vol. 8, Springer, New York, 1987. O. Guichard, A. Wienhard, [*Anosov representations: Domains of discontinuity and applications*]{}, Invent. Math. Vol. [**190**]{} (2012) no. 2, p. 357–438. J.E. Humphreys, “Reflection groups and Coxeter groups,” Cambridge studies in advanced mathematics, Vol. 29, 1990. M. Kapovich, [*Kleinian groups in higher dimensions*]{}, In “Geometry and dynamics of groups and spaces. In memory of Alexander Reznikov”, M.Kapranov et al (eds). Birkhäuser, Progress in Mathematics, Vol. 265, 2007, p. 485–562. M. Kapovich, [*Non-coherence of arithmetic hyperbolic lattices*]{}, Geometry and Topology, Vol. [**17**]{} (2013) p. 39–71. M. Kapovich, B. Leeb, J.J. Millson, [*Convex functions on symmetric spaces, side lengths of polygons and the stability inequalities for weighted configurations at infinity*]{}, Journal of Differential Geometry, Vol. [**81**]{} (2009), p. 297–354. B. Kleiner, B. Leeb, [*Rigidity of quasi-isometries for symmetric spaces and Euclidean buildings*]{}, Inst. Hautes Études Sci. Publ. Math. No. [**86**]{} (1997) p. 115–197. B. Kleiner, B. Leeb, [*Rigidity of invariant convex sets in symmetric spaces*]{}, Invent. Math. Vol. [**163**]{}, No. 3, (2006) p. 657–676. F. Labourie, [*Anosov flows, surface groups and curves in projective space*]{}, Invent. Math. Vol. [**165**]{}, No. 1, (2006) p. 51–114. B. Leeb, [*A characterization of irreducible symmetric spaces and Euclidean buildings of higher rank by their asymptotic geometry*]{}, Bonner Mathematische Schriften 326 (2000), see also arXiv:0903.0584 (2009). J. Millson, [*On the first Betti number of a constant negatively curved manifold*]{}, Ann. of Math. (2) Vol. [**104**]{} (1976), no. 2, p. 235–247. M. Mj, [*Cannon-Thurston Maps for Kleinian Groups*]{}, Preprint, arXiv:1002.0996, 2010. D. Mumford, J. Fogarty, F. Kirwan, “Geometric Invariant Theory,” Springer Verlag, 1994. M. Ramachandran, in preparation. D. Sullivan, [*Quasiconformal homeomorphisms and dynamics. II. Structural stability implies hyperbolicity for Kleinian groups*]{}, Acta Math. Vol. [**155**]{} (1985), no. 3-4, p. 243–260. J. Tits, [*Free subgroups in linear groups*]{}, J. Algebra, Vol. [**20**]{} (1972) p. 250–270. P. Tukia, [*Conical limit points and uniform convergence groups*]{}, J. Reine Angew. Math. Vol. [**501**]{} (1998), p. 71–98. Addresses: M.K.: Department of Mathematics,\ University of California, Davis\ CA 95616, USA\ email: kapovich@math.ucdavis.edu B.L.: Mathematisches Institut, Universität München\ Theresienstr. 39\ D-80333 München, Germany\ email: b.l@lmu.de J.P.: Departament de Matemàtiques,\ Universitat Autònoma de Barcelona,\ E-08193 Bellaterra, Spain\ email: porti@mat.uab.cat [^1]: Choosing such point amounts to choosing a positive line bundle (up to a tensor power) in the setting of Geometric Invariant Theory, see Remark \[GIT\].

--- abstract: 'We present rapidly rising transients discovered by a high-cadence transient survey with Subaru telescope and Hyper Suprime-Cam. We discovered five transients at $z=0.384-0.821$ showing the rising rate faster than 1 mag per 1 day in the restframe near-ultraviolet wavelengths. The fast rising rate and brightness are the most similar to SN 2010aq and PS1-13arp, for which the ultraviolet emission within a few days after the shock breakout was detected. The lower limit of the event rate of rapidly rising transients is $\sim 9 \%$ of core-collapse supernova rates, assuming a duration of rapid rise to be 1 day. We show that the light curves of the three faint objects agree with the cooling envelope emission from the explosion of red supergiants. The other two luminous objects are, however, brighter and faster than the cooling envelope emission. We interpret these two objects to be the shock breakout from dense wind with the mass loss rate of $\sim 10^{-3}\ {M_{\odot}}$ yr$^{-1}$, as also proposed for PS1-13arp. This mass loss rate is higher than that typically observed for red supergiants. The event rate of these luminous objects is $\gsim 1 \%$ of core-collapse supernova rate, and thus, our study implies that more than $\sim 1 \%$ of massive stars can experience an intensive mass loss at a few years before the explosion.' author: - | Masaomi Tanaka, Nozomu Tominaga, Tomoki Morokuma, Naoki Yasuda, Hisanori Furusawa,\ Petr V. Baklanov, Sergei I. Blinnikov, Takashi J. Moriya, Mamoru Doi, Ji-an Jiang,\ Takahiro Kato, Yuki Kikuchi, Hanindyo Kuncarayakti, Tohru Nagao,\ Ken’ichi Nomoto, and Yuki Taniguchi title: 'Rapidly Rising Transients from Subaru Hyper Suprime-Cam Transient Survey[^1]' --- Introduction {#sec:introduction} ============ The transient sky has been intensively explored by various surveys in the last decade. Especially, optical surveys using wide-field cameras, such as Palomar Transient Factory [PTF, @law09; @rau09], Catalina Real-Time Transient Survey [CRTS, @drake09], and Pan-STARRS1 [PS1, [e.g., ]{} @kaiser10], have significantly contributed to building our knowledge on the transient phenomena in the Universe. One of the important discovery spaces for transient surveys is phenomena with a short timescale, [i.e., ]{}$\lsim 1$ day. There are, in fact, several theoretical expectations for such short-timescale transients. For supernovae (SNe), shock breakout emission should have timescale of $\sim 1$ hr for the case of red supergiant progenitors [[e.g., ]{} @falk78; @klein78sbo; @matzner99]. The subsequent cooling emission lasts for a few days [[e.g., ]{} @waxman07; @chevalier08; @nakar10]. For the case of blue supergiants or Wolf-Rayet stars, these timescale are even shorter. Other possible short-timescale transients include, for example, the disk outflow from black hole forming SNe [$<$ a few days, @kashiyama15] and accretion induced collapse of white dwarfs [$\sim$ 1 day, @metzger09]. In addition to these, there might also be unknown kind of transients with a short duration since our knowledge on the short-timescale transients is still limited. To explore the short-timescale transient sky, some dedicated high-cadence surveys have started. For example, Kiso Supernova Survey (KISS, @morokuma14 [@tanaka14agn], using 1.05m Schmidt telescope and $\sim 4$ deg$^{2}$ wide field camera, @sako12) and High-cadence Transient Survey (HiTS, @forster14, using 4m Blanco telescope and $\sim 3$ deg$^2$ Dark Energy Camera, @flaugher15) adopt $\sim$ 1 hr cadence aiming at the detection of SN shock breakout. There are also some ambitious surveys to explore even shorter timescales [[e.g., ]{} @becker04; @rau08; @berger13fot], although no extragalactic transients with $\lsim$ 30 min timescale have been detected. Recently, we have started a high-cadence transient survey with the 8.2m Subaru telescope and 1.77 deg$^2$ Hyper Suprime-Cam [HSC, @miyazaki06; @miyazaki12], as a part of Subaru HSC Survey Optimized for Optical Transients (SHOOT). SHOOT also adopts $\sim$ 1 hr cadence focusing on the detection of SN shock breakout [@tominaga15]. In this paper, we present rapidly rising transients discovered in SHOOT. Here we define rapidly rising transients as objects that rise more than 1 mag within restframe 1 day, [i.e., ]{}the rising rate [$|\Delta m/ \Delta t|$]{} &gt; 1 mag day$^{-1}$. We describe our observations and sample selection in Section \[sec:observations\]. Then, we compare the obtained light curves with previously known SNe and transients in Section \[sec:LC\]. Rising rates of various types of transients are summarized in Section \[sec:dmdt\]. Based on these comparison, we discuss the nature of these transients in Section \[sec:discussion\]. Finally we give conclusions in Section \[sec:conclusions\]. Throughout the paper, we assume the following cosmological parameters: $\Omega_M = 0.273$, $\Omega_{\Lambda}= 0.726$, and $H_0 = 70.5$ km s$^{-1}$ Mpc$^{-1}$ [@komatsu09]. All the magnitudes are given in AB magnitude. [lllll]{} 2014-07-02 & Day 1 & HSC & imaging ($g$,$r$) & 0.5\ 2014-07-03 & Day 2 & HSC & imaging ($g$,$r$) & 0.6\ 2014-08-05 & Day 35 & FOCAS & imaging ($g$,$r$) & 0.9\ & & & spectroscopy &\ 2014-08-06 & Day 36 & FOCAS & imaging ($g$,$r$) & 0.9\ & & & spectroscopy &\ 2015-05-24 & Day 327 & HSC$^{b}$ & imaging ($g$,$r$) & 1.0\ 2015-06-22 & Day 356 & FOCAS & spectroscopy & 0.5\ 2015-08-19 & Day 414 & HSC$^{b}$ & imaging ($r$) & 1.4\ \[tab:log\] Observations and sample selection {#sec:observations} ================================= HSC observations {#sec:HSC} ---------------- We performed a high-cadence transient survey with Subaru/HSC for two continuous nights, 2014 July 2 and 3 UT (hereafter Day 1 and 2, respectively). The log of our observations is given in Table \[tab:log\]. Seven field-of-views ($\simeq 12\ {\rm deg^2}$) were repeatedly visited with about 1 hr cadence. Our survey was carried out mostly in optical $g$-band, targeting the detection of the very early phase of SNe [@tominaga15]. Within one night, we had 3 or 4 visits in $g$-band (here one “visit” consists of five 2-min exposures). We also took 1 visit data in $r$-band in each night to obtain $g-r$ color. The HSC data were reduced using the HSC pipeline (version 3.6.1) developed based on the LSST pipeline [@ivezic08; @axelrod10]. After standard reduction for each frame, 5 exposure images were co-added. For astrometry and photometric calibration, we used the Sloan Digital Sky Survey DR8 catalog [@aihara11]. For stacked images for 1 visit ([i.e., ]{}10 min exposure), a typical limiting magnitude is about 26 mag (5 sigma limiting magnitude for point sources) in both $g$- and $r$-bands. We performed image subtraction using the HSC pipeline. The pipeline adopts the algorithm developed by @alard98 and @alard00, which are used for the ISIS package[^2] and the HOTPANTS package[^3]. The algorithm uses a space-varying convolution kernel to match the PSFs of two images. The optimal convolution kernel is derived by minimizing the difference between convoluted PSFs of two images. Although our 7 survey fields are selected based on the availability of the past imaging data, most of the survey fields lack imaging data that are deep and wide enough to be used as references for our new HSC images. Thus, we used the data taken at the first visit of Day 1 as reference images for sample selection. The data reduction described above was performed in realtime using the on-site data analysis system [@furusawa11] and a dedicated transient system [@tominaga15]. By using these systems, transient candidates were typically selected within the same night [@tominaga14atel1; @tominaga14atel2; @tominaga15atel1; @tominaga15atel2]. To obtain the final reference images, we also performed HSC imaging observations on 2015 May 24 UT (Day 327, for $g$- and $r$-band) and 2015 Aug 19 UT (Day 414, for $r$-band). All the photometric values given in this paper are derived by aperture photometry with 7 pixel radius (1.18 arcsec) in the difference images using these final reference images. [cccc]{}\ \ &\ &\ & &\ & &\ & & &\ & & &\ \[tab:classification\] -------------------- -------------------- -------------------- -------------------- --------------------      -------------------- -------------------- -------------------- -------------------- -------------------- [lcccc]{} SHOOT14gp & 23:20:20.80 & +28:25:00.54 & 0.635 & $> 3.10$\ SHOOT14or & 15:26:24.18 & +47:47:07.34 & 0.821 & $3.12^{+1.11}_{-0.70}$\ SHOOT14ha & 23:21:44.91 & +28:54:49.80 & 0.548 & $>1.19$\ SHOOT14jr & 16:33:49.99 & +34:28:05.36 & 0.384 & $1.61^{+0.39}_{-0.32}$\ SHOOT14ef & 21:31:08.77 & +09:32:54.10 & 0.560 & $>1.31$\ \[tab:objects\] [cccc]{}\ 56840.542 & $g$ & $>25.53$ & HSC\ 56840.577 & $g$ & $>25.57$ & HSC\ 56841.513 & $g$ & $23.74^{+0.09}_{-0.08}$ & HSC\ 56841.547 & $g$ & $23.72^{+0.08}_{-0.07}$ & HSC\ 56841.582 & $g$ & $23.70^{+0.08}_{-0.07}$ & HSC\ 56840.560 & $g$ & $>25.58$ & HSC$^{b}$\ 56841.548 & $g$ & $23.71^{+0.07}_{-0.06}$ & HSC$^{b}$\ 56874.475 & $g$ & $>25.45$ & FOCAS\ 56840.479 & $r$ & $>24.99$ & HSC\ 56841.456 & $r$ & $24.31^{+0.15}_{-0.13}$ & HSC\ 56874.463 & $r$ & $25.51^{+0.63}_{-0.39}$ & FOCAS\ \ 56840.287 & $g$ & $26.74^{+0.65}_{-0.40}$ & HSC\ 56840.332 & $g$ & $26.88^{+0.75}_{-0.44}$ & HSC\ 56841.283 & $g$ & $25.11^{+0.13}_{-0.12}$ & HSC\ 56841.326 & $g$ & $25.01^{+0.12}_{-0.10}$ & HSC\ 56841.487 & $g$ & $24.99^{+0.11}_{-0.10}$ & HSC\ 56840.310 & $g$ & $26.85^{+0.64}_{-0.40}$ & HSC$^{b}$\ 56841.365 & $g$ & $25.04^{+0.08}_{-0.07}$ & HSC$^{b}$\ 56873.315 & $g$ & $>25.69$ & FOCAS\ 56840.431 & $r$ & $>25.61$ & HSC\ 56841.412 & $r$ & $25.25^{+0.25}_{-0.20}$ & HSC\ 56873.276 & $r$ & $25.79^{+0.59}_{-0.38}$ & FOCAS\ \ 56840.542 & $g$ & $>25.77$ & HSC\ 56840.577 & $g$ & $>25.79$ & HSC\ 56841.513 & $g$ & $25.29^{+0.33}_{-0.25}$ & HSC\ 56841.547 & $g$ & $25.27^{+0.28}_{-0.22}$ & HSC\ 56841.582 & $g$ & $24.95^{+0.19}_{-0.16}$ & HSC\ 56840.560 & $g$ & $>25.87$ & HSC$^{b}$\ 56841.548 & $g$ & $25.11^{+0.20}_{-0.17}$ & HSC$^{b}$\ 56874.601 & $g$ & $>25.42$ & FOCAS\ 56840.479 & $r$ & $>25.48$ & HSC\ 56840.479 & $r$ & $>25.48$ & HSC\ 56841.456 & $r$ & $25.26^{+0.32}_{-0.25}$ & HSC\ 56873.501 & $r$ & $>25.03$ & FOCAS\ 56874.589 & $r$ & $>25.07$ & FOCAS\ \ 56840.299 & $g$ & $25.85^{+0.33}_{-0.25}$ & HSC\ 56840.342 & $g$ & $25.96^{+0.37}_{-0.27}$ & HSC\ 56840.526 & $g$ & $25.50^{+0.26}_{-0.21}$ & HSC\ 56841.293 & $g$ & $24.65^{+0.10}_{-0.09}$ & HSC\ 56841.338 & $g$ & $24.77^{+0.12}_{-0.11}$ & HSC\ 56841.500 & $g$ & $24.45^{+0.08}_{-0.08}$ & HSC\ 56840.389 & $g$ & $25.76^{+0.27}_{-0.21}$ & HSC$^{b}$\ 56841.377 & $g$ & $24.61^{+0.09}_{-0.08}$ & HSC$^{b}$\ 56840.442 & $r$ & $25.84^{+0.72}_{-0.43}$ & HSC\ 56841.422 & $r$ & $24.87^{+0.23}_{-0.19}$ & HSC\ 56873.262 & $r$ & $>25.36$ & FOCAS\ \ 56840.554 & $g$ & $>26.30$ & HSC\ 56840.591 & $g$ & $>26.41$ & HSC\ 56840.610 & $g$ & $>26.19$ & HSC\ 56841.525 & $g$ & $25.57^{+0.22}_{-0.18}$ & HSC\ 56841.559 & $g$ & $25.72^{+0.23}_{-0.19}$ & HSC\ 56841.596 & $g$ & $25.70^{+0.27}_{-0.21}$ & HSC\ 56841.615 & $g$ & $25.74^{+0.30}_{-0.23}$ & HSC\ 56840.585 & $g$ & $>26.50$ & HSC$^{b}$\ 56841.574 & $g$ & $25.67^{+0.20}_{-0.17}$ & HSC$^{b}$\ 56840.467 & $r$ & $>26.08$ & HSC\ 56841.445 & $r$ & $>26.06$ & HSC\ \[tab:data\] Sample selection {#sec:selection} ---------------- We adopted the following selection processes to select candidates for rapidly rising transients. As mentioned above, we used the first images taken on Day 1 as reference images for the selection process. Therefore, source detection in the subtracted images is sensitive only to objects showing variability within 2 nights. Detected sources in the subtracted images contain not only real astronomical sources but also fake sources such as spikes around bright stars, and artifacts due to mis-subtraction or mis-alignment [[e.g., ]{} @bailey07; @bloom12; @brink13]. Thus, we selected objects detected in the subtracted images at least twice with $>5\sigma$ significance. After this selection, 1407 sources remain. We first performed initial visual screening, resulting in 430 sources with SHOOT14XX names (412 independent sources because of 18 duplication in overlapped regions in the reduced images). Then, we further performed detailed classification. Results of the classifications are summarized in Table \[tab:classification\]. Among 412 independent sources, 215 sources are still fakes of the subtracted images while the other 197 sources are likely to be astronomical sources. The astronomical sources are dominated by stellar-shape sources, such as stars or quasars (166 sources). The remaining 31 sources are associated with extended sources (galaxies). Among these sources, 16 sources are located at the center of galaxies. Since they may be active galactic nuclei or tidal disruption events, we avoided these objects for follow-up observations. Since 8 out of 16 objects show declining flux, it is likely that the majority of these 16 sources are active galactic nuclei. Remaining 15 sources have an offset from the center of the galaxies, and selected as SN candidates. The final SN candidates consist of 14 brightening objects. From this final sample, we performed follow-up observations of most reliable 12 objects. Among these 12 objects, we measured redshifts for 8 objects while the other 4 objects (and their host galaxies) were too faint to take spectra. The remaining 2 objects were not observed. Note that the sample selection for spectroscopy was made based on the [*flux*]{} difference within 2 nights, not on the [*magnitude*]{} difference since the final reference images were not available and true magnitudes of the objects on Day 1 were not known at the time of spectroscopy (2014 Aug). Therefore, even after the selection processes, our initial samples could include not only rapidly rising transients but also normal SNe around the peak brightness if the flux difference within 2 nights is large enough. In fact, by our follow-up spectroscopic observations (Section \[sec:followup\]), 3 out of 8 objects were identified as normal SNe (at $z$=0.13, 0.25, and 0.40). In addition, after obtaining the final reference images on Day 327, we confirmed that these three objects are already bright on Day 1. The rising rates for these three objects are [$|\Delta m/ \Delta t|$]{}$< 1$ mag day$^{-1}$, which is also consistent with normal SNe. Therefore, we omit these three objects from our samples. Figure \[fig:image\] shows images of 5 rapidly rising transients, named as SHOOT14gp, 14or, 14ha, 14jr, and 14ef (Table \[tab:objects\]). Photometry of these 5 objects is shown in Table \[tab:data\]. ![ Spectra of host galaxies of SHOOT14gp, 14or, 14ha, 14jr, and 14ef (from top to bottom). The wavelengths of strong emission lines (\[\] $\lambda$3727, H$\beta$, and \[\] $\lambda$4959,5007) are marked with the dashed lines. The right panels show the data around these lines. []{data-label="fig:spec"}](f02.eps) Follow-up observations {#sec:followup} ---------------------- We performed imaging and spectroscopic observations of 5 objects (Table \[tab:objects\]) using the Faint Object Camera and Spectrograph [FOCAS, @kashikawa02] of the Subaru telescope. Observations of the four objects (SHOOT14gp, 14or, 14ha, and 14jr) were carried out on 2014 Aug 5 and 6 UT (Day 35 and 36, respectively) while observations of SHOOT14ef were on 2015 June 22 (Day 356, only for the host galaxy). For the FOCAS imaging data, we performed image subtraction with the final reference images using HOTPANTS package. SHOOT14gp and 14or were marginally detected only in $r$-band while they were not detected in $g$-band. The other objects were not detected both in $g$- and $r$-bands. A typical limiting magnitudes are $\simeq 25.0-25.5$ mag (Table \[tab:data\]). For spectroscopy, we used multi-object mode with $0\farcs 8$-width slit and long-slit mode with $1\farcs 0$-width slit (only for SHOOT14ef). With the 300B (300 lines mm$^{-1})$ grism and the SY47 order-sort filter, our configuration gives a wavelength coverage of 4700 - 9000 Åand a spectral resolution of $R = \lambda/\Delta \lambda \sim 600$. The data were reduced with the IRAF packages in a standard manner. The transient components are not detected in our spectra as expected from the results of imaging observations. Figure \[fig:spec\] shows the spectra of the host galaxies for these five objects. The \[\] $\lambda$3727 emission line is detected from all the host galaxies, which indicates that they are all star forming galaxies. The redshifts range from $z =0.384$ (SHOOT14jr) to $z=0.821$ (SHOOT14or). ------------------- -- -- -- --  ------------------- -- -- -- -- Light curves {#sec:LC} ============ Overview {#sec:overview} -------- Figure \[fig:LCabs\] shows light curves of our samples on Day 1 and Day 2. Hereafter, the epochs of stacked $g$-band data on Day 2 are taken to be $t=0$ unless otherwise mentioned. The photometry is performed in the subtracted images using the final references ([e.g., ]{}Day 1 $-$ Day 327 and Day 2 $-$ Day 327 for $g$-band). Throughout the paper, we do not take into account full $K$-correction for absolute magnitudes since only limited information about spectral energy distribution is available for our samples. Instead, we only correct the effect of redshifts, [i.e., ]{}$M = m - \mu + 2.5\log(1+z)$, where $M$ and $m$ are absolute and observed AB magnitudes (measured as $f_{\nu}$), $\mu$ is the distance modulus. The last term originates from the difference in the frequency bin in the restframe and observer frame, [i.e., ]{}$L_{\nu}(\nu_e) = [(4 \pi d^2)/(1+z)] f_{\nu}(\nu_o)$, where $\nu_e$ and $\nu_o$ are restframe and observer frame frequency, and $d$ is the luminosity distance [@hogg02]. The absolute magnitudes of the five objects range from $-16$ to $-19$ mag in the restframe near-ultraviolet (UV) wavelengths (2620Å$-$ 3450Å, depending on the redshifts). The photometric values of our samples are corrected for the extinction in our Galaxy but not for the extinction in the host galaxy. Therefore, intrinsic absolute magnitudes can be brighter than those shown in Figure \[fig:LCabs\]. All of the five objects show blue $g-r$ color on Day 2, $g-r \simeq -0.60, -0.21, -0.15$, and $-0.15$ mag for SHOOT14gp, 14or, 14ha, and 14jr, respectively. For SHOOT14ef, the color is $g-r < -0.39$ mag. This indicates that, for blackbody case, the peak of the spectra is located at wavelengths shorter than the wavelengths corresponding to the observed $r$-band. Therefore the blackbody temperatures for our objects are $T_{\rm BB} \gsim$ 13000, 15000, 13000, 11000, and 13000 K for SHOOT14gp, 14or, 14ha, 14jr, and 14ef, respectively. Note that the intrinsic colors can be bluer due to the extinction in the host galaxies. SHOOT14or and 14jr are detected in the images of Day 1 $-$ Day 327. We measure the rising rates from Day 1 to Day 2 using the $g$-band 1-day stacked images: [$|\Delta m/ \Delta t|$]{}= $3.12^{+1.11}_{-0.70}$ and $1.61^{+0.39}_{-0.32}$ mag day$^{-1}$ for SHOOT14or and 14jr, respectively (errors represent 1$\sigma$, Table \[tab:objects\]). Note that the rising rate is measured in the restframe, so the time interval used for the measurement varies with the source redshifts ($\Delta t = 0.55$ days for SHOOT14or while $\Delta t = 0.72$ days for SHOOT14jr). The other three objects (SHOOT14gp, 14ha, and 14ef) are not detected in the subtracted images of Day 1 $-$ Day 327. The 3 $\sigma$ lower limits of the rising rate measured in $g$-band are [$|\Delta m/ \Delta t|$]{} $>$ 3.10, 1.21, and 1.17 mag day$^{-1}$. These are also high enough to match our criterion for rapidly rising transients. In the following sections, we compare the light curves of our samples with those of previously known SNe and transients. -------------------- --------------------   -------------------- -------------------- ![Comparison of light curves with SLSNe [@pastorello10; @chomiuk11]. Observed $u$-band light curves are shown for SN 2010gx, while observed $g$- and $r$-band light curves are shown for PS1-10awh and PS1-10ky. For SLSNe, the peak epochs are shifted to $t=13$ days and magnitudes are corrected for only Galactic extinction. []{data-label="fig:LCSLSN"}](f05.eps) Comparison with SNe {#sec:SN} ------------------- Figure \[fig:LCall\] shows comparison of rapidly rising transients with normal SNe. Since the redshifts of our samples are moderately high, $z =0.384-0.821$, we compare our $g$- and $r$-band light curves with near-UV and $u$-band light curves of nearby SNe with good temporal coverage. We use the [*Swift*]{} $uvw1$- and $u$-band data from @brown12 and @pritchard14 with extinction correction (both in our Galaxy and host galaxies) using the extinction law by @brown10. Since the effective restframe wavelengths do not always match perfectly, we always give effective restframe wavelengths in parenthesis. Figure \[fig:LCall\] shows that the properties of our samples are not consistent with those of Type Ia SNe at any phase, and those of core-collapse SNe at $\gsim$ a few days after the explosion. The absolute magnitudes of our samples are as luminous as the peak magnitude of Type Ia SN 2011fe [@brown12] and Type IIP SN 2006bp [@pritchard14]. However, the rising rates for our samples are faster than the very early phase of SN 2011fe, one of the best observed Type Ia SNe. We also compare our objects with Type IIb SN 2008ax, Type IIn SN 2011ht, and Type Ib SN 2007Y [@pritchard14]. Their rising rates are slower than those of our samples at any epochs with available data, [i.e., ]{}$\gsim$ a few days after the explosion. In addition, the blue colors of our samples ($g-r \le -0.2$ mag) are not consistent with normal SNe after a few days from the explosion. For nearby SNe after a few days from the explosion, the $uvw1$ magnitude is generally fainter than the $u$ magnitude as shown in Figure \[fig:LCall\], [i.e., ]{}the color is $uvw1-u > 0$ mag. Our samples might correspond to the rising phase of much brighter SNe, such as superluminous SNe [SLSNe, @quimby11; @gal-yam12]. Figure \[fig:LCSLSN\] shows comparison of our samples with SLSN SN 2010gx, PS1-10awh, and PS1-10ky with a good temporal coverage [@pastorello10; @chomiuk11]. Our data on Days 1 and 2 could be interpreted as the very early phase of SLSNe, which have never been caught. However, the data on Days 35 and 36 are clearly inconsistent with the declining part of SLSNe. -------------------- --------------------   -------------------- -------------------- --------------------------------------- --   --------------------------------------- -- Comparison with very early phase of SNe {#sec:earlyCC} --------------------------------------- We compare our samples with earlier phases of SNe ($\lsim$ a few days after the explosion). First, we show comparison with Type IIP SN 2010aq [@gezari10] and PS1-13arp [@gezari15], with UV detection at the very early phase with $GALEX$. The early emission of SN 2010aq is consistent with cooling envelope emission after SN shock breakout [@gezari10]. The emission of PS1-13arp is brighter and shorter, which may indicate shock breakout emission from dense wind [@gezari15]. The upper panel of Figure \[fig:LCearly\] shows a similarity of the rising rate and brightness between our samples and SN 2010aq and PS1-13arp. SN 2010aq and PS1-13arp also show fast rise, [$|\Delta m/ \Delta t|$]{}$ > 0.989$ and $> 2.635$ mag day$^{-1}$, respectively. They reach about $-17$ - $-18$ mag, which is also similar to our samples. Note that the effective restframe wavelengths corresponding to the NUV filter of $GALEX$ (2130 Å and 1990 Å for SN 2010aq and PS1-13arp, respectively) are shorter than those for our samples ($\sim 2600-3500$ Å). For comparison, we also show non-filter magnitude of Type IIP SN 2006bp [@quimby0706bp], for which very early phases were observed (see also @rubin15 for recent larger samples). It also shows a fast rise, [$|\Delta m/ \Delta t|$]{}$= 2.3$ mag day$^{-1}$. Again, although the difference in the restframe wavelengths should be cautioned, these similarities suggest that our samples of rapidly rising transients are the very early phase of SNe. We also compare our samples with the very early part of Type Ic SN 2006aj and Type Ib SN 2008D. They are among the best-studied stripped-envelope SNe. SN 2006aj is associated with low luminosity gamma-ray burst (GRB) 060218, and thus, good optical to NUV data are available from soon after the explosion [[e.g., ]{} @campana06; @soderberg06; @pian06; @mazzali06; @sollerman06; @modjaz06; @mirabal06; @simon10]. SN 2008D is associated with X-ray transient 080109 [[e.g., ]{} @soderberg08; @mazzali08; @tanaka0908D; @tanaka0908Dneb; @modjaz09]. Emission at the first 2 days of SN 2006aj and SN 2008D is interpreted as cooling envelope emission [@waxman07; @soderberg08; @modjaz09; @chevalier08; @nakar15]. The lower panel of Figure \[fig:LCearly\] shows that the rising rate of SN 2006aj is as fast as our samples. The time to the peak is only $\sim 0.5$ days, which is as short as that inferred for our samples although we cannot not firmly determine the peak dates only with 2-night data. SN 2008D lacks the data at $\sim 1$ day after the explosion. Nevertheless, the rising rate of SN 2008D in [*Swift*]{} $u$-band (measured with 2-day interval) is similar to SHOOT14jr. Note that if the early part of SN 2008D is interpreted as cooling envelope emission, the peak would be around $\sim 1$ day after the explosion [@soderberg08; @modjaz09], and the rising rate in the first day is faster than that measured with 2-day interval. When we match our objects with core-collapse SNe within a few days after the explosion, our observations on Day 35 and 36 correspond to the plateau phase of Type IIP or the peak phase of Type Ibc SNe. As shown in Figure \[fig:LCCCSN\], the distribution of $uvw1$ brightness of core-collapse SNe at these epochs ranges from $-12$ to $-17$ mag. Since our limits in $g$-band correspond to $-17.0$ mag, non-detection in $g$-band on Days 35 and 36 is not surprising. SHOOT14gp and 14or are marginally detected in $r$-band (right panel of Figure \[fig:LCCCSN\]). Compared with [*Swift*]{} $u$-band data, their brightness is consistent with those of core-collapse SNe at the luminous end. -------------------- --------------------     -------------------- --------------------  Comparison with rapidly rising transients from PS1 {#sec:rapid} -------------------------------------------------- The rapid rising rates of our samples remind us of population of rapidly evolving and luminous transients from PS1, which are compiled by @drout14 [see also @poznanski10 [@kasliwal10; @drout13]]. These transients show rapid luminosity evolution both in rising and declining phases compared with normal SNe with a time above half-maximum of less than 12 days. Interestingly, they show a faster rising rate than a declining rate, which motivates the comparison with our samples. In addition, they have blue $g-r$ colors ($g-r < -0.2$ mag), similar to our samples. Since the PS1 samples have a wide luminosity range, we divide the samples into two classes with the absolute magnitude brighter (hereafter PS1 luminous samples) or fainter (PS1 faint samples) than $-19.0$ mag. @drout14 interpret their rapid transients to be either (1) the cooling envelope emission following shock breakout (especially for faint samples) or (2) shock breakout from dense wind (for luminous samples). Figure \[fig:LCrapid\] shows comparison of our samples with the PS1 samples [@drout14] which are detected at the rising part in $g$-band. The peak dates of the PS1 samples are taken to be $t=0$ day. It should be cautioned that the PS1 samples have a wider redshift range than ours, and thus the rest wavelengths corresponding the observed filters have a wider variety. For the PS1 luminous samples, $g$- and $r$-band data for our samples are compared with PS1 $g$- and $r$-band data, respectively. Since the PS1 faint samples have low redshifts ($z = 0.074$ and $0.113$ for PS1-10ah and PS1-10bjp, respectively), we compare our $g$- and $r$-band data with PS1 $g$-band data. The peak magnitudes of the PS1 luminous samples are brighter than the magnitudes of our sample on Day 2. Our samples could thus be interpreted to the rising part of the PS1 samples. The dashed lines in the upper left panel of Figure \[fig:LCrapid\] shows the extrapolation of the rising part by assuming the flux rises as $f = (t-t_0)^2$ (as often assumed for the early part of SNe, see [e.g., ]{}@nugent11 [@pastorello13; @prieto13; @yamanaka14]), where $t_0$ is the epoch with zero flux. Three of our samples (SHOOT14ha, 14jr, and 14ef) show a nice agreement with the extrapolated rising part if the epochs of these objects are shifted so that Day 2 corresponds to $t \sim -10$ days. However, with this assumption, the non detection of PS1-13duy before the peak in $r$-band is not consistent with our detection on Day 2. In addition, the brightness and upper limits at later epochs (Days 35 and 36) are much fainter than the magnitudes of PS1-11qr for which the data at the declining part is available. Therefore, our samples are not likely to be the same population as the PS1 luminous samples. Our samples show a better agreement with the PS1 faint samples (lower panels of Figure \[fig:LCrapid\]). The rising rates of the PS1 samples in $g$-band is [$|\Delta m/ \Delta t|$]{} &lt; 1 mag day$^{-1}$, which do not fulfill our criterion. However, PS1 data are taken with $\sim 3$ days cadence, and thus, the rising rate measured with a shorter interval can be faster. In fact, if the rising part is interpolated with $f = (t-t_0)^2$, the rising rate can be as fast as that measured for our samples. Especially, three of our samples (SHOOT14ha, 14jr, and 14ef) show a good match if the epochs of these objects are shifted so that Day 2 corresponds to $t \sim -2$ days. Then, our data at later epochs are also consistent with the PS1 samples at the declining phase. Since the estimated epoch of zero flux for PS1-10ah and PS1-10bjp is $t_0 \sim -4.2 $ days from the peak, the epochs of our observations correspond to $\sim 1.5 - 2.2$ days after the explosion. The agreement between the luminous 2 objects in our samples (SHOOT14gp and 14or) and PS1 faint samples is not as good as that for the faint 3 objects (SHOOT14ha, 14jr, and 14ef). Note that the direct comparison at the perfectly matched wavelengths is not possible ($< 3000$ Å for SHOOT14gp and 14or while $> 4000$ Å for the PS1 faint samples). Nevertheless, SHOOT14gp and 14or show faster rises than the PS1 faint samples. The rising rates of SHOOT14gp and 14or are $> 3.10$ and $3.12^{+1.11}_{-0.70}$ mag day$^{-1}$, respectively (Table \[tab:objects\]). On the other hand, the rising rate of the PS1 faint sample is [$|\Delta m/ \Delta t|$]{}$< 1.3$ mag day$^{-1}$ even at the fastest phase in the interpolated light curves (see dashed lines in Figures \[fig:LCrapid\] and \[fig:dmdt\]). The nature of these objects are discussed in Section \[sec:discussion\]. Rising rates of transients {#sec:dmdt} ========================== Figure \[fig:dmdt\] shows a summary of rising rate and absolute magnitudes of our samples and other transients shown in Figures \[fig:LCall\], \[fig:LCearly\], and \[fig:LCrapid\]. The figure is shown as a function of rising timescale $\tau_{\rm rise} \equiv 1/$ [$|\Delta m/ \Delta t|$]{}, time to have 1 mag rise. For our objects, SN 2010aq, PS1-13arp, and the PS1 samples, the rising rates are measured only at an interval on the rise as there are no time-series data before the peak. The time interval is $\Delta t \gsim 0.5$ days. For normal SNe, for which good time-series data are available, we measure the rising rate [$|\Delta m/ \Delta t|$]{} as a function of time (connected with lines in Figure \[fig:dmdt\]). In order to match the time interval with other objects, the time interval is kept to be $\Delta t \gsim 0.5$ days. For example, although fine time-series data are available for SN 2006aj before the peak, we measure the rising rate from $t$=0.082 and $t$=0.541 days from the burst ($\Delta t_{\rm rest} = 0.45$ days). For the PS1 faint samples (PS1-10ah and PS1-10bjp), the green dashed lines show the the rising rate measured with $\Delta t_{\rm rest} = 0.5$ days using the light curves interpolated with $f = (t-t_0)^2$. In this diagram, as also discussed in Section \[sec:SN\], it is clear that Type Ia SN shows the fast rise only at the very early phase with faint magnitudes. Core-collapse SNe after a few days from the explosion are located at the region with fainter magnitudes and longer timescales compared with our samples. Our samples share a region similar to SN 2010aq and PS1-13arp, SNe with early UV detection by [*GALEX*]{} [@gezari10; @gezari15], as expected from the comparison in the previous sections (Figure \[fig:LCearly\]). The early peak of SN 2006aj also has a similar rising rate, but it is brighter than our samples. The PS1 luminous samples [@drout14] is located at the region with brighter magnitudes and longer timescales. On the other hand, the PS1 faint samples are closer to the faint three objects in our samples (SHOOT14ha, 14jr, and 14ef). Especially, when the rising rate is measured with the interpolated light curves to have a similar $\Delta t_{\rm rest}$ with our samples, the brightness and the rising timescale of the PS1 faint samples shows fairly good agreement with SHOOT14ha, 14jr, and 14ef (see also Figure \[fig:LCrapid\]). Discussion {#sec:discussion} ========== The properties of our samples of rapidly rising transients are similar to those of very early core-collapse SNe, such as SN 2010aq, PS1-13arp, and SN 2006aj (Figure \[fig:dmdt\]). The faint three objects also show a similarity to the faint population (with $> -19$ mag) of the rapidly rising transients from PS1 [@drout14], which are also interpreted as the very early phase of SNe. For both cases, the best match is obtained when our samples are assumed to be $\sim 1-2$ days after the explosion. By these facts, although we do not have photometric follow-up and spectroscopic identification of our samples, we interpret that the rapidly rising transients presented in this paper are the very early phase of core-collapse SNe. In the following sections, we discuss the nature of the rapidly rising transients based on this interpretation. -------------------- --------------------     -------------------- --------------------   Constraints on the event rate ----------------------------- Event rates of rapidly rising transients shown in this paper are of interest. However, to estimate the event rates, we need detailed information about spectral energy distribution, light curve shape, and luminosity function, which are not available for our samples. Instead, we give crude constraints on how high event rate is required for short-timescale events to be detected with our short-period survey. We estimate the event rates by using a method based on $1/V_{\rm max}$ method [@schmidt68; @eales93], which is used for estimation of galaxy luminosity function. The event rates of transients $R$ can be written as $R = \sum_i R_i = \sum_i \frac{1}{p_i \tau_i V_{{\rm max},i}}$. Here, $p_i$ is a detection efficiency ($p_i < 1$), $\tau_i$ is the restframe time window for a rapidly rising transient to be detected with our survey, and $V_{{\rm max},i}$ is the maximum volume in which the transient is detectable with our survey. The summation is taken for all the detected objects. The difference from galaxy luminosity function is $\tau_i$ in the denominator to take into account the fact that transient event rate should be measured for a given time period. As the number of samples is small, we do not take into account redshift evolution of the event rate. We do [*not*]{} correct detection efficiency since the selection criteria are complicated: we need spectroscopic redshift to define the rapidly rising transients (Section \[sec:selection\]). Thus, we assume $p_i = 1$, so that the analysis gives a conservative lower limit for the event rate (see below for possible impact of this assumption). Then, the free parameter in this analysis is only $\tau_i$. For simplicity, we assume this parameter is the same ($\tau$) for all the objects by neglecting different redshifts. Here, $\tau$ means the duration for which transients show a rapid rise with sufficient brightness so that they are recognized as rapidly rising transients in our survey. For the two objects detected both on Days 1 and 2 (SHOOT14or and 14jr), the duration of the emission is about 1.2 days in the observed frame (0.67 and 0.86 days in the restframe, respectively), and thus, $\tau$ is not much shorter than 1 day. A smaller $\tau$ is not excluded for the other three objects but they do not show clear intranight variability for 1.6-3.1 hr in the observed frame (1.0-2.0 hr in the restframe). Comparison with previously known transients (Section \[sec:LC\]) and also with models (see Section \[sec:model\]) suggest that it is unlikely that the rising rate as high as [$|\Delta m/ \Delta t|$]{} $> 1$ mag day$^{-1}$ continues for $> 2$ days in restframe with sufficient brightness. Thus, we adopt $\tau = 1$ day as a fiducial value for all objects. A typical 3$\sigma$ limiting magnitude for the images used for candidate selection is $\simeq$ 26.0 mag. We use this value for the calculation of the maximum volume $V_{\rm max}$. In fact, for objects to be recognized as rapidly rising transients, they should be sufficiently brighter than the limiting magnitude on Day 2. Thus, the effective limiting magnitude for the rapidly rising transients tends to be shallower than 26.0 mag. Since analysis with a shallower limiting magnitude gives a smaller maximum volume and a higher event rate, our choice of deep limiting magnitude gives conservative estimates for the event rate. It is noted that the extinction in the host galaxy is not corrected and the true absolute magnitude of our samples should be brighter. However, if the extinction for the current samples represents an average degree of extinction, the estimate of $V_{\rm max}$ is not significantly affected ([i.e., ]{}our estimate crudely includes the effect of extinction). We estimate pseudo event rate for each object ($R_i$). For example, the maximum redshift, in which our survey would have detected SHOOT14gp, is $z_{\rm max} = 1.87$ with the limiting magnitude of 26.0 mag using absolute magnitude of $M = -18.67$ mag and crude K-correction (the term of $2.5\log(1+z)$) as in Section \[sec:LC\]. The comoving volume within this redshift in 12 deg$^2$ survey area is $V_{{\rm max},i} = $ 0.16 Gpc$^3$. For this object to be detected with our survey, the required event rate should be $R_i \simeq 1/\tau_i V_{{\rm max},i} \simeq 0.23 \times 10^{-5} (\tau/ 1 {\rm day})^{-1} $ yr$^{-1}$ Mpc$^{-3}$. Similar analysis for SHOOT14or, 14ha, 14jr and 14ef give $z_{\rm max} =$ 1.28, 0.70, 0.82, and 0.62, and the event rates are $R_i \simeq $ 0.47, 1.9, 1.3, 2.5 $\times 10^{-5} (\tau/ 1 {\rm day})^{-1} $ yr$^{-1}$ Mpc$^{-3}$, respectively. By summing up the pseudo rates, the lower limit of the total event rate is $R \simeq 6.4 \times 10^{-5} \ (\tau/ 1 {\rm day})^{-1} $ yr$^{-1}$ Mpc$^{-3}$. It corresponds to about 9 % of core-collapse SN rate at $z \sim 1$ (the core-collapse SN rate is $(3-7) \times 10^{-4}$ yr$^{-1}$ Mpc$^{-3}$ at $z=0-1$, @dahlen04 [@botticella08; @li11; @dahlen12]). Note that the event rate is dominated by the less luminous object with smaller maximum volumes. The event rate for the two luminous events (SHOOT14gp and 14or) is $R = 0.7 \times 10^{-5} \ (\tau/ 1 {\rm day})^{-1} $ yr$^{-1}$ Mpc$^{-3}$ ($\sim 1$ % of the core-collapse SN rate at $z \sim 1$), while the event rate for the three faint events (SHOOT14ha, 14jr, and 14ef) is $R = 5.7 \times 10^{-5} \ (\tau/ 1 {\rm day})^{-1} $ yr$^{-1}$ Mpc$^{-3}$ ($\sim 8$ % of the core-collapse SN rate). It is worthy to mention that the event rate of the rapid transients from PS1 is estimated to be $4\% - 7\%$ of core-collapse SN rate @drout14, which is broadly consistent with our estimate. As described above, our estimate involve crude approximation, mainly due to (1) incompleteness of the sample, (2) a choice of simple magnitude limit, and (3) unknown transient duration. To anchor a possible range of uncertainties, we here discuss impacts of each effect. (1) As discussed in Section \[sec:selection\], we could not take spectra of 6 SN candidates. If all of them satisfy the criteria of rapid transients, the total number of the objects is 11 instead of 5. Actual impact to the event rate depends on their luminosity and redshifts, but if all of them are assumed to be similar to our faint samples (with a high event rate), the total event rate can be increased at most by a factor of about 2.2 (11/5). (2) If a shallow magnitude limit is adopted, it results in a smaller $V_{\rm max}$ and a higher event rate. By adopting 25.5 mag limit, which is the possible shallowest limit to detect SHOOT14ef, the event rate is increased by a factor of 1.7. (3) The effect of duration ($\tau$) is crudely expressed in a term of $\tau^{-1}$ and it can either reduce or increase the event rate. The event rate is reduced by 2 for the duration of $\tau = 2$ days, while it is increased by a factor of 1.4 for the duration of $\tau = 0.7$ days (SHOOT14or). In summary, our rate estimate is uncertain by a factor of $\sim 2$ for reduction and $\sim 5$ for increase. In either case, the event rate is not totally negligible compared with the core-collapse SN rate. Given the crude approximation in the estimate, the true event rate can be comparable to the SN rate, [i.e., ]{}the rapidly rising phase can be associated with all core-collapse SNe. Nature of the rapidly rising transients {#sec:model} --------------------------------------- [*Shock breakout:*]{} The electromagnetic signal from SNe starts with shock breakout emission. Shock breakout occurs when the diffusion timescale of photons in front of the shock wave becomes as short as the dynamical timescale [@falk78; @klein78sbo]. A typical duration of the shock breakout is light crossing time of the progenitor size, [i.e., ]{}$\sim 1000$ sec for a red supergiant progenitor with 500 ${R_{\odot}}$ [[e.g., ]{} @matzner99; @ensman92; @tominaga09sbo; @tominaga11] and shorter for more compact progenitors. Timescales of shock breakout emission are much shorter than the observed timescale for SHOOT14or and 14jr, which are detected both on Days 1 and 2 (0.55-0.72 days in restframe). Therefore, they can not be shock breakout emission. On the other hand, the other three objects (SHOOT14gp, 14ha, and 14ef) are not detected on Day 1, and thus, the possibilities of the shock breakout are not ruled out. However, they do not show significant intranight variability within 1.0, 1.1, and 2.0 hr (restframe) on Day 2, respectively, and there is no supportive signature for shock breakout interpretation (see @tominaga15 for the detection of a transient with an extremely rapid decline, which is interpreted to be shock breakout emission). [*Cooling envelope emission:*]{} Following shock breakout emission, SNe show emission from cooling envelope [@waxman07; @chevalier08; @nakar10; @rabinak11]. This phase is believed to have been detected for SNe with very early detection, such as SNe 2006aj and 2008D [@waxman07; @soderberg08; @modjaz09; @chevalier08; @nakar15 but see @bersten13 for caveats on SN 2008D]. The early UV detection of SN 2010aq (Figure \[fig:LCearly\]) is also interpreted as a cooling emission [@gezari10]. @drout14 also showed that, among their rapid transients from PS1, the faint objects such as PS1-10ah can be interpreted as the cooling envelope emission. In addition to these very early detection, the tail of the cooling phase is sometimes observed in some other SNe, such as SNe 1993J, 1999ex, and 2011dh, at later phases [[e.g., ]{} @lewis94; @richmond94; @stritzinger02; @arcavi11; @marion14]. Figure \[fig:models\] shows light curves of cooling envelope emission for red supergiant cases by @nakar10, compared with light curves of our samples, SN 2010aq, and PS1-13arp. We divide these objects into 4 classes according to effective restframe wavelengths (2000, 2500, 3000, and 3500 Å). The black dashed lines show the fiducial model with the ejecta mass ${M_{\rm ej}}= 15 {M_{\odot}}$, progenitor radius $R = 500 {R_{\odot}}$, and explosion energy $E = 1.0 \times 10^{51}$ erg. Other lines show models with different mass, radius, and energy: upper gray dashed line (${M_{\rm ej}}, R, E$) = ($15 {M_{\odot}}, 1000 {R_{\odot}}, 1.0 \times 10^{51}$ erg), lower gray dashed ($25 {M_{\odot}}, 500 {R_{\odot}}, 1.0 \times 10^{51}$ erg), and red dashed ($15 {M_{\odot}}, 500 {R_{\odot}}, 5.0 \times 10^{51}$ erg). The epochs of observed data are arbitrarily shifted to match the models. The brightness of observed samples is consistent or brighter than the red supergiant models. Since the cooling envelope emission from explosions of more compact progenitor tend to be fainter than red supergiant case in UV at $\sim 1$ day [@nakar10], models with blue supergiant or Wolf-Rayet star progenitors do not give better agreement. The light curve of SHOOT14jr is qualitatively consistent with a model of cooling envelope emission. SHOOT14ha and 14ef can also be explained by the models, although they are detected only Day 2. Since the cooling envelope emission peaks at a epoch when $h \nu \sim 3kT$ is fulfilled, the spectral peak at the rising phase is located at shorter wavelengths than the observed wavelengths. This is also consistent with the blue color of our objects. Note that comparison with the models suggest an explosion energy higher than $1.0 \times 10^{51}$ erg. In addition, due to possible extinction in the host galaxies, the true absolute magnitudes of our objects can be even brighter. These situations are also the case for SN 2010aq, where a model brighter than our fiducial model by $1.5$ mag gives the best match with the observed data without host extinction correction [@gezari10]. To understand possible varieties in the models, we also show selected numerical models for the early phase of Type IIP SNe. The models are calculated with the multigroup radiation hydrodynamics code STELLA [@blinnikov06]. For the purpose of parametric studies, quasi-polytrope pre-SN models are constructed in hydrostatic equilibrium by assuming the solar metallicity and a power-law dependence of the temperature on the density as in @baklanov05 [@baklanov15]. In Figure \[fig:models\], magnitudes in [*Swift*]{} $uvw1$ and $u$-filters are shown in the panels of 2500 Å and 3500 Å data. Black and red solid lines show the models with similar parameters to those for analytic models: (${M_{\rm ej}}, R, E$) = ($15 {M_{\odot}}, 500 {R_{\odot}}, 1.2 \times 10^{51}$ erg) and ($15 {M_{\odot}}, 500 {R_{\odot}}, 4.0 \times 10^{51}$ erg), respectively. Although there are some discrepancy between analytic and numerical models, the trend is similar: SHOOT14jr can be consistent with models while SHOOT14or is brighter and faster than the models. Figure \[fig:dmdt\_model\] shows the rising timescales and absolute magnitudes (as in Figure \[fig:dmdt\]) compared with those of analytic (dashed) and numerical (solid) models. The black and red lines show the fiducial models and models with a higher energy. As also shown in Figure \[fig:models\], the light curve models are consistent with the faint three objects in our samples at $\lsim 1-2$ days after the shock breakout. In summary, the three faint objects (SHOOT14ha, 14jr, and 14ef) out of our five samples are interpreted to be the cooling envelope emission of red supergiant explosion. The epochs of our detection is likely to be $\lsim 1-2$ days after the shock breakout. [*Shock breakout from dense wind:*]{} SHOOT14gp and SHOOT14or, two luminous objects in our samples, are brighter and faster than the cooling envelope models. In fact, this difficulty is also found for the case of PS1-13arp, and @gezari15 suggested that it is shock breakout from a dense wind since the luminosity of the shock breakout from the wind can be more luminous than cooling envelope emission by factor of $\gsim 10$ [@ofek10; @moriya11; @chevalier11; @balberg11]. For the shock breakout from the wind, the timescale to the peak luminosity reflects the diffusion timescale in the wind, $t_p = 6.6 \ (\kappa/0.34\ {\rm cm^2\ g^{-1}}) (\dot{M}/10^{-2}\ {\rm {M_{\odot}}\ yr^{-1}}) (v_{\rm wind} / 10\ {\rm km\ s^{-1}})$ days [@chevalier11], where $\dot{M}$ and $v_{\rm wind}$ is the mass loss rate and wind velocity, respectively. For our samples, the time to the peak is not tightly constrained, but it is longer than $0.55$ days for SHOOT14or. Therefore, the required mass loss rate is the order of $10^{-3}\ {M_{\odot}}$ yr$^{-1}$ for the wind velocity of $v_{\rm wind} = 10$ [km s$^{-1}$]{}. A typical epoch when such a mass loss rate is required is $t_{\rm wind} \sim 2.7 \ (v_{\rm SN}/10,000 \ {\rm km\ s^{-1}}) (v_{\rm wind}/10 \ {\rm km\ s^{-1}})^{-1} (t_{\rm SN}/ 1\ {\rm day})$ years before the explosion, where $v_{\rm SN}$ and $t_{\rm SN}$ are shock velocity of SN and observed time after the explosion, respectively. The inferred mass loss rate is as high as enhanced, episodic mass loss rate estimated for VY Canis Majoris [$(1-2) \times 10^{-3}\ {M_{\odot}}$ yr$^{-1}$, @smith09RSG], and higher than that typically estimated for red supergiants, $\dot{M} \lsim 10^{-4} {M_{\odot}}$ [@vanloon05; @mauron11]. If our interpretation is the case, our study implies that $\gsim 1 \%$ of massive stars can have such a high mass loss rate at the very end of the stellar evolution ([i.e., ]{}a few years before the explosion). @drout14 also suggested that the PS1 luminous samples are the shock breakout from the wind. PS1 luminous samples show longer timescale than those for our two luminous samples and PS1-13arp (Figure \[fig:dmdt\]). This may be understood as the different mass loss rates of the wind: the PS1 luminous samples require a higher mass loss rates $\sim 10^{-2}\ {M_{\odot}}$ yr$^{-1}$ [@drout14]. Conclusions {#sec:conclusions} =========== We perform a high-cadence transient survey using Subaru/HSC. In the observations of two continuous nights, we detected five rapidly rising transients at $z =0.384 - 0.821$ with the rising rate faster than 1 mag per 1 day in restframe ([$|\Delta m/ \Delta t|$]{}&gt; 1 mag day$^{-1}$). The absolute magnitudes of the five objects range from $-16$ to $-19$ mag in the restframe near-UV wavelengths, and they all show blue colors, $g-r \lsim -0.2$ mag. To our knowledge, the rising rate and brightness of our samples are the most similar to those of the very early phase ($<$ a few days after the explosion) of core-collapse SNe, such as SN 2010aq and PS1-13arp detected by [*GALEX*]{} at the very early phases [@gezari10; @gezari15], and the faint population of rapid transients from PS1 [@drout14]. A conservative estimates suggest that the event rate of rapidly rising transients is $\gsim$ 9 % of core-collapse SN rates, assuming a typical duration of the fast rising phase in the near-UV wavelengths to be 1 day. The true event rate can be comparable to the core-collapse SN rate. Although spectroscopic identification is not available, the rapidly rising transients presented in this paper are interpreted to be the very early phase of core-collapse SNe. The observed light curves of faint three objects (SHOOT14ha, 14jr, and 14ef) are qualitatively consistent with the cooling envelope emission from the explosion of red supergiants. The comparison with the analytic and numerical models shows that the epochs of our observations correspond to $\lsim 1-2$ days after the shock breakout. The other two luminous objects (SHOOT14gp and 14or) are brighter and faster than the expectation of the cooling envelope models. We interpret that they are shock breakout emission from the dense wind, as also suggested for PS1-13arp. The required mass loss rate is $\sim 10^{-3}\ {M_{\odot}}$ yr$^{-1}$. The event rate of these luminous events is higher than $\sim 1 \%$ of core-collapse SN rate. Therefore, if our interpretation is correct, it implies that more than $\sim 1 \%$ of massive stars can experience such a strong mass loss at a few years before the explosion. We thank the anonymous referee for constructive comments that improved the paper. This research was in part supported by Grants-in-Aid for Scientific Research of JSPS (23224004, 23740157, 24740117, 25800103, 26400222, 15H02075, 15H05440), MEXT (25103515, 15H00788), the World Premier International Research Center Initiative, MEXT, Japan, the research grant program of Toyota foundation (D11-R-0830), and the RFBR-JSPS bilateral program. S.B. and P.B. are supported in the work on STELLA code by the Russian Science Foundation Grant No. 14-12-00203. TJM is supported by JSPS Postdoctoral Fellowships for Research Abroad (2651). Support for HK is provided by the Ministry of Economy, Development, and Tourism’s Millennium Science Initiative through grant IC120009, awarded to The Millennium Institute of Astrophysics, MAS. HK acknowledges support by CONICYT through FONDECYT grant 3140563. This paper makes use of software developed for the LSST. We thank the LSST Project for making their code available as free software at http://dm.lsstcorp.org. [99]{} natexlab\#1[\#1]{} , H., [et al.]{} 2011, , 193, 29 , C. 2000, , 144, 363 , C., & [Lupton]{}, R. H. 1998, , 503, 325 , I., [et al.]{} 2011, , 742, L18 , T., [Kantor]{}, J., [Lupton]{}, R. H., & [Pierfederici]{}, F. 2010, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 7740, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, 15 , S., [Aragon]{}, C., [Romano]{}, R., [Thomas]{}, R. C., [Weaver]{}, B. A., & [Wong]{}, D. 2007, , 665, 1246 , P. V., [Blinnikov]{}, S. I., & [Pavlyuk]{}, N. N. 2005, Astronomy Letters, 31, 429 , P. V., [Sorokina]{}, E. I., & [Blinnikov]{}, S. I. 2015, Astronomy Letters, 41, 95 , S., & [Loeb]{}, A. 2011, , 414, 1715 , A. C., [et al.]{} 2004, , 611, 418 , E., [et al.]{} 2013, , 779, 18 , M. C., [Tanaka]{}, M., [Tominaga]{}, N., [Benvenuto]{}, O. G., & [Nomoto]{}, K. 2013, , 767, 143 , S. I., [R[ö]{}pke]{}, F. K., [Sorokina]{}, E. I., [Gieseler]{}, M., [Reinecke]{}, M., [Travaglio]{}, C., [Hillebrandt]{}, W., & [Stritzinger]{}, M. 2006, , 453, 229 , J. S., [et al.]{} 2012, , 124, 1175 , M. T., [et al.]{} 2008, , 479, 49 , A. A., [Landsman]{}, W., [Holland]{}, S. T., [Roming]{}, P., [Kuin]{}, N. P. M., & [Page]{}, M. J. 2011, in American Institute of Physics Conference Series, Vol. 1358, American Institute of Physics Conference Series, ed. J. E. [McEnery]{}, J. L. [Racusin]{}, & N. [Gehrels]{}, 373–376 , H., [Richards]{}, J. W., [Poznanski]{}, D., [Bloom]{}, J. S., [Rice]{}, J., [Negahban]{}, S., & [Wainwright]{}, M. 2013, , 435, 1047 , P. J., [et al.]{} 2012, , 753, 22 —. 2010, , 721, 1608 , S., [et al.]{} 2006, , 442, 1008 , R. A., & [Fransson]{}, C. 2008, , 683, L135 , R. A., & [Irwin]{}, C. M. 2011, , 729, L6+ , L., [et al.]{} 2011, , 743, 114 , T., [Strolger]{}, L.-G., [Riess]{}, A. G., [Mattila]{}, S., [Kankare]{}, E., & [Mobasher]{}, B. 2012, , 757, 70 , T., [et al.]{} 2004, , 613, 189 , A. J., [et al.]{} 2009, , 696, 870 , M. R., [et al.]{} 2014, , 794, 23 —. 2013, , 774, 58 , S. 1993, , 404, 51 , L., & [Burrows]{}, A. 1992, , 393, 742 , S. W. 1978, , 225, L133 , B., [et al.]{} 2015, ArXiv e-prints , F., [Maureira]{}, J. C., [Gonzalez-Gaitan]{}, S., & [Galbany]{}, L. 2014, The Astronomer’s Telegram, 5949, 1 , H., [et al.]{} 2011, , 63, 585 , A. 2012, Science, 337, 927 , S., [et al.]{} 2015, , 804, 28 —. 2010, , 720, L77 , D. W., [Baldry]{}, I. K., [Blanton]{}, M. R., & [Eisenstein]{}, D. J. 2002, arXiv:astro-ph/0210394 , Z., [et al.]{} 2008, arXiv:0805.2366 , N., [et al.]{} 2010, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 7733, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, 0 , N., [et al.]{} 2002, , 54, 819 , K., & [Quataert]{}, E. 2015, , 451, 2656 , M. M., [et al.]{} 2010, , 723, L98 , R. I., & [Chevalier]{}, R. A. 1978, , 223, L109 , E., [et al.]{} 2009, , 180, 330 , N. M., [et al.]{} 2009, , 121, 1395 , J. R., [et al.]{} 1994, , 266, L27 , W., [et al.]{} 2011, , 412, 1441 , G. H., [et al.]{} 2014, , 781, 69 , C. D., & [McKee]{}, C. F. 1999, , 510, 379 , N., & [Josselin]{}, E. 2011, , 526, A156 , P. A., [et al.]{} 2006, , 645, 1323 —. 2008, Science, 321, 1185 , B. D., [Piro]{}, A. L., & [Quataert]{}, E. 2009, , 396, 1659 , N., [Halpern]{}, J. P., [An]{}, D., [Thorstensen]{}, J. R., & [Terndrup]{}, D. M. 2006, , 643, L99 , S., [et al.]{} 2006, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 6269, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series , S., [et al.]{} 2012, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 8446, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, 0 , M., [et al.]{} 2009, , 702, 226 —. 2006, , 645, L21 , T., [Tominaga]{}, N., [Blinnikov]{}, S. I., [Baklanov]{}, P. V., & [Sorokina]{}, E. I. 2011, , 415, 199 , T., [et al.]{} 2014, , 66, 114 , E. 2015, , 807, 172 , E., & [Sari]{}, R. 2010, , 725, 904 , P. E., [et al.]{} 2011, , 480, 344 , E. O., [et al.]{} 2010, , 724, 1396 , A., [et al.]{} 2013, , 767, 1 —. 2010, , 724, L16 , E., [et al.]{} 2006, , 442, 1011 , D., [et al.]{} 2010, Science, 327, 58 , J. L., [Brimacombe]{}, J., [Drake]{}, A. J., & [Howerton]{}, S. 2013, , 763, L27 , T. A., [Roming]{}, P. W. A., [Brown]{}, P. J., [Bayless]{}, A. J., & [Frey]{}, L. H. 2014, , 787, 157 , R. M., [et al.]{} 2011, , 474, 487 , R. M., [Wheeler]{}, J. C., [H[ö]{}flich]{}, P., [Akerlof]{}, C. W., [Brown]{}, P. J., & [Rykoff]{}, E. S. 2007, , 666, 1093 , I., & [Waxman]{}, E. 2011, , 728, 63 , A., [et al.]{} 2009, , 121, 1334 , A., [Ofek]{}, E. O., [Kulkarni]{}, S. R., [Madore]{}, B. F., [Pevunova]{}, O., & [Ajello]{}, M. 2008, , 682, 1205 , M. W., [Treffers]{}, R. R., [Filippenko]{}, A. V., [Paik]{}, Y., [Leibundgut]{}, B., [Schulman]{}, E., & [Cox]{}, C. V. 1994, , 107, 1022 , A., [et al.]{} 2015, ArXiv e-prints , S., [et al.]{} 2012, Proc. SPIE, 8446, 84466L, 0 , M. 1968, , 151, 393 , N., [Hinkle]{}, K. H., & [Ryde]{}, N. 2009, , 137, 3558 , A. M., [et al.]{} 2008, , 453, 469 —. 2006, , 442, 1014 , J., [et al.]{} 2006, , 454, 503 , M., [et al.]{} 2002, , 124, 2100 , M., [et al.]{} 2014, , 793, L26 —. 2009, , 692, 1131 —. 2009, , 700, 1680 , N., [Blinnikov]{}, S., [Baklanov]{}, P., [Morokuma]{}, T., [Nomoto]{}, K., & [Suzuki]{}, T. 2009, , 705, L10 , N., [Morokuma]{}, T., [Blinnikov]{}, S. I., [Baklanov]{}, P., [Sorokina]{}, E. I., & [Nomoto]{}, K. 2011, , 193, 20 , N., [et al.]{} 2015, submitted to —. 2015, The Astronomer’s Telegram, 7565, 1 —. 2014, The Astronomer’s Telegram, 6291, 1 —. 2014, The Astronomer’s Telegram, 6763, 1 —. 2015, The Astronomer’s Telegram, 7927, 1 , V., [Pizzichini]{}, G., & [Hudec]{}, R. 2010, , 523, A56 , J. T., [Cioni]{}, M.-R. L., [Zijlstra]{}, A. A., & [Loup]{}, C. 2005, , 438, 273 , E., [M[é]{}sz[á]{}ros]{}, P., & [Campana]{}, S. 2007, , 667, 351 , M., [et al.]{} 2014, , 782, L35 [^1]: Based \[in part\] on data collected at Subaru Telescope, which is operated by the National Astronomical Observatory of Japan [^2]: <http://www2.iap.fr/users/alard/package.html> [^3]: <http://www.astro.washington.edu/users/becker/v2.0/hotpants.html>

--- abstract: 'We present a method to map the absolute electromagnetic field strength *inside* photonic crystals. We apply the method to map the electric field component $E_z$ of a two-dimensional photonic crystal slab at microwave frequencies. The slab is placed between two mirrors to select Bloch standing waves and a subwavelength spherical scatterer is scanned inside the resulting resonator. The resonant Bloch frequencies shift depending on the electric field at the position of the scatterer. To map the electric field component $E_z$ we measure the frequency shift in the reflection and transmission spectrum of the slab versus the scatterer position. Very good agreement is found between measurements and calculations without any adjustable parameters.' address: | $^1$ Laser Physics and Nonlinear Optics, MESA+ Institute for Nanotechnology, University of Twente, P.O.Box 217, 7500 AE Enschede, The Netherlands\ $^2$ Complex Photonic Systems (COPS), MESA+ Institute for Nanotechnology, University of Twente, P.O.Box 217, 7500 AE Enschede, The Netherlands author: - 'T. Denis,$^{1,*}$ B. Reijnders,$^1$ J. H. H. Lee,$^1$ P. J. M. van der Slot,$^1$ W. L. Vos,$^2$ and K.-J. Boller$^1$' title: Mapping individual electromagnetic field components inside a photonic crystal --- [99]{} S. John, “Strong localization of photons in certain disordered dielectric superlattices,” Phys. Rev. Lett. [**58,**]{} 2486–2489 (1987). E. Yablonovitch, “Inhibited spontaneous emission in solid-state physics and electronics,” Phys. Rev. Lett. [**58,**]{} 2059–2062 (1987). J. D. Joannopoulos, S. G. Johnson, J. N. Winn, and R. D. Meade, [*Photonic crystals: molding the flow of light*]{} (Princeton University Press, 2008). N. W. Ashcroft and N. D. Mermin, [*Solid state physics*]{} (Holt, Rinehard & Winston, 1976). R. Sprik, B. A. van Tiggelen, and A. Lagendijk, “Optical emission in periodic dielectrics,” Europhys. Lett. [**35,**]{} 265–270 (1996). P. Lodahl, A. F. van Driel, I. S. Nikolaev, A. Irman, K. Overgang, D. Vanmaekelbergh, and W. L. Vos, “Controlling the dynamics of spontaneous emission from quantum dots by photonic crystals,” Nature [**430,**]{} 654–657 (2004). M. Fujita, S. Takahashi, Y. Tanaka, T. Asano, and S. Noda, “Simultaneous inhibition and redistribution of spontaneous light emission in photonic crystals,” Science [**308,**]{} 1296–1298 (2005). H. Caglayan, I. Bulu, and E. Ozbay, “Highly directional enhanced radiation from sources embedded inside three-dimensional photonic crystals,” Opt. Express [**13,**]{} 7645–7652 (2005). A. F. Koenderink, M. Kafesaki, C. M. Soukoulis, and V. Sandoghdar, “Spontaneous emission in the near-field of two-dimensional photonic crystals,” Opt. Lett. [**30,**]{} 3210–3212 (2005). A. Rodenas, G. Zhou, D. Jaque, and M. Gu, “Rare-earth spontaneous emission control in three-dimensional lithium niobate photonic crystals,” Adv. Mater. [**21,**]{} 3526–-3530 (2009). L. Sapienza, H. Thyrrestrup, S. Stobbe, P. D. Garcia, S. Smolka, and P. Lodahl, “Cavity quantum electrodynamics with Anderson-localized modes,” Science [**327,**]{} 1352–1355 (2010). M. R. Jorgensen, J. W. Galusha, and M. H. Bart, “Strongly modified spontaneous emission rates in Diamond-structured photonic crystals,” Phys. Rev. Lett. [**107,**]{} 143902 (2011). M. D. Leistikow, A. P. Mosk, E. Yeganegi, S. R. Huisman, A. Lagendijk, and W. L. Vos, “Inhibited spontaneous emission of quantum dots observed in a 3D photonic band gap,” Phys. Rev. Lett. [**107,**]{} 193903 (2011). T. Yoshie, A. Scherer, J. Hendrickson, G. Khitrova, H. M. Gibbs, G. Rupper, C. Ell, O. B. Shchekin, and D. G. Deppe, “Vacuum Rabi splitting with a single quantum dot in a photonic crystal nanocavity,” Nature [**432,**]{} 200–203 (2004). O. Painter, R. K. Lee, A. Scherer, A. Yariv, J. D. O’Brien, P. D. Dapkus, and I. Kim, “Two-dimensional photonic band-gap defect mode laser,” Science [**284,**]{} 1819–1821 (1999). H.-G. Park, S.-H. Kim, S.-H. Kwon, Y.-G. Ju, J.-K. Yang, J.-H. Baek, S.-B. Kim, and Y.-H. Lee, “Electrically driven single-cell photonic crystal laser,” Science [**305,**]{} 1444–1447 (2004). H. Altug, D. Englung, and J. Vuckovic, “Ultrafast photonic crystal nanocavity laser,” Nature Phys. [**2,**]{} 484–488 (2006). K. Nozaki, S. Kita, and T. Baba, “Room temperature continuous wave operation and controlled spontaneous emission in ultrasmall photonic crystal nanolaser,” Opt. Express [**15,**]{} 7506–7514 (2007). P. J. M. van der Slot, T. Denis, and K.-J. Boller, “The photonic FEL: toward a handheld THz FEL,” in [*Proc. of the FEL 2008,*]{} V. Schaa, ed. (JACoW, 2008), pp. 231–234. H. K. Park, J. R. Oh, Y. R. Do, “2D SiN$_\mathrm{x}$ photonic crystal coated $\mathrm{Y_3Al_5O_{12}:Ce^{3+}}$ ceramic plate phosphor for high-power white light-emitting diodes,” Opt. Express [**19,**]{} 25593–25601 (2011). M. Florescu, H. Lee, I. Puscasu, M. Pralle, L. Florescu, D. Z. Ting, and J. P. Dowling, “Improving solar cell efficiency using photonic band-gap materials,” Sol. Energy Mater. Sol. Cells [**91,**]{} 1599–1610 (2007). D.-H. Ko, J. R. Tumbleston, L. Zhang, S. Williams, J. M. DeSimone, R. Lopez, and E. T. Samulski, “Photonic crystal geometry for organic solar cells,” Nano Lett. [**9,**]{} 2742–2746 (2009). A. F. Oskooi, D. Roundy, M. Ibanescu, P. Bermel, J. D. Joannopoulos and S. G. Johnson, “MEEP: A flexible free-software package for electromagnetic simulations by the FDTD method,” Comput. Phys. Commun. [**181,**]{} 687–702 (2010). A. F. Koenderink and W. L. Vos, “Optical properties of real photonic crystals: anomalous diffuse transmission,” J. Opt. Soc. Am. B [**22,**]{} 1075–1084 (2005). M. L. M. Balistreri, H. Gersen, J. P. Korterik, L. Kuipers, and N. F. van Hulst, “Tracking femtosecond laser pulses in space and time” Science [**294,**]{} 1080–1082 (2001). S. I. Bozhevolnyi, V. S. Volkov, J. Arentoft, A. Boltasseva, T. Sondergaard, and M. Kristensen, “Direct mapping of light propagation in photonic crystal waveguides,” Opt. Commun. [**212,**]{} 51–-55 (2002). L. Okamoto, M. Loncar, T. Yoshie, A. Scherer, Y. Qiu, and P. Gogna, “Near-field scanning optical microscopy of photonic crystal nanocavities,” Appl. Phys. Lett. [**82,**]{} 1676–-1678 (2003). P. Kramper, M. Agio, C. M. Soukoulis, A. Birner, F. Müller, R. B. Wehrspohn, U. Gösele, and V. Sandoghdar, “Highly directional emission from photonic crystal waveguides of subwavelength width” Phys. Rev. Lett. [**92,**]{} 113903 (2004). H.-H. Tao, R.-J. Liu, Z.-Y. Li, S. Feng, Y.-Z. Liu, C. Ren, B.-Y. Cheng, D.-Z. Zhang, H.-Q. Ma, L.-A. Wu, and Z.-B. Zhang, “Mapping of complex optical field patterns in multimode photonic crystal waveguides by near field scanning optical microscopy,” Phys. Rev. B [**74,**]{} 205111 (2006). M. Abashin, P. Tortora, I. Märki, U. Levy, W. Nakagawa, L. Vaccaro, H. Herzig, and Y. Fainman, “Near-field characterization of propagating optical modes in photonic crystal waveguides,” Opt. Express [**14,**]{} 1643–1657 (2006). S. Vignolini, F. Intonti, F. Riboli, D. S. Wiersma, L. Balet, L. H. Li, M. Francardi, A. Gerardino, A. Fiore, and M. Gurioli, “Polarization-sensitive near-field investigation of photonic crystal microcavities,” Appl. Phys. Lett. [**94,**]{} 163102 (2009). J. Dahdah, M. Pilar-Bernal, N. Courjal, G. Ulliac, and F. Baida, “Near-field observations of light confinement in a two dimensional lithium niobate photonic crystal cavity,” J. Appl. Phys. [**110,**]{} 074318 (2011). E. Flück, N. F. van Hulst, W. L. Vos, and L. Kuipers, “Near-field optical investigation of three-dimensional photonic crystals,” Phys. Rev. E [**68,**]{} 015601 (2003). P. Burgos, Z. Lu, A. Ianoul, C. Hnatovsky, M. L. Viriot, L. J. Johnston, R. S. Taylor, “Near-field scanning optical microscopy probes: a comparison of pulled and double-etched bent NSOM probes for fluorescence imaging of biological samples,” J. Microsc. [**211,**]{} 37–47 (2003). N. E. Dickenson, E. S. Erickson, M. L. Olivia, and R. C. Dunn, “Characterization of power induced heating and damage in fiber optic probes for near-field scanning optical microscopy,” Rev. Sci. Instrum. [**78,**]{} 053712 (2007). B. Hecht, H. Bielefeldt, Y. Inouye, D. W. Pohl, and L. Novotny, “Facts and artifacts in near-field optical microscopy,” J. Appl. Phys. [**81,**]{} 2492–2498 (1997). R. Carminati, A. Madrazo, M. Nieto-Vesperinas, and J.-J. Greffet, “Optical content and resolution of near-field optical images: Influence of the operating mode,” J. Appl. Phys. [**82,**]{} 501–509 (1997). P. J. Valle, J.-J. Greffet, and R. Carminati, “Optical contrast, topographic contrast and artifacts in illumination-mode scanning near-field optical microscopy,” J. Appl. Phys. [**86,**]{} 648–656 (1999). K. D. Weston and S. K. Buratto, “A reflection near-field scanning optical microscope technique for subwavelength resolution imaging of thin organic films,” J. Phys. Chem. B [**101,**]{} 5684–5691 (1997). D. C. Kohlgraf-Owens, S. Sukhov, and A. Dogariu, “Optical-force-induced artifacts in scanning probe microscopy,” Opt. Lett. [**36,**]{} 4758–4760 (2011). M. Labardi, S. Patane, and M. Allegrini, “Artifact-free near-field optical imaging by apertureless microscopy,” Appl. Phys. Lett. [**77,**]{} 621–623 (2000). M. Esslinger, J. Dorfmüller, W. Khunsin, R. Vogelgesang, and K. Kern, “Background-free imaging of plasmonic structures with cross-polarized apertureless scanning near-field optical microscopy,” Rev. Sci. Instrum. [**83,**]{} 033704 (2012). L. C. Maier, “Field strength measurements in resonant cavities,” (Massachusetts Institute of Technology, 1949). L. C. Maier and J. C. Slater, “Field strength measurements in resonant cavities,” J. Appl. Phys., [**23,**]{} 68–77 (1952). C. C. Johnson, [*Field and wave electrodynamics*]{} (McGraw-Hill, 1965). R. A. Marsh, M. A. Shapiro, R. J. Temkin, V. A. Dolgashev, L. L. Laurent, J. R. Lewandowski, A. D. Yeremian, and S. G. Tantawi, “X-band photonic band-gap accelerator structure breakdown experiment,” Phys. Rev. STAB [**14,**]{} 021301 (2011). T. Denis P. J. M. van der Slot, and K.-J. Boller, “Experimental design of a single beam photonic free-electron laser,” in [*Proc. of the FEL 2009*]{}, S. Waller, ed. (JACoW, 2009), pp. 431–434. Concerto V7.5, Cobham Ltd., UK, http://www.cobham.com B. Guru and H. Hiziroglu, [*Electromagnetc field theory and fundamentals*]{} (Pws Pub Co, UK, 1997). H. Guo, Y. Carmel, W. R. Lou, L. Chen, J. Rodgers, D. K. Abe, A. Bromborsky, W. Destler, and V. Granatstein, “A novel highly accurate synthetic technique for determination of the dispersive characteristics in periodic slow wave circuits,” IEEE Trans. Microwave Theory Tech. [**40,**]{} 2086–2094 (1992). M. Kageshima, H. Jensenius, M. Dienwiebel, Y. Nakayama, H. Tokumoto, S. P. Jarvis, and T. H. Oosterkamp, “Noncontact atomic force microscopy in liquid environment with quartz tuning fork and carbon nanotube probe,” Appl. Surf. Sci. [**188,**]{} 440–444 (2002). M. Frimmer, Y. Chen, and A. F. Koenderink, “Scanning emitter lifetime imaging microscopy for spontaneous emission control,” Phys. Rev. Lett. [**107,**]{} 123602 (2011). Introduction ============ Photonic crystals attract a tremendous deal of interest as they offer to radically control the light propagation [@John1987] and emission of light [@Yablonovitch1987]. In photonic crystals the dielectric constant varies periodically on the order of the wavelength [@Joannopoulos2008]. Due to this periodicity, the light propagates in the form of Bloch modes [@Joannopoulos2008; @Ashcroft1976]. An intriguing capability of photonic crystals is to shape the local density of electromagnetic states (LDOS) inside the crystal , which is the key for controlling the interaction of light with matter [@Sprik1996]. Manipulating the LDOS allows, for example, the inhibition or the enhancement of spontaneous emission of embedded light sources \[6-13\]. This forms the basis for investigating the strong-coupling cavity regime in quantum electrodynamics [@Yoshie2004]. Such manipulations also have far-reaching technological implications, such as the development of efficient micro scale lasers, LEDs or solar cells \[15-22\]. The field strength of the Bloch modes inside the photonic crystal at the locations of the emitters determine the local character of the LDOS [@Sprik1996]. Bloch mode fields of ideal photonic crystals, *i.e.*, assuming a perfect periodicity, can be calculated by numerical methods such as finite-difference time domain (FDTD) [@Oskooi2010]. However, all real photonic crystals suffer inevitably from unpredictable non-periodic local deviations both due to fabrication errors and, also fundamentally, due to thermodynamical arguments [@Ashcroft1976; @Koenderink2005]. Such deviations can not effectively be included into any kind of numerical calculations to date. A measurement is the only way to analyze the electromagnetic field inside a real photonic crystal. To eventually also characterize the LDOS of photonic crystals the field measurement method should resolve the direction and absolute value of the individual field components ($E_x$, $E_y$, $E_z$, $H_x$, $H_y$, $H_z$). A method to measure the absolute field strength of an eigenmode *inside* a real photonic crystal, however, has not been reported to date. So far the only optical method to map local fields is near-field scanning optical microscopy (NSOM) \[25-42\]. The technique relies on scanning a small tip of a tapered optical fiber *above* the surface of a photonic crystal that collects part of the evanescent field with subwavelength resolution. However, NSOM suffers from some several drawbacks. First, the method is restricted to probe local fields *outside* the crystal near its surfaces while the field deep *inside* the crystal cannot be mapped. Thus one needs to make assumptions about the scattering from the internal fields under study to the detected evanescent fields. In case of three-dimensional crystals, relating the fields at the surface to the derived fields in the bulk is even more challenging [@Flueck2003] than in the widely studied two-dimensional slab systems. Second, the fabrication and handling of the fiber tip is difficult to reproduce and its precise shape and size are hard to control [@Burgos2003; @Dickenson2007]. Hence, a pure NSOM tip probes an unknown superposition of electromagnetic field components. Third, NSOM measurements are strongly affected by a number of background effects \[36-40\] making absolute field measurements very difficult. Even in most sophisticated NSOM methods which provide nearly background free detection schemes absolute fields have not been reported [@Labardi2000; @Esslinger2012]. What is desirable is to devise a method that can probe the absolute strength of the electromagnetic field *inside* a photonic crystal. Such a method can eventually also provide a separate mapping of the individual field components, to determine the LDOS in a real photonic crystal. Here we demonstrate a method to measure the absolute strength of the electromagnetic field distribution *inside* a photonic crystal. Our method relies on measuring the resonant frequencies of Bloch standing waves in a photonic crystal of finite length that is enclosed by two mirrors. A subwavelength scatterer placed inside the crystal scatters the electromagnetic field which shifts the frequency of the Bloch resonances proportionally to the square of the electric and magnetic field strength at the scatterer position \[43-46\]. By measuring the frequency shifts as a function of the spatial position of the scatterer, we obtain maps of the field strengths versus position. We demonstrate the method at microwave frequencies where fabrication errors are relatively small. Furthermore, in this frequency range the typical structure sizes are sufficiently large that a scatterer can be conveniently scanned through the crystal. To simplify the demonstration we deliberately choose a photonic crystal design where a single field component inside the microwave photonic crystal slab dominates throughout most of the crystal. Using a spherical, metallic bead as a scatterer we thus map the dominant electric field component $E_z$ *inside* a photonic crystal slab. Measurement method ================== To map the electromagnetic field, a photonic crystal is placed between two mirrors. The mirrors restrict the electromagnetic field in the photonic crystal to discrete longitudinal Bloch modes with associated resonance frequencies $\nu_0$ and wave numbers $k_z$. To measure the electromagnetic field we place a scatterer inside the photonic crystal and measure the shift of the resonance frequencies. We chose a scatterer in the Rayleigh regime with a small size parameter $x\ll1$, with $x = 2\pi R/\lambda$ . Here, $R$ is the radius of the object and $\lambda$ the wavelength. In this regime the field is approximately constant throughout the scatterer volume, thereby the scattering can be treated within the electrostatic approximation to calculate the resulting frequency shift $\Delta\nu$ due to the scatterer. Using a spherical metallic scatterer with a radius $R$ we obtain \[43-45\]. $$\begin{aligned} \Delta\nu(\mathbf{r})= \frac{\frac{1}{2}\mu_0\mathbf{H}(\mathbf{r})^2 - \epsilon_0\mathbf{E}(\mathbf{r})^2}{U}\pi R^3\nu_{0} \label{eq:shif}\end{aligned}$$ where $\mathbf{E}(\mathbf{r})$ and $\mathbf{H}(\mathbf{r})$ are the unperturbed electric and magnetic fields at the location $\mathbf{r}$ of the scatterer, respectively. $U$ is the total energy stored inside the unperturbed cavity, $\epsilon_0$ is the permittivity and $\mu_0$ the permeability of free space. If we measure the frequency shift versus the scatterer location $\Delta\nu(\mathbf{r})$ we will map, in a general photonic crystal, the electromagnetic field quantity $\left(\frac{1}{2}\mu_0\mathbf{H}(\mathbf{r})^2 - \epsilon_0\mathbf{E}(\mathbf{r})^2\right)$. In certain photonic crystal geometries, however, certain field components strongly dominate, hence these components can be mapped. For instance if a crystal is designed to have the $E_z$ field dominate, *i.e.*, the $E_z$ field strength is much greater than all other field components, the frequency shift becomes $$\begin{aligned} \Delta\nu(\mathbf{r})\approx -\frac{\epsilon_0E_z(\mathbf{r})^2}{U}\pi R^3\nu_{0} \label{eq:shift_full}\end{aligned}$$ Solving for the $E_z$ component yields $$\begin{aligned} E_z(\mathbf{r})=\sqrt{\frac{-\Delta\nu(\mathbf{r}) U}{\pi R^3\epsilon_0\nu_{0}}}. \label{eq:field}\end{aligned}$$ Equation (\[eq:field\]) shows that it is possible to map the absolute strength of the $E_z$ field component by measuring the frequency shift of the longitudinal resonances $\Delta\nu$ versus the bead position and by determining the total energy stored in the cavity $U$ for a specific input power $P_{in}$. The photonic crystal slab ========================= The unit cell of the photonic crystal slab we used (Fig. \[figure1\]a) is designed to provide a dominant $E_z$ component in the structure, which is also required for its intended application in a photonic free-electron laser [@Denis2009]. The unit cell is a supercell which is surrounded by a metallic waveguide creating a two dimensional photonic crystal slab. A rectangular lattice of metal rods with a central line defect at $x=0$ forms the basis of the supercell. Along the z-direction the rectangular lattice has a lattice constant of $a_z=7.5\,\mathrm{mm}$ and the lattice constant is $a_x=6.75\,\mathrm{mm}$ along the $x$-direction. Inside the supercell the third transverse row of rods is missing. Thus the supercell consists of 12 rods and has a length of $a_{z,\mathrm{eff}\,}=22.5\,\mathrm{mm}$, indicated in Fig. \[figure1\]a. The diameter of the rods is 4mm and the surrounding waveguide has a width of $w=47.25\,\mathrm{mm}$ and a height of $h=20\,\mathrm{mm}$. ![(a) Schematic three-dimensional view of the photonic crystal slab indicating the defining geometry parameters. Metallic rods are placed inside a rectangular metallic waveguide. The dashed line indicates the size of the supercell. (b) Calculated band structure of the photonic crystal slab showing the four lowest TE-like modes having a non-zero $E_z$ component of the electric field.[]{data-label="figure1"}](figure1.pdf){width="0.98\columnwidth"} To calculate the band structure of the photonic crystal slab a finite-difference time-domain (FDTD) method is used [@Concerto]. In the calculations the photonic crystal slab is taken to be infinitely long along the $z$-direction by applying appropriate periodic boundary conditions to the unit cell. All metal parts are treated as perfect electric conductors which is well justified in the microwave range. Figure \[figure1\]b shows the results for the four lowest-frequency TE-like modes, *i.e.*, modes with a non-zero longitudinal electric field component $E_z$. Due to the $z$-periodicity of the slab, the dispersion in the first Brillouin zone, *i.e.*, for normalized wave numbers ($a_{z,\mathrm{eff}}\,k_z$) between 0 and $2\pi$, repeats with increasing wave number [@Joannopoulos2008]. Furthermore, the finite transverse size of the waveguide results in a lowest allowed normalized frequency of 0.61 for mode 1. No other TE-like modes exist below this cut-off frequency. ![Transverse $E_z$-eigenmode patterns of the photonic crystal slab for mode 1 and mode 2 at two cross sections (xy-plane) inside the unit cell. First, through the first row of rods (at $z=0.5a_z$, (a) and (b)). Second, at a cross section through the empty part of the waveguide (at $z=2.5a_z$, (c) and (d)). The normalized wave number for both patterns is $k_za_{z,\mathrm{eff}}\,=0$ and the corresponding normalized frequency for (a) is 0.61 and for (b) 0.71.[]{data-label="figure2"}](figure2.pdf){width="0.90\columnwidth"} For a comparison with experimental field mapping data we calculate the local $E_z$ field distribution of mode 1 and mode 2. The FDTD method is applied at the resonant frequency of mode 1 and 2 for a normalized wave number of $a_{z,\mathrm{eff}}\,k_z = 0$ which corresponds to a normalized frequency of 0.61 and 0.71 respectively. Figure \[figure2\]a through \[figure2\]d show the $E_z$ field pattern of the modes at two transverse planes at different $z$ coordinates. The first plane is taken through the center of the first row of rods ($z=0.5a_z$) and the second plane is taken in the empty row of the unit cell ($z=2.5a_z$) where mapping is performed. As expected, the field pattern is symmetric due to the symmetry of the photonic crystal slab. Furthermore, throughout all field patterns it is observed that mode 2 has a field node at the center of the waveguide which will explain a certain effect in the experimental data later on. Experimental setup ================== A schematic view of the experimental setup is shown in Fig. \[figure3\]. The photonic crystal, described in section 3, is placed between two highly reflective alumimun mirrors to form a resonator. The longitudinal resonator contains 15 unit cells of the photonic crystal in total. The first mirror is positioned at a distance of $0.5a_z$ from the center of the first row of rods and the second mirror is positioned at a distance of $1.5a_z$ from the last row of rods. ![Schematic view of the setup to measure the electromagnetic field inside the photonic crystal slab by a scatterer. The figure shows a cross section through the photonic crystal slab at $y=0$. The photonic crystal slab is sandwiched between two aluminum mirrors (bold black). The input and output antennas are mounted at the center of both mirrors. To map the $E_z$ field component along the $x$-direction a spherical metallic scatterer, which is mounted on a string, can be moved throughout the photonic crystal slab.[]{data-label="figure3"}](figure3.pdf){width="0.75\columnwidth"} The photonic crystal slab is fabricated as follows. A channel with a rectangular cross section of $47.25\,\mathrm{mm}\times20\,\mathrm{mm}$ is milled into a solid aluminum block (bottom plate) and is covered with an aluminum top plate. Both parts contain holes for mounting the rods. The rods are hollow brass cylinders with an inner diameter of 2mm and an outer diameter of 4mm. With screws extending from the top plate through the rods into the bottom plate the rods are positioned. This also provides an electrical connection between the rods and the waveguide. The total positioning accuracy for the rods is estimated to a maximum of about $100\,\mu\mathrm{m}$, corresponding to a high precision of about 1.4% relative to the inter rod distances. To measure the resonant frequencies of the photonic crystal slab cavity, a Hertzian dipole antenna [@Guru] is mounted at the center of each mirror (see Fig. \[figure3\]). Both antennas point along the $z-$direction to excite or detect modes with a non-zero $E_z$ field component. One antenna acts as an emitter – labeled “in” in Fig. \[figure3\] – while the other antenna acts as a receiver for transmission measurements – labeled “out” in Fig. \[figure3\]. As a compromise between minimal loading of the resonator by the antennae and sufficient coupling to the modes, the length of both antennas was selected as 4mm. Note that with the antenna position in the center of the mirror, modes with an $E_z$ field node in the center cannot be excited such as mode 2 shown in Fig. \[figure2\]b. The other end of both antennae is connected to a network analyzer via coaxial SMA cables. The network analyzer is formed by a tunable microwave source with a maximum frequency of 20GHz (Wiltron, model 69147A), two directional couplers (Krytar, model 2610) and two microwave power meters (Anritsu, model ML1438A with power heads MA2444A and MA2424B). To compensate for frequency dependent losses in cables and other components the network analyzer is calibrated before the measurements. The accuracy is estimated to about 10%, as re-connecting coaxial cables typically has an effect in this magnitude. Using this setup we measure the reflection and transmission spectra of the photonic crystal slab cavity. For the measurements the frequency resolution is set to 250kHz and the input power $P_{in}$ from the network analyzer is 1mW. To map the electromagnetic field inside the photonic crystal slab the scatterer is scanned through various locations inside the resonator. The scatterer is a stainless steel bead with a radius of $R=2\,\mathrm{mm}$ which sits on a 0.3mm thick nylon string. As shown in Fig. \[figure3\] the string runs through two small holes ($300\,\mu\mathrm{m}$) in the opposing side walls. The holes are positioned in the center of the side walls ($y=0$) and at $z=2.5a_z$ inside the 3rd unit cell, as counted from the input side. At the position of the holes in the inner surface of the side walls a 4mm deep cylindrical cavity with a diameter of 4.5mm is fabricated in which the bead can fit completely. One end of the string is attached to a weight to keep the string straight via tension, and the other end of the string is mounted on a translation stage. The translation stage is used to position the bead with a relative accuracy of better than $0.05\,\mathrm{mm}$. The absolute position is calibrated using the position at which the scatterer just completely disappears within the cylindrical cavity, and the error in this position is estimated to be smaller than $0.1\,\mathrm{mm}$. Dispersion measurement ====================== Before measuring the $E_z$ field component of the photonic crystal slab resonator we verified the appropriate description of the slab by the FDTD model. Specifically, to confirm the crystal dispersion we measured the resonance frequencies of the different longitudinal and transverse modes without a scatterer inside the photonic crystal. Measuring the dispersion is based on determining the longitudinal resonance frequencies of a finite-length photonic crystal consisting of $n$ unit cells and assigning wave numbers to each observed resonance frequency [@Guo1992]. For the assignment we consider the phase advance $\delta\phi$ of standing waves per round trip in the resonator. At each longitudinal resonance the phase advance per round trip along the $z$-direction is a multiple of $\pi$: $$\begin{aligned} n\delta\phi = na_{z,\mathrm{eff}}\,k_z = m\pi\qquad\qquad m={1,2...} \label{wave}\end{aligned}$$ Here $L=na_z$ is the geometrical length of the resonator and $k_z$ is the wave number. As the resonator mirrors enclose 15 unit cells of the photonic crystal, $n=15$ longitudinal resonances with a finite wavelength are expected [@Guo1992] having a phase advance $\delta\phi$ or normalized wave number $a_{z,\mathrm{eff}}\,k_z$ of $$\begin{aligned} 0<\delta\phi=a_{z,\mathrm{eff}}\,k_z=\frac{m}{n}\pi\leq\pi. \label{eq:wavenumber}\end{aligned}$$ ![Transmission and reflection spectrum of the photonic crystal slab without a scatterer between 8.0GHz and 13.5GHz. The peaks correspond to the various longitudinal modes for each transverse mode.[]{data-label="figure4"}](figure4.pdf){width="0.68\columnwidth"} from the calculations shown in Fig. \[figure1\]b the frequency of the considered modes is seen to be a monotonously increasing or decreasing function of the wave number which renders the resonances denumerable. By using eq. (\[eq:wavenumber\]), a normalized wave number can be associated to each longitudinal resonance frequency, from which the dispersion of that transverse mode is obtained. Figure \[figure4\] shows the measured transmission and reflection power spectrum of the unperturbed photonic crystal slab resonator in the range from $8\,\mathrm{GHz}$ to $14\,\mathrm{GHz}$ on a logarithmic power scale. For most of the measured frequencies the reflection is close to 0dB meaning that it is equal to the input power of 1mW. Due to the resonator a radiation field can only build up at frequencies where longitudinal resonances of a transverse photonic crystal mode exist. In both spectra we clearly observe the resonances belonging to the longitudinal resonances. As expected for such resonances the frequency of reflection and transmission resonances agree very well with each other. By comparing the transmission levels of the various detected resonances we identify which resonance belongs to the same transverse mode. Four frequency ranges can be distinguished in Fig. \[figure4\]. From $8.0\,\mathrm{GHz}$ to $9.8\,\mathrm{GHz}$ sharp resonances appear with an averaged normalized transmission of about -15 dB. From $9.9\,\mathrm{GHz}$ to $10.7\,\mathrm{GHz}$ the transmission is very low ($< -40\,\mathrm{dB}$) and only one weak and broad resonance is observed. However, this is not a stop band of the slab, but a mode that cannot be effectively excited with a centered Hertzian dipole antenna due to its mode symmetry (Fig. \[figure2\]). From $10.8\,\mathrm{GHz}$ to $11.8\,\mathrm{GHz}$ another set of resonances appears. This set of resonances is easily distinguished from the following set between $11.8\,\mathrm{GHz}$ to $13.3\,\mathrm{GHz}$ whose average transmission level ($-10\,\mathrm{dB}$) is much higher. In total three different sets of resonances are identified as three different transverse modes. To retrieve the dispersion of each transverse mode we assign a wave number to the longitudinal resonances of each transverse mode. As an example for this process we concentrate on the frequency range from $8.0\,\mathrm{GHz}$ to $9.8\,\mathrm{GHz}$. Fig. \[figure5\]a shows a zoom into the reflection and transmission spectrum for this frequency range. For our photonic crystal slab of 15 unit cells, we expect from theory to observe 15 longitudinal resonances with a finite wavelength belonging to mode 1. Indeed, inspecting the transmission and reflection spectra, 15 resonances are observed. The first resonance is only clearly visible in the reflection spectrum. Furthermore, the resonances above the tenth one are only visible in the transmission spectrum due to a lower signal-to-noise ratio in the reflected signal. ![(a) Zoom into transmission and reflection spectrum of the photonic crystal slab depicted in Fig. \[figure4\]. The spectra clearly show the different longitudinal modes for transverse mode 1. The labeling depicts the number of anti-nodes, $m$, along the propagation direction. (b) Measured band structure of the photonic crystal slab for the four lowest frequency TE-like eigenmodes (symbols) compared to the calculated values from Fig. \[figure1\]b[]{data-label="figure5"}](figure5.pdf){width="0.98\columnwidth"} Using these considerations we have used eq. (\[eq:wavenumber\]) to assign wave numbers to each measured resonance frequency of mode 1, as plotted in Fig. \[figure5\]b. The agreement with the theoretical dispersion (solid lines) is excellent. To assign also wave numbers to higher order modes we take into account, as was explained above, that mode 2 is not effectively excited with the centered Hertzian dipole antenna. Furthermore, the resonances of mode 4 can only be partly observed. Mode 4 overlaps in frequency with mode 3, but mode 3 couples better to the antenna as can be seen from the higher transmission level of mode 3. Hence, only mode 3 is detected where both modes overlap. With these considerations we can assign normalized wave numbers to mode 3 and mode 4, as well. Also for these modes the agreement with the theoretical dispersion (solid lines) is excellent. In conclusion, the excellent agreement for mode 1 to mode 4 indicates an appropriate description of the fabricated slab with FDTD calculations. Electric field measurements =========================== We measure the longitudinal electric field inside the photonic crystal slab by measuring the frequency shift of an individual longitudinal resonance due to a spherical metal scatterer inside the crystal. By scanning the position of the scatterer we can map the electric field distribution. Note that the measured frequency shift $\Delta\nu$ is referenced to the resonance frequency of the resonator loaded by the nylon string alone. The effect of the nylon string is small as it results in a shift of only $250\,\mathrm{kHz}$. ![(a) Measured frequency shift $\Delta\nu$ induced by placing the spherical scatterer at that location. Shown is the frequency shift $\Delta\nu$ for mode 1 at a longitudinal resonance with $\nu=8.16\,\mathrm{GHz}$ and $k_za_{z,\mathrm{eff}}\,=\frac{1}{15}\pi$. (b) Resulting electric field component $E_z$ from the measured frequency shift shown in (a) versus position. The measured electric field component $E_z$ (symbols) is compared to the calculated electric field (lines).[]{data-label="figure6"}](figure6.pdf){width="0.68\columnwidth"} Figure \[figure6\]a and \[figure7\]a show two examples of the measured frequency shift $\Delta\nu$ for mode 1 due to the scatterer. The frequencies of the two longitudinal resonances are $8.16\,\mathrm{GHz}$ and $8.42\,\mathrm{GHz}$, respectively. The uncertainty of the frequency shift is $250\,\mathrm{kHz}$ due to the frequency resolution of the network analyzer. In both measurements the frequency shift remains always smaller or equal to zero, as we expect from eq.(\[eq:shift\_full\]) for a mode with a dominating $E_z$ field component. Towards the edge of the waveguide the frequency shift $\Delta\nu$ approaches zero. As can be seen from Fig. \[figure2\], the electric field $E_z$ is at these location close to zero and consequently also the frequency shift. The strongest frequency shift of about $4.75\,\mathrm{MHz}$ and $2.25\,\mathrm{MHz}$, respectively, is reached in both examples at about $7.0\,\mathrm{mm}$ from the center. In between two rows of rods the strongest $E_z$ field component along the x-direction is generally not located at the center of the waveguide but located slightly off the center, see Fig. \[figure2\]. To calculate the longitudinal electric field $E_z$ from the measured frequency shift we apply eq. (\[eq:field\]). However, this requires to determine the total energy stored inside the cavity $U$. To retrieve $U$ we use the definition of the quality factor: $$Q=2\pi\nu_0\frac{U}{P_{diss}}.$$ Exactly at resonance the dissipated power per cycle $P_{diss}$ is equal to the input power $P_{in}$ from the network analyzer. To retrieve the experimental Q-values we determine the full width half maximum of each transmission resonance $\nu_{FWHM}$ shown in Fig. \[figure5\]a and use $Q=\nu_0/\nu_{FWHM}$. ![(a) Measured frequency shift $\Delta\nu$ induced by placing the spherical scatterer at that location. Shown is the frequency shift $\Delta\nu$ for mode 1 at a longitudinal resonance with $\nu=8.42\,\mathrm{GHz}$ and $k_za_{z,\mathrm{eff}}\,=\frac{6}{15}\pi$ (b) Resulting electric field component $E_z$ from the measured frequency shift shown in (a) versus position. The measured electric field component $E_z$ (symbols) is compared to the calculated electric field (lines).[]{data-label="figure7"}](figure7.pdf){width="0.68\columnwidth"} Figure \[figure6\]b and \[figure7\]b shows the resulting electric field strength $E_z$ (square dots) determined for the two longitudinal resonances of mode 1 at a frequency of $8.16\,\mathrm{GHz}$ and $8.42\,\mathrm{GHz}$, respectively. For a comparison, the figure displays also the $E_z$ field strength from the FDTD calculations presented in section 2 (black line). Uncertainties for the measured field values are determined by using the Gaussian error propagation law. We assume that the dominating errors are the uncertainty in input power of about 10% and the frequency resolution of 250kHz. All other uncertainties are much smaller and do not significantly contribute to the error bars of the measurement. Figure \[figure6\] shows the results for a longitudinal resonance with a frequency of $8.16\,\mathrm{GHz}$. The overall shape of the electric field shows an excellent agreement to the calculated field strength. Furthermore, also the absolute values agree within the range of the measurement accuracy, although no adjustable parameter is used in the calculations. Figure \[figure7\] shows an example at a higher frequency of $8.42\,\mathrm{GHz}$. The agreement between FDTD calculations and experiment is again very good. This holds both for the shape of the profile and also for the absolute values. Minor deviations are visible at the waveguide center. The small difference between electric field maximum and central electric field value at the center is calculated to be about 7V/mm which is not resolved in the measurements. In addition, the measured absolute field value is slightly lower. We tentatively attribute both effects to the fact that at a higher frequency (shorter wavelength) influences due to non-periodic variations in the photonic crystal become more important. Nevertheless, in both examples a good agreement between experiment and theory is found. The two discussed examples for mapping the electric field component $E_z$ illustrate the capability of the method to map the absolute value of a specific electric field component *inside* a photonic crystal. The good agreement between the measurements and calculations further clearly demonstrates that a field mapping *inside* a photonic crystal is possible. Summary and outlook =================== We have demonstrated for the first time a method for mapping the absolute strength of an electromagnetic field component *inside* a photonic crystal. The method relies on measuring the change in resonance frequency when the photonic crystal is placed inside a resonator and the field inside the photonic crystal is perturbed by a sub-wavelength scatterer. A spherical scatterer is applied to measure the dominating longitudinal electric field $E_z$ in a specific photonic crystal slab. We observe a good agreement between measured and calculated electric field strength $E_z$ without using any adjustable parameters in the calculations. Note that even if all six components would be of comparable strength, such as in an arbitrary photonic crystal or at specific locations, it is still possible to separately measure each field component. This can be achieved by selecting for the scatterer a suitable material, shape and orientation such that only a single field component contributes to the frequency shift [@Maier1949; @Maier1952]. For example, a thin metallic needle would short-circuit and thus probe the electric field along the orientation of the needle while leaving all other field components unaffected. In the future, this method could be applied in the near infrared or optical domain by scaling down the photonic crystal and the bead on a string. To induce a space dependent frequency shift a metallic or dielectric scatterer could be placed on a carbon nanotube acting as a string. To control the nanotube it could be attached to an atomic force microscope tip. Mounting a carbon nanotube on an atomic force microscope has been demonstrated [@Kageshima2002]. Further, atomic force microscopy allows a sufficiently high spatial resolution, demonstrated by recent measurements where an emitter directly mounted to an atomic force microscope has been used to map the emitter lifetime around a single nanorod [@Frimmer2011]. Combining these results it should be possible to measure the shift in resonance frequency in the transmission spectrum to determine the absolute field strength inside the air voids of a photonic crystal at optical frequencies. Acknowledgment ============== This research is supported by the Dutch Technology Foundation STW, applied science division of NWO and the Technology Program of the Ministry of Economic Affairs. The authors further thank ESA-ESTEC for providing part of the RF-equipment. The research is also part of the strategic research orientation Applied Nanophotonics within the MESA+ Research Institute.

--- abstract: | We present photometric observations in $B$ and $V$ as well as spectroscopic observations of the detached, eccentric 6.6-day double-lined eclipsing binary GG Ori, a member of the Orion OB1 association. Absolute dimensions of the components, which are virtually identical, are determined to high accuracy (better than 1% in the masses and better than 2% in the radii) for the purpose of testing various aspects of theoretical modeling. We obtain $M_A = 2.342 \pm 0.016$ M$_{\sun}$ and $R_A = 1.852 \pm 0.025$ R$_{\sun}$ for the primary, and $M_B = 2.338 \pm 0.017$ M$_{\sun}$ and $R_B = 1.830 \pm 0.025$ R$_{\sun}$ for the secondary. The effective temperature of both stars is $9950 \pm 200$ K, corresponding to a spectral type of B9.5. GG Ori is very close to the ZAMS, and comparison with current stellar evolution models gives ages of 65-82 Myr or 7.7 Myr depending on whether the system is considered to be burning hydrogen on the main sequence or still in the final stages of pre-main sequence contraction. Good agreement is found in both scenarios for a composition close to solar. We have detected apsidal motion in the binary at a rate of $\dot\omega = 0.00061 \pm 0.00025$ deg cycle$^{-1}$, corresponding to an apsidal period of $U = 10700 \pm 4500$ yr. A substantial fraction of this ($\sim$70%) is due to the contribution from General Relativity, and our measurement is entirely consistent with theory. The eccentric orbit of GG Ori is well explained by tidal evolution models, but both theory and our measurements of the rotational velocity of the components are as yet inconclusive as to whether the stars are synchronized with the orbital motion. author: - Guillermo Torres - 'Claud H. Sandberg Lacy' - Antonio Claret - 'Jeffrey A. Sabby' title: 'Absolute dimensions of the unevolved B-type eclipsing binary GG Orionis' --- To appear in *The Astronomical Journal*, December 2000 Introduction ============ The discovery of GG Orionis (HD 290842, Tycho 4767 857 1, $V = 10.4-11.1$, , $\alpha = 05^h 43^m 10\fs2$, $\delta = -00\arcdeg 41\arcmin 15\arcsec$, epoch and equinox J2000) as a variable star is due to Hoffmeister (1934), who observed the object photographically at the Sonneberg Observatory. The correct period of 6.631 days was first given by Kordylewski (1951) based on visual and photographic times of minimum. This author obtained a mean visual light curve, and established that the orbit is eccentric from the displacement of the secondary minimum. Aside from the occasional measurement of the times of eclipse by a number of authors, GG Ori has remained until recently a rather neglected system. Double lines in the spectrum were detected by Lacy (1984), who described them as being narrow and of nearly equal strength, but no detailed spectroscopic study has been made to date. The first photoelectric light curves were published by Zakirov (1997), who presented light elements for this well detached binary solved by the method of Lavrov (1993). Based on the fact that the orbit is eccentric, it is expected that the system may present a measurable apsidal motion. This effect is of great interest in the study of detached eclipsing binaries because it provides information on the internal structure of stars that may be compared with predictions from theory. From its spectral type and other known properties, GG Ori was listed by Giménez (1985) as a good candidate for the study of the contribution of General Relativity to the secular displacement of the line of apsides, given that the relativistic effect is expected to be dominant in this particular case. GG Ori is located in the Orion OB1 association (see, e.g., Blaauw 1964; Warren & Hesser 1977), a complex region of star formation that has been the subject of numerous studies to determine the properties of the population of young stars and surrounding gas. The binary is located not far from the Belt of Orion, and therefore there is reason to expect that the system might also be quite young, adding to its interest. In this paper we present new high-quality photoelectric light curves in two passbands, which we analyze together with other published photometry. We also report the results of our intensive spectroscopic monitoring of GG Ori that, combined with the light curves, enable us to derive highly precise absolute dimensions for both components of the system. The stars turn out to be nearly identical in all their properties. Our current knowledge of the internal structure and evolution of stars is such that observed stellar properties determined with errors of even 5% are of little value to constrain the models in any significant way, since the differences between competing calculations or the effects of slight changes in the input physics or atomic constants are below this level (see Andersen 1991, 1998). The masses we obtain here have errors smaller than 1%, and the radii are determined to better than 2%. We use these determinations for a comparison with the predictions of recent stellar evolution models, which suggest a very young age for the system. An analysis of all available eclipse timings along with our radial velocities leads to a small but apparently significant apsidal motion detection. We discuss the importance of the relativistic effect mentioned above compared to classical effects from rotational and tidal distortions. In addition, we examine the predictions from tidal theory regarding spin-orbit synchronization and the circularization of the orbit. Spectroscopic observations and reductions ========================================= GG Ori was observed spectroscopically at the Kitt Peak National Observatory (KPNO) and at the Harvard-Smithsonian Center for Astrophysics (CfA). Observations at KPNO were made with the 2.1m and coudé-feed spectrometer from 1983 December to 1999 March. A total of 15 spectrograms were obtained with a variety of CCD detectors. A spectral resolution of 0.02 nm or 2 pixels was used before May 1998, and 0.03 nm or 3 pixels after that date. The spectral coverage was about 10 nm in the first 12 spectrograms, and 32 nm in the last 3 spectrograms. The central wavelength of those observations was 450 nm. Projected rotational velocities ($v \sin i$) were determined by comparing line widths of the binary components’ unblended features with artificially-broadened features in the spectra of comparison stars with known values of $v \sin i$. These comparison stars were chosen to be of nearly the same spectral type as GG Ori, and were observed with the same instrumental configuration as the binary. The reference stars used were $o$ Peg (HR 8641, , $v \sin i = 10{\ifmmode{~\rm km{\thinspace}s^{-1}}\else ~km{\thinspace}s$^{-1}$\fi}$; Fekel 1998, private communication), 68 Tau (HR 1389, A2 IV-V, $v \sin i = 18{\ifmmode{~\rm km{\thinspace}s^{-1}}\else ~km{\thinspace}s$^{-1}$\fi}$; Hoffleit 1982), and HR 8404 (21 Peg, , $v \sin i = 4{\ifmmode{~\rm km{\thinspace}s^{-1}}\else ~km{\thinspace}s$^{-1}$\fi}$; Fekel 1999). From these comparisons in 10 of the spectrograms, values of $24 \pm 2{\ifmmode{~\rm km{\thinspace}s^{-1}}\else ~km{\thinspace}s$^{-1}$\fi}$ and $23 \pm 2{\ifmmode{~\rm km{\thinspace}s^{-1}}\else ~km{\thinspace}s$^{-1}$\fi}$ were determined for the primary and secondary in GG Ori, respectively, based mainly on the 448.1 nm line. The uncertainties given account for the scatter of the line width measurements as well as the agreement between the results using different standard stars. The component of GG Ori we call here the “primary" (also star “A") is the one eclipsed at phase 0.0 in the light curve (see §4). Formally it is also the more massive one, but only marginally so, as we describe later. Radial velocities of the components were measured by cross-correlation with the [FXCOR]{} task in IRAF[^1], using standard stars (same as above) with known radial velocities as templates. Radial velocities for the low-amplitude spectroscopic binary star $o$ Peg were taken from the orbit of Fekel (1999). The radial velocity of HR 8404 was taken as $+0.2{\ifmmode{~\rm km{\thinspace}s^{-1}}\else ~km{\thinspace}s$^{-1}$\fi}$ based on 12 velocities by Fekel (1999), and for 68 Tau we adopted the value $+39.0{\ifmmode{~\rm km{\thinspace}s^{-1}}\else ~km{\thinspace}s$^{-1}$\fi}$, from the same source. Our measurements for GG Ori from the 11 spectra with relatively unblended lines are given in Table 1. In addition to the radial and rotational velocities, we determined the line strength ratio from 17 line pairs in spectra of good signal-to-noise (S/N) ratio. This may be used as a proxy for the light ratio between the stars. We obtained ${\rm EW}_A/{\rm EW}_B = 1.01 \pm 0.02$, corresponding to the blue region of the spectrum. Because the two components are virtually identical, no correction for the difference in temperature is needed, and the ratio of the equivalent widths, ${\rm EW}_A/{\rm EW}_B$, can be assumed to be identical to the light ratio. Our observations of GG Ori at the CfA were collected mostly with the echelle spectrograph on the 1.5m Tillinghast reflector at the F. L.Whipple Observatory on Mt. Hopkins (Arizona), where the system was monitored from 1996 March to 2000 March. Occasionally we observed with an identical spectrograph on the Multiple Mirror Telescope (also atop Mt. Hopkins, Arizona). A total of 42 spectra were recorded with a photon-counting Reticon detector covering a single echelle order centered at 518.7 nm. The resolution is 0.015 nm ($\lambda/\Delta\lambda \sim 35,\!000$), and the spectra span about 4.5 nm. The S/N ratios range from about 25 to 40 per resolution element. The zero point of the velocity system was monitored by means of nightly exposures of the dusk and dawn sky, in the manner described by Latham (1992). The accuracy of the CfA velocity system, which is within about 0.1[ kms$^{-1}$]{} of the reference frame defined by minor planets in the solar system, is documented in the previous citation and also by Stefanik et al. (1999). Radial velocities were determined using the CfA implementation of TODCOR (Zucker & Mazeh 1994), a two-dimensional cross-correlation algorithm that uses two templates, one for each component of the binary. The templates were selected from a large library of synthetic spectra based on the latest model atmospheres by R. L. Kurucz, computed for us specifically for the wavelength of our observations (Morse & Kurucz 1998, private communication). The instrumental profile is explicitly included by applying a Gaussian convolution with a FWHM of 8.5[ kms$^{-1}$]{}, corresponding to our spectral resolution. These synthetic spectra have been calculated over a wide range of effective temperatures, projected rotational velocities, metallicities, and surface gravities. Initially we used templates with a temperature of $T_{\rm eff} = 10000$ K for both stars, surface gravity of $\log g = 4.5$ (cgs), and solar metallicity. The projected rotational velocities of the components were determined by running extensive grids of correlations against templates covering a range of values of $v \sin i$, seeking the best match to the observed spectra as determined by the correlation coefficient averaged over all exposures. We obtained $v \sin i = 16 \pm 1$[ kms$^{-1}$]{} for both components, which is significantly different than our determination based on KPNO spectra. We discuss this further in §5 and §6. For the final templates we adopted a temperature of 9950 K for both stars and $\log g = 4.3$, based on the results from the light curve solutions and spectroscopic orbits presented below. Systematic errors in the radial velocities resulting from line blending are always a concern, particularly when the goal is to achieve the highest possible precision and accuracy in the mass determinations. Although the use of a two-dimensional correlation technique such as TODCOR in principle minimizes those errors (see Latham et al. 1996), experience has shown us that this must be checked on a case-by-case basis, particularly in view of the narrow spectral window of the CfA observations (e.g., Torres et al.1997; 2000). For this we generated artificial composite spectra by adding together synthetic spectra for the two components with Doppler shifts appropriate for each actual time of observation, computed from a preliminary orbital solution. We adopted also a light ratio close to that for the real stars (see below). We then processed these simulated spectra with TODCOR in the same manner as the real spectra, and compared the input and output velocities. The differences derived in this way vary systematically with phase (or radial velocity), as expected, and were applied to the real velocities as corrections, even though in the case of GG Ori they are quite small ($\leq 0.5$[ kms$^{-1}$]{}) and hardly affect the results. The final radial velocities from the CfA spectra with the corrections included are given in Table 2. The light ratio was derived from these observations using TODCOR as described by Zucker & Mazeh (1994). Because of the relatively small number of lines in the 4.5 nm spectral window observed at CfA, significant errors can be introduced in the light ratio when features of each component shift in and out of this region by up to 0.2 nm due to orbital motion. To estimate and correct for this effect we followed a procedure analogous to that described above for the velocities, and found the magnitude of the systematic error to be about 3.5%. After correcting for this we obtain $L_A/L_B = 1.05 \pm 0.03$ at a mean wavelength of 518.7 nm, which is sufficiently close to the visual band that we will assume it is the light ratio in $V$, given that the stars are virtually identical. Preliminary double-lined orbital solutions based on our radial velocities were computed separately for the KPNO and CfA measurements, and are compared in Table 3. Initially we adopted the eclipse ephemeris given in the General Catalog of Variable Stars (GCVS) (Kholopov 1985): Min I (HJD) $ = 2,\!433,\!596.496 + 6.63147 \cdot E$. However, the residuals from the orbit for the CfA data showed a systematic pattern and the scatter was significantly larger than expected from previous experience with similar spectroscopic material. This suggested that the ephemeris is not accurate enough to be propagated forward by $\sim50$ yr (2600 cycles) to the epoch of our velocity measurements. Therefore, a new linear ephemeris computed from times of minimum collected from the literature was derived (see next section) and used here. The possibility of a systematic difference between the center-of-mass velocities of the primary and secondary components was explored by solving for an offset simultaneously with the orbital elements. This was done independently for the two data sets. No shift was expected because the stars are so similar. The results in the sense $\langle$primary minus secondary$\rangle$ were $+1.42 \pm 0.97{\ifmmode{~\rm km{\thinspace}s^{-1}}\else ~km{\thinspace}s$^{-1}$\fi}$ for KPNO and $-0.74 \pm 0.53{\ifmmode{~\rm km{\thinspace}s^{-1}}\else ~km{\thinspace}s$^{-1}$\fi}$ for CfA. Since these are indeed not significantly different from zero, the velocities for the two components can be considered for all practical purposes to be on the same reference frame. As seen in Table 3, there are slight differences in some of the elements of these two orbital solutions, in particular in the velocity amplitude of the secondary component ($K_B$), but we do not consider them to be significant in view of the small number of KPNO observations (only 11). Therefore, for the final solution discussed below we have merged the two data sets. Apsidal motion analysis and final spectroscopic solution ======================================================== As indicated earlier, the noticeable eccentricity of the orbit of GG Ori suggests the possibility of apsidal motion, even though none has been reported for this system. In extreme examples this can affect the spectroscopic solutions quite significantly, as in the case of V477 Cyg (Popper 1968). Before computing the final orbital fit, we therefore searched the literature for eclipse timings, both to improve the ephemeris and to investigate the apsidal motion. Table 4 lists all times of minimum available to us, of which there are 25 primary minima and 32 secondary minima covering about 65 yr (1930-1995). The majority were determined by visual or photographic means, and only the 6 recent photoelectric timings and one visual timing have published uncertainties. For the rest we determined the errors iteratively, based on the mean residuals for each type of observation from preliminary fits: $\sigma_{\rm pg} = 0.038$ days for the photographic minima, and $\sigma_{\rm v} = 0.028$ days for the visual minima. One visual estimate was found to give a large residual (0.14 days, $\sim5\sigma$), and was excluded. It is indicated in parentheses in Table 4. A linear ephemeris fit to these data gives Min I (HJD) $ = 2,\!449,\!717.66253(21) + 6.6314948(16) \cdot E$, with a phase difference between the primary and secondary minima of $\Delta\Phi = 0.42252 \pm 0.00024$. This is the ephemeris we used above in our preliminary orbital solutions, and the period is very similar to that reported by Zakirov (1997) based on a smaller number of observations. Although no unusual trends were seen in the timing residuals, we investigated the possibility of apsidal motion by computing the apparent periods separately for the primary and the secondary. We obtained $P_I = 6.6314971 \pm 0.0000021$ days and $P_{II} = 6.6314890 \pm 0.0000030$ days, which are different at the 2.2$\sigma$ level. Next we submitted these same data to an apsidal motion analysis using the method by Lacy (1992b). For this we assumed a fixed eccentricity equal to the average of the values derived in columns (2) and (3) of Table 3, and a fixed inclination angle of $89\arcdeg$ from preliminary light curve solutions. The result is only marginally significant ($\dot\omega = 0.00050 \pm 0.00027$ deg cycle$^{-1}$), largely due to the weak constraint provided by the early photographic and visual times of minimum, which have relatively low weight. Changes in those weights have little effect. Our radial velocities extend the time base provided by the times of minimum by another 5 yr, and contain valuable information on the longitude of periastron that can also be used. In addition, they provide a much stronger constraint on the eccentricity, which must otherwise be fixed when using only eclipse timings. The optimal solution, therefore, is to combine both kinds of measurements into a single least squares fit, solving simultaneously for the spectroscopic orbital elements and the apsidal motion. In this way the information on the period contained in the times of minimum is implicitly taken into account for the spectroscopic elements, rather than having to adopt a fixed ephemeris as we have done in §2. At the same time, the effect of the possible rotation of the line of apsides on the spectroscopic elements is also accounted for. In addition, we have allowed for an arbitrary shift between the KPNO and CfA velocity systems in view of the difference in the way the two zero points were established. The offset turns out to be negligible: $\langle$CfA$-$KPNO$\rangle$ $= -0.06 \pm 0.39{\ifmmode{~\rm km{\thinspace}s^{-1}}\else ~km{\thinspace}s$^{-1}$\fi}$. The results of this simultaneous fit are given in column (4) of Table 3. The relative weights of the primary and secondary velocities in each data set, as well as the weights of the times of minimum, have been iterated until reaching convergence for a reduced $\chi^2$ of unity. Tests show once again that the weights assigned to the visual and photographic timings, or to the velocities, have little effect on the results. The velocity residuals from this solution are listed separately in Table 1 and Table 2, and the timing residuals are given in Table 4. The final spectroscopic orbital solution is represented graphically in Fig. 1, along with the observations. The apsidal motion resulting from the final fit is $\dot\omega = 0.00061 \pm 0.00025$ deg cycle$^{-1}$, which is significant at the $2.5\sigma$ level, and differs from our preliminary estimate above by less than half of its uncertainty. Also listed in Table 4 are the sidereal and anomalistic periods ($P_s$ and $P_a$), as well as the apsidal motion period $U = 10700 \pm 4500$ yr. Periastron passage occurs at a photometric phase of 0.06. A plot of the $O\!-\!C$ deviations of the times of minimum from the linear terms of the apsidal motion is given in Fig. 2 along with the predicted deviations. The top panel draws attention to the fact that the observations cover less than 1% of a full cycle, while the bottom panel expands the region around the observations, with the coverage provided by the radial velocity measurements being indicated in the upper right. Photometric observations and light curve solutions ================================================== Differential light curves were obtained by CHSL at the Cerro Tololo Inter-American Observatory (CTIO) with the 24-inch Lowell telescope in the Johnson $B$ and $V$ bands during 1993-1995. Absolute indices were also measured separately based on measurements that were carefully tied to the standard $\ubv$ system through observations of secondary standards from Landolt (1973), typically 30-40 standards per night. The absolute photometry was made at Mount Laguna Observatory near San Diego (California) in the fall season of 1989, and at CTIO in 1988-1990 and 1993-1995. The procedures that were used are described in detail by Lacy (1992a). The $\ubv$ indices on the standard system are shown in Table 5, and the 257 differential $BV$ observations are given in Table 6 and Table 7. The differential magnitudes are in the sense $\langle$variable$-$comparison$\rangle$ and are referred to BD $-1\arcdeg$1013 (HD 38165, ). The comparison stars (see Table 5) were found to be constant at a level of about 0.008 mag for the standard deviation of the differences between comparison stars. The precision of the CTIO differential observations, based on previous results, is estimated to be about 0.006 mag in both $B$ and $V$ at the magnitudes of the program stars. In addition to the absolute photometric indices mentioned above, which we have computed separately for the two observing intervals, $uvby$ indices are available from Hilditch & Hill (1975). The results from the three different sources are somewhat inconsistent. The $V$ magnitude is slightly brighter by about 2.5% in the 1988-1990 results compared to the 1993-1995 results and the value of Hilditch & Hill (1975), which is $V = 10.380 \pm 0.015$. The $\bv$ color index during 1993-1995 is considerably bluer than the value measured in 1988-1990. These discrepancies cannot be explained by differences in sky conditions (which were always photometric), nor by differences in aperture size, centering errors, or seasonal variations. Furthermore, the scatter from the light curve fits described later in this section is significantly larger than expected based on our previous results with the equipment used at CTIO at these magnitude levels. The standard deviation of the residuals from the fitted photometric orbit of FS Mon (Lacy et al. 2000), for example, was 0.006 and 0.005 mag in $B$ and $V$, respectively, whereas the corresponding figures for GG Ori (see below for details) are 0.014 and 0.009 mag. Both binaries were observed during the same runs. Intrinsic variability of one or both of the components of GG Ori is therefore suspected. Differential light curves in the Johnson $U$, $B$, $V$, and $R$ bands have also been obtained in Uzbekistan by the group led by M. Zakirov. They used 0.6m telescopes at the Maidanak Observatory, in the southern part of Uzbekistan. Their differential magnitudes in the sense $\langle$variable$-$comparison$\rangle$ relative to the star BD $-1\arcdeg$1013 have been described by Zakirov (1997). The transformation of these measurements to the standard system is somewhat less secure than for our own observations, but there are marginal indications that the $\bv$ index of GG Ori became bluer by a few hundredths of a magnitude in the interval 1993-1994, toward the end of their observations. Though far from being conclusive evidence of intrinsic variability, we note that this is in the same direction as the trend suggested in Table 5. We have re-analyzed the differential observations by Zakirov with the same methods we use for our own measurements. The light curve fitting was done with the NDE model as implemented in the code EBOP (Etzel 1981; Popper & Etzel 1981), and the ephemeris adopted is that of §3. The main adjustable parameters are the ratio of the central surface brightness of the secondary star ($J_B$) in units of that of the primary, the relative radius of the primary ($r_A$) in units of the separation, the ratio of the radii ($k \equiv r_B/r_A$), the inclination of the orbit ($i$), and the geometric factors $e\cos\omega$ and $e\sin\omega$ which account for the orbital eccentricity. As usual, we have allowed also for a photometric scale factor and a phase shift in all solutions. Auxiliary quantities needed in the analysis include the gravity-brightening coefficients, for which we adopt the values 1.00, 0.85, 0.70, and 0.59 in $U$, $B$, $V$, and $R$, respectively, from Martynov (1973). For the reflection coefficients we adopted the value 1.0, as appropriate for stars with radiative envelopes. Limb-darkening coefficients ($u$) were included as one of the variables to be fitted, although in one case (noted below) it was fixed to an appropriate value from the tables of Wade and Rucinski (1985). The mass ratio ($q\equiv M_B/M_A = 0.9982$) was adopted from the spectroscopic analysis in §3. Preliminary solutions revealed a few outliers with residuals greater than $5\sigma$ (one of the CTIO measurements in $B$, two in $V$, and one of Zakirov’s measurements in each of his four bands), which were given zero weight in subsequent iterations. The fitting procedure converged in the general solutions of the CTIO light curves with all variables adjusted simultaneously (see columns 2 and 3 in Table 8). As is often the case in the analysis of partially eclipsing binary stars with nearly equal components, the fitted values of the parameter $k$, the ratio of radii, are rather uncertain, but near unity. Attempts at general solutions of the Zakirov light curves failed to converge. In cases such as this light ratios estimated from spectrograms are very important for constraining the photometric solutions. For GG Ori we have at our disposal two spectroscopic determinations of the light ratio, one approximately in the $B$ band (from KPNO) and the other close to the $V$ band (from CfA). A grid of EBOP solutions was run on the CTIO light curves for fixed values of $k$ between 0.94 and 1.10, in steps of 0.01. Within this range of about 15%, the rms residuals from the fits were found to change very little as a function of $k$ ($\leq 0.00005$ mag in $B$ and $\leq 0.00006$ mag in $V$), supporting our concerns about the indeterminacy of $k$ from the light curves alone. The sum of the relative radii changes by less than 0.7% over the entire range of $k$ values. The light ratios in $B$ and $V$ were computed from each of these solutions, and then by interpolation we determined the $k$ values that correspond to our measured spectroscopic light ratios: $k = 0.996 \pm 0.011$ from the KPNO spectra, and $k = 0.972 \pm 0.015$ from the CfA spectra (see Fig. 3). The value we adopt is the weighted average, $\langle k\rangle = 0.988 \pm 0.009$. For comparison, the average ratio of the radii derived from the CTIO light curve solutions without the constraint from spectroscopy is $k = 1.03 \pm 0.04$ (see Table 8), which is different by $1\sigma$. With the ratio of the star sizes fixed to the weighted average given above, we repeated the light curve fits separately for the CTIO and Zakirov (1997) photometry, and the adopted solutions are given in Table 8 and Table 9. Attempts to include third light as a parameter showed that, within uncertainties of about 1%, it was not significantly different from zero. The CTIO observations in $B$ and $V$ and the best fit models are shown in Fig. 4, and the data by Zakirov (1997) in $\ubvr$ along with the corresponding fitted light curves are shown in Fig. 5. Table 10 gives the weighted averages of the light elements in each band obtained from the two data sets. In addition to $r_A$ and $k$, we list also the sum of the relative radii, $r = r_A + r_B$, and the radius of the secondary, $r_B$. The errors for these quantities were derived as described in the Appendix. There is good agreement between the photometric determinations of $e \cos\omega$ and $e \sin\omega$ and the corresponding spectroscopic values, which are $-0.1200 \pm 0.0017$ and $+0.1865 \pm 0.0020$, respectively. The results in Table 10 show that the components of GG Ori are nearly indistinguishable in size and of virtually the same luminosity, which is consistent with the fact that they are also very similar in mass (§3). The departure from the spherical shape is insignificant. Both eclipses are partial, with approximately 91% of the light of the primary blocked at phase 0.0 (which is a transit), and 90% of the light of the secondary eclipsed at the other minimum. The earlier study by Zakirov (1997) reported the primary eclipse to be total (larger secondary star blocking the primary), although the ratio of the radii was much closer to 1.0. The uncertainties assigned to the adjusted quantities in Table 10 are generally the internal errors produced by EBOP. As is well known, the formal uncertainties in least-squares solutions in which one or more quantities have been held fixed are typically too small because correlations between the elements are artificially eliminated (as we have done by fixing $k$, for example). The quantities $k$, $r_A$, and $i$ are the most important for determining the absolute dimensions of the components. The error in $k$ is based on the uncertainty in the spectroscopic light ratios, and therefore does not suffer from this shortcoming. For $r_A$ and $i$ we have increased the formal uncertainties to account for all contributions to the error. The rms deviations given in Table 8 for the CTIO observations are somewhat larger than expected, based on previous experience, and suggest that one or both components of GG Ori may be intrinsically variable. We have searched for patterns in the residuals by computing the power spectra in the $B$ and $V$ bands, but no significant periodicities were found. Absolute dimensions =================== The combination of the spectroscopic results in Table 3 and the light curve results in Table 10 leads to the absolute masses and radii for GG Ori, shown in Table 11. The masses are determined with a precision of 0.7%, and the radii are good to about 1.4%. The fact that the surface brightness ratio $J_B$ is near unity over a wide range of wavelengths indicates that the effective temperatures of the stars are essentially the same. From $J_B$ in the $V$ band the difference in visual surface brightness is $\Delta F^{\prime}_V = 0.0002 \pm 0.0004$ (Popper 1980), which translates into a completely negligible color difference $\Delta(\bv) = 0.0005 \pm 0.0009$, and a formal temperature difference of only 8 K. The discrepancies in the absolute photometry at different epochs pointed out earlier in §4 complicate the derivation of the effective temperature of the components. Because the light variations may be real, we have chosen to adopted the straight average of the largest and smallest $\ubv$ indices for GG Ori in Table 5: $V = 10.372 \pm 0.013$, $\bv = 0.511 \pm 0.021$, $\ub = 0.324 \pm 0.004$. The errors for the first two of these quantities are simply half of the difference between the 1988-1990 and 1993-1995 measurements. The reddening was derived using the reddening-free index $Q = (\ub) - 0.72 (\bv)$ (Johnson & Morgan 1953) and the calibration by Deutschman, Davis & Schild (1976). We obtained $(\bv)_0 = -0.036 \pm 0.006$, $(\ub)_0 = -0.070 \pm 0.020$, $E(\bv) = 0.547 \pm 0.022$, and $A_V = 1.696 \pm 0.068$ (using $A_V = 3.1\cdot E(\bv)$). The formal errors do not account for uncertainties in the calibrations themselves, which are difficult to quantify. These color indices correspond to the combined light, but since the components are virtually identical they are also the indices of the individual stars. An independent measure of the reddening may be obtained from the $uvby$ photometry by Hilditch & Hill (1975): $\by = 0.408 \pm 0.015$, $c_1 = 1.002 \pm 0.030$, $m_1 = +0.016 \pm 0.020$. The calibration by Crawford (1978) for B-type stars leads to $(\by)_0 = -0.030$ and $E(\by) = 0.438$, from which $E(\bv) = 0.592$. This is slightly larger than the reddening derived from the $\ubv$ photometry. However, GG Ori is near the end of the range of validity of the $uvby$ calibrations, where they become rather uncertain. Also, we note that the $(\by)_0$ index corresponds to a spectral type of B8.5 according to Table 1 by Popper (1980). But the fact that we see no sign of the 447.1 nm line in our KPNO spectra of the object strongly suggests that the spectral type cannot be earlier than B9.5. We have therefore chosen to rely only on the Johnson photometry above. The de-reddened $\bv$ index corresponds to a temperature of $9950 \pm 200$ K, and a spectral type of B9.5 (Popper 1980). Further properties of the stars are listed in Table 11, including the mean density ($\bar\rho$), the absolute visual magnitudes, and the distance. The latter two are based on the visual surface brightness parameter $F^{\prime}_V = 3.974 \pm 0.003$ derived from the intrinsic color $(\bv)_0$ and the tabulation by Popper (1980), and are thus independent of bolometric corrections. The distance modulus of GG Ori, $m\!-\!M = 8.20 \pm 0.10$ ($d = 438 \pm 20$ pc), is similar to other determinations for various subregions of the Orion OB1 complex (Warren & Hesser 1977; Brown, de Geus, & de Zeeuw 1994; Brown, Walter & Blaauw 1999; de Zeeuw et al. 1999), and lends support to its association. Also listed in Table 11 are the projected rotational velocities expected if the axial rotations were synchronized with the mean orbital motion of the binary and with the orbital motion at periastron (pseudo-synchronization). The two independent measurements of $v \sin i$ in §2 based on our KPNO and CfA spectra disagree, with the KPNO results being closer to the pseudo-synchronous values, while the CfA determinations suggest synchronization with the mean orbital motion. We discuss this further in §6 in the context of tidal evolution theory. No metallicity determination is available for GG Ori. A rough estimate can be obtained from our CfA spectra following a procedure analogous to that used to derive $v \sin i$, and the result is \[m/H\] $= -0.15 \pm 0.20$, consistent with the solar abundance. Discussion ========== In this section we compare the properties of GG Ori as listed in Table 11 with predictions from models. The three aspects of theory we focus on are stellar evolution, internal structure, and tidal evolution. For each of these we require calculations of the basic properties of the components, which we have taken from evolutionary tracks computed specifically for the exact masses we determined for the stars. The evolutionary code we used is that by Claret (1995). Further details on the input physics are described by Claret & Giménez (1992). Convection in these models is treated with the standard mixing-length prescription, with a fixed mixing-length parameter of 1.52$H_p$ that gives the best fit between a solar model and the observed properties of the Sun. A moderate amount of core overshooting is assumed ($\alpha_{\rm ov} = 0.20 H_p$), although in the case of GG Ori this effect is insignificant due to the unevolved nature of the system. Stellar evolution ----------------- Consistent with the young environment surrounding the object (the Orion OB1 association), we find that GG Ori is indeed very close to the Zero Age Main Sequence (ZAMS). This is illustrated in Fig. 6, which shows a greatly enlarged section of the evolutionary tracks as they approach the main sequence. The full extent of the main sequence band is shown in the inset. The binary is so close to the ZAMS that, given the uncertainties, it is difficult to determine whether the stars are already burning hydrogen in their cores, or whether they are still in the final stage of contraction towards the main sequence. The error bars shown reflect the uncertainties in the effective temperatures of the components and their surface gravities. The error in the placement of the tracks due to the uncertainty in the measured masses is indicated in the lower left corner. The models plotted in Fig. 6 correspond to the case where the system is already on the main sequence (MS). Tracks for two different chemical compositions ($Z$, $Y$) are shown, giving metallicities very close to solar (\[m/H\] $= +0.03$, $dY/dZ = 1.65$, solid lines; and \[m/H\] $= +0.05$, $dY/dZ = 1.95$, dashed lines) consistent with the indication in §5. The enrichment laws ($dY/dZ$) implied by the best fit values of the helium abundance ($Y$) are within the range found in other determinations (see, e.g., Peimbert 1993; Renzini 1994; Pagel & Portinari 1998; Izotov & Thuan 1998), and are also similar to the values favored by other eclipsing binaries (Ribas et al. 2000), using the same set of evolutionary models. Our attempts to find evolutionary tracks that fit the location of GG Ori assuming it is nearing the end of the pre-main sequence (PMS) phase also gave acceptable results (see Fig. 7), considering the error bars, although the agreement is formally not as good as the main-sequence case. The PMS tracks shown in the figure are for \[m/H\] $= +0.07$ and $dY/dZ = 2.09$. The evolutionary ages derived for the components are quite different in the two scenarios (see Table 12). In the main sequence case the average age for the two stars is 82 Myr (for $Z = 0.020$, $Y = 0.2729$) or 65 Myr (for $Z = 0.021$, $Y = 0.2810$), and the difference in age between the components is similar to the formal uncertainty, which is derived from the error in $\log g$. On the other hand, if GG Ori is on the PMS the models give an average age of only 7.7 Myr. Few cases are known of eclipsing binaries with well-determined absolute dimensions in which at least one component is a bona-fide PMS star: EK Cep (Popper 1987; Claret, Giménez, & Martín 1995), TY CrA (Casey et al.1998), and possibly RS Cha (Clausen & Nördstrom 1980; Pols et al. 1997; Mamajek, Lawson, & Feigelson 2000). Such systems are particularly valuable to test models of PMS evolution, where theory remains essentially unchallenged by observations so far. The stars in GG Ori are so similar and so close to the ZAMS, however, that the constraint on PMS evolution is not very strong. Unevolved systems such as this are much more useful for testing opacity and metallicity effects in the models when coupled with an accurate spectroscopic determination of the metal abundance, which unfortunately is not yet available for GG Ori. Whether or not the system has already settled on the main sequence, it is undoubtedly young. In this connection we note, incidentally, that light variations such as those hinted at in §4 are not entirely unexpected. It is a remarkable coincidence that among the eclipsing systems with the best known absolute dimensions, no fewer than *six* have at least one component with virtually the same mass as the stars in GG Ori, to within 1%. These systems are V451 Oph, YZ Cas, $\beta$ Aur, WX Cep, SZ Cen (Andersen 1991), and V364 Lac (Torres et al. 1999). Thus, a total of eight stars (including GG Ori) can in principle be compared with *a single evolutionary track*, and they span the entire main sequence band from the ZAMS (GG Ori) to the shell hydrogen burning phase (SZ Cen, age $\sim$800 Myr). In Fig. 8 we show this comparison against one of the models used above for GG Ori ($Z = 0.020$, $Y = 0.2729$), for a composition close to solar. The agreement with theory is very good, indicating that all these systems are well represented by a single metallicity. Internal structure and General Relativity ----------------------------------------- Our detection of the apsidal motion of the binary in §3 provides the opportunity to test models of the internal structure of stars. In addition, the relatively short orbital period and large masses of the components along with the fact that the orbit is eccentric led Giménez (1985) to propose GG Ori as a good candidate for the study of the general-relativistic (GR) contribution to the apsidal motion. At the current level of the uncertainties in the measurement of apsidal motion in binaries, the theoretical GR contribution is separable from the classical (Newtonian) terms due to the gravitational quadrupole moment induced by rotation and tides. The total apsidal motion can then be expressed as $\dot\omega_{tot} = \dot\omega_N + \dot\omega_{GR}$. The two effects cannot be measured separately, though, so we have chosen here to compare the total theoretical value with the measured quantity. In the case of GG Ori, the GR term is predicted to be about 2.5 times larger than the classical term, contributing $\sim$70% to $\dot\omega_{\rm tot}$. Following Giménez (1985), we obtain $\dot\omega_{GR} = 0.000454 \pm 0.000002$ deg cycle$^{-1}$. To compute $\dot\omega_N$ we have used the internal structure constants from theory (including the higher order terms $k_3$ and $k_4$), which are calculated at each point along the evolutionary track of each component. The contribution of the rotational distortions to the Newtonian apsidal motion rate $\dot\omega_N$ depends on the ratio between rotational angular velocity and the Keplerian angular velocity. This ratio is often derived under the assumption that the stars are already synchronized at periastron. For a system as young as GG Ori this may not necessarily be the case, and in fact the observational evidence we have is conflicting (§5; see also §6.3). We have therefore relied on our own measurements of the projected rotational velocities, $v \sin i$, even though our two independent determinations from KPNO and CfA spectra are somewhat different (§2). The KPNO values ($v_A \sin i = 24 \pm 2{\ifmmode{~\rm km{\thinspace}s^{-1}}\else ~km{\thinspace}s$^{-1}$\fi}$, $v_B \sin i = 23 \pm 2{\ifmmode{~\rm km{\thinspace}s^{-1}}\else ~km{\thinspace}s$^{-1}$\fi}$) lead to $\dot\omega_N = 0.000209 \pm 0.000012$ deg cycle$^{-1}$, while the CfA measurements ($v \sin i = 16 \pm 1{\ifmmode{~\rm km{\thinspace}s^{-1}}\else ~km{\thinspace}s$^{-1}$\fi}$ for both components) give $\dot\omega_N = 0.000183 \pm 0.000009$ deg cycle$^{-1}$. When adding this to the GR term, we obtain $\dot\omega_{tot} = 0.000663 \pm 0.000012$ deg cycle$^{-1}$ (KPNO) and $\dot\omega_{tot} = 0.000637 \pm 0.000010$ deg cycle$^{-1}$ (CfA), in which the errors account for all contributions from measured quantities. These predictions are to be compared with the observed value, $\dot\omega_{obs} = 0.00061 \pm 0.00025$ deg cycle$^{-1}$ (§3). There is good agreement with theory, although the uncertainty in the observed value is large enough that the constraint on the models is very weak. Further measurements of times of minimum are required to improve $\dot\omega_{obs}$. Tidal evolution --------------- Tidal forces within a binary tend to synchronize the rotation of each component with the orbital motion, and to make the orbit circular. The two main mechanisms proposed to describe these effects (Tassoul & Tassoul 1997, and references therein; Zahn 1992, and references therein) make somewhat different predictions for the timescales of these processes. In this section we compare our observations with both, although the hydrodynamical mechanism by Tassoul has often been found to be too efficient (see, e.g., Claret, Giménez, & Cunha 1995). Synchronization times and circularization times have been computed as described in the latter reference and also by Claret & Cunha (1997), by integrating the differential equations describing the evolution of the rotation and eccentricity. All time-dependent properties of the stars were interpolated directly from the evolutionary tracks for each component. In performing these calculations for young stars such as GG Ori a number of complications arise regarding the pre-main sequence phase where the radii of the stars are much larger than on the main sequence. One of them is that in principle the orbital period (or semimajor axis) is also changing during this phase. For example, if the period of GG Ori is assumed to be fixed at the current value of 6.6 days, we find that the size of the orbit is too small to accommodate the very large radii predicted for the stars during the early stages of contraction along the Hayashi tracks. Thus the equations for the orbital evolution (period and eccentricity) and rotational evolution are coupled and must be integrated simultaneously (see, e.g., Duquennoy, Mayor, & Mermilliod 1992). Zahn & Bouchet (1989) have done this for late type stars (0.5 $M_{\sun} \leq M \leq 1.25~M_{\sun}$) and concluded that all binaries with periods up to 7 or 8 days arrive on the main sequence with their orbits already circularized, and that further evolution of the eccentricity on the main sequence is negligible. However, eccentric systems on the main sequence with periods shorter than this *are* actually observed (see, e.g., Mathieu et al. 1992). For earlier-type binaries such as GG Ori the efficiency of the tidal mechanisms may be different, and a detailed investigation of the evolution in the PMS phase is beyond the scope of this paper. Tidal theory is still largely in development, and many aspects of it remain somewhat controversial. Nevertheless, useful comparisons with the observations are still possible under certain simplifying assumptions, and have indeed been made in a number of cases. Verbunt & Phinney (1995), Claret, Giménez, & Cunha (1995), Claret & Cunha (1997), and others, have compared the predictions of the main mechanisms mentioned above with the properties of detached eclipsing systems with well-determined absolute dimensions. Their calculations were done in detail (i.e., by *integrating* the corresponding differential equations rather than simply using the *timescales* for syncronization and circularization), but avoided the PMS problem by excluding it altogether (Verbunt & Phinney 1995) or by including only the final stages of contraction, and assumed also that the orbital period does not change significantly, which may be a valid approximation when the stars are already close to the ZAMS. We have followed a similar prescription here and considered only the final loop in the evolutionary tracks preceding the ZAMS (see Fig. 8). We start our integrations at the onset of the nuclear reaction $^{12}{\rm C}(p,\gamma)~^{13}{\rm N}(\beta+\nu)~^{13}{\rm C}(p,\gamma)~^{14}{\rm N}$. At this stage this is not the only source of energy, since contraction is still taking place. The synchronization and circularization times ($t_{sync}$, $t_{circ}$) are not strongly dependent on the chemical composition in the range we have considered above. Using the formalism by Tassoul we obtain $\log t_{sync} = 6.875$ for both components of GG Ori. The predictions from the mechanism by Zahn, on the other hand, give $\log t_{sync}^A = 8.888$ and $\log t_{sync}^B = 8.890$, which are significantly longer. The circularization times are nearly identical in both theories: $t_{circ} = 8.888$ (Tassoul) and $t_{circ} = 8.889$ (Zahn). These values are to be compared with the mean age of the system counted from the starting point of the integrations, which differs by 5.4 Myr from the true ages given in §6.1. The modified age is then $\log t = 7.884$ ($Z = 0.020$, $Y = 0.2729$) or $\log t = 7.775$ ($Z = 0.021$, $Y = 0.2810$), assuming GG Ori is already burning hydrogen on the main sequence. We illustrate this in a slightly different way in Fig. 9, where we focus on $\log g$ as a sensitive measure of stellar evolution. The value of the surface gravity at which synchronization or circularization is achieved ($\log g_{\rm crit}$) is shown as a function of orbital period for the two main tidal theories. Stars evolve upwards in this diagram. The measured values of $\log g$ for the primary and secondary of GG Ori are represented by crosses, with the error bars being of the same size as the symbols. According to the theory by Zahn the stars should not yet have reached synchronous rotation, which agrees with the measured $v \sin i$ values from our CfA spectra. Also, the orbit should not yet be circular, as is indeed observed. The theory by Tassoul also predicts that the orbit should still be eccentric, but indicates that the stars should already be rotating synchronously with the orbital motion, which appears to be more consistent with the $v \sin i$ values from our KPNO spectra. Although as mentioned above the mechanism by Tassoul is often found to be too efficient, a more accurate measurement of the projected rotational velocities in GG Ori could shed more light on this issue. Conclusions =========== New photometric and spectroscopic observations of the eccentric binary GG Ori combined with a reanalysis of data from the literature have allowed us to derive definitive orbital parameters and physical properties of the components. Our determinations have formal errors smaller than 1% in the masses and smaller than 2% in the radii. GG Ori thus joins the elite of stars with the best established absolute dimensions. The system is very young, and in fact it is so close to the ZAMS that within the observational errors the comparison with current stellar evolution models is unable to tell us whether it has already settled on the main sequence (age $\sim 65$–82 Myr, depending on the metallicity), or whether it is still at the end of the contraction phase (age $=7.7$ Myr). We find good agreement with both scenarios for a chemical composition close to solar. The tentative detection of apsidal motion in GG Ori with a period of $U = 10700$ yr has allowed us to make an initial test of interior structure models. This system is particularly interesting in that the contribution from General Relativity should be substantial ($\sim$70% of the total apsidal motion). Although the agreement with theory is good, further observations to improve the error in $\dot\omega$ are necessary for a more stringent test. Current mechanisms that describe the tidal evolution of binary properties disagree as to whether the rotation of the components of GG Ori should be synchronized with the orbital motion. Unfortunately our own measurements of $v \sin i$ have not settled the issue. Although this area of theory is perhaps the weakest of the comparisons we have discussed, the models do explain the significant eccentricity of the orbit system at the relatively short period of 6.6 days. We thank P. Berlind, M. Calkins, D. W. Latham, A. Milone, and R.P. Stefanik, who obtained many of the spectroscopic observations used here, and R. Davis, who maintains the CfA database of radial velocities. We are also grateful to Daryl Willmarth at KPNO for assistance with the spectroscopic observations there. Paul Etzel provided helpful comments on aspects of the light curve solutions. We thank the referee for useful suggestions. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France, and of NASA’s Astrophysics Data System Bibliographic Services. Appendix ======== The determination of the error in the relative radius of the secondary component, $r_B$, in eclipsing binaries has often been a source of confusion among users of EBOP. Results are occasionally published in which that uncertainty is derived by propagating errors in the expression $r_B = k\cdot r_A$, where $k$ (defined as $r_B/r_A$), the primary radius $r_A$, and their associated uncertainties are usually adopted directly from the output of the program. Aside from the fact that the formal errors in the latter two quantities are typically underestimated (see §4), this procedure is incorrect, as already pointed out clearly in the documentation for EBOP (Etzel 1980). The reason is that $k$ and $r_A$ are, as a rule, strongly (negatively) correlated, and simply adding their uncertainties by quadratures does not account for this correlation. It is still possible to make correct use of the expression above if the correlation is explicitly taken into account by adding the corresponding term from the off-diagonal elements of the covariance matrix in the least squares solution. However, this information is not readily available in the standard output from EBOP. Alternative procedures for estimating $\sigma_{r_B}$ are discussed in the documentation for the code (see also Popper & Etzel 1981). For example, when $k$ is fixed in the least squares solution one may use as an approximation the expression $$\sigma_{r_B}= k\cdot \sigma_{r_A},$$ although this explicitly ignores the error in $k$, which in our case is well known independently of EBOP. The reason for the strong correlation between $k$ and $r_A$ is that the *sum* of the radii, $r \equiv r_A + r_B$, which is directly related to the form of the light curve, is usually very well determined, and therefore if $k$ is increased the primary radius $r_A$ must decrease to maintain the sum. This suggests that a more sensible way to approach the problem might be to consider the expressions $$r_A = \left({1\over 1+k}\right)r~,~~~~~~~~~r_B = \left({k\over 1+k}\right)r~,$$ in which the independent variables are $r$ and $k$, which are *not* strongly correlated in general. For GG Ori we have shown this to be the case in §4. Propagation of errors in these equations may therefore be expected to give more reasonable estimates of $\sigma_{r_A}$ and $\sigma_{r_B}$: $$\sigma_{r_A} = {1\over 1+k}\left[\left({r\over 1+k}\right)^2 \sigma_k^2 + \sigma_r^2\right]^{1/2}~, ~~~~\sigma_{r_B} = {1\over 1+k}\left[\left({r\over 1+k}\right)^2 \sigma_k^2 + k^2 \sigma_r^2\right]^{1/2}.$$ Note that $\sigma_{r_B}$ will be smaller or larger than $\sigma_{r_A}$ depending only on the value of $k$, as in eq.(1). In fact, setting $\sigma_k = 0$ and combining the expressions above leads back to (1). Practical application of these equations requires an estimate of $\sigma_r$. Although $r$ is not one of the adjustable variables in EBOP, it is not difficult to obtain an estimate of its error by experimenting with the light curve solutions and assessing the sensitivity of $r$ to the various input quantities, as we have in fact done with $r_A$. However, to ensure consistency with the error in $r_A$, we have chosen to use the first of the equations in (3) to solve for $\sigma_r$, since $\sigma_{r_A}$ is already known and so is $\sigma_k$. Once we determined the value of $\sigma_r$ needed to reproduce $\sigma_{r_A}$, we used it in the expression for $\sigma_{r_B}$. Andersen, J. 1991, , 3, 91 Andersen, J. 1998, in Fundamental Stellar Properties: The Interaction between Observation and Theory, IAU Symp. 189, eds.T. R. Bedding, A. J. Booth & J. Davis (Dordrecht: Reidel), 99 Blaauw, A. 1964, , 2, 213 Braune, W., Hübscher, J., & Mundry, E. 1977, AN, 298, 121 Brown, A. G. A., de Geus, E. J., & de Zeeuw, P.T. 1994, , 289, 101 Brown, A. G. A., Walter, F. M., & Blaauw, A. 1999, in The Orion Complex Revisited, ASP Conf. Ser., ed. M. J.McCaughrean & A. Burkert (San Francisco: ASP), in press Busch, H. 1975, Mitt. der Bruno-H.-Burgel Sternwarte Hartha, 8, 18 Casey, B. W., Mathieu, R. D., Vaz, L. P. R., Andersen, J., & Suntzeff, N. B. 1998, , 115, 1617 Claret, A. 1995, , 109, 441 Claret, A., & Cunha, N. C. S. 1997, , 318, 187 Claret, A., & Giménez, A. 1992, , 96, 255 Claret, A., Giménez, A., & Cunha, N. C. S. 1995, , 299, 724 Claret, A., Giménez, A., & Martín, E. L. 1995, , 302, 741 Clausen, J. V., & Nördstrom, B. 1980, , 83, 339 Crawford, D. L. 1978, , 83, 48 Deutschman, W. A., Davis, R. J., & Schild, R. E. 1976, , 30, 97 Diethelm, R. 1975, BBSAG Bull., 21, 4 Diethelm, R. 1976, BBSAG Bull., 25, 2 Diethelm, R. 1983, BBSAG Bull., 65, 4 Diethelm, R. 1992, BBSAG Bull., 21, 1 Duquennoy, A., Mayor, M., & Mermilliod, J.-C. 1992, in Binaries as Tracers of Stellar Formation, eds. A. Duquennoy & M. Mayor (Cambridge: Cambridge Univ. Press), 52 Etzel, P. B. 1980, EBOP User’s Guide, Univ. of California, Los Angeles Etzel, P. B. 1981, in Photometric and Spectroscopic Binary Systems, ed. E. B. Carling & Z. Kopal (Dordrecht: Reidel), 111 Fekel, F. C. 1999, in Precise Radial Velocities, ASP Conf. Ser. 185, eds. J. B. Hearnshaw & C. D. Scarfe (San Francisco: ASP), 378 Giménez, A. 1985, , 297, 405 Hilditch, R. W., & Hill, G. 1975, , 79, 101 Hoffleit, D. 1982, The Bright Star Catalogue (New Haven: Yale Univ. Obs.) Hoffmeister, C. 1934, AN, 253, 195 Izotov, Y. I., & Thuan, T. X. 1998, , 500, 188 Johnson, H. L., & Morgan, W. W. 1953, , 117, 313 Kholopov, P. N. 1985, General Catalog of Variable Stars (Moscow: Nauka), Vol. II Kordylewski, K. 1951, Acta Astr. Ser. C, 4, 134 Lacy, C. H. S. 1984, IBVS, No. 2489 Lacy, C. H. S. 1992a, , 104, 801 Lacy, C. H. S. 1992b, , 104, 2213 Lacy, C. H. S., & Fox, G. W. 1994, IBVS, No.4009 Lacy, C. H. S., Ibanoglu, C., Tunca, Z., Evren, S., Akan, C., Keskin, V., Zakirov, M., Arzumanyants, G., Ishankulov, R., & Kharchenko, V. 1995, IBVS, No. 4194 Lacy, C. H. S., Torres, G., Claret, A., Stefanik, R. P., Latham, D. W., & Sabby, J. A. 2000, , 119, 1389 Landolt, A. U. 1973, , 78, 959 Latham, D. W. 1992, in IAU Coll. 135, Complementary Approaches to Double and Multiple Star Research, ASP Conf. Ser. 32, eds. H. A. McAlister & W. I. Hartkopf (San Francisco: ASP), 110 Latham, D. W., Nordström, B., Andersen, J., Torres, G., Stefanik, R. P., Thaller, M., & Bester, M. 1996, , 314, 864 Lavrov, M. I. 1993, Trudy Kazansk. Gor. Astron.Obs., 53, 34 Mamajek, E. E., Lawson, W. A., Feigelson, E. D. 2000, , in press Martynov, D. Y., 1973, in Eclipsing Variable Stars, ed. V. P. Tsesevich (New York: Halsted), 146 Mathieu, R. D., Duquennoy, A., Latham, D. W., Mayor, M., Mazeh, T., & Mermilliod, J.-C. 1992, in Binaries as Tracers of Stellar Formation, ed. A. Duquennoy & M. Mayor (Cambridge: Cambridge Univ. Press), 278 Nikulina, T. G. 1958, Astr. Circ., 189, 16 Pagel, B. E. J., & Portinari, L. 1998, , 298, 747 Peimbert, M. 1993, Rev. Mex. Astr. Astrof., 27, 9 Pols, O. R., Tout, C. A., Schröder, K.-P., Eggleton, P. P., & Manners, J. 1997, , 289, 869 Popper, D. M. 1968, , 154, 191 Popper, D. M. 1987, , 313, L81 Popper, D. M. 1980, , 18, 115 Popper, D. M., & Etzel, P. B. 1981, , 86, 102 Renzini, A. 1994, , 285, L5 Ribas, I., Jordi, C., Torra, J., & Giménez, A. 2000, , 313, 99 Soloviev, A. 1945, Astr. Circ., 41, 8 Stefanik, R. P., Latham, D. W., & Torres, G. 1999, in Precise Stellar Radial Velocities, IAU Coll. 170, ASP Conf.Ser., 185, eds. J. B. Hearnshaw & C. D. Scarfe (San Francisco: ASP), 354 Szafraniec, R. 1952, Acta Astr. Ser. C, 5, 7 Szafraniec, R. 1955, Acta Astr. Ser. C, 5, 189 Szafraniec, R. 1956, Acta Astr., 6, 141 Szafraniec, R. 1958, Acta Astr., 8, 189 Tassoul, M., & Tassoul, J.-L. 1997, , 481, 363 Tshuprina, P. I. 1957, Astr. Circ., 183, 16 Torres, G., Andersen, J., Nördstrom, B., & Latham, D. W. 2000, , 119, 1942 Torres, G., Lacy, C. H. S., Claret, A., Zakirov, M.M., Arzumanyants, G. C., Bayramov, N., Hojaev, A. S., Stefanik, R.P., Latham, D. W., & Sabby, J. A. 1999, , 118, 1831 Torres, G., Stefanik, R. P., Andersen, J., Nördstrom, B., Latham, D. W., & Clausen, J. V. 1997, , 114, 2764 Verbunt, F., & Phinney, E. S. 1995, , 296, 709 Winiarski, M. 1972, Acta Astr., 24, 89 Wade, R. A., & Rucinski, S. M. 1985, , 60, 471 Warren, W. H. Jr., & Hesser, J. E. 1977, , 34, 115 Zakirov, M. M. 1997, Pis’ma Astr. Zh., 23, 626 Zahn, J.-P. 1992, in Binaries as Tracers of Stellar Formation, ed. A. Duquennoy & M. Mayor (Cambridge: Cambridge Univ. Press), 253 Zahn, J.-P., & Bouchet, L. 1989, , 223, 112 de Zeeuw, P. T., Hoogerwerf, R., de Bruijne, J. H.J., Brown, A. G. A., & Blaauw, A. 1999, , 117, 354 Zessewitsch, W. P. 1945, Astr. Circ., 40, 7 Zucker, S., & Mazeh, T. 1994, , 420, 806 [^1]: IRAF is distributed by the National Optical Astronomy Observatories, which is operated by the Association of Universities for Research in Astronomy, Inc., under contract with the National Science Foundation.

--- abstract: 'Despite nearly a century of study of the $S=1/2$ Heisenberg model on the square lattice, there is still disagreement on the nature of its high-energy excitations. By tuning toward the Heisenberg model from the exactly soluble Ising limit, we find that the strongly attractive magnon interactions of the latter naturally account for a number of spectral features of the Heisenberg model. This claim is backed up both numerically and analytically. Using the density matrix renormalization group method, we obtain the dynamical structure factor for a cylindrical geometry, allowing us to continuously connect both limits. Remarkably, a semi-quantitative description of certain observed features arises already at the lowest non-trivial order in perturbation theory around the Ising limit. Moreover, our analysis uncovers that high-energy magnons are localized on a single sublattice, which is related to the entanglement properties of the ground state.' author: - Ruben Verresen - Frank Pollmann - Roderich Moessner bibliography: - 'SLAFHM\_v2.bib' title: 'Quantum dynamics of the square-lattice Heisenberg model' --- At its ripe age of 90 years[@Heisenberg28], the square lattice antiferromagnetic Heisenberg model has had its dynamical properties studied intensely. Spin wave theory was in large part developed to investigate this model’s low-energy properties[@Hulthen36; @Anderson52; @Harris71]. Its anomalous terms lead to a ‘relativistic’ (linear) low-energy dispersion, related to coupling the two sublattices of the ground state’s spontaneous Néel ordering[@Keffer53]. While evading an exact treatment, its dynamics has been studied numerically via quantum Monte Carlo simulations[@Sandvik01; @Shao17] and exact diagonalization[@Luescher09]; also several high-order perturbative expansions have been devised, such as around the Ising limit[@Singh89; @Gelfand96; @Singh95; @Zheng05] or via continuous unitary transformations (CUT)[@Powalski15; @Powalski17]. In addition, a number of ad-hoc approaches have been motivated by a range of different physical pictures. Moreover, this model is related to experimental systems, not least to the parent states of the cuprates[@Anderson87; @Rice87; @Chakravarty88; @Ho01], the study of which has led to much of the modern theory of quantum magnetism. It may thus seem all the more surprising that there is still no consensus on the appropriate physical picture for certain regions in momentum space. Proposals include strongly-interacting magnons[@Powalski15; @Powalski17], deconfined spinons[@Hsu90; @Ho01; @Syljuaasen02; @DallaPiazza15; @Ferrari18], all the way to a connection to deconfined quantum criticality [@Shao17]. The disagreement is not limited to the underlying physical mechanism, but also pertains to quantitative aspects of spectral properties. One uniting factor, at least, is a need to go beyond a perturbatively-dressed single-magnon picture. Our intention is not to propose yet another scenario. Rather, we adopt the perspective that away from the unassailable hydrodynamic limit—accounting for the low-energy Goldstone modes[@Keffer53; @Nielsen76]—other features which have caught the attention of the community may not even be uniquely described by one picture as opposed to another. Instead, we seek to provide a simple account of salient features of the intermediate and high-energy part of the spectrum. Perhaps the most controversial region concerns magnons with momenta $|k_x|+|k_y| = \pi$. Spin wave theory predicts that these are dispersionless, in disagreement with both experiment[@Ronnow01; @Christensen07; @DallaPiazza15] and theoretical methods. What is instead observed, is a local mininum at $\bm k = (\pi,0)$—commonly referred as the *roton mode*, in analogy with the quasi-particle dispersion in liquid Helium[@Landau41]. Moreover, spin wave theory is unable to account for the high spectral weight in the continuum just above this mode. In this paper, we advance along two complementary tracks. First, we determine the dynamical structure factor using a method based on the density matrix renormalization group (DMRG)[@White92; @White93; @Schollwoeck11; @Stoudenmire13; @Zaletel15; @Gohlke17], with systematic errors distinct from those of previous approaches. This novel method has at least two welcome features: it confirms that the phenomenology of the roton mode is indeed beyond the dressed single-magnon picture, and it uncovers a hitherto-unrecognized property of magnons with $|k_x|+|k_y| = \pi$, which we refer to as *sublattice-localization*. We also clarify how the latter is related to the entanglement of the ground state. Second, we use existing data from an Ising expansion developed by Singh and Gelfand[@Singh89; @Gelfand96; @Singh95]—and pushed further by Zheng, Oitmaa and Hamer[@Zheng05]—to point out that some known results at the isotropic point are already semi-quantitatively accounted for by the lowest non-trivial order. Moreover, our numerical method allows us to study such an XXZ model (with dominant easy-axis anisotropy) without any perturbative approximation. The main message of our paper is that aspects of the attractive magnon interactions, i.e. the physics beyond spin wave theory, arise naturally from domain-wall-counting in the Ising limit, sometimes connecting all the way to the isotropic Heisenberg point. In particular, the numerics shows this at a phenomenological level, but a simple perturbative calculation also sheds light on, e.g., the small-yet-nonzero magnitude and shape of the roton mode’s dispersion. Moreover, even the aforementioned phenomenon of sublattice-localization can be accounted for within a low-order perturbative picture. In addition, we provide a quantitative analysis of the roton mode. The remainder of this paper is structured as follows. In section \[sec:model\] we give a brief overview of the model’s salient features, relating them to previous literature whenever possible. The spectral functions obtained using DMRG are shown in section \[sec:spectral\]: first for the Heisenberg model, which is then connected to the Ising limit. Section \[sec:perturbative\] supplements this by showing how various features, such as the roton minimum or sublattice-localization, naturally arise within a low-order perturbative picture. The apparently hitherto-unexplored phenomenon of sublattice-localization is studied numerically in section \[sec:entanglement\], emphasizing its link to entanglement (or absence thereof). Section \[sec:roton\] contains a quantitative analysis of the roton mode with comparison to results from the literature. Square lattice Heisenberg model {#sec:model} =============================== We study the spin-$\frac{1}{2}$ antiferromagnetic Heisenberg model (AFH) on the square lattice, allowing for easy-axis anisotropy: $$H = J \sum_{\langle\bm n, \bm m \rangle} \Big( S^z_{\bm n} S^z_{\bm m} + \lambda \left[S ^x_{\bm n} S^x_{\bm m} + S^y_{\bm n} S^y_{\bm m} \right] \Big) \label{eq:model}$$ where $J>0$. We are principally interested in the isotropic point $\lambda=1$, where the Néel order of the ground state spontaneously breaks the $SU(2)$ symmetry down to a $U(1)$ group generated by $S^z_\textrm{tot} = \sum_{\bm n} S^z_{\bm n} $ (where we define the ordering direction to be along the spin $z$-axis). As we will argue, it is also useful to consider $0 \leq \lambda < 1$, where the model is in a gapped Ising phase which spontaneously breaks the $\mathbb Z_2$ symmetry $R^x_\pi = \prod_{\bm n} \exp{\left(-i \pi S^x_{\bm n} \right)}$. Dynamical structure factor and quantum numbers ---------------------------------------------- Spectral functions give direct insight into the properties of excitations. In this work we focus on the dynamical structure factor, which is experimentally accessible through, for example, inelastic neutron scattering. It can be expressed in terms of the dynamical correlation functions $C^{\gamma\gamma}(\bm r,t) = \langle \sigma^\gamma_{\bm r}(t) \sigma^\gamma_{\bm 0}(0) \rangle = 4\langle S^\gamma_{\bm r}(t) S^\gamma_{\bm 0}(0) \rangle $: $$\mathcal S^{\gamma \gamma}(\boldsymbol k,\omega) = \frac{1}{2\pi} \sum_{\mathbf r} \int_{-\infty}^\infty e^{i(\omega t-\mathbf k \cdot \mathbf r)} C^{\gamma\gamma}(\mathbf r, t) \; \mathrm dt, \label{eq:spectral}$$ which is normalized as $\int \mathcal S^{\gamma \gamma}(\boldsymbol k,\omega) \; \mathrm d \mathbf k \mathrm d \omega = (2\pi)^d$. We focus on the transverse spectral function: $$\mathcal S^t(\bm k,\omega) = \mathcal S^{xx}(\bm k,\omega) + \mathcal S^{yy}(\bm k,\omega) . \label{eq:transverse}$$ This object gives direct insight into the excitations above the ground state. If $\gamma =x,y$, one can show[^1] that $$\mathcal S^{\gamma \gamma} (\bm k, \omega) = \sum_\alpha \delta(\omega - (\omega_\alpha-\omega_0)) \; |\langle \alpha | \tilde S^\gamma_{\bm k} |0\rangle|^2 \label{eq:Lehmann}$$ where $\tilde S^\gamma_{\bm k} = \sum_{\bm r} e^{i\bm{k\cdot r}} S_{\bm r}^\gamma$. It is natural to choose a basis $|\alpha\rangle = |\bm k , S^z_\textrm{tot}, \beta\rangle$, where $\bm k$ is the momentum with respect to the translation symmetry $T_{1,\pm 1}$ of the *two-site unit cell*. Eq.  tells us that the spectral function gives information about the existence of energy eigenstates with momentum $\bm k$ and and $S^z_\textrm{tot} = \pm 1$. Note that when labeling states, $\bm k$ lives in the reduced (magnetic) Brillouin zone, $|k_x| + |k_y| \leq \pi$, but the spectral function itself is periodic only with respect to the original (lattice) Brillouin zone, $-\pi \leq k_x, k_y \leq \pi$ (taking the lattice constant to be unity). Spin wave theory ---------------- In terms of the above quantum numbers, spin wave theory predicts two bands[^2]. These exactly coincide and are distinguished by $S^z_\textrm{tot} = \pm 1$. The dispersion relation to order[^3] $1/S^0$, i.e. linear spin wave theory (LSWT), is[@Anderson52; @Kubo52] $$\varepsilon^\textrm{LSWT}_{\bm k} = \sqrt{4 - \lambda^2 (\cos(k_x) + \cos(k_y))^2}.$$ Hence for $\lambda=1$, there are two linearly-dispersing Goldstone modes at the zone center (in sectors $S^z_\textrm{tot} = \pm 1$), consistent with two of three generators of $SU(2)$ being spontaneously broken[@Nielsen76]. However, based on general sum rules[@Stringari94], it is known that $|| S^{x,y}_{\bm k} |0\rangle || \sim|\bm k|$ as $\bm k \to \bm 0$, such that the Goldstone modes will have vanishing intensity in the transverse spectral function $\mathcal S^t(\bm k ,\omega)$ at the zone center. Instead, they show up near the ordering wavevector M $= (\pi,\pi)$, since the same sum rules imply an (integrable) divergence $|| S^{x,y}_{\textrm{M}+\bm k} |0\rangle || \sim 1/|\bm k|$ as $\bm k \to \bm 0$. The first order corrections to the dispersion within spin wave theory are[@Oguchi60; @Zheng91] $$\varepsilon^\textrm{LSWT+$1/S$}_{\bm k} = a \; \varepsilon^\textrm{LSWT}_{\bm k} - (1-\lambda^2) \;b \left( \frac{2}{\varepsilon^\textrm{LSWT}_{\bm k}} - \frac{\varepsilon^\textrm{LSWT}_{\bm k}}{2} \right)$$ where $a$ and $b$ are ($\lambda$-dependent) constants[^4]. At the isotropic point, this correction is only a momentum-independent rescaling. Higher-order corrections ($1/S^2$ and $1/S^3$) are also known[@Hamer92; @Zheng93; @Syrom10], which we discuss in section \[sec:roton\]. Phenomenology of diagonal magnons: a short review ------------------------------------------------- The purpose of this section is to give a brief (and, unavoidably, partial) overview of some of the salient features which have been the focus of much of the previous work on the excitations of this model. There is a peculiar property of the LSWT prediction: $\varepsilon_{\bm k}$ is constant along $|k_x| + |k_y| = \pi$. For convenience, we refer to magnons with these momenta as being *diagonal*. This dispersionless feature is a consequence of the more basic fact that at these momenta, the low-order spin wave Hamiltonian vanishes. This also means that the Bogoliubov rotation, which normally mixes the bosons of the two sublattices, is absent there. We thus arrive at the fact that, within LSWT, the diagonal magnons are purely *localized on a single sublattice*. In fact, this one-dimensional flatness in the spectrum means one has a freedom in choosing a basis of energy eigenstates. By Fourier transforming the momentum eigenstates along one direction, one can thus construct eigenstates which are spatially *localized onto a single diagonal* (of a given sublattice), with alternating signs along this diagonal. In summary, at low order in SWT, diagonal magnons are localized on both a sublattice and a diagonal. (In section \[sec:perturbative\], we show that the same features arise naturally at low order in the Ising expansion.) Despite being flat in LSWT and LSWT+$1/S$, the diagonal magnons acquire a finite but very small dispersion at higher order[@Hamer92; @Zheng93; @Syrom10; @Uhrig13]. Equivalently, this means that magnons can no longer be confined onto a single diagonal. However, it has not yet been investigated whether the aforementioned sublattice-localization persists. We study this both numerically (section \[sec:entanglement\]) and perturbatively (section \[subsec:states\]). The SWT predictions at diagonal momenta, $|k_x|+|k_y|=\pi$, do not agree well with other methods or experiments[@Ronnow01; @Christensen07; @DallaPiazza15]—both with respect to single- and multi-magnon features. Examples of previous studies include methods based on quantum Monte Carlo (QMC) combined with analytic continuation[@Sandvik01; @Shao17], series expansions in $\lambda$ (up to 14th order)[@Singh95; @Zheng05], exact diagonalization (ED)[@Luescher09] and the continuous unitary transform (CUT) [@Powalski15; @Powalski17]. All these methods predict a more pronounced local minimum of the magnon at $\bm k =\left(\pi,0\right) = X$, referred to as the roton mode, although they do not agree on its exact magnitude or shape (a quantitative discussion is deferred to section \[sec:roton\]). More strikingly, they also predict an unusually large weight in the continuum above this local minimum. The latter phenomenology is also observed in experiment[@Christensen07; @Headings10; @DallaPiazza15], and exotic scenarios have been given to explain it. For example, it has been argued that near $\bm k \approx X$, the magnon can be seen as two (nearly) deconfined spinons [@Ho01; @DallaPiazza15; @Shao17; @Ferrari18]. This interpretation has subsequently been challenged by the CUT method [@Powalski15; @Powalski17], which reproduces various salient features based on a picture of strongly-interacting magnons. The intuitive nature of said strong interactions, however, has not yet been clarified. We will argue that an Ising-like domain-wall interaction naturally accounts for it. Spectral functions {#sec:spectral} ================== In this section we discuss the transverse spectral function $\mathcal S^t(\bm k,\omega)$ as defined in Eq. . For this, we use the numerical method introduced in Ref. , which we briefly outline here. Firstly, the model  is put on an infinitely long cylinder whose finite, periodic direction is along a zigzag/staircase path. We define the circumference $L_\textrm{circ}$ in Manhattan distance, i.e. the minimal number of bonds needed to wrap around the cylinder. In this work, $L_\textrm{circ} = 8,10$. The infinite density matrix renormalization group (iDMRG) method [@White92; @White93; @Kjaell13] is used to obtain the ground state[@Stoudenmire13]. The dynamical spin-spin correlations $C^{\gamma,\gamma}(\bm r,t)$ can then be calculated by using a matrix-product-operator-based time evolution[@Zaletel15]. The spectral function follows directly from Eq.  [@Schollwoeck11]. Let us mention a few technical details before discussing the results. To minimize the effects of Fourier transforming a finite-time window, we use linear prediction[@White08] to increase the time window, after which we multiply the data with a Gaussian envelope[^5]. This effectively introduces an artificial broadening of the spectral function with full-width-at-half-maximum $2.355 \sigma_\omega$. For a given circumference, we confirm that our results are converged in both bond dimension and inverse time-step by increasing both until the results no longer change. Due to the expensive nature of time-evolving large cylinders, in this work we are limited to bond dimension $\chi \approx 400$ for the largest circumference considered ($L_\textrm{circ}=10$). A typical size that we used for the time-step is $dt = 0.01/J$. The conservation of $S^z_\textrm{tot}$ was implemented explicitly. Isotropic/Heisenberg model\[subsec:spectral\] --------------------------------------------- Fig. \[fig:isotropic\] shows the transverse spectral function at the isotropic point ($\lambda = 1$). Because of the cylindrical geometry, momentum is discrete along one direction and continuous along the other. This is indicated by the red lines in the Brillouin zone in Fig. \[fig:isotropic\]. Since the periodic direction is along a zigzag/staircase path, the momentum cuts are lines of constant $k_x - k_y$. This means we can directly access a line of diagonal magnons (as defined in section \[sec:model\]), in the figure denoted by the line segment Y–X, where $X = (\pi,0)$ and $Y = \left( \frac{\pi}{2},-\frac{\pi}{2} \right) \cong \left( \frac{\pi}{2},\frac{\pi}{2} \right)$ (by symmetry). In fact, while it is true that $ \left( \frac{\pi}{2},-\frac{\pi}{2} \right)$ and $\left( \frac{\pi}{2},\frac{\pi}{2} \right)$ are symmetry-equivalent in 2D, this is not strictly true on the cylinder geometry. However, such finite-size effects turn out to be small, as discussed in Appendix \[app:analysis\]. The same line of diagonal magnons, X–Y, can be accessed for $L_\textrm{circ} = 10$ if there are antiperiodic boundary conditions along the finite direction, shifting the momentum cuts as shown. We numerically observe the Goldstone modes at the zone center and the ordering wavevector $\textrm{M}=(\pi,\pi)$. Moreover, the intensity vanishes at the zone center, and diverges at M, consistent with the sum rules discussed in section \[sec:model\]. This agrees with the Goldstone modes predicted by LSWT+$1/S$ (solid blue line), whereas the naive evaluation of the series expansion data (up to $\lambda^{12}$) does not reproduce this[^6] (dashed black line). On the other hand, along the Y–X line, the series expansion data fares better at reproducing the local minimum at $\textrm{X} = (\pi,0)$. As discussed in section \[sec:model\], SWT predicts a flat dispersion along Y-X. Moreover, even in this linear color scale, we can see spectral weight above the single-magnon curve near $\bm k \approx \textrm{X}$. These single- and multi-magnon features at the isotropic point are analyzed in more detail in section \[sec:roton\]. Here, we limit ourselves to a few general, conceptual remarks. For $L_\textrm{circ} = 8$, the system is gapless at the zone center, hence the multi-magnon continuum starts at the one-magnon branch (up to the miniscule gap introduced by using a finite bond dimension). For the $L_\textrm{circ}=10$ geometry, however, the antiperiodic boundary condition imply that we do not pass through the Goldstone mode, such that the multi-magnon continuum is separated from the one-magnon branch. Nevertheless, these antiperiodic boundary conditions have some useful side-effects. Due to now simulating a gapped system, it is easier to converge the numerics in the bond dimension parameter of the matrix product state describing the ground state. Moreover, it allows the ground state to spontaneously break the symmetry, even at the isotropic point. This is non-trivial given our set-up, since the cylinder is effectively a one-dimensional system (with a large unit cell), such that the Mermin-Wagner-Coleman theorem[@Mermin66; @Coleman73] should prevent ordering. The catch is that the antiperiodic boundary conditions explicitly break the $SU(2)$ symmetry, although this is not locally noticeable. This effective gap for $L_\textrm{circ} = 10$ can give us further insight into the physics beyond that of a single magnon. In Fig. \[fig:isotropic\_log\], we show the same data in log-scale. We see a continuum right above the single-magnon branch near $\bm k \approx X$. However, this continuum does *not* fall within the frequency region of the kinematic (non-interacting) three-magnon continuum[^7]. To emphasize this, we have plotted the three-magnon continuum *for this cylinder geometry* in the grey shaded region. Hence, the continuum above the roton mode is instead related to (quasi-)bound states. More precisely, using the insights from the upcoming section \[subsec:evolution\], this continuum is a combination of closely packed three-magnon (quasi-)bound states and a continuum made out of a single magnon and a two-magnon (quasi-)bound state. This is strongly suggestive that the roton mode arises by being repelled from these strongly-interacting states. This agrees with the conclusions of the CUT approach[@Powalski15; @Powalski17]. Interpolating between the Ising and Heisenberg limits\[subsec:evolution\] ------------------------------------------------------------------------- ![The transverse spectral function as a function of $\lambda$ ($L_\textrm{circ} = 8$, log-scale). Spectral features at the $SU(2)$ point ($\lambda=1$) can be connected back to the Ising limit ($\lambda=0$). Near the Ising limit, we identify the one-magnon branch, three-magnon bound states and three-magnon continuum based on domain-wall counting. Solid blue line is SWT$+1/S$ [@Oguchi60; @Zheng91], dashed black line is series expansion[@Zheng05; @Oitmaa18] up to order $\lambda^{12}$. For the labeling of the momentum directions, see Fig. \[fig:isotropic\]. \[fig:spectral\_evolution\]](spectral_evol1 "fig:") ![The transverse spectral function as a function of $\lambda$ ($L_\textrm{circ} = 8$, log-scale). Spectral features at the $SU(2)$ point ($\lambda=1$) can be connected back to the Ising limit ($\lambda=0$). Near the Ising limit, we identify the one-magnon branch, three-magnon bound states and three-magnon continuum based on domain-wall counting. Solid blue line is SWT$+1/S$ [@Oguchi60; @Zheng91], dashed black line is series expansion[@Zheng05; @Oitmaa18] up to order $\lambda^{12}$. For the labeling of the momentum directions, see Fig. \[fig:isotropic\]. \[fig:spectral\_evolution\]](spectral_evol2 "fig:") ![The transverse spectral function as a function of $\lambda$ ($L_\textrm{circ} = 8$, log-scale). Spectral features at the $SU(2)$ point ($\lambda=1$) can be connected back to the Ising limit ($\lambda=0$). Near the Ising limit, we identify the one-magnon branch, three-magnon bound states and three-magnon continuum based on domain-wall counting. Solid blue line is SWT$+1/S$ [@Oguchi60; @Zheng91], dashed black line is series expansion[@Zheng05; @Oitmaa18] up to order $\lambda^{12}$. For the labeling of the momentum directions, see Fig. \[fig:isotropic\]. \[fig:spectral\_evolution\]](spectral_evol3 "fig:") ![The transverse spectral function as a function of $\lambda$ ($L_\textrm{circ} = 8$, log-scale). Spectral features at the $SU(2)$ point ($\lambda=1$) can be connected back to the Ising limit ($\lambda=0$). Near the Ising limit, we identify the one-magnon branch, three-magnon bound states and three-magnon continuum based on domain-wall counting. Solid blue line is SWT$+1/S$ [@Oguchi60; @Zheng91], dashed black line is series expansion[@Zheng05; @Oitmaa18] up to order $\lambda^{12}$. For the labeling of the momentum directions, see Fig. \[fig:isotropic\]. \[fig:spectral\_evolution\]](spectral_evol4 "fig:") ![The transverse spectral function as a function of $\lambda$ ($L_\textrm{circ} = 8$, log-scale). Spectral features at the $SU(2)$ point ($\lambda=1$) can be connected back to the Ising limit ($\lambda=0$). Near the Ising limit, we identify the one-magnon branch, three-magnon bound states and three-magnon continuum based on domain-wall counting. Solid blue line is SWT$+1/S$ [@Oguchi60; @Zheng91], dashed black line is series expansion[@Zheng05; @Oitmaa18] up to order $\lambda^{12}$. For the labeling of the momentum directions, see Fig. \[fig:isotropic\]. \[fig:spectral\_evolution\]](spectral_evol5 "fig:") Fig. \[fig:isotropic\] confirms that the DMRG method can reproduce the roton mode and the strong presence of multi-magnon features near $\bm k \approx X$. However, to gain insight into *how* and *from where* these features appear, it is useful to interpolate from the Ising limit ($\lambda=0$) to the $SU(2)$-symmetric point ($\lambda=1$). This is illustrated in Fig. \[fig:spectral\_evolution\] for $L_\textrm{circ}=8$. In particular, we demonstrate that the features at the isotropic point can be traced back to those in the Ising limit. To this end, let us first identify the spectral features close to the Ising limit ($\lambda=\frac{1}{3}$). We can do this by using simple energetic arguments. Note that when $\lambda = 0$, a ‘magnon’ corresponds to a single localized spin flip with energy cost $\frac{J}{2}$ per bond, totaling $2J$. In Fig. \[fig:spectral\_evolution\], we see that for for $\lambda = \frac{1}{3}$ the magnon has gained some dispersion, but its energy is still roughly $2J$. Since the transverse spectral function picks up states with $S^z_\textrm{tot} = \pm 1$, the next excitation contains three magnons. Energetically, these magnons prefer to form a bound state whose domain wall counts eight bonds. Indeed, we observe bound states at energy $8 \times \frac{J}{2} = 4J$. There are several such states at this energy due to the internal degree of freedom corresponding to orientation and shape. At even higher energies, there is the kinematic continuum made out of a two-magnon bound state ($6\times\frac{J}{2}$) and a free magnon ($2J$) with total energy around $5J$. Having identified all spectral features for $\lambda=\frac{1}{3}$, we track their evolution as we tune $\lambda \to 1$ in Fig. \[fig:spectral\_evolution\]. At $\lambda = \frac{1}{2}$, some of the three-magnon bound states have merged. When $\lambda \approx 0.77$, several of the three-magnon bound states have already been absorbed into the three-magnon continuum. Closer to the isotropic point, $\lambda \approx 0.91$, the three-magnon continuum continues to come down in energy. This trend gradually continues up to $\lambda \approx 0.99$. We see that the spectral features vary continuously as a function of $\lambda$. In particular, we see that there is no restructuring of the magnon near $\bm k \approx (\pi,0)$ for any $\lambda < 1$. This was a priori not a given. Read backwards, this means that the features near the isotropic point can be continuously traced back to those in the Ising limit. Relatedly, it is worth pointing out that even at the isotropic point, the multi-magnon continuum is not featureless. We discuss this substructure more quantitatively in section \[sec:roton\]. In Fig. \[fig:isotropic\_log\], we saw how for $L_\textrm{circ}=10$, there is a continuum above the roton mode which is not made out of kinematic combinations of magnons. In section \[subsec:spectral\], we claimed that it is instead a continuum made up out of (quasi-)bound states. The justification for this claim is that by smoothly decreasing $\lambda$, the observed continuum indeed splits up into three-magnon bound states and a continuum made up out of a magnon and a two-magnon bound state. Perturbative understanding from the Ising limit {#sec:perturbative} =============================================== In section \[sec:spectral\] we saw that we could connect spectral features of the isotropic model to those near the Ising limit. The purpose of this section is to complement this by gaining insights from low-order perturbation theory in $\lambda$. The point is not to see how well the isotropic point can be described *quantitatively* by a series expansion in $\lambda$ [@Singh95; @Zheng05]. Instead, we ask what the lowest order processes are that *qualitatively* explain certain features at the isotropic point. Intriguingly, this already naturally leads to a *semi-quantitative* description. We rewrite Hamiltonian as $H = H_0 + \lambda V$ with $$\begin{array}{cl} H_0 &= J \sum_{\langle\bm n, \bm m \rangle} S^z_{\bm n} S^z_{\bm m}, \\ V &= \frac{J}{2} \sum_{\langle\bm n, \bm m \rangle} \left( S ^+_{\bm n} S^-_{\bm m} + S^-_{\bm n} S^+_{\bm m} \right). \end{array}$$ The Ising limit $\lambda = 0$ is exactly solvable: the ground state is a product Néel state, and the single-magnon excitations consist of localized spin flips. The perturbation $\lambda V$ introduces dynamics to these static excitations. Before going through this in detail, let us give the broad brush strokes to emphasize the simplicity of both the ingredients and results. ![Perturbation theory around the Ising limit. We define the A-(B-)sublattice where spins point down (up) in the ground state. (a) An A-magnon can hop at order $\lambda^2$ through a virtual three-magnon bound state; (b) this one-magnon state which is localized on a single diagonal cannot hop off it at order $\lambda^2$ due to the destructive interference shown in (c); (d) no destructive interference at order $\lambda^4$: different intermediate five-magnon bound states do not have the same energy.\[fig:pert\]](perturbation_theory_v2) Overview and summary of the perturbative picture \[subsec:overview\] -------------------------------------------------------------------- As we will argue, the effective Hamiltonian has contributions at *even* order in $\lambda$ only. The Ising magnons start to hop at order $\lambda^2$ by going through a virtual three-magnon bound state (see Fig. \[fig:pert\](a)). Nevertheless, as we explain in section \[subsec:leading\], magnons with diagonal momentum $|k_x| + |k_y| = \pi$ are still dispersionless at this order. This is equivalent to the statement that if one builds a one-magnon state which is entirely localized on a single diagonal and has momentum $\pi$ along it (see Fig. \[fig:pert\](b)), then it cannot hop off due to destructive interference (see Fig. \[fig:pert\](c)). The key to the destructive interference traces back to the fact that all three-magnon bound states have the same energy, and hence the virtual paths—half of which come with opposite signs due to the $\pi$-momentum—can cancel exactly. Thus from the viewpoint of the Ising expansion, such destructive interference and the resulting flatness of the diagonal magnons seems accidental. It is hence not surprising that if one goes to next-to-leading order, i.e. $\lambda^4$, the diagonal magnons acquire a dispersion. Indeed, now virtual five-magnon bound states appear, which can have differing energies (see Fig. \[fig:pert\](d)). Since the emergence of the roton mode is due to the physics of (virtual) bound states, one can indeed say that this phenomenology is due to the attractive interactions between magnons. Since this interaction is so natural in the Ising language, the qualitatively correct physics arises rather easily. Indeed, the resulting dispersion at order $\lambda^4$ does not just correctly reproduce the qualitative features of having a local minimum at $\bm k = (\pi,0)$ and a maximum at $\bm k = \left( \frac{\pi}{2} , \frac{\pi}{2} \right)$, but evaluating it at $\lambda = 1$ even gives a semi-quantitative description for the isotropic model, as we discuss in section \[subsec:nexttoleading\]. It is moreover in remarkable proximity to the CUT prediction[@Powalski15; @Powalski17], which is a sophisticated framework for strongly-interacting magnons. This success at relatively low order is in contrast with higher-order SWT. That yet-higher-order corrections don’t radically change the physics at hand can be confirmed by repurposing results from previous studies[@Singh95; @Zheng05]. In Fig. \[fig:higherorder\], we show how the dispersion along the line of diagonal magnons has certain ‘harmonics’ generated at distinct orders in $\lambda^n$ (the first non-trivial harmonic appearing at $\lambda^4$). We see that the higher harmonics die off exponentially fast, justifying a low-order picture. Note that such an exponential decay is a priori not obvious, considering that perturbation theory generically leads to an exponential proliferation of the number of terms. ![The size of the Fourier coefficients of the dispersion of the diagonal magnons, $\varepsilon_{\pi-k,k} = J \sum_m a_{2m} \cos(2m k)$, obtained from series expansion up to different orders $\lambda^n$, evaluated at $\lambda=1$. This is based on data from Refs. . We observe that the higher harmonics die off exponentially fast. \[fig:higherorder\]](roton_series_exp_v3) Lastly, we also consider the perturbed wavefunctions at leading order. In particular, the ground state is dressed by pairs of correlated spin flips, introducing entanglement. This would usually allow one to create magnons associated with one sublattice by acting on the other sublattice. Surprisingly, our perturbative analysis implies that this is *not* possible for diagonal magnons. In other words, they seem to be localized on a given sublattice. This is discussed in section \[subsec:states\]. (We also revisit and confirm such sublattice-localization numerically in section \[sec:entanglement\].) We now provide the details of the story as just described. Ising limit and defining magnons -------------------------------- In the Ising limit $\lambda = 0$, the ground state is a product Néel state. Let us define the A-(B-)sublattice to be where spins point *down* (*up*) in the ground state. For any product state in the spin-$z$ basis, we can count the number of flipped spins on the A-sublattice, relative to the ground state, which we denote by $N_A$ (similarly for $N_B$). Hence, the ground state corresponds to $N_A = 0 = N_B$, whereas the single-magnon states have $N_A = 1$, $N_B = 0$ (called A-magnons) or $N_A = 0$, $N_B = 1$ (B-magnons). The perturbation $V$ will mix states with different $N_A$ and $N_B$, however $N_A - N_B = S^z_\textrm{tot}$ remains a well-defined quantum number. If we perturb the system such that, for a given momentum, the one-magnon energy scale does not overlap with multi-magnon states, we can *non-perturbatively* label the single-magnon states by $N_A-N_B =1$ or $N_A- N_B = -1$, referred to as A- and B-magnons, respectively. Since $V$ thus cannot connect A-magnons to B-magnons, we can limit our study to A-magnons. More precisely, we are interested in the *effective* Hamiltonian $H_\textrm{eff}$ on $\mathcal H_0$, the Hilbert space of states satisfying $N_A= 1$ and $N_B = 0$. Note that these Ising magnons are exactly those encountered in Fig. \[fig:spectral\_evolution\]. As discussed in section \[sec:spectral\], a single magnon has a domain wall crossing four bonds and hence has energy $E_0 = 4\times\frac{J}{2} = 2J$ relative to the ground state. Dispersionless diagonal magnons at leading-order \[subsec:leading\] ------------------------------------------------------------------- The single-magnon states are completely static and localized in the Ising limit. They moreover stay immobile at first order in $\lambda$. More precisely, denoting the projector onto $\mathcal H_0$ as $P_0$, then at first order we have $P_0 V P_0 = 0$. This is because $V$ creates a pair of A- and B-magnons out of the vacuum: it, e.g., maps $|\uparrow\downarrow\rangle \xrightarrow{V} \frac{J}{2} |\downarrow \uparrow\rangle$, whereas it *annihilates* ferromagnetic bonds. More generally, $V$ flips the *parity* of $N_{A,B}$, hence the conservation of $N_A - N_B$ shows that there are *no* contributions to $H_\textrm{eff}$ at any odd order $\lambda^{2n+1}$. Thus, by standard perturbation theory, the lowest-order effective Hamiltonian on $\mathcal H_0$ is $$H_\textrm{eff} = E_0 P_0 + \lambda^2 P_0 V G_0 V P_0 + \mathcal O(\lambda^4),$$ where $G_0 = (E-H_0)^{-1}$. This indeed introduces hopping, as shown in Fig. \[fig:pert\](a): the A-magnon can hop to any of the eight nearest sites (on the same sublattice) by going through a virtual three-magnon bound state. As discussed in section \[sec:spectral\], such a bound state involves eight *ferromagnetic* bonds, with total energy cost $8 \times \frac{J}{2} = 4J$—whereas the cost of a single magnon is $E_0 = 2J$. Thus, the path shown in Fig. \[fig:pert\](a) carries a weight $\frac{\lambda J}{2} \times \frac{1}{2J-4J} \times \frac{\lambda J}{2} = -\lambda^2 \frac{J}{8} $. Magnons can thus hop at order $\lambda^2$. However, certain superpositions are immobile due to destructive interference. Consider, for example, the state shown in Fig. \[fig:pert\](b): it is localized on a single A-diagonal, with an alternating sign structure. (We can say its momentum along the diagonal is $\pi$.) The magnon is unable to hop off the diagonal at order $\lambda^2$. This is illustrated in Fig. \[fig:pert\](c), showing two destructively interfering paths. It is important that both paths go through a virtual three-magnon bound state (both with energy $4J$), such that the two weights cancel exactly. Equivalently, all A-magnon momentum eigenstates with $|k_x| + |k_y| = \pi$—which we referred to as diagonal magnons in the previous sections—have constant energy. In summary, the diagonal magnons are dispersionless in the Ising expansion up to order $\mathcal O(\lambda^4)$. It is interesting to note that this coincides with the LSWT$(+1/S)$ predictions in section \[sec:model\]. Roton mode at next-to-leading order \[subsec:nexttoleading\] ------------------------------------------------------------ It is hence important to see what happens at sub-leading order in $\lambda$. Here we follow the perturbative scheme by Kato[@Kato49] and Takahashi[@Takahashi77]. In terms of $H_0$ and $V$, they constructed a general-purpose unitary mapping $\Gamma_\lambda: \mathcal H_0 \to \mathcal H$ which embeds the unperturbed states into the space spanned by the true eigenstates (the latter being $\lambda$-dependent). This object gives us access to the effective Hamiltonian $$H_\textrm{eff} := \Gamma_\lambda^\dagger H \Gamma_\lambda =: \sum_{n=0}^\infty \lambda^n H_\textrm{eff}^{(n)}.$$ As argued before, $H_\textrm{eff}^{(2n+1)} = 0$. Moreover, $H_\textrm{eff}^{(0)} = E_0 P_0$ and $H_\textrm{eff}^{(2)} = P_0 V G_0 V P_0 $. From knowing the aforementioned object $\Gamma_\lambda$ (which, for completeness, we reproduce as a function of $H_0$ and $V$ in Appendix \[app:pert\]), one can derive that $$H_\textrm{eff}^{(4)} = P_0 V \tilde G_0 V \tilde G_0 V \tilde G_0 V P_0 - \frac{1}{2} \left\{ H^{(2)}_\textrm{eff}, P_0 V \tilde G_0^2 V P_0 \right\}$$ where $\tilde G_0 = Q_0 G_0 Q_0 = Q_0 (E-H_0)^{-1} Q_0$ and $Q_0 = 1-P_0$. From this, one can calculate that the diagonal magnons acquire a dispersion at this order. The reason for this is in fact simple, as hinted at in section \[subsec:overview\]. Fig. \[fig:pert\](d) shows two possible virtual five-magnon bound states that can appear as intermediate states. These two states have domain walls extending over, respectively, twelve and ten bonds. Their energy is thus different, and one should not expect perfect destructive interference. More precisely, the resulting dispersion at order $\lambda^4$ is described by a simple cosine-like dispersion for the diagonal magnons: $\varepsilon_{\pi-k,k} =a + b\lambda^2 - c\lambda^4 (d+ \cos(2k)) + \mathcal O(\lambda^6)$ with $a,b,c,d>0$. This has a local (roton) minimum at $\bm k = (\pi,0)$ and a maximum at $\bm k = \left( \frac{\pi}{2},\frac{\pi}{2} \right)$. Moreover, evaluating this at $\lambda = 1$ already gives a semi-quantitative description of the isotropic model. We refer the reader interested in quantitative details to section \[sec:roton\], which is devoted to an in-depth comparison between various different methods. In summary, the roton mode naturally appears in the Ising expansion. Through the property of all three-magnon bound states having the same energy, the local minimum at $\bm k = (\pi,0)$ is absent at leading order, already indicating that it is less pronounced at the isotropic point. At the same time, its salient features readily appear at next-to-leading order, leading to a semi-quantitatively correct description. From this point of view, it is indeed an interacting-magnon effect, where the interaction is based on a simple domain-wall counting in the Ising limit. ![Dressing of the Ising limit ground state and a single A-magnon at leading order in perturbation theory. The gray boxes denote the A-sublattice. The entanglement structure of the above states can be used to understand why the diagonal magnons seem to be localized on a given sublattice, as discussed and observed numerically in section \[sec:entanglement\]. \[fig:perturbation\_states\]](multimagnon_v2) Sublattice-localization of diagonal magnons \[subsec:states\] ------------------------------------------------------------- Aside from looking at the effective Hamiltonian, it can be instructive to consider how the eigenstates evolve with $\lambda$. This is exactly the information encoded in $\Gamma_\lambda$. In Fig. \[fig:perturbation\_states\] we show the leading-order results, both for the ground state as well as a localized A-magnon. (One would have to Fourier transform the latter to obtain an energy eigenstate.) We see that these states are dressed with ‘pair fluctuations’ whilst staying within a well-defined $N_A - N_B =\sum S^z_\textrm{tot} = 0,1$ sector. Having access to the perturbed states, we can ask what excitations are created upon acting with a local operator on the ground state. Fig. \[fig:perturbation\_states\] shows that at this order, a $\sigma^x$ operator does not just create a single magnon, but also three-magnon bound states. This is to be expected and is directly in line with the spectral weight observed in Fig. \[fig:spectral\_evolution\]. It is more interesting to consider what happens when applying $\sigma^-$ on the A-sublattice. This brings us into the sector $N_A - N_B = -1$. In other words, by acting with this operator on the A-sublattice, we create a B-magnon. This is not possible in the product state Ising limit $\lambda \to 0$, where acting with $\sigma^-$ on the A-sublattice annihilates the ground state. But as shown in Fig. \[fig:perturbation\_states\], the perturbation $V$ introduces entanglement, such that $\sigma^-_{\bm n} |\psi_\textrm{gs}\rangle$ is nonzero and has a B-magnon on the four B-sites adjacent to the original site $\bm n \in A$. However, something unusual happens for diagonal momenta. Note that for any given B-site, there are four adjacent A-sites. If the operator we acted with on the A-sublattice has momentum $|k_x| + |k_y| =\pi$, then half of these four sites carry a positive sign, and half a negative sign, so that there would be perfect cancellation. In other words: we are not able to create a B-magnon with $|k_x| + |k_y| = \pi$ by acting on the A-sublattice. The above used an explicit calculation, but the essential mechanism at play should hold at all orders. If we act on a given site of the A-sublattice, then by the $90^\circ$ symmetry of the model, the signs and weights will be the same in all four directions. However, this is incompatible with the alternating sign structure of diagonal momenta. This argument suggests that as long as we act on a single site of the A-sublattice, we *cannot* create a B-magnon with a diagonal momentum. We investigate this claim of *sublattice-localization* non-perturbatively in the following section, detailing its relationship to entanglement. Entanglement and sublattice-localization of diagonal magnons {#sec:entanglement} ============================================================ ![Sublattice-localization of diagonal magnons. (a) If there is entanglement in the ground state (due to pair fluctuations), then it is generically possible to create a B-magnon by acting on the A-sublattice. (b) The *sublattice spectral function* $\mathcal S_{A\to B}(\bm k, \omega)$ obtained with DMRG is shown: this measures whether acting on the A-sublattice can create B-magnons. As defined in Eq. , we act with $S^-$ exclusively on the A-sublattice (where the expectation value already points down). The ground state entanglement is responsible for the non-zero spectral weight on the single-magnon branch (dashed line). Surprisingly, there is *no* weight on the diagonal magnons (edge of shaded square). The plot is for $\lambda = 0.95$ and $L_\textrm{circ}=8$. The same conclusion seems to hold when acting with multi-site operators. \[fig:no\_entanglement\]](entanglement) In this section we discuss a peculiar property of the entanglement in this model. As before, we define the A-sublattice where the spins point down in the ground state (opposite for the B-sublattice). The ground state is in the sector $S^z_\textrm{tot} = 0$, and a magnon associated with the B-sublattice, a *B-magnon*, is in the sector $S^z_\textrm{tot} = -1$. If the ground state had *no* entanglement between the two sublattices, then by acting on the A-sublattice, one could *not* create a B-magnon. The intuitive idea is sketched in Fig. \[fig:no\_entanglement\](a), but the more precise wording is as follows: if $|\Psi_\textrm{gs}\rangle = |\psi_A\rangle \otimes |\psi_B\rangle$ (where $|\psi_{\alpha = A,B}\rangle$ lives on the $\alpha$-sublattice), then, e.g., $S^-_{\bm n}$—which puts us in the $S^z_\textrm{tot}$ sector of a B-magnon—*annihilates* the ground state if $\bm n \in A$. To argue this, note that since the ground state is an eigenstate of $S^z_\textrm{tot}$, then the factorization implies that $|\psi_{\gamma = A,B}\rangle$ must be an eigenstate of $S^z_{\gamma =A,B} := \sum_{\bm n \in \gamma} S^z$. Moreover, since the product Néel state has a finite[^8] overlap with $|\Psi_\textrm{gs}\rangle$, this fixes the eigenvalue of $S^z_A$ to be as (algebraically) small as possible. Hence, acting with the lowering operator on A must annihilate the state. The actual ground state will of course have entanglement; for example, at leading order in $\lambda$, there are two-site spin-flips (‘pair fluctuations’) which entangle the two sublattices, see e.g. Fig. \[fig:no\_entanglement\](a) or Fig. \[fig:perturbation\_states\]. Thus, it is indeed generically possible to create a B-magnon by acting on the A-sublattice. However, this does not seem to be true for diagonal magnons, i.e. when $|k_x| + |k_y| = \pi$. To make this precise, we introduce what we call the *sublattice spectral function*, $$\mathcal S_{A \to B}(\bm k, \omega) = \sum_\alpha \delta(\omega - (\omega_\alpha-\omega_0)) \; |\langle \alpha | \tilde S^-_{A,\bm k} |0\rangle|^2 \label{eq:S_ent}$$ where $\tilde S^-_{A,\bm k} = \sum \limits_{\bm r \in A} e^{i \bm{k\cdot r}} S^-_{\bm r}$. There are two crucial differences that distinguish it from the usual transverse spectral function as in Eq. . Firstly, we only act on the A-sublattice, where spins point down in the symmetry-broken ground state. Secondly, we act with the lowering operator $S^-$, putting us in the $S^z_\textrm{tot}$ sector of B-magnons. In Fig. \[fig:no\_entanglement\], we show this sublattice spectral function $\mathcal S_{A \to B}(\bm k,\omega)$. For convenience, we consider $\lambda = 0.95$ instead of the isotropic point, as this allows us to tell the one-magnon branch straightforwardly apart from the multi-magnon sector. We see the response is non-zero on almost the whole single-magnon branch (dashed lines). However, the spectral weight is exactly zero for any of the diagonal magnons (i.e. along the border of the shaded region in the Brillouin zone). We conclude that the diagonal magnons appear to be localized on their respective sublattices. In section \[subsec:states\] we gave a symmetry-based argument for the sublattice-localization within the perturbative framework, using the fact that we act with a single-site operator. However, we have also numerically confirmed that the same absence of spectral weight occurs even if we act with *multi-site* operators localized on the A-sublattice (not shown). We have not found an explanation for this and it would be interesting to study this in more detail. It is an open question whether there is a probe that could directly access $\mathcal S_{A \to B} (\bm k, \omega)$ in an experimental set-up. Quantitative analysis at diagonal momenta {#sec:roton} ========================================= In this section, we analyze the roton mode at $\bm k = (\pi,0) \cong X$ and its associated multi-magnon features in more quantitative detail, including a comparison to previous work. Depth of the anomalous mode at $k = (\pi,0)$ \[subsec:amplitude\] ----------------------------------------------------------------- ![Depth of the roton mode as a function of $\lambda$. Yellow dots are the DMRG results ($L_\textrm{circ}=10$). Dashed lines are series expansion results to various orders[@Singh95; @Zheng05]. We obtain the solid line by rewriting series expansions in terms of $\delta = 1-\sqrt{1-\lambda^2}$ and subsequently making a Padé approximant (details in main text). At the isotropic point ($\lambda=1$), we compare to QMC[@Sandvik01; @Shao17], CUT[@Powalski15; @Powalski17], ED[@Luescher09] and SWT[@Anderson52; @Kubo52; @Oguchi60; @Zheng91; @Hamer92; @Zheng93; @Syrom10]. \[fig:roton\_evolution\]](plot_evolution_dispersion_v13){width="0.85\linewidth"} In Fig. \[fig:roton\_evolution\], we consider the depth of the roton mode, i.e. the maximum of the dispersion at $\bm k = \left( \frac{\pi}{2}, \frac{\pi}{2} \right) \cong Y$ relative to the local minimum at $\bm k = (\pi,0) \cong X$. This is shown as a function of the parameter $\lambda$. As derived in section \[sec:perturbative\], we expect the dispersion to scale as $\sim \lambda^4$ for small $\lambda$. For that reason, we scale our axis accordingly. The numerical results obtained with our DMRG-based method for $L_\textrm{circ}= 10$ (up to $\chi=400$) are shown as yellow dots. At the isotropic point ($\lambda=1$), we plot the predictions of QMC[@Sandvik01; @Shao17] (extrapolated from up to $N =48 \times 48$), CUT[@Powalski15; @Powalski17], ED[@Luescher09] (extrapolated from up to $N = 36$) and SWT[@Anderson52; @Kubo52; @Oguchi60; @Zheng91; @Hamer92; @Zheng93; @Syrom10]. This quantity is extracted from the spectral function by fitting the single-magnon response with a Gaussian[^9]. Near the isotropic point, this fitting is somewhat subtle, as the one-magnon response is not easily separated from the multi-magnon weight. Fitting the (quasi-)bound states just above the single-magnon branch—which are highly relevant near $\bm k \approx (\pi,0)$—as well, we obtain results which are stable with respect to the numerical parameters. At the isotropic point, the method that DMRG is closest to is QMC. For $L_\textrm{circ} = 10$, we obtain $\varepsilon_{\frac{\pi}{2},\frac{\pi}{2}} \approx 2.40J$ and $\varepsilon_{\pi,0} \approx 2.06J$-$2.07J$. This can be compared to the QMC[@Shao17] extrapolations $\varepsilon_{\frac{\pi}{2},\frac{\pi}{2}} \approx 2.40J$ and $\varepsilon_{\pi,0} \approx 2.13J$. We are unable to perform a finite-circumference analysis, since for $L_\textrm{circ} = 6$ there are domain-wall excitations (wrapping around the circumference), while for $L_\textrm{circ} = 8$ the system is gapless such that we expect stronger finite-circumference effects. However, one can compare our results to the finite-size QMC data with linear dimension $10$, corresponding to our largest cylinder circumference. In that case, QMC obtains[@Shao17] $\varepsilon_{\pi,0} \approx 2.2 J$. Taking this to give a rough finite-size estimate, we note that we are within the same distance to the extrapolated QMC data (although at the opposite extreme). It is illuminating to not just focus on the isotropic case, but to track the evolution as a function of $\lambda$ as shown in Fig. \[fig:roton\_evolution\]. The dashed lines are from series expansions[@Singh95; @Zheng05] to different orders in $\lambda$. If we track the lowest-order result $\sim \lambda^4$ toward the isotropic point $\lambda =1$, we already obtain the correct order of magnitude. This is moreover in striking proximity to the CUT prediction. As we include higher order terms, we see that the dispersion gradually creeps up, showing no real sign of convergence. However, since SWT results are analytic in the modified parameter $\delta = 1-\sqrt{1-\lambda^2}$, it is suggestive to rewrite the series expansion in terms of $\delta$. Doing so, and building a Padé approximant out of it, we obtain the solid line in Fig. \[fig:roton\_evolution\]. We find that the Padé approximant is remarkably robust: the approximants $[3,3]$, $[4,2]$, $[5,2]$, $[4,3]$, $[3,4]$ all give virtually indistinguishable results! This stability suggests that the solid line could be a reasonable prediction for the true evolution of the dispersion as a function of $\lambda$. Exactly *at* the isotropic point, there is a small caveat: any power series in $\lambda$, when rewritten in terms of $\delta$, generically predicts a diverging slope at $\delta = 1$. Hence, also in this case, we find that the dashed curve is finite at $\delta = 1$ but its slope is vertical. It is not clear whether this particular feature is physical or not. Other than that, we expect that the Padé approximant should be reliable and we are encouraged by the fact that our numerical results agree so well with the Padé curve, indicating that finite-size corrections for $L_\textrm{circ} = 10$ are already rather small. It would be interesting to investigate to what extent the Padé approximant gives the correct prediction. In particular, it might be worthwhile to test and compare the other methods[^10] at $0 < \lambda <1$. Dispersion relation ------------------- ![Dispersion of the diagonal magnons. The solid black line is the DMRG result ($L_\textrm{circ}=10$). The green lines are the series expansion results: the dashed line[@Zheng05; @Oitmaa18] is the result to order $\lambda^{12}$ evaluated at $\lambda=1$, whereas the dotted green line is extracted from the plot in Ref. . We also compare to QMC[@Sandvik01; @Shao17] and CUT[@Powalski15; @Powalski17]. Inset: rescaled dispersions to compare functional forms. \[fig:roton\_comparison\]](roton_comparison_v4) Aside from studying the depth of the roton mode, we also consider its shape. Our numerical result is shown in Fig. \[fig:roton\_comparison\] (solid black line). We compare it to the functional forms obtained by CUT, QMC and series expansion. As discussed in section \[sec:perturbative\], the lowest-order non-trivial prediction from the Ising expansion is a simple cosine-like dispersion. This is not significantly altered at higher orders in $\lambda$ (or at the very least, only very slowly so), as was shown in Fig. \[fig:higherorder\]. The purpose of the inset is to show a comparison with this simple cosine. Remarkably, the CUT result is perfectly fit by a single harmonic. In conjunction with section \[subsec:amplitude\], we thus conclude that the fourth order extrapolation from the Ising expansion seems to be in striking proximity to the CUT prediction of the roton dispersion. Our result, on the other hand, while being dominated by the same cosine, also contains the higher harmonics (which are qualitatively generated in the Ising expansion). We point out that the QMC curve has more structure near $\bm k \approx X$, with a possible small subsidiary maximum at the X point itself; it would be interesting to investigate its origin. Multi-magnon features --------------------- Lastly, in Fig. \[fig:spectralcut\] we show a more detailed slice of the transverse spectral function $\mathcal S^t(\bm k,\omega)$ first shown in in Fig. \[fig:isotropic\]. At two values of the momentum, $\bm k = (\pi,0) \cong X$ and $\bm k = \left( \frac{\pi}{2} , \frac{\pi}{2} \right) \cong Y$, we show the spectral weight as a function of $\omega$. This numerical data is for the smaller circumference $L_\textrm{circ} = 8$, since as discussed in section \[sec:spectral\], in that case the system is gapless. Being gapless, one expects there to be more significant finite-size effects on a *quantitative* level, but the *qualitative* shape should look more like the 2D limit than the $L_\textrm{circ}=10$ data would. For either momentum, we clearly see the single-magnon peak (broadened due to our finite-time window, as explained in section \[sec:model\]) and a broad three-magnon continuum. Moreover, for $\bm k = (\pi,0) \cong X$, we recognize a second, smaller peak. This is a three-magnon (quasi-)bound state. We would like to comment on the following two features of Fig. \[fig:spectralcut\]. Firstly, there is considerable weight in the multi-magnon sector at $\bm k = (\pi,0)$, and not at $\bm k = \left( \frac{\pi}{2},\frac{\pi}{2} \right)$. Due to not having a tight grasp on finite-size effects for $L_\textrm{circ}=8$, we do not believe there is great value in quoting precise numbers, but at $\bm k = (\pi,0)$, only roughly half the weight is on the single magnon. Secondly, there is certainly a substructure to the multi-magnon weight. This seems to be in contrast to the featureless spectral function observed in a recent Monte Carlo study[@Shao17], which however considers the sum of the transverse and longitudinal spectral function. We do not expect the longitudinal contribution to completely smear out the substructure; in fact, the CUT analysis indicates the presence of strong resonances in the latter[@Powalski15; @Powalski17]. Due to the absence of a finite-size analysis, the substructure we observe is not conclusive and it would be interesting to investigate this further. ![Transverse spectral function $\mathcal S^t(\bm k, \omega)$ for the Heisenberg model ($\lambda = 1$) with $L_\textrm{circ} = 8$ and broadening $\sigma_\omega \approx 0.055J$. The second peak for $\bm k = (\pi,0) \cong X$ is the three-magnon (quasi-)bound state.\[fig:spectralcut\]](spectral_cut_L8) Conclusion {#sec:conclusion} ========== We have studied the spectral properties of the Heisenberg model, allowing for Ising anisotropy, using two complementary methods: a DMRG-based approach to obtain the unbiased dynamical structure factor (for certain circumferences), and a low-order perturbative expansion around the Ising limit to give insights into the physical mechanisms at play. One of our key messages is that some of its salient dynamical features are naturally accounted for starting from the Ising limit. The exactly soluble Ising limit itself has strongly-attractive magnon interactions based on simple domain-wall-counting. One does not expect such an Ising-based picture to be applicable to the *low-energy* hydrodynamic Goldstone modes as $\lambda \to 1$, but our work shows that it does remain relevant for the *high-energy* magnons. In particular, the lowest non-trivial order in the Ising expansion already captures the physics of the roton mode—i.e. the dispersion along $|k_x|+|k_y|=\pi$—at a semi-quantitative level. Furthermore, we clarified its physical origin in terms of the properties of virtual bound states which mediate the magnon’s hopping. On a more phenomenological level, the spectral function for the Heisenberg model obtained with DMRG shows that the anomalous local minimum at $\bm k = (\pi,0)$ grows monotonically when coming from the Ising limit. Moreover, even the strong continuum above this mode is continuously connected to spectral features near the Ising limit. The spectral function on the geometry with $L_\textrm{circ}=10$ directly supports the point of view that the physics near $\bm k \approx (\pi,0)$ is beyond that of a perturbatively-dressed magnon: in Fig. \[fig:isotropic\_log\] we saw that there is a continuum directly above the magnon which is *not* a standard kinematic three-magnon continuum. Instead, the relevant physics is due to (quasi-)bound states arising from attractive interactions of magnons ‘sharing’ their domain walls. This agrees with the insights of the CUT-based analysis[@Powalski15; @Powalski17]. It would be interesting to further explore the potential link between the interactions arising in the CUT framework and that of the Ising-based picture. Remarkably, as far as the roton mode is concerned, a low-order Ising expansion gives predictions close to that of the CUT approach, but at this point it is not clear whether this is accidental or not. Our study has also uncovered a curious spectral property. It turns out that magnons with $|k_x| + |k_y| = \pi$ are localized on their respective sublattices. This means that any operator localized on the A-sublattice cannot create a diagonal magnon associated with the B-sublattice. We have emphasized that this is rather unusual, since entanglement in the ground state would generically allow for it. Interestingly, such sublattice-localization is also predicted at low order in spin wave theory. It is not yet clear to what extent it is compatible with higher-order corrections in $1/S$. Having established a link between the spectral properties of the Heisenberg model and an Ising-based picture, several questions can be raised. Firstly, does the latter simple picture also give an intuitive explanation for *why* the (quasi-)bound states bunch up near $\bm k \approx (\pi,0)$? Secondly, as already mentioned in the introduction, the spectral features under discussion may not be uniquely described by one picture as opposed to another. Hence, we are not proposing the Ising-based picture as a complete framework, but rather as a very simple account of various features. Hence, it leaves open a link between the Ising limit and the other approaches that have been explored thus far. This could be interesting to investigate further. For example, a recent work interpreted the properties of the Heisenberg model in the context of the larger $J-Q$ model[@Shao17], and it could be worthwhile to explore its interplay with an Ising anisotropy. Finally, we note that although our analysis concerned zero temperature properties, much of the physics that we have discussed should stay relevant at low temperatures. In particular, while it is true that the Mermin-Wagner-Coleman theorem[@Mermin66; @Coleman73] prevents ordering at $T\neq 0$, it has been calculated that magnons with momentum $|\bm k| \gg 1/\xi(T)$ (where $\xi(T) \sim \exp(\textrm{const.}/T)$) remain well-defined[@Dyson56; @Harris71; @Kopietz90; @Ty90]. Efficiently extending the two-dimensional DMRG-based algorithm to finite temperatures remains a challenge for the future[@Vidal03; @Verstraete04; @White09; @Barthel09; @Stoudenmire10; @Berta17; @Hauschild2017]. Acknowledgements ================ The authors would like to thank Ehud Altman, Sylvain Capponi, Fabian Essler, Efstratios Manousakis, Kai Schmidt and Götz Uhrig for helpful discussions. We are particularly indebted to Alexander Chernyshev for various discussions and useful comments on the manuscript, and to Jan Oitmaa for sharing the series expansion data of Ref.  up to twelfth order. RV was supported by the German Research Foundation (DFG) through the Collaborative Research Centre SFB 1143 and FP acknowledges the support of the DFG Research Unit FOR 1807 through grants no. PO 1370/2-1, TRR80, the Nanosystems Initiative Munich (NIM) by the German Excellence Initiative, and the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program (grant agreement no. 771537). effective Hamiltonians to arbitrary order {#app:pert} ========================================= For completeness, we reproduce the general perturbative scheme that allows to obtain a well-defined effective Hamiltonian to any order. As in the main text, we consider a Hamiltonian of the form $H = H_0 + \lambda V$. Let $\mathcal H_0$ be the Hilbert space associated with the degenerate eigenvalue $E_0$ of $H_0$. Moreover, let $P_0$ be the projector onto $\mathcal H_0$, and $Q_0 = 1-P_0$. Note that we can decompose the total Hilbert space as $\mathcal H = \mathcal H_0 \oplus \mathcal H_0^\perp$. Suppose that for $0 \leq \lambda \leq \lambda_c$, we can decompose the Hilbert space into $\mathcal H = \mathcal H_\lambda \oplus \mathcal H_\lambda^\perp$ in such a way that 1. $\left(\mathcal H_{\lambda} \right)\large|_{\lambda=0} = \mathcal H_0$; 2. $\mathcal H_\lambda$ is a smooth function of $\lambda \in [0,\lambda_c]$; 3. the Hamiltonian respects the decomposition. Physically speaking, this formalizes the idea that we want the the energy levels of the sector we are interested to stay separated from the remaining levels; otherwise the idea of an effective Hamiltonian is misguided. If those conditions hold, the work by Kato[@Kato49] and Takashi[@Takahashi77] showed that for the same range $0 \leq \lambda \leq \lambda_c$, one can explicitly construct a smooth unitary mapping $\Gamma_\lambda: \mathcal H_0 \to \mathcal H$ which maps the unperturbed eigenstates into the perturbed ones. Hence, the desired effective Hamiltonian on $\mathcal H_0$ is then simply $H_\textrm{eff} := \Gamma_\lambda^\dagger H \Gamma_\lambda$. To perturbatively express $\Gamma_\lambda$ as a function of the known quantities $H_0$, $\lambda$ and $V_0$, it is useful to define a few other quantities. Firstly, let $P_\lambda :\mathcal H \to \mathcal H_\lambda$ be the projector onto $\mathcal H_\lambda$; we will derive a perturbative expression for this object as well. Secondly, define $$\begin{aligned} S^0 &:= - P_0, \\ S^k &:= \tilde G_0(E_0)^k := \left(Q_0 \frac{1}{E_0 - H_0} Q_0\right)^k \quad (k \neq 0 ).\end{aligned}$$ Note that $S^k$ is expressed in terms of *known* quantities. One can then derive[@Kato49; @Takahashi77] that $$P_\lambda P_0 = \sum_{n=0}^\infty \lambda^n \sum_{\substack{k_1+k_2+\cdots+k_n = n,\\k_i \geq 0}} S^{k_1} V S^{k_2} V \cdots S^{k_n} V P_0.$$ Moreover, it can be shown that the following function then has the desired properties: $$\Gamma_\lambda := P_\lambda P_0 \left( P_0 + \sum_{n=1}^\infty \frac{(2n-1)!!}{(2n)!!} [P_0 - P_0 P_\lambda P_0]^n \right).$$ It can be proven that $\Gamma_\lambda$ as thus defined indeed satisfies $\Gamma_\lambda^\dagger \Gamma_\lambda = P_0$. In terms of the above quantities, we thus have that $$\begin{aligned} H_\textrm{eff} &:= \Gamma_\lambda^\dagger H \Gamma_\lambda = E_0 P_0 + \lambda \; \Gamma_\lambda^\dagger V \Gamma_\lambda - \Gamma_\lambda^\dagger S^{-1} \Gamma_\lambda \; . \label{eq:appHeff}\end{aligned}$$ The result in Eq. , namely that $\Gamma_\lambda^\dagger H_0 \Gamma_\lambda = E_0 P_0 - \Gamma_\lambda^\dagger S^{-1} \Gamma_\lambda$, is a direct consequence of $H_0 = H_0 P_0 + H_0 Q_0 = E_0 P_0 + (E_0Q_0 - S^{-1}) = E_0 - S^{-1}$ and the fact that $\Gamma_\lambda^\dagger \Gamma_\lambda = P_0$. finite-circumference effects for symmetry-equivalent points \[app:analysis\] ============================================================================ ![For a given $L_\textrm{circ}$, we consider the DMRG result for the one-magnon dispersion $\varepsilon_{\bm k}$ and find the maximum energy difference between momenta $\{\bm k, \bm {k^*}\}$ for which $\varepsilon_{\bm k} = \varepsilon_{\bm {k^*}}$ in the 2D limit. \[fig:max\_symmetry\_inequiv\]](max_symmetry_equiv_v2){width="0.7\linewidth"} The two-dimensional models enjoys symmetries which are broken when putting the model on a cylinder. Rotating the square lattice by $90^\circ$ gives an example. One can turn this curse into a blessing, since it gives us a direct probe of the finite-circumference effects. More precisely, suppose $\bm k$ and $\bm{k^*}$ are two distinct momenta which are symmetry-equivalent in the two-dimensional limit, but which are *not* symmetry-equivalent on a cylinder with circumference $L_\textrm{circ}$. Hence, $|\varepsilon_{\bm k} - \varepsilon_{\bm{k^*}}|/J$ gives us a rough sense of how strong the finite-circumference effects are. Fig. \[fig:max\_symmetry\_inequiv\] plots the maximum of this over all pairs of symmetry-equivalent points. We see that it goes down as $L_\textrm{circ} \to \infty$, as expected. [^1]: To argue that , one can use that has a well-defined quantum number with respect to the *modified* translation operator . [^2]: One could instead work in a local frame where there is only one band, but this obscures some of the physics. In particular, then $S^z_\textrm{tot}$ is no longer a good quantum number. [^3]: To wit, the Holstein-Primakoff transformation is $S^+ = \sqrt{2S}\sqrt{1-\frac{a^\dagger a}{2S}} a$. The dominant term of the square root $\sqrt{1-\frac{a^\dagger a}{2S}}$ is of order $1/S^0$, which leads to LSWT. [^4]: and , where the integration is over $[-\pi,\pi]\times [-\pi,\pi]$. [^5]: The width of the Gaussian is chosen such that there is only little weight on the data generated by linear prediction. [^6]: More sophisticated extrapolation techniques could be and have been used for the low-energy modes near the isotropic point [@Singh95; @Zheng05], however in this work we focus on the high-energy modes. [^7]: Note that due to parity symmetry, the transverse spectral function does not pick up the two-magnon continuum. [^8]: For overlaps to be well-defined, one can apply the argument to finite systems. [^9]: Note that the width of the Gaussian is known; see section \[sec:spectral\]. [^10]: We note that applying the CUT method developed in Refs.  to $\lambda < 1$ would be distinct from the CUT method that was applied to the XXZ model in Ref. : in the latter case, the CUT method was perturbative in the parameter $\lambda$, hence being an alternative way of calculating the series expansion coefficients.

--- author: - | [**André Voros**]{}\ \ CEA, Service de Physique Théorique de Saclay\ CNRS URA 2306\ F-91191 Gif-sur-Yvette CEDEX (France)\ [ E-mail : [voros@spht.saclay.cea.fr]{}]{}\ \ and\ \ Institut de Mathématiques de Jussieu–Chevaleret\ CNRS UMR 7586\ Université Paris 7\ 2 place Jussieu, F-75251 Paris CEDEX 05 (France) title: | A sharpening of Li’s criterion\ for the Riemann Hypothesis --- [Exact and asymptotic formulae are displayed for the coefficients $\lambda _n$ used in Li’s criterion for the Riemann Hypothesis. In particular, we argue that if (and only if) the Hypothesis is true, $\lambda _n \sim n(A \log n +B)$ for $n \to \infty$ (with explicit $A>0$ and $B$). The approach also holds for more general zeta or $L$-functions. ]{} Alternative expressions are presented for the coefficients $\lambda _n$ introduced by Li [@LI; @BL] to recast the Riemann Hypothesis as $ \lambda_n >0 \ (\forall n)$: a first representation (\[LZS\]) as a finite oscillatory sum, then a closely related integral representation (\[LZI\]), upon which the saddle-point method finally yields definite $n \to +\infty$ asymptotic estimates in one of two sharply distinct forms, (\[RS\]) or (\[REs\]), depending on the falsity or truth of the Riemann Hypothesis. (Only the main ideas are indicated here.) Background. =========== The coefficients $\lambda_n$ are defined as $ \lambda_n = \sum\limits _\rho \, [1-(1-1/\rho)^n] $, or equivalently via the generating function $$\label{LId} {\d \over \d z} \log \Xi \Bigl( {1 \over 1-z}\Bigr) \equiv \sum_{n=1}^\infty \lambda _n z^{n-1} , \qquad \qquad \bigl( \Xi(s) = s(s-1) \Gamma (s/2) \pi^{-s/2} \zeta(s) \bigr).$$ The Riemann zeros are listed in pairs, as $$\label{ZER} \{\rho = \hf \pm \mi \tau_k \}_{k=1,2,\ldots}, \quad \{ \Re \, \tau_k \} \mbox{ positive and non-decreasing;}$$ sums and products over zeros are performed with $\rho$ and $(1-\rho)$ paired together, as usual; we parametrize each such pair by the single number $x_k=\rho(1-\rho)=\qt+{\tau_k}^2$. A “secondary" zeta function is $$\label{ZDef} Z(\sigma) = \sum_{k=1}^\infty {x_k}^{-\sigma}, \qquad \Re \sigma > \hf .$$ It extends to a meromorphic function in $\mathbb C$, and all its poles lie at the negative half-integers except $\sigma=+\hf$ [@K], which has the polar part [@Vz] $$\label{DPol} Z(\hf+\eps) = R_{-2} \,\eps^{-2} + R_{-1} \,\eps^{-1} + O(1)_{\eps \to 0}, \qquad R_{-2} = {1 \over 8\pi} , \quad R_{-1} = -{\log 2\pi \over 4\pi} .$$ Exact forms. ============ To reexpress the $\lambda _n$, we use a symmetrical Hadamard product form of $\Xi(s)$ [@Ed Sec.1.10][@Vz], namely $\Xi(s) = \prod_{k=1}^{\infty} \, \bigl[ 1 - s(1-s)/x_k \bigr]$, instead of the standard form [@Da chap. 12]. Hence, $$\label{HAL} \Xi \Bigl( {1 \over 1-z} \Bigr) = \prod_{k=1}^{\infty} \biggl[ 1 + {z \over (1-z)^2 x_k} \biggr], \qquad \log \Xi \Bigl( {1 \over 1-z} \Bigr) = - \sum_{j=1}^{\infty} {(-1)^j \over j} \biggl[ {z \over (1-z)^2 } \biggr] ^j Z(j);$$ then, expanding $(1-z)^{-2j}$ by the generalized binomial formula, reordering in powers of $z$ and identifying the output with (\[LId\]), we get as first result $$\label{LZS} \lambda_n = - n \sum_{j=1}^n {(-1)^j \over j}{n+j-1 \choose 2j-1} \, Z(j) .$$ Remark: other such linear relations already exist, involving the cumulants of the Stieltjes constants [@BL] and/or the sums ${\mathscr Z}_j = \sum_\rho \rho^{-j}$ instead of $Z(j)$; see [@Vz Sec.3.3] and references therein. An advantage of (\[LZS\]) is that the lesser known factors $Z(j)$ are positive, and smooth (they sample the holomorphic function $Z(\sigma)$). Still, this evaluation of $\lambda _n$ involves cancellations that increase with $n$. An integral representation equivalent to (\[LZS\]) simply by residue calculus, but much more manageable, is $$\label{LZI} \lambda_n = {(-1)^n n \,\mi \over \pi} \int_C I(\sigma) \,\d \sigma, \qquad I(\sigma) = { \Gamma (\sigma+n) \Gamma (\sigma-n) \over \Gamma (2 \sigma+1) } \, Z(\sigma) ,$$ where $C$ is a positive contour encircling (only) the subset of poles $\sigma=+1, \cdots, +n$ of the integrand $I(\sigma)$. Asymptotic forms. ================= The integral formula readily suggests an asymptotic ($n \to \infty$) evaluation by the classic [*saddle-point method*]{} [@Er Sec.2.5]. First, the contour $C$ is deformed in the direction of decreasing $|I(\sigma)|$ until it crosses the nearest saddle-points of $|I(\sigma)|$. Among these saddle-points, the one(s) where $|I(\sigma)|$ is the largest give the dominant large-$n$ contributions to the integral, through local formulae at each such saddle-point (the integrand itself may be asymptotically approximated as well: e.g., by a Stirling formula for $\Gamma (z)$). [*We caution that our subsequent assertions, involving an integrand which cannot be described in fully closed form, partly retain an experimental character and would warrant further confirmation.*]{} In the present problem and for large $n$, the landscape of the function $|I(\sigma)|$ near the contour $C$ is dominantly controlled by the $\Gamma $-ratio: the resulting deformation is a dilation of $C$ away from the segment $[1,n]$, then the nearest saddle-points can be of two types here (once $n$ is large enough). 1\) For $\sigma$ on the segment $(\hf,1)$, $ |I(\sigma)| \sim \pi \, [\sin \pi \sigma \, \Gamma (2 \sigma+1)]^{-1} n^{2\sigma -1} Z(\sigma)$ always has one [*real*]{} minimum $ \sigma_{\rm r}(n) $ (tending to $\hf$ as $n \to \infty$; other real saddle-points lie below $\sigma=\hf$, and are subdominant). 2\) There may exist [*complex*]{} saddle-points $\sigma_{\rm c}(n)$, with imaginary parts proportional to $n$; we may focus just on the upper half-plane, as the lower half-plane gives complex-conjugate (“c.c.") contributions. Wherever $Z(\sigma)$ is dominated by the term ${x_k}^{-\sigma}$ from one Riemann zero $x_k$ (naively, $x_1$), the saddle-point equation (in the Stirling approximation for the $\Gamma $-ratio) is $0= {\d \over \d \sigma} \log |I(\sigma)| \sim \log(\sigma^2-n^2)- 2 \log 2\sigma -\log x_k$, yielding the [*formal*]{} saddle-point location $$\label{CSP} %\sigma_{\rm c} (n) = {n \, \mi \over 2 \tau_k} . \sigma_{\rm c} (n) = n \, \mi \, / \, 2 \tau_k .$$ Here, any zero $x_k$ [*on*]{} the critical axis ($\tau_k$ real) yields a purely imaginary $\sigma_{\rm c}(n)$, not eligible: it lies outside the domain of convergence of (\[ZDef\]), and its contribution would be subdominant in any case. [*The discussion then fundamentally splits depending on the presence or absence of zeros off the critical axis.*]{} \[RH false\] If the zero $x_k$ lies [*off*]{} the critical axis (and selecting $\arg \tau_k >0$), the associated complex saddle-point $\sigma_{\rm c}(n)$ is relevant: it lies inside the domain of convergence $\{ \Re \sigma>\hf \}$ as soon as ${n \, |\Im 1/\tau_k| >1}$, and it will exponentially dominate the real saddle-point (as seen later): by the standard calculation, its asymptotic contribution to $\lambda _n$ is $\bigl[ (\tau_k+\mi/2 )/( \tau_k-\mi/2) \bigr] ^n$, which [*grows exponentially*]{} in modulus and fluctuates in phase. This result can also be rigorously confirmed directly from (\[LId\]): by a conformal mapping argument [@BL], the function ${\d \over \d z} \log \Xi \bigl( {1 \over 1-z}\bigr)$ has precisely the points $z_k={(\tau_k - \mi/2) (\tau_k + \mi/2)^{-1}}$ and $z_{-k}=\overline z_k$ as simple poles of residue 1 in the unit disk; by a general Darboux theorem [@Di chap. VII §2], this implies the asymptotic form $\sum_k (z_{\pm k})^{-n}$ for the Taylor coefficients $\lambda _n$ of that function, i.e., $$\label{RS} \lambda _n \sim \sum_{\{ \arg \tau_k >0 \}} \Bigl( {\tau_k+\mi/2 \over \tau_k-\mi/2} \Bigr) ^n \quad [{}+ {\rm c.c.}] \quad \pmod {o(\e^{cn}) \ \forall c>0}, \quad n \to \infty. \qquad \quad [{\rm RH\ false}]$$ (An infinite set of these zeros poses no extra problem: their contributions form an asymptotic sequence.) On the other hand, the Darboux approach does not resolve the case \[RH true\], in which the poles $z_{\pm k}$ all lie at the same (unit) distance, and cluster at $z=1$ ! \[RH true\] By contrast, the saddle-point analysis of (\[LZI\]) still appears to work. Now all the $\tau_k$ are real, $Z(\sigma)=O(Z(\Re \sigma) \, |\Im \sigma|^{-3/2})$ in $\{ \Re \sigma > \hf \}$, and the contour $C$ can be freely moved towards the boundary of the half-plane without crossing any of the $\sigma_{\rm c}(n)$ (all purely imaginary). Hence the only dominant saddle-point in this case is $\sigma_{\rm r}(n) \in (\hf,1)$; it is shaped by the double pole of $Z(\sigma)$ at $\hf$ (itself generated by the totality of Riemann zeros), in the form $ \sigma_{\rm r}(n) \sim \hf + {1 \over \log n}$. One may then proceed as usual in the quadratic approximation of $\log I(\sigma)$ around $ \sigma_{\rm r}(n) $, but this is not so fit for a confluent case ($\sigma=\hf$ is a singular point). Here, it is at once simpler and more accurate to keep on deforming a portion of the contour $C$ nearest to $ \sigma=\hf$ until it fully encircles this pole (now clockwise), and to note that the resulting additions to the integral are asymptotically negligible. Hence for \[RH true\], the result is $$\label{REs} \lambda _n \sim (-1)^n 2n \Res_{\sigma=1/2} \Bigl[ { \Gamma (\sigma+n) \Gamma (\sigma-n) \over \Gamma (2 \sigma+1) } Z(\sigma) \Bigr] \sim 2 \pi n \,\bigl[ 2 R_{- 2} \log n - 2 R_{-2} (1-\gamma ) + R_{-1} \bigl] \ ;$$ ($\gamma = $ Euler’s constant); specifically for the Riemann zeros, using the explicit values (\[DPol\]), $$\label{RER} \lambda _n \sim \hf n (\log n - \log 2\pi - 1+\gamma ), \qquad n \to \infty. \qquad \qquad [{\rm RH\ true}]$$ The form (\[REs\]) covers more general situations: the zeros of, e.g., Dedekind zeta functions and some Dirichlet $L$-functions [@VO], in their \[RH true\] case ($R_{-2}$ is always positive, in agreement with Li’s criterion). We see a good agreement of (\[RER\]) with numerical data [@M] for $n<3300$ – still a bit short to give full confidence in (\[RER\]), however. (This agreement is even improved in the mean if we include the contribution like (\[REs\]) but from the next pole, $$\delta \lambda _n = (-1)^n 2n \Res_{\sigma=0} \Bigl[ { \Gamma (\sigma+n) \Gamma (\sigma-n) \over \Gamma (2 \sigma+1) } Z(\sigma) \Bigr] = 2Z(0) = +{7 \over 4},$$ although this term should not be asymptotically meaningful – larger oscillatory terms are also present.) Finally, the result becomes even simpler for some [*pure linear combinations of the Stieltjes cumulants*]{}. The latter may be defined in degree $n$ as $$\label{SC} g_n^{\rm c} = (-1)^{n-1} {\d^n \over \d s^n} [\log(s \, \zeta(1+s)]_{s=0} = - \lim_{M \to +\infty} \Biggl\{ \sum_{m=1}^M {\Lambda(m) (\log m)^{n-1} \over m} -{ (\log M)^n \over n} \Biggl\}$$ (cf. formula (4.1) in [@BL]: this relates to the Euler factorization of $\zeta(s)$ over the primes). Our $g_n^{\rm c}$ corresponds to $(-1)^n (n-1)! \, \eta_{n-1}$ in [@BL]: e.g., $g_1^{\rm c} = \gamma = -\eta_0$. Theorem 2 in [@BL] evaluates the differences $(\lambda_n-S_n)$ where $S_n= -\sum\limits_{j=1}^n {n \choose j} \, \eta_{j-1} = \sum\limits_{j=1}^n {(-1)^{j-1} \over (j-1)!} {n \choose j}\, g_j^{\rm c}$. As a consequence for $n \to \infty$, (\[RER\]) implies $$S_n = o(n) ; \qquad \qquad \qquad \qquad \qquad \qquad \qquad [{\rm RH\ true}]$$ whereas in the opposite case, we expect $S_n \sim \lambda_n$ as in (\[RS\]): it will oscillate between [*exponentially large*]{} values, negative [*and positive*]{}, but appreciable in absolute size only beyond $n \approx \min\limits_{\{ \arg \tau_k >0 \}} \{ [\Im 1/\tau_k]^{-1} \}$. [00]{} E. Bombieri and J.C. Lagarias, Complements to Li’s criterion for the Riemann Hypothesis, J. Number Theory 77 (1999) 274–287. H. Davenport, Multiplicative Number Theory (3rd ed., revised by H.L. Montgomery), Graduate Texts in Mathematics 74, Springer-Verlag (2000). R.B. Dingle, Asymptotic Expansions: their Derivation and Interpretation, Academic Press (1973). H.M. Edwards, Riemann’s Zeta Function, Academic Press (1974). A. Erdélyi, Asymptotic Expansions, Dover (1956). N. Kurokawa, Parabolic components of zeta functions, Proc. Japan Acad. 64, Ser. A (1988) 21–24, and Special values of Selberg zeta functions, in: Algebraic K-theory and algebraic number theory (Proceedings, Honolulu 1987), M.R. Stein and R. Keith Dennis eds., Contemp. Math. 83, Amer. Math. Soc. (1989) 133–149. X.-J. Li, The positivity of a sequence of numbers and the Riemann Hypothesis, J. Number Theory 65 (1997) 325–333. K. Maślanka, Effective method of computing Li’s coefficients and their properties, submitted to Experimental Mathematics . A. Voros, Zeta functions for the Riemann zeros, Ann. Inst. Fourier, Grenoble 53 (2003) 665–699. A. Voros, Zeta functions over zeros of general zeta and $L$-functions, in: Zeta functions, topology and quantum physics (Proceedings, Osaka, March 2003), T. Aoki et al. eds., Kluwer Academic Publishers (to appear) .

--- abstract: | Consistent interactions that can be added to a two-dimensional, free abelian gauge theory comprising a special class of BF-type models and a collection of vector fields are constructed from the deformation of the solution to the master equation based on specific cohomological techniques. The deformation procedure modifies the Lagrangian action, the gauge transformations, as well as the accompanying algebra of the interacting model. PACS number: 11.10.Ef author: - | E. M. Cioroianu[^1] and S. C. Săraru[^2]\ Faculty of Physics, University of Craiova,\ 13 A. I. Cuza Str., Craiova 200585, Romania title: 'Two-dimensional interactions between a BF-type theory and a collection of vector fields ' --- Introduction ============ A key point in the development of the BRST formalism was its cohomological understanding, which allowed, among others, a useful investigation of many interesting aspects related to the perturbative renormalization problem [@4]–[@5], the anomaly-tracking mechanism [@5]–[@6], the simultaneous study of local and rigid invariances of a given theory [@7], as well as to the reformulation of the construction of consistent interactions in gauge theories [@7a]–[@7d] in terms of the deformation theory [@8a]–[@8c], or, actually, in terms of the deformation of the solution to the master equation. The scope of this paper is to investigate the consistent interactions that can be added to a free, abelian, two-dimensional gauge theory consisting of a collection of vector fields and a BF-type model [@13] involving a set of scalar fields, two collections of one-forms and a system of two-forms. This work enhances the previous Lagrangian [@mpla] and Hamiltonian [@ijmpa]–[@jhep] results on the study of self-interactions in certain classes of BF-type models. The resulting interactions are accurately described by a gauge theory with an open algebra of gauge transformations. The interacting model reveals a geometric interpretation in terms of a Poisson structure present in various models of two-dimensional gravity [@grav1]–[@grav4] and also some interesting algebraic features. The analysis of Poisson Sigma Models, including their relationship to two-dimensional gravity and the study of classical solutions, can be found in [@psm1]–[@psmn] (see also [@ikeda12]). Our strategy goes as follows. Initially, we determine in Section 2 the antifield-BRST symmetry of the free model, that splits as the sum between the Koszul-Tate differential and the exterior derivative along the gauge orbits, $s=\delta +\gamma $. Next, in Section 3 we determine the consistent deformations of the solution to the master equation for the free model. The first-order deformation belongs to the local cohomology $H^{0}(s|d)$, where $% d$ is the exterior space-time derivative. The computation of the cohomological space $H^{0}(s|d)$ proceeds by expanding the co-cycles according to the antighost number, and by further using the cohomological groups $H(\gamma )$ and $H(\delta |d)$. We find that the first-order deformation is parametrized by some functions of the undifferentiated scalar fields, which become restricted to fulfill certain equations in order to produce a consistent second-order deformation. With the help of these equations, we then infer that the remaining deformations, of order three and higher, can be taken to vanish. The identification of the interacting model is developed in Section 4. The cross-couplings between the collection of vector fields and the BF field spectrum are described, among others, by generalized cubic and quartic Yang-Mills vertices in some “backgrounds” of scalar fields. Meanwhile, both the gauge transformations corresponding to the coupled model and their algebra are deformed with respect to the initial abelian theory in such a way that the new gauge algebra becomes open. Section 5 comments on two classes of solutions to the equations satisfied by the various functions of the scalar fields (that parametrize the deformed solution to the master equation) and Section 6 closes the paper with the main conclusions. Free model. Antibracket-antifield BRST symmetry =============================================== The starting point is represented by the free Lagrangian action $$S_{0}\left[ A_{\mu }^{a},H_{\mu }^{a},\varphi _{a},B_{a}^{\mu \nu },V_{\mu }^{A}\right] =\int d^{2}x\left( H_{\mu }^{a}\partial ^{\mu }\varphi _{a}+% \frac{1}{2}B_{a}^{\mu \nu }F_{\mu \nu }^{\prime a}-\frac{1}{4}F_{\mu \nu }^{A}F_{A}^{\mu \nu }\right) , \label{f1}$$ where we used the notations $$F_{\mu \nu }^{A}=\partial _{[\mu }V_{\nu ]}^{A},\;F_{\mu \nu }^{\prime a}=\partial _{[\mu }A_{\nu ]}^{a}, \label{f1a}$$ and the symbol $\left[ \mu \nu \cdots \right] $ denotes full antisymmetrization with respect to the indices between brackets, but without normalization factors. Capital Latin indices $A$, $B$, etc., are raised with a constant, symmetric and field-independent, non-degenerate matrix $k^{AB}$. We observe that (\[f1\]) is written like a sum between the action of a two-dimensional abelian BF theory (involving two sets of one-forms $\left\{ A_{\mu }^{a},H_{\mu }^{a}\right\} $, a collection of scalar fields $\left\{ \varphi _{a}\right\} $ and a sequence of two forms $\left\{ B_{a}^{\mu \nu }\right\} $) and the action corresponding to a set of abelian vector fields $% \left\{ V_{\mu }^{A}\right\} $. The collection indices $a$, $b$, etc., and respectively $A$, $B$, etc., are assumed to run independently ones to the others. A generating set of gauge transformations for the action (\[f1\]) can be taken under the form $$\begin{aligned} \delta _{\epsilon }A_{\mu }^{a} &=&\partial _{\mu }\epsilon ^{a},\;\delta _{\epsilon }H_{\mu }^{a}=\partial ^{\nu }\epsilon _{\mu \nu }^{a}, \label{f2} \\ \delta _{\epsilon }\varphi _{a} &=&0=\delta _{\epsilon }B_{a}^{\mu \nu },\;\delta _{\epsilon }V_{\mu }^{A}=\partial _{\mu }\epsilon ^{A}, \label{f3}\end{aligned}$$ where all the gauge parameters are bosonic, with $\epsilon _{\mu \nu }^{a}$ antisymmetric in their Lorentz indices. In $D=2$ spacetime dimensions, where this model evolves, the abelian gauge transformations (\[f2\]–\[f3\]) are irreducible. In conclusion, (\[f1\]) and (\[f2\]–\[f3\]) describe a linear (the field equations are linear in the fields) gauge theory of Cauchy order equal to two. In order to construct the antifield-BRST symmetry for the free gauge theory under study, we need to identify the algebra on which the BRST differential acts. The generators of the BRST algebra are, besides the original bosonic fields $$\Phi ^{\alpha _{0}}=\left( A_{\mu }^{a},H_{\mu }^{a},\varphi _{a},B_{a}^{\mu \nu },V_{\mu }^{A}\right) , \label{f10a}$$ the fermionic ghosts $$\eta ^{\alpha _{1}}=\left( \eta ^{a},\eta _{\mu \nu }^{a},C^{A}\right) , \label{f10b}$$ respectively associated with the gauge parameters from (\[f2\]–\[f3\]), as well as their antifields $$\Phi _{\alpha _{0}}^{*}=\left( A_{a}^{*\mu },H_{a}^{*\mu },\varphi ^{*a},V_{A}^{*\mu },B_{\mu \nu }^{*a}\right) ,\;\eta _{\alpha _{1}}^{*}=\left( \eta _{a}^{*},\eta _{a}^{*\mu \nu },C_{A}^{*}\right) . \label{f11}$$ The Grassmann parity of an antifield is opposite to that of the corresponding field/ ghost. It is understood that $\eta _{\mu \nu }^{a}$ and $\eta _{a}^{*\mu \nu }$ are antisymmetric, just like the gauge parameters $% \epsilon _{\mu \nu }^{a}$. Since the gauge generators of this model are field-independent, it follows that the BRST differential $s$ simply reduces to $$s=\delta +\gamma , \label{f11a}$$ where $\delta $ represents the Koszul-Tate differential, graded by the antighost number $\mathrm{agh}$ ($\mathrm{agh}\left( \delta \right) =-1$), and $\gamma $ stands for the exterior derivative along the gauge orbits, whose degree is named pure ghost number $\mathrm{pgh}$ ($\mathrm{pgh}\left( \gamma \right) =1$). Naturally, these two degrees do not interfere ($\mathrm{% agh}\left( \gamma \right) =0$, $\mathrm{pgh}\left( \delta \right) =0$). The overall degree that grades the BRST complex, known as the ghost number ($% \mathrm{gh}$), is defined like the difference between the pure ghost number and the antighost number, such that $\mathrm{gh}\left( s\right) =\mathrm{gh}% \left( \delta \right) =\mathrm{gh}\left( \gamma \right) =1$. According to the standard rules of the BRST method, the corresponding degrees of the generators from the BRST complex are valued like $$\begin{aligned} \mathrm{pgh}\left( \Phi ^{\alpha _{0}}\right) &=&\mathrm{pgh}\left( \Phi _{\alpha _{0}}^{*}\right) =0,\;\mathrm{pgh}\left( \eta ^{\alpha _{1}}\right) =1,\;\mathrm{pgh}\left( \eta _{\alpha _{1}}^{*}\right) =0, \label{f8} \\ \mathrm{agh}\left( \Phi ^{\alpha _{0}}\right) &=&0,\;\mathrm{agh}\left( \Phi _{\alpha _{0}}^{*}\right) =1,\;\mathrm{agh}\left( \eta ^{\alpha _{1}}\right) =0,\;\mathrm{agh}\left( \eta _{\alpha _{1}}^{*}\right) =2. \label{f9}\end{aligned}$$ Actually, (\[f11a\]) is a decomposition of the BRST differential according to the antighost number and it shows that $s$ contains only components of antighost number equal to minus one and zero. Consequently, the equation expressing the nilpotency of $s$ projected on the distinct values of the antighost number is equivalent with the nilpotency and anticommutation of its components $$s^{2}=0\Leftrightarrow \left( \delta ^{2}=0,\;\delta \gamma +\gamma \delta =0,\;\gamma ^{2}=0\right) . \label{f11b}$$ The Koszul-Tate differential is imposed to realize an homological resolution of the algebra of smooth functions defined on the stationary surface of the field equations for the action (\[f1\]), while the exterior longitudinal derivative is related to the gauge symmetries (\[f2\]–\[f3\]) through its cohomology at pure ghost number zero computed in the cohomology of $% \delta $, which is required to be the algebra of physical observables for the free model under consideration. The actions of $\delta $ and $\gamma $ on the generators from the BRST complex, which enforce all the above mentioned properties, are given by $$\delta \Phi ^{\alpha _{0}}=0=\delta \eta ^{\alpha _{1}}, \label{f12}$$ $$\delta A_{a}^{*\mu }=\partial _{\alpha }B_{a}^{\alpha \mu },\;\delta H_{a}^{*\mu }=-\partial ^{\mu }\varphi _{a},\;\delta \varphi ^{*a}=\partial ^{\mu }H_{\mu }^{a},\;\delta B_{\mu \nu }^{*a}=-\frac{1}{2}F_{\mu \nu }^{\prime a}, \label{f13}$$ $$\delta V_{A}^{*\mu }=-\partial _{\alpha }F_{A}^{\alpha \mu },\;\delta \eta _{a}^{*}=-\partial _{\mu }A_{a}^{*\mu },\;\delta \eta _{a}^{*\mu \nu }=\frac{% 1}{2}\partial ^{[\mu }H_{a}^{*\nu ]},\;\delta C_{A}^{*}=-\partial _{\mu }V_{A}^{*\mu }, \label{f14}$$ $$\gamma \Phi _{\alpha _{0}}^{*}=0=\gamma \eta _{\alpha _{1}}^{*},\;\gamma V_{\mu }^{A}=\partial _{\mu }C^{A}, \label{f15}$$ $$\gamma A_{\mu }^{a}=\partial _{\mu }\eta ^{a},\;\gamma H_{\mu }^{a}=\partial ^{\nu }\eta _{\mu \nu }^{a},\;\gamma \varphi _{a}=\gamma B_{a}^{\mu \nu }=\gamma \eta ^{\alpha _{1}}=0. \label{f16}$$ The Lagrangian BRST differential admits a canonical action in a structure named antibracket and defined by decreeing the fields/ghosts conjugated with the corresponding antifields, $s\cdot =\left( \cdot ,S\right) $, where $% \left( ,\right) $ signifies the antibracket and $S$ denotes the canonical generator of the BRST symmetry. It is a bosonic functional of ghost number zero, involving both field/ghost and antifield spectra, that obeys the master equation $$\left( S,S\right) =0. \label{mastorig}$$ The master equation is equivalent with the second-order nilpotency of $s$, where its solution $S$ encodes the entire gauge structure of the associated theory. Taking into account the formulas (\[f12\]–\[f16\]), as well as the standard actions of $\delta $ and $\gamma $ in canonical form, we find that the complete solution to the master equation for the model under study reads $$\begin{aligned} S &=&S_{0}\left[ A_{\mu }^{a},H_{\mu }^{a},\varphi _{a},B_{a}^{\mu \nu },V_{\mu }^{A}\right] \nonumber \\ &&+\int d^{2}x\left( A_{a}^{*\mu }\partial _{\mu }\eta ^{a}+H_{a}^{*\mu }\partial ^{\nu }\eta _{\mu \nu }^{a}+V_{A}^{*\mu }\partial _{\mu }C^{A}\right) , \label{f5}\end{aligned}$$ such that it contains pieces of antighost number zero and one. The absence of components with antighost numbers higher than one is due to the abelianity and irreducibility of the chosen generating set of gauge transformations. If the gauge algebra were non-abelian, then the solution to the master equation would also include terms of antighost number two that are quadratic in the ghosts (\[f10b\]): ones linear in the antifields $% \eta _{\alpha _{1}}^{*}$ and proportional with the structure functions appearing at the commutators between the gauge generators, and others quadratic in the antifields $\Phi _{\alpha _{0}}^{*}$. The latter kind of elements is present only if these commutators strictly close on-shell, *i.e.*, on the stationary surface of field equations, or, which is the same, only if the gauge algebra is open. In the case where the gauge algebra is open, the solution to the master equation may in principle continue to be non-vanishing at antighost numbers higher than two, the corresponding elements being related to the higher-order structure functions and to the identities satisfied by them. The main ingredients of the antifield-BRST symmetry derived in this section will be useful in the sequel at the analysis of consistent interactions that can be added to the action (\[f1\]) without changing its number of independent gauge symmetries. Deformation of the solution to the master equation ================================================== General idea ------------ A consistent deformation of the free action (\[f1\]) and of its gauge invariances (\[f2\]–\[f3\]) defines a deformation of the corresponding solution to the master equation that preserves both the master equation and the field/ antifield spectra. Let us denote by $g$ the coupling constant and assume that the local functional $\bar{S}_{0}\left[ A_{\mu }^{a},H_{\mu }^{a},\varphi _{a},B_{a}^{\mu \nu },V_{\mu }^{A}\right] =S_{0}+g\int d^{2}xa_{0}+g^{2}\int d^{2}xb_{0}+O\left( g^{3}\right) $ represents a consistent deformation of (\[f1\]), subject to the deformed gauge transformations $\bar{\delta}_{\epsilon }A_{\mu }^{a}=\partial _{\mu }\epsilon ^{a}+g\lambda _{\mu }^{a}+g^{2}\lambda _{\mu }^{\prime a}+O\left( g^{3}\right) $, $\bar{\delta}_{\epsilon }H_{\mu }^{a}=\partial ^{\nu }\epsilon _{\mu \nu }^{a}+g\rho _{\mu }^{a}+g^{2}\rho _{\mu }^{\prime a}+O\left( g^{3}\right) $, $\bar{\delta}_{\epsilon }\varphi _{a}=g\sigma _{a}+g^{2}\sigma _{a}^{\prime }+O\left( g^{3}\right) $, $\bar{\delta}% _{\epsilon }B_{a}^{\mu \nu }=g\sigma _{a}^{\mu \nu }+g^{2}\sigma _{a}^{\prime \mu \nu }+O\left( g^{3}\right) $ and $\bar{\delta}_{\epsilon }V_{\mu }^{A}=\partial _{\mu }\epsilon ^{A}+gv_{\mu }^{A}+g^{2}v_{\mu }^{\prime A}+O\left( g^{3}\right) $ (by consistent we mean that $\bar{S}_{0}$ is invariant under the modified gauge transformations $\bar{\delta}% _{\epsilon }\Phi ^{\alpha _{0}}$ at all orders in the coupling constant). Accordingly, we find that $$\bar{S}=S+g\int d^{2}xa+g^{2}\int d^{2}xb+O\left( g^{3}\right) , \label{f4}$$ is a consistent deformed solution to the master equation for the interacting theory, *i.e.*, it satisfies the equation $$\left( \bar{S},\bar{S}\right) =0, \label{mastdef}$$ at all orders in the coupling constant (with $S$ given by (\[f5\])). Moreover, the objects $a$ and $b$ start like $$\begin{aligned} a &=&a_{0}+A_{a}^{*\mu }\bar{\lambda}_{\mu }^{a}+H_{a}^{*\mu }\bar{\rho}% _{\mu }^{a}+V_{A}^{*\mu }\bar{v}_{\mu }^{A}+\varphi ^{*a}\bar{\sigma}% _{a}+B_{\mu \nu }^{*a}\bar{\sigma}_{a}^{\mu \nu }+``\mathrm{more}", \label{f6} \\ b &=&b_{0}+A_{a}^{*\mu }\bar{\lambda}_{\mu }^{\prime a}+H_{a}^{*\mu }\bar{% \rho}_{\mu }^{\prime a}+V_{A}^{*\mu }\bar{v}_{\mu }^{\prime A}+\varphi ^{*a}% \bar{\sigma}_{a}^{\prime }+B_{\mu \nu }^{*a}\bar{\sigma}_{a}^{\prime \mu \nu }+``\mathrm{more}", \label{f7}\end{aligned}$$ where the “bar” quantities are obtained by replacing the gauge parameters $% \epsilon ^{a}$, $\epsilon _{\mu \nu }^{a}$ and $\epsilon ^{A}$ respectively with the fermionic ghosts $\eta ^{a}$, $\eta _{\mu \nu }^{a}$ and $C^{A}$ in the functions $\lambda _{\mu }^{a}$, $\lambda _{\mu }^{\prime a}$, $\rho _{\mu }^{a}$, $\rho _{\mu }^{\prime a}$, $\sigma _{a}$, $\sigma _{a}^{\prime }$, $\sigma _{a}^{\mu \nu }$, $\sigma _{a}^{\prime \mu \nu }$, $v_{\mu }^{A}$ and $v_{\mu }^{\prime A}$ contained in the deformed gauge transformations. First-order deformation ----------------------- ### Basic cohomological results Using the development (\[f4\]) and the Eq. (\[mastorig\]) satisfied by $S$, we obtain that the master equation (\[mastdef\]) of the deformed theory holds to order $g$ if and only if $$sa=\partial _{\mu }j^{\mu }, \label{f17}$$ for some local $j^{\mu }$. This means that the non-integrated density of the first-order deformation of the solution to the master equation, $a$, belongs to the local cohomology of the BRST differential, $H^{0}\left( s|d\right) $, where $d$ is the exterior spacetime derivative. In the case where $a$ is a $s $ coboundary modulo $d$ ($a=sc+\partial _{\mu }k^{\mu }$), then the deformation is trivial (it can be eliminated by a redefinition of the fields). As a consequence, $a$ is unique only up to replacing it with an element from the same cohomological class, $a\rightarrow a+sc+\partial _{\mu }k^{\mu }$, and, on the other hand, if $a$ is purely trivial, $a=sc+\partial _{\mu }k^{\mu }$, then it can be removed from $\bar{S}$ by setting $a=0$. For obvious reasons, we are interested only in smooth, local, Lorentz-covariant and Poincaré-invariant deformations. In order to investigate the solution to (\[f17\]), we develop $a$ according to the antighost number $$a=a_{0}+a_{1}+\cdots +a_{J},\;\mathrm{agh}\left( a_{k}\right) =k=\mathrm{pgh}% \left( a_{k}\right) ,\;\varepsilon \left( a_{k}\right) =0, \label{f18}$$ and assume that the expansion (\[f18\]) stops at a finite, but otherwise arbitrary, value of the antighost number, $J$. (The notation $\varepsilon \left( F\right) $ signifies the Grassmann parity of $F$.) This result can be argued like in  [@gen2] (Section 3), under the sole assumption that the interacting Lagrangian at the first order in the coupling constant, $a_{0}$, is local, so it contains a finite, but otherwise arbitrary, number of derivatives. Replacing (\[f18\]) into the Eq. (\[f17\]) and taking into account the decomposition (\[f11a\]), we obtain that the Eq. (\[f17\]) is equivalent to a tower of local equations, corresponding to the different decreasing values of the antighost number $$\begin{aligned} \gamma a_{J} &=&\partial _{\mu }j_{J}^{\mu }, \label{f18b} \\ \delta a_{J}+\gamma a_{J-1} &=&\partial _{\mu }j_{J-1}^{\mu }, \label{f18c} \\ \delta a_{k}+\gamma a_{k-1} &=&\partial _{\mu }j_{k-1}^{\mu },\;J-1\geq k\geq 1, \label{f18d}\end{aligned}$$ where $\left( j_{k}^{\mu }\right) _{k=\overline{0,J}}$ are some local currents with $\mathrm{agh}\left( j_{k}^{\mu }\right) =k$. As $\mathrm{pgh}% \left( a_{J}\right) =J$, the Eq. (\[f18b\]) shows that $a_{J}$ belongs to the local cohomology of the exterior derivative along the gauge orbits at pure ghost number $J$, $H^{J}\left( \gamma |d\right) $. Following a reasoning similar to that from [@gen2]–[@lingr], it can be shown that one can replace the Eq. (\[f18b\]) at strictly positive antighost numbers with $$\gamma a_{J}=0,\;J>0. \label{f18a}$$ In other words, for $J>0$ the last representative from (\[f18\]) can always be considered to pertain to the cohomological group of the exterior derivative along the gauge orbits at pure ghost number $J$, $H^{J}\left( \gamma \right) $. As a consequence, it is unique up to $\gamma $-exact contributions, $a_{J}\rightarrow a_{J}+\gamma c_{J}$, while the purely $% \gamma $-trivial solutions $a_{J}=\gamma c_{J}$ can be safely removed by taking $a_{J}=0$. In conclusion, the Eq. (\[f17\]) associated with the local form of the first-order deformation is completely equivalent to the tower of Eqs. (\[f18a\]) and (\[f18c\]–\[f18d\]). Thus, we need to know the cohomology of $\gamma $, $H\left( \gamma \right) $, in order to determine the terms of highest antighost number in $a$. From the definitions (\[f15\]–\[f16\]) it is simple to see that this cohomology is generated by $F_{\mu \nu }^{A}$, $F_{\mu \nu }^{\prime a}$, $% \varphi _{a}$, $B_{a}^{\mu \nu }$, $\partial ^{\mu }H_{\mu }^{a}$, by the antifields $\left\{ \Phi _{\alpha _{0}}^{*}\right\} $, by all their derivatives, as well as by the undifferentiated ghosts $\eta ^{\alpha _{1}}$. (The derivatives of the ghosts are $\gamma $-exact, as can be observed from the last relation in (\[f15\]) and the first two formulas in (\[f16\]), so we can discard them as being trivial in $H\left( \gamma \right) $.) If we denote by $e^{M}\left( \eta ^{\alpha _{1}}\right) $ the elements of pure ghost number equal to $M$ of a basis in the space of polynomials in $% \eta ^{\alpha _{1}}$, which is finite-dimensional due to the anticommutation of the ghosts, it follows that the general local solution to the Eq. (\[f18a\]) takes the form (up to irrelevant, $\gamma $-exact contributions) $$a_{J}=\alpha _{J}\left( \left[ F_{\mu \nu }^{A}\right] ,\left[ F_{\mu \nu }^{\prime a}\right] ,\left[ \varphi _{a}\right] ,\left[ B_{a}^{\mu \nu }\right] ,\left[ \partial ^{\mu }H_{\mu }^{a}\right] ,\left[ \Phi _{\alpha _{0}}^{*}\right] ,\left[ \eta _{\alpha _{1}}^{*}\right] \right) e^{J}\left( \eta ^{\alpha _{1}}\right) ,\;J>0, \label{f19}$$ where $\mathrm{agh}\left( \alpha _{J}\right) =J$ for $a_{J}$ to have the ghost number equal to zero, and $\alpha _{J}$ must display the same Grassmann parity like $e^{J}$ in order to ensure that $a_{J}$ is bosonic. Here and in the sequel the notation $f\left( \left[ q\right] \right) $ signifies that $f$ depends on $q$ and its spacetime derivatives up to a finite order. The index-notation $J$ is generic, in the sense that it may include unspecified Lorentz and/or collection indices. As they have both finite antighost number and derivative order, the elements $\alpha _{J}$, which are non-trivial in $H^{0}\left( \gamma \right) $, are polynomials in the antifields, their derivatives, and in the allowed derivatives of the fields, but may contain an indefinite number of undifferentiated fields $% \varphi _{a}$ and $B_{a}^{\mu \nu }$. They will be called “invariant polynomials”. At zero antighost number, the invariant polynomials are nothing but the local, gauge-invariant quantities of the free model under study. The fact that we can replace the Eq. (\[f18b\]) for $J>0$ with (\[f18a\]) is a consequence of the triviality of the cohomology of the exterior spacetime differential in the space of invariant polynomials at strictly positive antighost numbers. So, if $\alpha _{J}$ is an invariant polynomial with $\mathrm{agh}\left( \alpha _{J}\right) =J>0$ that is $d$-closed, $d\alpha _{J}=0$, then $\alpha _{J}=d\beta _{J}$, with $\beta _{J}$ also an invariant polynomial. Inserting the expression (\[f19\]) into the Eq. (\[f18c\]) and recalling the definitions (\[f15\]–\[f16\]), we find that a necessary condition for the existence of (non-trivial) solutions $a_{J-1}$ is that the invariant polynomials $\alpha _{J}$ are (non-trivial) elements from the local cohomology group of the Koszul-Tate differential at pure ghost number zero and at strictly positive antighost number $J$, $H_{J}\left( \delta |d\right) $[^3] $$\delta \alpha _{J}=\partial _{\mu }k_{J-1}^{\mu },\;\mathrm{agh}\left( k_{J-1}^{\mu }\right) =J-1\geq 0. \label{f19a}$$ By “trivial elements of $H_{J}\left( \delta |d\right) $” we understand $% \delta $-exact modulo $d$ objects, hence of the form $\delta d_{J+1}+\partial _{\mu }m_{J}^{\mu }$. As a consequence of (\[f19a\]), we need to investigate some of the main properties of the cohomology $H\left( \delta |d\right) $ at strictly positive antighost numbers in order to fully determine the component $a_{J}$ of highest antighost number from the first-order deformation. As we have discussed in Section 2, the free model under study is a linear gauge theory of Cauchy order equal to two. In agreement with the general results from [@gen1] (also see [@gen2]–[@lingr]), one can state that $H\left( \delta |d\right) $ (at pure ghost number zero) is trivial at antighost numbers strictly greater than its Cauchy order. The same result holds for the local cohomology of the Koszul-Tate differential in the space of invariant polynomials, $H^{\mathrm{% inv}}\left( \delta |d\right) $, so we actually have that $$H_{J}\left( \delta |d\right) =0,\;H_{J}^{\mathrm{inv}}\left( \delta |d\right) =0,\;J>2. \label{f19b}$$ An element of $H_{J}^{\mathrm{inv}}\left( \delta |d\right) $ is defined via an equation similar to (\[f19a\]), but with the corresponding current $% k_{J-1}^{\mu }$ an invariant polynomial. Moreover, it can be shown [@gen2]–[@lingr] that if the invariant polynomial $\alpha _{J}$ with $% \mathrm{agh}\left( \alpha _{J}\right) =J\geq 2$ is trivial in $H_{J}\left( \delta |d\right) $, then it can be taken to be trivial also in $H_{J}^{% \mathrm{inv}}\left( \delta |d\right) $, *i.e.*, $$\left( \alpha _{J}=\delta d_{J+1}+\partial _{\mu }m_{J}^{\mu },\;\mathrm{agh}% \left( \alpha _{J}\right) =I\geq 2\right) \Rightarrow \alpha _{J}=\delta \beta _{J+1}+\partial _{\mu }\gamma _{J}^{\mu }, \label{f19c}$$ with both $\beta _{J+1}$ and $\gamma _{J}^{\mu }$ invariant polynomials. With the help of the definitions (\[f12\]–\[f14\]), we find that the most general non-trivial representative from $H_{2}\left( \delta |d\right) $, which, essentially, has the same status in $H_{2}^{\mathrm{inv}}\left( \delta |d\right) $, is $$\alpha _{2}^{0}=K^{\Delta }M_{\Delta }+K^{\Delta ^{\prime }\mu \nu }N_{\Delta ^{\prime }\mu \nu }+K^{\Delta ^{\prime \prime }}P_{\Delta ^{\prime \prime }}, \label{f20}$$ where $M_{\Delta }$, $N_{\Delta ^{\prime }\mu \nu }$ and $P_{\Delta ^{\prime \prime }}$ are respectively given by $$\begin{aligned} M_{\Delta } &=&\frac{\partial W_{\Delta }}{\partial \varphi _{m}}\eta _{m}^{*}-\frac{\partial ^{2}W_{\Delta }}{\partial \varphi _{m}\partial \varphi _{n}}\left( B_{m}^{\mu \nu }\eta _{n\mu \nu }^{*}+H_{m\mu }^{*}A_{n}^{*\mu }\right) \nonumber \\ &&-\frac{1}{2}\frac{\partial ^{3}W_{\Delta }}{\partial \varphi _{m}\partial \varphi _{n}\partial \varphi _{p}}H_{m\mu }^{*}H_{n\nu }^{*}B_{p}^{\mu \nu }, \label{f21}\end{aligned}$$ $$N_{\Delta ^{\prime }\mu \nu }=\frac{\partial U_{\Delta ^{\prime }}}{\partial \varphi _{m}}\eta _{m\mu \nu }^{*}+\frac{1}{2}\frac{\partial ^{2}U_{\Delta ^{\prime }}}{\partial \varphi _{m}\partial \varphi _{n}}H_{m\mu }^{*}H_{n\nu }^{*}, \label{f22}$$ $$P_{\Delta ^{\prime \prime }}=f_{\Delta ^{\prime \prime }}^{A}C_{A}^{*}+\frac{% \partial f_{\Delta ^{\prime \prime }}^{A}}{\partial \varphi _{m}}\left( V_{A}^{*\mu }H_{m\mu }^{*}+\eta _{m\mu \nu }^{*}F_{A}^{\mu \nu }\right) +% \frac{1}{2}\frac{\partial ^{2}f_{\Delta ^{\prime \prime }}^{A}}{\partial \varphi _{m}\partial \varphi _{n}}H_{m\mu }^{*}H_{n\nu }^{*}F_{A}^{\mu \nu }, \label{f23}$$ and $K^{\Delta }$, $K^{\Delta ^{\prime }\mu \nu }$ and $K^{\Delta ^{\prime \prime }}$ denote some constant coefficients, with $K^{\Delta ^{\prime }\mu \nu }$ antisymmetric in their Lorentz indices. The generic indices $\Delta $, $\Delta ^{\prime }$ and $\Delta ^{\prime \prime }$ are exclusively composed of collection indices (of the type $a$, $b$, etc., and/or $A$, $B$, etc.). All the functions $W_{\Delta }$, $U_{\Delta ^{\prime }}$ and $% f_{\Delta ^{\prime \prime }}^{A}$ involved in (\[f21\]–\[f23\]) depend in an arbitrary manner on the undifferentiated scalar fields $\varphi _{a}$. Moreover, the objects $M_{\Delta }$, $N_{\Delta ^{\prime }\mu \nu }$ and $% P_{\Delta ^{\prime \prime }}$ separately satisfy the equations $$\delta M_{\Delta }=\partial _{\mu }k_{\Delta }^{\mu },\;\delta N_{\Delta ^{\prime }\mu \nu }=\frac{1}{2}\partial _{\left[ \mu \right. }k_{\Delta ^{\prime }\left. \nu \right] },\;\delta P_{\Delta ^{\prime \prime }}=\partial _{\mu }k_{\Delta ^{\prime \prime }}^{\mu }, \label{f24}$$ where the corresponding currents are also invariant polynomials $$\begin{aligned} k_{\Delta }^{\mu } &=&-\left( \frac{\partial W_{\Delta }}{\partial \varphi _{m}}A_{m}^{*\mu }+\frac{\partial ^{2}W_{\Delta }}{\partial \varphi _{m}\partial \varphi _{n}}B_{m}^{\mu \nu }H_{n\nu }^{*}\right) , \label{f24a} \\ k_{\Delta ^{\prime }\nu } &=&\frac{\partial U_{\Delta ^{\prime }}}{\partial \varphi _{m}}H_{m\nu }^{*}, \label{f24b} \\ k_{\Delta ^{\prime \prime }}^{\mu } &=&-V_{A}^{*\mu }f_{\Delta ^{\prime \prime }}^{A}+\frac{\partial f_{\Delta ^{\prime \prime }}^{A}}{\partial \varphi _{m}}F_{A}^{\mu \nu }H_{m\nu }^{*}, \label{f24c}\end{aligned}$$ and hence we have that $$\delta \alpha _{2}^{0}=\partial _{\mu }k_{1}^{\mu }, \label{f24e}$$ where $k_{1}^{\mu }$ is the invariant polynomial $$k_{1}^{\mu }=K^{\Delta }k_{\Delta }^{\mu }+K^{\Delta ^{\prime }\mu \nu }k_{\Delta ^{\prime }\nu }+K^{\Delta ^{\prime \prime }}k_{\Delta ^{\prime \prime }}^{\mu }. \label{f24f}$$ The previous results on $H\left( \delta |d\right) $ and $H^{\mathrm{inv}% }\left( \delta |d\right) $ at strictly positive antighost numbers are important because they control the obstructions to removing the antifields from the first-order deformation. Indeed, due to (\[f19b\]–\[f19c\]) and to the triviality of the cohomology of the exterior spacetime differential in the space of invariant polynomials at strictly positive antighost numbers, it follows that we can successively remove all the pieces with $J>2$ from the non-integrated density of the first-order deformation by adding to it only trivial terms. In conclusion we can take, without loss of non-trivial objects, the maximum value $J=2$ of the antighost number in the decomposition (\[f18\]). ### Determination of the first-order deformation For $J=2$, the first-order deformation (\[f18\]) reduces to $$a=a_{0}+a_{1}+a_{2}, \label{f24d}$$ where its last representative ($\gamma a_{2}=0$) is of the form (\[f19\]). The elements of pure ghost number equal to two of a basis in the space of polynomials in $\eta ^{\alpha _{1}}$ are spanned by $$e^{2}:\left( \eta ^{a}\eta ^{b},\eta _{\mu \nu }^{a}\eta _{\rho \lambda }^{b},C^{A}C^{B},\eta ^{a}\eta _{\mu \nu }^{b},\eta ^{a}C^{A},\eta _{\mu \nu }^{a}C^{A}\right) , \label{f24g}$$ and therefore we can write (up to $\gamma $-exact contributions) that $$\begin{aligned} a_{2} &=&\frac{1}{2}\left( \alpha _{ab}\eta ^{a}\eta ^{b}+\alpha _{ab}^{\mu \nu \rho \lambda }\eta _{\mu \nu }^{a}\eta _{\rho \lambda }^{b}+\alpha _{AB}C^{A}C^{B}\right) \nonumber \\ &&+\alpha _{ab}^{\mu \nu }\eta ^{a}\eta _{\mu \nu }^{b}+\alpha _{aA}\eta ^{a}C^{A}+\alpha _{aA}^{\mu \nu }\eta _{\mu \nu }^{a}C^{A}. \label{f25}\end{aligned}$$ According to the previous discussion, the objects $\alpha _{ab}$, $\alpha _{ab}^{\mu \nu \rho \lambda }$, $\alpha _{AB}$, $\alpha _{ab}^{\mu \nu }$, $% \alpha _{aA}$ and $\alpha _{aA}^{\mu \nu }$ necessarily belong to $H_{2}^{% \mathrm{inv}}\left( \delta |d\right) $, so $$\begin{aligned} \delta \alpha _{ab} &=&\partial _{\mu }k_{ab}^{\mu },\;\delta \alpha _{ab}^{\mu \nu \rho \lambda }=\partial _{\beta }k_{ab}^{\beta \mu \nu \rho \lambda },\;\delta \alpha _{AB}=\partial _{\mu }k_{AB}^{\mu }, \label{f27a} \\ \delta \alpha _{ab}^{\mu \nu } &=&\partial _{\beta }k_{ab}^{\beta \mu \nu },\;\delta \alpha _{aA}=\partial _{\mu }k_{aA}^{\mu },\;\delta \alpha _{aA}^{\mu \nu }=\partial _{\beta }k_{aA}^{\beta \mu \nu }, \label{f27b}\end{aligned}$$ for some currents that are invariant polynomials of antighost number one. In addition, they are subject to the “symmetry” conditions (due to the anticommutation of the ghosts) $$\begin{aligned} \alpha _{ab} &=&-\alpha _{ba},\;\alpha _{ab}^{\mu \nu \rho \lambda }=-\alpha _{ba}^{\rho \lambda \mu \nu },\;\alpha _{AB}=-\alpha _{BA}, \label{f26a} \\ \alpha _{ab}^{\mu \nu } &=&-\alpha _{ab}^{\nu \mu },\;\alpha _{aA}^{\mu \nu }=-\alpha _{aA}^{\nu \mu },\;\alpha _{ab}^{\mu \nu \rho \lambda }=-\alpha _{ab}^{\nu \mu \rho \lambda }=-\alpha _{ab}^{\mu \nu \lambda \rho }. \label{f26b}\end{aligned}$$ If we insert the expression (\[f25\]) into the Eq. (\[f18c\]) for $% J=2$ $$\delta a_{2}+\gamma a_{1}=\partial _{\mu }j_{1}^{\mu }, \label{f27}$$ use the formulas (\[f27a\]–\[f27b\]) and recall the definitions (\[f15\]–\[f16\]), we obtain that the existence of $a_{1}$ demands further restrictions on the currents, namely, $$\begin{aligned} k_{ab}^{\beta \mu \nu \rho \lambda }\partial _{\beta }\left( \eta _{\mu \nu }^{a}\eta _{\rho \lambda }^{b}\right) &=&\sigma _{ab}^{\nu \rho \lambda }\left( \partial ^{\mu }\eta _{\mu \nu }^{a}\right) \eta _{\rho \lambda }^{b}+\sigma _{ab}^{\mu \nu \lambda }\eta _{\mu \nu }^{a}\left( \partial ^{\rho }\eta _{\rho \lambda }^{b}\right) , \label{f28a} \\ k_{ab}^{\beta \mu \nu }\partial _{\beta }\eta _{\mu \nu }^{b} &=&\mu _{ab}^{\nu }\partial ^{\mu }\eta _{\mu \nu }^{b},\;k_{aA}^{\beta \mu \nu }\partial _{\beta }\eta _{\mu \nu }^{a}=\mu _{aA}^{\nu }\partial ^{\mu }\eta _{\mu \nu }^{a}, \label{f28b}\end{aligned}$$ for some $\sigma $ and $\mu $. On the other hand, in agreement with the result (\[f24\]), every function from $H_{2}^{\mathrm{inv}}\left( \delta |d\right) $ entering the solution (\[f25\]) can only be constructed out of the three different kinds of invariant polynomials (\[f21\]–\[f23\]). Expressing now each function of the type $\alpha $ from $a_{2}$ in terms of the allowed elements (\[f21\]–\[f23\]) and imposing the supplementary Eqs. (\[f28a\]–\[f28a\]) at the level of the accompanying currents, after some computation we infer the solutions $$\begin{aligned} \alpha _{ab} &=&M_{ab}+P_{ab}=\frac{\partial W_{ab}}{\partial \varphi _{m}}% \eta _{m}^{*}-\frac{\partial ^{2}W_{ab}}{\partial \varphi _{m}\partial \varphi _{n}}\left( B_{m\mu \nu }\eta _{n}^{*\mu \nu }+H_{m}^{*\mu }A_{n\mu }^{*}\right) \nonumber \\ &&-\frac{1}{2}\frac{\partial ^{3}W_{ab}}{\partial \varphi _{m}\partial \varphi _{n}\partial \varphi _{p}}H_{m}^{*\mu }H_{n}^{*\nu }B_{p\mu \nu }+f_{ab}^{A}C_{A}^{*} \nonumber \\ &&+\frac{\partial f_{ab}^{A}}{\partial \varphi _{m}}\left( V_{A}^{*\mu }H_{m\mu }^{*}+\eta _{m}^{*\mu \nu }F_{A\mu \nu }\right) +\frac{1}{2}\frac{% \partial ^{2}f_{ab}^{A}}{\partial \varphi _{m}\partial \varphi _{n}}% H_{m}^{*\mu }H_{n}^{*\nu }F_{A\mu \nu }, \label{f29a}\end{aligned}$$ $$\alpha _{ab}^{\mu \nu \rho \lambda }=0=\alpha _{aA}^{\mu \nu }, \label{f29b}$$ $$\begin{aligned} \alpha _{AB} &=&P_{AB}=f_{AB}^{C}C_{C}^{*}+\frac{\partial f_{AB}^{C}}{% \partial \varphi _{m}}\left( V_{C}^{*\mu }H_{m\mu }^{*}+\eta _{m}^{*\mu \nu }F_{C\mu \nu }\right) \nonumber \\ &&+\frac{1}{2}\frac{\partial ^{2}f_{AB}^{C}}{\partial \varphi _{m}\partial \varphi _{n}}H_{m}^{*\mu }H_{n}^{*\nu }F_{C\mu \nu }, \label{f29c}\end{aligned}$$ $$\alpha _{ab}^{\mu \nu }=N_{ab}^{\mu \nu }=\frac{\partial U_{ab}}{\partial \varphi _{m}}\eta _{m}^{*\mu \nu }+\frac{1}{2}\frac{\partial ^{2}U_{ab}}{% \partial \varphi _{m}\partial \varphi _{n}}H_{m}^{*\mu }H_{n}^{*\nu }, \label{f29d}$$ $$\begin{aligned} \alpha _{aA} &=&P_{aA}=g_{aA}^{C}C_{C}^{*}+\frac{\partial g_{aA}^{C}}{% \partial \varphi _{m}}\left( V_{C}^{*\mu }H_{m\mu }^{*}+\eta _{m}^{*\mu \nu }F_{C\mu \nu }\right) \nonumber \\ &&+\frac{1}{2}\frac{\partial ^{2}g_{aA}^{C}}{\partial \varphi _{m}\partial \varphi _{n}}H_{m}^{*\mu }H_{n}^{*\nu }F_{C\mu \nu }, \label{f29e}\end{aligned}$$ where $W_{ab}$, $f_{ab}^{A}$, $f_{AB}^{C}$, $U_{ab}$ and $g_{aA}^{C}$ depend only on the scalar fields $\varphi _{a}$, with $W_{ab}$, $f_{ab}^{A}$ and $% f_{AB}^{C}$ antisymmetric in their lower indices in order to enforce the “symmetry” properties (\[f26a\]–\[f26b\]) $$W_{ab}=-W_{ba},\;f_{ab}^{A}=-f_{ba}^{A},\;f_{AB}^{C}=-f_{BA}^{C}. \label{fas}$$ In conclusion, the full expression of the last component from the first-order deformation (\[f24d\]), such that it leads to a consistent solution $a_{1}$ to the Eq. (\[f27\]), has the form (\[f25\]), with the invariant polynomials $\alpha _{ab}$, $\alpha _{ab}^{\mu \nu \rho \lambda }$, $\alpha _{AB}$, $\alpha _{ab}^{\mu \nu }$, $\alpha _{aA}$ and $% \alpha _{aA}^{\mu \nu }$ from $H_{2}^{\mathrm{inv}}\left( \delta |d\right) $ precisely given by (\[f29a\]–\[f29e\]). With $a_{2}$ at hand, direct calculations provide the piece of antighost number one from (\[f24d\]) like $$\begin{aligned} a_{1} &=&\left( U_{ab}\left( B^{*a\mu \nu }\eta _{\mu \nu }^{b}+\varphi ^{*b}\eta ^{a}\right) -\frac{\partial U_{ab}}{\partial \varphi _{c}}% H_{c}^{*\nu }\left( A^{a\mu }\eta _{\mu \nu }^{b}+H_{\nu }^{b}\eta ^{a}\right) \right) \nonumber \\ &&+\left( f_{ab}^{A}F_{A}^{\mu \nu }-\frac{\partial W_{ab}}{\partial \varphi _{c}}B_{c}^{\mu \nu }\right) \eta ^{a}B_{\mu \nu }^{*b}+g_{aA}^{C}F_{C}^{\mu \nu }B_{\mu \nu }^{*a}C^{A} \nonumber \\ &&+\left( \frac{\partial f_{ab}^{A}}{\partial \varphi _{c}}F_{A}^{\mu \nu }H_{c\nu }^{*}-\left( \frac{\partial W_{ab}}{\partial \varphi _{c}}% A_{c}^{*\mu }+\frac{\partial ^{2}W_{ab}}{\partial \varphi _{c}\partial \varphi _{d}}B_{c}^{\mu \nu }H_{d\nu }^{*}+f_{ab}^{A}V_{A}^{*\mu }\right) \right) \eta ^{a}A_{\mu }^{b} \nonumber \\ &&+\left( -f_{AB}^{C}V_{C}^{*\mu }+\frac{\partial f_{AB}^{C}}{\partial \varphi _{c}}F_{C}^{\mu \nu }H_{c\nu }^{*}\right) C^{A}V_{\mu }^{B} \nonumber \\ &&+\left( -g_{aA}^{C}V_{C}^{*\mu }+\frac{\partial g_{aA}^{C}}{\partial \varphi _{b}}F_{C}^{\mu \nu }H_{b\nu }^{*}\right) \left( \eta ^{a}V_{\mu }^{A}-A_{\mu }^{a}C^{A}\right) . \label{f30}\end{aligned}$$ The last step in completing the first-order deformation is the resolution of the equation (\[f18d\]) for $k=1$ $$\delta a_{1}+\gamma a_{0}=\partial _{\mu }j_{0}^{\mu }. \label{f30a}$$ Evaluating $\delta a_{1}$, we find that the Eq. (\[f30a\]) possesses solutions with respect to $a_{0}$ if the functions $f_{AB}^{C}$ and $% g_{aA}^{C}$ of the undifferentiated scalar fields obey the conditions $$f_{ABC}=-f_{BAC}=-f_{ACB},\;g_{aAB}=-g_{aBA}, \label{f31}$$ where $f_{ABC}$ and $g_{aAB}$ are defined by $$f_{ABC}=k_{AE}f_{BC}^{E},\;g_{aAB}=k_{AE}g_{aB}^{E}, \label{f32}$$ and $k_{AE}$ denote the elements of the matrix inverse to $k^{AE}$ (used to raise the collection indices of the vector fields). Then, we get the interacting Lagrangian at the first order in the coupling constant like $$\begin{aligned} a_{0} &=&\frac{1}{2}\left( \frac{\partial W_{ab}}{\partial \varphi _{c}}% B_{c}^{\mu \nu }-f_{ab}^{A}F_{A}^{\mu \nu }\right) A_{\mu }^{a}A_{\nu }^{b}-% \frac{1}{2}f_{BC}^{A}F_{A}^{\mu \nu }V_{\mu }^{B}V_{\nu }^{C} \nonumber \\ &&-g_{aB}^{A}F_{A}^{\mu \nu }A_{\mu }^{a}V_{\nu }^{B}-U_{ab}A^{a\mu }H_{\mu }^{b}. \label{f33}\end{aligned}$$ So far we have completely determined the first-order deformation of the solution to the master equation for the model under study $$S_{1}=\int d^{2}x\left( a_{0}+a_{1}+a_{2}\right) , \label{f33a}$$ where its components are expressed by (\[f25\]) (with the corresponding invariant polynomials of the form (\[f29a\]–\[f29e\])), (\[f30\]) and (\[f33\]). Moreover, the various functions of the undifferentiated scalar fields are taken to obey the properties (\[fas\]) and (\[f31\]). Higher-order deformations ------------------------- ### Second-order deformation Using the notations from (\[f4\]), the master equation (\[mastdef\]) holds to order $g^{2}$ if and only if $$\Delta =-2sb+\partial _{\mu }u^{\mu }, \label{f33b}$$ where $\left( S_{1},S_{1}\right) =\int d^{2}x\Delta $. In other words, the consistency of the deformed solution to the master equation at the second order in the coupling constant requires that the integrand of $\left( S_{1},S_{1}\right) $ should (locally) be written like a $s$-co-boundary modulo $d$. Relying on the expression of $S_{1}$ deduced in the above, we can emphasize a $s$-exact part in $\Delta $ if and only if the functions $% U_{ab}$ and $W_{ab}$ coincide $$U_{ab}=W_{ab}, \label{f34}$$ in which case we find that $$\begin{aligned} \Delta &=&-2sb+\left( t_{bcd}u^{bcd}+\frac{\partial t_{bcd}}{\partial \varphi _{e}}u_{e}^{bcd}+\frac{\partial ^{2}t_{bcd}}{\partial \varphi _{e}\partial \varphi _{f}}u_{ef}^{bcd}+\frac{\partial ^{3}t_{bcd}}{\partial \varphi _{e}\partial \varphi _{f}\partial \varphi _{g}}u_{efg}^{bcd}\right) \nonumber \\ &&+\left( \alpha _{abc}^{A}v_{|A}^{abc}+\frac{\partial \alpha _{abc}^{A}}{% \partial \varphi _{e}}v_{e|A}^{abc}+\frac{\partial ^{2}\alpha _{abc}^{A}}{% \partial \varphi _{e}\partial \varphi _{f}}v_{ef|A}^{abc}\right) \nonumber \\ &&+\left( \alpha _{BCD}^{A}w_{|A}^{BCD}+\frac{\partial \alpha _{BCD}^{A}}{% \partial \varphi _{e}}w_{e|A}^{BCD}+\frac{\partial ^{2}\alpha _{BCD}^{A}}{% \partial \varphi _{e}\partial \varphi _{f}}w_{ef|A}^{BCD}\right) \nonumber \\ &&+\left( \alpha _{abB}^{A}z_{|A}^{abB}+\frac{\partial \alpha _{abB}^{A}}{% \partial \varphi _{e}}z_{e|A}^{abB}+\frac{\partial ^{2}\alpha _{abB}^{A}}{% \partial \varphi _{e}\partial \varphi _{f}}z_{ef|A}^{abB}\right) \nonumber \\ &&+\left( \alpha _{aBC}^{A}q_{|A}^{aBC}+\frac{\partial \alpha _{aBC}^{A}}{% \partial \varphi _{e}}q_{e|A}^{aBC}+\frac{\partial ^{2}\alpha _{aBC}^{A}}{% \partial \varphi _{e}\partial \varphi _{f}}q_{ef|A}^{aBC}\right) . \label{f36}\end{aligned}$$ The expression of $b$ in (\[f36\]) is $$b=-\frac{1}{4}Q_{\mu \nu }^{A}k_{AB}Q^{B\mu \nu }, \label{f42}$$ where we performed the notation $$\begin{aligned} Q_{\mu \nu }^{A} &=&\left( \frac{\partial f_{ab}^{A}}{\partial \varphi _{c}}% \eta _{c\mu \nu }^{*}+\frac{1}{2}\frac{\partial ^{2}f_{ab}^{A}}{\partial \varphi _{c}\partial \varphi _{d}}H_{c\mu }^{*}H_{d\nu }^{*}\right) \eta ^{a}\eta ^{b}-\left( f_{ab}^{A}A_{\mu }^{a}A_{\nu }^{b}+f_{BC}^{A}V_{\mu }^{B}V_{\nu }^{C}\right) \nonumber \\ &&-\left( \frac{\partial f_{ab}^{A}}{\partial \varphi _{c}}H_{c[\mu }^{*}A_{\nu ]}^{b}+2f_{ab}^{A}B_{\mu \nu }^{*b}\right) \eta ^{a}-\frac{% \partial f_{BC}^{A}}{\partial \varphi _{c}}H_{c[\mu }^{*}V_{\nu ]}^{C}C^{B}-g_{aB}^{A}A_{[\mu }^{a}V_{\nu ]}^{B} \nonumber \\ &&+\left( \frac{\partial f_{BC}^{A}}{\partial \varphi _{c}}\eta _{c\mu \nu }^{*}+\frac{1}{2}\frac{\partial ^{2}f_{BC}^{A}}{\partial \varphi _{c}\partial \varphi _{d}}H_{c\mu }^{*}H_{d\nu }^{*}\right) C^{B}C^{C}+2g_{aB}^{A}B_{\mu \nu }^{*a}C^{B} \nonumber \\ &&+2\left( \frac{\partial g_{aB}^{A}}{\partial \varphi _{c}}\eta _{c\mu \nu }^{*}+\frac{1}{2}\frac{\partial ^{2}g_{aB}^{A}}{\partial \varphi _{c}\partial \varphi _{d}}H_{c\mu }^{*}H_{d\nu }^{*}\right) \eta ^{a}C^{B} \nonumber \\ &&+\frac{\partial g_{aB}^{A}}{\partial \varphi _{c}}\left( H_{c[\mu }^{*}A_{\nu ]}^{a}C^{B}-H_{c[\mu }^{*}V_{\nu ]}^{B}\eta ^{a}\right) . \label{f43}\end{aligned}$$ The coefficients denoted by $t$ and $\alpha $ from (\[f36\]) involve only the undifferentiated scalar fields and are given by $$t_{abc}=W_{e[a}\frac{\partial W_{bc]}}{\partial \varphi _{e}}, \label{f35a}$$ $$\alpha _{abc}^{A}=f_{e[a}^{A}\frac{\partial W_{bc]}}{\partial \varphi _{e}}% +W_{e[a}\frac{\partial f_{bc]}^{A}}{\partial \varphi _{e}}% -f_{[ab}^{E}g_{c]E}^{A}, \label{f35b}$$ $$\alpha _{BCD}^{A}=f_{E[B}^{A}f_{CD]}^{E}, \label{f35c}$$ $$\alpha _{abB}^{A}=g_{eB}^{A}\frac{\partial W_{ab}}{\partial \varphi _{e}}% +f_{EB}^{A}f_{ab}^{E}+W_{e[a}\frac{\partial g_{b]B}^{A}}{\partial \varphi _{e}}+\left( g_{aB}^{E}g_{bE}^{A}-g_{bB}^{E}g_{aE}^{A}\right) , \label{f35d}$$ $$\alpha _{aBC}^{A}=g_{aE}^{A}f_{BC}^{E}-W_{ea}\frac{\partial f_{BC}^{A}}{% \partial \varphi _{e}}+\left( f_{EB}^{A}g_{aC}^{E}-f_{EC}^{A}g_{aB}^{E}\right) , \label{f35e}$$ while the functions of the type $u$, $v$, $w$, $z$ and $q$ contain only undifferentiated ghosts and antifields. Their expressions are listed below $$u^{bcd}=\left( A^{b\mu }A^{c\nu }-B^{*b\mu \nu }\eta ^{c}\right) \eta _{\mu \nu }^{d}-\left( A^{b\mu }H_{\mu }^{d}+\varphi ^{*b}\eta ^{d}\right) \eta ^{c}, \label{f37a}$$ $$\begin{aligned} u_{e}^{bcd} &=&\left( B_{e}^{\mu \nu }A_{\mu }^{b}A_{\nu }^{d}+A_{e}^{*\mu }A_{\mu }^{b}\eta ^{d}\right) \eta ^{c}+\left( B_{e}^{\mu \nu }B_{\mu \nu }^{*b}-H_{e}^{*\mu }H_{\mu }^{b}-\frac{1}{3}\eta _{e}^{*}\eta ^{b}\right) \eta ^{c}\eta ^{d} \nonumber \\ &&+\left( H_{e}^{*\nu }A^{b\mu }-\eta _{e}^{*\mu \nu }\eta ^{b}\right) \eta ^{c}\eta _{\mu \nu }^{d}, \label{f37b}\end{aligned}$$ $$u_{ef}^{bcd}=\left( \eta _{e}^{*\mu \nu }B_{f\mu \nu }\eta ^{b}-B_{e\mu \nu }H_{f}^{*\nu }A^{b\mu }-\frac{1}{2}H_{e}^{*\mu }H_{f}^{*\nu }\eta _{\mu \nu }^{b}-A_{e\mu }^{*}H_{f}^{*\mu }\eta ^{b}\right) \eta ^{c}\eta ^{d}, \label{f37c}$$ $$u_{efg}^{bcd}=\frac{1}{6}B_{e\mu \nu }H_{f}^{*\mu }H_{g}^{*\nu }\eta ^{b}\eta ^{c}\eta ^{d}, \label{f37d}$$ $$v_{|A}^{abc}=-\left( \frac{1}{3}C_{A}^{*}\eta ^{a}+V_{A}^{*\mu }A_{\mu }^{a}+F_{A}^{\mu \nu }B_{\mu \nu }^{*a}\right) \eta ^{b}\eta ^{c}+F_{A}^{\mu \nu }\eta ^{a}A_{\mu }^{b}A_{\nu }^{c}, \label{f38a}$$ $$v_{e|A}^{abc}=-\left( \frac{1}{3}V_{A}^{*\mu }H_{e\mu }^{*}\eta ^{a}+\frac{1% }{3}F_{A}^{\mu \nu }\eta _{e\mu \nu }^{*}\eta ^{a}+F_{A}^{\mu \nu }H_{e\mu }^{*}A_{\nu }^{a}\right) \eta ^{b}\eta ^{c}, \label{f38b}$$ $$v_{ef|A}^{abc}=-\frac{1}{6}F_{A}^{\mu \nu }H_{e\mu }^{*}H_{f\nu }^{*}\eta ^{a}\eta ^{b}\eta ^{c}, \label{f38c}$$ $$w_{|A}^{BCD}=-\left( \frac{1}{3}C_{A}^{*}C^{B}+V_{A}^{*\mu }V_{\mu }^{B}\right) C^{C}C^{D}+F_{A}^{\mu \nu }C^{B}V_{\mu }^{C}V_{\nu }^{D}, \label{f39a}$$ $$w_{e|A}^{BCD}=-\left( \frac{1}{3}V_{A}^{*\mu }H_{e\mu }^{*}C^{B}+\frac{1}{3}% F_{A}^{\mu \nu }\eta _{e\mu \nu }^{*}C^{B}+F_{A}^{\mu \nu }H_{e\mu }^{*}V_{\nu }^{B}\right) C^{C}C^{D}, \label{f39b}$$ $$w_{ef|A}^{BCD}=-\frac{1}{6}F_{A}^{\mu \nu }H_{e\mu }^{*}H_{f\nu }^{*}C^{B}C^{C}C^{D}, \label{f39c}$$ $$\begin{aligned} z_{|A}^{abB} &=&\left( F_{A}^{\mu \nu }A_{\mu }^{a}A_{\nu }^{b}+2V_{A}^{*\mu }A_{\mu }^{b}\eta ^{a}-2F_{A}^{\mu \nu }B_{\mu \nu }^{*a}\eta ^{b}\right) C^{B} \nonumber \\ &&-F_{A}^{\mu \nu }V_{[\mu }^{B}A_{\nu ]}^{b}\eta ^{a}-\left( C_{A}^{*}C^{B}+V_{A}^{*\mu }V_{\mu }^{B}\right) \eta ^{a}\eta ^{b}, \label{f40a}\end{aligned}$$ $$\begin{aligned} z_{e|A}^{abB} &=&-\left( V_{A}^{*\mu }H_{e\mu }^{*}C^{B}+F_{A}^{\mu \nu }\eta _{e\mu \nu }^{*}C^{B}+F_{A}^{\mu \nu }H_{e\mu }^{*}V_{\nu }^{B}\right) \eta ^{a}\eta ^{b} \nonumber \\ &&-F_{A}^{\mu \nu }H_{e[\mu }^{*}A_{\nu ]}^{b}C^{B}\eta ^{a}, \label{f40b}\end{aligned}$$ $$z_{ef|A}^{abB}=-\frac{1}{2}F_{A}^{\mu \nu }H_{e\mu }^{*}H_{f\nu }^{*}C^{B}\eta ^{a}\eta ^{b}, \label{f40c}$$ $$\begin{aligned} q_{|A}^{aBC} &=&\left( C_{A}^{*}\eta ^{a}+V_{A}^{*\mu }A_{\mu }^{a}+F_{A}^{\mu \nu }B_{\mu \nu }^{*a}\right) C^{B}C^{C} \nonumber \\ &&-\left( 2V_{A}^{*\mu }V_{\mu }^{C}C^{B}+F_{A}^{\mu \nu }V_{\mu }^{B}V_{\nu }^{C}\right) \eta ^{a}+F_{A}^{\mu \nu }A_{[\mu }^{a}V_{\nu ]}^{C}C^{B}, \label{f41a}\end{aligned}$$ $$\begin{aligned} q_{e|A}^{aBC} &=&\left( F_{A}^{\mu \nu }\eta _{e\mu \nu }^{*}\eta ^{a}+F_{A}^{\mu \nu }H_{e\mu }^{*}A_{\nu }^{a}+V_{A}^{*\mu }H_{e\mu }^{*}\eta ^{a}\right) C^{B}C^{C} \nonumber \\ &&+F_{A}^{\mu \nu }H_{e[\mu }^{*}V_{\nu ]}^{C}\eta ^{a}C^{B}, \label{f41b}\end{aligned}$$ $$q_{ef|A}^{aBC}=\frac{1}{2}F_{A}^{\mu \nu }H_{e\mu }^{*}H_{f\nu }^{*}\eta ^{a}C^{B}C^{C}. \label{f41c}$$ Since none of the terms from the right-hand side of (\[f36\]) containing the quantities $t_{abc}$, $\alpha _{abc}^{A}$, $\alpha _{BCD}^{A}$, $\alpha _{abB}^{A}$, $\alpha _{aBC}^{A}$ and their derivatives with respect to the scalar fields can be written in a $s$-exact modulo $d$ form, it follows that the second-order deformation of the solution to the master equation exists if and only if all these vanish $$t_{abc}=0,\;\alpha _{abc}^{A}=0,\;\alpha _{BCD}^{A}=0,\;\alpha _{abB}^{A}=0,\;\alpha _{aBC}^{A}=0. \label{relcoef}$$ In conclusion, the deformed solution to the master equation is consistent to order $g^{2}$ if and only if the functions $W_{ab}$, $f_{BC}^{A}$, $% f_{ab}^{A}$ and $g_{aB}^{A}$ satisfy the Eqs. (\[relcoef\]), in which case the second-order deformation is expressed like in (\[f42\]). We will comment more on the Eqs. (\[relcoef\]) in Section 5, where we explicitly compute some particular solutions, which allow a nice geometric and algebraic interpretation. ### Third- and higher-order deformations If we denote the third-order deformation by $S_{3}=\int d^{2}x\,c$, the master equation (\[mastdef\]) holds to order $g^{3}$ if and only if $$\Lambda =-sc+\partial _{\mu }w^{\mu }, \label{f44}$$ where$\left( S_{1},S_{2}\right) =\int d^{2}x\,\Lambda $. Taking into account the expressions of the first- and second-order deformations obtained in the above, after some computation we infer that $$\begin{aligned} \Lambda &=&\left( \eta ^{a}g_{aE}^{A}+\frac{1}{2}C^{D}f_{DE}^{A}\right) Q^{E\mu \nu }k_{AB}Q_{\mu \nu }^{B} \nonumber \\ &&+\left( \alpha _{abc}^{A}\bar{v}_{|A}^{abc}+\frac{\partial \alpha _{abc}^{A}}{\partial \varphi _{e}}\bar{v}_{e|A}^{abc}+\frac{\partial ^{2}\alpha _{abc}^{A}}{\partial \varphi _{e}\partial \varphi _{f}}\bar{v}% _{ef|A}^{abc}\right) \nonumber \\ &&+\left( \alpha _{BCD}^{A}\bar{w}_{|A}^{BCD}+\frac{\partial \alpha _{BCD}^{A}}{\partial \varphi _{e}}\bar{w}_{e|A}^{BCD}+\frac{\partial ^{2}\alpha _{BCD}^{A}}{\partial \varphi _{e}\partial \varphi _{f}}\bar{w}% _{ef|A}^{BCD}\right) \nonumber \\ &&+\left( \alpha _{abB}^{A}\bar{z}_{|A}^{abB}+\frac{\partial \alpha _{abB}^{A}}{\partial \varphi _{e}}\bar{z}_{e|A}^{abB}+\frac{\partial ^{2}\alpha _{abB}^{A}}{\partial \varphi _{e}\partial \varphi _{f}}\bar{z}% _{ef|A}^{abB}\right) \nonumber \\ &&+\left( \alpha _{aBC}^{A}\bar{q}_{|A}^{aBC}+\frac{\partial \alpha _{aBC}^{A}}{\partial \varphi _{e}}\bar{q}_{e|A}^{aBC}+\frac{\partial ^{2}\alpha _{aBC}^{A}}{\partial \varphi _{e}\partial \varphi _{f}}\bar{q}% _{ef|A}^{aBC}\right) , \label{f46}\end{aligned}$$ where we employed the notations $$\bar{v}_{|A}^{abc}=\frac{1}{2}k_{AM}Q^{M\mu \nu }\left( B_{\mu \nu }^{*a}\eta ^{a}\eta ^{b}-\eta ^{a}A_{\mu }^{b}A_{\nu }^{c}\right) , \label{f47a}$$ $$\bar{v}_{e|A}^{abc}=\frac{1}{2}k_{AM}Q^{M\mu \nu }\left( \frac{1}{3}\eta _{e\mu \nu }^{*}\eta ^{a}+\frac{1}{2}H_{e[\mu }^{*}A_{\nu ]}^{a}\right) \eta ^{b}\eta ^{c}, \label{f47b}$$ $$\bar{v}_{ef|A}^{abc}=\frac{1}{12}k_{AM}Q^{M\mu \nu }H_{e\mu }^{*}H_{f\nu }^{*}\eta ^{a}\eta ^{b}\eta ^{c}, \label{f47c}$$ $$\bar{w}_{|A}^{BCD}=-\frac{1}{2}k_{AM}Q^{M\mu \nu }C^{B}V_{\mu }^{C}V_{\nu }^{D}, \label{f48a}$$ $$\bar{w}_{e|A}^{BCD}=\frac{1}{2}k_{AM}Q^{M\mu \nu }\left( \frac{1}{3}\eta _{e\mu \nu }^{*}C^{B}+\frac{1}{2}H_{e[\mu }^{*}V_{\nu ]}^{B}\right) C^{C}C^{D}, \label{f48b}$$ $$\bar{w}_{ef|A}^{BCD}=\frac{1}{12}k_{AM}Q^{M\mu \nu }H_{e\mu }^{*}H_{f\nu }^{*}C^{B}C^{C}C^{D}, \label{f48c}$$ $$\bar{z}_{|A}^{abB}=-\frac{1}{2}k_{AM}Q^{M\mu \nu }\left( \eta ^{a}A_{[\mu }^{b}V_{\nu ]}^{B}+A_{\mu }^{a}A_{\nu }^{b}C^{B}+2B_{\mu \nu }^{*b}\eta ^{a}C^{B}\right) , \label{f49a}$$ $$\bar{z}_{e|A}^{abB}=\frac{1}{2}k_{AM}Q^{M\mu \nu }\left[ \left( \eta _{e\mu \nu }^{*}C^{B}+\frac{1}{2}H_{e[\mu }^{*}V_{\nu ]}^{B}\right) \eta ^{a}\eta ^{b}-\frac{1}{2}H_{e[\mu }^{*}A_{\nu ]}^{b}\eta ^{a}C^{B}\right] , \label{f49b}$$ $$\bar{z}_{ef|A}^{abB}=\frac{1}{4}k_{AM}Q^{M\mu \nu }k_{AM}Q^{M\mu \nu }H_{e\mu }^{*}H_{f\nu }^{*}\eta ^{a}\eta ^{b}C^{B}, \label{f49c}$$ $$\bar{q}_{|A}^{aBC}=-\frac{1}{2}k_{AM}Q^{M\mu \nu }\left( A_{[\mu }^{a}V_{\nu ]}^{C}C^{B}+B_{\mu \nu }^{*a}C^{B}C^{C}-\eta ^{a}V_{\mu }^{B}V_{\nu }^{C}\right) , \label{f50a}$$ $$\bar{q}_{e|A}^{aBC}=-\frac{1}{2}k_{AM}Q^{M\mu \nu }\left[ \left( \eta _{e\mu \nu }^{*}\eta ^{a}+H_{e[\mu }^{*}A_{\nu ]}^{a}\right) C^{B}C^{C}+H_{e[\mu }^{*}V_{\nu ]}^{C}\eta ^{a}C^{B}\right] , \label{f50b}$$ $$\bar{q}_{ef|A}^{aBC}=-\frac{1}{4}k_{AM}Q^{M\mu \nu }k_{AM}Q^{M\mu \nu }H_{e\mu }^{*}H_{f\nu }^{*}\eta ^{a}C^{B}C^{C}. \label{f50c}$$ It is now clear that none of the terms in the right-hand side of (\[f46\]) can be written like in (\[f44\]). However, if we take into account the Eqs. (\[relcoef\]) and the antisymmetry properties (\[f31\]) of the functions (\[f32\]), we find that $\Lambda =0$, so we can take $c=0$ in (\[f44\]), and consequently obtain that $$S_{3}=0. \label{f100}$$ The equation that governs the fourth-order deformation $S_{4}$ reads as $$sS_{4}+\left( S_{3},S_{1}\right) +\frac{1}{2}\left( S_{2},S_{2}\right) =0. \label{f100a}$$ On the one hand, the result (\[f100\]) implies that $\left( S_{3},S_{1}\right) =0$ and, on the other hand, if we compute $\left( S_{2},S_{2}\right) $, where $S_{2}=\int d^{2}x\,b$, with $b$ given in (\[f42\]), it follows that $\left( S_{2},S_{2}\right) =0$, so we can set $$S_{4}=0. \label{f100b}$$ Meanwhile, we remark that the equations responsible for the deformations $% \left( S_{k}\right) _{k>4}$ involve only the solutions $\left( S_{j}\right) _{j\geq 3}$, which further allows us to put $$S_{k}=0,\;k>4. \label{f100c}$$ In conclusion, among the higher-order deformations of the solution to the master equation, only that of second-order is non-vanishing and non-trivial. Identification of the interacting theory ======================================== Putting together the results deduced in the previous section, we can write the full deformed solution to the master equation (\[mastdef\]), that is consistent to all orders in the coupling constant, under the form $$\begin{aligned} \bar{S} &=&S+gS_{1}+g^{2}S_{2}=\int d^{2}x\left( H_{\mu }^{a}D^{\mu }\varphi _{a}+\frac{1}{2}B_{a}^{\mu \nu }\bar{F}_{\mu \nu }^{^{\prime }a}+A_{a}^{*\mu }\left( D_{\mu }\right) _{\;\;b}^{a}\eta ^{b}\right. \nonumber \\ &&-gW_{ab}\varphi ^{*a}\eta ^{b}+gB^{*a\mu \nu }\left( W_{ab}\eta _{\mu \nu }^{b}-\frac{\partial W_{ab}}{\partial \varphi _{c}}B_{c\mu \nu }\eta ^{b}\right) \nonumber \\ &&+H_{a}^{*\mu }\left( \left( D^{\nu }\right) _{\;\;b}^{a}\eta _{\mu \nu }^{b}-g\left( \frac{\partial W_{bc}}{\partial \varphi _{a}}H_{\mu }^{c}-% \frac{\partial ^{2}W_{bc}}{\partial \varphi _{a}\partial \varphi _{d}}% B_{d\mu \nu }A^{c\nu }\right) \eta ^{b}\right) \nonumber \\ &&+V_{A}^{*\mu }\left( \left( D_{\mu }\right) _{\;\;B}^{A}C^{B}+\left( D_{\mu }\right) _{\;\;a}^{A}\eta ^{a}\right) \nonumber \\ &&-\frac{1}{4}\left( F_{\mu \nu }^{A}-Q_{\mu \nu }^{A}\right) k_{AB}\left( F^{B\mu \nu }-Q^{B\mu \nu }\right) \nonumber \\ &&+\frac{g}{2}\left( f_{ab}^{A}C_{A}^{*}+\frac{\partial W_{ab}}{\partial \varphi _{c}}\eta _{c}^{*}+\frac{\partial f_{ab}^{A}}{\partial \varphi _{c}}% V_{A}^{*\mu }H_{c\mu }^{*}+\frac{\partial ^{2}W_{ab}}{\partial \varphi _{c}\partial \varphi _{d}}A_{d}^{*\mu }H_{c\mu }^{*}\right. \nonumber \\ &&\left. -\frac{\partial ^{2}W_{ab}}{\partial \varphi _{c}\partial \varphi _{d}}B_{d\mu \nu }\eta _{c}^{*\mu \nu }-\frac{1}{2}\frac{\partial ^{3}W_{ab}% }{\partial \varphi _{c}\partial \varphi _{d}\partial \varphi _{e}}B_{c\mu \nu }H_{d}^{*\mu }H_{e}^{*\nu }\right) \eta ^{a}\eta ^{b} \nonumber \\ &&+g\left( \frac{\partial W_{ab}}{\partial \varphi _{c}}\eta _{c}^{*\mu \nu }+\frac{1}{2}\frac{\partial ^{2}W_{ab}}{\partial \varphi _{c}\partial \varphi _{d}}H_{c}^{*\mu }H_{d}^{*\nu }\right) \eta ^{a}\eta _{\mu \nu }^{b} \nonumber \\ &&+\frac{g}{2}\left( f_{BC}^{A}C_{A}^{*}+\frac{\partial f_{BC}^{A}}{\partial \varphi _{c}}V_{A}^{*\mu }H_{c\mu }^{*}\right) C^{B}C^{C} \nonumber \\ &&\left. +g\left( g_{aB}^{A}C_{A}^{*}+\frac{\partial g_{aB}^{A}}{\partial \varphi _{b}}V_{A}^{*\mu }H_{b\mu }^{*}\right) \eta ^{a}C^{B}\right) , \label{f51}\end{aligned}$$ where we performed the notations $$\begin{aligned} D^{\mu }\varphi _{a} &=&\partial ^{\mu }\varphi _{a}+gW_{ab}A^{b\mu }, \label{f52a} \\ \bar{F}_{\mu \nu }^{^{\prime }a} &=&\partial _{[\mu }A_{\nu ]}^{a}+g\frac{% \partial W_{bc}}{\partial \varphi _{a}}A_{\mu }^{b}A_{\nu }^{c}, \label{f52e} \\ \left( D^{\mu }\right) _{\;\;b}^{a} &=&\delta _{\;\;b}^{a}\partial ^{\mu }-g% \frac{\partial W_{bc}}{\partial \varphi _{a}}A_{\mu }^{c}, \label{f52b} \\ \left( D_{\mu }\right) _{\;\;B}^{A} &=&\delta _{\;\;B}^{A}\partial _{\mu }-g\left( f_{BC}^{A}V_{\mu }^{C}-g_{aB}^{A}A_{\mu }^{a}\right) , \label{f52c} \\ \left( D_{\mu }\right) _{\;\;a}^{A} &=&-g\left( f_{ab}^{A}A_{\mu }^{b}+g_{aB}^{A}V_{\mu }^{B}\right) . \label{f52d}\end{aligned}$$ We note that the deformed solution (\[f51\]) contains components of antighost numbers ranging from zero to four, unlike the solution (\[f5\]) of the master equation for the free model, which stopped at antighost number one. We stress again that the coefficients $W_{ab}$, $f_{ab}^{A}$, $% f_{BC}^{A}$ and $g_{aB}^{A}$ are all functions of the undifferentiated scalar fields, that *must* obey the antisymmetry properties (\[fas\]) and (\[f31\]), as well as the Eqs. (\[relcoef\]), where the functions $t_{abc}$, $\alpha _{abc}^{A}$, $\alpha _{BCD}^{A}$, $\alpha _{abB}^{A}$ and $\alpha _{aBC}^{A}$ are defined in the formulas (\[f35a\]–\[f35e\]). At this stage, we have all the information necessary at the identification of the interacting gauge theory behind our deformation procedure. According to the general rules of the antifield-BRST formalism, the Lagrangian action that describes the coupled model is nothing but the antighost number zero piece from (\[f51\]), so it has the expression $$\bar{S}_{0}\left[ A_{\mu }^{a},H_{\mu }^{a},\varphi _{a},B_{a}^{\mu \nu },V_{\mu }^{A}\right] =\int d^{2}x\left( H_{\mu }^{a}D^{\mu }\varphi _{a}+% \frac{1}{2}B_{a}^{\mu \nu }\bar{F}_{\mu \nu }^{\prime a}-\frac{1}{4}\bar{F}% _{\mu \nu }^{A}\bar{F}_{A}^{\mu \nu }\right) , \label{f53}$$ where we used the notation $$\bar{F}_{\mu \nu }^{A}=\partial _{[\mu }V_{\nu ]}^{A}+g\left( f_{BC}^{A}V_{\mu }^{B}V_{\nu }^{C}+f_{ab}^{A}A_{\mu }^{a}A_{\nu }^{b}+g_{aB}^{A}A_{[\mu }^{a}V_{\nu ]}^{B}\right) . \label{f54}$$ The terms of antighost number one from (\[f51\]) offer us the generating set of deformed gauge transformations corresponding to the Lagrangian action (\[f53\]), or, in other words, the gauge symmetries of the interacting action, namely, $$\bar{\delta}_{\epsilon }A_{\mu }^{a}=\left( D_{\mu }\right) _{\;\;b}^{a}\epsilon ^{b}, \label{f55a}$$ $$\begin{aligned} \bar{\delta}_{\epsilon }H_{\mu }^{a} &=&\left( D^{\nu }\right) _{\;\;b}^{a}\epsilon _{\mu \nu }^{b}-g\left( \frac{\partial W_{bc}}{\partial \varphi _{a}}H_{\mu }^{c}-\frac{\partial ^{2}W_{bc}}{\partial \varphi _{a}\partial \varphi _{d}}B_{d\mu \nu }A^{c\nu }\right) \epsilon ^{b} \nonumber \\ &&+\bar{F}_{A\mu \nu }\left( \frac{\partial \left( D^{\nu }\right) _{\;\;b}^{A}}{\partial \varphi _{a}}\epsilon ^{b}+\frac{\partial \left( D^{\nu }\right) _{\;\;B}^{A}}{\partial \varphi _{a}}\epsilon ^{B}\right) , \label{f55b}\end{aligned}$$ $$\bar{\delta}_{\epsilon }\varphi _{a}=-gW_{ab}\epsilon ^{b},\;\bar{\delta}% _{\epsilon }V_{\mu }^{A}=\left( D_{\mu }\right) _{\;\;B}^{A}\epsilon ^{B}+\left( D_{\mu }\right) _{\;\;a}^{A}\epsilon ^{a}, \label{f55c}$$ $$\bar{\delta}_{\epsilon }B_{a}^{\mu \nu }=g\left( W_{ab}\epsilon ^{b\mu \nu }-% \frac{\partial W_{ab}}{\partial \varphi _{c}}B_{c}^{\mu \nu }\epsilon ^{b}\right) +g\left( g_{aB}^{A}\epsilon ^{B}+f_{ab}^{A}\epsilon ^{b}\right) \bar{F}_{A}^{\mu \nu }. \label{f55d}$$ There also appear two types of antighost number two elements in (\[f51\]). As we have stated in the end of Section 2, their presence indicates that the gauge algebra associated with the deformed gauge transformations is open, so the commutators among the gauge transformations (\[f55a\]–\[f55d\]) only close on the stationary surface of the field equations corresponding to the action (\[f53\]). Indeed, let $\epsilon ^{\alpha _{1}}=\left( \epsilon ^{a},\epsilon _{\mu \nu }^{a},\epsilon ^{A}\right) $ and $\xi ^{\alpha _{1}}=\left( \xi ^{a},\xi _{\mu \nu }^{a},\xi ^{A}\right) $ be two different sets of gauge parameters. Then, the expressions of the commutators between the deformed gauge transformations (\[f55a\]–\[f55d\]) associated with these parameters are completely determined from the antighost number two objects in (\[f51\]) under the form $$\left[ \bar{\delta}_{\epsilon },\bar{\delta}_{\xi }\right] \varphi _{a}=\bar{% \delta}_{\Lambda }\varphi _{a}, \label{com1}$$ $$\left[ \bar{\delta}_{\epsilon },\bar{\delta}_{\xi }\right] A_{\mu }^{a}=\bar{% \delta}_{\Lambda }A_{\mu }^{a}+g\frac{\delta \bar{S}_{0}}{\delta H^{d\mu }}% \frac{\partial ^{2}W_{bc}}{\partial \varphi _{a}\partial \varphi _{d}}% \epsilon ^{b}\xi ^{c}, \label{com2}$$ $$\begin{aligned} &&\left[ \bar{\delta}_{\epsilon },\bar{\delta}_{\xi }\right] B_{a}^{\mu \nu }=\bar{\delta}_{\Lambda }B_{a}^{\mu \nu }+2g^{2}\frac{\delta \bar{S}_{0}}{% \delta B_{d\mu \nu }}k_{AB}\left( \left( g_{aC}^{A}g_{dD}^{B}-g_{aD}^{A}g_{dC}^{B}\right) \epsilon ^{C}\xi ^{D}\right. \nonumber \\ &&\left. +f_{a[b}^{A}f_{c]d}^{B}\epsilon ^{b}\xi ^{c}-\left( g_{aC}^{A}f_{bd}^{B}+f_{ab}^{A}g_{dC}^{B}\right) \left( \epsilon ^{b}\xi ^{C}-\xi ^{b}\epsilon ^{C}\right) \right) \nonumber \\ &&+g^{2}\delta _{\alpha }^{[\mu }\delta _{\beta }^{\nu ]}\frac{\delta \bar{S}% _{0}}{\delta H_{\beta }^{d}}k_{AB}\left( A^{b\alpha }\left( g_{aC}^{A}\frac{% \partial g_{bD}^{B}}{\partial \varphi _{d}}-g_{aD}^{A}\frac{\partial g_{bC}^{B}}{\partial \varphi _{d}}\right) \epsilon ^{C}\xi ^{D}\right. \nonumber \\ &&-V^{E\alpha }g_{a[C}^{A}\frac{\partial f_{D]E}^{B}}{\partial \varphi _{d}}% \epsilon ^{C}\xi ^{D}-\left( A^{e\alpha }f_{a[b}^{A}\frac{\partial f_{c]e}^{B}}{\partial \varphi _{d}}+V^{E\alpha }f_{a[b}^{A}\frac{\partial g_{c]E}^{B}}{\partial \varphi _{d}}\right) \epsilon ^{b}\xi ^{c} \nonumber \\ &&+\left( A^{c\alpha }\left( g_{aC}^{A}\frac{\partial f_{bc}^{B}}{\partial \varphi _{d}}+f_{ab}^{A}\frac{\partial g_{cC}^{B}}{\partial \varphi _{d}}% \right) \right. \nonumber \\ &&\left. \left. -V^{D\alpha }\left( g_{aC}^{A}\frac{\partial g_{bD}^{B}}{% \partial \varphi _{d}}-f_{ab}^{A}\frac{\partial f_{CD}^{B}}{\partial \varphi _{d}}\right) \right) \left( \epsilon ^{b}\xi ^{C}-\xi ^{b}\epsilon ^{C}\right) \right) , \label{com3}\end{aligned}$$ $$\begin{aligned} &&\left[ \bar{\delta}_{\epsilon },\bar{\delta}_{\xi }\right] H_{\mu }^{a}=% \bar{\delta}_{\Lambda }H_{\mu }^{a}-g\frac{\delta \bar{S}_{0}}{\delta A^{d\mu }}\frac{\partial ^{2}W_{bc}}{\partial \varphi _{a}\partial \varphi _{d}}\epsilon ^{b}\xi ^{c} \nonumber \\ &&+g\frac{\delta \bar{S}_{0}}{\delta V^{A\mu }}\left( \frac{\partial f_{BC}^{A}}{\partial \varphi _{a}}\epsilon ^{B}\xi ^{C}+\frac{\partial f_{bc}^{A}}{\partial \varphi _{a}}\epsilon ^{b}\xi ^{c}+\frac{\partial g_{bB}^{A}}{\partial \varphi _{a}}\left( \epsilon ^{b}\xi ^{B}-\xi ^{b}\epsilon ^{B}\right) \right) \nonumber \\ &&+g^{2}\frac{\delta \bar{S}_{0}}{\delta B_{d}^{\mu \alpha }}k_{AB}\left( A^{b\alpha }\left( g_{dC}^{A}\frac{\partial g_{bD}^{B}}{\partial \varphi _{a}% }-g_{dD}^{A}\frac{\partial g_{bC}^{B}}{\partial \varphi _{a}}\right) \epsilon ^{C}\xi ^{D}\right. \nonumber \\ &&-V^{E\alpha }g_{d[C}^{A}\frac{\partial f_{D]E}^{B}}{\partial \varphi _{a}}% \epsilon ^{C}\xi ^{D}-\left( A^{e\alpha }f_{d[b}^{A}\frac{\partial f_{c]e}^{B}}{\partial \varphi _{a}}+V^{E\alpha }f_{d[b}^{A}\frac{\partial g_{c]E}^{B}}{\partial \varphi _{a}}\right) \epsilon ^{b}\xi ^{c} \nonumber \\ &&+\left( A^{c\alpha }\left( g_{dC}^{A}\frac{\partial f_{bc}^{B}}{\partial \varphi _{a}}+f_{db}^{A}\frac{\partial g_{cC}^{B}}{\partial \varphi _{a}}% \right) -V^{D\alpha }\left( g_{dC}^{A}\frac{\partial g_{bD}^{B}}{\partial \varphi _{a}}-f_{db}^{A}\frac{\partial f_{CD}^{B}}{\partial \varphi _{a}}% \right) \right) \nonumber \\ &&\left. \times \left( \epsilon ^{b}\xi ^{C}-\xi ^{b}\epsilon ^{C}\right) \right) +g\frac{\delta \bar{S}_{0}}{\delta H_{\mu }^{e}}\left( -\frac{% \partial ^{2}W_{bc}}{\partial \varphi _{a}\partial \varphi _{e}}\left( \epsilon ^{b}\xi _{\mu \nu }^{c}-\xi ^{b}\epsilon _{\mu \nu }^{c}\right) \right. \nonumber \\ &&+\left( \frac{\partial ^{3}W_{bc}}{\partial \varphi _{a}\partial \varphi _{d}\partial \varphi _{e}}B_{d\mu \nu }-\frac{\partial ^{2}f_{bc}^{A}}{% \partial \varphi _{a}\partial \varphi _{e}}\bar{F}_{A\mu \nu }\right) \epsilon ^{b}\xi ^{c} \nonumber \\ &&\left. -\bar{F}_{A\mu \nu }\left( \frac{\partial ^{2}g_{bB}^{A}}{\partial \varphi _{a}\partial \varphi _{e}}\left( \epsilon ^{b}\xi ^{B}-\xi ^{b}\epsilon ^{B}\right) +\frac{\partial ^{2}f_{BC}^{A}}{\partial \varphi _{a}\partial \varphi _{e}}\epsilon ^{B}\xi ^{C}\right) \right) \nonumber \\ &&+g^{2}\delta _{\mu }^{[\alpha }\delta _{\nu }^{\beta ]}M_{\alpha }^{ab\nu }% \frac{\delta \bar{S}_{0}}{\delta H^{b\beta }}, \label{com4}\end{aligned}$$ $$\begin{aligned} &&\left[ \bar{\delta}_{\epsilon },\bar{\delta}_{\xi }\right] V_{\mu }^{A}=% \bar{\delta}_{\Lambda }V_{\mu }^{A}-g\frac{\delta \bar{S}_{0}}{\delta H^{a\mu }}\left( \frac{\partial f_{BC}^{A}}{\partial \varphi _{a}}\epsilon ^{B}\xi ^{C}+\frac{\partial f_{bc}^{A}}{\partial \varphi _{a}}\epsilon ^{b}\xi ^{c}\right. \nonumber \\ &&\left. +\frac{\partial g_{bB}^{A}}{\partial \varphi _{a}}\left( \epsilon ^{b}\xi ^{B}-\xi ^{b}\epsilon ^{B}\right) \right) , \label{com5}\end{aligned}$$ where $$\Lambda ^{\alpha _{1}}=\left( \Lambda ^{a},\Lambda _{\mu \nu }^{a},\Lambda ^{A}\right) , \label{not0}$$ with $$\Lambda ^{a}\equiv g\frac{\partial W_{bc}}{\partial \varphi _{a}}\epsilon ^{b}\xi ^{c}, \label{not1}$$ $$\begin{aligned} \Lambda _{\mu \nu }^{a} &\equiv &g\left( \frac{\partial f_{bc}^{A}}{\partial \varphi _{a}}\bar{F}_{A\mu \nu }-\frac{\partial ^{2}W_{bc}}{\partial \varphi _{a}\partial \varphi _{d}}B_{d\mu \nu }\right) \epsilon ^{b}\xi ^{c}+g\frac{% \partial f_{BC}^{A}}{\partial \varphi _{a}}\bar{F}_{A\mu \nu }\epsilon ^{B}\xi ^{C} \nonumber \\ &&+g\frac{\partial g_{bB}^{A}}{\partial \varphi _{a}}\bar{F}_{A\mu \nu }\left( \epsilon ^{b}\xi ^{B}-\xi ^{b}\epsilon ^{B}\right) -g\frac{\partial W_{bc}}{\partial \varphi _{a}}\left( \epsilon ^{b}\xi _{\mu \nu }^{c}-\xi ^{b}\epsilon _{\mu \nu }^{c}\right) , \label{not2}\end{aligned}$$ and respectively $$\Lambda ^{A}\equiv g\left( f_{BC}^{A}\epsilon ^{B}\xi ^{C}+f_{ab}^{A}\epsilon ^{a}\xi ^{b}+g_{bB}^{A}\left( \epsilon ^{b}\xi ^{B}-\xi ^{b}\epsilon ^{B}\right) \right) . \label{not3}$$ At the same time, the function from (\[com4\]) denoted by $M_{\alpha }^{ab\nu }$ reads $$\begin{aligned} M_{\alpha }^{ab\nu } &=&-k_{AB}\left( \left( \left( \frac{\partial g_{cC}^{A}% }{\partial \varphi _{b}}V_{\alpha }^{C}+\frac{\partial f_{ce}^{A}}{\partial \varphi _{b}}A_{\alpha }^{e}\right) \left( \frac{\partial g_{dC}^{B}}{% \partial \varphi _{a}}V^{C\nu }+\frac{\partial f_{df}^{B}}{\partial \varphi _{a}}A^{f\nu }\right) \right. \right. \nonumber \\ &&\left. -\left( \frac{\partial g_{dC}^{A}}{\partial \varphi _{b}}V_{\alpha }^{C}+\frac{\partial f_{de}^{A}}{\partial \varphi _{b}}A_{\alpha }^{e}\right) \left( \frac{\partial g_{cC}^{B}}{\partial \varphi _{a}}V^{C\nu }+\frac{\partial f_{cf}^{B}}{\partial \varphi _{a}}A^{f\nu }\right) \right) \epsilon ^{c}\xi ^{d} \nonumber \\ &&+\left( \left( \frac{\partial f_{CE}^{A}}{\partial \varphi _{b}}V_{\alpha }^{E}-\frac{\partial g_{eC}^{A}}{\partial \varphi _{b}}A_{\alpha }^{e}\right) \left( \frac{\partial f_{DF}^{B}}{\partial \varphi _{a}}V^{F\nu }-\frac{\partial g_{dD}^{B}}{\partial \varphi _{a}}A^{d\nu }\right) \right. \nonumber \\ &&\left. -\left( \frac{\partial f_{DE}^{A}}{\partial \varphi _{b}}V_{\alpha }^{E}-\frac{\partial g_{cD}^{A}}{\partial \varphi _{b}}A_{\alpha }^{e}\right) \left( \frac{\partial f_{CF}^{B}}{\partial \varphi _{a}}V^{F\nu }-\frac{\partial g_{dC}^{B}}{\partial \varphi _{a}}A^{d\nu }\right) \right) \epsilon ^{C}\xi ^{D} \nonumber \\ &&+\left( \left( \frac{\partial g_{cD}^{A}}{\partial \varphi _{b}}V_{\alpha }^{D}+\frac{\partial f_{ce}^{A}}{\partial \varphi _{b}}A_{\alpha }^{e}\right) \left( \frac{\partial f_{CF}^{B}}{\partial \varphi _{a}}V^{F\nu }-\frac{\partial g_{dC}^{B}}{\partial \varphi _{a}}A^{d\nu }\right) \right. \nonumber \\ &&\left. -\left( \frac{\partial f_{CE}^{A}}{\partial \varphi _{b}}V_{\alpha }^{E}-\frac{\partial g_{eC}^{A}}{\partial \varphi _{b}}A_{\alpha }^{e}\right) \left( \frac{\partial g_{cD}^{B}}{\partial \varphi _{a}}V^{D\nu }+\frac{\partial f_{cd}^{B}}{\partial \varphi _{a}}A^{d\nu }\right) \right) \times \nonumber \\ &&\left. \times \left( \epsilon ^{c}\xi ^{C}-\xi ^{c}\epsilon ^{C}\right) \right) . \label{not7}\end{aligned}$$ From the terms of antighost numbers three and four present in (\[f51\]) we can recover the higher-order structure functions due to the open character of the deformed gauge algebra, as well as the accompanying identities. They have an intricate (but not illuminating) form and consequently we will omit writing their concrete expressions. At this point, we have all the information on the gauge structure of the deformed model, whose free limit is given by the Lagrangian action (\[f1\]), together with the abelian and irreducible gauge symmetries (\[f3\]). We observe that there are two main types of vertices in the deformed action (\[f53\]). The first kind $$g\left( H_{\mu }^{a}W_{ab}A^{b\mu }+\frac{1}{2}B_{a}^{\mu \nu }\frac{% \partial W_{bc}}{\partial \varphi _{a}}A_{\mu }^{b}A_{\nu }^{c}\right) , \label{purebf}$$ corresponds to the self-interactions among the purely BF fields in the absence of the vector fields $\left\{ V_{\mu }^{A}\right\} $, being given only by terms of order one in the coupling constant. Such terms have been previously obtained in the literature and we will not insist on their structure (for a detailed analysis, see for instance [@mpla]). The second kind of vertices can be written in the form $$\begin{aligned} &&-gk_{AB}\left( \partial ^{\mu }V^{\nu A}\right) \left( f_{CD}^{B}\left( \varphi \right) V_{\mu }^{C}V_{\nu }^{D}+f_{ab}^{B}\left( \varphi \right) A_{\mu }^{a}A_{\nu }^{b}+g_{aC}^{B}\left( \varphi \right) A_{[\mu }^{a}V_{\nu ]}^{C}\right) \nonumber \\ &&-\frac{g^{2}}{4}k_{AD}\left( f_{BC}^{A}\left( \varphi \right) V_{\mu }^{B}V_{\nu }^{C}+f_{ab}^{A}\left( \varphi \right) A_{\mu }^{a}A_{\nu }^{b}+g_{aB}^{A}\left( \varphi \right) A_{[\mu }^{a}V_{\nu ]}^{B}\right) \times \nonumber \\ &&\times \left( f_{EF}^{D}\left( \varphi \right) V^{E\mu }V^{F\nu }+f_{cd}^{D}\left( \varphi \right) A^{c\mu }A^{d\nu }+g_{cE}^{D}\left( \varphi \right) A^{c[\mu }V^{\nu ]E}\right) . \label{crosscoup}\end{aligned}$$ We note that (\[crosscoup\]) contains a vertex involving only the BF fields, namely $-\frac{1}{4}k_{AD}f_{ab}^{A}\left( \varphi \right) f_{cd}^{D}\left( \varphi \right) A_{\mu }^{a}A_{\nu }^{b}A^{c\mu }A^{d\nu }$, whose existence is induced by the presence of the vector fields $\left\{ V_{\mu }^{A}\right\} $. Indeed, if the vector fields $\left\{ V_{\mu }^{A}\right\} $ were absent ($k_{AB}=0$), then this term would vanish. The remaining terms reveal the cross-couplings between the BF fields and the vector fields $\left\{ V_{\mu }^{A}\right\} $. Among the cross-coupling pieces in (\[crosscoup\]), we find generalized cubic and quartic Yang-Mills-like vertices in ‘backgrounds’ of the scalar fields. We remark that neither the one-forms $\left\{ H_{\mu }^{a}\right\} $ nor the two-forms $\left\{ B_{a}^{\mu \nu }\right\} $ can be coupled in a consistent, non-trivial manner to the vector fields. Related to the deformed gauge transformations (\[f55a\]–\[f55d\]), there appears a complementary situation, in the sense that among the BF fields, only the one-forms $% \left\{ H_{\mu }^{a}\right\} $ and the two-forms $\left\{ B_{a}^{\mu \nu }\right\} $ gain gauge symmetries involving the parameters $\epsilon ^{B}$. The deformed gauge transformations of the vector fields $\left\{ V_{\mu }^{A}\right\} $ have a rich structure, including, besides other terms, the generalized covariant derivative $\left( \delta _{\;\;B}^{A}\partial _{\mu }-gf_{BC}^{A}\left( \varphi \right) V_{\mu }^{C}\right) $ with respect to the parameters $\epsilon ^{B}$. Some solutions to the Eqs. (\[relcoef\]) ======================================== We have seen that the deformation procedure developed so far essentially relies on the existence of four types of functions depending on the undifferentiated scalar fields, namely, $W_{ab}$, $f_{BC}^{A}$, $f_{ab}^{A}$ and $g_{aB}^{A}$, which are subject on the one hand to the conditions (\[fas\]) plus (\[f31\]) and on the other hand to the Eqs. (\[relcoef\]). In the sequel we analyze two classes of solutions to the above conditions and equations and emphasize that they admit an interesting geometric and algebraic interpretation. The first class of solutions corresponds to a non-vanishing $W_{ab}\left( \varphi \right) $, while the second kind is associated with $W_{ab}=0$. Related to the former type of solutions, it is clear that the first equation from (\[relcoef\]) together with the first antisymmetry property in (\[fas\]) $$W_{e[a}\frac{\partial W_{bc]}}{\partial \varphi _{e}}=0,\;W_{ab}=-W_{ba}, \label{f43a}$$ shows that the antisymmetric functions $W_{ab}$ of the undifferentiated scalar fields satisfy the Jacobi’s identity for a nonlinear algebra. Let us see the geometric meaning of $W_{ab}$. To this end, we briefly review the basic notions on Poisson manifolds. If $N$ denotes an arbitrary Poisson manifold, then this is equipped with a Poisson bracket $\left\{ ,\right\} $ that is bilinear, antisymmetric, subject to a Leibnitz-like rule and satisfies a Jacobi-type identity. If $\left\{ X^{i}\right\} $ are some local coordinates on $N$, then there exists a two-tensor $\mathcal{P}^{ij}\equiv \left\{ X^{i},X^{j}\right\} $ (the Poisson tensor) that uniquely determines the Poisson structure together with the Leibnitz rule. This two-tensor is antisymmetric and transforms covariantly under coordinate transformations. Jacobi’s identity for the Poisson bracket $\left\{ ,\right\} $ expressed in terms of the Poisson tensor reads as $\mathcal{P}_{,k}^{ij}\mathcal{P}^{kl}+% \mathrm{cyclic}\left( i,j,l\right) =0$, where $\mathcal{P}_{,k}^{ij}\equiv \partial \mathcal{P}^{ij}/\partial X^{k}$. Now, the geometric origin of $% W_{ab}$ is obvious. If, for instance, we choose a concrete form for the antisymmetric functions $W_{ab}\left( \varphi \right) $ that satisfy (\[f43a\]), then we can interpret the dynamical scalar fields $\left\{ \varphi _{a}\right\} $ precisely like some local coordinates on a target manifold endowed with a prescribed Poisson structure (up to the plain convention that the lower index $a$ is a ‘covariant’ index of the type $i$). Conversely, any given Poisson manifold parametrized in terms of some local coordinates $% \left\{ \varphi _{a}\right\} $ (within the same index convention) prescribes a Poisson tensor $W_{ab}\left( \varphi \right) $ which is antisymmetric and satisfies (\[f43a\]). Once we have fixed the functions $W_{ab}$, it is easy to see that a solution for the remaining coefficients ($f_{BC}^{A}$, $% f_{ab}^{A}$ and $g_{aB}^{A}$) is represented by $$f_{BC}^{A}=\bar{f}_{BC}^{A},\;g_{aB}^{A}=\bar{f}_{BE}^{A}M^{E}X_{a}\left( \varphi _{c}\right) , \label{f44a}$$ $$f_{ab}^{A}=M^{A}\left( X_{c}\frac{\partial W_{ab}}{\partial \varphi _{c}}% +W_{ca}\frac{\partial X_{b}}{\partial \varphi _{c}}+W_{bc}\frac{\partial X_{a}}{\partial \varphi _{c}}\right) , \label{f44b}$$ where $M^{E}$ are some real constants, $\left\{ X_{a}\right\} $ stands for a set of arbitrary functions depending only on the undifferentiated scalar fields $\varphi _{a}$, and $\bar{f}_{BC}^{A}$ are some real, antisymmetric constants, that obey the identity $$\bar{f}_{E[B}^{A}\bar{f}_{CD]}^{E}=0. \label{f45}$$ Accordingly, $\bar{f}_{BC}^{A}$ can be viewed like the structure constants of a semi-simple Lie algebra, endowed with the Killing-Cartan metric $k_{AB}$, while $M^{A}$ can be seen like the components of an arbitrary element from this Lie algebra. In this situation, the deformed Lagrangian action (\[f53\]) also includes self-interactions among the vector fields $\left\{ V_{\mu }^{A}\right\} $ precisely described by cubic and quartic Yang-Mills vertices. Accordingly, the gauge transformations of $V_{\mu }^{A}$ contain the well-known covariant derivative of the gauge parameters $\epsilon ^{A}$$$\bar{\delta}_{\epsilon }V_{\mu }^{A}=\left( \delta _{\;\;B}^{A}\partial _{\mu }-g\bar{f}_{BC}^{A}V_{\mu }^{C}\right) \epsilon ^{B}+``\mathrm{more}". \label{f45abcdef}$$ Next, we examine the latter kind of solutions (corresponding to $W_{ab}=0$), in which case the Eqs. (\[relcoef\]) become $$f_{E[B}^{A}f_{CD]}^{E}=0, \label{relc1}$$ $$f_{[ab}^{E}g_{c]E}^{A}=0,% \;f_{EB}^{A}f_{ab}^{E}+g_{aB}^{E}g_{bE}^{A}-g_{bB}^{E}g_{aE}^{A}=0, \label{relcn}$$ $$g_{aE}^{A}f_{BC}^{E}+f_{EB}^{A}g_{aC}^{E}-f_{EC}^{A}g_{aB}^{E}=0. \label{relc3}$$ The solution to (\[relcn\]–\[relc3\]) takes the form $$f_{ab}^{A}=f_{BC}^{A}M_{a}^{B}M_{b}^{C},\;g_{aB}^{A}=f_{CB}^{A}M_{a}^{C}, \label{f56}$$ where $M_{b}^{B}$ are some arbitrary functions of the undifferentiated scalar fields and $f_{BC}^{A}$ verify the Eq. (\[relc1\]). In order to solve the remaining equation, namely, (\[relc1\]), let $\bar{M}_{A}^{a}$ be some functions of the scalar fields such that $\bar{M}_{A}^{a}M_{b}^{A}=% \delta _{b}^{a}$. Then, the solution of (\[relc1\]) reads as $$f_{BC}^{A}=f_{bc}^{a}M_{a}^{A}\bar{M}_{B}^{b}\bar{M}_{C}^{c}, \label{f57}$$ where $f_{bc}^{a}$ are the structure constants of a semi-simple Lie algebra with the Killing-Cartan metric $k_{ab}$. It is easy to see that the Jacobi identity (\[relc1\]) is a consequence of the Jacobi identity for the structure constants $f_{bc}^{a}$. For the functions $g_{aB}^{A}$ and $% f_{BC}^{A}$ given in (\[f56\]–\[f57\]) to satisfy the antisymmetry properties (\[fas\]) and (\[f31\]), it is necessary that $k_{AB}$ and $% k_{ab}$ are correlated through some relations of the type $$k_{AB}M_{a}^{A}M_{b}^{B}=k_{ab}\Phi \left( \varphi \right) , \label{f58}$$ with $\Phi \left( \varphi \right) $ a non-vanishing, but otherwise arbitrary function of the scalar fields. We remark that, although the functions $% f_{BC}^{A}$ from (\[f57\]) depend in general on the scalar fields, they however verify the Jacobi identity (\[relc1\]). Accordingly, these functions can be regarded like some ‘structure constants’ of a Lie algebra whose generators depend on the scalar fields (generalized Lie algebra). It is interesting to note that the gauge algebra is open also for the latter kind of solutions. In both cases, the entire gauge structure of the interacting model can be obtained by substituting the solutions (\[f44a\]–\[f44b\]) and respectively (\[f56\]–\[f57\]) in the formulas (\[f53\]) and (\[f55a\]–\[com5\]). Conclusion ========== To conclude with, in this paper we have investigated the consistent two-dimensional interactions that can be introduced among a set of scalar fields, two types of one-forms, a system of two-forms and a collection of vector fields, described in the free limit by an abelian BF theory and a sum of Maxwell actions. Starting with the BRST differential for the free theory, $s=\delta +\gamma $, we compute the consistent first-order deformation of the solution to the master equation with the help of some cohomological techniques, and obtain that it is parametrized by five kinds of functions depending on the undifferentiated scalar fields. Next, we investigate the second-order deformation, whose existence reduces the number of independent types of functions on the scalar fields to four and, meanwhile, requires that these are subject to certain equations. Based on these restrictions, we determine the expression of the second-order deformation and, moreover, show that we can take all the remaining higher-order deformations to vanish. As a consequence of our procedure, we are led to an interacting gauge theory with deformed gauge transformations and a non-abelian gauge algebra that only closes on-shell. The presence of the collection of vector fields brings in a rich structure of non-trivial terms if compared with the self-interactions that can be added to a two-dimensional abelian BF theory [@mpla]. Finally, we give two classes of solutions to the equations satisfied by the various functions of the scalar fields that parametrize the deformed solution to the master equation, which can be interpreted in terms of Poisson manifolds and respectively of generalized Lie algebras. Acknowledgment {#acknowledgment .unnumbered} ============== The authors wish to thank C. Bizdadea and S. O. Saliu for useful discussions and comments. This work has been supported by the type $\mathrm{A}_{\mathrm{T% }}$ grant 33547/2003, code 302/2003, from the Romanian Council for Academic Scientific Research (CNCSIS) and the Romanian Ministry of Education, Research and Youth (MECT). [99]{} B. Voronov and I. V. Tyutin, *Theor. Math. Phys.* **50**, 218 (1982); **52**, 628 (1982). J. Gomis and S. Weinberg, *Nucl. Phys.* **B469**, 473 (1996). S. Weinberg, *The Quantum Theory of Fields* (Cambridge University Press, Cambridge, 1996). O. Piguet and S. P. Sorella, *Algebraic Renormalization: Perturbative Renormalization, Symmetries and Anomalies* (Springer, Berlin, 1995), Lecture Notes in Physics, Vol. 28. P. S. Howe, V. Lindstrőm and P. White, *Phys. Lett.* **B246**, 430 (1990). W. Troost, P. van Nieuwenhuizen and A. van Proeyen, *Nucl. Phys.* **B333**, 727 (1990). G. Barnich and M. Henneaux, *Phys. Rev. Lett.* **72**, 1588 (1994). G. Barnich, *Mod. Phys. Lett.* **A9**, 665 (1994); *Phys. Lett.* **B419**, 211 (1998). F. Brandt, M. Henneaux and A. Wilch, *Phys. Lett.* **B387**, 320 (1996). R. Arnowitt and S. Deser, *Nucl. Phys.* **49**, 133 (1963). J. Fang and C. Fronsdal, *J. Math. Phys.* **20**, 2264 (1979). F. A. Berends, G. H. Burgers and H. Van Dam, *Z. Phys.* **C24**, 247 (1984); *Nucl. Phys.* **B260**, 295 (1985). A. K. H. Bengtsson, *Phys. Rev.* **D32**, 2031 (1985). G. Barnich and M. Henneaux, *Phys. Lett.* **B311**, 123 (1993). J. D. Stasheff, q-alg/9702012; hep-th/9712157. J. A. Garcia and B. Knaepen, *Phys. Lett.* **B441**, 198 (1998). D. Birmingham, M. Blau, M. Rakowski and G. Thompson, *Phys. Rept.* **209**, 129 (1991). C. Bizdadea, *Mod. Phys. Lett.* **A15**, 2047 (2000). C. Bizdadea, E. M. Cioroianu and S. O. Saliu, *Int. J. Mod. Phys.* **A17**, 2191 (2002). C. Bizdadea, C. C. Ciobirca, E. M. Cioroianu, S. O. Saliu and S. C. Sararu, *JHEP* **0301**, 049 (2003). M. O. Katanaev and I. V. Volovich, *Phys. Lett.* **B175**, 413 (1986); *Annals Phys.* **197**, 1 (1990). S. N. Solodukhin, *Class. Quantum Grav.* **10**, 1011 (1993). H.-J. Schmidt, *J. Math. Phys.* **32**, 1562 (1991). N. Ikeda and K. I. Izawa, *Prog. Theor. Phys.* **90**, 237 (1993). D. Grumiller, W. Kummer and D. V. Vassilevich, *Phys. Rept.* **369**, 327 (2002). J. Brown, *Lower Dimensional Gravity* (World Scientific, Singapore, 1988). C. Teitelboim, *Phys. Lett.* **B126**, 41 (1983); R. Jackiw, *Nucl. Phys.* **B252**, 343 (1985). N. Ikeda, *Annals Phys.* **235**, 435 (1994). T. Strobl, *Phys. Rev.* **D50**, 7346 (1994). P. Schaller and T. Strobl, *Mod. Phys. Lett.* **A9**, 3129 (1994). A. Yu. Alekseev, P. Schaller and T. Strobl, *Phys. Rev.* **D52**, 7146 (1995). A. S. Cattaneo and G. Felder, *Mod. Phys. Lett.* **A16**, 179 (2001); *Commun. Math. Phys.* **212**, 591 (2000). T. Kloesch and T. Strobl, *Class. Quantum Grav.* **13**, 965 (1996); **13**, 2395 (1996); **14**, 1689 (1997). T. Strobl, hep-th/0011240. N. Ikeda, *JHEP* **0011**, 009 (2000); **0107**, 037 (2001). G. Barnich, F. Brandt and M. Henneaux, *Commun. Math. Phys.* **174**, 93 (1995). N. Boulanger, T. Damour, L. Gualtieri and M. Henneaux, *Nucl. Phys.* **B597**, 127 (2001). X. Bekaert, N. Boulanger and M. Henneaux, *Phys. Rev.* **D67**, 044010 (2003). G. Barnich, F. Brandt and M. Henneaux, *Commun. Math. Phys.* **174**, 57 (1995); *Phys. Rept.* **338**, 439 (2000). M. Henneaux, *Commun. Math. Phys.* **140**, 1 (1991). [^1]: E-mail address: manache@central.ucv.ro [^2]: E-mail address: scsararu@central.ucv.ro [^3]: We note that the local cohomology group of the Koszul-Tate differential at strictly positive pure ghost *and* antighost numbers is trivial, so the notations $H_{J}\left( \delta |d\right) $ and $H\left( \delta |d\right) $ automatically take into consideration only objects of pure ghost number zero (see, for instance, [@gen1] and [@commun1]).

--- abstract: 'Unsupervised node embedding methods (e.g., DeepWalk, LINE, and node2vec) have attracted growing interests given their simplicity and effectiveness. However, although these methods have been proved effective in a variety of applications, none of the existing work has analyzed the robustness of them. This could be very risky if these methods are attacked by an adversarial party. In this paper, we take the task of link prediction as an example, which is one of the most fundamental problems for graph analysis, and introduce a data poisoning attack to node embedding methods. We give a complete characterization of attacker’s utilities and present efficient solutions to adversarial attacks for two popular node embedding methods: DeepWalk and LINE. We evaluate our proposed attack model on multiple real-world graphs. Experimental results show that our proposed model can significantly affect the results of link prediction by slightly changing the graph structures (e.g., adding or removing a few edges). We also show that our proposed model is very general and can be transferable across different embedding methods. Finally, we conduct a case study on a coauthor network to better understand our attack method.' author: - | Mingjie Sun$^{1}{}$[^1] , Jian Tang$^2$, Huichen Li$^3$, Bo Li$^3$, Chaowei Xiao$^4$, Yao Chen$^5$, Dawn Song$^6$\ $^1$Tsinghua University  $^2$Mila & HEC Montreal, Canada  $^3$UIUC  $^4$University of Michigan\ $^5$Tecent  $^6$University of California, Berkeley\ bibliography: - 'iclr2019\_conference.bib' title: Data Poisoning attack against Unsupervised Node Embedding Methods --- Introduction ============ Node representations, which represent each node with a low-dimensional vector, have been proved effective in a variety of applications such as node classification [@perozzi2014deepwalk], link prediction [@grover2015node2vec], and visualization [@tang2016visualizing]. Some popular node embedding methods include DeepWalk [@perozzi2014deepwalk], LINE [@tang2015LINE], and node2vec [@grover2015node2vec]. These methods learn the node embeddings by preserving graph structures, which do not depend on specific tasks. As a result, the learned node embeddings are very general and can be potentially useful to multiple downstream tasks. However, although these methods are very effective and have been used for a variety of tasks, none of the existing work has studied the robustness of these methods. As a result, these methods are susceptible to a risk of being maliciously attacked. Take the task of link prediction in a social network (e.g., Twitter) as an example, which is one of the most important applications of node embedding methods. A malicious party may create malicious users in a social network and attack the graph structures (e.g., adding and removing edges) so that the effectiveness of node embedding methods is maximally degraded. For example, the attacker may slightly change the graph structures (e.g., following more users) so that the probability of a specific user to be recommended/linked can be significantly increased or decreased. Such a kind of attack is known as *data poisoning*. In this paper we are interested in the robustness of the node embedding methods w.r.t. data poisoning, and their vulnerability to the adversarial attacks in the worst case. We are inspired by existing literature on adversarial attack, which has been extensively studied for different machine learning systems [@szegedy2013intriguing; @goodfellow2014explaining; @moosavi2016deepfool; @carlini2017towards; @xiao2018spatially; @xiao2018generating; @xiao2018characterizing; @xie2017adversarial; @cisse2017houdini; @yang2018realistic]. Specifically, it has been shown that deep neural networks are very sensitive to adversarial attacks, which can significantly change the prediction results by slightly perturbing the input data. However, most of existing work on adversarial attack focus on image [@szegedy2013intriguing; @goodfellow2014explaining; @moosavi2016deepfool; @carlini2017towards; @xiao2018spatially] and text data [@cheng2018seq2sick; @jia2017adversarial], which are independently distributed while this work focuses on graph data. There are some very recent work which studied adversarial attack for graph data [@ICML2018graph; @KDD2018graph]. However, these work mainly studied graph neural networks, which are supervised methods, and the gradients for changing the output label can be leveraged. Therefore, in this paper we are looking for an approach that is able to attack the unsupervised node embedding methods for graphs. In this paper, we introduce a systematic approach to adversarial attacks against unsupervised node embedding methods. We assume that the attacker can poison the graph structures by either removing or adding edges. Two types of adversarial goals are studied including *integrity attack*, which aims to attack the probabilities of specific links, and *availability attack*, which aims to increase overall prediction errors. We propose a unified optimization framework based on projected gradient descent to optimally attack both goals. In addition, we conduct a case study on a coauthor network to better understand our attack method. To summarize, we make the following contributions: - We formulate the problem of attacking unsupervised node embeddings for the task of link prediction and introduce a complete characterization of attacker utilities. - We propose an efficient algorithm based on projected gradient descent to attack unsupervised node embedding algorithms, specifically DeepWalk and LINE, based on the first order Karush Kuhn Tucker (KKT) conditions. - We conduct extensive experiments on real-world graphs to show the efficacy of our proposed attack model on the task of link prediction. Moreover, results show that our proposed attack model is transferable across different node embedding methods. - Finally, we conduct a case study on a coauthor network and give an intuitive understanding of our attack method. [^1]: Work done while visiting UC Berkeley.

--- abstract: 'We present infrared observations of the young, oxygen-rich supernova remnant 1E 0102.2-7219 (E0102) in the Small Magellanic Cloud, obtained with the [*Spitzer Space Telescope*]{}. The remnant is detected at 24 $\mu$m but not at 8 or 70 $\mu$m and has a filled morphology with two prominent filaments. We find evidence for the existence of up to $8\times10^{-4}$ M$_{\odot}$ of hot dust ($T_{d}\sim120$ K) associated with the remnant. Most of the hot dust is located in the central region of E0102 which appears significantly enhanced in infrared and radio continuum emission relative to the X–ray emission. Even if [*all*]{} of the hot dust was formed in the explosion of E0102, the estimated mass of dust is at least 100 times lower that what is predicted by some recent theoretical models.' author: - 'Snežana Stanimirović$^{1}$, Alberto D. Bolatto$^{1}$, Karin Sandstrom$^{1}$, Adam K. Leroy$^{1}$, Joshua D. Simon$^{1,2}$, B. M. Gaensler$^{3}$, Ronak Y. Shah$^{4}$, James M. Jackson$^{4}$' title: '[*Spitzer Space Telescope*]{} detection of the young supernova remnant 1E 0102.2-7219' --- Introduction ============ Young supernova remnants (SNRs) are exciting laboratories for studying energetic phenomena occurring in the interstellar medium (ISM) such as cosmic ray acceleration, nucleosynthesis, dust formation and destruction, and collisionless shock physics [@McKee01; @Vink04]. One small subclass of young SNRs, with less than a dozen members, is the so called oxygen–rich SNRs. These remnants have very high–velocity debris ($V>1000$ ) and high abundances of oxygen, neon, carbon, and magnesium. They are believed to be remnants of the most massive stars (typical mass of $>15$ M$_{\odot}$). The oxygen-rich SNRs arise mainly from Type Ib/c supernovae, and their typical representative is Cas A. Another very young oxygen-rich SNR, found in the Small Magellanic Cloud (SMC), is 1E 0102.2-7219 (hereafter E0102), commonly considered as dynamically very similar to Cas A. E0102 is a well studied object at optical, radio, and X–ray wavelengths. Its kinematic age is approximately 1000 years, making it one of the youngest SNRs known. E0102 is located $\sim15$ pc in projected distance NE of the edge of the massive star–forming region LHA 115-N 76C (N 76, Henize 1956). The remnant is a prominent radio source at 6 cm, with a shell-like morphology and a diameter of $40\arcsec$ [@Amy93]. X–ray images of E0102 show a faint, filled circular structure that marks the location of hot gas associated with the forward moving blast wave. The bright X–ray ring — mainly due to strong emission lines of O, Ne and Mg — marks the interaction of the reverse shock with the stellar ejecta and shows significant substructure: a bright knot in the southwest with a radial extension or “spoke”, a bright arc in the southeast, and a bright linear feature or “shelf” to the north [@Gaetz00; @Flanagan04]. All observations point to a very massive progenitor star ($>20$ M$_{\odot}$). We report here the detection of E0102 at 24 $\mu$m in the [*Spitzer Space Telescope (SST)*]{} Survey of the SMC[^1] (S$^{3}$MC). We summarize observations and data reduction strategy in §2. The infrared (IR) morphology of E0102 is presented in §3, and compared with observations at other wavelengths in §4. We discuss the origin of the IR emission associated with the remnant in §5, and summarize our results in §6. Observations and Data Processing {#s:obs} ================================ Images presented here were obtained as a part of the S$^{3}$MC program that images the SMC in all Infrared Array Camera (IRAC) and Multiband Imaging Photometer (MIPS) bands (Bolatto et al., in preparation). The [*SST*]{} mosaics were constructed using the MOPEX data reduction package [@Makovoz05]. The angular resolution is 2, 6and 18, at 8, 24 and 70 $\mu$m, respectively. All images were flat-fielded to remove the residual background. For the 24 $\mu$m image, a fourth order polynomial function was modeled for each image row around the SNR and subtracted to properly separate the IR emission associated with the remnant from the surrounding diffuse emission associated with N76. The final images have a noise level of 0.07, 0.08, and 0.5 MJy sr$^{-1}$ at 8, 24, and 70 $\mu$m, respectively. 1E 0102.2-7219 at 24 and 70 $\mu$m ================================== E0102 is prominent in the 24 $\mu$m [*SST*]{} image, while it is not clearly detected in any of the other [*SST*]{} bands at the depth of our observations. The 24 $\mu$m MIPS image of E0102 is shown in Figure \[f:24m\_70m\]a. The remnant displays an almost filled morphology, with some limb brightening along the west side, and two bright elongated knots of emission to the south. The two bright knots, with a mean intensity of $\sim3$ MJy sr$^{-1}$, have major axes parallel to each other. The bigger and slightly brighter knot to the west corresponds to the “spoke” feature seen in X–rays, while the smaller knot to the east is spatially coincident with the bright arc seen in the X-ray ring [@Gaetz00; @Flanagan04]. Conversely, at 70 $\mu$m and with the resolution provided by the [*SST*]{} (18), it is not easy to extricate the remnant from the extended IR emission from the nearby region N 76. Figure \[f:24m\_70m\]b shows the MIPS image of N 76 at 70 $\mu$m, where the white box represents the area around E0102 that is shown in Figure \[f:24m\_70m\]a. E0102 is not obviously detected at 70 $\mu$m, although there is a small bump present in the low-level IR contours exactly at the remnant’s position. The measured flux density of the whole remnant at 24 $\mu$m is $80\pm4$ mJy, while 1-$\sigma$ upper limits at 8 and 70 $\mu$m are 3 and 20 mJy, respectively. The resultant flux density ratios are $S_{70}/S_{24}\la0.25$, and $S_8/S_{24}\la0.04$. Comparison with X-ray and radio continuum data ============================================== Observations of young SNRs at different wavelengths trace different regions/stages of the SNR/ISM interaction. The X–ray emission arises from both forward and reverse shocks. The hotter outer ring identifies the forward shock, while the bright, inner ring originates from the reverse-shocked stellar ejecta, and is seen in various emission lines [@Hayashi94; @Flanagan04]. The 6 cm radio continuum image traces synchrotron emission from ultra-relativistic electrons accelerated by the expanding supernova shock wave, and marks the extent of the forward shock [@Amy93]. In young SNRs, strong radio continuum emission can also be found in the contact region between the forward and reverse shocks [@DeLaney04]. The IR emission associated with young SNRs can originate from several different mechanisms (see Section \[s:IR-origin\] for discussion). The comparison between the 24 $\mu$m and X–ray images of E0102 reveals surprisingly similar structures despite the very different wavelengths and emission mechanisms. The [*SST*]{} 24 $\mu$m image of E0102 overlaid with the [*Chandra*]{} X–ray contours from [@Gaetz00] is shown in Figure \[f:24m\_Xray\]a. The position and curvature of the two IR knots agree remarkably well with those seen in the X–rays. The “spoke” has a very similar size in both the IR and X–ray data while the bright arc is more confined in the IR. The region on the northeast side has the lowest intensity at 24 $\mu$m, again in agreement with X-ray observations. Despite these similarities, some differences are also apparent. For example, while the X–ray image has a high contrast, edge–brightened ring structure, the 24 $\mu$m image displays a lower contrast between the ring and the central region. In addition, the intensity of the IR emission in the limb does not decrease as sharply towards the edge of the SNR as the X-ray emission does. The morphology of the IR emission is less similar to that seen in the radio continuum distribution. Figure  \[f:24m\_Xray\]b shows a comparison of the 24 $\mu$m emission and the 6 cm image by [@Amy93]. The IR emission extends all the way to the edge of the radio continuum distribution. Only in the southwestern part is the radio continuum emission significantly more extended than the IR emission. The IR knots do not have corresponding features in the radio continuum. For a more quantitative comparison we convolved and regridded the X–ray and 6 cm images to match the resolution and pixel size of the 24 $\mu$m image. We then normalized each image to its peak value (each normalized image has a peak value equal to 1). The normalization factors were: 4.7 MJy sr$^{-1}$, $1.1\times10^{-3}$ counts arcsec$^{-2}$ s$^{-1}$, and 177 mJy beam$^{-1}$, for the 24 $\mu$m, X–ray, and 6 cm images, respectively. Three ratio images were then derived from the normalized images: 24 $\mu$m/X–ray (Figure \[f:ratios\]a), 24 $\mu$m/6 cm (Figure \[f:ratios\]b), and 6 cm/X–ray (not shown). Figure \[f:ratios\]a emphasizes the good spatial correlation between the 24 $\mu$m emission and the bright X–ray ring. The almost featureless white ring has a typical pixel value of $\sim0.7-1$. The bright arc is almost invisible in this ratio image, while only the north part of the “spoke” is noticeable, with a slightly higher ratio of $\sim1.2$. The SNR’s edges are mainly dark, with a higher ratio ($>1.5$), emphasizing that the 24 $\mu$m distribution decreases less sharply than the X–ray distribution. The striking feature in this figure is the central region of the SNR, particularly between the two IR knots, which has the highest IR intensity relative to the X–ray intensity, $\sim2$. At the same location in the IR/radio ratio image (Figure \[f:ratios\]b) there is a relatively low ratio ($\sim1$), indicating enhanced synchotron emission (at 6 cm). The ratio image in Figure \[f:ratios\]b shows a thinner ring with a larger radius, reaching the edge of the forward shock. This image has many more significant features, emphasizing the generally poor spatial correlation between the IR and radio continuum distributions. The two knots are prominent in this image as their IR intensity is in significant excess relative to the radio continuum intensity. The region on the northeast side of the SNR with the lowest intensity at 24 $\mu$m stands out in this ratio image as it appears faint in the radio relative to the IR. Discussion {#s:discussion} ========== Origin of IR emission {#s:IR-origin} --------------------- E0102 is the first SNR in the SMC detected in the IR, and it is a prominent object at 24 $\mu$m. Our comparison of the 24 $\mu$m image with radio and X–ray observations suggests that the IR emission is associated with both forward and reverse shocks. The noticeable excess of the IR emission relative to the bright X–ray ring suggests the presence of thermal dust continuum emission associated with the SNR and originating from the circumstellar and/or interstellar dust being heated by the X–ray emitting plasma. For collisionally heated interstellar dust grains, with a grain radius of 0.01 $\mu$m and immersed in a hot gas with $n\sim1-10$ cm$^{-3}$ and $T\sim10^{7}$ K, the expected dust temperature is $T_{d}\ga130$ K (based on Dwek 1987). However, circumstellar or interstellar dust grains embedded in the hot SN cavity are destroyed by sputtering and their lifetime is given by $\tau_{\rm sput}\approx 10^{6} a/n$ yr, where $n$ is the density of the hot post-shock cavity (in cm$^{-3}$), $a$ is the radius of dust grains (in $\mu$m), and the assumed temperature is $\ga10^{6}$ K [@Dwek81]. Assuming a typical size for small dust grains in the SMC $a=0.005-0.05$ $\mu$m [@Stanimirovic00] and knowing that the grains have to survive for the age of E0102 (10$^{3}$ yr), we estimate $n\la5-10$ cm$^{-3}$ for the density of the post-shock cavity. This is in agreement with estimated densities from modeling UV/optical emission lines [@Blair00]. Hot dust in 1E 0102.2-7219 -------------------------- Our measured flux density at 24 $\mu$m and upper limits at 8 and 70 $\mu$m suggest that the IR emission peaks around 24 $\mu$m and sharply falls off toward shorter and longer wavelengths. This is suggestive of the existence of hot dust associated with the SNR, with a temperature of $T_{d}\sim120$ K. In this scenario, our 1-$\sigma$ upper limits at 8 and 70 $\mu$m are slightly higher than the expected IR flux densities for a quasi-blackbody spectrum at $T_{d}\sim120$ K ($\propto \nu^{2} B_{\nu}(T_{d})$). We have taken into account the contribution from the synchrotron emission, extrapolated from radio measurements [@Payne04], which is not significant (only $\sim1$ mJy at 24 $\mu$m). The estimated temperature of hot dust in E0102 is in accord with expectations for the collisionally heated interstellar and/or circumstellar dust. However, the two bright X–ray knots and a significant fraction of the bright X–ray ring structure correlate spatially very well with the IR distribution. This suggests that a fraction of IR emission is most likely associated with the reverse shock as well, and may even be largely due to line emission. The MIPS band at 24 $\mu$m contains emission lines of \[OIV\] at 25.88 $\mu$m and \[FeII\] at 25.98 $\mu$m. Both lines were observed in the young Galactic SNRs RCW103 [@Oliva99] and Cas A [@Arendt99]. The fact that E0102 is detected only at 24 $\mu$m, as well as the fact that this is the only SNR we detect in these [*SST*]{} observations (for discussion see Bolatto et al., in preparation), also argue in favor of a significant contribution from emission lines of \[OIV\] and/or \[FeII\]. As Fe is found to be almost absent, to a level of $\sim10^{-2}-10^{-3}$ relative to O, in the ejecta of E0102 [@Hayashi94; @Blair00] the line emission of \[OIV\] is more likely to affect our 24 $\mu$m measurements. Future [*SST*]{} spectroscopic observations (J. Rho, private communication) will be able to tell us the exact contribution of \[OIV\] to the IR flux at 24 $\mu$m. For a preliminary estimate we use the 24 $\mu$m/X–ray ratio image (Figure \[f:ratios\]a) and assume that all the IR emission in excess to X–ray emission (pixels with values $>1-1.2$, mainly in the central region and at the SNR’s edges) correspond to the dust continuum emission. All other IR emission (IR emission that correlates well with the X–ray emission) is predominantly due to \[OIV\] emission. By integrating the 24 $\mu$m image over these two exclusive subsets, we estimate that $\la60$% of the total IR emission from E0102 could be due to the line emission. If we assume the dust mass absorption coefficient $\kappa_{\lambda}=2.5(\lambda/{\rm 450}\mu m)^{-2.0}$ cm$^2$ g$^{-1}$ [@Draine84] and a distance of 60 kpc [@Westerlund91], we estimate the total mass of hot dust associated with the SNR to be $M_{d}=8\times10^{-4}$ M$_{\odot}$. As shown above, up to 60% of the 24 $\mu$m flux could be due to the line emission. This gives a lower limit on the mass of hot dust associated with the SNR, $M_{d}\ga3\times10^{-4}$ M$_{\odot}$. Even if [*all*]{} the hot dust traced at 24 $\mu$m was formed in the explosion of E0102, the estimated mass of dust is significantly lower than what is expected from some recent theoretical models. For example, [@Todini01] suggested that a large amount of dust could be formed in explosions of core-collapse SNRs even in early galaxies with very low metallicity; for a progenitor star with a mass of 12–35 M$_{\odot}$ this model predicts $0.08<M_{d}<0.3$ M$_{\odot}$. This range is further predicted to increase by a factor of 2–3 as the metallicity approaches the solar value. With the metallicity of $\sim1/10$ the solar value, the amount of dust we estimate for E0102 is at least a factor of 100 lower than the lowest predicted value. It is interesting, though, that the amount of hot dust in E0102 is only a few times lower than that found in the recent measurement of hot dust in Cas A [@Hines04], although the two SNRs occurred in very different interstellar environments. Conclusions =========== We have obtained IR observations of the young, oxygen-rich SNR E0102 as a part of a large program to image the SMC with the [*SST*]{} in all available bands. The remnant is detected only at 24 $\mu$m, and its IR distribution has a filled morphology with two prominent elongated filaments. Our comparison of IR observations with the existing X–ray and radio continuum images suggests that the IR emission is most likely associated with both reverse and forward shocks. We find evidence for the existence of up to $8\times10^{-4}$ M$_{\odot}$ of hot dust at a temperature of $\sim120$ K, associated with the remnant. The \[OIV\] emission line may be contributing significantly (up to 60%) to our flux measurement at 24 $\mu$m. The hot dust is mainly located in the central region of E0102 which appears significantly enhanced in IR and radio continuum emission relative to the X–ray emission. The amount of dust associated with E0102, under the assumption that [*all*]{} the dust was formed during the SN explosion, is at least 100 times lower than what is predicted by some recent theoretical models for dust production in the ejecta of core-collapse supernovae. We thank Carl Heiles for stimulating discussions and an anonymous referee for valuable suggestions. This work is based \[in part\] on observations made with the [*Spitzer Space Telescope*]{}, which is operated by the Jet Propulsion Laboratory, California Institute of Technology under NASA contract 1407. Support for this work was provided by NASA through Contract Number 1264151 issued by JPL/Caltech. We also acknowledge support by NSF grants AST-0097417 and AST-0228963. [19]{} natexlab\#1[\#1]{} , S. W. & [Ball]{}, L. 1993, , 411, 761 , R. G., [Dwek]{}, E., & [Moseley]{}, S. H. 1999, , 521, 234 , W. P., [Morse]{}, J. A., [Raymond]{}, J. C., [Kirshner]{}, R. P., [Hughes]{}, J. P., [Dopita]{}, M. A., [Sutherland]{}, R. S., [Long]{}, K. S., & [Winkler]{}, P. F. 2000, , 537, 667 , T., [Rudnick]{}, L., [Fesen]{}, R. A., [Jones]{}, T. W., [Petre]{}, R., & [Morse]{}, J. A. 2004, , 613, 343 , B. T. & [Lee]{}, H. M. 1984, , 285, 89 , E. & [Werner]{}, M. W. 1981, , 248, 138 , K. A., [Canizares]{}, C. R., [Dewey]{}, D., [Houck]{}, J. C., [Fredericks]{}, A. C., [Schattenburg]{}, M. L., [Markert]{}, T. H., & [Davis]{}, D. S. 2004, , 605, 230 , T. J., [Butt]{}, Y. M., [Edgar]{}, R. J., [Eriksen]{}, K. A., [Plucinsky]{}, P. P., [Schlegel]{}, E. M., & [Smith]{}, R. K. 2000, , 534, L47 , I., [Koyama]{}, K., [Ozaki]{}, M., [Miyata]{}, E., [Tsumeni]{}, H., [Hughes]{}, J. P., & [Petre]{}, R. 1994, , 46, L121 , K. G. 1956, , 2, 315 , D. C. & et al. 2004, , 154, 290 Makovoz, D. & Marleau, F. R. 2005, , in press, astro-ph/0507007 , C. F. 2001, in AIP Conf. Proc. 565: Young Supernova Remnants, 17–28 , E., [Moorwood]{}, A. F. M., [Drapatz]{}, S., [Lutz]{}, D., & [Sturm]{}, E. 1999, , 343, 943 , J. L., [Filipovi[' c]{}]{}, M. D., [Reid]{}, W., [Jones]{}, P. A., [Staveley-Smith]{}, L., & [White]{}, G. L. 2004, , 355, 44 Stanimirović, S., Staveley-Smith, L., van der Hulst, J. M., Bontekoe, T. R., Kester, D. J. M., & Jones, P. A. 2000, MNRAS, 315, 791 , P. & [Ferrara]{}, A. 2001, , 325, 726 , J. 2004, Nuclear Physics B Proceedings Supplements, 132, 21 , B. E. 1991, in IAU Symp. 148: The Magellanic Clouds, 15 \[lastpage\] [^1]: http://celestial.berkeley.edu/spitzer

--- abstract: 'We provide generalisations of two Schwarz-type lemmas — the first a result of Globevnik and the other due to Ransford and White — for holomorphic mappings into the spectral unit ball. The first concerns mappings of the unit disc in ${\mathbb{C}}$ into the spectral unit ball, while the second concerns self-mappings. The aforementioned results apply to holomorphic mappings that map the origin to the origin. We extend these results to general holomorphic mappings into the spectral unit ball. We also show that our results are sharp.' address: 'Department of Mathematics, Indian Institute of Science, Bangalore – 560 012' author: - Gautam Bharali title: | Sharp Schwarz-type lemmas for the\ spectral unit ball --- Introduction and Statement of Results {#S:intro} ===================================== The [**spectral unit ball**]{} (denoted by ${\Omega}_n$) is defined as $${\Omega}_n \ := \ \{W\in M_n({\mathbb{C}}):r(W)<1\},$$ where $r(W)$ denotes the spectral radius of the $n\times n$ matrix $W$. In this paper, $D$ will denote the unit disc in ${\mathbb{C}}$ and, if ${\Omega}$ and $G$ are complex domains, ${\mathcal{O}}({\Omega};G)$ will denote the class of holomorphic mappings of ${\Omega}$ into $G$. We present two Schwarz-type lemmas for the spectral unit ball. These lemmas are inspired by the renewed interest in the function theory on the spectral unit ball — the reader is referred to [@baribeauRansford:nlspm00], [@rostand:asub03], [@aglerYoung:2b2sNPp04] and [@edgrnZwnk:gsp05], to name just a few recent papers. This interest stems, to a large extent, from recent work on the spectral version of the Nevanlinna-Pick interpolation problem. We shall not address this problem directly in this paper; although Theorem \[T:disc\] below might have some bearing on the [*two-point*]{} interpolation problem. We begin by considering the following result by Globevnik [@globevnik:Slsr74], which is perhaps the earliest Schwarz-type lemma for the spectral unit ball: $$\label{E:globevnik} F\in{\mathcal{O}}(D;{\Omega}_n) \ \text{and} \ F({\zeta}_1)=0 \ {\Longrightarrow}\ r(F({\zeta}_2))\leq{\left|\frac{{{\zeta}_1}-{{\zeta}_2}}{1-\overline{{{\zeta}_2}}{{\zeta}_1}}\right|}=:{\mathcal{M}}({\zeta}_1,{\zeta}_2).$$ One would like to generalise this result to the case when $F({\zeta}_j)$ is not necessarily $0$, $j=1,2$. The second Schwarz-type lemma that motivates this paper — this time a result on [*self-mappings*]{} of ${\Omega}_n$ — is the following result of Ransford and White [@ransfordWhite:hsmsub91]: $$\label{E:rnsfrdWht} G\in{\mathcal{O}}({\Omega}_n;{\Omega}_n) \ \text{and} \ G(0)=0 \ {\Longrightarrow}\ r(G(X))\leq r(X) \;\; \forall X\in{\Omega}_n.$$ Incidentally, we refer to the above results as “Schwarz-type lemmas” because they relate the growth of a holomorphic mapping to the growth of its argument(s). One would like to generalise in the similar manner that the Schwarz-Pick lemma generalises the Schwarz lemma for $D$ — i.e. by formulating an inequality that is valid without assuming that the holomorphic mapping in question has a fixed point. What is the key idea needed to generalise and in appropriate ways ? The following example shows how the conclusion of can fail, even in the simple situation where $r(F({\zeta}_1))=0$, if $F({\zeta}_1)\neq 0$. However, it also suggests a way forward. \[Ex:keyex\] For $n\geq 3$ and $d=2,\dots,n-1$, define the holomorphic map ${F_d}:D{\longrightarrow}{\Omega}_n$ by $${F_d}({\zeta}) \ := \ \begin{bmatrix} \ 0 & {} & {} & {\zeta}\ &\vline & \ {} \ \\ \ 1 & 0 & {} & 0 \ &\vline & {} \ \\ \ {} & \ddots & \ddots & \vdots \ &\vline & \text{\LARGE{0}} \ \\ \ {} & {} & 1 & 0 \ &\vline & {} \ \\ \hline \ {} & {} & {} & {} \ &\vline & {} \ \\ \ {} & {} & \text{\LARGE{0}} & {} \ &\vline & {\zeta}\mathbb{I}_{n-d} \ \end{bmatrix}_{n\times n}, \qquad {\zeta}\in D,$$ where $\mathbb{I}_{n-d}$ denotes the identity matrix of dimension $n-d$ for $1<d<n$. One easily computes that $r({F_d}({\zeta}))=|{\zeta}|^{1/d}$. Hence $$\begin{aligned} r({F_d}({\zeta}))^d \ &= \ |{\zeta}|\ =\ {\mathcal{M}}(0,{\zeta})\quad \forall{\zeta}\in D, \label{E:nilpot} \\ \text{but, for each $q<d$,} \ r({F_d}({\zeta}))^q \ &> \ {\mathcal{M}}(0,{\zeta})\quad \forall{\zeta}\neq 0. \notag\end{aligned}$$ In particular, $r({F_d}({\zeta}))>{\mathcal{M}}(0,{\zeta})\quad \forall{\zeta}\neq 0$, in contrast with . The above example is rather suggestive when one notices that the exponent occurring in the left-hand side of is the degree of the minimal polynomial of ${F_d}(0)$. While $r(F_d({\zeta}_1))=0$ for each $d=2,\dots,n-1$ (take ${\zeta}_1=0$ in this discussion), what differs in each case is the degree of the minimal polynomial of $F_d({\zeta}_1)$. Presumably, this information should be encoded in any generalisation of . This idea is the key to establishing a result with the following features: - It has a Schwarz-type structure: i.e. we get an expression in $F({\zeta}_1)$ and $F({\zeta}_2)$ — call it $\mathcal{E}(F({\zeta}_1),F({\zeta}_2))$ — such that $$F\in{\mathcal{O}}(D;{\Omega}_n) {\Longrightarrow}\mathcal{E}(F({\zeta}_1),F({\zeta}_2)) \ \leq \ {\mathcal{M}}({\zeta}_1,{\zeta}_2).$$ - Globevnik’s result is recovered when we set $F({\zeta}_1)=0$ in the above inequality. - The above Schwarz-type inequality is [*sharp*]{} in the sense that this inequality is the best one can achieve. In more precise terms: given $z\neq w\in D$, we can find a ${\mathscr{F}^{z,w}}\in{\mathcal{O}}(D;{\Omega}_n)$ such that $\mathcal{E}({\mathscr{F}^{z,w}}(z),{\mathscr{F}^{z,w}}(w))={\mathcal{M}}(z,w)$. Before stating this result, let us, for any compact subset $K\varsubsetneq D$ and ${\zeta}\in D$, define $${{\sf dist}_{\mathcal{M}}}({\zeta};K) \ := \ \min_{z\in K}{\left|\frac{{{\zeta}}-{z}}{1-\overline{{z}}{{\zeta}}}\right|}.$$ Our first result is as follows. \[T:disc\] Let $F\in{\mathcal{O}}(D;{\Omega}_n)$, $n\geq 2$, and let ${\zeta}_1,{\zeta}_2\in D$. Let $W_j=F({\zeta}_j), \ j=1,2$, and define $$d_j \ := \ \text{the degree of the minimal polynomial of $W_j$},$$ for $j=1,2$. Then $$\label{E:SchwarzIneq} \max\left\{\max_{\mu\in\sigma(W_2)}[{{\sf dist}_{\mathcal{M}}}(\mu;\sigma(W_1))^{d_1}], \ \max_{\lambda\in\sigma(W_1)} [{{\sf dist}_{\mathcal{M}}}(\lambda;\sigma(W_2))^{d_2}]\right\} \ \leq \ {\left|\frac{{{\zeta}_1}-{{\zeta}_2}}{1-\overline{{{\zeta}_2}}{{\zeta}_1}}\right|}.$$ Furthermore, is sharp in the sense that given any two points $z,w\in D$, there exists a mapping ${\mathscr{F}^{z,w}}\in{\mathcal{O}}(D;{\Omega}_n)$ such that $$\begin{gathered} \max\left\{\max_{\mu\in\sigma({\mathscr{F}^{z,w}}(w))}[{{\sf dist}_{\mathcal{M}}}(\mu;\sigma({\mathscr{F}^{z,w}}(z)))^{d(z)}], \ \max_{\lambda\in\sigma({\mathscr{F}^{z,w}}(z))}[{{\sf dist}_{\mathcal{M}}}(\lambda;\sigma({\mathscr{F}^{z,w}}(w)))^{d(w)}]\right\} \\ = \ {\left|\frac{{z}-{w}}{1-\overline{{w}}{z}}\right|},\notag\end{gathered}$$ where $d(z)$ (resp. $d(w)$) is the degree of the minimal polynomial of ${\mathscr{F}^{z,w}}(z)$ (resp. ${\mathscr{F}^{z,w}}(w)$). Note that Globevnik’s result is recovered when we set $F({\zeta}_1)=0$ in the above theorem. This is because if $W_1=0$, then, in the notation of Theorem \[T:disc\], $d_1=1$, and $$\begin{aligned} \max_{\lambda\in\sigma(W_1)}[{{\sf dist}_{\mathcal{M}}}(\lambda;\sigma(W_2))^{d_2}] \ &= \ \min_{\mu\in\sigma(W_2)}|\mu|^{d_2}\notag \\ &\leq \ \max_{\mu\in\sigma(W_2)}[{{\sf dist}_{\mathcal{M}}}(\mu;\sigma(W_1))^{d_1}] \ = \ r(W_2).\notag\end{aligned}$$ Hence , in this case, is identical to the conclusion of Globevnik’s result. The proof of Theorem \[T:disc\] is presented in Section \[S:proofDisc\]. The consideration of a pertinent minimal polynomial turns out to equally relevant to our next result. Essentially, the proofs of both results exploit the minimal polynomial of a key matrix lying in the ranges of $F$, respectively $G$, to transform the maps $F$ and $G$ to maps to which the results and , respectively, are applicable. One is led to do this when examining the basic example of a mapping $G\in{\mathcal{O}}({\Omega}_n;{\Omega}_n)$ where $G(0)\neq 0$ and $G$ fails to satisfy the inequality in , [*even though*]{} $r(G(0))=0$. Since we do not wish to prolong this already protracted introduction, we refer the reader to the counterexample immediately following Theorem 2 in the Ransford-White paper [@ransfordWhite:hsmsub91] (or to the end of Section \[S:proofSpec\] of this paper). The idea hinted at above leads to a new result that: - Just like , provides a bound on the growth of the spectral radius of $G(X)$ in terms of $r(X)$, and specialises precisely to when we set $G(0)$; - Is sharp in a manner analogous to our discussion of “sharpness” of our previous result. More precisely, we have the following: \[T:spec\] Let $G\in{\mathcal{O}}({\Omega}_n;{\Omega}_n), \ n\geq 2$, and define $d_G:=$ the degree of the minimal polynomial of $G(0)$. Then: $$\label{E:growthBound} r(G(X)) \ \leq \ \frac{r(X)^{1/d_G}+r(G(0))}{1+r(G(0))r(X)^{1/d_G}} \quad\forall X\in{\Omega}_n.$$ Furthermore, the inequality is sharp in the sense that there exists a non-empty set ${\mathfrak{S}}_n\subset{\Omega}_n$ such that given any $A\in{\mathfrak{S}}_n$ and $d=1,\dots,n$, we can find a ${\mathfrak{G}^{A,d}}\in{\mathcal{O}}({\Omega}_n;{\Omega}_n)$ such that $$\begin{aligned} d_{{\mathfrak{G}^{A,d}}} \ &= \ d, \;\; \text{and}\notag\\ r({\mathfrak{G}^{A,d}}(A)) \ &= \ \frac{r(A)^{1/d}+r({\mathfrak{G}^{A,d}}(0))}{1+r({\mathfrak{G}^{A,d}}(0))r(A)^{1/d}}. \label{E:sharp}\end{aligned}$$ It is quite obvious why the Ransford-White bound is recovered when we set $G(0)=0$ in the above theorem. When $G(0)=0$, then $d_G=1$ and $r(G(0))=0$, whence is identical to the conclusion of . The Proof of Theorem \[T:disc\] {#S:proofDisc} =============================== The proofs in this section depend crucially on a theorem by Vesentini. The result is as follows: \[R:vesentini\] Let ${\mathcal{A}}$ be a complex, unital Banach algebra and let $r(x)$ denote the spectral radius of any element $x\in{\mathcal{A}}$. Let $f\in{\mathcal{O}}(D;{\mathcal{A}})$. The the function ${\zeta}\longmapsto r(f({\zeta}))$ is subharmonic on $D$. The following result is the key lemma of this section. The proof of Theorem \[T:disc\] is reduced to a simple application of this lemma. \[L:key\] Let $F\in{\mathcal{O}}(D;{\Omega}_n)$ and let ${\lambda}_1,\dots,{\lambda}_s$ be the distinct eigenvalues of $F(0)$. Define $m(j):=$the multiplicity of the factor $({\lambda}-{\lambda}_j)$ in the minimal polynomial of $F(0)$. Define the Blaschke product $$B({\zeta}) \ := \ \prod_{j=1}^s{\left(\frac{{{\zeta}}-{{\lambda}_j}}{1-\overline{{{\lambda}_j}}{{\zeta}}}\right)}^{m(j)}, \quad{\zeta}\in D.$$ Then $|B({\lambda})|\leq|{\zeta}| \ \forall{\lambda}\in\sigma(F({\zeta}))$. The Blaschke product $B$ induces a matrix function $\widetilde{B}$ on ${\Omega}_n$: for any matrix $A\in{\Omega}_n$, we set $$\widetilde{B}(A) \ := \ \prod_{j=1}^s({\mathbb{I}}-\overline{{\lambda}_j} A)^{-m(j)}(A-{\lambda}_j{\mathbb{I}})^{m(j)},$$ which is well-defined on ${\Omega}_n$ because whenever ${\lambda}_j\neq 0$, $$({\mathbb{I}}-\overline{{\lambda}}_j A) \ = \ \overline{{\lambda}_j}({\mathbb{I}}/\overline{{\lambda}}_j-A)\in GL(n,{\mathbb{C}}).$$ Furthermore, since ${\zeta}\longmapsto({\zeta}-{\lambda}_j)/(1-\overline{{\lambda}_j}{\zeta})$ has a power-series expansion that converges uniformly on compact subsets of $D$, it follows from standard arguments that $$\label{E:specB} \sigma(\widetilde{B}(A)) \ = \ \{B({\lambda}):{\lambda}\in\sigma(A)\}\quad\text{for any $A\in{\Omega}_n$.}$$ By the definition of the minimal polynomial, $\widetilde{B}\circ F(0)=0$. At this point, we could apply Globevnik’s lemma — i.e. above — to complete the proof. The actual argument, however, is very elementary, and we provide it here. Since $\widetilde{B}\circ F(0)=0$, there exists a holomorphic map $\Phi\in{\mathcal{O}}(D;M_n({\mathbb{C}}))$ such that $\widetilde{B}\circ F({\zeta})={\zeta}\Phi({\zeta})$. Note that $$\label{E:specRelation} \sigma(\widetilde{B}\circ F({\zeta})) \ = \ \sigma({\zeta}\Phi({\zeta})) \ = \ {\zeta}\sigma(\Phi({\zeta}))\quad\forall{\zeta}\in D.$$ Since $\sigma(\widetilde{B}\circ F({\zeta}))\subset D$, the above equations give us: $$\label{E:circBound} r(\Phi({\zeta})) \ < \ 1/R \quad\forall{\zeta}:|{\zeta}|=R, \ R\in(0,1).$$ Taking ${\mathcal{A}}=M_n({\mathbb{C}})$ in Vesentini’s theorem, we see that ${\zeta}\longmapsto r(\Phi({\zeta}))$ is subharmonic on the unit disc. Applying the Maximum Principle to and taking limits as $R\longrightarrow 1^-$, we get $$\label{E:oneBound} r(\Phi({\zeta})) \ \leq \ 1\quad\forall{\zeta}\in D.$$ In view of , and , we get $$|B({\lambda})| \ \leq \ |{\zeta}|r(\Phi({\zeta})) \ \leq \ |{\zeta}|\quad\forall\lambda\in\sigma(F({\zeta})).$$ We are now in a position to provide Define the disc automorphisms $$M_j({\zeta}) \ := \ \frac{{\zeta}-{\zeta}_j}{1-\overline{{\zeta}_j}{\zeta}}, \quad j=1,2,$$ and write $\Phi_j=F\circ M_j^{-1}, \ j=1,2$. Note that $\Phi_1(0)=W_1$. Let ${\lambda}_1,\dots,{\lambda}_r$ be the distinct eigenvalues of $W_1$ and define $m_1(j):=$the multiplicity of the factor $({\lambda}-{\lambda}_j)$ in the minimal polynomial of $W_1$. Define $$B_1({\zeta}) \ := \ \prod_{j=1}^r{\left(\frac{{{\zeta}}-{{\lambda}_j}}{1-\overline{{{\lambda}_j}}{{\zeta}}}\right)}^{m_1(j)}, \quad{\zeta}\in D.$$ Applying Lemma \[L:key\], we get $$\begin{aligned} {\left|\frac{{{\zeta}_1}-{{\zeta}_2}}{1-\overline{{{\zeta}_2}}{{\zeta}_1}}\right|} \ = \ |M_1({\zeta}_2)| \ &\geq \ \prod_{j=1}^r{\left|\frac{{\mu}-{{\lambda}_j}}{1-\overline{{{\lambda}_j}}{\mu}}\right|}^{m_1(j)} \label{E:1stpart} \\ &\geq \ {{\sf dist}_{\mathcal{M}}}(\mu;\sigma(W_1))^{d_1} \quad\forall\mu\in\sigma(\Phi_1(M_1({\zeta}_2))) =\sigma(W_2).\notag\end{aligned}$$ Now, swapping the roles of ${\zeta}_1$ and ${\zeta}_2$ and applying the same argument to $$B_2({\zeta}) \ := \ \prod_{j=1}^s{\left(\frac{{{\zeta}}-{\mu_j}}{1-\overline{{\mu_j}}{{\zeta}}}\right)}^{m_2(j)}, \quad{\zeta}\in D,$$ where $\mu_1,\dots,\mu_s$ are the distinct eigenvalues of $W_2$ and $m_2(j):=$the multiplicity of the factor $({\lambda}-\mu_j)$ in the minimal polynomial of $W_2$, we get $$\label{E:2ndpart} {\left|\frac{{{\zeta}_1}-{{\zeta}_2}}{1-\overline{{{\zeta}_2}}{{\zeta}_1}}\right|} \ \geq \ {{\sf dist}_{\mathcal{M}}}({\lambda};\sigma(W_2))^{d_2} \quad\forall{\lambda}\in\sigma(W_1).$$ Combining and , we get $$\max\left\{\max_{\mu\in\sigma(W_2)}[{{\sf dist}_{\mathcal{M}}}(\mu;\sigma(W_1))^{d_1}], \ \max_{\lambda\in\sigma(W_1)} [{{\sf dist}_{\mathcal{M}}}(\lambda;\sigma(W_2))^{d_2}]\right\} \ \leq \ {\left|\frac{{{\zeta}_1}-{{\zeta}_2}}{1-\overline{{{\zeta}_2}}{{\zeta}_1}}\right|}.$$ In order to prove the sharpness of , fix an $n\geq 2$, and choose any $z,w\in D$. Next, define $M({\zeta}):=({\zeta}-z)(1-{\overline{z}}{\zeta})^{-1}$. Pick any $d=1,\dots,n$, and define $$N_d({\zeta}) \ := \ \begin{cases} \quad [M({\zeta})], & \text{if $d=1$}, \\ \ \begin{bmatrix} \ 0 & {} & {} & M({\zeta}) \ \\ \ 1 & 0 & {} & 0 \ \\ \ {} & \ddots & \ddots & \vdots \ \\ \ {} & {} & 1 & 0 \ \end{bmatrix}_{d\times d}, & \text{if $d\geq 2$}, \end{cases}$$ and, for the chosen $d$, define ${\mathscr{F}^{z,w}}$ by the following block-diagonal matrix $${\mathscr{F}^{z,w}}({\zeta}) \ := \begin{bmatrix} \ N_d({\zeta}) & {} \ \\ \ {} & M({\zeta})\mathbb{I}_{n-d} \ \end{bmatrix} \quad\forall{\zeta}\in D.$$ Note that - ${\mathscr{F}^{z,w}}(z)$ is nilpotent of degree $d$, whence $d(z)=d$; and - Since $|M(w)|^{1/d}>|M(w)|$, $$\max_{\mu\in\sigma({\mathscr{F}^{z,w}}(w))}[{{\sf dist}_{\mathcal{M}}}(\mu;\sigma({\mathscr{F}^{z,w}}(z)))^{d(z)}] \ = \ |M(w)| \ = \ {\left|\frac{{z}-{w}}{1-\overline{{w}}{z}}\right|}.$$ A similar argument yields $$\max_{\lambda\in\sigma({\mathscr{F}^{z,w}}(z))}[{{\sf dist}_{\mathcal{M}}}(\lambda;\sigma({\mathscr{F}^{z,w}}(w)))^{d(w)}] \ = \ |M(w)|^{d+1} \ = \ {\left|\frac{{z}-{w}}{1-\overline{{w}}{z}}\right|}^{d+1}.$$ Hence, we have the equality $$\begin{gathered} \max\left\{\max_{\mu\in\sigma({\mathscr{F}^{z,w}}(w))}[{{\sf dist}_{\mathcal{M}}}(\mu;\sigma({\mathscr{F}^{z,w}}(z)))^{d(z)}], \ \max_{\lambda\in\sigma({\mathscr{F}^{z,w}}(z))}[{{\sf dist}_{\mathcal{M}}}(\lambda;\sigma({\mathscr{F}^{z,w}}(w)))^{d(w)}]\right\} \\ = \ {\left|\frac{{z}-{w}}{1-\overline{{w}}{z}}\right|}.\notag\end{gathered}$$ ${\mathscr{F}^{z,w}}$ is therefore the desired map that establishes the sharpness of . The Proof of Theorem \[T:spec\] {#S:proofSpec} =============================== In order to prove Theorem \[T:spec\], we shall need the following elementary \[L:mobiusTrans\] Given a M[ö]{}bius transformation $T(z):=(az+b)/(cz+d)$, if $T({\partial D})\Subset{\mathbb{C}}$, then $T({\partial D})$ is a circle with $${\rm centre}(T({\partial D})) \ = \ \frac{b\overline{d}-a\overline{c}}{|d|^2-|c|^2}, \qquad {\rm radius}(T({\partial D})) \ = \ \frac{|ad-bc|}{||d|^2-|c|^2|}.$$ We are now in a position to present Let $G\in{\mathcal{O}}({\Omega}_n;{\Omega}_n)$ and let ${\lambda}_1,\dots,{\lambda}_s$ be the distinct eigenvalues of $G(0)$. Define $m(j):=$the multiplicity of the factor $({\lambda}-{\lambda}_j)$ in the minimal polynomial of $G(0)$. Define the Blaschke product $$B_G({\zeta}) \ := \ \prod_{j=1}^s{\left(\frac{{{\zeta}}-{{\lambda}_j}}{1-\overline{{{\lambda}_j}}{{\zeta}}}\right)}^{m(j)}, \quad{\zeta}\in D.$$ $B_G$ induces the following matrix function which, by a mild abuse of notation, we shall also denote as $B_G$ $$B_G(Y) \ := \ \prod_{j=1}^s({\mathbb{I}}-\overline{{\lambda}_j} Y)^{-m(j)}(Y-{\lambda}_j{\mathbb{I}})^{m(j)} \quad\forall Y\in{\Omega}_n,$$ which is well-defined on ${\Omega}_n$ precisely as explained in the proof of Lemma \[L:key\]. Once again, owing to the analyticity of $B_G$ on $D$, $$\sigma(B_G(Y)) \ = \ \{B_G({\lambda}):{\lambda}\in\sigma(Y)\}\quad\forall Y\in{\Omega}_n,$$ whence $B_G:{\Omega}_n{\longrightarrow}{\Omega}_n$. Therefore, if we define $$H(X) \ := \ B_G\circ G(X) \quad\forall X\in{\Omega}_n,$$ then $H\in{\mathcal{O}}({\Omega}_n;{\Omega}_n)$ and, by construction, $H(0)=0$. By the Ransford-White result, $r(H(X))\leq r(X)$, or, more precisely $$\max_{\mu\in G(X)}\left\{\prod_{j=1}^s{\left|\frac{{\mu}-{{\lambda}_j}}{1-\overline{{{\lambda}_j}}{\mu}}\right|}^{m(j)}\right\} \ \leq \ r(X) \quad\forall X\in{\Omega}_n.$$ In particular: $$\max_{\mu\in G(X)}\left[{{\sf dist}_{\mathcal{M}}}(\mu;\sigma(G(0)))^{d_G}\right] \ \leq \ r(X) \quad\forall X\in{\Omega}_n.$$ For the moment, let us fix $X\in{\Omega}_n$. For each $\mu\in\sigma(G(X))$, let ${\lambda}_\mu$ be an eigenvalue of $G(0)$ such that ${|({\mu}-{{\lambda}_\mu})(1-\overline{{{\lambda}_\mu}}{\mu})^{-1}|}={{\sf dist}_{\mathcal{M}}}(\mu;\sigma(G(0)))$. Now fix $\mu\in\sigma(G(X))$. The above inequality leads to $$\label{E:hyperIneq1} {\left|\frac{{\mu}-{{\lambda}_\mu}}{1-\overline{{{\lambda}_\mu}}{\mu}}\right|} \ \leq \ r(X)^{1/d_G}.$$ Applying Lemma \[L:mobiusTrans\] to the M[ö]{}bius transformation $$T(z) \ = \ \frac{|\mu|z-{\lambda}_\mu}{1-\overline{{\lambda}_\mu}|\mu|z},$$ we deduce that $${\left|\frac{{{\zeta}}-{{\lambda}_\mu}}{1-\overline{{{\lambda}_\mu}}{{\zeta}}}\right|} \ \geq \ \frac{||\mu|-|{\lambda}_\mu||}{1-|\mu||{\lambda}_\mu|}\quad\forall{\zeta}:|{\zeta}|=|\mu|.$$ Applying the above fact to , we get $$\begin{aligned} \frac{|\mu|-|{\lambda}_\mu|}{1-|\mu||{\lambda}_\mu|} \ &\leq \ r(X)^{1/d_G} \notag \\ \Rightarrow\quad |\mu| \ &\leq \frac{r(X)^{1/d_G}+|{\lambda}_\mu|}{1+|{\lambda}_\mu|r(X)^{1/d_G}}, \quad\mu\in\sigma(G(X)). \label{E:specIneq}\end{aligned}$$ Note that the function $$t\longmapsto\frac{r(X)^{1/d_G}+t}{1+r(X)^{1/d_G}t}, \quad t\geq 0,$$ is an increasing function on $[0,\infty)$. Combining this fact with , we get $$|\mu| \ \leq \ \frac{r(X)^{1/d_G}+r(G(0))}{1+r(G(0))r(X)^{1/d_G}},$$ which holds $\forall\mu\in\sigma(G(X))$, while the right-hand side is independent of $\mu$. Since this is true for any arbitrary $X\in{\Omega}_n$, we conclude that $$r(G(X)) \ \leq \ \frac{r(X)^{1/d_G}+r(G(0))}{1+r(G(0))r(X)^{1/d_G}} \quad\forall X\in{\Omega}_n.$$ In order to prove the sharpness of , let us fix an $n\geq 2$, and define $${\mathfrak{S}}_n \ := \ \{A\in{\Omega}_n:A \ \text{has a single eigenvalue of multiplicity $n$} \}.$$ Pick any $d=1,\dots,n$, and define $$M_d(X) \ := \ \begin{cases} \quad [{{\sf tr}}(X)/n], & \text{if $d=1$}, \\ \ \begin{bmatrix} \ 0 & {} & {} & {{\sf tr}}(X)/n \ \\ \ 1 & 0 & {} & 0 \ \\ \ {} & \ddots & \ddots & \vdots \ \\ \ {} & {} & 1 & 0 \ \end{bmatrix}_{d\times d}, & \text{if $d\geq 2$}, \end{cases}$$ and, for the chosen $d$, define ${\mathfrak{G}^{(d)}}$ by the following block-diagonal matrix $${\mathfrak{G}^{(d)}}(Y) \ := \begin{bmatrix} \ M_d(X) & {} \ \\ \ {} & \dfrac{{{\sf tr}}(X)}{n}\mathbb{I}_{n-d} \ \end{bmatrix} \quad\forall X\in{\Omega}_n.$$ For our purposes ${\mathfrak{G}^{A,d}}={\mathfrak{G}^{(d)}}$ for each $A\in{\mathfrak{S}}_n$; i.e., the equality will will hold with the same function for each $A\in{\mathfrak{S}}_n$. To see this, note that - $r({\mathfrak{G}^{(d)}}(X))=|{{\sf tr}}(X)/n|^{1/d}$; and - ${\mathfrak{G}^{(d)}}(0)$ is nilpotent of degree $d$, whence $d_{{\mathfrak{G}^{(d)}}}=d$. Therefore, $$\frac{r(A)^{1/d}+r({\mathfrak{G}^{(d)}}(0))}{1+r({\mathfrak{G}^{(d)}}(0))r(A)^{1/d}} \ = \ r(A)^{1/d} \ = \ r({\mathfrak{G}^{(d)}}(A)) \quad\forall A\in{\mathfrak{S}}_n,$$ which establishes [99]{} J. Agler and N.J. Young, [*The two-by-two spectral Nevanlinna-Pick problem*]{}, Trans. Amer. Math. Soc. [**356**]{} (2004), 573-585. L. Baribeau and T. Ransford, [*Non-linear spectrum-preserving maps*]{}, Bull. London Math. Soc. [**32**]{} (2000), 8-14. A. Edigarian and W. Zwonek, [*Geometry of the symmetrized polydisc*]{}, Arch. Math. (Basel) [**84**]{} (2005), 364-374. J. Globevnik, [*Schwarz’s lemma for the spectral radius*]{}, Rev. Roumaine Math. Pures Appl. [**19**]{} (1974), 1009-1012. T.J. Ransford and M.C. White, [*Holomorphic self-maps of the spectral unit ball*]{}, Bull. London Math. Soc. [**23**]{} (1991), 256-262. J. Rostand, [*On the automorphisms of the spectral unit ball*]{}, Studia Math. [**155**]{} (2003), 207-230. E. Vesentini, [*On the subharmonicity of the spectral radius*]{}, Boll. Un. Mat. Ital. (4) [**1**]{} 1968, 427-429.

--- author: - Lihua Liu$^1$ Zhengjun Cao$^2$ title: | A Note on “Efficient Algorithms for Secure\ Outsourcing of Bilinear Pairings" --- > **Abstract**. We show that the verifying equations in the scheme \[Theoretical Computer Science, 562 (2015), 112-121\] cannot filter out some malformed values returned by the malicious servers. We also remark that the two untrusted programs model adopted in the scheme is somewhat artificial, and discuss some reasonable scenarios for outsourcing computations. > > **Keywords.** bilinear pairing, outsourcing computation, semi-honest server. Introduction ============ Very recently, Chen et al. [@C15] have put forth a scheme for outsourcing computations of bilinear pairings in two untrusted programs model which was introduced by Hohenberger and Lysyanskaya [@HL05]. In the scheme, a user $T$ can indirectly compute the pairing $e(A, B)$ by outsourcing some expensive work to two untrusted servers $U_1$ and $U_2$ such that $A$, $B$ and $e(A, B)$ are kept secret. Using the returned values from $U_1$, $U_2$ and some previously stored values, the user $T$ can recover $e(A, B)$. The Chen et al.’s scheme is derived from the Chevallier-Mames et al.’s scheme [@CC10] by storing some values in a table in order to save some expensive operations such as point multiplications and exponentiations. Besides, the new scheme introduces two servers $U_1$ and $U_2$ rather than the unique server $U$ in the Chevallier-Mames et al.’s scheme. The authors [@C15] claim that the scheme achieves the security *as long as one of the two servers is honest*. In other word, a malicious server cannot obtain either $A$ or $B$. Unfortunately, the assumption cannot ensure that the scheme works well, because a malicious server can return some random values while the user $T$ cannot detect the malicious behavior. As a result, $T$ outputs a false value. In this note, we show that the verifying equations in the scheme [@C15] cannot filter out some malformed values returned by the malicious servers. To fix this drawback, we should specify that the servers are semi-honest. We also point out that the two untrusted programs model adopted in the scheme is somewhat artificial and discuss some reasonable scenarios for outsourcing computations. Review of the scheme ==================== Let $\mathbb{G}_1$ and $\mathbb{G}_2$ be two cyclic additive groups with a large prime order $q$. Let $G_3$ be a cyclic multiplicative group of the same order $q$. A bilinear pairing is a map $e : \mathbb{G}_1\times \mathbb{G}_2 \rightarrow \mathbb{G}_3$ with the following properties. (1) Bilinear: $e(aR, bQ) =e(R, Q)^{ab}$ for all $R \in \mathbb{G}_1$, $Q \in \mathbb{G}_2$, and $a, b \in \mathbb{Z}_q^*$. (2) Non-degenerate: There exist $R \in \mathbb{G}_1$ and $Q\in \mathbb{G}_2$ such that $e(R, Q) \not = 1$. (3) Computable: There is an efficient algorithm to compute $e(R, Q)$ for all $R \in \mathbb{G}_1, Q\in \mathbb{G}_2$. The Chen et al.’s scheme [@C15] uses two untrusted servers $U_1, U_2$. The outsourcer $T$ queries some pairings to the two servers. The scheme can be described as follows. *Setup*. A trusted server computes a table *Rand* which consists of the elements of random and independent six-tuple $(W_1, W_2, w_1W_1, w_2W_1, w_2W_2, e(w_1W_1, w_2W_2))$, where $w_1, w_2 \in_R \mathbb{Z}_q^*, W_1\in_R \mathbb{G}_1$, and $W_2\in_R \mathbb{G}_2$. The table is then loaded into the memory of $T$. - *Look-up table*. Given $A \in \mathbb{G}_1, B \in \mathbb{G}_2$, where $A$ and $B$ may be secret or protected and $e(A, B)$ is always secret or protected. $T$ looks up *Rand* to create $$(V_1, V_2, v_1V_1, v_2V_1, v_2V_2, e(v_1V_1, v_2V_2)),$$ $$(X_1, X_2, x_1 X_1, x_2 X_1, x_2 X_2, e(x_1 X_1, x_2 X_2)),$$ $$(Y_1, Y_2, y_1Y_1, y_2Y_1, y_2Y_2, e(y_1Y_1, y_2Y_2)).$$ - *Interaction with $U_1$*. $T$ sends $\{(A + v_1V_1, B + v_2V_2), (v_1V_1 + v_2V_1, V_2), (x_1X_1, x_2X_2), (y_1Y_1, y_2Y_2)\}$ to $U_1$. $U_1$ returns $$\alpha_1=e(A + v_1V_1, B + v_2V_2),\ \delta=e(V_1, V_2)^{v_1+v_2},\ \beta_1=e(x_1X_1, x_2X_2), \ \beta_2=e(y_1Y_1, y_2Y_2).$$ - *Interaction with $U_2$*. $T$ sends $\{(A + V_1, v_2V_2), (v_1V_1, B + V_2), (x_1X_1, x_2X_2), (y_1Y_1, y_2Y_2)\}$ to $U_2$. $U_2$ returns $$\alpha_2=e(A + V_1, v_2V_2),\ \alpha_3= e(v_1V_1, B + V_2),\ \widehat{\beta_1}=e(x_1X_1, x_2X_2), \ \widehat{\beta_2}= e(y_1Y_1, y_2Y_2).$$ - *Verification*. $T$ checks that both $U_1$ and $U_2$ produce the correct outputs by verifying that $$\beta_1=\widehat{\beta_1}\ \mbox{and}\ \beta_2=\widehat{\beta_2}.$$ If not, $T$ outputs “error". - *Computation*. $T$ computes $e(A, B) =\alpha_1\alpha^{-1}_2 \alpha^{-1}_3 \delta \cdot e(v_1V_1, v_2V_2)^{-1}.$ **Remark 1**. In the original description of the scheme, the step 2 (see section 4.2 in Ref.[@C15]) has not specified any actions. It only explains that the pairing $e(A, B)$ can be composed by the related values. The authors have confused the explanation with steps of the scheme (it is common that a step of a scheme should specify some actions performed by a participator), which makes the original description somewhat obscure. The checking mechanism in the scheme fails ========================================== In the scheme, to check whether the returned values $\alpha_1,\, \delta,\, \beta_1, \, \beta_2$ and $\alpha_2,\, \alpha_3,\, \widehat{\beta_1}, \, \widehat{\beta_2}$ are properly formed, the user $T$ has to check the verifying equations $$\beta_1=\widehat{\beta_1}\ \mbox{and}\ \beta_2=\widehat{\beta_2}.$$ We now want to stress that the checking mechanism *cannot filter out some malformed values*. The drawback is due to that the protected values $A, B$ are not involved in the equations at all. For example, upon receiving $\{(A + v_1V_1, B + v_2V_2),\ (v_1V_1 + v_2V_1, V_2),\ (x_1X_1, x_2X_2), \ (y_1Y_1, y_2Y_2)\}$, $U_1$ picks a random $ \rho\in \mathbb{Z}_q^*$ and returns $$\alpha_1=e(A + v_1V_1, B + v_2V_2),\ \rho,\ \beta_1=e(x_1X_1, x_2X_2), \ \beta_2=e(y_1Y_1, y_2Y_2)$$ to $T$. In such case, we have $ \alpha_1\alpha^{-1}_2 \alpha^{-1}_3 e(v_1V_1, v_2V_2))^{-1} \rho =e(A, B)e(V_1, V_2)^{-v_1-v_2}\rho.$ That means $T$ obtains $e(A, B)e(V_1, V_2)^{-v_1-v_2}\rho$ instead of $e(A, B)$. To fix the above drawback, we have to specify that *both two servers are semi-honest*. The term of semi-honest here means that a server can copy the involved values and always returns the correct outputs, but cannot conspire with the other server. Under the reasonable assumption, the original scheme can be greatly simplified. We now present a revised version of it as follows. - *Look-up table*. Given $A \in \mathbb{G}_1, B \in \mathbb{G}_2$, where $A$ and $B$ may be secret or protected and $e(A, B)$ is always secret or protected. $T$ looks up *Rand* to create $$(V_1, V_2, v_1V_1, v_2V_1, v_2V_2, e(v_1V_1, v_2V_2)).$$ - *Interaction with $U_1$*. $T$ sends $\{(A + v_1V_1, B + v_2V_2),\ (v_1V_1 + v_2V_1, V_2)\}$ to $U_1$. $U_1$ returns $$\alpha_1=e(A + v_1V_1, B + v_2V_2),\ \delta=e(V_1, V_2)^{v_1+v_2}.$$ - *Interaction with $U_2$*. $T$ sends $\{(A + V_1, v_2V_2),\ (v_1V_1, B + V_2)\}$ to $U_2$. $U_2$ returns $$\alpha_2=e(A + V_1, v_2V_2),\ \alpha_3= e(v_1V_1, B + V_2).$$ - *Computation*. $T$ computes $e(A, B) =\alpha_1\alpha^{-1}_2 \alpha^{-1}_3 e(v_1V_1, v_2V_2)^{-1} \delta.$ The checking mechanism in the Chevallier-Mames et al.’s scheme works well ========================================================================= As we mentioned before, the Chen et al.’s scheme [@C15] is derived from the Chevallier-Mames et al.’s scheme [@CC10]. But the new scheme misses the feature of the checking mechanism in the Chevallier-Mames et al.’s scheme. We think it is helpful for the later practitioners to explain the feature. In the Chevallier-Mames et al.’s scheme, the outsourcer $T$ wants to compute the pairing $e(A, B)$ with the help of the untrusted server $U$ such that $A, B$ and $e(A, B)$ are kept secret. The scheme can be described as follows (See Table 1). Table 1: The Chevallier-Mames et al.’s scheme -------------------------------------------------------------------------------- -------------------------------------------------------------------------- ------------------------------------------- The outsourcer $T$ The server $U$ {$P_1 \in \mathbb{G}_1, P_2 \in \mathbb{G}_2$, $e(P_1, P_2)$} Input: $A\in \mathbb{G}_1, B\in \mathbb{G}_2$ Pick $g_1, g_2, a_1, r_1, a_2, r_2\in Z_q^*$, compute $ A + g_1P_1, B + g_2P_2$ Compute $a_1 A +r_1P_1, a_2B +r_2P_2$ $\stackrel{(A + g_1P_1, P_2)}{-------\rightarrow}$ $\alpha_1=e(A + g_1P_1, P_2)$ and query them. $\stackrel{(P_1, B + g_2P_2)}{-------\rightarrow}$ $\alpha_2=e(P_1, B + g_2P_2)$ $\stackrel{(A + g_1P_1, B + g_2P_2)}{-------\rightarrow}$ $\alpha_3=e(A + g_1P_1, B + g_2P_2)$ Compute $\stackrel{(a_1 A +r_1P_1, a_2B +r_2P_2)}{-------\rightarrow}$ $\alpha_4=e(a_1 A +r_1P_1, a_2B +r_2P_2)$ $ e(A, B) =\alpha_1^{-g_2}\alpha_2^{-g_1}\alpha_3 e(P_1, P_2)^{g_1g_2}$ $\stackrel{\alpha_1, \alpha_2, \alpha_3, \alpha_4 }{\leftarrow -------}$ and return them. Check that $ \alpha_4\stackrel{?}{=} e(A, B)^{a_1a_2} \alpha_1^{a_1r_2}\alpha_2^{a_2r_1}$ $\cdot e(P_1, P_2)^{r_1r_2-a_1g_1r_2-a_2g_2r_1}$ If true, output $e(A, B)$. -------------------------------------------------------------------------------- -------------------------------------------------------------------------- ------------------------------------------- Notice that the true verifying equation is $$\begin{aligned} \alpha_4&=& \left(\alpha_1^{-g_2}\alpha_2^{-g_1}\alpha_3 e(P_1, P_2)^{g_1g_2}\right)^{a_1a_2} \alpha_1^{a_1r_2}\alpha_2^{a_2r_1} e(P_1, P_2)^{r_1r_2-a_1g_1r_2-a_2g_2r_1}\\ &=& \alpha_1^{-g_2a_1a_2+a_1r_2}\alpha_2^{-g_1a_1a_2+a_2r_1}\alpha_3^{a_1a_2} e(P_1, P_2)^{g_1g_2a_1a_2+r_1r_2-a_1g_1r_2-a_2g_2r_1} \end{aligned}$$ where $\alpha_1, \alpha_2, \alpha_3, \alpha_4$ are generated by the server $U$, and the session keys $g_1, g_2, a_1, r_1, a_2, r_2$ are randomly picked by the outsourcer $T$. Clearly, the server $U$ cannot generate the four-tuple $(\alpha_1, \alpha_2, \alpha_3, \alpha_4)$ satisfying the above verifying equation because the exponents $$a_1r_2-g_2a_1a_2, \ a_2r_1-g_1a_1a_2,\ a_1a_2,\ g_1g_2a_1a_2+r_1r_2-a_1g_1r_2-a_2g_2r_1$$ are not known to the server. The intractability of the above equation can be reduced to the following general challenge: $$\mbox{Without knowing a secret exponent}\, \theta, \mbox{find}\, X, Y \in Z_q^*, \, X\neq 1, Y\neq 1, \mbox{such that}\, X^{\theta}=Y.$$ The remote and shared servers ============================= The authors stress that the two servers $U_1$ and $U_2$, in the real-world applications, can be viewed as two copies of one advertised software from two different vendors. We would like to remark that the two copies are neither nearby nor private. They must be remote and shared by many outsourcers. Otherwise, the user $T$ equipped with two private copies of one software can be wholly viewed as an *augmented user*. But the situation is rarely considered in practice. We now consider the situation that the outsourcer $T$ has to communicate with two remote and shared servers. If the data transmitted over channels are not encrypted, then an adversary can obtain $A + v_1V_1, B + v_2V_2$ by tapping the communication between $T$ and $U_1$, and get $v_1V_1, v_2V_2$ by tapping the communication between $T$ and $U_2$. Hence, he can recover $A$ and $B$. Thus, it is reasonable to assume that all data transmitted over channels are encrypted. From the practical point of view, the communication costs (including that of authentication of the exchanged data, the underlying encryption/decryption, etc.) could be far more than the computational gain in the scheme. The authors have neglected the comparisons between the computational gain and the incurred communication costs. Taking into account this drawback, we think the scheme is somewhat artificial. A nearby and trusted server --------------------------- Girault and Lefranc [@GL05] have described some situations in which a chip has only a small computation capability is connected to a powerful device. - In a GSM mobile telephone, the more sensitive cryptographic operations are performed in the so-called SIM (Subscriber Identification Module), which is already aided by the handset chip, mainly to decipher the over-the-air enciphered conversation. - In a payment transaction, a so-called SAM (Secure Access Module) is embedded in a terminal already containing a more powerful chip. - A smart card is plugged into a personal computer, seeing that many PCs will be equipped with smart card readers in a near future. We find that in all these situations (a SIM vs. a handset, a SAM vs. a powerful terminal, a smart card vs. a personal computer) the servers are nearby and trusted, not remote and untrusted. Conclusion ========== The true goal of outsourcing computation in the Chen et al.’s scheme is to compute bilinear pairings. In view of that pairings spread everywhere in pairing-based cryptograph, we do not think that the trick of equipping a low capability chip with two untrusted softwares is useful. In practice, we think, it is better to consider the scenario where a portable chip has access to a nearby and trusted server. Otherwise, the communication costs could overtake the computational gain of the outsourced computations. [4]{} X. Chen, et al., Efficient algorithms for secure outsourcing of bilinear pairings, Theoretical Computer Science, 562 (2015) 112-121. B. Chevallier-Mames, et al., Secure delegation of elliptic-curve pairing, in: CARDIS 2010, in: LNCS, vol. 6035, 2010, pp.24-35. M. Girault, D. Lefranc, Server-aided verification: theory and practice, in: ASIACRYPT 2005, in: LNCS, vol. 3788, 2005, pp.605-623. S. Hohenberger, A. Lysyanskaya, How to securely outsource cryptographic computations, in: TCC 2005, in: LNCS, vol. 3378, pp.264-282.

--- abstract: 'We compute and compare the decay lengths of several correlation functions and effective coupling constants in the many-body localized (MBL) phase. To this end, we consider the distribution of the *logarithms* of these couplings and correlators: in each case the log-coupling follows a normal distribution with mean and variance that grow linearly with separation. Thus, a localization length is asymptotically sharply defined for each of these quantities. These localization lengths differ numerically from one another, but all of them remain short up to the numerically observed MBL transition, indicating stability of the MBL phase against isolated ergodic inclusions. We also show how these broad distributions may be extracted using interferometric probes such as double electron-electron resonance (DEER) and the statistics of local spin precession frequencies.' author: - 'V. K. Varma' - 'A. Raj' - 'S. Gopalakrishnan' - 'V. Oganesyan' - 'D. Pekker' bibliography: - 'Ref.bib' date: - - title: 'Length scales in the many-body localized phase and their spectral signatures' --- Introduction {#sec:Intro} ============ Many-body localized (MBL) systems violate many of our expectations from equilibrium statistical mechanics: they do not thermalize under their own intrinsic dynamics [@BAA; @Oganesyan:2007; @ZnidaricPrelovsek; @PalHuse], have extensively many quasi-local conserved quantities [@lbitVadim; @Abanin:2013; @PekkerClark2017PRB], and retain the memory of their initial state at arbitrarily late times [@HuseReview]. These properties of MBL systems are stable to arbitrary (static) local perturbations; in this sense, MBL systems constitute a dynamical phase of matter, the properties of which have been extensively studied in the past decade [@ZnidaricPrelovsek; @Bardarson; @SerbynSlowGrowth; @Schreiber2]. Most of these studies have considered one-dimensional systems, for which the MBL phase has been proved to exist under minimal assumptions [@Imbrie1; @Imbrie2]. The properties of the MBL phase are often characterized in terms of a localization length $\xi$. For instance, the growth of entanglement entropy starting from a product state follows[@ZnidaricPrelovsek; @Bardarson; @SerbynSlowGrowth] $S(t) \sim \xi \log t$; the high-temperature limit of the low-frequency a.c. conductivity is expected to behave as $\sigma(\omega) \sim \omega^{2 - \xi \log 2}$ for a spin-$1/2$ system [@SarangMarkus]; and an instability to isolated ergodic grains is believed[@deroeckHuveneers; @Roeck2] to set in when $\xi = 1/\ln 2$. We have used the same symbol for these various quantities, but they are related to different correlation functions, and there is reason to doubt whether all correlations decay with the same $\xi$. Indeed, whether it makes sense to posit a well-defined $\xi$ in the MBL phase is unclear: based on numerical studies [@DavidVadim; @Rademaker0; @Rademaker; @monthus2016flow; @Pollet; @PalSimon; @ChandranKim], analyses of rare-region effects [@KartiekReview], and the structure of the locator perturbation theory [@Ros], we expect all physical quantities in the MBL phase to exhibit strong fluctuations. ![ Spatial decay rates as a function of inverse disorder strength for effective coupling $J^z$ and correlators xx and xX in the MBL phase (see also Fig. \[fig:Jz-vs-xX\] and Fig. \[fig:Jmed2\], MBL/ergodic transition $\frac{1}{W_c}\approx 0.14$). Dashed lines are fits to expected strong disorder behavior ($\frac{45}{W^{2}}, \frac{2.3}{W}, \frac{3.3}{W}$). Estimated uncertainty is indicated by shading. Gray dotted lines correspond to localization lengths $1/\log 2$ and $1/\log 4$, which are distinct estimates for the onset of the avalanche instability (Sec. 3). Inset: Plot of $\sqrt{R}\sim 1/W$ for $J^z$ coupling. []{data-label="fig:pxi"}](Fig1a_rescaled.eps){width="1\columnwidth"} The present work addresses and clarifies these issues, by analyzing how the probability distributions of correlation functions and effective coupling constants evolve with spatial separation in the MBL phase. These distributions are extracted numerically, using exact diagonalization and the Wegner-Wilson Flow (WWF) method. We find that in each case, the logarithm of the correlation function or coupling follows a normal distribution, with a mean and variance that grow linearly with spatial separation. These features are common to MBL and Anderson insulators; however, in the MBL case no simple relation seems to exist between the coefficients controlling the growth of the mean and variance. (By contrast, in Anderson insulators, for suitably defined quantities, the two coefficients are related by single-parameter scaling [@DMPK1; @DMPK2; @ATAF].) The fact that variance only grows linearly in separation, $r$, guarantees that the inverse localization length for each coupling is sharply defined, with a distribution that narrows at large $r$. Thus one can ask how this quantity varies depending on the coupling. The corresponding spatial decay rate $R = \textrm{exp}(-1/\xi)$ exhibits simple dependencies on disorder strength anticipated by the structure of the perturbative “forward approximation”[@Ros] (Fig. \[fig:pxi\]). Importantly, all these lengths remain quite short inside the MBL phase previously identified in numerical studies. Our results suggest that this regime of disorder is also stable against rare disorder effects nucleating ergodic runaways [@deroeckHuveneers]. Although the distribution for $1/\xi$ narrows for large separation, the couplings themselves become increasingly broadly distributed (log-normal), as noted above, with variance growing linearly in r. We develop protocols for extracting these broad distributions using an interferometric probe related to double-electron-electron resonance (DEER) [@chuang_RMP; @SarangPRL; @kucsko_deer]. We show that a log-normal distribution of couplings implies that the disorder-averaged DEER response decays logarithmically in time, and also affects the statistics of local precession frequencies. Sec. 2 sets up the notation, including model and methods, and also observables of interest. In Sec. 3 we present evidence for the ubiquity of broadening log-normal distributions that enable our definition of length scales shown in Fig. 1. Results on local spectra and spin echoes are presented in Secs. 4 and 5, respectively. We conclude with the discussion of the likely significance of our results and some future directions. definitions {#sec:model} =========== Model Hamiltonian: P-bits and L-bits ------------------------------------ We consider Heisenberg spin chains subject to random fields in the $z$ direction, described by the following Hamiltonian: $$\label{eq:p-bit} H=\sum_{i=1}^{L-1} \sigma_{i}^{x}\sigma_{i+1}^{x}+\sigma_{i}^{y}\sigma_{i+1}^{y}+\sigma_{i}^{z}\sigma_{i+1}^{z} + \sum_{i=1}^{L}h^{\phantom{z}}_i \sigma_i^z,$$ where $\sigma_i^\alpha$ are Pauli matrices, $h_i$ is drawn from the distribution $h \in [-W, W]$, and the system size is $L$. We are interested mainly in the so-called Full MBL (FMBL) regime, where the entire spectrum of the system is MBL for $W\agt 10$ [@ZnidaricPrelovsek; @PalHuse; @Alet; @Devakul] but will also include a value of $W=8$ in the transition regime. Although the model conserves total $\sigma^z$ magnetisation and we may restrict ourselves to one with a fixed value of magnetisation, many of the correlators of interest mix magnetisation sectors; hence our results will be presented for the full model taking all sectors into account. As described earlier, the FMBL phase possesses a complete set of “local" L-bits[@lbitVadim] $[\tau^{z}_i,H]=0$ for $i \in [1,L]$ and corresponding Pauli raising and lowering operators $\tau^\pm_i$. The Hamiltonian can be re-written as [@lbitVadim; @Abanin:2013] $$\label{eq:lbit1} \tilde{H} = E_0 + \sum_i B_i \tau_i^z + \sum_{i>j} J_{ij}\tau^z_i\tau^z_j + \sum_{i>j>k} J_{ijk}\tau^z_i\tau^z_j\tau^z_k + \ldots .$$ The conserved charges $\tau^z$’s maybe thought of as obtained from $\sigma^z$ with appropriate dressing by (small) quantum fluctuations due to off-diagonal terms (in the z-basis)[@Ros]. Multispin interactions $J_{ijk\ldots}$ induced by these virtual exchange processes should decay exponentially with the end-to-end distance among the spins (and therefore with the number of spins involved). In the regime where $W \gg 1$ we expect the dressing to be weak; therefore we expect the local fields $B_i$ to be close to the microscopic $h_i$, and the high-order terms in Eq.  to fall off rapidly. Observables {#sec:observables} ----------- ### Couplings {#sec:Jzs} An alternative and useful representation of the L-bit Hamiltonian is obtained by focusing on subsystems of few spins treating the rest as a static environment, thereby trading infinitely many multispin interactions for *distributions* of few spin terms. For example, if the subsystem is just a single L-bit at site $j$ $$H_{j|\text{env}}=B_{j|\text{env}}\tau^z_j, \label{eq:H1}$$ where the total number of $B_{j}$’s (and $H_{j}$’s) is $2^{L-1}$, with each rearrangement of environment’s spins, $|\text{env}$, contributing to a spectral shift of an otherwise sharp local line. The statistics of these local fields is interesting and will be examined in Section \[sec:LSF\]. If we instead look at two-spin subsystems with separation $r=j-k$ (henceforth we drop the subscript $|\text{env}$) $$H_{j,k}=B_{j}\tau^z_j+B_{k}\tau^z_{k}+J^z_{j,k}\tau^z_{j}\tau^z_{k} \label{eq:H2}$$ we can access the distribution of the two spin couplings, which is expected to exhibit a “flow" previously described as an evolution to a broad $1/f$ law in the MBL phase[@DavidVadim] as $r\to \infty$. The flow reverses towards narrow (approximately Gaussian) distributions in the ergodic phase with the critical regime appearing as a family of non-flowing “scale invariant" distributions[@DavidVadim]. Importantly, the bulk of previously computed results[@DavidVadim] did *not* use the effective two spin distributions described here but rather a variant with only the fluctuations in the intervening region (i.e. only $r$ spins bookended by the two-spin subsystem) accounted for. While the prior choice was physically motivated, e.g. with distributions’ cardinality growing $\sim 2^r$ as the more and more spins “mediate" the 2-spin coupling, we found empirically in this work that including *all* spins (and thereby fixing the cardinality to $2^{L-2}$ for all r) significantly changes (reduces) finite size effects and vastly improves the overall quality of simple exponential fits, enabling for unambiguous extraction of length scales. ### Transverse correlators In addition to distributions of one- and two- L-bit terms in $H$ we will be interested in two transverse correlators, $\sigma^x_j\sigma^x_{k}$ and $\sigma^x_j\tau^x_{k}$, which we will refer to as xx and xX (note that XX is trivial by construction). Both of these correlators are interesting, albeit for different reasons: xx involves physical P-bits and can therefore be measured directly, while xX is important in studying effects of MBL subsystems weakly coupled to other degrees of freedom, e.g. the argument and the analysis of the instability due to isolated ergodic grains[@deroeckHuveneers] makes plausible assumptions about xX. *Importantly*, in what follows we only consider the *amplitude* of these correlators, i.e. averages of $|\langle n|\sigma^x_i \sigma^x_j|n\rangle|$. These are *not* instantaneous equilibrium observables, as they can only be extracted from the Edwards-Anderson type “persistence in time" order parameter, $|\mathcal{O}|\equiv\sqrt{\lim_{T\to \infty} \int^T_0 dt \langle \mathcal{O}(t)\mathcal{O}(0)\rangle/T}$, e.g. with $\mathcal{O}=\sigma^x_i \sigma^x_j$ – this point has been extensively discussed and used in the context of eigenstate order in Hamiltonian problems [@PekkerPRX; @SarangPRL; @SarangFermi]. Finally, we note that all averaging in this paper is done by computing expectation values in eigenstate and then performing the equal weight (infinite temperature) Gibbs average and finally averaging over disorder realizations. Wegner-Wilson flow ------------------ A priori there is no unique way for constructing the dressed operators $\tau^z$; this is because the similarity transformation that diagonalizes $H$ needs to come with a single, consistent labeling scheme that assigns an L-bit label to each of the $2^L$ eigenstates of $H$ (out of the $2^L!$ possible labeling schemes). However, a posteriori, physical constraints help sieve out a good method for the constructing this transformation: well-defined spatial locality of couplings $J$, and tightness of fields $B$ about the physical onsite potentials. By any definition of localization, these are reasonable requirements to be satisfied. The renormalization group-like technique of Wegner-Wilson flow (WWF) [@Wegner; @Kehrein; @DavidVadim] achieves these admirably. In localized systems specifically, there are clear reasons for why WWF works well, as discussed in the literature – this has to do with the order in which off-diagonal matrix elements are eliminated and the reversibility of the flow. [^1] Let the WWF parameter be labelled as $\kappa$; then the flow equation that transforms $H \rightarrow \tilde{H}$ is given by $$\label{eq:flow1} \frac{d H(\kappa)}{d \kappa} = \mathcal{G}(\kappa, H),$$ where the function $\mathcal{G} := [H, \eta]$ is chosen so that (a) there is good stability in the flow as $\kappa : 0\rightarrow \infty$, and (b) the parts of the spectrum that we want to eliminate fall off quicker if they have larger energy separation. A canonical choice of the generator $\eta$ [@Wegner; @Kehrein], in particular for diagonalising the entire system [@DavidVadim], is given by $$\label{flow:Gen} \eta (\kappa) = [V(\kappa), H(\kappa)],$$ where $V(\kappa)$ is the off-diagonal part of $H(\kappa)$, and $\eta(\kappa) = -\eta(\kappa)^{\dagger}$ is an antiunitary generator. With these flow equations (and the understanding that $H(0) = H$ and $H(\infty) = \tilde{H}$) the transformation relating the two representations Eq. and Eq. is given by $$\label{eq:Transform} \tilde{H} = U^{-1}HU,$$ where columns of $U$ contain the L-bit representations of the eigenstates in the correct order as defined in the original *P-bit* representation. Then the dressed spin operators are given by $$\label{eq:STransform} \tau^{\alpha} = U \sigma^{\alpha} U^{-1},$$ for $\alpha = x, y, z$ The transformation matrix $U$ that holds these L-bit representations is governed by a similar flow equation $$\label{eq:Uflow} \frac{d U(\kappa)}{d \kappa} = U \eta,$$ with $U(0) = \mathbbm{1}$. Columns of $U$ along with the diagonal of $\tilde{H}$ completely characterises the spectrum of the problem. [^2] Distributions and localisation lengths {#sec:lengths} ====================================== \ This section makes two essential points. First, log-normal distributions are natural and pervasive in localized problems. Second, unlike in Anderson localized (single particle) case, multiple length scales are required to characterize distributions in many-body problems. Warmup – single particle and non-interacting fermions ----------------------------------------------------- It is well known that the wavefunction intensity $G=|\psi|^2$ (or measurable conductance) of Anderson localized single particle states in 1D is log-normally distributed [@DMPK1; @DMPK2; @ATAF] $$\label{eq:DMPK} P(\log G) = \sqrt{\frac{\xi_2}{8\pi r}}\textrm{exp}\left[ -\frac{(\log G + 2r/\xi_1)^2}{8 r /\xi_2}\right],$$ with $r$ denoting spatial separation. Remarkably, both the mean and variance of $G$ are controlled by the same $\xi=\xi_1=\xi_2$ – the localization length – that fully characterises the localised phase[@ATAF; @DMPK1; @DMPK2]. That this length (or rather its inverse) is sharply defined is made clear by considering a distribution for $1/\xi\equiv-\log G/(2 r)$ which narrows to a $\delta$ function in the limit $r\to\infty$. The fluctuations in $1/\xi$ remain important for some observable quantitites, e.g. they are known to alter the prefactor to the Mott a.c. conductivity of disordered systems [@MottHybrid]. Importantly, $\xi$ in Eq. \[eq:DMPK\] varies somewhat with energy even at strong disorder and appreciably at weak disorder. This, combined with specifics of averaging, may result in significantly different statistics in non-interacting many-body problems, e.g. at low but finite temperature in a weakly interacting regime. With an eye to interacting spin chains, we consider infinite-temperature correlators in the noninteracting many-fermion problem in order to understand the distinction between xx and xX correlators. We define locator fermion annihilation operator on site $j$, $c_j$, and also that of the unique eigenstate “attached" to site $j$ in perturbation theory $\gamma_j$. The point we wish to make is that xx and xX spin correlators defined above are analogous (sans Jordan-Wigner phase) to $$\begin{aligned} \langle n| c_i^+ c_{j} |n \rangle &= \sum_k f(k) \psi^*_k(i) \psi_k(j) \label{eq:fermxx} \\ \langle n| c_i^+ \gamma_{j}|n \rangle&=f(j)\psi^*_j(i), \label{eq:fermxX}\end{aligned}$$ respectively. (Note: $f(\alpha)=0,1$ is the occupation of each single-particle state $\alpha$ in $|n\rangle$) Comparing typical decay rates of the two correlators we expect the first one to decay slower since it involves a sum over many orbitals and is dominated by the slowest-decaying orbitals that contribute. Fluctuations in the single-particle localization length, and particularly its energy-dependence, imply that the length-scale in Eqs. should systematically exceed that in Eq. . Many-body case -------------- We now present one of the central results: in Fig. \[fig:Jz-vs-xX\] – the distributions of xx and xX correlators and of $J^z_{i,j}$ effective couplings sampled over many-body eigenstates and disorder realizations. Their apparently log-normal shape and linear growth (with separation) of both mean-log and var-log naturally leads to a spectrum of length scale parameters, two per observable. Following the non-interacting example we expect xx to decay slower than xX and also display weaker fluctuations, as it “pre-averages" over single particle fluctuations. These expectations are clearly borne out (see Figs. \[fig:pxi\], \[fig:Jz-vs-xX\] for averaged quantities, and Fig. \[fig:SingleSample\] for a single sample.). Finally, we expect and observe that the spatial decay rate of the typical $J^z$ is $e^{-1/\xi_1}\sim 1/W^2$. At strong disorder, the leading contribution to the typical $J^z$ comes from Hartree shifts of the nearly site-localized orbitals. These Hartree shifts induced by site $i$ at site $j$ scale as the *square* of the amplitude of orbital $i$ at site $j$; since the amplitudes decay as $(1/W)^{|j - i|}$, the effective interaction $J^z$ decays as $(1/W^2)^{|i - j|}$. This difference in dependence on $1/W$ is borne out as seen in Fig. \[fig:pxi\]; the error bars are determined by the uncertainties of fitting various segments and various system sizes. Two comments are in order. First, all length scales are quite short, suggesting that most of the numerically observed localized phase (for $W \agt 10$, corresponding to $W \agt 5$ in the notation of Ref. [@PalHuse]) is stable against the inclusion of ergodic grains. The exact threshold for the critical localization length in the generic interacting case is not fully settled. In the case of noninteracting l-bits [@Roeck2], the instability happens when $\xi = 1/\log 2$. When the l-bits are interacting, a given l-bit at a distance $r$ from the ergodic inclusion can couple to the inclusion via many distinct processes [^3] (depending on whether the $r - 1$ intermediate l-bits get flipped or not) and a conservative estimate of when the instability sets in is $\xi = 1/\log 4$. Both criteria are marked in Fig. \[fig:pxi\]. Second, much of the ensemble-averaged physics is present in single samples as well; see Fig. \[fig:SingleSample\] where we present the decays of the three observables for two different samples with different disparate $\xi$ values, one small and one big. We find that the xx and xX decays still follow the same trend with respect to each other, in particular Eq. , as we see from the middle column of plots. Moreover the right column of Fig. \[fig:SingleSample\] shows, upon comparing with its middle column, that quicker decays of couplings (stronger localization) results in a broader spread of local field splittings (see next section for its definitions) and hence broader spread of $J^z$ couplings. To summarize, each two-point object of interest may be efficiently “labeled" using two length scales, $\xi_1$ and $\xi_2$, encoding typical decay and growth of fluctuations. The first of these connects to the established body of work where properties of the MBL phase were understood using L-bit phenomenology, albeit with a potentially important caveat that different observables are controlled by numerically different decay lengths. It is an interesting exercise to re-examine and correct these prior results to account for fluctuations. In some cases we expect the corrected answer to be qualitatively similar to the mean-field one. For example, the logarithmic growth of entanglement will now come with strong but subdominant fluctuations – the so called “logarithmic lightcone" will be smoothed out on scales $\sim \sqrt{\log t}$, i.e. there is a broad but still discernible boundary demarcating entangled spins in real space. Unless the two lengths turn out vastly different (they are not in our case but may be in other models) the mean-field theory is still a reasonable if incomplete description in such cases. The more interesting possibility, perhaps (of course!), is to find examples of phenomena where these fluctuations invalidate mean-field expectations entirely – we focus on two such cases next. Local spectral functions {#sec:LSF} ======================== \ Local spectra can be defined and measured experimentally via autocorrelations of single spin operators, e.g. Gibbs-averaged $$A_{\textrm{xx}}(\omega)=\int dt e^{i \omega t} \langle \sigma^x_j (t)\sigma^x_j (0)\rangle. \label{eq:Axx}$$ These are generally complicated convolutions of spectra and matrix elements. For a chain of $L$ spins we expect $\sim 2^{2L}$ broadly distributed contributions to $A_{\textrm{xx}}(\omega)$. In MBL systems, with well- (or at least usefully) defined L-bits, $\sim2^L$ of these values are parametrically larger than the rest. We can vastly simplify (“clean up") the situation by considering spectral functions of L-bits (its relationship to $A_{\textrm{xx}}(\omega)$ will be explained at the end of this Section) $$A_{\textrm{XX}}(\omega)=\int dt e^{i \omega t} \langle \tau^x_j (t)\tau^x_j (0)\rangle, \label{eq:AXX}$$ thereby removing the fluctuations of the matrix elements and vastly reducing the number of terms, down to $2^L$, with $A_{\textrm{XX}}$ equal to the probability distribution of local fields in Eq. \[eq:H1\], $P(B_{j})$. The spectral line starts infinitely sharp in non-interacting systems and becomes splintered by L-bit interactions. We will focus on the distribution of *spectral shifts* (i.e., spacings between adjacent frequencies), $P(\delta B_{j}=B_{j+1}-B_{j})$, to elucidate local spectral correlations. We find that $P(\delta B)$ appears to follow log-normal statistics with a dramatic overall increase in the dynamical range upon entering the MBL phase (see Fig. \[fig:SpectralProps\](a)) implying a simple $1/f$ powerlaw as before, albeit with system size $L$ providing regularization instead of separation $r$ in two point observables. We argue below that to reproduce this power law one must include fluctuations of the $J^z$ couplings. To start, it is helpful to visualize the local spectrum as a splintering process using a spectral tree[@NGH], see Fig. \[fig:SpectralProps\](b). Here the root of the tree is the frequency of the isolated site $j$ (average of $B_{j}$ over all the environment spins, corresponding to the local onsite potential), and each generation corresponds to incrementally turning on (exponentially decaying) interactions to further neighbors (or equivalently, only averaging interactions with progressively distant spins). Since each site has a spectral tree associated with it, we have chosen the rightmost boundary site from each sample (so that the starting value in each tree is approximately the onsite potential at $x=10$); note that the boundary sites show the strongest splitting because of the largest available distances. Deep inside the MBL phase we expect a very rapid decay of the coupling and the tree not to cross itself, with half of $\delta B$’s *equal* to (twice) some interaction term with the most distant spin, a quarter of $\delta B$’s corresponding to sums and differences of two interaction terms etc. Put differently, there is a considerable “which path" or “branching" memory which should manifest in a fractal-like structure of local MBL spectra and a non-trivial non-universal powerlaw distribution $P(\delta B)\sim |\delta B|^{-1-\xi_1 \log 2}$ (see appendix). By contrast, when interactions are strong, we do not expect any branching memory, with each $\delta B$ obtained from a random combination of many $J$’s – adjacent frequencies correspond to configurations that differ by several spin-flips, hence $P(\delta B)$ should obey Poisson statistics in the ergodic phase. This transition takes place, within the mean-field picture, at $e^{-1/\xi_1}=1/2$ (note that here we are considering the special case of an edge spin [@NGH]). The degree of self-crossing in the spectral tree can be quantified, in fact, if we examine the distribution of Hamming distances corresponding to each $\delta B$ (see Appendix for details on its computation). These appear quite short in the MBL phase (see Fig. \[fig:SpectralProps\](c)), and is smaller than the fully ordered mean-field prediction of $\langle \textrm{Hd} \rangle = 2$, which can occur due to rearrangements of L-bits. The fully-ordered mean-field case is easily seen to have a distribution of $P_{\textrm{MF}}(\text{Hd}) = 2^{-\text{Hd}}$. Away from this fully-ordered mean-field limit, we observe (Fig. \[fig:SpectralProps\](c)) that the actual data exhibits a generalized exponential behavior $$\label{eq:PHd} P(\text{Hd}) = (\kappa-1) \kappa^{-\text{Hd}},$$ with $\kappa = \kappa(W)$, a nondecreasing function of disorder strength. Note that Eq. is a valid probability distribution defined at the positive integers, with $\langle \text{Hd} \rangle = \frac{\kappa}{\kappa-1}$. We therefore observe that while the tree remains essentially non-crossing as anticipated by the mean-field model, the spectral statistics are much more universal, with a simple powerlaw of $-1$. To understand this result we now allow multiplicative disorder in the coupling strength, to mimique log-normal distributions with growing variance (as detailed in the previous section). Thusly modified L-bit description may be simulated straightforwardly numerically, see Fig. \[fig:SpectralProps\](d). We clearly observe the existence of a simple $1/f$ powerlaw. Indeed, in principle, there should also be a crossover from a non-universal to the asymptotic $1/f$ powerlaw (at smallest $\delta B$’s) – this requires $\xi_2/\xi_1 \ll 1$ which is not the case for spin chains studied here. While the local L-bit spectral function is not directly measurable, when $\xi_1 < 1$ one can approximate it well by taking a physical spectral function, measured by standard spectroscopic means, and dropping all spectral lines below a certain threshold intensity when computing gaps. In practice, extremely small splittings will not be resolvable, so one can only measure the “tree” out to a depth set by experimental resolution. Assuming that the experimental resolution is $500$ times the microscopic energy scales (as is reasonable for present-day experiments with ultracold atoms and superconducting qubits [@Google1; @Google2]), and that $1/\xi_1 = 2/3$, one can resolve up to four generations of the tree, which should be adequate to test the predicted hierarchy of gaps. The protocol to measure the energy spectrum of a set of spins or qubits consists of creating local excitations (through, say, a magnetic $\pi/2$ pulse) and measuring the time-dependent vibrational response: a simple Fourier transform will then reveal the characteristic modes (eigenenergies) of the system [@Google1]. In order to construct the tree, and the corresponding splittings, up to some desired level, the strongest $2^L$ spectral lines may be retained. Echoes and L-bit dynamics {#sec:echoes} ========================= ![ P-bit DEER signal for $W=15$ and $L=8$ compared against Fourier transform of the distribution of L-bit $J^z$ (Sec. \[sec:Jzs\]). We rescaled the former signal by a small factor to line-up the traces at short-test times. []{data-label="fig:pxiDEER"}](Fig1b_rescaled.eps){width="0.99\columnwidth"} The existence of L-bits may be demonstrated using several related but inequivalent dynamical protocols [@SarangPRL; @Rademaker; @VasseurMoore; @ChandranKim]. The one of interest here is a *local* Hahn echo which consists of a familiar (e.g. from NMR) sequence of pulses, however with all three pulses applied on the same site $j$ of the chain $$|t\rangle=R_j^{\pi/2} e^{-iHt/2} R_j^{\pi} e^{-iHt/2} R_j^{\pi/2} |\uparrow\rangle, \label{eq:Hahnt}$$ where $R_j^{\phi}=\exp(i \phi \sigma_y/2)$ and $|\uparrow\rangle$ is a state with site $j$ fully polarized (either via quench or by pre-measurement.) With these manipulations the persistent echo is obtained in the MBL phase $$D_j(t\to \infty)\equiv\langle t| \sigma_j^z | t \rangle>0 \label{eq:Hahn}$$ The initial state may be a unique state, e.g. a particular product state that is easy to prepare or an eigenstate which might be subsequently averaged over to simulate a thermal distribution (which may be imposed by coupling weakly to the environment). The only difference with textbook NMR discussion of Hahn echo is that there the pulses apply globally and can only rephase the decoherence from chemical shifts. Applying pulses locally in MBL systems effectively performs Hahn rephasing in the total static field comprised of local single body and interaction components. As previously discussed[@SarangPRL] the existence of the echo may be interpreted similarly to a finite quasiparticle residue in Fermi liquids which guarantees that calculations done in the renormalized model are directly measureable by coupling to actual degrees of freedom. MBL states in finite chains even at moderate disorder tend to support a reasonably visible echoe amplitude, e.g. at $W=8$ typical $D_j$ is about 0.5 (it is 1 for the perfect echo performed with exact L-bit rotations). Building on this we design a two-spin echo \[similar to the double electron-electron resonance (DEER)\] to extract the information about the distribution of couplings $J_{jk}$ between two L-bits at a specified distance $r=|j-k|$. To start we ignore the difference between L-bits and P-bits, i.e. assume $D_j=1$. In this protocol, one performs a local Hahn spin echo on spin $j$; however, simultaneous with the $\pi$-pulse applied to spin $j$ in this protocol, one also applies a $\pi$-pulse to spin $k$ $$|t\rangle=R_j^{\pi/2} e^{-iHt/2} R_j^{\pi} R_k^{\pi} e^{-iHt/2} R_j^{\pi/2} |\uparrow\rangle. \label{eq:Deer}$$ All couplings acting on $j$ *except* that due to $k$ remain echoed out as above, allowing one to measure $J_{jk}$ without having it masked by the stronger couplings due to spins closer to $j$ which allows for the signal to decohere in time following $$D_{jk}(t)=\langle e^{i J_{jk} t}\rangle=\int dJ P_{jk}(J) e^{i J t}. \label{eq:ftdeer}$$ The time-dependence of the *averaged* DEER response is precisely the Fourier transform (i.e., characteristic function) of the probability distribution of $J_{jk}$ so the DEER protocol allows for a concrete test of our predictions concerning the distribution functions. What should we expect for $[\mathcal{D}_{jk}(t)]$ assuming a log-normal distribution of $J$? We expect that $[\mathcal{D}_{jk}(t)]$ decays from nearly 1 to zero albeit logarithmically slowly, as implied by the $1/J$ prefactor in the log-normal distribution. It is especially illuminating to consider the dependence of the decay profile on the separation between the two spins in the echo $r=|j-k|$ – signal’s half-life, $t_{1/2}$, is directly determined by the typical (log-mean) coupling at that separation ($\log t_{1/2}\approx \log 1/J_{typ}$), while the log-slope reflects the fluctuations - it decreases as inverse root variance $$D_{jk}(t)\propto -\sqrt\frac{\xi_2}{r}\log t$$ Our results at moderate disorder $W=15$ (Fig. \[fig:pxiDEER\]) are clearly consistent with these expectations. Discussion ========== In this work we explored multiple ways of characterizing localization lengths and their distributions in the MBL phase. We found that localization lengths extracted from distinct observables do not coincide in general, but all of them remain short throughout the MBL phase. Thus the apparent MBL phase at small system sizes seems stable with respect to rare configurations of disorder hosting thermalizing grains; this is consistent with a scenario in which the true MBL critical point occurs at comparable disorder to the numerically observed one. The spatial correlation functions and couplings from which we extracted localization lengths share the feature of having log-normal distributions at large separation, with a width that broadens as the separation increases. This feature was noticed in previous work as an approach to a $1/f$ distribution; here, we identify it as a broadening log-normal, a type of behavior that is qualitatively similar to what happens in Anderson localization [@DMPK1; @DMPK2]. The interplay between interaction effects and these broad distributions gives rise to qualitatively modified spectral signatures: both the statistics of local spectral lines and the response to the “DEER” spin echo protocol differ qualitatively from naive predictions that ignore the broadening of distributions. We expect similar qualitative modifications for other physical quantities (e.g., post-quench response functions and a.c. conductivity) in which localization lengths appear in the exponent; these consequences will be explored in future work. *Note added*.—While this manuscript was being completed, a related work was posted [@paola], which presented a different algorithm for extracting l-bits and the distribution of localization lengths. Acknowledgments =============== We thank W. Bialek, T. Can, D. A. Huse, and A. Scardicchio for discussions. D.P. and V.O. also thank B. K. Clark and E. Kapit, respectively, for prior exploratory collaborations that nucleated some of the ideas in this work. VKV and VO acknowledge support from the NSF DMR Grant No. 1508538 and US-Israel BSF Grant No. 2014265. SG acknowledges support from NSF Grant No. DMR-1653271. Appendix {#appendix .unnumbered} ======== Further details on Sec. \[sec:observables\]: 2-spin coupling protocol --------------------------------------------------------------------- In our work, to expound on Eq. , the representation of the L-bit representation is by subsuming multispin interactions into a function that further dresses an effective two-spin model: $$\label{eq:lbit2} \tilde{H} = E_0 + \sum_i B_i \tau_i^z + \sum_i \sum_{r} J^z_r \tau_i^z \left( \sum_{m=1}^{r-1} \prod_{k=1}^{m}\tau^z_{i+k} \right) \tau^z_{i+r}.$$ Here too the interactions are beyond nearest neighbour; with this re-representation, however, we will find a much more systematic variation of the couplings with system size, and thence a better definition of localisation regions in space. In this protocol we are interested in generating the couplings $J^z_r$ as written in Eq. . For a given sample and any range $r$, there are $2^{L-2}$ such couplings. For each state $\psi^{(L-2)}$ in the $2^{L-2}$-dimensional Hilbert space, we take $$\label{eq:constraint} \psi_{r_1, r_2} = \psi^{(L-2)} \otimes \{r_1, r_2 \},$$ where $r_1, r_2$ denote up or down spins (4 combinations), with the constraint that their site positions $x(r_i)$ on the lattice are given by $$x(r_1) - x(r_2) = r.$$ Then for each $\psi^{(L-2)}$ we solve the linear equation $$\begin{bmatrix} +1 & -1 & -1 & +1 \\ +1 & +1 & -1 & -1 \\ +1 & -1 & +1 & -1 \\ +1 & +1 & +1 & +1 \\ \end{bmatrix} \times \left[ \begin{array}{c} E_0 \\ J_1 \\ J_2 \\ \tilde{J}_{12} \end{array} \right] = \left[ \begin{array}{c} E_{\psi_{0,0}} \\ E_{\psi_{0,1}} \\ E_{\psi_{1,0}} \\ E_{\psi_{1,1}} \end{array} \right],$$ where the $E_{\psi_{r1,r2}}$’s are the eigenvalues corresponding to that $l$-bit configuration obtained from Wegner flow.\ Then $J^z := \tilde{J}_{12}$ gives the renormalized coupling from the four states $\psi_{r_1, r_2}$. This is repeated over all $\psi^{(L-2)}$, and $x(r_1), x(r_2)$ (again such that Eq. is satisfied).\ \ \ Further details on Sec. \[sec:observables\]: multispin coupling protocol ------------------------------------------------------------------------ In this protocol we are interested in generating the couplings $J^z_r$ as in previous section but with an additional index that denotes the number of coupled $\tau^z$ operators. For a given sample and range $r$, there are $2^{r}$ such couplings. Consider L-bits $0000, 0001, 0010, \ldots$ such that $\tau_i |0\rangle = +1|0\rangle$. Then after diagonalizing: $$\begin{bmatrix} +1 & +1 & \ldots & +1 & +1 & \ldots \\ -1 & +1 & \ldots & -1 & -1 & \ldots \\ \vdots & \vdots & \vdots\end{bmatrix} \times \left[ \begin{array}{c} E_0 \\ J_1 \\ \vdots \\ J_{12} \\ J_{13} \\ \vdots \end{array} \right] = \left[ \begin{array}{c} E_{0000} \\ E_{0001} \\ \vdots \end{array} \right]$$ The $J_{i}, J_{i,j}, J_{i,j,k} \ldots$ can be obtained by solving above linear equation.\ The right-hand side are simply eigenvalues from WWF. The left-hand-side matrix is known a priori because of ordering of L-bits in the given sector. For $r=L$ there are equal number $2^{L-2}$ of $J^z$ and $J^z_{\textrm{ms}}$ couplings; using $L=3, 4$ we can show that these two sets of couplings, written below as a vector, are related by $$\label{eq:DEERvsTEXAS} J^z_{\textrm{ms}} = \left(\prod_{i=1}^{L-2} \otimes a \right) J^z,$$ where $a = \frac{1}{2}\begin{bmatrix}1 & 1\\ -1 & 1\end{bmatrix}$. So we see that even for the largest range, the two sets of couplings will be different numerically: however their decay tendencies are *qualitatively* similar. Spectral trees: mean-field and beyond ------------------------------------- In a general MBL system, the location of the spectral line depends on the global configuration of L-bits, via the interaction terms in Eq. . The state of every L-bit $j \neq i$ affects that of L-bit $i$, though the effect falls off exponentially with the separation as $\exp(-|j - i|/\xi)$. The effects of the other spins on the local field at site $i$ can be understood in terms of a spectral “tree” with $2^{L - 1}$ nodes at the edge, each corresponding to the effective local field in a many-body eigenstate. The structure of this spectral tree was first discussed in Ref. [@NGH], in a treatment that implicitly assumed a unique, sharply defined localization length; we refer to this, below, as a “mean-field” treatment. According to this treatment, when $\xi < 1/(2\ln 2)$, the splittings due to distant lines fall off sufficiently fast that the thermally averaged spectral function does not fill in; instead, it forms a fractal structure with spectral gaps at all frequency scales [@NGH]. This can be seen as follows: the typical splittings at scale $L$ are $\sim \exp(-L/\xi)$, and those at scale $L + 1$ are $\exp(-(L+1)/\xi)$. When $\exp(-1/\xi) \alt 1/4$, the branches at stage $L+1$ coming from the four different nodes at stage $L$ typically do not cross. The resulting thermally averaged spectral function has gaps at scales $\exp(-n/\xi)$ for all $n$, analogous to a Cantor set. This “mean-field” treatment can be used to derive the distribution of gaps between adjacent spectral lines in the local, thermally averaged L-bit spectral function. A simple version of this argument, valid for $\xi \ll 1$, is as follows. Consider the probability of a given spectral gap $\delta$ exceeding some threshold $\delta_0$, i.e., $P(\delta > \delta_0)$; these gaps correspond to splittings at very early stages in the spectral tree, and thus to events at distances $r \alt r_0 \approx \text{const.} - \xi \log \delta_0$. The total number of such splittings scales as $P(\delta > \delta_0) \approx 2^{r_0} \sim \delta^{-\xi \log 2}$. Differentiating this to get the probability distribution, we find that $$\label{eq:MF-PdB} P(\delta) \sim \delta^{-1 - \xi \log 2}.$$ This analytic prediction agrees with numerical simulations of the mean-field theory (including multiplicative noise of order unity, which is inevitably present because of the randomness of matrix elements). While the mean-field theory predicts a continuously varying exponent that is always greater than unity, numerically we find that $P(\delta) \sim 1/\delta$ throughout the MBL phase. In the main text, we have “fixed” this mean-field picture by including noise in the exponent, so that there is a crossover from Eq. to $1/\delta$ behavior at a given $\delta B$ (which depends on the strength of the noise.); see Fig. \[fig:SpectralProps\](b). Beyond mean-field, in our case, the local fields are given by $B_i$ of Eq. . They are generated by taking the difference in eigenvalues of L-bit states that are flipped at site $i$: $$\label{eq:Bdef} B_i = E_{\ldots 1_i \ldots} - E_{\ldots 0_i \ldots};$$ this equation is only true on average, because upon taking the difference on RHS a whole bunch of $J^z$ couplings will also enter into the mix that dress the $B_i$ fields.\ To be contrete, at first order in perturbation theory of couplings, these are split into $B_i \pm 2\Delta$ for $i$ in the bulk and $B_i \pm \Delta$ for $i$ at the boundaries. Already at this level, we see how the level spacing $\delta B$ is proportional to the coupling $J$. Introducing further spins generates more splittings; this will become clearer if we employ the multispin couplings $J^z_{ijk\ldots}$ for a boundary spin. At first order the splitting is $\pm J^z_{12} \approx \Delta$; at next order the upper branch ($B_i + J^z_{12}$) has new splittings of $\pm (J^z_{13} + J^z_{123})$, whereas the lower branch ($B_i - J^z_{12}$) has new splittins of $\pm (J^z_{13} - J^z_{123})$, and so on. A spectral tree will thus be built up at each site, whose splittings are determined by combinations of $J^z$ couplings. Hamming distance ---------------- The Hamming distance is defined as the number of spin flips required to go from one spin configuration (whether L-bit or p-bit) to another. For example the states $1011$ and $0111$ have a Hamming distance of two between them. Clearly there is some arbitrariness in how we choose from which pair the states to construct the Hamming distance; however following our discussion of local fields and their splittings (especially Fig. \[fig:SpectralProps\](a),(b)) we see that adjacent pairs of L-bit states (ordered by their local fields) are a good indicator of localisation, i.e. broader the distribution of $P(\delta B)$, smaller the typical $\xi$, and stronger the localisation.\ This immediately implies that adjacent states in a strongly localised system will be connected to each other by fewer number of spin flips than in an ergodic system (where the median field splitting is much larger). This expectation is borne out, as displayed in the inset of Fig. \[fig:SpectralProps\](d) where we display the mean Hamming distances as a function of disorder strength for three system sizes $L=8,10,12$. There appear to be stronger finite size effects as we approach the ergodic phase, which considerably decrease as we wade deeper into the localised phase.\ Moreover, interestingly, in the main panel of Fig. \[fig:SpectralProps\](d) we see that the Hamming distances, sorted over all samples and all states, follow a Poisson distribution as we enter deep into the localised phase. This means that there is an emergent pair-wise statistical independence of spin-flips among adjacent states as we enter deeper into the localised phase. The exponential distribution of spin-flips conspicuously breaks down as the disorder is decreased, and the ergodic phase is entered. The reason for this change in distribution is unclear to us for now but it provids us a clear spectroscopic probe of localisation that is readily amenable to spin-flip experiments.\ The main upshot of this Appendix is that the Hamming distances clearly demarcate localised vs. ergodic physics, whether through their means or the distributions.\ Perturbation theory ------------------- \ Our starting point is a three site chain with strong onsite potentials, with the Hamiltonian $H_0$ diagonal in these fields and the interaction of strength $\Delta$. Using second order perturbation theory in xy terms $H_1 = J_{xy}(\sigma^x_0\sigma^x_1 + \sigma^y_0\sigma^y_1 + \sigma^x_1\sigma^x_2 + \sigma^y_1\sigma^y_2)$ we may perturbatively construct the eigenstates up to second order in $J_{xy}$. Let us denote the difference in local disorder fields between sites to be $\delta_{02}, \delta_{10}, \delta_{12}, \delta_{02}$; with $J_p = 2J_{xy}$, the fields of $\sigma^x \sigma^x$ and $\sigma^x \tau^x$ correlators are $$\begin{aligned} \label{eq: Pertxx} \langle \sigma^x_0 \sigma^x_1 \rangle &=& \{ 0, 0, \frac{J_p^3}{4\delta_{02}(\delta_{12} \pm \Delta)^2}, \pm \frac{J_p}{\delta_{10} + \Delta}, \pm \frac{J_p}{\delta_{10} - \Delta}\} \nonumber \\ \langle \sigma^x_0 \sigma^x_2 \rangle &=& \{ 0, 0, \frac{J_p^2}{2\delta_{02}(\delta_{12} \pm \Delta)}, \frac{J_p^2}{2(\delta_{12}\pm \Delta)(\delta_{10}\pm \Delta)}, -\frac{J_p^2}{2\delta_{02}(\delta_{10} \pm \Delta)} \}; \nonumber \\ \langle \sigma^x_0 \tau^x_1 \rangle &=& \{ \frac{J_p}{2(\delta_{10} \pm \Delta)}, -\frac{J_p}{2(\delta_{10} \pm \Delta)}, \frac{J_p}{2(\delta_{10} \pm \Delta)}, -\frac{J_p}{2(\delta_{10} \pm \Delta)} \} \nonumber \\ \langle \sigma^x_0 \tau^x_2 \rangle &=& \{ 0, 0, \frac{J_p^2}{4\delta_{02}(\delta_{12} \pm \Delta)}, 2 \times \frac{J_p^2}{4(\delta_{10}\pm \Delta)}\left[\frac{1}{\delta_{12} \mp \Delta} - \frac{1}{\delta_{02}} \right] \}. \end{aligned}$$ Note that it is vital to include the $J_p^3$ contribution in the first line lest the zeros dominate whilst taking the typical values of the above fields: this is because the denominators of these $J_p^3$ terms can skew these otherwise parametrically smaller fields to be comparable to the other terms proportional to $J_p$. (This may be checked for a given small $L$ system comparing numerics and the above expressions.)\ Taking $J_{xy} = J^z = 1/W$, with $W \gg 1$, we see that the typical value of $-(\log{\frac{\langle \sigma^x_0 \sigma^x_2 \rangle}{\langle \sigma^x_0 \sigma^x_1 \rangle}})^{-1} \approx 0.54, 0.73$ and $-(\log{\frac{\langle \sigma^x_0 \tau^x_2 \rangle}{\langle \sigma^x_0 \tau^x_1 \rangle}})^{-1} \approx 0.33, 0.41$ from these perturbative treatments for $W=25, 15$ respectively; these perturbation theory predictions are plotted in Fig. \[fig:Jz-vs-xX\](a) and shown as dot-dashed thick brown lines, both for the means and variances. while the analytical predictions for typical $\sigma^x \sigma^x$ correlators are in reasonable agreement with the numerical values for $W \gg 1$, the agreement for the $\sigma^x \tau^x$ correlators is much less impressive. This is presumably due to nonperturbative effects in the latter. Moreover we see that the initial upturn (strong-finite size effect) in the variance for $\sigma^x \sigma^x$ correlators at small $r$ is nicely captured by the perturbation theory, as is the absence of this upturn in the $\sigma^x \tau^x$ correlators; these are indicated by the thick brown triangles joined by the thick dot-dashed lines. Therefore after averaging this inverse localisation lengths we find generically that $$\label{eq:xxVsxX} \xi^{-1}_{\textrm{xx}} < \xi^{-1}_{\textrm{xX}},$$ thereby making the $\langle \sigma^x_0 \tau^x_2 \rangle$ correlator decay faster than $\langle \sigma^x_0 \sigma^x_2 \rangle$. We have seen this to be true whether averaged all samples (Fig. \[fig:pxi\]) or in single samples (Fig. \[fig:SingleSample\]).\ Moreover we see from perturbation theory that as $\Delta \rightarrow 0$, (i) the two localisation lengths for $\sigma^x \sigma^x$ and $\sigma^x \tau^x$ will be different; (ii) $\xi \sim \Delta^{\kappa}$, with $\kappa \approx 0.25 < 1$. The linear dependence on $\Delta$ (i.e. $\kappa = 1$) is satisfied only when $\Delta/W \approx 0.1$. Variance measures ----------------- Unlike the medians or means, the variances over states and samples do not commute i.e. it depends whether we perform the variance measurements over all states and samples, or first mean over states and then variance over samples (intersample variance), or first variance over states and then mean over samples (intrasample variance). Therefore in principle from the variances alone there will be three length scales $\xi_{\textrm{var.}}, \xi_{\textrm{var. mean}}, \xi_{\textrm{mean var.}}$, which correspond to the above three ways respectively. In the main text we presented only the first of the three. However it is clear that the full variance over samples and states is simply the sum of the other two.\ Moreover we see that it is not always the case that intersample fluctuations dominate; this seems only true for $\langle \sigma^x_0 \sigma^x_r \rangle$ and $\langle \sigma^x_0 \tau^x_r \rangle$ correlators but not for $J^z$ couplings. While the latter two do not have strong transients at small $r$, the p-bit correlator does. This is explained by the perturbative structure of fields, already seen in Eq. where there is a larger proportion of exact zero values in the correlators at a given order. Indeed we find that the results from perturbation theory at small distances captures the numerical results for variances too, qualitatively and quantitatively, discriminating between intersample and intrasample fluctuations.\ Finally in all three measures of variances, we find that a linearly growing variance leads to a broad distribution of these couplings as well, whether in the intrasample or intersample measure i.e. the presence of broad distributions is a persistent effect. [^1]: WWF can also be considered as an adiabatic flow between infinite and finite disorder where the invariant object is the entire set of bitstring labels of all many-body eigenstates – these are conserved by definition under the flow and allow to interpret the results even outside the localized phase. Once defined, WWF continues to work across the phase diagram and allows for a seamless discussion of observables and their distributions – that is very convenient, but may be confusing or ambiguous, e.g. if the L-bits are strongly smeared. [^2]: A technical note is in order here: while taking the limit $\kappa \rightarrow \infty$ is not practically possible, we implemenent the flow using an adaptive 4,5 Dormand-Pince algorithm until the error estimate of $H(\kappa) \rightarrow H(\kappa + d\kappa)$ is below some threshold $-\log_{10} \epsilon = 3\sim 6$, concomitantly chopping away off-diagonal elements that are smaller than $n\epsilon$, where $n = \mathcal{O}(1) > 1$. The latter aspect is required in order to speed up the convergence to the effective Hamiltonian; $n<1$, on the other hand, will result in very minimal chopping away, and hence much slower convergence. As long as $n = \mathcal{O}(1)$, the chopping is not too aggressive. Due to these approximations, the eigenvalues and eigenvectors are only approximate; they are matched to machine-precision eigenvalues and eigenvectors obtained from exact diagonalisation (ED) using the Hungarian matching algorithm [@Kuhn] i.e. bipartite graph matching for maximum sum of weights in the overlap of $U U_{ED}^{T}$. [^3]: D. A. Huse, private communication

--- abstract: 'We explore different (and complementary) views of spinors and their exotic counterparts, linking the very existence of the later to the presence of black holes. Moreover, we investigate the effects of the exotic term in the Hawking radiation emission rate, as well as its extremes, for asymptotically flat black holes solutions of general relativity. We show that, under certain circumstances, the emission rate extreme condition fixates an equation from which the exotic term could be inferred.' address: - | Departamento de Física e Química, Universidade Estadual Paulista (UNESP)\ Av. Ariberto Pereira da Cunha, 333\ CEP 12516-410\ Guaratinguetá, SP\ Brazil - | Departamento de Física e Química, Universidade Estadual Paulista (UNESP)\ Av. Ariberto Pereira da Cunha, 333\ CEP 12516-410\ Guaratinguetá, SP\ Brazil - | Departamento de Física e Química, Universidade Estadual Paulista (UNESP)\ Av. Ariberto Pereira da Cunha, 333\ CEP 12516-410\ Guaratinguetá, SP\ Brazil author: - 'Beghetto, D.' - 'Cavalcanti, R. T.' - 'Hoff da Silva, J. M.' date: 'July 1, 2018' title: Exotic Spinorial Structure and Black Holes in General Relativity --- Introduction ============ It is almost impossible to give the right emphasis on the importance of spinor fields in the construction of our understanding of the physical phenomena. Spinor fields constitute an irremediable essential tool without which high energy physics cannot be described. By the same token, it is equally hard to envisage the usefulness of the comprehensive mathematical construction behind the very concept of spinors. In fact, the general algebraic structure supporting the spinor concept has abundant applications in physics, from condensed matter to cosmology, going through quantum field theory. In spite of that, the construction of a solid bridge between a well established concept at the algebraic level and its counterpart in physics is highly nontrivial. In other words, the precise use of spinors in physics is hardly reached when the departure point is the algebraic definition of the spinorial quantity in question. Obviously, the point raised in the previous paragraph is not relevant within a more pragmatic point of view. In some cases, it is indeed irrelevant whether the spinor is understood as a mathematical object belonging to a section of a specific fiber bundle or as a four component “column”, as far as it helps to give an observable agreeing with experience in a good level. However, a pragmatic approach, while quite satisfactory and acceptable in some cases, is not always free of the brevity narrowness. In several cases the formal emphasis lead to possibilities other than the highlighted by means of the usual thought. In the formal mathematical treatment concerning spinors, two seminal approaches are of particular interest in physics: the concept of spinors as “pre-geometric” quantities, in a manner of speaking, pointed out by Cartan [@ECARTAN], in which the spacetime points themselves can be understood as generated by spinors components (a program which has been generalized eminently by Penrose [@Pe84]) and, on the other hand, the profound algebraic view of spin structures in a given manifold, codified in the Čech cohomology class [@naka]. Apart from its mathematical rigidity, both approaches may be connected in a tentative intuitive fashion providing an interesting picture to explore exotic spinors in a physical ground. We shall explore this connection in this paper, further developing this point of view and applying the resulting construction in the investigation of a physical system which we believe may serve as an interesting laboratory to explore questions about nontrivial topology, its changes, and impact in high energy physics. The idea is to explore at a physical level a consequence of linking the (time) variation of a specific additional term appearing in the connection (hence, the dynamics) of exotic fermions, and the emission of radiation by black hole through the Hawking process. Even further, by understanding the very existence of black holes as inductors of nontrivial topologies, we investigate how the extremes of emission rate of exotic spinors are influenced by nearly adiabatic changes in the black hole area. The present paper is organized as follows: Section II is devoted to the study of exotic spinors, not only reviewing its formal aspects but also extending an intuitive approach to this subject. In Section III we investigate a physical output of a time-variable topology performed by a slowly varying black hole radius, analyzing the Hawking radiation emission rate in this case. In order to clarify the different physical effects of gauge and exoticness, in the Appendix \[append\] we show that the exotic term induces modifications on the dispersion relation incompatible with the ones due to the gauge field. In the final Section we conclude and discuss the results. Exotic spinors ============== In the case of a nontrivial topology for the base manifold, there is not one but many spinorial structures different from the usual. Spinors belonging to these additional structures have, as a net effect of the nontrivial topology, a slightly different dynamics, as the derivative operator encodes a new term coming from such non triviality. We shall explore two approaches to this key correction. The first one is basically a short review on the formal aspects of exotic spinors[^1]. The idea is to pinpoint the strictly necessary formalism to evince the existence and crucial effects related to the exoticity. Subsequently, we also revisit an intuitive approach to exotic spinors, taking advantage of the Cartan’s systematic approach to spinorial quantities, which will be further explored here. Essential Formal aspects, a very brief review --------------------------------------------- Not only one spin structure arise from different and inequivalent patching of local coverings for a given manifold $M$ [@Hawking:1977ab; @Geroch:1968zm; @Geroch:1970uv] . These possible nontrivial spin structures are labeled by elements of the well known group of homomorphisms of $\pi_1(M)$ into $\mathbb{Z}_2$, i. e., the first cohomology group $H^1(M,\mathbb{Z}_2)$ [@trinta; @trintaseis; @Asselmeyer]. Although the mathematical perspective is intricate, some basilar aspects may be pinpointed. For instance, when dealing with a simply connected base manifold, the fundamental group in indeed trivial and, hence, there is only one spin structure. In multiple connected manifolds, however, the situation is quite peculiar. Let us start with an appropriate definition of a spin structure on $M$. It is composed by a principal fiber bundle, say $\pi_s$, from the orthonormal coframe bundle $P_{Spin_{1,3}}(M)$ to $M$ along with the two-fold covering $s: P_{Spin_{1,3}}\rightarrow P_{SO_{1,3}}$ in such a way that $\pi_s=\pi\circ s$ for $\pi: P_{SO_{1,3}}\rightarrow M$. The assertion of the previous paragraph may be read in the following terms: if $H^1(M,\mathbb{Z}_2)\neq 0$ (notation for non-triviality), then different spin structures $(\tilde{P}_{Spin_{1,3}},\tilde{s})$ are in order. Noticed that a non-triviality on $M$ leads to an explicit dependence on different base manifold coverings for the definition of $\pi$, $\pi_s$. This dependence gives rise, in a manner of speaking, to its tilde counterparts. Now, the orthonormal coframe bundle act as a substrate for the covariant spinor bundle of $P_{Spin_{1,3}}(M)\times_\rho \mathbb{C}^4$ of which a classical spinor $\psi$ is a section. Therefore, it is indeed expected the existence of $\tilde{\psi}\in \sec \tilde{P}_{Spin_{1,3}}(M)\times_\rho \mathbb{C}^4$. In the notation here adopted, $\rho$ stands for the Weyl representation space $(1/2,0)\oplus(0,1/2)$ as a whole (or, eventually, to a single part of it). The equivalence between two given spin structures $(P_{Spin_{1,3}},s)$ and $(\tilde{P}_{Spin_{1,3}},\tilde{s})$ is performed by means of a mapping $q:P_{Spin_{1,3}}\rightarrow \tilde{P}_{Spin_{1,3}}$, such that $q=s\circ \tilde{s}$. Take now a point $x\in M$ and let $U_i$ and $U_j$ be two open sets on $M$ whose intersection encompass $x$. From $U_i\cap U_j\subset M$ it is possible to define two mappings $m_{ij}$ and $\tilde{m}_{ij}$ such that $$\begin{aligned} U_i\cap U_j\subset M \underbrace{\longrightarrow}_{m_{ij}} Spin_{1,3} \hspace{2cm} U_i\cap U_j\subset M \underbrace{\longrightarrow}_{\tilde{m}_{ij}} \tilde{Spin}_{1,3}, \label{1}\end{aligned}$$ with a common transition, say $\sigma$, to $SO_{1,3}$. An additional mapping can be defined [@map] such that $m_{ij}(x)=\tilde{m}_{ij}(x)c_{ij}$ with the requirement that $c_{ij}: U_i\cap U_j\rightarrow \mathbb{Z}_2\hookrightarrow Spin_{1,3}$. In order to define a meaningful covariant derivative, it is generally assumed the existence of unimodular functions $\xi_i:U_i \subset M \rightarrow \mathbb{C}$ such that $\xi_i(x)\in U(1)$ [@trinta; @trintaseis]. In fact, in the absence of torsion, the existence of such functions is guaranteed [@trinta; @novo]. As we have mentioned, a spinor field $\psi$ is an element of $\sec P_{Spin_{1,3}}(M)\times_\rho \mathbb{C}^4$. However, different patching of local coverings resulting from nontrivial topology are labeled by different elements of $H^1(M, \mathbb{Z}_2)$. To each element of $H^1(M, \mathbb{Z}_2)$ it is associated a connection $\nabla$, giving rise to a one-to-one correspondence between inequivalent spin structures (and consequently inequivalent spinor fields) and elements of the first cohomology group. Having said that, it is a matter of working through the definition of a bundle mapping $$\begin{aligned} f:\tilde{P}_{Spin_{1,3}}\times_\rho \mathbb{C}^4 &\rightarrow& P_{Spin_{1,3}}\times_\rho \mathbb{C}^4\nonumber\\ \tilde{\psi}&\mapsto &q(\tilde{\psi})=\psi,\label{uia}\end{aligned}$$ in such a way that $$\begin{aligned} \tilde{\nabla}_X f(\tilde{\psi})=f(\nabla_X\tilde{\psi})+\frac{1}{2}(X, \xi^{-1}d\xi)f(\tilde{\psi}),\label{eia}\end{aligned}$$ for all $\psi \in P_{Spin_{1,3}}\times_\rho \mathbb{C}^4$ for all vectorial field $X$ consistently defined over $M$. In Eq. (\[eia\]) the brackets in the second term stands for an specific contraction not important to our general argumentation. The relevant point is that the unimodular field engenders an additional term to the covariant derivative, coming genuinely from the nontrivial topology. Intuitive aspects ----------------- We intend here to give a complete account of an intuitive argument firstly presented in Ref.[^2] [@uni], along with a proper generalization. The underlying idea is to argue on the derivative term correction coming from the nontrivial topology. We would like to revisit and extend this intuitive point of view in this very section. Starting from the Cartan’s spinorial approach we motivate the insertion of the second term in (\[eia\]). Also, at the end of this section, bearing in mind that a black hole itself may be the inductor of a nontrivial topology, we make a link with the idea we are going to explore in the next section. A spacetime vector ${\bf v}\in\mathbb{R}^{1,3}$ can be expressed by ${\bf v} = x^0 {\mathbf{e}}_0 + x^1 {\mathbf{e}}_1 + x^2 {\mathbf{e}}_2 + x^3 {\mathbf{e}}_3,$ where ($x^0, x^1, x^2, x^3$) denote components of ${\bf v}$ with respect to an orthonormal basis {${\mathbf{e}}_0, {\mathbf{e}}_1, {\mathbf{e}}_2, {\mathbf{e}}_3$}. Null vectors are isotropic vectors, and satisfy $(x^0)^2 - (x^1)^2 - (x^2)^2 - (x^3)^2 = 0$. They present null directions in $\mathbb{R}^{1,3}$ with respect to the origin $\mathcal{O}$ of an arbitrary frame in $\mathbb{R}^{1,3}$. The space of null directions that are future \[past\] pointed[^3] are denoted by $\mathcal{S}^+$ \[$\mathcal{S}^-$\], and represented by the intersections $\mathbb{E}^+$ \[$\mathbb{E}^-$\] of the future \[past\] light cones with the hyperplanes $x^0 = 1$ \[$x^0 = -1$\]. The space $\mathbb{E}^\pm$ is a sphere with equation $x^2 + y^2 + z^2 = 1,$ where ($x, y, z$) are coordinates in $\mathbb{E}^\pm$ [@Pe84]. Despite this flat spacetime approach, the generalization for curved spacetimes shall be direct by means of the tangent bundle. Returning to our main exposition, more generally, the direction of any null vector ${\bf v}\in\mathbb{R}^{1,3}$, unless such vector is an element of the plane defined by the equation $x^0 = 0$, can be represented by two points. Such description results from the intersection of ${\bf v}$ and the hyperplanes $x^0 = \pm 1$. The future-pointed ${\bf v}$ is thus represented by ($x^1/\|x^0\|, x^2/\|x^0\|, x^3/\|x^0\|$). The inner points of $\mathbb{E}^+$ ($\mathbb{E}^-$) represent the set of future-pointed (past-pointed) light-like directions. By considering $\mathbb{E}^+$, by performing a stereographic projection on the Argand-Gauss plane, we obtain a representation of the union between the set of complex numbers and the point at the infinity, that corresponds to the north pole of $\mathbb{E}^+$. By defining the complex number $$\label{be1}\beta = \frac{x + iy}{1-z}\,,$$ it yields $\beta \overline\beta = \frac{x^2 + y^2}{(1-z)^2}$, and consequently $$\label{xyz} x = \frac{\beta + \overline\beta }{\beta \overline\beta + 1},\quad y = \frac{\beta - \overline\beta }{i(\beta \overline\beta + 1)},\quad z=\frac{\beta \overline\beta - 1}{\beta \overline\beta + 1}.$$ The correspondence between points of $\mathbb{E}^+$ and the Argand-Gauss plane is injective if the point $\beta \sim \infty$ is added to the complex plane, making it to correspond to the north pole with components $(1,0,0,1)$. However, to avoid this point, it is more convenient to associate a point of $\mathbb{E}^+$ not to a complex number $\beta $, but to a pair of complex numbers [^4] $(\xi, \eta)$, where $$\label{be2}\beta = \xi/\eta.$$ The pairs $(\xi, \eta)$ and $(\lambda\xi, \lambda\eta), \lambda \in {\mathbb{C}}$, represent the same point in $\mathbb{E}^+$. Such components are called [*projective coordinates*]{}. These projective coordinates resulting from the stereographic projection of the Riemann sphere on the complex plane, collected in a pair, gives rise to what we call by spinor. It is important to remark that the used Riemann sphere is a quite special one: it results from the intersection of a time constant plane with the (future oriented here) light cone. The point $\beta = \xi/\eta \sim \infty$ corresponds to the point of coordinates ${\xi\choose\eta} = {1\choose 0}$. Eqs. (\[xyz\]) can be expressed as $$\begin{aligned} \label{xyw} x = \frac{\xi\overline\eta + \eta\overline\xi}{\xi\overline\xi + \eta\overline\eta},\quad y = \frac{\xi\overline\eta - \eta\overline\xi}{i(\xi\overline\xi + \eta\overline\eta)},\quad z = \frac{\xi\overline\xi - \eta\overline\eta}{\xi\overline\xi + \eta\overline\eta},\end{aligned}$$ explaining the claim of spinors pre-geometric quantities, or the “square-root of a point”. The point $P=(1,x,y,z)$ is an arbitrary point of the light-cone transversal section with constant time and represents a null future-pointed direction, that can be represented by any point of the line ${\mathcal{O}} P$. In particular, if a point $R$ is taken in the line ${\mathcal{O}} P$ by multiplying $P$ by the factor $(\xi\overline\xi + \eta\overline\eta)/\sqrt{2}$, then $R$ has coordinates $$\begin{aligned} \label{xt}x^1 &=& \frac{1}{\sqrt{2}}(\xi\overline\eta + \eta\overline\xi), \quad x^2 = \frac{1}{i\sqrt{2}}(\xi\overline\eta - \eta\overline\xi),\nonumber\\ x^3 &=& \frac{1}{\sqrt{2}}(\xi\overline\xi - \eta\overline\eta), \quad x^0 = \frac{1}{\sqrt{2}}(\xi\overline\xi + \eta\overline\eta).\end{aligned}$$ Notice that contrary to the point $P$, the point $R$ is not invariant under $(\xi, \eta) \mapsto (r\xi, r\eta), r \in {\mathbb{R}}$, although it is independent of phases $(\xi, \eta) \mapsto (e^{i\theta}\xi, e^{i\theta}\eta), \theta \in {\mathbb{R}}$. Consider now the following complex linear transformation $$\begin{array}{l} \label{tlc} \xi \mapsto \tilde\xi = \alpha\xi + \mu\eta,\\ \eta \mapsto \tilde\eta = \gamma\xi + \delta\eta, \end{array}$$ where $\alpha, \mu, \gamma, \delta \in {\mathbb{C}}$ satisfy $\alpha\delta - \mu\gamma \neq 0$, in order to such transformation being invertible. It can be rewritten as $$\label{mobius} \beta \mapsto f(\beta ) = \frac{\alpha\beta + \mu}{\gamma\beta + \delta},$$ and named a Möbius transformation, from the set ${\mathbb{C}}\backslash\{-\delta/\gamma\}$ to ${\mathbb{C}}\backslash \{\alpha/\gamma\}$. Moreover, if $f(-{\delta/\gamma}) \sim \infty$ and $f(\infty) \sim {\alpha/\gamma}$, then $f$ is an injective function from the complex plane, compactified by the point at the infinity, denoted by (${\mathbb{C}}\cup \{\infty\}$). Hence the space of light-like vectors on Minkowski spacetime is naturally a Riemann sphere. The restricted Lorentz group ${{\mathcal}L}^+$ is, on the other hand, the automorphism group of the Riemann sphere. Eq. (\[tlc\]) with the condition $$\alpha\delta - \mu \gamma = 1$$ are called spinor transformations, where $\beta = $ $ \xi/\eta $ is related to the null vectors by Eqs. (\[xt\]), implying that $$\beta = \frac{x^1 + ix^2}{x^0-x^3} = \frac{x^0-x^3}{x^1-ix^2}.$$ The spinor matrix ${\bf{A}} \in {\rm SL}(2, {\mathbb{C}})$ is defined as $$\label{mspin} {\bf{A}} = \left(\begin{array}{cc} \alpha & \mu\\ \gamma & \delta \end{array} \right),\quad {\rm det}\;{\bf{A}} = 1.$$ Eqs. (\[tlc\]), with respect to $\bf{A}$, read $$\left( \begin{array}{c} \tilde\xi\\ \tilde\eta \end{array} \right) = \bf{A} \left( \begin{array}{c} \xi\\ \eta \end{array} \right)\,.$$ The spinor matrices $\{\pm\bf{A}\}$ induce the same transformation of $\beta = \xi/\eta$. Eq.(\[xt\]) yields $$\label{vesp}\frac{1}{\sqrt{2}}\left( \begin{array}{cc} x^0 + x^3 & x^1 + ix^2\\ x^1 - ix^2 & x^0 - x^3 \end{array} \right) = \left( \begin{array}{cc} \xi\overline\xi\ & \xi\overline\eta\\ \eta\overline\xi & \eta\overline\eta \end{array} \right) = \left( \begin{array}{c} \xi\\ \eta \end{array} \right) (\overline\xi \quad\overline\eta).$$ Hence, up to a factor $1/\sqrt{2}$, it follows that $$\begin{aligned} \left( \begin{array}{cc} x^0 \!+\! x^3 & x^1 \!+\! ix^2\\ x^1 \!-\! ix^2 & x^0 \!-\! x^3 \end{array} \right) &\mapsto& \left( \begin{array}{cc} \tilde{x^0} \!+\! \tilde{x^3} & \tilde{x^1}\! +\! i\tilde{x^2}\\ \tilde{x^1} \!-\! i\tilde{x^2} & \tilde{x^0} \!-\! \tilde{x^3} \end{array} \right)\nonumber\\ &=& \bf{A}\left( \begin{array}{cc} x^0 \!+\! x^3 & x^1 \!+\! ix^2\\ x^1 \!-\! ix^2 & x^0 \!-\! x^3 \end{array} \right)\bf{A^\dagger}. \end{aligned}$$ Such transformation acting on the point ${\bf v} = (x^0,x^1,x^2,x^3)$ is real and preserves the light cone structure $(x^0)^2 - (x^1)^2 - (x^2)^2 - (x^3)^2 = 0$. Thus the above relation defines a restricted Lorentz transformation. In fact, the group SL(2,${\mathbb{C}}$) is the two-fold covering of the restricted Lorentz group SO$_+(1,3) \simeq \mathcal{L}^+$. As mentioned before, all the previous formalization can be performed in a curved space tangent bundle. Some adaptation shall be expected, however, if the base manifold is endowed with a nontrivial topology, for instance, engendered by a black-hole. Indeed, there is no meaning in associate geometrical points to the interior of a given black-hole, evincing thus its existence of nontrivial topology. Nevertheless this means, by its turn, that the tangent bundle itself is not simply connected. Hence, the appearance of different spinors, resulting from different patches in the tangent bundle, are in order. Moreover the spinor dynamics, as partially dictated by the connection, shall also be affected. The net effect of the non-trivial topology is labelled by an integer number reflecting, as explained in the previous subsection, the non-triviality of the first cohomology group. This label is recovered as a macroscopic effect by an integration upon a closed curve. From these considerations, the term appearing in the new derivative may be recast into the form $$\begin{aligned} \gamma^\mu\tilde{\nabla}_\mu=\gamma^\mu\nabla_\mu+\frac{1}{2\pi i }\xi^{-1}d\xi,\label{ulti}\end{aligned}$$ where spin and geometric connections are embedded into $\nabla$. Since it must be fulfilled the following requirement $$\begin{aligned} \frac{1}{2\pi i}\oint \xi^{-1}d\xi \in \mathbb{Z}, \label{ma}\end{aligned}$$ we have $\xi=e^{in\theta(x)} \in U(1)$, with $n\in \mathbb{Z}$, in agreement with last subsection. Finally, upon a simple rescaling of the field $\theta(x)$ and bearing in mind that $\gamma^\mu$ is an adequate basis for the orthonormal frame we have $$d\theta=\gamma^\mu\partial_\mu \theta,\label{qua}$$ and the dynamic equation reads $$\begin{aligned} (i\gamma^{\mu}\nabla_\mu+i\gamma^\mu\partial_\mu\theta-m)\tilde{\psi}=0.\label{se}\end{aligned}$$ In the next section we shall pursue an interesting effect connecting the non triviality of the spacetime topology and black holes, in adequate conditions, regarding the Hawking radiation emission rate. Hawking radiation vs exoticity ============================== Hawking radiation is one of the main achievements of semi-classical gravity. In fact, it is a widely accepted feature of gravity beyond the Einstein’s theory, pointing towards a theory encompassing gravity and quantum mechanics. An interesting and not very explored aspect of the Hawking’s result regards its relationship with the exotic structure introduced in the previous section. As we pointed out, the very existence of a black hole makes the topology of the space-time nontrivial, implying the possible existence of exotic spinors. As mentioned in [@uni], the exotic term does not change the Hawking temperature, however it affects the emission rate from where the temperature is derived. This fact, which is not as spread as it should be, can change the expected life time of black holes. In this section we are going to investigate the possible extremes of the emission rate $\Gamma$ and how it is affected by the exotic term $\theta$. For this reason we must consider the emission of spinors, as their dynamics is affected by the presence of $\theta$. Due to its generalness and compatibility with the Einstein’s theory, we have chosen to adopt the emission rate of the Kerr-Newman solution. The Hawking radiation emission rate, also interpreted as the tunneling probability [@Parikh:1999mf; @Angheben:2005rm; @Arzano:2005rs; @Jiang:2005ba], is quite straightforwardly derived by using the Hamilton-Jacobi approach to the tunneling method. Such method is based upon the particle description of Hawking radiation, under the assumption that the emitted particle action does satisfy the relativistic Hamilton-Jacobi equation. It relies on allowing particles to travel along classically forbidden trajectories, from just behind the horizon onward to infinity. A comprehensive review of the tunneling method can be found in [@Vanzo:2011wq]. The Kerr-Newman solution of the Einstein’s field equations, in Boyer-Lindquist coordinates $( t,r,\vartheta,\phi)$, reads $$\begin{aligned} \nonumber ds^2=&-\left(\frac{\Delta-a^2\sin^2\vartheta}{\Sigma}\right)dt^2-\frac{2(r^2+a^2-\Delta)a\sin^2\vartheta}{\Sigma}dtd\phi+\frac{\Sigma}{\Delta}dr^2+\Sigma d\vartheta^2+\\ &+\frac{(r^2+a^2)^2-a^2\Delta\sin^2\vartheta}{\Sigma}\sin^2\vartheta d\phi^2,\end{aligned}$$ where $a=\frac{J}{M}$, $\Sigma=r^2+a^2\cos^2\theta$ and $\Delta=r^2+a^2+{Q^2}-{2Mr}$. We are going to depart from its emission rate of fermionic particles [@Vanzo:2011wq; @uni], taking into account the exotic term $$\Gamma=\exp\left[-4\pi\left(\frac{r_+^2+a^2}{r_+-r_-}\right)(\omega-j\Omega-q\Phi+\dot{\theta})\right].$$ Here $r_+$ $[r_-]$ denotes the outer \[inner\] horizon, given by $$r_\pm=M\pm\sqrt{M^2-Q^2-a^2}.$$ A preliminary analysis could start from assuming that the horizon radii, as well as the emitted particle and black hole parameters, do not change appreciably in time. In fact, bearing in mind astrophysical black holes, this fact seems quite reasonable. In such a simplified case, however, the condition for the existence of extremes of $\Gamma$ is simply $\ddot{\theta}=0$, that is, the exotic term should be linear. The situation becomes more interesting by relaxing the assumption of constant radii. From now we are going to distinguish two cases, characterized by the existence or not of an external electromagnetic potential $A_\mu$. The tunneling/emission rate will be denoted by $\Gamma_{DN}$ for a vanishing external potential and $\Gamma_{DE}$ for a non vanishing external potential. For Dirac exotic spinors, that is, solutions to the Dirac equation with the additional exotic term, and with no external interaction, the tunneling probability is given, as earlier, by [@Vanzo:2011wq; @uni] $$\label{gammaDN} \Gamma_{DN} = \exp{\left[ -4\pi \left( \frac{r_{+}^2 + a^2}{r_{+} - r_{-}}\right) (\omega - j\Omega - q \Phi + \dot{\theta}) \right]},$$ and its first derivative by $$\begin{aligned} \label{derivada1gammaDN} \dot{\Gamma}_{DN} = \frac{4 \pi \Gamma_{DN}}{(r_- - r_+)^2} & \left\{ (\omega - j\Omega - q \Phi + \dot{\theta}) \left[(a^2 + r_+^2) (-\dot{r}_- + \dot{r}_+)+\right. \right. \nonumber \\& \left. \left. + 2 (r_- - r_+) (a \dot{a} + r_+ \dot{r}_+) \right] + (r_- - r_+) (a^2 + r_+^2) \ddot{\theta} \right\}.\end{aligned}$$ Here we are taking $\omega$, $j\Omega$ and $q\Phi$ as constant in time and setting $\dot{r}^2, \dot{\theta}^2, \ddot{r} \rightarrow 0$. The later meaning adiabatic variation, i. e., despite the possible extreme emission rate, the radii varies slowly. The extreme condition $\dot{\Gamma}_{DN} = 0$ leads to the equation $$\label{asterisco} \ddot{\theta} + (\omega - j\Omega - q \Phi + \dot{\theta}) \frac{d}{dt} \left[ \ln{ \left( \frac{r_+^2 + a^2}{r_+ - r_-} \right) } \right] = 0.$$ Note that, in spite of being a particular case, it gives us a second order differential equation for the unknown exotic term $\theta$. Furthermore, we emphasize that this case describes the emission rate of exotic dark spinors (Elko) [@uni; @Cavalcanti:2015nna; @exo], a prime candidate to describe dark matter, as it interacts only with gravity and the Higgs field (see [@Ahluwalia:2016rwl] and references therein for a general discussion on Elko dark spinors). For the case of exotic Dirac spinors interacting with an $A_\mu$ field, with $\slashed A \equiv \gamma_\mu A^\mu$, one can express the tunnelling probability $\Gamma_{DE}$ for the interacting exotic Dirac spinor as $$\label{gammaDE} \Gamma_{DE} = \exp{\left[ -4\pi \left( \frac{r_{+}^2 + a^2}{r_{+} - r_{-}}\right) (\omega - j\Omega - q \Phi + \dot{\theta} + \slashed A) \right]}.$$ We point out that, apart from its very nature, the exotic term on the equation above can not be incorporated by the gauge field. As shown in the Appendix \[append\], the exotic term induces modifications on the dispersion relation incompatible with the ones due to the gauge field. From Eqs. (\[gammaDN\]) and (\[gammaDE\]) we can write $$\Gamma_{DE} = \Gamma_{DN} \mathcal{R},$$ where $ \mathcal{R} \equiv \exp{\left[ -4\pi \left( \frac{r_{+}^2 + a^2}{r_{+} - r_{-}} \right) \slashed A \right]}$, following $$\label{derivada1gammaDE} \dot{\Gamma}_{DE} = \dot{\Gamma}_{DN} \mathcal{R} + \Gamma_{DN} \dot{\mathcal{R}}.$$ Here the extreme condition gives two possibilities: $\dot{\Gamma}_{DN} = 0$ or $\dot{\Gamma}_{DN} \neq 0$. The first one results in $$\label{derivadaRzero} \dot{\slashed A} + \slashed A \frac{d}{dt} \left[ \ln{ \left( \frac{r_+^2 + a^2}{r_+ - r_-} \right) } \right] = 0.$$ Comparing Equations (\[asterisco\]) and (\[derivadaRzero\]) one finds $$\label{condicaoDNzero} \slashed A - \dot{\theta} = \omega - j\Omega - q \Phi.$$ On the other hand, the condition $\dot{\Gamma}_{DN} \neq 0$ leads to $$\label{condicaoDNnaozero} - \slashed A - \dot{\theta} = \omega - j\Omega - q \Phi.$$ Again the extreme condition fixates a differential equation for the exotic term, allowing the time variation of $\theta$ being completely determined by the external field. Our next step is to investigate the second derivatives of the tunnelling rates. From Equation (\[derivada1gammaDN\]) follows $$\begin{aligned} \ddot{\Gamma}_{DN} = \frac{4\pi \Gamma_{DN}}{(r_- - r_+)^4} & \left\{ \left[ 4\pi [\mathcal{X} (\alpha + 2\beta) + \sigma \ddot{\theta}] - 2(\dot{r}_- - \dot{r}_+)(r_- - r_+) \right] \right. \nonumber \\ & \times \left. \left[ \mathcal{X}(\alpha + 2\beta) + \sigma \ddot{\theta} \right] + + (r_- - r_+)^2 \left[ 2\ddot{\theta} (\alpha + 2\beta) + \sigma \dot{\ddot{\theta}} \right] \right\}, \end{aligned}$$ with $$\begin{aligned} \alpha &\equiv (a^2 + r_+^2)(-\dot{r}_- + \dot{r}_+)\\ \beta &\equiv (r_- - r_+)(a\dot{a} + r_+ \dot{r}_+)\\ \sigma &\equiv (r_- - r_+)(a^2 + r_+^2)\\ \mathcal{X} &\equiv (\omega - j\Omega - q \Phi + \dot{\theta}).\end{aligned}$$ Starting from the case $\dot{\Gamma}_{DN} = 0 = \dot{\Gamma}_{DE}$ we define $$\begin{aligned} \ddot{\tilde{\Gamma}}_{DE} \equiv & \ddot{\Gamma}_{DE} \Big\rvert_{\dot{\Gamma}_{DN}, \dot{\Gamma}_{DE} = 0},\\ \ddot{\tilde{\Gamma}}_{DN} \equiv & \ddot{\Gamma}_{DN} \Big\rvert_{\dot{\Gamma}_{DN}, \dot{\Gamma}_{DE} = 0}, \end{aligned}$$ and after some simple yet lengthy calculations, noticing that $\dot{\ddot{\theta}} = \ddot{\mathcal{X}}$ and $\ddot{\theta} = \dot{\mathcal{X}}$, on finds $$\begin{aligned} \label{derivada2gammaDEDNdoisextremos} \ddot{\tilde{\Gamma}}_{DE} = & \frac{4\pi \Gamma_{DE}}{(r_- - r_+)^2} \left[ f(\mathcal{X}) + f(\slashed A) \right],\\ \ddot{\tilde{\Gamma}}_{DN} = & \frac{4\pi \Gamma_{DN}}{(r_- - r_+)^2} f(\mathcal{X}), \end{aligned}$$ where $f(y) \equiv 2\dot{y} (\alpha + 2\beta) + \sigma \ddot{y}.$ The signs of $\ddot{\tilde{\Gamma}}_{DE}$ and $\ddot{\tilde{\Gamma}}_{DN}$ are then strongly related to the signs of $f(\mathcal{X})$ and $f(\slashed A)$. Note that in this case, by Eq. (\[condicaoDNzero\]), one has $\mathcal{X} = \slashed A$. Therefore, the sign of $f(\mathcal{X}) = f(\slashed A)$ defines identical signs for $\ddot{\tilde{\Gamma}}_{DE}$ and $\ddot{\tilde{\Gamma}}_{DN}$. This means that, for our particular Kerr-Newman black hole situation, the presence of an external field $A_\mu$ does not change the general behaviour of the emission extremes of Hawking radiation for an exotic particle, remaining it maximum \[minimum\] if it was already maximum \[minimum\]. In order to establish rather the emission rate is maximum or minimum for the $\dot{\Gamma}_{DN} = 0 = \dot{\Gamma}_{DE}$ case we must know the signs of the $f(y)$ coefficients. Even though $\alpha < 0$ and $\beta > 0$, there is no general sign for $\alpha + 2\beta$. In fact, with $z \equiv M^2 - (\frac{J}{M})^2 - q^2 > 0$ and $\frac{dM}{dt} < 0$, we can write $$\begin{aligned} \label{termoescroto} \alpha + 2\beta = \frac{dM}{dt} & \left[ \left( \frac{4M^2}{\sqrt{z}} + 4 - \frac{2q^2}{\sqrt{z}} - 8M\sqrt{z} \right) \left( M + \frac{J^2}{M^3} \right) + \right. \nonumber \\ & + \left. \left( - 8M^2\sqrt{z} - 12Mz - 8z^{\frac{3}{2}} - 8M^3 + \frac{4J^2 \sqrt{z}}{M^3} - \frac{8J^2}{M} \right) \right]. \end{aligned}$$ Since we do not have constraints on $J$, $M$ and $q$ other then $z > 0$, the sign of (\[termoescroto\]) is undetermined. However, we can set the conditions that would lead to maximum or minimum values of the tunnelling probabilities, in terms of the signs of the derivatives of $\dot{\theta}$ and $\slashed A$. We have summarized all the possible cases below (adopting the notation $y \equiv \mathcal{X} = \slashed A$): 1. $\alpha + 2\beta < 0$: 1. $\dot{y} < 0$, $\ddot{y} < 0$: $f(y) > 0$, minimum emission rate; 2. $\dot{y} > 0$, $\ddot{y} > 0$: $f(y) < 0$, maximum emission rate; 3. $\dot{y} > 0$, $\ddot{y} < 0$: both cases are possible; 4. $\dot{y} < 0$, $\ddot{y} > 0$: both cases are possible. 2. $\alpha + 2\beta > 0$: 1. $\dot{y} < 0$, $\ddot{y} < 0$: both cases are possible; 2. $\dot{y} > 0$, $\ddot{y} > 0$: both cases are possible; 3. $\dot{y} > 0$, $\ddot{y} < 0$: $f(y) > 0$, minimum emission rate; 4. $\dot{y} < 0$, $\ddot{y} > 0$: $f(y) < 0$, maximum emission rate. Finally, there are two conditions ruling these scenarios: all maximum emission rate are related to the condition $\displaystyle \frac{\dot{\sigma}}{\sigma} < - \frac{1}{2} \frac{\ddot{y}}{\dot{y}}$, while the minimum emission rates are linked to $\displaystyle \frac{\dot{\sigma}}{\sigma} > - \frac{1}{2} \frac{\ddot{y}}{\dot{y}}$. Note also that $\dot{\sigma} = \alpha + 2\beta$. The same results concerning the behaviour of the tunnelling probabilities are valid for all asymptotically flat black hole solution in general relativity, since setting $q = 0$ and/or $J = 0$ does not affect any of our conclusions under the same assumptions ($\omega$, $j\Omega$ and $q\Phi$ constants in time and $\dot{r}^2, \dot{\theta}^2, \ddot{r} \rightarrow 0$). Also, note that in all calculations we did not fixed any restriction to $\slashed A$ rather than $\slashed A(0) = 0$. Conclusion ========== In this paper we investigated the relationship between exotic spinorial structures and the emission rate of spinors by asymptotically flat black holes from general relativity. We established general conditions for the emission rate being extreme in two different cases, considering the existence or not of an external field $A_\mu$. From general conditions for the emission rate extremes we found differential equations for the exotic term. In particular, we found the dynamics of the exotic term $\dot{\theta}$ to the dynamics of $A_\mu$, depending on having or not extreme values for the emission of Dirac exotic spinors. Indeed, note that the right-handed side of Eqs. (\[condicaoDNzero\]) and (\[condicaoDNnaozero\]) are both constants in time. More accurately, let us call $DN-$particles the ones related to the Dirac exotic spinor fields that have no interaction with $A_\mu$, and suppose the situation that $\slashed A$ is increasing with time. Then, if it happens at the moment of extreme emission of $DN-$particles, the exotic term $\dot{\theta}$ also increases. On the other hand, if the emission of $DN-$particles is not in its extreme state, the exotic term $\dot{\theta}$ decreases. This relates the dynamics of the external field $A_\mu$ to the exoticiness of the spacetime itself, in the case of any black hole solution in general relativity. Moreover, in the case of $\dot{\Gamma}_{DN} \neq 0$ and $\dot{\Gamma}_{DE} = 0$ simultaneously, the Eq. (\[gammaDE\]) results in $\Gamma_{DE} = 1$, i.e., in this case we have maximum emission of Hawking radiation for $DE-$particles, with $100\%$ of tunnelling probability. Therefore, the only possibility to reach a minimum tunnelling probability for $DE-$particles is with $DN-$particles in the same situation (i.e., also with minimum radiation): this scenario happens if $\dot{\Gamma}_{DN} = 0 = \dot{\Gamma}_{DE}$. In order to check the conditions for maximum and minimum for this case by analysing the possible signs of $f(\mathcal{X}) = f(\slashed A)$, we have found that the sign of (\[termoescroto\]) is undetermined. Although, we have found a condition to separate these situations in terms of inequalities: maximum emission rate is happening when $\displaystyle \frac{\dot{\sigma}}{\sigma} < - \frac{1}{2} \frac{\ddot{y}}{\dot{y}}$, while minimum emission rate is related to the inequality with opposite sign. The conditions that lead to these relations are displayed in the (a-d) cases for the possible signs of $\alpha + 2\beta$. In this sense, the maximum and minimum emission rates are constrained by the exotic term. Then, by means of Hawking radiation emission rates, we have found relations between the black hole’s parameters mass, charge and angular momentum (via the $\dot{\sigma} = \alpha+2\beta$) with the spacetime topology (encoded in the exotic term $\dot{\theta}$) and the behaviour of external fields near to the black hole itself. It can help us to give a step ahead into the understanding of the yet not well known (up to our knowledge) behaviour of the exotic topological term, potentially enlightening the full dynamics of exotic spinors. Exotic dispersion relations {#append} =========================== There is an important history in similarities and differences concerning additional vectorial terms in Dirac equation. In this Appendix we would like to evince one more point regarding this history. In fact, we shall start from the old, but important, fact that every connected group have an unitary representation. This simple observation is crucial to understand the root of the difference between the exotic term appearing in the Dirac equation and the gauge interacting field. As well known, the gauge group of the electromagnetism is $U(1)$ and, hence, its elements are disposed as $e^{i\Lambda(x)}$ with $\Lambda \in \mathbb{R}$. Now, departing from the usual Dirac equation $(i\gamma^\mu \partial_\mu - m)\Psi = 0,$ and performing the transformation $\partial_\mu \rightarrow \partial_\mu - iA_\mu$, one finds $i\gamma^\mu \partial_\mu \rightarrow i\gamma^\mu \partial_\mu + \gamma^\mu A_\mu$, which turns the Dirac equation, along with the exotic extra-term, into $$\label{exoticdirac} (i\gamma^\mu \partial_\mu + \gamma^\mu A_\mu + i\gamma^\mu \partial_\mu \theta - m)\Psi = 0.$$ Performing a gauge transformation $A_\mu \rightarrow A_\mu - \partial_\mu \Lambda$, it would be necessary to have $\Lambda = i\theta$ to absorb the exotic term into a gauge transformation. However $\theta \in \mathbb{R}$, as it can be read from the very construction of the exotic term, thus it is not possible to have the exotic term acting as a shift on the gauge field. The exotic term and the gauge field have different nature themselves. In fact, both are vectorial fields, but there is not a gauge tuning which encompass the exotic term. Besides their very different origins, the gauge field and the exotic term lead to different dispersion relations, emphasizing that $\partial_\mu \theta$ is a vector but not a gauge field. From now on, we shall investigate a simple toy model evincing such a point. Using $\gamma^\mu (A_\mu + i\partial_\mu \theta) \equiv \gamma^\mu \tilde{A}_\mu$ in the Eqn. (\[exoticdirac\]) leads to $$\label{diractilde} (i\gamma^\mu - m + \gamma^\mu \tilde{A}_\mu) \Psi = 0.$$ Imposing a plane wave solution $$\Psi = \omega(p) e^{\mp i px},$$ Eqn. (\[diractilde\]) provides $$\label{diracplane} (\pm \slashed p - m + \gamma^\mu \tilde{A}_\mu) \omega(p) = 0.$$ Thus, in order to have nontrivial solution, there cannot exists an inverse for $(\pm \slashed p - m + \gamma^\mu \tilde{A}_\mu)$. Certainly, one can write $$(\pm \slashed p - m + \gamma^\mu \tilde{A}_\mu)^{-1} = \frac{(\pm \slashed p + m + \epsilon \gamma^\mu \tilde{A}_\mu)}{(\pm \slashed p - m + \gamma^\mu \tilde{A}_\mu)(\pm \slashed p + m + \epsilon \gamma^\nu \tilde{A}_\nu)},$$ with $\epsilon$ being a real parameter. Then, rewriting the denominator, we find the condition $$\label{impose} p^2 - m^2 \pm 2 \epsilon p^\mu \tilde{A}_\mu + (1-\epsilon)(m\gamma^\mu \tilde{A}_\mu \pm \gamma^\mu \tilde{A}_\mu \slashed p) + \epsilon \tilde{A}^\mu \tilde{A}_\mu = 0$$ to be sufficient to get nontrivial solution for Equation (\[diracplane\]). Now we are able to consider some particular cases of the above equation. - $\tilde{A}_\mu \rightarrow 0$: In this case, the imposition (\[impose\]) leads to $p^2 = m^2$, i.e., the usual dispersion relation $E^2 = \vec{p}^2 + m^2$, expected for plane waves. - $\partial_\mu \theta \rightarrow 0$ (or, equivalently, $\theta$ is constant): Here we have $\tilde{A}_\mu \rightarrow A_\mu$. With $\epsilon = 1$, one has $p^2 = m^2 \mp 2p^\mu A_\mu - A^2$, which leads, using the Coulomb gauge, to $$\label{dispersionGauge} E^2 = m^2 + (\vec{p} \pm \vec{A})^2,$$ revealing an unusual spectrum for the plane wave solution. Either way, it is a feasible observation signature typical for plane wave like fermions interacting with the electromagnetic field, without exotic term. - $A_\mu \rightarrow 0$: This case leads to $$p^2 - m^2 + \epsilon \partial^\mu \theta \partial_\mu \theta + i [\pm 2\epsilon p^\mu \partial_\mu \theta + (1-\epsilon)\gamma^\mu \partial_\mu \theta (m \pm \slashed p)] = 0,$$ and there are no choices for $\epsilon$ providing a real dispersion relation. Notice, however, that with $\epsilon = 1$ we have $$p^2 - m^2 + \partial^\mu \theta \partial_\mu \theta \pm 2i p^\mu \partial_\mu \theta = 0$$ and a topology that generates a $\theta$ such that[^5] $p^\mu \partial_\mu \theta = 0$ can lead to a relation of the type $p^2 = m^2 - \partial^\mu \theta \partial_\mu \theta$. In this situation, we have $$E^2 = \vec{p}^2 + m^2 - \partial^\mu \theta \partial_\mu \theta.$$ Moreover, due to the constraint $p^\mu \partial_\mu \theta = 0$, we find that the relation $E^2 = \frac{(\vec{p} \boldsymbol{\cdot} \vec{\nabla}\theta)^2}{\dot{\theta}^2}$ must be considered and, eventually, used to constraint the exotic term. The relevant point here is that the dispersion relation $$E^2 = \vec{p}^2 + m^2 - \dot{\theta}^2 + \vec{\nabla} \theta \boldsymbol{\cdot} \vec{\nabla} \theta$$ obtained in this case is different from that obtained in Equation (\[dispersionGauge\]), which had only took into account the gauge field interaction. Finally, the case with all terms (gauge field and exotic) may also be analysed, leading to a corresponding and different dispersion relation. Despite the simplicity, such analysis demonstrates that there exists a possible observable difference between an interacting and a topological Dirac operator. Acknowledgment {#acknowledgment .unnumbered} -------------- The authors are grateful to Professor Roldão da Rocha for bringing us together into the spinorial issues. JMHS thanks to CNPq for partial support. RCT thanks the UNESP-Guaratinguet´a Post-Graduation program and CAPES. DB thanks to CAPES for the financial support. [1]{} E. Cartan, *The Theory of Spinors*, MIT Press, Cambridge, 1967. R. Penrose and W. Rindler, *Two-Spinor Calculus and Relativistic Fields*, Second Edition, Cambridge Univ. Press, Cambridge (1986). M. Nakahara, *Geometry, Topology, and Physics*, Second Edition, IoP Publishing Ltd, London (2003). R. da Rocha, A. E. Bernardini and J. M. Hoff da Silva, JHEP [**1104**]{}, 110 (2011). S. W. Hawking and C. N. Pope, Phys. Lett.  [**73B**]{}, 42 (1978). R. P. Geroch, J. Math. Phys.  [**9**]{}, 1739 (1968). R. P. Geroch, J. Math. Phys.  [**11**]{}, 343 (1970). S. J. Avis and C. J. Isham, Nucl. Phys. B [**156**]{}, 441 (1979); S. J. Avis and C. J. Isham, Commun. Math. Phys. [**72**]{}, 103 (1980). H. R. Petry, J. Math. Phys. [ bf 20]{}, 231 (1979). Asselmeyer-Maluga T. [*Exotic smoothness and physics*]{} World Scientific (2007). T. Friedrich, [*Dirac operators in Riemannian geometry*]{}, Graduate Studies in Mathematics 25, American Mathematical Society, Providence U.S.A. (2000). C. J. Isham, Proc. R. Soc. London, Ser. A [**362**]{}, 383 (1978); C. J. Isham, Proc. R. Soc. London, Ser. A [**364**]{}, 591 (1978). J. M. Hoff da Silva, C. H. Coronado Villalobos, and R. da Rocha, Universe [**2**]{}, 8 (2016). M. K. Parikh and F. Wilczek, Phys. Rev. Lett.  [**85**]{}, 5042 (2000). M. Angheben, M. Nadalini, L. Vanzo and S. Zerbini, JHEP [**0505**]{}, 014 (2005). M. Arzano, A. J. M. Medved and E. C. Vagenas, JHEP [**0509**]{}, 037 (2005). Q. Q. Jiang, S. Q. Wu and X. Cai, Phys. Rev. D [**73**]{}, 064003 (2006); Erratum: \[Phys. Rev. D [**73**]{}, 069902 (2006)\]. L. Vanzo, G. Acquaviva and R. Di Criscienzo, Class. Quant. Grav.  [**28**]{}, 183001 (2011). R. T. Cavalcanti and R. da Rocha, Adv. High Energy Phys.  [**2016**]{}, 4681902 (2016). D. V. Ahluwalia, Adv. Appl. Clifford Algebras [**27**]{}, no. 3, 2247 (2017). [^1]: See Ref. [@exo] for a more complete account. [^2]: A special collection, entitled “Open Questions in Black Hole Physics”, Ed. Gonzalo J. Olmo. [^3]: Such space is nothing but a Riemann sphere. [^4]: With the condition that both numbers are not simultaneously zero. [^5]: A sufficient condition for $p^\mu \partial_\mu \theta = 0$ is $\partial_\mu \theta \sim a_\mu$, the 4-acceleration.

--- abstract: 'Starting from the tri-Hamiltonian formulation of the Lagrange top in a six-dimensional phase space, we discuss the possible reductions of the Poisson tensors, the vector field and its Hamiltonian functions on a four-dimensional space. We show that the vector field of the Lagrange top possesses, on the reduced phase space, a quasi-bi-Hamiltonian formulation, which provides a set of separation variables for the corresponding Hamilton-Jacobi equation.' --- \[section\] \[prop\] \[prop\] \[prop\] addtoreset[equation]{}[section]{} [The quasi-bi-Hamiltonian formulation of the Lagrange top]{} [Carlo Morosi${}^1$, Giorgio Tondo${}^2$ ]{}\ ${}^1$ Dipartimento di Matematica, Politecnico di Milano,\ P.za L. da Vinci 32, I-20133 Milano, Italy\ e–mail: carmor@mate.polimi.it\ ${}^2$ Dipartimento di Scienze Matematiche, Università di Trieste,\ via A. Valerio 12/1, I-34127 Trieste, Italy\ e–mail: tondo@univ.trieste.it\ **Keywords:** Lagrange top, Hamiltonian formulation, Hamilton-Jacobi separability. **AMS 2000 Subject classifications:** 37K10, 37J35, 53D17, 70E40, 70H06. Introduction {#intro} ============ The classical theory of separation of variables for the Hamilton-Jacobi equation provides the most effective tool to solve the equations of motion of a given Hamiltonian system. In this framework, the main problem is to have an efficient (possibly algorithmic) way to [*produce*]{} a set of separation variables. To this purpose, two new approaches, stemming from soliton theory, have been recently introduced: the “magic Sklyanin’s recipe"[@Sk], based on the Lax representation of the equations of the motion, and the bi-Hamiltonian ($bH$) approach to separation of variables [@bro; @mt; @bla; @FMPZ], based on the bi-Hamiltonian structures associated with the equations of motion. A remarkable feature of the latter approach is that if the Hamiltonian system admits a [*[quasi-bi-Hamiltonian]{}*]{} ($qbH$) formulation, then a set of separation variables can be algorithmically computed [@mt]; moreover, the $qbH$ property is independent of the coordinate system in which the $bH$ structure is written down. The aim of this paper is to apply the approach based on the $qbH$ property to the classical Lagrange top $(LT)$; in particular, we show how the (complex) separation variables for $LT$, introduced in [@med] in an algebraic-geometric setting, arise quite naturally as distinguished functions for its tri-Hamiltonian structure. The starting point of our analysis is the fact that, on a six-dimensional phase space $M$, the vector field $X_L$ admits a tri-Hamiltonian formulation $X_L= P_\alpha dh_\alpha$ (throughout the paper, the index $\alpha$ takes values $0, 1, 2$), each one of the three compatible Poisson tensors $P_\alpha$ possessing two independent Casimir functions. When one tries to eliminate the Casimirs by fixing their values, one is faced with a typical situation, occurring also for other $bH$ finite-dimensional integrable systems [@GT1; @FMT; @FMPZ]: to each one of the symplectic leaves $S_\alpha$ one can restrict only the vector field $X_L$ and the corresponding pair ($P_\alpha, h_\alpha$), but not the entire triple of the Poisson tensors, so that the tri-Hamiltonian formulation of $X_L$ is lost under restriction. Nevertheless, using a more general reduction process [*à la Marsden-Ratiu*]{}, we will show that the symplectic leaf $S_0$ of the Poisson tensor $P_0$ can be endowed with a Poisson-Nijhenuis structure [@MM; @FoFu] (hence a $bH$ structure) and that $X_L$ can be given a $qbH$ formulation. So, the separability of $LT$ is obtained from its Hamiltonian structures as a natural outcome of the reduction process. The paper is organised as follows. In Section \[rev\] the tri-Hamiltonian structure of is shortly reviewed; in Section \[reminder\] the main properties of the $qbH$ model are discussed in view of application to $LT$. In Sections \[reduno\] and \[reddue\], respectively, the reduction of the Poisson tensors $P_{\alpha}$ and of the vector field $X_L$ with its Hamiltonian functions are considered; the $qbH$ formulation for $X_L$ is explicitly constructed, together with a solution of the corresponding Hamilton-Jacobi equation. Our results are summarised in Section 6, where some potential extensions of this work are pointed out. 0.4cm The multi-Hamiltonian structure of the Lagrange top {#rev} =================================================== A modern formulation of $LT$ can be found in [@aud; @gav]; as usual in this framework, the components of vectors and covectors and the entries of matrices are referred to the comoving frame, whose axes are the principal inertia axes of the top, with fixed point $O$. The phase space $M$ of is parametrised by the pair $m=(\omega,\gamma)$, where $\omega = (\omega_1, \omega_2, \omega_3)^T~$ and $\gamma = (\gamma_1, \gamma_2, \gamma_3)^T~$ are the angular velocity and the vertical unit vector, respectively. The following notations are introduced: $\mu$ is the mass of the top, $g$ the acceleration of gravity, $J=\mbox{diag}(A, A, cA)$ the principal inertia matrix $(c\neq 1)$, $G=(0,0,a)^T$ is the center of mass; at last, normalisations are chosen so that $\mu a g/A=1$. The Euler-Poisson equations are $dL_o/dt=M_o$ (change of the angular momentum) and $d\gamma/dt=0$ (invariance of the vertical unit vector); with the above notations and normalisations, these equations take the well-known form =X\_L(m) ,X\_L(m)= ( [c]{} (1-c) \_2 \_3-\_2\ -(1-c) \_3 \_1+\_1\ 0\ \_2 \_3- \_3 \_2\ \_3 \_1- \_1 \_3\ \_1 \_2- \_2 \_1\ ) . \[moto\] The vector field $X_L$ can be given a tri-Hamiltonian formulation X\_L= P\_0 dh\_0= P\_1 dh\_1= P\_2 dh\_2 ; \[tri\] the compatible Poisson tensors $P_\alpha$, written in matrix block form, are \_0= ( [cc]{} 0 & B\ B & C\ ) , \_1= ( [cc]{} -B & 0\ 0 &\ ) , \_2= ( [cc]{} T & R\ -R\^T & 0\ ) , \[poisson\] where $B$, $C$, $\Gamma$, $T$ and $R$ are $3\times 3$ matrices $$B= \left( \begin{array}{ccc} 0 & -1 & 0 \\ 1 & 0 & 0 \\ 0 & 0& 0 \\ \end{array} \right), \quad C= \left( \begin{array}{ccc} 0 & c ~\omega_3 & -\omega_2 \\ - c ~\omega_3 & 0 & \omega_1 \\ \omega_2 & -\omega_1 & 0 \\ \end{array} \right), \quad \Gamma = \left( \begin{array}{ccc} 0 & \gamma_3 & -\gamma_2 \\ -\gamma_3 & 0 & \gamma_1 \\ \gamma_2 & -\gamma_1 & 0 \\ \end{array} \right),$$ T= ( [ccc]{} 0 & - c  \_3 & \_2/ c\ c  \_3 & 0 & - \_1 / c\ - \_2 / c & \_1/ c & 0\ ) , R = ( [ccc]{} 0 & -\_3 & \_2\ \_3 & 0 & -\_1\ -\_2/ c & \_1 /c & 0\ )  . \[bcgamma\] The Hamiltonian functions $h_\alpha$ can be written as h\_0=[12]{} F\_4+2 c F\_1 F\_3,h\_1=c\^2 F\_[1]{}\^3- F\_3-2 c F\_1 F\_2,h\_2=F\_2, \[ham\] where $\sigma= {c-1\over 2c }$ and F\_1=\_3,F\_2=[12]{} (\_[1]{}\^2+\_[2]{}\^2+ c  \_[3]{}\^2) -\_3, \[integrali\] $$F_3= \omega_1\gamma_1+\omega_2\gamma_2+ c~ \omega_3\gamma_3~,\qquad F_4= \gamma_{1}^2+\gamma_{2}^2+ \gamma_{3}^2~.$$ As it is known, the functions $F_i$ $(i=1,...,4)$ are integrals of motion for Eq.[(\[moto\])]{}; they are independent and in involution w.r.t. each one of the three Poisson tensors. Moreover, $(F_1, F_2)$ are Casimir functions of $P_0$, $(F_1, F_4)$ of $P_1$ and $(F_3, F_4)$ of $P_2$. The vector field $X_L$ can be immersed in two different $bH$ chains, starting and ending with the Casimirs of the Poisson tensors $P_\alpha$: P\_0 dF\_2=0,P\_2 dF\_2=P\_0 dh\_0=X\_L, P\_2 dh\_0=P\_0 d(-F\_3\^2), P\_2 d(-F\_3\^2)=0 ; $$P_0 dF_2=0,\quad P_1 dF_2=P_0 dh_1, \quad P_1 dh_1=P_0 dh_0=X_L, \quad P_1 dh_0= P_0 d(-\sigma c F_1 F_4),$$ $$\quad P_1 d( -\sigma c F_1 F_4)=0~.$$ 0.2cm \[rem1\] [**[Remark.]{}**]{} The Hamiltonian formulation of w.r.t. $P_2$ is classical (see, e.g.,[@gav]). The formulation w.r.t. $(P_0, P_2)$ was introduced in [@rat] in the semidirect product $\mathfrak{so}(3)\times \mathfrak{so}(3)$, and was later recovered in [@med] in an algebraic-geometric setting. The tri-Hamiltonian formulation w.r.t. $(P_0, P_1, P_2)$ was constructed in [@mag], by a suitable reduction of the Lie-Poisson pencil defined in the direct sum of three copies of $\mathfrak{so}(3)$. (To compare the above-quoted results, let us recall that the angular momentum and the vertical unit vector are taken as dynamical variables in [@gav; @rat; @mag], whereas the angular momentum is replaced by the angular velocity $\omega$ in [@med] and in the present paper.) 0.4cm The quasi-bi-Hamiltonian model {#reminder} ============================== The $qbH$ model was introduced in [@cab; @bro] and developed in [@mt; @tm] (see also [@bla] and references therein). Here we summarise some facts to be used in the rest of the paper.Let $Q_0$, $Q_1$ be two compatible Poisson tensors on a manifold $M$; a vector field $X$ is said to admit a $qbH$ formulation w.r.t. $Q_0$ and $Q_1$ if there are three functions $\rho$, $H$, $K$ such that X= [Q]{}\_0  dH= [1]{}  [Q]{}\_1  dK . \[ro\] In other words, $X$ is Hamiltonian w.r.t. $Q_0$ with Hamiltonian function $H$, and it is [*quasi-Hamiltonian*]{} ($qH$) w.r.t. $Q_1$, with $qH$ function $K$ and conformal factor $1/\rho$. In spite of the presence of $\rho$, equation (\[ro\]) implies that $H$ and $K$ are in involution w.r.t. both Poisson brackets corresponding to $Q_0$ and $Q_1$ (as well as in the particular $bH$ case $\rho=1$). If $\mbox{dim}~M=2n$, the $qbH$ formulation is said to be of maximal rank if at each point $m\in M$ the Poisson tensors $Q_0$, $Q_1$ are non degenerate and the associated tensor $N=Q_1~ Q_{0}^{-1}$ (with vanishing Nijenhuis torsion) has $n$ independent eigenvalues $\lambda_1(m),..., \lambda_n(m)$. In this case, one can introduce a local chart $(\lambda_i, \mu_i)$ $(i=1,2,...,n)$, called a Darboux-Nijenhuis chart [@mm], such that $Q_0$, $Q_1$ and $N$ take the canonical form Q\_0= ( [cc]{} 0 & I\_n\ -I\_n& 0\ ), Q\_1= ( [cc]{} 0 &\ -& 0\ ),N= ( [cc]{} &0\ 0&\ ) , \[diag\] with $\Lambda=\mbox{diag} (\lambda_1,...,\lambda_n)$; in general, the coordinate functions $\mu_i$, canonically conjugated to $\lambda_i$, can be computed by quadratures. At last, the $qbH$ formulation is said to be of Pfaffian type if $\rho=\prod_{i=1}^n \lambda_i$. The following result has been proved in [@mt] for a Pfaffian $qbH$ vector field. \[hkgen\] **Proposition.** The general solution of Eq.[(\[ro\])]{} for the Pfaffian case is given by functions $H$ and $K$ which, in a Darboux-Nijenhuis chart $(\lambda_i, \mu_i)$, take the “canonical” form H=\_[i=1]{}\^n [f\_i \_i]{}, K=\_[i=1]{}\^n [\_i]{} [f\_i \_i]{} , \_i= \_[ji]{} (\_i-\_j) , \[gen\] where each $f_i$ is an arbitrary function, depending at most on the pair $(\lambda_i, \mu_i)$. Moreover, the Hamilton-Jacobi equations for both $H$ and $K$ are separable. 0.2cm This Proposition has a straightforward consequence. \[rem2\] [**Corollary.**]{} Let $X=Q_0~dH$ be a Hamiltonian vector field; if in a $Q_0$-Darboux chart $(x,y)$ the Hamiltonian $H$ takes the canonical form [(\[gen\])]{}, then $X$ admits a Pfaffian $qbH$ formulation w.r.t. a Poisson tensor $Q_1$ and a $qH$ function $K$ of the form [(\[diag\])]{} and [(\[gen\])]{}, respectively. Viceversa, let $X=1 / \rho~ Q_1~dK$ be a $qH$ vector field w.r.t. $Q_1$; if, in a chart $(x,y)$, $Q_1$ and $K$ take the canonical forms [(\[diag\])]{} [(\[gen\])]{} and $\rho=\prod_{i=1}^n x_i$, then it is also $X=Q_0~dH$ with $Q_0$ and $H$ given by (\[diag\]) (\[gen\]), respectively. Hence, the chart $(x,y)$ is a Darboux-Nijenhuis chart for the Poisson pair $Q_0, Q_1$. For $n=2$, this Corollary can be slightly generalised, in a way that is useful for subsequent applications to . \[hkgen1\] **Proposition.** Let $S$ be a four-dimensional manifold and $Y=Q_0~ dH $ be a Hamiltonian vector field w.r.t. a non degenerate Poisson tensor $Q_0$. Let there is a Darboux chart $(x, y)$ such that the Hamiltonian $H$ can be written as a linear combination of two functions $\hat{H}$, $\hat{K}$ with the canonical form [(\[gen\])]{}, i.e., H(x,y)= (x,y) + (x,y) =const  , \[y\] $$\hat{H}(x,y)={1\over x_1-x_2} \left(\hat{f}_1(x_1, y_1)- \hat{f}_2(x_2, y_2)\right)~,$$ $$\hat{K}(x,y)={1\over x_1-x_2} \left(x_2 \hat{f}_1(x_1, y_1)- x_1 \hat{f}_2(x_2, y_2)\right)~.$$ Then, the vector field $Y$ admits the Pfaffian $qbH$ formulation (\[ro\])-(\[gen\]); a Darboux-Nijenhuis chart $(\lambda,\mu)$ is given by the following map: :(x,y)(,)\_i=[1x\_i+]{} ,\_i=-y\_i ( x\_i+)\^2 (i=1,2) . \[fi\] Hence, $H$ is separable in the chart $(\lambda,\mu)$. Moreover, $H$ is separable also in the chart $(x,y)$ and the corresponding Hamilton-Jacobi equation H(x\_1, x\_2, W /x\_1, W /x\_2)=h \[hamjac\] has the complete solution $ W(x_1, x_2; \hat{h},\hat{k})= W_1(x_1; \hat{h},\hat{k})+ W_2(x_2; \hat{h},\hat{k})$ , $W_1$ and $W_2$ fulfilling the Sklyanin separation equations [@Sk] \_1(x\_1, W’\_1(x\_1))= x\_1 -, \_2(x\_2, W’\_2(x\_2))= x\_2 -  , \[eqSk\] with $\beta\,\hat{h} + \,\hat{k}= h~$. [**Proof.**]{} It is straightforward to check that the map $ \Phi: (x,y)\mapsto (\lambda,\mu)$ is a Darboux map for $Q_0$; moreover, since $x_1-x_2=-(\lambda_1-\lambda_2) /{ \lambda_1 \lambda_2}$, the Hamiltonian $H$ takes the canonical form (\[gen\]): H (x(,),y(,))= (x(,),y(,)) + (x(,),y(,))= $$= { 1\over \lambda_1-\lambda_2} \left( - \lambda_1 \hat{f}_1 ({1\over\lambda_1} -\beta, -\lambda_1^2\mu_1 ) + \lambda_2 \hat{f}_2 ({1\over\lambda_2} -\beta, -\lambda_2^2 \mu_2) \right)=$$ $$= {1\over \lambda_1-\lambda_2} \Big(f_1(\lambda_1, \mu_1) -f_2(\lambda_2, \mu_2)\Big)~,$$ where f\_1(\_1 , \_1)=-\_1\_1 ([1\_1]{} -, -\_1\^2\_1) , f\_2(\_2 , \_2)=-\_2\_2 ([1\_2]{} -, -\_2\^2\_2) . On account of Corollary \[rem2\], the vector field $Y= Q_0 dH$ admits the $qH$ formulation $Y=1/\rho~ Q_1 dK$ and $H$ is separable. Obviously enough, $H$ is separable also in the chart $(x,y)$, since the map $\Phi$ is a [*separated*]{} map [@Be], i.e., it maps separated coordinates into separated ones. Indeed, taking into account the form [(\[y\])]{} of the function $H$, it is easily checked that the Hamilton-Jacobi equation $H(x,\partial W/\partial x)= h$ has a complete solution $ W(x_1, x_2; \hat{h}, \hat{k})=$ $W_1(x_1; \hat{h}, \hat{k})+ W_2(x_2; \hat{h}, \hat{k})$, with $\beta\,\hat{h} +\,\hat{k}= h~$, and that $W_1$, $W_2$ fulfil the Sklyanin separation equations (\[eqSk\]) for the Hamilton-Jacobi equations $\hat{H}(x,\partial W/\partial x)=\hat{h}$,   $\hat{K}(x,\partial W/\partial x)=\hat{k}$. The reduction of the tri-Hamiltonian structure of the Lagrange top {#reduno} ================================================================== If a vector field $X$ on a manifold $M$ is $bH$ w.r.t. a pair of degenerate Poisson tensors $(P_0, P_1)$, a preliminary step in analysing its integrability is trying to reduce the vector field, its Hamiltonian functions and the Poisson tensors on a lower-dimensional manifold $M'$, where one of the two Poisson tensors, say $P_0$, be invertible. A natural way to do that is to fix the values of the Casimir functions of $P_0$. Of course, both $P_0$ and $X$ can be properly [*restricted*]{} to a symplectic leaf $S_0$, giving rise to a Poisson tensor $P'_0$ and to a vector field $X'= P'_0~dH'$, $H'$ being the restriction to $S_0$ of the original Hamiltonian $H$. However, without additional assumptions, $P_1$ is not assured to restrict to $S_0$, so that $X'$ loses the original $bH$ formulation. This situation occurs also for the tri-Hamiltonian structure of . Each one of the three Poisson tensors $P_\alpha$ has two independent Casimir functions, and the generic symplectic leaves $S_\alpha$ are four-dimensional submanifolds of $M$. On account of Eq.[(\[integrali\])]{}, they are defined as S\_0={ mM   \_3=[a\_1 2c]{},   \_[1]{}\^2+\_[2]{}\^2+ c \_[3]{}\^2-2\_3=2a\_2 } , \[leave0\] $$S_1=\{ m\in M~\vert~ \omega_3={a_1 \over 2c}, ~~\gamma_{1}^2+\gamma_{2}^2+ \gamma_{3}^2=a_4\}~,$$ $$S_2=\{ m\in M~\vert~ \omega_1\gamma_1+\omega_2\gamma_2+ c \omega_3\gamma_3= -{1\over 2}a_3~,~~\gamma_{1}^2+\gamma_{2}^2+ \gamma_{3}^2=a_4\}~,$$ where $a_1$, $a_2$, $a_3$ and $a_4$ are arbitrary constants. Each Poisson tensor $P_\alpha$ can be properly restricted to a corresponding symplectic leaf $S_\alpha$, but the other two tensors do not restrict to the same leaf. Nevertheless, a quite general reduction technique given by the Marsden-Ratiu theorem  [@MR] can be applied; it will enables us to construct on $S_\alpha$ a Poisson-Nijhenuis structure [@MM; @FoFu] induced by the tri-Hamiltonian structure on $M$, and on $S_0$ a $qbH$ formulation for the vector field $X_L'$. Essentially, one considers a Poisson manifold $(M, P)$, a submanifold $S\hookrightarrow M$ and a distribution $D\subset TM_{\vert_S}$ such that $E:=D\cap TS$ is a regular foliation with a good quotient ${{\cal N}}= S/E$. Then, the theorem states that the Poisson tensor $P$ is reducible to ${{\cal N}}$ if the following conditions hold: the functions on $M$ which are invariant along $D$ form a Poisson subalgebra of $C^\infty(M)$; $P(D^\circ)\subset TS + D$ ($D^\circ$ being the annihilator of $D$ in $T^*M$). Analogously to previous applications of this procedure to $bH$ structures [@FMT; @FMPZ; @FMP], let us choose as the submanifold $S$ a generic symplectic leaf $S_\alpha$ of the Poisson tensor $P_\alpha$ and a distribution $D$ such that at each point $s_\alpha \in S_\alpha$ the following decomposition holds: $$\label{DTS} {T_{s_{\alpha}}} M={T_{s_{\alpha}}} S_\beta \oplus {D_{s_{\alpha}}} \ ,$$ $S_\beta$ being the symplectic leaf of $P_\beta \ (\beta=0,1,2)$ passing through $s_\alpha $. This assumption assures that [**ii)**]{} is trivially fulfilled and that $E=0$, so that the reduction procedure becomes a [*submersion*]{} $\Pi: M \rightarrow S_\alpha$ onto the manifold $S_\alpha$; then, it allows us to endow $S_\alpha$ with a non degenerate tri-Hamiltonian structure, since the kernels of the reduced Poisson tensors $P'_\beta $ vanish. Indeed, if $\Pi^*$ denotes the (injective) pull-back of the submersion $\Pi$, it is P’\_=(\^\*)\^[-1]{} ([Im\_[s\_]{}]{}\^\*P\_\^[-1]{}([D\_[s\_]{}]{} S\_)) $$\stackrel{ (\ref{DTS})}{=} (\Pi^*)^{-1}({Im_{s_{\alpha}}}\Pi^*\cap {Ker_{s_{\alpha}}} P_\beta)=0 ,$$ where we have taken into account that \^\*D\^, D\^ P\_=D\^ P\_)\^=D\^S\_)\^0. In the $LT$ case, the distribution is as follows. \[ld:lem\] **Lemma.** Let $D$ be the distribution given by the vector fields $$Z_1=-i c\frac{{{\partial}}}{{{\partial}}\omega_2}+ \frac{{{\partial}}}{{{\partial}}\omega_3}, \qquad Z_2=i\frac{{{\partial}}}{{{\partial}}\gamma_2}- \frac{{{\partial}}}{{{\partial}}\gamma_3}$$ ($i=\sqrt{-1}$). Moreover, let $\varphi_1, \varphi_2$ be two generic functions. Then, for each Poisson tensor $P_\alpha$ there are two vector fields $W_{1\alpha}$ and $W_{2\alpha}$ (depending on $\varphi_1$ and $\varphi_2$) such that $$\label{eq:lieZP} L_{\varphi_1 Z_1+ \varphi_2Z_2}\,(P_\alpha)=Z_1\wedge W_{1\alpha}+Z_{2}\wedge W_{2\alpha}$$ ($L_Z$ and $\wedge$ denoting the Lie derivative along the flow of the vector field $Z$ and the exterior product of vector fields, respectively). [**Proof.**]{} It is easy to check that $L_{ Z_j}P_\alpha= Z_1\wedge Y_{1j\alpha}+Z_{2}\wedge Y_{2j\alpha}$ ($j=1,2$), with suitable vector fields $Y_{1j\alpha}, Y_{2j\alpha}$. This result, together with the identity $L_{f X}(P)=fL_{X}(P)+X\wedge P\,df$, implies (\[eq:lieZP\]), the vector fields $W_{j\alpha}$ being $W_{j\alpha}=\varphi_1 Y_{j1\alpha} +\varphi_2 Y_{j2\alpha}+P_\alpha d\varphi_j$. Eq.(\[eq:lieZP\]) implies the assumption [**i)**]{}, since if $f$ and $g$ are invariant functions along $D$ and $Z \in D$, then $L_{\varphi Z}\{f,g\} =<df, L_{\varphi Z}(P) dg>\stackrel {(\ref{eq:lieZP})}{=}0$ for each function $ \varphi$. Moreover, condition (\[DTS\]) is generically satisfied as it can be easily verified. Hence, conditions [**i)**]{}, [**ii**]{}) are fulfilled and the Marsden-Ratiu reduction technique can be applied on each symplectic leaf $S_\alpha$. In conclusion, we have proved the following. \[prpo4\] **Proposition.** The tri-Hamiltonian structure $P_\beta$ is [*reducible*]{} to a non degenerate tri-Hamiltonian structure $P'_\beta$ on each one of the symplectic leaves $S_\alpha$. To express the reduced tensors in a particularly simple and useful form, it is convenient to adapt the coordinates on $M$ to the distribution $D$, introducing a parametrisation including coordinate functions which span the subalgebra of the functions invariant along $D$. Let us choose the chart $(u, v, w)$, related to $(\omega, \gamma)$ by the map $\Psi:M \rightarrow M:$ $(\omega, \gamma)\mapsto (u, v, w)$ \[map\] u\_1=c\_3-i \_2 ,u\_2=i\_2-\_3 , $$v_1=\omega_1,\quad v_2=-\gamma_1,\qquad w_1=i\omega_2+c \omega_3,\quad w_2=-i\gamma_2-\gamma_3~.$$ Taking into account the tri-Hamiltonian structure $P_\alpha$ given by [(\[poisson\])]{} and the definition [(\[leave0\])]{} of $S_\alpha$, a straightforward (though lengthy) calculation allows one to verify that the chart ($u, v$) gives a parametrisation on each one of the symplectic leaves $S_\alpha$; the reduced Poisson tensors $P'_{\beta}$ and the tensor $N$ take the form $$\label{eq:P0} P'_0=i\left( \begin{array} {cccc} 0& 0& 0&1 \\ 0& 0& 1& u_1 \\ 0& -1& 0& 0\\ -1& - u_1&0& 0 \end{array} \right) \ , \qquad P'_1=i\left( \begin{array} {cccc} 0& 0& 1& 0 \\ 0& 0&0&-u_2\\ -1 & 0& 0& 0\\ 0 &u_2& 0&0 \end{array} \right) \ ~,$$ $$P'_2=i\left( \begin{array} {cccc} 0& 0& -u_1& -u_2 \\ 0& 0& -u_2& 0 \\ u_1 & u_2& 0& 0\\ u_2& 0& 0&0 \end{array} \right) \ .$$ **Remark.** \[hankel\] By a direct inspection, one easily concludes that the tensor $N':=P'_1 {P'_0}^{-1}$ (with vanishing Nijenhuis torsion) is such that $P'_1= N' P'_0$ and $P'_2=N' P'_1$. The matrix representation of $P'_0$ and of the adjoint tensor ${N'}^*$ of $N'$ are formed by Hankel and Frobenius blocks, respectively, so that $(u,v)$ are Hankel-Frobenius coordinates, in the terminology of [@FMT]. \[conclude\] **Proposition.** Let us consider the map $\Psi: S_\alpha \rightarrow S_\alpha:$ $(u, v)\mapsto (x, y) $ \[another\] x\_1=[12]{} (-u\_1-) ,x\_2=[12]{} (-u\_1+) , $$y_1={1\over 2} (2 v_2-u_1 v_1- v_1 \sqrt{u_{1}^2-4u_2})~,\qquad y_2={1\over 2} (2 v_2-u_1 v_1+v_1 \sqrt{u_{1}^2-4u_2})~.$$ The chart $(x,y)$ is a Darboux-Nijenhuis chart for the tri-Hamiltonian structure on $S_\alpha$; the reduced Poisson tensors $P'_\alpha$ have the matrix block form P’\_0=-i ( [cc]{} 0 & I\ -I& 0\ ) P’\_1=-i ( [cc]{} 0 & X\ -X & 0\ ) P’\_2=-i ( [cc]{} 0 & [X]{} \^2\ -[X ]{}\^2& 0\ ) , \[ppp\] where $ {\cal X} =\mbox{diag} (x_1, x_2)$ . [**[Proof.]{}**]{} A straightforward computation, taking into account Eq.s [(\[eq:P0\])]{} and [(\[another\])]{}. (To be more precise, in order to have the Darboux-Nijenhuis chart defined in Section \[reminder\] one should eliminate the factor $(-i)$ in Eq.[(\[ppp\])]{}, via the map $x \mapsto i x$, $y \mapsto y)$. 0.4cm The reduction of the vector field and the Hamiltonians of the Lagrange top. {#reddue} =========================================================================== Having established the projection of the tri-Hamiltonian structure on each one of the symplectic leaves $S_\alpha$, the next step is to consider the reduction of the vector field $X_L$ and of the corresponding Hamiltonian functions $h_\alpha$. Unfortunately, they [*do not project*]{} onto $S_\alpha$, since $X_L$ does not preserve the distribution $D$ and the Hamiltonians $h_\alpha$ are not invariant along $D$; hence, the tri-Hamiltonian formulation of $X_L$ is lost on $S_\alpha$. Nevertheless, each pair $(X_L, h_\alpha)$ can be [*restricted*]{} to the corresponding symplectic leaf $S_\alpha$, so that Eq.[(\[moto\])]{}, restricted to $S_\alpha$, keeps a Hamiltonian formulation. Furthermore, if we consider the reduction on a symplectic leaf $S_0$, we can recover, as a reminder of the original tri-Hamiltonian formulation, a $qbH$ formulation for $X_L$; this suffices to provide a set of separation variables. Indeed, the following holds. \[trivial\] **Proposition.** The vector field $X_L$, restricted to $S_0$, takes the form X\_L=P’\_0 dH=-i Q\_0  dH . \[cinqueuno\] Its Hamiltonian $H={h_0}_{\vert S_0}$ takes the form H(x, y)= a\_1  (x, y)+ (x, y)  , \[cinquedue\] where (x, y)=[1x\_1-x\_2]{}((x\_1, y\_1)-(x\_2, y\_2)) , \[cinquetre\] $$\hat{K}(x, y)={1\over x_1-x_2}\left(x_2 \hat{f}(x_1, y_1)-x_1 \hat{f}(x_2, y_2)\right)~,$$ $$\hat{f}(\xi, \eta)= -{1\over 2}~ \eta^2 + {1\over 2}~ \xi^4 + {1\over 2}~ a_1 \xi^3 + (a_2+\sigma~{a_{1}^2\over 4})\xi^2~.$$ [**[Proof.]{}**]{} A straightforward computation. On account of this result, we are just in the situation considered in Proposition \[hkgen1\], with =a\_1, \_1=\_2= . So, ${X}_L$ admits a $qbH$ formulation; the Darboux-Nijenhuis coordinates $(\lambda, \mu)$ are obtained from $(x, y)$ via the map [(\[fi\])]{}: \_i=(x\_i+ a\_1)\^[-1]{}, \_i=- y\_i (x\_i+ a\_1)\^2(i=1,2). \[trinv\] As it follows from the general results of Propositions \[hkgen\], \[hkgen1\], $H$ and $K$ are separable both in the Darboux-Nijenhuis chart $(\lambda, \mu)$ and in the chart $(x,y)$. Using the latter, let us compute a solution $W$ of the Hamilton-Jacobi equations for $H$ and $K$ H(x\_1,x\_2, [Wx\_1]{}, [Wx\_2]{})= h, K(x\_1, x\_2, [Wx\_1]{}, [Wx\_2]{})= k ; \[HK\] taking into account the expression [(\[cinquetre\])]{} of $\hat{f}$ and the fact that the $qH$ function $K$ given by [(\[gen\])]{} turns out to be $K=\hat{H}$, we have W(x\_1,x\_2; h,k)= \^[x\_1]{}   d+ \^[x\_2]{}   d ,$$g(\xi)= \xi^4+ a_1 \xi^3 + (2 a_2 +\sigma {a_{1}^4\over2} ) \xi^2 - 2 k \xi + 2(h-\sigma a_1 k). \label{wx}$$ Concluding remarks ================== The first result in this paper is that, reducing [*à la Marsden-Ratiu*]{} the tri-Hamiltonian structure ($P_0$, $P_1$, $P_2$) of $LT$ onto a generic symplectic leaf $S_\alpha$ of each Poisson tensor, a non degenerate Poisson-Nijenhuis structure is obtained. The reduction depends essentially on the distribution $D$ fulfilling (\[DTS\]) and (\[eq:lieZP\]); since $D$ may be not unique, possibly different Poisson-Nijenhuis structures can be constructed on the symplectic leaf. This point deserves further investigations. The second step of the reduction procedure is the restriction of the $LT$ vector field and Hamiltonian functions to the invariant submanifold $S_0$, discussed in Section 5. This produces a quasi-bi-Hamiltonian formulation for the $LT$ vector field and consequently, as a necessary outcome, a set of separation variables. An open question is whether the restriction of the $LT$ vector field to other invariant submanifolds, such as the symplectic leaves $S_1$ and $S_2$ of the Poisson tensors $P_1$ and $P_2$, gives rise to different sets of separation variables. As a last remark, we observe that the tri-Hamiltonian structure of $LT$ has a deformation in the original phase space $M$ (see, e.g., [@mag]). In fact, there is a vector field $\tau$ such that $L_\tau(P_2) = 2 P_1$, $L_\tau(P_1) = P_0$, $L_\tau(P_0) = 0; $ in the chart $(\omega, \gamma)$ chosen in this paper, $\tau$ is given by $\tau=(0,0,-2/c, \omega_1,\omega_2,c\, \omega_3)^T$. On the contrary, a recursion operator $N$ relating the Poisson tensors does not exist in $M$. Once the reduction onto the submanifold $S_0$ has been performed under the submersion $\Pi: M \rightarrow S_0$, the deformation process is preserved since the vector field $\tau$ is projectable onto $S_0$. Hence, the previous relations hold for $P'_0,~P'_1,~P'_2$ w.r.t. the projected vector field $\tau'$, given by $\tau'=-(2, u_1, 0, v_1,)^T$ in the chart $(u, v)$. As observed in Remark \[hankel\], on $S_0$ there is also an invertible recursion operator $N'$ such that $P'_1 = N' P'_0$ and $P'_2 = N' P'_1$ (consequently, one has a whole sequence of compatible Poisson tensors $P'_{j+1}=N' P'_j$ for each integer $j~$). One may wonder whether the recursion scheme based on $N'$ could be inferred from the existence of the deformation scheme on the initial phase space, and under which conditions on the deformation vector field $\tau$ . At the best of our knowledge, this question (which is not peculiar of $LT$ only) has not yet received a satisfactory answer; in our opinion, it deserves further investigations in the general framework of the reduction theory for multi-Hamiltonian manifolds. [**Acknowledgments.**]{} This work was partially supported by Italian M.I.U.R., under the research project [*Geometry of Integrable Systems*]{}, and by INDAM (G.N.F.M.). We thank an anonymous referee for useful suggestions about the style of the paper. [99]{} E.K. Sklyanin *Separation of variables. New Trends*, Progr. Theor. Phys. Suppl. [**118**]{} (1995), 35–60. R. Brouzet, R. Caboz, J. Rabenivo, V. Ravoson *Two degrees of freedom bi-Hamiltonian systems*, J.Phys. A, **29** (1996), 2069-2075. C. Morosi, G. Tondo *Quasi-bi-Hamiltonian systems and separability*, J.Phys. A **30** (1997), 2799-2806. M. Blaszak *Multi-Hamiltonian theory of dynamical systems*, Springer, Berlin (1998). G. Falqui, F. Magri, M. Pedroni, G. Zubelli *A bi-Hamiltonian theory for stationary KDV flows and their separability*, Regul. Chaotic Dyn. 5 (2000), 33–52. C. Médan *The bi-Hamiltonian structure of the Lagrange top*, Phys. Lett. A **215** (1996), 176–180. G. Tondo *On the integrability of stationary and restricted flows of the KdV hierarchy*, J. Phys. A [**28**]{} (1995), 5097–5115. G. Falqui, F. Magri, G. Tondo *Reduction of bihamiltonian systems and separation of variables: an example from the Boussinesq hierarchy,* Theor. Math. Phys., [**122**]{} (2000), 176–192. F. Magri, C. Morosi *A geometrical characterization of integrable Hamiltonian systems through the theory of Poisson-Nijenhuis manifolds*, Quaderno 19/S, Dip. Mat., Università di Milano (1984). A. Fokas, B. Fuchssteiner *Symplectic structures, Bäcklund transformations and hereditary symmetries,* Physica D[**4**]{} (1981), 47–66. M. Audin *Spinning tops*, Cambridge Univ.Press, Cambridge (1996) L. Gavrilov, A. Zhivkov *The complex geometry of the Lagrange top*, Enseign. Math. **44** (1998), 133–170. T. Ratiu *Euler-Poisson Equations on Lie algebras and the $n$-dimensional heavy rigid body*, Amer. Jour. Math. [**104**]{} (1982), 409–448. G. Magnano *Bihamiltonian approach to Lax equations with spectral parameters*, Acc. Sc. Torino-Mem. Sc. Fis. [**19-20**]{} (1995–1996), 159–209. R. Caboz, V. Ravoson, L. Gavrilov *Bi-Hamiltonian structure of an integrable Hénon-Heiles system*, J.Phys. A, **24** (1991), L523-L525. G. Tondo, C. Morosi *Bi-Hamiltonian manifolds, quasi-bi-Hamiltonian systems and separation variables*, Rep. Math. Phys. **44** (1999), 255–266. F. Magri, T. Marsico *Some developments of the concept of Poisson manifolds in the sense of A. Lichnerowitz*, in: Gravitation, Electromagnetism and Geometric structures (G. Ferrarese ed.), 207-222, Pitagora, Bologna (1996). S. Benenti *Intrinsic characterization of the variable separation in the Hamilton-Jacobi equation,* J. Math. Phys., [**38**]{} (1997), 6578–6602. J.E. Marsden, T. Ratiu *Reduction of Poisson Manifolds,* Lett. Math. Phys. [**11**]{} (1986), 161–169. G. Falqui, F. Magri, M. Pedroni *Bihamiltonian geometry and separation of variables for Toda lattices,* J. Nonlinear Math. Phys. [**8**]{} (2001), 118–127.

--- abstract: | In a previous paper we formulated a set of necessary conditions for the spherically symmetric weakly charged dust to avoid Big Bang/Big Crunch, shell crossing and permanent central singularities. However, we did not discuss the properties of the energy density, some of which are surprising and seem not to have been known up to now. A singularity of infinite energy density does exist – it is a point singularity situated on the world line of the center of symmetry. The condition that no mass shell collapses to $R = 0$ if it had $R > 0$ initially thus turns out to be still insufficient for avoiding a singularity. Moreover, at the singularity the energy density $\epsilon$ is direction-dependent: $\epsilon \to - \infty$ when we approach the singular point along a $t =$ const hypersurface and $\epsilon \to + \infty$ when we approach that point along the center of symmetry. The appearance of negative-energy-density regions turns out to be inevitable. We discuss various aspects of this property of our configuration. We also show that a permanently pulsating configuration, with the period of pulsation independent of mass, is possible only if there exists a permanent central singularity. author: - 'Andrzej Krasiński and Krzysztof Bolejko[^1]' title: 'Can a charged dust ball be sent through the Reissner–Nordström wormhole?' --- N. Copernicus Astronomical Center, Polish Academy of Sciences Bartycka 18, 00 716 Warszawa, Poland email: akr@camk.edu.pl The problem =========== In our previous paper [@KrBo2006] (Paper I) we formulated a set of necessary conditions for the spherically symmetric weakly charged dust to avoid Big Bang/Big Crunch, shell crossing and permanent central singularities. In fact, it had already been proven by Ori [@Ori1990; @Ori1991] that weakly charged dust with no central singularity necessarily evolves toward a shell crossing. However, Ori assumed that the absolute value of the charge density (in geometric units) is strictly smaller than the mass density throughout the volume, including the center. His proof did not include the subcase when the values of these two densities at the center of symmetry are equal. We discussed precisely this subcase. We found that shell crossings are still unavoidable when the energy function $E \geq 0$, but the conditions did not lead to a contradiction when $E < 0$. We also provided an example of a solution of these conditions, which was in fact incorrect (see below). However, we did not discuss the properties of the energy density. While we started to investigate more properties of the numerical example presented in Paper I, it turned out that this density was becoming negative in a vicinity of the center of symmetry, for a brief period around the bounce instant. Then, exact consideration showed that this is a general problem – a weakly charged spherically symmetric dust ball will always have such a negative-energy-density region if the charge and mass densities become equal in absolute value at the center. This fact is hidden in the equations in a rather tricky way. In the general case treated by Ori this problem can be avoided by an appropriate choice of the arbitrary functions. The explicit example we gave in Paper I had a problem - it did not obey the condition that the ratio $R/{\cal M}^{1/3}$ must have a finite nonzero limit at the center of symmetry. For our example, this limit was zero, which means that there was a permanent central singularity in it (the limit of the mass density at the center was actually $- \infty$). The permanent infinity is easily cured by changing $x^{4/3}$ to $x^{5/3}$ in the formulae for $\Gamma$ and $E$. The whole subsequent reasoning then applies with only little quantitative changes,[^2] but a direction-dependent point singularity at the center necessarily appears at the instant of maximal compression. In the present paper we fill in the gaps left out by previous investigation. We show that the energy density of the charged dust must become negative within a finite time-interval containing the instant of maximal compression. We also show that the point on the world line of the center of symmetry that is the limit of the points of maximal compression is a direction-dependent singularity of infinite energy density. The energy density tends to [*minus*]{} infinity when we approach the singular point along the hypersurface of maximal compression, and to plus infinity when we approach the same point along the center of symmetry. Thus, the condition that the areal radius $R(t, r)$ never goes down to zero if it was nonzero initially turns out to be still insufficient for avoiding singularities. These are the most striking and most important results of the present paper. In order to make the paper independently readable, we quote the basic results of Paper I in the next section, without proof. All the proofs (or references to the proofs) and details of the calculation can be found in Paper I. Section \[paper1\] also contains remarks on the interpretation of the energy density, which is nontrivial in the presence of charges, and was not properly discussed in the earlier papers. In section \[negdens\] we show that the quantity $u = \Gamma - QQ,_N / R$ must become negative in a vicinity of the instant of maximal compression whenever the conditions ${Q,_N}^2 = G / c^4$ at the center and ${Q,_N}^2 < G / c^4$ elsewhere are obeyed. In section \[negucons\] we discuss the various consequences of this fact and show that this necessarily implies a negative energy density for a certain period around the bounce instant. In section \[novac\] we show that the negative energy density regions cannot be eliminated by assuming that the charged dust ball contains an empty (Minkowski) vacuole around the center of symmetry because such a vacuole cannot exist. In section \[transient\] we prove the existence of the point-singularity on the world line of the center of symmetry. In section \[perm\] we show that, even with negative energy density allowed, a solution with the period of oscillation independent of mass does not exist. A permanently nonsingular configuration of nonstatic weakly charged dust is thus impossible, unless one allows a permanent central singularity. We also compare our conclusion with the properties of the uncharged case. Section \[conclu\] summarizes the conclusions. Basic formulae and results of Paper I {#paper1} ===================================== For a spherically symmetric spacetime in comoving coordinates, the metric can be put in the form $$\label{2.1} {\rm d} s^2 = {\rm e}^{C(t,r)} {\rm d} t^2 - {\rm e}^{A(t, r)} {\rm d} r^2 - R^2(t,r)\left[{\rm d} \vartheta^2 + \sin^2(\vartheta) {\rm d} \varphi^2\right].$$ In the generic case $R,_r \neq 0$, assuming there are no magnetic charges, the Einstein–Maxwell equations yield the following result [@KrBo2006; @PlKr2006; @Vick1973]. The only independent nonzero component of the electromagnetic field is $$\begin{aligned} F^{01} &=& Q(r) {\rm e}^{- (A + C)/2} / R^2, \label{2.2} \\ Q,_r &=& (4\pi / c) \rho_e {\rm e}^{A/2} R^2, \label{2.3}\end{aligned}$$ where $Q(r)$ is an arbitrary function – the electric charge within the $r$-surface, and $\rho_e$ is the electric charge density; $$\label{2.4} \frac {\kappa} 2 \epsilon R^2 {\rm e}^{A/2} = \frac G {c^4}\ N,_r, \qquad \kappa {\ {\overset {\rm def} =}\ }8 \pi G/c^4,$$ where $\epsilon$ is the energy density and $N,_r$ is an arbitrary function of integration. The $N$ so defined corresponds, in the electrically neutral case, to the energy equivalent to the sum of rest masses within the $r$-surface. With charges present, this interpretation involves a subtle point, see below after eq. (\[2.10\]). The ratio $Q,_r/N,_r = \rho_e/(c \epsilon)$ is time-independent. $$\label{2.5} {\rm e}^{A/2} = \frac {R,_r} {\Gamma(r) - QQ,_N / R},$$ where $Q,_N$ is just an abbreviation for $Q,_r/N,_r$; $$\label{2.6} C,_r = 2 \frac {{\rm e}^{A/2}} {R^2} QQ,_N;$$ $$\label{2.7} {\rm e}^{- C}{R,_t}^2 = \Gamma^2 - 1 + \frac {2M(r)} R + \frac {Q^2\left({Q,_N}^2 - G/c^4\right)} {R^2} - \frac 1 3 \Lambda R^2,$$ where $\Lambda$ is the cosmological constant, and $M(r)$ is an arbitrary function. By analogy with the uncharged Lemaître – Tolman model, $(\Gamma^2 - 1)/2$ is often denoted as $E(r)$. The function $M(r)$ is the [*effective mass*]{}, and it [*need not be positive*]{}. It is connected with the previously defined arbitrary functions by $$\label{2.8} \frac G {c^4} \Gamma N,_r = \left(M + QQ,_N\Gamma\right),_r.$$ The quantity $$\label{2.9} {\cal M} {\ {\overset {\rm def} =}\ }M + QQ,_N\Gamma,$$ is the active gravitational mass. Thus, via (\[2.8\]), $\Gamma$ determines by how much ${\cal M}$ increases when a unit of rest mass is added to the source, i.e. $\Gamma$ is a measure of the gravitational mass defect/excess. The energy density of the dust is: $$\label{2.10} \kappa \epsilon = \frac {2GN,_r} {c^4 R^2 R,_r}\ \left(\Gamma - \frac {QQ,_N} R\right).$$ Now note the subtle point. In the electrically neutral case, $Q = 0$, the quantity $\epsilon$ is the density of rest-mass (in energy units), so that the 3-space integral $\int_V \sqrt{- g} \epsilon {\rm d}_3x$ simply equals the sum of rest masses in the volume $V$ divided by $c^2$. However, with charges present, $\epsilon$ contains a contribution from the charges. As will be shown later in this paper, $QQ,_N \equiv \frac 1 2 \left(Q^2\right),_r / N,_r$ must be positive in a vicinity of the center of symmetry, and $R > 0$ everywhere for geometrical reasons. Thus, the presence of charges always [*decreases*]{} the energy density. It will turn out later that, in the class of models considered here, $\epsilon$ necessarily becomes negative for a brief period around the instant of maximal compression along each world-line except the central one. The functions $R(t, r)$ and $C(t, r)$ are implicitly defined by the set (\[2.6\]) – (\[2.7\]), which can in general be solved only numerically. If the configuration considered here is matched to the Reissner – Nordström (R–N) metric across a hypersurface $r = r_b$, then the following must hold [@PlKr2006; @Vick1973; @KrBo2006]: $$\label{2.11} e = \frac {\sqrt{G}} {c^2} Q(r_b), \qquad m = \left(M + QQ,_N \Gamma\right)_{r = r_b},$$ where $e$ and $m$ are the R–N charge and mass parameters, respectively. Unlike in the electrically neutral case, in charged dust the Big Bang/Big Crunch (BB/BC) singularity can be avoided, i.e. there exist solutions of (\[2.6\]) – (\[2.7\]) in which the function $R(t, r)$ never goes down to zero if it was nonzero initially.[^3] We give the conditions for the existence of such solutions only in the case $\Lambda = 0$. One of those conditions is $$\label{2.12} M^2 \geq 2EQ^2\left({Q,_N}^2 - G/c^4\right),$$ which is fulfilled identically when $E = 0$. With (\[2.12\]) fulfilled and $E \neq 0$, the right-hand side of (\[2.7\]) has two roots, given by $$\label{2.13} R_{\pm} = - \frac M {2E} \pm \frac 1 {2E} \sqrt{M^2 - 2EQ^2\left({Q,_N}^2 - G/c^4\right)}.$$ The second condition for avoiding the BB/BC singularity depends on $E$: \(a) When $E \leq 0$, the condition is $$\label{2.14} {Q,_N}^2 < G/c^4 \qquad {\rm and}\qquad M > 0.$$ With (\[2.14\]) fulfilled and $E < 0$, $R$ oscillates between the nonzero minimum at $R = R_+$ and the maximum at $R = R_-$. With (\[2.14\]) and $E = 0$, $R$ goes down from infinity to the finite minimal value $R_{\rm min} = Q^2 \left(G/c^4 - {Q,_N}^2\right) / (2M)$ and then increases to infinity again. \(b) When $E > 0$ and $M > 0$, the BB/BC singularity does not occur if ${Q,_N}^2 < G/c^4$. When $E > 0$ and $M < 0$, nonsingular solutions exist with no further conditions, provided $R > R_+$ initially. The bounce with $M < 0$ is nonrelativistic, since it occurs also in Newton’s theory, under the same conditions ($M < 0$ means that the electric repulsion of the charges spread throughout the volume of the dust prevails over the gravitational attraction of the mass). The inequality ${Q,_N}^2 < G / c^4$ translates into $\left|\rho_e\right| < \sqrt{G} \epsilon / c$, which means that, in geometric units, the absolute value of the charge density is [*smaller*]{} than the mass density.[^4] There is another subtle point here. The conditions (\[2.14\]) guarantee that a particle that had $R > 0$ initially will not hit the set $R = 0$ in the future or in the past. But, as we will see in Sec. \[transient\], the configurations obeying (\[2.14\]) contain a cleverly hidden singularity of a type hitherto unknown in dust solutions. On the world line of the center of symmetry, where $R(t, r) = 0$ permanently, there is a point in which $\epsilon \to + \infty$ for a single instant. This instant is the limit at $R \to 0$ of the hypersurface $S_{\rm min}$ consisting of the instants in which the mass shells with $R > 0$ attain their minimal sizes. However, if we approach the same spacetime location along the hypersurface $S_{\rm min}$, then $\epsilon \to - \infty$. The surface of the charged sphere obeys the equation of radial motion of a charged particle in the Reissner–Nordström spacetime. Thus, the surface of a collapsing sphere must continue to collapse until it crosses the inner R–N horizon $r_-$, and can bounce at $R \leq r_-$. Then, however, it cannot re-expand back into the same spacetime region from which it collapsed, as this would require motion backward in time. The surface would thus continue through the tunnel between the singularities and re-expand into another copy of the asymptotically flat region. At small charge density (${Q,_N}^2 < G / c^4$ throughout the volume) a shell crossing is unavoidable, and it will block the passage through the R–N tunnel, as shown by Ori [@Ori1990; @Ori1991]. A nonsingular bounce might be possible only if ${Q,_N}^2 > G / c^4$ everywhere or if ${Q,_N}^2 \to G / c^4$ at the center, while ${Q,_N}^2 < G / c^4$ elsewhere. The first case was dealt with by Ori (unpublished [@Oripriv]). We consider the second case in Section 3, and the result is that a nonsingular bounce might happen only if the energy density becomes negative for a period around the bounce instant. Shell crossings are most conveniently discussed in the mass-curvature coordinates $({\cal M}, R)$, first introduced by Ori [@Ori1990]. Details of the transformation can be found in Ref. [@PlKr2006]. The $({\cal M}, R)$ coordinates allow the Einstein–Maxwell equations with $\Lambda = 0$ to be explicitly integrated, but it happens at a price. The spacetime points are identified by the values of ${\cal M}$ and $R$, which means a given point is defined by saying “it is the place where the shell containing the mass ${\cal M}$ has the radius $R$”. The information on the time-dependence of $R$ is lost, and can be regained only by reverting to the comoving coordinates – but the transformation equations are equivalent to (\[2.6\]) – (\[2.7\]) and cannot be explicitly integrated. The solution of the Einstein – Maxwell equations is given below. The velocity field has only one contravariant component: $$\label{2.15} u^R = \pm \sqrt{\Gamma^2 - 1 + \frac {2M} R + \frac {Q^2\left({Q,_N}^2 - G/c^4\right)} {R^2} - \frac 1 3 \Lambda R^2}$$ ($+$ for expansion, $-$ for collapse). We define the auxiliary quantities $$\label{2.16} u {\ {\overset {\rm def} =}\ }\Gamma - QQ,_N/R, \qquad \Delta {\ {\overset {\rm def} =}\ }1 - \frac {2 {\cal M}} R + \frac {GQ^2} {c^4 R^2} + \frac 1 3 \Lambda R^2,$$ and we get for the metric $$\label{2.17} g_{{\cal M} {\cal M}} = F^2 \Delta, \qquad g_{{\cal M} R} = F u / u^R, \qquad g_{R R} = 1 / (u^R)^2,$$ where the function $F({\cal M}, R)$ is given by $$\label{2.18} F,_R = - \frac 1 {\left(u^R({\cal M}, R)\right)^3} \left\{\Gamma,_{\cal M} + \frac 1 {R \Gamma}\left[1 - \frac {c^4} G \left({Q,_N}^2 + QQ,_{NN}\right)\right]\right\}.$$ With $\Lambda = 0$, the integral of this is elementary, but to give its explicit form several cases have to be considered separately (see Ref [@Ori1990] for a list). Shell crossings occur at the zeros of the function $F$. Thus, to avoid shell crossings, the arbitrary functions must be chosen so that $F \neq 0$ everywhere. The only nonvanishing components of the electromagnetic tensor in the $({\cal M}, R)$ coordinates are $$\label{2.19} F^{{\cal M} R} = - F^{R {\cal M}} = \frac Q {FR^2}, \qquad F_{{\cal M} R} = - F_{R {\cal M}} = - \frac {FQ} {R^2},$$ while the charge density and the energy-density are $$\label{2.20} \frac {4\pi \rho_e} c = - \frac {Q,_{\cal M}} {R^2 F u^R}, \qquad \kappa \epsilon = - \frac 2 {\Gamma R^2 F u^R}.$$ The $({\cal M}, R)$-coordinates cover only such a region where $R,_t$ has a constant sign. The function $F$ changes sign where $R,_t$ does, and so does $u^R = {\rm e}^{- C/2} R,_t$. Thus, $Fu^R$ preserves its sign when collapse turns to expansion and vice versa. In the $({\cal M}, R)$ coordinates eq. (\[2.8\]) reads $$\label{2.21} {\cal M},_N = (G \Gamma / c^4).$$ Just as in the L–T model, the set $R = 0$ in charged dust consists of the Big Bang/Crunch singularity (which is now avoidable), and of the center of symmetry, which may or may not be singular. The conditions for the absence of a permanent central singularity are the following (not all of them are independent, but this is the full list): $$\begin{aligned} N(r_c) &=& 0 = Q(r_c) = {\cal M}\left(r_c\right) = M\left(r_c\right), \label{2.22} \\ \lim_{r \to r_c} R/{\cal M}^{1/3} &=& \beta(t) \neq 0, \label{2.23} \\ \lim_{r \to r_c} \Gamma^2(r) &=& 1 \Longrightarrow \lim_{r \to r_c} E(r) = 0, \label{2.24} \\ \lim_{r \to r_c} 2E / {\cal M}^{2/3} &=& \lim_{r \to r_c} \left(\Gamma^2(r) - 1\right) / {\cal M}^{2/3} {\ {\overset {\rm def} =}\ }D = {\rm const}, \label{2.25}\end{aligned}$$ and this last constant may be zero. In the mass-curvature coordinates, with $\Lambda = 0$, it is easy to solve the evolution equation ${{{\rm d} {R}} / {{\rm d} {s}}} = u^R$. We quote here the only one case, for which we found in Ref. [@KrBo2006] that there may exist solutions avoiding both kinds of singularity, $E < 0$ (see Fig. 5 in Paper I). We take $\ell = +1$ for expansion and $\ell = -1$ for collapse at the initial instant of evolution, and we denote $$\label{2.26} \Phi {\ {\overset {\rm def} =}\ }\frac {Q^2 \left({Q,_N}^2 - G / c^4\right)} {2E}.$$ When $E < 0$, a solution exists only with $\Phi < M^2 / \left(4E^2\right)$, and it is given by the parametric equations $$\begin{aligned} \label{2.27} R &=& - \frac M {2E} - \sqrt{\frac {M^2} {4 E^2} - \Phi}\ \cos \omega, \nonumber \\ s - s_B({\cal M}) &=& \frac {\ell} {\sqrt{- 2E}}\ \left(- \frac M {2E}\ \omega - \sqrt{\frac {M^2} {4 E^2} - \Phi}\ \sin \omega\right),\end{aligned}$$ where $\omega$ is a parameter and $s_B({\cal M})$ is an arbitrary function of integration. This will avoid a BB/BC singularity only if $M > 0$ and $\Phi > 0$. At $s = s_B$, $R$ starts off with the minimal value $R_{\rm min} = R_+$ (see eq. (\[2.13\])), and periodically returns to this value, never going down to zero if $R_{\rm min} \neq 0$. The period is $$\label{2.28} T_p = 2 \pi M / (- 2E)^{3/2}.$$ The period given by (\[2.28\]) is with respect to the proper time of the given shell of constant ${\cal M}$. To calculate the period in the time coordinate $t$, we must first transform the variables in eq. (\[2.7\]), to make $({\cal M}, R)$ the independent variables and $t$ the unknown function. We obtain $$\label{2.29} T_c = 2 \int_{R_+}^{R_-} \frac {{\rm e}^{-C({\cal M}, R)/2} {\rm d} R} {\sqrt{\Gamma^2 - 1 + 2M/R + Q^2 \left({Q,_N}^2 - G / c^4\right) / R^2 - (1/3) \Lambda R^2}}.$$ Note that each of the periods, $T_p$ and $T_c$, is a function of ${\cal M}$ only. We found in Paper I that even for a single bounce to be free of shell crossings, it is necessary that the bounce occurs in a time-symmetric manner, i.e. all the mass shells have to go through their minimal sizes at the same instant of the coordinate time $t$. Finally, we recall the 9 necessary conditions that the functions defining the model must obey in order that even a single nonsingular bounce can occur. We denote $$\label{2.30} F_1 {\ {\overset {\rm def} =}\ }1 - \left(c^4 / G\right) \left({Q,_N}^2 + QQ,_{NN}\right),$$ and the conditions are: \(1) $E < 0$; \(2) $E \geq -1/2$; \(3) $\lim_{r \to r_c} F_1/{\cal M}^{1/3} = 0$; \(4) $\Gamma,_{\cal M} < 0$; \(5) ${Q,_N}^2 < G / c^4$ at $N > 0$ and ${Q,_N}^2 = G / c^4$ at $N = 0$; \(6) $M \equiv {\cal M} - QQ,_N \Gamma > 0$; \(7) $M^2 - 2E Q^2 \left({Q,_N}^2 - G / c^4\right) > 0$; \(8) $$\label{2.31} F_1 > \frac {\Gamma \Gamma,_{\cal M}} {2E}\ M.$$ \(9) $$\label{2.32} F_1 > \frac {\Gamma \Gamma,_{\cal M}} {2E}\ \left[M + \sqrt{M^2 - 2EQ^2 \left({Q,_N}^2 - G / c^4\right)}\right].$$ Conditions (1) – (9) must hold in a [*neighbourhood*]{} of the center. [*At*]{} the center, the left-hand sides of conditions (1) and (6 – 9) must have zero limits. In addition to this, all the regularity conditions at the center must be obeyed. Note that the conditions (\[2.22\]) and (\[2.24\]), together with eq. (\[2.21\]), imply that $$\label{2.33} \lim_{N \to 0} M/N = 0$$ – this will prove useful later. For practical calculations, the most convenient radial coordinate is $N(r)$, and it will be used in some of the following sections. In Paper I we gave an example of a configuration that was supposed to obey conditions 1 – 9 and all the regularity conditions. In fact, it did not obey the condition (\[2.23\]). This oversight has already been corrected in an erratum, and in the version stored in the gr-qc archive, see Ref. [@KrBo2006]. That paper contained one more error: in the caption to Fig. 10 we stated that the two curves that seemed tangent were merely adjacent to each other. In truth, they [*were*]{} tangent – the minimal radius at bounce must be equal to the radius of the inner R–N horizon at every local extremum of the latter. The proof is given here in Appendix \[extrem\]. This is just a correction of an error that has no relation to the other results of the present paper. Inevitability of negative values of $u$ {#negdens} ======================================= From now on, we assume $\Lambda = 0$. From eq. (\[2.10\]) we see that if the quantity $u$ defined in (\[2.16\]) changes sign at a certain point $(t, r)$, while $R,_r$ does not, then the energy density also changes sign at that point. We will discuss the consequences of the change of sign of $u$ in the next section. In the present section we will show that, with the 9 conditions listed in Sec. 2 fulfilled, $u$ is necessarily negative within an interval containing the instant of maximal compression (minimal size) of every mass shell on which ${\cal M} \neq 0$. We first prove that negative $u$ appears in the hypersurface of maximal compression, and later we will show that $u / R,_r < 0$ for a certain period around the instant of maximal compression at every ${\cal M} \neq 0$. At first sight, it seems that the conditions (1) – (9) given at the end of Sec. 2 secure $u > 0$ in a vicinity of the center, since $\Gamma \to 1$, $Q/N \to \pm \sqrt{G}/c^2$ and $R/N^{1/3} \to \beta > 0$ as $r \to r_c$. However, the situation changes if $R = R_+$, i.e. if one approaches the center along the locus of the inner turning points. Fig. 12 in Ref. [@KrBo2006] (and also Fig. \[drawcycles2\] further here) shows that along this locus $R$ tends to zero faster than along any ${\cal M} =$ const line, and eq. (\[2.13\]) confirms this. At $R = R_+$ we have $$\label{3.1} u = \Gamma + \frac {Q,_N} {Q \left({Q,_N}^2 - G/c^4\right)} \left[M + \sqrt{M^2 - 2EQ^2 \left({Q,_N}^2 - G/c^4\right)}\right].$$ To guarantee $\epsilon > 0$, this should be positive everywhere, including the center $N = 0$. This requirement can be easily fulfilled if ${Q,_N}^2 < G/c^4$ everywhere including the center. Then, it is enough to choose such $Q(N)$ that the coefficient in front of the square brackets in (\[3.1\]) is small and tends to zero as $N \to 0$, for example $Q(N) = \pm (\sqrt{G}N_0/c^2) {\rm e}^{- a x^2}$, where $a > 0$ is a constant and $x {\ {\overset {\rm def} =}\ }N/N_0$. In our case, when $Q,_N$ must obey condition (5), it turns out that the limit of $u$ at $N \to 0$ is necessarily negative. This is seen as follows. We first observe that $$\label{3.2} \lim_{N \to 0} \left[- 2EQ^2 \left({Q,_N}^2 - G/c^4\right)\right] / M^2 = 0,$$ i.e. that the second term under the square root in (\[3.1\]) can be neglected compared to $M^2$ in the limit $N \to 0$ (see Appendix \[negli\]). Then, from (\[3.1\]) $$\label{3.3} \lim_{N \to 0} u = 1 + \lim_{N \to 0} \frac {2MQ,_N} {Q \left({Q,_N}^2 - G/c^4\right)}.$$ Since, by condition (5), $Q,_N(0) \neq 0$, we find $$\begin{aligned} \label{3.4} \lim_{N \to 0} u &=& 1 + 2 Q,_N(0) \lim_{N \to 0} \frac {{\cal M} - QQ,_N \Gamma} {Q \left({Q,_N}^2 - G/c^4\right)} \nonumber \\ &=& 1 + 2 Q,_N(0) \lim_{N \to 0} \frac {G\Gamma/c^4 - {Q,_N}^2 \Gamma - QQ,_{NN} \Gamma - QQ,_N \Gamma,_N} {Q,_N \left({Q,_N}^2 - G/c^4\right) + 2QQ,_N Q,_{NN}} \nonumber \\ &=& -1 + 2 \lim_{N \to 0} \frac {QQ,_{NN} \Gamma - QQ,_N \Gamma,_N} {{Q,_N}^2 - G/c^4 + 2QQ,_{NN}} \nonumber \\ &\equiv& -1 + 2 \lim_{N \to 0} \frac {Q,_{NN} \Gamma - Q,_N \Gamma,_N} {\left({Q,_N}^2 - G/c^4\right)/Q + 2Q,_{NN}} = -1 + \frac {Q,_{NN}(0) - Q,_N(0) \Gamma,_N(0)} {2Q,_{NN}(0)} \nonumber \\ &\equiv& - \frac 1 2 - \frac {Q,_N(0) \Gamma,_N(0)} {2Q,_{NN}(0)}.\end{aligned}$$ (In deriving this, we first applied the de l’Hopital rule, then did an algebraic simplification in the result, and again applied the de l’Hopital rule in the denominator.) Now, by condition (5), if $Q,_N(0) > 0$, then $Q,_{NN}(0) \leq 0$, and if $Q,_N(0) < 0$, then $Q,_{NN}(0) \geq 0$. By condition (4), $\Gamma,_N(0) \leq 0$. Thus, if $\Gamma,_N(0) < 0 \neq Q,_{NN}(0)$, then the second term in the last line of (\[3.4\]) can never be positive. If $\Gamma,_N(0) < 0 = Q,_{NN}(0)$, then the limit in (\[3.4\]) is $- \infty$ because $Q,_{NN} < 0$ in a vicinity of $N = 0$. If $\Gamma,_N(0) = 0 \neq Q,_{NN}(0)$, then the limit is $- 1/2 < 0$. If $\Gamma,_N(0) = 0 = Q,_{NN}(0)$, then the limit of $Q,_N \Gamma,_N / Q,_{NN}$ at $N \to 0$ is still negative because of conditions (4) and (5). Consequently, $\lim_{N \to 0} u < 0$ in every case, $\square$. Since the inequality is sharp (actually, $\lim_{N \to 0} u \leq -1/2$), $u$ will be negative already at some $N > 0$. If we approach the center of symmetry along the locus of the outer turning points, $R = R_-$, then $u$ remains positive up to the very center (see the proof in the final part of Appendix \[negli\]). This shows that the negative-$u$ region does not exist permanently, but appears only for finite time intervals around the bounce instant. Thus, a nonsingular bounce of spherically symmetric charged dust is possible only when there exists a region of negative $u$ for a finite time-interval around the bounce instant. We discuss the implications of this fact in the next section. Consequences of $u < 0$ {#negucons} ======================= The only place in the metric where $u$ explicitly appears is via (\[2.5\]), which shows that ${\rm e}^A = \left(R,_r / u\right)^2$. This is insensitive to the sign of $u$. One might thus suspect that it is enough to take ${\rm e}^{A/2} = - R,_r / u$ instead of (\[2.5\]) to cure the problem. However, with such changed ${\rm e}^{A/2}$, the Einstein equations imply that $C,_r = - 2QQ,_N {\rm e}^{A/2} / R^2$ instead of (\[2.6\]) (see the derivations in Refs. [@PlKr2006] and [@KrBo2006]), i.e. $C$ is sensitive to the sign of $u$. The changes propagate through all the equations, and the resulting formula for energy-density, eq. (\[2.10\]), remains unchanged. This means that $\epsilon$ becomes negative when $u < 0$, unless this change of sign of $u$ can be offset by a change of sign of $R,_r$ or $N,_r$. We show below that such an offset is impossible. The radial coordinate $r$ is arbitrary, and all the formulae are covariant under the transformation $r = f(r')$, where $f$ is an arbitrary function. The values of $r$ label flow-lines of the charged dust, all flow lines of the same $r$ form a 3-cylinder in spacetime, whose sections of constant $t$ are 2-spheres. Thus, for logical clarity, it is convenient to assume that the labeling is such that increasing $r$ corresponds to receding from the center of symmetry.[^5] Suppose that $r = r_c$ corresponds to the center of symmetry. Then, $R,_r < 0$ is impossible in a vicinity of $r = r_c$: since $R = 0$ at $r = r_c$, $R,_r < 0$ at $r > r_c$ would imply $R < 0$ in a vicinity of the center, which is a geometrical nonsense. Thus, $R,_r > 0$ at the center, and so $R,_r / u < 0$. But then the energy density will be negative at the center, unless $N,_r < 0$. However, the physical interpretation of eq. (\[2.11\]) suggests that $N,_r$ cannot be negative: since $N = 0$ at the center, $N,_r < 0$ would imply $N < 0$ around the center. The final conclusion is that the energy density $\epsilon / c^2$ must be negative where $u < 0$, at least in some neighbourhood of the center (but see further in this section). Is the set $u = 0$ a singularity? The answer to this question depends not only on $u$, but also on the behaviour of $R,_r$. If $R,_r = 0$ at $u \neq 0$, then this is a shell crossing, which is a curvature singularity. The 9 conditions listed at the end of section \[paper1\] were derived from the requirement that $R,_r \neq 0$ (actually, $R,_r/u \neq 0$) throughout the evolution. All of them are necessary conditions for $R,_r/u \neq 0$. A sufficient condition is not known, but anyway we wish to avoid $R,_r/u = 0$, and will not consider this case here. If $u/R,_r \neq 0$ at $u = 0$, then there is no singularity at this location, but the energy density will be negative where $u/R,_r < 0$. We will now investigate the behaviour of $\epsilon$ in the set $S_{\rm min}$, which is a 3-space composed of those points where $R(t,r)$ assumes its minimal value $R_+$ given by (\[2.13\]). At those values, $R,_t = 0$. We have already shown that $u/R,_r < 0$ (implying $\epsilon < 0$) in a vicinity of the center $R = 0$ in $S_{\rm min}$. Where does $S_{\rm min}$ intersect $u = 0$? At the intersection, from (\[2.13\]) and (\[2.16\]), $R$ must obey two equations: $$\begin{aligned} \label{4.1} R = R_+ &=& - \frac 1 {2E} \left[M - \sqrt{M^2 - 2EQ^2 \left({Q,_N}^2 - G / c^4\right)}\right] \nonumber \\ &=& \frac {QQ,_N} {\Gamma}.\end{aligned}$$ We substitute for $M$ from (\[2.9\]), recall that $\Gamma^2 = 2E + 1$, and solve (\[4.1\]) for $Q,_N$: $$\begin{aligned} \label{4.2} Q,_N &=& \frac {\Gamma} Q \left({\cal M} - \sqrt{{\cal M}^2 - G Q^2 / c^4}\right) \nonumber \\ \Longrightarrow R = R_i &=& {\cal M} - \sqrt{{\cal M}^2 - G Q^2 / c^4}.\end{aligned}$$ This $R_i$ is equal to the radius of the inner Reissner – Nordström event horizon corresponding to the mass shell ${\cal M}$.[^6] Equation (\[4.2\]) should be understood as follows. The functions ${\cal M}(r)$ and $Q(r)$ are arbitrary functions in the model, and $\Gamma$ is determined by ${\cal M}$ via (\[2.8\]). Thus, (\[4.2\]) is an additional condition imposed on functions that are otherwise arbitrary, and the first of (\[4.2\]) may possibly have no solution. However, [*if*]{} it has a solution, then this point must have the $R$ value given by the second of (\[4.2\]). Then we calculate $R,_r$ at the point determined by (\[4.2\]), and obtain, using (\[2.8\]) and the first of (\[4.2\]): $$\label{4.3} \left(R,_r\right)_{R = R_i} = 0.$$ Thus, within the set $S_{\rm min}$ the locus of $u = 0$ (if it exists at all) coincides with the locus of $R,_r = 0$. However, the equations do not allow us to determine the sign of $u/R,_r$ at this location in the general case. In the explicit example given in Sec. 8, $u/R,_r < 0$ throughout $S_{\rm min}$, but this may be a property of that particular model only. In general, we can only say that there exists a finite neighbourhood of the center $R = 0$ within $S_{\rm min}$ in which $u / R,_r < 0$. Analogously, we could consider the set $S_{\rm max}$ consisting of those points where $R = R_-$, i.e. where all the mass shells attain their maximal radii. We showed in Appendix \[negli\] that $u > 0$ in a vicinity of the center within $S_{\rm max}$. At the center, $R,_r > 0$, so $u / R,_r > 0$ in a neighbourhood of the center. The whole calculation above can now be repeated with $R_-$ substituted for $R_+$, and the result will be that also within $S_{\rm max}$ the set $u = 0$ coincides with $R,_r = 0$, but this time at the value $R = R_o = {\cal M} + \sqrt{{\cal M}^2 - G Q^2 / c^4}$. This means that there exists a finite neighbourhood of the center $R = 0$ within $S_{\rm max}$ in which $u / R,_r > 0$, implying $\epsilon > 0$ in the same neighbourhood. All this implies that a volume of charged dust evolves from negative energy density in $S_{\rm min}$ to positive energy density in $S_{\rm max}$. Consequently, the regions of $\epsilon < 0$ and $\epsilon > 0$ are not separated by a comoving hypersurface – the matter particles proceed from one to the other. This should indicate that strong electric fields induce a hitherto unknown physical process inside matter. This concludes the discussion of the case when $u / R,_r \neq 0$ at $u = 0$. Since we have found that $u / R,_r < 0$ in a neighbourhood of $R = 0$ in the hypersuface $R = R_+$ and $u / R,_r > 0$ in a neighbourhood of $R = 0$ in the hypersurface $R = R_-$, there must be some intermediate set between them on which $u / R,_r = 0 = \epsilon$. This set must be a subset of the $u = 0$ set. Note that in a singularity-free model $R,_r$ can vanish only where $u = 0$,[^7] but does not have to vanish everywhere on the $u = 0$ set. Across the set on which $u = 0 \neq R,_r$, the energy density changes sign. In Sec. \[anexample\] we will trace all these boundaries on a numerical example. There is no geometric singularity at $u/R,_r = 0$ {#nosing} ================================================= Now we consider the case when $u / R,_r = 0$ at $u = 0$. This is a singularity of the metric because $\left|g_{11}\right| = \left|{\rm e}^A\right| \to \infty$ there. However, none of the curvature or flow scalars become infinite at those locations, as we show below. The components of tensors referred to below are scalar components with respect to the orthonormal tetrad defined by the metric (\[2.1\]): $$\label{5.1} e^0 = {\rm e}^{C/2} {\rm d} t, \qquad e^1 = {\rm e}^{A/2} {\rm d} r, \qquad e^2 = R {\rm d} \vartheta, \qquad e^3 = R \sin \vartheta {\rm d} \varphi.$$ These components are thus scalars. In this frame, the nonzero components of the Ricci tensor and of the scalar curvature are[^8] $$\begin{aligned} \label{5.2} R_{00} &=& \Lambda + \frac G {c^4}\ \left(\frac {Q^2} {R^4} + \frac {u N,_r} {R^2 R,_r}\right), \nonumber \\ R_{11} &=& R_{22} = R_{33} = - \Lambda + \frac G {c^4}\ \left(- \frac {Q^2} {R^4} + \frac {u N,_r} {R^2 R,_r}\right), \nonumber \\ R &=& 4 \Lambda - \frac {2 G} {c^4}\ \frac {u N,_r} {R^2 R,_r}.\end{aligned}$$ The nonzero tetrad components of the Weyl tensor are $$\begin{aligned} \label{5.3} C_{0 1 0 1} &=& - C_{2 3 2 3} = - 2W, \nonumber \\ C_{0 2 0 2} &=& C_{0 3 0 3} = - C_{1 2 1 2} = - C_{1 3 1 3} = W, \nonumber \\ W &{\ {\overset {\rm def} =}\ }& \frac G {c^4}\ \left(\frac {Q^2} {R^4} + \frac {u N,_r} {3 R^2 R,_r}\right) - \frac {\cal M} {R^3}.\end{aligned}$$ None of these are singular at $u / R,_r = 0$. The scalars defined by the flow – the square of the acceleration vector $\dot{u}^{\alpha} \dot{u}_{\alpha}$, the expansion $\theta = {u^{\alpha}};_{\alpha}$ and the shear $\sigma$ given by $\sigma^2 = \frac 1 2 \sigma^{\alpha \beta} \sigma_{\alpha \beta}$, where $$\label{5.4} \sigma_{\alpha \beta} = u_{(\alpha; \beta)} - \dot{u}_{(\alpha} u_{\beta)} - \frac 1 3 \theta \left(g_{\alpha \beta} - u_{\alpha} u_{\beta}\right),$$ are as follows $$\begin{aligned} \dot{u}^{\alpha} \dot{u}_{\alpha} &=& - \frac {Q^2 {Q,_N}^2} {R^4}, \label{5.5} \\ \theta &=& \frac 1 2 {\rm e}^{- C/2} \left(A,_t + 4 \frac {R,_t} R\right), \label{5.6} \\ \sigma &=& \frac 1 {2 \sqrt{3}} {\rm e}^{- C/2} \left(2 \frac {R,_t} R - A,_t\right). \label{5.7}\end{aligned}$$ The quantities $\dot{u}^{\alpha} \dot{u}_{\alpha}$ and ${\rm e}^{- C/2} R,_t / R$ are seen to be nonsingular at $u = 0$ (the latter from (\[2.7\])), so we investigate ${\rm e}^{- C/2} A,_t$. We find from (\[2.5\]) and (\[2.7\]): $$\begin{aligned} \label{5.8} {\rm e}^{- C/2} A,_t &=& {\rm e}^{- C/2} \left(\frac {2 R,_{t r}} {R,_r} - \frac {2 Q Q,_N R,_t} {u R^2}\right) \nonumber \\ &=& \frac {2 {\rm e}^{C/2}} {R,_t R,_r} \left({\rm e}^{- C} R,_t R,_{t r} - \frac {QQ,_N R,_r} {uR^2}\ {\rm e}^{- C} {R,_t}^2\right) \nonumber \\ &=& \frac {2 {\rm e}^{C/2}} {R,_t} \left\{\left[\Gamma,_r + \frac {GN,_r} {c^4 R} - \frac {{Q,_N}^2 N,_r} R - \frac {QQ,_{NN} N,_r} R\right]\ \frac u {R,_r}\right. \nonumber \\ &-& \left.\left[\frac {{\cal M} - QQ,_N \Gamma} {R^2} + \frac {Q^2 \left({Q,_N}^2 - G/c^4\right)} {R^3}\right]\right\}.\end{aligned}$$ This is seen to be nonsingular at $u / R,_r = 0$. Thus, all the flow scalars are nonsingular at $u / R,_r = 0$, just like all the curvature scalars. This shows that the singularity in $g_{11} = - {\rm e}^A$ is merely a coordinate singularity. Note the following identity $$\label{5.9} u^2 - \Delta = \left(u^R\right)^2 > 0.$$ Thus, at $u = 0$, $\Delta$ must be non-positive. This means, if the region $u \leq 0$ exists, then the outer surface of this region either lies between the two R–N event horizons or touches one of them. (As shown before, the boundary of the region $u \leq 0$ touches the horizons when $u^R = 0$ at $u = 0$.) It is puzzling that this negative-$u$ region is invisible in the mass-curvature coordinates. It is not surprising that the coordinate singularity at $u = 0$ disappears – the mass-curvature coordinates turn thus out to be those that remove it. However, we showed above that at $u / R,_r = 0$ the mass density changes sign, and the change of sign of a scalar is an invariant property that should be visible in all coordinate systems. Equation (\[2.20\]) shows that this can happen only by changing the sign of $F$. But then there is no way in which the $\epsilon$ of (\[2.20\]) could go through zero – $F$ is finite everywhere, except at the turning points, at which $F u^R$ is finite. This seems to imply that the set $u = 0$ is not covered by the mass-curvature coordinates. This is indeed likely. Equations (\[2.5\]) – (\[2.6\]) show that (\[2.6\]) cannot be integrated across the $u = 0$ set. Then, the definition of $F$ via the comoving coordinates is $F = R,_r / \left(uu^R {\cal M},_r\right)$, i.e. the transformation to the $({\cal M}, R)$-coordinates is singular at $u / R,_r = 0$. Thus, $F$ should suffer a discontinuous jump from positive to negative values across the set $u / R,_r = 0$, but this change remains invisible if we use the $({\cal M}, R)$-coordinates throughout. A vacuole around the center of symmetry is no remedy to $u < 0$ {#novac} ============================================================ Since, with the 9 conditions of Sec. \[paper1\], the region $u < 0$ cannot be eliminated by a choice of the arbitrary functions, can we get rid of it by matching the charged dust metric to vacuum across a hypersurface $r =$ const [*on the inside*]{}? i.e. by cutting an empty vacuole around the center of symmetry of the charged dust ball? We shall consider the matching only in the case $\Lambda = 0$. The electrovacuum spacetime matched to spherically symmetric charged dust must be Reissner – Nordström (R–N). However, if the R–N solution is to be applied around the center of symmetry, then it will contain a singularity unless its mass and charge parameters are both zero. In this case, the R–N solution reduces to the Minkowski spacetime. Can the charged dust metric be matched to the Minkowski metric, with all the matching conditions fulfilled? As eq. (\[2.11\]) implies, the matching requires that $Q = {\cal M} = 0$ at the matching sphere $r = r_i$. Then, (\[2.7\]) shows that if $\Gamma(r_i) = 1$, then $R,_t(t, r_i) = 0$, i.e. the charged dust particles that are initially on the matching hypersurface will remain on it permanently. Equation (\[2.6\]) shows that $C,_r(r_i) = 0$, thus $C(t, r_i) = C_0(t)$ and can be made zero, so that ${\rm e}^C = 1$, by a transformation of $t$. This all should happen at a nonzero $R$. For the dust particles remaining permanently on the matching hypersurface, the minimal and maximal radius given by (\[2.13\]) should coincide and both be nonzero. In order to achieve this, the limit of $- M / E$ at $N = 0$ must be positive, and in addition, to avoid $R_+$ being zero, the terms $(M/2E)^2$ and $Q^2 \left({Q,_N}^2 - G / c^4\right)/(2E)$ must be of the same order (so that their sum at $r = r_i$ does not equal $(M/2E)^2$). These conditions are impossible to fulfil – see Appendix \[novacu\]. The conditions of a bounce at a nonzero $R$ can be fulfilled in the cases $E \geq 0$. However, then the arguments used in Refs. [@Ori1990; @PlKr2006] and [@KrBo2006] still hold: even if $Q = {\cal M} = 0$ occurs at $R > 0$, the shell crossings will be unavoidable because the function $F$ necessarily changes sign somewhere in the vicinity of the $Q = 0$ surface. A transient singularity at the center of symmetry {#transient} ================================================= Paper I and the present paper were concerned with avoiding the BB/BC, shell crossings and permanent central singularities. It turns out that there is one more kind of singularity that has slipped through all the tests applied so far. We noted in Sec. \[negucons\] that $R \to 0$ faster than anywhere else if we approach the center of symmetry along the hypersurface $S_{\rm min}$, in which $R = R_+$. The energy density given by (\[2.10\]) will be finite at the center provided the product $u N,_r / \left(R^2 R,_r\right) < \infty$. We have already shown in Sec. \[negdens\] that $u$ is finite at all points of the world line of the center, albeit negative when the center is approached from within $S_{\rm min}$. Thus, to avoid a singularity at all points of the center, we must have $\lim_{r \to r_c} N,_r / \left(R^2 R,_r\right) \equiv \lim_{r \to r_c} 3 / \left(R^3\right),_N < \infty$, i.e. $\lim_{r \to r_c} \left(R^3\right),_N \neq 0$. However, as we show below, $\lim_{r \to r_c} \left(R^3\right),_N = 0$ if the center of symmetry is approached along the hypersurface $S_{\rm min}$, and also $\lim_{r \to r_c} \left(R^3\right),_N = 0$ if the hypersurface $S_{\rm min}$ is approached along the center of symmetry. A nonzero limit of $\left(R^3\right),_N$ at the center, where $R = 0 = N$, means that $R^3$ must tend to zero as fast as $N$, i.e. that $R$ must tend to zero as fast as $N^{1/3}$ and no faster. Let us then take the first of (\[2.27\]). The sufficient condition for $\lim_{r \to r_c} R/N^{1/3} \neq 0$ is then $$\label{7.1} \lim_{N \to 0} \frac M {2E N^{1/3}} \neq 0,$$ which does the job at those points where $\cos \omega \neq 1$. With $\cos \omega = 1$, however, where $R = R_+$, we have $$\label{7.2} R = R_+ = \frac M {(- 2E)} \left[1 - \sqrt{1 - \frac {2EQ^2 \left({Q,_N}^2 - G / c^4\right)} {M^2}}\right].$$ We proved in Appendix \[negli\] that the fraction under the square root always tends to zero at $r \to r_c$ if the other regularity conditions are obeyed. Thus, the expression in square brackets also tends to zero at $r \to r_c$, implying that $R / N^{1/3} {\ {\underset {r \to r_c} {\longrightarrow}}\ } 0$, i.e. a singularity at the center. We can approach the singular point along the world line of the center of symmetry. Assuming that $\lim_{r \to r_c} M / \left(- 2E N^{1/3}\right) = \beta_0 \neq 0$, we obtain $$\label{7.3} \lim_{r \to r_c} \frac {R(t, r)} {N^{1/3}} = \beta_0 (1 - \cos \omega) \neq 0$$ (the dependence on time is now hidden in $\omega$). Taking the limit of (\[7.3\]) as $\cos \omega \to 1$, we again obtain the zero limit of $R / N^{1/3}$ at the feral point. Let us recall: this transient singularity results because the $R_+$ given by (\[7.2\]) goes to zero at $r \to r_c$ faster than $N^{1/3}$. Can we avoid the singularity by forcing $R_+$ to go to zero at $r \to r_c$ as fast as $N^{1/3}$ or slower? The following simple consideration provides a negative answer. Equation (\[7.2\]) can be equivalently written as $$\label{7.4} R_+ = - \frac {Q^2 \left({Q,_N}^2 - G / c^4\right)} {M \left[1 + \sqrt{1 - {\displaystyle \frac {2EQ^2 \left({Q,_N}^2 - G / c^4\right)} {M^2}}}\right]}.$$ We can now use the reasoning in Appendix \[negli\] to show that $\lim_{N \to 0} R_+/N^{1/3} = 0$ independently of the forms of the functions $Q$, $E$ and $M$, provided conditions (1) – (9) of sec. 2 are obeyed. Equation (\[b.3\]) tells us what forms $p$ and $\left({p,_x}^2 - 1\right)$ must have in a vicinity of the center. Thus we have $$\begin{aligned} \label{7.5} &-&\frac {Q^2 \left({Q,_N}^2 - G / c^4\right)} M \nonumber \\ &=& \frac {x^2 \left(1 + A_p x^{\alpha_p - 1} + {\cal O}_{\alpha_p - 1}\right)^2 \left(2 A_p x^{\alpha_p - 1} + {\cal O}_{\alpha_p - 1}\right)} {A_p \left(\alpha_p + 1\right) x^{\alpha_p} + {\cal O}_{\alpha_p}} \nonumber \\ &=& \frac {2x} {\alpha_p + 1} + {\cal O}_1 \equiv \frac {2N} {N_0 \left(\alpha_p + 1\right)} + {\cal O}_1.\end{aligned}$$ Because of (\[7.5\]) and (\[7.1\]), the expression under the square root in (\[7.4\]) has the limit 1 at $r \to r_c$. Consequently $$\label{7.6} \frac {R_+} {N^{1/3}} = \frac {N^{2/3}} {N_0 \left(\alpha_p + 1\right)} + {\cal O}_{2/3} {\ {\underset {N \to 0} {\longrightarrow}}\ } 0,$$ $\square$. Note that since $u < 0$ as the singularity is approached from within $S_{\rm min}$, and at the center necessarily $R > 0$ and $R,_N > 0$, the density in the singularity becomes [*minus infinity*]{}. At the same time, if we start at a point of the center of symmetry different from the maximal compression instant, then, from (\[2.21\]) – (\[2.24\]) and ${Q,_N}^2(r_c) = G / c^4$ we have $u(r_c) = 1 > 0$ everywhere on $N = 0$, so $\epsilon > 0$ everywhere on $N = 0$, including the limiting point of maximal compression. Consequently, $\epsilon \to + \infty$ at the point where the center of symmetry hits the maximal compression hypersurface. This singularity is thus direction-dependent. This transient singularity exists within each $S_{\rm min}$ hypersurface. However, with all other conditions of regularity obeyed, the limit of the period of oscillations of the solution (\[2.27\]) at $r \to r_c$ is infinite, as will be shown below. Thus, the transient singular point can be reached only once, where the bounce set $s = s_B({\cal M})$ hits the center of symmetry at a finite $s$. Every other hypersurface $S_{\rm min}$ escapes to the infinite future or the infinite past when it approaches the center of symmetry. Here is the proof that $T_p {\ {\underset {r \to r_c} {\longrightarrow}}\ } \infty$, where $T_p$ is the period of oscillation given by (\[2.28\]). We know that $M / (- 2E)$ must tend to zero at the center as fast as $N^{1/3}$. We know from (\[2.33\]) that $M$ tends to zero faster than $N$, so let $M \propto N^{1 + \varepsilon}$, where $\varepsilon > 0$. Consequently, $(- 2E)$ tends to zero as fast as $N^{2/3 + \varepsilon}$. This means that $T_p = M / (- 2E)^{3/2}$ behaves as $N^{- \varepsilon / 2}$, i.e. tends to infinity at the center. $\square$ Since we were interested in sending a ball of dust through the Reissner – Nordström throat, we set up the initial conditions so that the state of maximal compression (minimal size) was attained simultaneously by all mass shells. In such a configuration, illustrated in Figs. 12 and 13 of Paper I (and also in Fig. \[drawcycles2\] further here), the hypersurfaces $S_{\rm max}$ (consisting of the instants of maximal size of all the shells) approach the center of symmetry in the infinite past and in the infinite future. Consequently, it is not in general guaranteed that the center of symmetry will ever reach the region where $u > 0$ and $\epsilon > 0$, but in our explicit example, given in the next section, the region $u < 0$ is contained in a small neighbourhood of $S_{\rm min}$. However, we could take exactly the same example and set up the initial conditions so that the state of minimal compression (maximal size) is attained simultaneously. Then the dust ball would never emerge from the R–N throat because shell crossings would appear immediately after the bounce. However, initially and for some (perhaps infinite) time to the future and to the past, the center of symmetry would be in the region of positive $u$ and $\epsilon$. An example {#anexample} ========== To illustrate the statements of the preceding sections we will use the same example that we introduced in Paper I. To recall, in the example the function $N$ was (and still will be) used as the radial coordinate. In the exemplary configuration, the charge function was $$\label{8.1} Q(N) = q \frac {\sqrt{G} N_0} {c^2}\ p(x),$$ where $q = \pm 1$, to allow for any sign of the charge, $x {\ {\overset {\rm def} =}\ }N/N_0$, $N_0$ is an arbitrary constant, and $$\label{8.2} p(x) = x / (1 + x)^2.$$ Then $$\label{8.3} Q,_N = q \frac {\sqrt{G}} {c^2}\ \frac {1 - x} {(1 + x)^3},$$ and $$\label{8.4} F_1(x) {\ {\overset {\rm def} =}\ }1 - \frac {c^4} G\ \left({Q,_N}^2 + QQ,_{NN}\right) = 1 - \frac {3x^2 - 6x + 1} {(1 + x)^6}.$$ The original definition of the function $E(x)$ was incorrect, as it implied a permanent central singularity. We quote here the corrected $E(x)$ and the formulae dependent on it, from the gr-qc version of Paper I. The functions defined along the way obey conditions (1) – (9) of Sec. \[paper1\] – for proofs see the gr-qc version of Paper I. Thus $$\label{8.5} 2E = - \frac {b x^{5/3}} {1 + b x^{5/3}},$$ where $b$ is another arbitrary constant. With such $E$ we have $$\label{8.6} \Gamma(x) = \frac 1 {\sqrt{1 + b x^{5/3}}},$$ $$\label{8.7} {\cal M}(x) = \frac {GN_0} {c^4} \int_0^x \frac {{\rm d} x'} {\sqrt{1 + b {x'}^{5/3}}} {\ {\overset {\rm def} =}\ }\frac {GN_0} {c^4} \mu(x).$$ Then, further: $$\begin{aligned} \label{8.8} M &\equiv& {\cal M} - QQ,_N \Gamma = \frac {GN_0} {c^4}\ F_2(x), \nonumber \\ F_2(x) &{\ {\overset {\rm def} =}\ }& \mu(x) - \frac {x (1 - x)} {(1 + x)^5 \sqrt{1 + b x^{5/3}}}.\end{aligned}$$ For checking condition (7) we need the function $F_3(x)$ defined below. This condition is equivalent to $$\label{8.9} F_3(x) > 0, \qquad F_3(x) {\ {\overset {\rm def} =}\ }{F_2}^2(x) - F_6(x),$$ where $$\label{8.10} F_6(x) {\ {\overset {\rm def} =}\ }- 2E p^2 \left(1 - {p,_x}^2\right) = \frac {b x^{11/3}} {\left(1 + b x^{5/3}\right) (1 + x)^4}\ \left[1 - \frac {(1 - x)^2} {(1 + x)^6}\right].$$ Fig. \[chadu6fig\] shows the graphs of the functions defined above with $b = 2.5$ (why this value – see below). This is a corrected version of Fig. 7 from Paper I. ${}$\ ${}$\ For checking the remaining conditions we need the functions $F_4(x)$ and $F_5(x)$ defined below. Condition (8) is equivalent to: $$\label{8.11} F_4(x) > 0, \qquad {\rm where} \qquad F_4(x) {\ {\overset {\rm def} =}\ }F_1(x) - \frac 5 {6x \sqrt{1 + b x^{5/3}}}\ F_2(x),$$ and condition (9) is equivalent to: $$\label{8.12} F_5(x) > 0, \qquad {\rm where}\ F_5(x) {\ {\overset {\rm def} =}\ }F_1(x) - \frac 5 {6 x \sqrt{1 + b x^{5/3}}} \left[F_2(x) + \sqrt{F_3(x)}\right].$$ Fig. \[chadu7fig\] shows the graphs of $F_4(x)$ and $F_5(x)$ with $b = 2.5$. The reason for choosing $b = 2.5$ was this: the graph of $F_5$ is sensitive to the value of $b$. With decreasing $b$, the local minimum of $F_5$ becomes smaller, and for $b$ small enough (for example, $b = 0.75$) $F_5 < 0$ around the minimum. With $b \geq 2.5$, the minimum is clearly positive. The value $b = 2.5$ was used in all the subsequent figures. Figs. 9 and 10 of Paper I do not change in any noticeable way after the correction done in $E(x)$, so they need not be repeated here (see the gr-qc version of Paper I). However, Figs. 11 and 12 do change significantly (also in consequence of the different value of $b$ adopted here), and their correct versions are shown in Figs. \[plotperiod\] and \[drawcycles2\] here. Fig. \[drawcycles2\] also shows the light cone of the transient singularity at $(t, R) = (t_B, 0)$ – to demonstrate that all mass shells enter this cone soon after going through the minimal size. The other qualitative conclusions presented in Paper I remain unchanged. Below, we present the new results found in the present paper. In all the subsequent figures, the radial coordinate $r$ was defined as $r = N^{1/3}$, so that one can instantly see whether there is a central singularity ($R,_r = 0$ at the center) or not ($R,_r > 0$ at the center). Fig. \[uRpratR+\] shows the functions $u(x)$ and $R,_r(x)$ (left graph) and the energy density $\epsilon(x)$ (right graph) in the surface $R = R_+$. The functions $u(x)$ and $R,_r(x)$ have a zero at the same value of $x$ and opposite signs in all other points, except at $x = 0$ where $R,_r \to 0$ and $u < 0$, so $u / R,_r \to - \infty$. As follows from the calculations in Secs. \[negucons\] and \[transient\], in this surface the energy density is negative in a vicinity of the center and goes to $- \infty$ at the center. The fact that the point where $u = 0 = R,_r$ does not seem to be singular is rather mysterious because it is a limiting point of the contour $R,_r = 0$ in the $(t, x)$-plane, and the rest of the contour is a shell crossing (see Fig. \[u=0\]), where the energy density does go to $\pm \infty$. But the curves in Fig. \[drawcycles2\] are all for smaller values of $x$, so the singularity is not in its range. ${}$\ Fig. \[uRpratR-\] shows the functions $u(x)$ and $R,_r(x)$ (left graph) and the energy density $\epsilon(x)$ (right graph) in the surface $R = R_-$. Within this surface $u(x)$ and $R,_r(x)$ are everywhere positive.[^9] The energy density is everywhere positive and finite (it does not tend to zero, but to a small positive value as $r \to 0$, as shown in the inset). ${}$\ Fig. \[u=0\] shows the curves $u = 0$ (lower part) and $R,_r = 0$ (upper part) in the $(t, x)$-plane (actually, each curve is one half of the full contour, which is mirror-symmetric with respect to the $x$-axis). The function $u$ is negative to the left of the lower contour and positive to the right of it. The function $R,_r$ is negative to the left of the upper contour and positive to the right of it. As stated above, the set $R,_r = 0$ is a shell crossing. Its presence proves that the 9 necessary conditions listed at the end of Sec. \[paper1\] were not sufficient. Note the characteristic features of the curves, consistent with the calculations of Secs. \[negucons\] and \[nosing\]: 1\. The lower curve hits the $t$-axis (the center of symmetry) at $t = 0$ (in the figure, we chose $t_B = 0$, so $t = 0$ is the simultaneously achieved state of minimal size). Thus $u$ remains positive all the time as we proceed toward $t = 0$ along the center of symmetry. 2\. This curve intersects the $x$-axis at some $x > 0$. This intersection point coincides with the point where the $u(x)$ and $R,_r(x)$ curves intersect in Fig. \[uRpratR+\]. The upper curve begins at the same point. 3\. To the right of the lower curve, we have $u > 0$, and to the right of the upper curve $R,_r > 0$. The energy density is positive in the area to the right of both curves, and negative to the left of any of them, including the $x$-axis. In order to better visualise the variation of the functions $u$, $R,_r$ and $\epsilon$ in the $(t, R)$-plane, we provide three further graphs. Fig. \[uRpratsmalt\] shows the functions $u(x)$ and $R,_r(x)$ along the line $t = 0.004$ in Fig. \[u=0\]. Fig. \[den2atsmalt\] shows the energy density along the same section. It changes sign every time when the line $t = 0.004$ crosses one contour or the other. Where $u = 0$, it changes from negative to positive by smoothly going through zero; where $R,_r = 0$ it changes sign by jumping from $+ \infty$ to $- \infty$ or the other way round. Fig. \[uRpratmidt\] shows the same functions along the line $t = t_2$, where $t_2$ lies between $t = 0.004$ and the right end of the the $u = 0$ contour in Fig. \[u=0\]; $t_2 = 0.025$. Finally, Fig. \[uRprdeninter\] shows the functions $u(x)$, $R,_r(x)$ and $\epsilon(x)$ along the line $t = t_3$, where $t_3$ lies completely to the right of the $u = 0$ contour in Fig. \[u=0\]. In each case, the functions behave exactly in the way in which Fig. \[u=0\] implies they should. ${}$\ Fig. 13 of Paper I, which showed the schematic Penrose diagram of the evolution of our exemplary model, is still qualitatively correct. However, now it has to be supplemented with a few new elements – the images of the surfaces $u = 0$ and the light cone of the point singularity at the center. The complete version of that diagram is shown in Fig. \[chadumax2\], and here is the explanation. The diagram is written into the background of the Penrose diagram for the maximally extended Reissner–Nordström spacetime (thin lines). C is the center of symmetry, Sb is the surface of the charged ball, S$_{\rm RN}$ is the Reissner–Nordström singularity. The interior of the body is encompassed by the lines C, E, Sb and B; the only singularity that occurs within this area is the point singularity at the center, marked PS. The dotted lines issuing from PS mark the approximate position of the future and past light cone of the singularity (i.e. the Cauchy horizon of the nonsingular part of the surface $R = R_+$). Lines B and E connect the points in spacetime where the shell crossings occur at different mass shells. N1 (N2) are the past- (future-) directed null geodesics emanating from the points in which the shell crossings reach the surface of the body (compare Fig. \[drawcycles2\]). The line Sb should be identified with the uppermost curve in Fig. \[drawcycles2\]. The top end of Sb is where the corresponding curve in Fig. \[drawcycles2\] first crosses another curve, the middle point of Sb is at $t = 0$ in Fig. \[drawcycles2\]. The two dotted arcs marked “$u$” symbolise the hypersurfaces $\epsilon = 0$, on which the energy density changes from positive (below the lower curve) to negative and then to positive again (above the upper curve). The distance between these arcs is greatly exaggerated; at the proper scale of the figure they would seem to coincide. The arcs are parts of the $u = 0$ contour shown in Fig. \[u=0\]. In Figs. \[drawcycles2\] and \[chadumax2\], the surface of the body is at smaller $x$ than the top of the $u = 0$ contour in Fig. \[u=0\]. Every $N \neq 0$ world line of the dust necessarily enters the region of negative energy density for a finite time interval around the instant of maximal compression. Permanent avoidance of singularity is impossible {#perm} ================================================ In Paper I we hypothesised that a permanently nonsingular configuration of charged dust might exist – provided the period of oscillations is independent of the mass ${\cal M}$. We show below that in our class of configurations the period cannot be independent of ${\cal M}$ because conditions (1) – (9) prohibit this. The condition $T_{c,{\cal M}} = 0$, if considered by explicitly differentiating (\[2.29\]), leads to a very complicated integral equation that we were not able to handle. (The integral in (\[2.29\]) cannot be explicitly calculated without knowing the explicit form of $C({\cal M}, R)$.) However, let us recall that both $T_p$ given by (\[2.28\]) and $T_c$ given by (\[2.29\]) are in general functions of ${\cal M}$ only. Thus, we conclude that $T_c$ is a function of $T_p$, and so the condition for the special case of $T_c$ being constant is $$\label{9.1} {\frac {{\rm d} {T_c}} {{\rm d} {\cal M}}} = {\frac {{\rm d} {T_c}} {{\rm d} {T_p}}} {\frac {{\rm d} {T_p}} {{\rm d} {\cal M}}} = 0.$$ Hence, $T_c$ will be constant when $$\label{9.2} {\frac {{\rm d} {T_p}} {{\rm d} {\cal M}}} = 0 \Longleftrightarrow {\frac {{\rm d} {T_p}} {{\rm d} {N}}} = 0,$$ since (\[2.21\]) implies that $N,_{\cal M} \neq 0$. From (\[2.28\]), (\[2.9\]) and $T_p = C =$ const we then find $$\label{9.3} (- 2E)^{3/2} \equiv \left(1 - \Gamma^2\right)^{3/2} = C^{3/2} \left({\cal M} - QQ,_N \Gamma\right).$$ From here $$\label{9.4} QQ,_N = - \frac 1 {\Gamma} \left(\frac {1 - \Gamma^2} C\right)^{3/2} + \frac {\cal M} {\Gamma}.$$ Given $\Gamma(N)$ and ${\cal M} (N)$ found from (\[2.21\]), this determines $Q(N)$ simply by integration. We now verify whether (\[9.4\]) is consistent with the regularity conditions (\[2.22\]) – (\[2.25\]) and the 9 conditions listed after (\[2.30\]). From (\[2.28\]) we see that $T_p = C =$ const implies $$\label{9.5} \lim_{{\cal M} \to 0} \frac M {\cal M} = \frac {D^{3/2} C} {2 \pi},$$ where, as stated after (\[2.5\]), $D$ may be zero. Then, comparing (\[2.23\]) with (\[2.27\]), we see that $$\label{9.6} \lim_{{\cal M} \to 0} \frac M {(-2 E) {\cal M}^{1/3}} = \alpha \neq 0, \qquad \alpha < \infty,$$ where $\alpha$ is a constant. This, together with (\[2.25\]), means $$\label{9.7} \lim_{{\cal M} \to 0} \frac M {D {\cal M}} = \alpha < \infty,$$ i.e. that $D \neq 0$. But with $CD \neq 0$, eq. (\[9.5\]) is in contradiction with $\lim_{{\cal M} \to 0} M / {\cal M} \neq 0$ (see the remark at the end of Sec. \[paper1\]). We have thus proven that a configuration with $E < 0$ cannot pulsate with the period of pulsations being independent of the mass ${\cal M}$, while being singularity-free for ever. Different periods for different mass shells will necessarily cause shell crossings, during the second cycle at the latest (Fig. \[drawcycles2\] illustrates this). Permanently nonsingular oscillations of weakly charged dust, with the period independent of mass, are possible only when there is a central singularity. An example of such a configuration results when $\Gamma = 1 / \cosh(bx)$, where $b$ is a constant. A spherically symmetric charged dust configuration can be permanently nonsingular only if it is static. Such configurations do indeed exist with special forms of the arbitrary functions, as pointed out in Refs [@PlKr2006] and [@KrBo2006], but they are not interesting from the point of view of avoiding singularities. Conclusions {#conclu} =========== The conclusion of this paper is: the weakly charged spherically symmetric dust distribution considered here (${Q,_N}^2 < G / c^4$ at $N > 0$ and ${Q,_N}^2 = G / c^4$ at $N = 0$) must contain at least one of the following features: 1\. A Big Bang/Big Crunch singularity; 2\. A permanent central singularity; 3\. A shell crossing singularity in a vicinity of the center 4\. A finite time interval around the bounce instant in which the energy density becomes negative, and a transient momentary singularity of infinite energy density at a single point on the world line of the center of symmetry. A fully nonsingular bounce is possible for a strongly charged configuration, ${Q,_N}^2 > G/c^4$, which can exist only with $E > 0$ – see the comments after (\[2.14\]). This would be a collapse followed by a single bounce and re-expansion to infinite size. This phenomenon occurs also in the Newtonian limit – the bounce here is caused by the prevalence of electrostatic repulsion over gravitational attraction. Such an example was reportedly found, but never published, by Ori and coworkers [@Oripriv]. A permanently nonsingular pulsating configuration of spherically symmetric charged dust does not exist. At most, a single full cycle of nonsingular bounce can occur, and shell crossings will necessarily appear during the second collapse phase. The nonsingular bounce occurs at $R > 0$, but the momentary isolated singularity at the center of symmetry is still there. The possibility of $\epsilon$ going negative in the presence of electric charges does not seem to have been noticed and may need further work on its interpretation. If the negative energy density region existed permanently in some part of the space, with comoving boundary, then we might suspect that this is a consequence of a bad choice of parameters that implies unphysical properties in that region. However, here we have the case in which the energy density is positive for some time, and then these same matter particles acquire negative energy density in a time-interval around the bounce instant. This suggests that there is some physical process involved in this, which should be further investigated. A question arises now. The uncharged limit of the family of configurations defined by eqs. (\[2.1\]) – (\[2.10\]), $Q = 0$, is the Lemaître – Tolman (L–T) model [@Lema1933; @Tolm1934]. For the latter, shell crossings can be avoided, as is well-known since long ago [@HeLa1985]. Why, then, are they unavoidable with $Q \neq 0$? The answer is this: in the L–T model, the Big Bang (BB) or Big Crunch (BC) are unavoidable (both are present when $E < 0$). In the cases that are colloquially called “free of shell crossings”, in reality the shell crossings are not removed, but shifted to the epoch before the BB or after the BC, or both. Thus, the shell crossings are no longer in the domain of physical applicability of the model. When the BB/BC singularities are replaced with a smooth bounce in charged dust, the shell crossings that were hidden on the other side of BB/BC become physically accessible, and they terminate the evolution of the configuration. The proof that $R_+(N)$ coincides with $H_{\rm min}(N)$ at the local extrema of $H_{\rm min}(N)$ {#extrem} ================================================================================================ We stated in the caption to Fig. 10 in Paper I, that the curves $H_{\rm min}(N)$ and $R_+(N)$ were not really tangent at the point where $H_{\rm min}(N)$ has its maximum, but just close to one another. In fact, they not only were tangent at that point, but [*had to*]{} be tangent, independently of the explicit forms of the two functions. We show here that this is a general law: at every local extremum of $H_{\rm min}(N)$ (call it $N_e$) we have $H_{\rm min}(N_e) = R_+(N_e)$. The radius of the inner R–N horizon as a function of $N$ is $$\label{a.1} H_{\rm min}(N) = {\cal M} - \sqrt{{\cal M}^2 - GQ^2/c^4},$$ where ${\cal M}$ and $Q$ are assumed to depend on $N$ through $x = N/N_0$. We do not assume any explicit form of the functions ${\cal M}$ and $Q$, we only use the general properties (\[2.9\]), (\[2.15\]) and (\[2.23\]). At every local extremum $N = N_e$ we have $H_{{\rm min},N} = 0$. Using (\[2.23\]), this means $$\label{a.2} \left[\Gamma \sqrt{{\cal M}^2 - GQ^2/c^4}\right]_e = \left[{\cal M} \Gamma - QQ,_N\right]_e,$$ where the subscript $e$ denotes the value at the extremum. From here $$\label{a.3} \left[\frac {GQ^2} {c^4}\right]_e = \left[\frac {QQ,_N} {\Gamma^2} \left(2 {\cal M} \Gamma - QQ,_N\right)\right]_e.$$ Substituting this value of $GQ^2/c^4$ in (\[a.1\]) wet get $$\label{a.4} H_{\rm min}(N_e) = QQ,_N/\Gamma.$$ Substituting the same value of $GQ^2/c^4$ in (\[2.13\]) for $R_+$ we find $$\label{a.5} R_+(N_e) = QQ,_N/\Gamma = H_{\rm min}(N_e).$$ $\square$ Proof of (\[3.2\]) {#negli} ================== For simplicity, to avoid physical coefficients, we introduce the functions $p(x)$, $\mu(x)$ and $F_2(x)$ (where $x = N/N_0$, $N_0 =$ const) by $$\label{b.1} Q(N) {\ {\overset {\rm def} =}\ }\pm \frac {\sqrt{G} N_0} {c^2}\ p(x), \qquad {\cal M} {\ {\overset {\rm def} =}\ }\frac {GN_0} {c^4}\ \mu(x) \qquad M(N) {\ {\overset {\rm def} =}\ }\frac {GN_0} {c^4}\ F_2(x).$$ Then, to prove (\[3.2\]) we must prove that $$\label{b.2} \lim_{x \to 0} \left[-2E p^2 \left({p,_x}^2 - 1\right)\right]/{F_2}^2 = 0.$$ By condition (5), in a vicinity of $x = 0$, $p$ must be of the form $$\label{b.3} p(x) = x + A_p x^{\alpha_p} + {\cal O}_{\alpha_p}, \Longrightarrow {p,_x}^2 - 1 = 2A_px^{\alpha_p - 1} + {\cal O}_{\alpha_p - 1},$$ where $A_p$ and $\alpha_p > 1$ are constants, and ${\cal O}_{\alpha_p}$ is an unspecified function with the property $\lim_{x \to 0} {\cal O}_{\alpha_p}/x^{\alpha_p} = 0$. Then, by the regularity condition (\[2.24\]), $$\label{b.4} \Gamma = 1 + A_E x^{\alpha_E} + {\cal O}_{\alpha_E} \Longrightarrow -2E = -2 A_E x^{\alpha_E} + {\cal O}_{\alpha_E},$$ where $\alpha_E \geq 2/3$, by (\[2.25\]). It follows from (\[2.21\]) and (\[2.8\]) – (\[2.9\]) that $$\label{b.5} \mu(x) = x + A_m x^{\alpha_E + 1} + {\cal O}_{\alpha_m}.$$ Hence, $$\label{b.6} F_2 \equiv \mu - pp,_x \Gamma = \left(A_m - A_E\right) x^{\alpha_E + 1} - A_p \left(\alpha_p + 1\right)x^{\alpha_p} - {\cal O}_{\alpha_p} - {\cal O}_{\alpha_E + 1}.$$ Thus $F_2$ is of order $\alpha_p$ or $\alpha_E + 1 \geq 5/3$, whichever limit is lower. However, the limit via $\alpha_E$ leads to a central singularity, since then, from (\[2.27\]), $R/N^{1/3} \to 0$ as $N \to 0$, so (\[2.23\]) will not hold. Consequently, $F_2$ is of order $\alpha_p$, while the second term under the square root in (\[3.1\]), $-2E p^2 \left({p,_x}^2 - 1\right)$ is of order $\alpha_2 = \alpha_E + \alpha_p + 1$. Suppose that $\alpha_2 \leq 2 \alpha_p$, so that the second term under the square root in (\[3.1\]) is of lower order than the first one. This translates to $\alpha_p \geq \alpha_E + 1$. But condition (\[2.23\]), in combination with (\[2.27\]), requires that $\alpha_p = \alpha_E + 1/3$, or else a central singularity will appear. Thus, $-2E p^2 \left({p,_x}^2 - 1\right)$ must be of higher order in $x$ than ${F_2}^2$ in eq. (\[3.1\]), i.e. (\[3.2\]) must hold. $\square$ Now consider the case when we approach the center of symmetry along the locus of the outer turning points, $R = R_-$. Then $u_- = \Gamma - QQ,_N / R_-$ remains positive up to the center, as is seen from the reasoning below. The conclusion that $-2E p^2 \left({p,_x}^2 - 1\right)$ is of higher order in $x$ than ${F_2}^2$ still applies, so we have $$\begin{aligned} \label{b.7} \lim_{N \to 0} u_- &\equiv& \lim_{N \to 0} \left[\Gamma + \frac {2EQQ,_N} {M + \sqrt{M^2 - 2EQ^2 \left({Q,_N}^2 - G/c^4\right)}}\right] \nonumber \\ &=& 1 + Q,_N(0) \lim_{N \to 0} EQ/M = 1 + Q,_N(0) \lim_{x \to 0} Ep/F_2.\end{aligned}$$ We know from the above that $p(x) = x + A_p x^{\alpha_p} + {\cal O}_{\alpha_p}$, $F_2 = - A_p \left(\alpha_p + 1\right)x^{\alpha_p} - {\cal O}_{\alpha_p}$ and $E = A_E x^{\alpha_E} + {\cal O}_{\alpha_E}$, while the regularity condition (\[2.23\]) implies, via (\[2.27\]), that $\alpha_E = \alpha_p - 1/3$. This, taken together, implies that $ \lim_{x \to 0} Ep/F_2 = 0$, i.e. that $\lim_{N \to 0} u_- = 1 > 0$, i.e. that $u_-$ remains positive up to the very center. Impossibility of a vacuole around $R = 0$ with $E < 0$ {#novacu} ====================================================== Let us assume $E(x)$ as in (\[b.4\]) and $p(x)$ in a form similar to (\[b.3\]): $$\label{c.1} p(x) = \alpha x + A_p x^{\alpha_p} + {\cal O}_{\alpha_p}, \Longrightarrow p,_x = \alpha + A_p \alpha_p x^{\alpha_p - 1} + {\cal O}_{\alpha_p - 1},$$ For the function $\mu(x)$ this implies $$\label{c.2} \mu = x + \frac {A_E} {\alpha_E + 1} x^{\alpha_E + 1} + {\cal O}_{\alpha_E + 1}.$$ Then $$\begin{aligned} \label{c.3} F_2(x) &\equiv& \mu - p p,_x \Gamma = \left(1 - \alpha^2\right) x - A_E \left(2 \alpha^2 - \frac 1 {\alpha_E + 1}\right) x^{\alpha_E + 1} - \alpha A_p \left(\alpha_p + 1\right) x^{\alpha_p} \nonumber \\ &+& {\cal O}_{\alpha_p} + {\cal O}_{\alpha_E + 1}.\end{aligned}$$ Now we have two cases: $\alpha^2 \neq 1$ and $\alpha^2 = 1$. In the first case we have $F_2 = \left(1 - \alpha^2\right) x + {\cal O}_{1}(x)$. Thus, in order that $M / E$ has a nonzero limit at $x \to 0$, it follows that $\alpha_E = 1$. Then, if the terms $(M/2E)^2$ and $Q^2 \left({Q,_N}^2 - G / c^4\right)/(2E)$ in (\[2.13\]) are to be of the same order, the ratio $p^2 \left({p,_x}^2 - 1\right) / (2E)$ should have a nonzero finite limit at $x \to 0$. But, with $\alpha^2 \neq 1$ and $\alpha_E = 1$ we have $$\begin{aligned} \label{c.4} &&\frac {p^2 \left({p,_x}^2 - 1\right)} {2E} = \frac {x^2 \left(\alpha + A_p x^{\alpha_p - 1} + {\cal O}_{\alpha_p - 1}\right)^2 \left(\alpha^2 - 1 + 2 A_p \alpha_p x^{\alpha_p - 1} + {\cal O}_{\alpha_p - 1}\right)} {2 A_E x + {\cal O}_{1}} \nonumber \\ && {\ {\underset {x \to 0} {\longrightarrow}}\ } 0,\end{aligned}$$ which has the consequence that $R_+ = 0$, i.e. the conditions $Q = {\cal M} = 0$ cannot be imposed at $R > 0$. In the second case, $\alpha^2 = 1$, we can assume $\alpha = 1$ (nothing in the equations depends on the sign of $Q$); then $F_2$ is of the order of either $x^{\alpha_E + 1}$ or $x^{\alpha_p}$, whichever exponent is smaller. With $\left(\alpha_E + 1\right)$ being smaller, the result is immediately seen: If $F_2$ is of the order $x^{\alpha_E + 1}$ while $E$ is of the order $x^{\alpha_E}$, then $\lim_{x \to 0} M/E = 0$, and $\lim_{x \to 0}R_+ = 0$. With $\alpha_p$ being smaller we have $$\label{c.5} \frac {p^2 \left({p,_x}^2 - 1\right)} {2E} = \frac {x^{\alpha_p + 1} \left(1 + A_p x^{\alpha_p - 1} + {\cal O}_{\alpha_p - 1}\right)^2 \left(2 A_p \alpha_p + {\cal O}_{0}\right)} {2 A_E x^{\alpha_E} + {\cal O}_{\alpha_E}}.$$ This will have a finite limit at $x = 0$ if $\alpha_E = \alpha_p + 1$. But we are considering the case when $F_2$ is of order $x^{\alpha_p}$, which means that the limit of $F_2/E$, i.e. the limit of $M/E$ at $x \to 0$ will be infinite, which again shows that the matching conditions cannot be fulfilled at a finite nonzero $R$. $\square$ [**Acknowledgement**]{} We are grateful to the referee for pointing out several errors in an earlier version of this paper. [aa]{} A. Krasiński, K. Bolejko, [*Phys. Rev.*]{} [**D73**]{}, 124033 (2006) + erratum [*Phys. Rev.*]{} [**D75**]{}, 069904 (2007). Corrected version: gr-qc 0602090. A. Ori, [*Class. Q. Grav.*]{} [**7**]{}, 985 (1990). A. Ori, [*Phys. Rev.*]{} [**D44**]{}, 2278 (1991). J. Plebański and A. Krasiński, [*An Introduction to General Relativity and Cosmology*]{}. Cambridge University Press 2006. P. A. Vickers, [*Ann. Inst. Poincarè*]{} [**A18**]{}, 137 (1973). A. Ori, private communication. A. Krasiński, [*Gen. Rel. Grav.*]{} [**33**]{}, 145 (2001). G. Lemaître, [*Ann. Soc. Sci. Bruxelles*]{} [**A53**]{}, 51 (1933); English translation, with historical comments: [*Gen. Rel. Grav.*]{} [**29**]{}, 637 (1997). R. C. Tolman, [*Proc. Nat. Acad. Sci. USA*]{} [**20**]{}, 169 (1934); reprinted, with historical comments: [*Gen. Rel. Grav.*]{} [**29**]{}, 931 (1997). C. Hellaby and K. Lake, [*Astrophys. J.*]{} [**290**]{}, 381 (1985) + erratum in [*Astrophys. J.*]{} [**300**]{}, 461 (1985). [^1]: This research was supported by the Polish Ministry of Education and Science grant no 1 P03B 075 29. [^2]: This error has already been pointed out in an erratum to Paper I, and a fully corrected text is available from the gr-qc archive, see Ref. [@KrBo2006]. [^3]: $R(t, r) = 0$ permanently at the center of symmetry. [^4]: This includes also the case $Q,_N = 0$, i.e. zero charge density, provided $Q \neq 0$, i.e. nonzero total charge. Such a configuration is neutral dust moving in the exterior electric field of a spherically symmetric source. Also in this case, the BB/BC singularity is avoided. [^5]: This order of labels is reversed on the other side of a shell crossing, but we are considering models without shell crossings. [^6]: Formally, the solution of (\[4.1\]) includes also the outer R–N horizon, with “+” in front of the square root. However, here we consider eq. (\[4.1\]) within the set $S_{\rm min}$ that consists of the inner turning points of the various mass shells. Within this set, the equation $R = {\cal M} + \sqrt{{\cal M}^2 - G Q^2 / c^4}$ has no solutions, as is known from the general properties of the R–N solution [@PlKr2006]. [^7]: Because $R,_r = 0 \neq u$ would be a shell crossing that was excluded by Conditions (1) – (9). [^8]: We substituted the solutions of the Einstein – Maxwell equations for $C$ and $A$ from (\[2.5\]) – (\[2.9\]). These formulae were calculated by the computer-algebra system Ortocartan [@Kras2001]. [^9]: We recall that the surface $R = R_-$ in our example never intersects the worldline of the center of symmetry, but approaches it asymptotically as $t \to + \infty$ or $t \to - \infty$, depending on whether it lies to the future or to the past of $R = R_+$.

--- abstract: 'We present new far-infrared ($70-500\,\mu$m) Herschel PACS and SPIRE imaging observations as well as new mid-IR Gemini/T-ReCS imaging (8.7 and $18.3\,\mu$m) and spectroscopy of the inner Lindblad resonance (ILR) region ($R<2.5\,$kpc) of the spiral galaxy NGC 1365. We complemented these observations with archival Spitzer imaging and spectral mapping observations. The ILR region of NGC 1365 contains a Seyfert 1.5 nucleus and a ring of star formation with an approximate diameter of 2kpc. The strong star formation activity in the ring is resolved by the Herschel/PACS imaging data, as well as by the Spitzer $24\,\mu$m continuum emission, \[Ne[ ii]{}\]$12.81\,\mu$m line emission, and 6.2 and $11.3\,\mu$m PAH emission. The AGN is the brightest source in the central regions up to $\lambda \sim 24\,\mu$m, but it becomes increasingly fainter in the far-infrared when compared to the emission originating in the infrared clusters (or groups of them) located in the ring. We modelled the AGN unresolved infrared emission with the [clumpy]{} torus models and estimated that the AGN contributes only to a small fraction ($\sim 5\%$) of the infrared emission produced in the inner $\sim 5\,$kpc. We fitted the non-AGN $24-500\,\mu$m spectral energy distribution of the ILR region and found that the dust temperatures and mass are similar to those of other nuclear and circumnuclear starburst regions. Finally we showed that within the ILR region of NGC 1365 most of the on-going star formation activity is taking place in dusty regions as probed by the $24\,\mu$m emission.' author: - | \ $^{1}$Instituto de Física de Cantabria, CSIC-UC, Avenida de los Castros s/n, 39005 Santander, Spain\ $^{2}$Augusto González Linares Senior Research Fellow\ $^{3}$Herschel Science Centre, INSA/ESAC, E-28691 Villanueva de la Cañada, Madrid, Spain\ $^{4}$Instituto de Astrofísica de Canarias, C/Vía Láctea s/n, 38205 La Laguna, Tenerife, Spain\ $^{5}$Departamento de Astrofíısica, Universidad de La Laguna, 38205 La Laguna, Tenerife, Spain\ $^{6}$Centro de Astrobiología, INTA-CSIC, 28850 Madrid, Spain\ $^{7}$Istituto di Astrofisica e Planetologia Spaziali, INAF-IAPS, Via Fosso del Cavaliere 100, 00133 Rome, Italy\ $^{8}$Departamento de Física Moderna, Universidad de Cantabria, Avenida de Los Castros s/n, 39005 Santander, Spain\ $^{9}$Observatorio Astronómico Nacional (OAN)-Observatorio de Madrid, Alfonso XII 3, 28014 Madrid, Spain\ $^{10}$Gemini Observatory, Casilla 603, La Serena, Chile\ $^{11}$European Southern Observatory, Karl-Schwarzschild-Str. 2, 85748 Garching bei München, Germany\ $^{12}$Instituto de Astrofísica de Andalucía, IAA-CSIC, C/ Glorieta de la Astronomía s/n, 18008 Granada, Spain\ $^{13}$Astronomy Department, University of Florida, 211 Bryant Science Center, PO Box 112055, Gainesville, FL 32611-2055, USA\ date: 'Accepted — . Received — ; in original form — ' title: 'Probing the Nuclear and Circumnuclear Activity of NGC 1365 in the Infrared[^1]' --- \[firstpage\] Galaxies: evolution — Galaxies: nuclei — Galaxies: Seyfert — Galaxies: structure — Infrared: galaxies — Galaxies: individual: NGC 1365. Introduction ============ ------------ ------------ --------------- ----------------------- ---------- Obsid Instrument Bands Start time Duration () (UTC) (s) 1342201436 SPIRE-P 250, 350, 500 2010-07-14 20:20:45.0 999 1342222495 PACS-P 70, 160 2011-06-11 12:59:25.0 1217 1342222496 PACS-P 70, 160 2011-06-11 13:20:45.0 1217 1342222497 PACS-P 100, 160 2011-06-11 13:42:05.0 1217 1342222498 PACS-P 100, 160 2011-06-11 14:03:25.0 1217 ------------ ------------ --------------- ----------------------- ---------- The fueling of the nuclear and circumnuclear activity of galaxies has been a topic of extensive discussion. Such activity not only includes the accretion of matter onto a supermassive black hole with the accompanying active galactic nucleus (AGN), but also the presence of intense nuclear and circumnuclear starbursts. Both types of activity require gas to be transported from the host galaxy on physical scales of a few kiloparsecs down to less than one kiloparsec for the nuclear starburst activity and even further in for the nuclear activity. Interactions, mergers, and large-scale bars, among others, have been proposed as possible mechanisms to transport gas from kiloparsec scales to the nuclear and circumnuclear regions [see the review by @Jogee2006 and references therein]. In isolated galaxies with a large-scale bar the gas is believed to flow inwards between corotation and the inner Lindblad resonance (ILR). Indirect evidence of this is the presence of star formation rings and hot spots near the ILR of barred galaxies [see the review of @Buta1996]. The direction of the flows inside the ILR is generally outwards. This implies the existence of a gravity torque barrier at this resonance. However, if there is a “spatially-extended” ILR region, this translates into the existence of an inner ILR (IILR) and outer ILR (OILR). Numerical simulations predict that in this case a gas ring can be formed in between these two resonances. Furthermore, in the scenario of a double ILR the dynamical decoupling of an embedded nuclear bar can drive the gas further in [@Shlosman1989; @Hunt2008; @GarciaBurillo2009]. The combined action of gravity torques due to embedded structures (bars-within-bars) and viscous torques could be a viable mechanism to drive the gas to the inner few parsecs and feed the AGN [@GarciaBurillo2005; @Hopkins2011]. NGC 1365 is a giant isolated barred galaxy at a distance of 18.6Mpc [@Lindblad1999 therefore $1\arcsec=90$pc]. This galaxy hosts a Seyfert 1.5 nucleus [@Schulz1999]. @Jungwiert1997 showed that in NGC 1365 there is also a nuclear bar ($R<10\arcsec$) embedded in the the large-scale bar ($R\sim 100\arcsec$). @Lindblad1996 used hydrodynamical simulations to reproduce the kinematics and the offset dust lanes along the large-scale bar of this galaxy with an outer ILR at a radius of $R_{\rm OILR}= 27\arcsec = 2.4\,$kpc. There is also an inner ILR at a radius of $R_{\rm IILR} \sim 0.3\,$kpc. Henceforth we refer to the ILR region as the region interior to the OILR of NGC 1365. As predicted by simulations, there is a ring of star formation inside the ILR region of this galaxy. The star formation activity in the central regions of NGC 1365 has been revealed by the presence of hot spots [@Sersic1965], intense H$\alpha$ emission [@Alloin1981; @Kristen1997], non-thermal radio continuum sources associated with H[ii]{} regions and supernova remnants [@Sandqvist1995; @Forbes1998], large amounts of molecular gas [@Sakamoto2007], and point-like and diffuse extended X-ray emission not associated with the AGN [@Wang2009]. Moreover, there is evidence that a significant fraction of the circumnuclear star formation activity might be obscured based on the prominent dust lane crossing the nuclear region (see Fig. \[fig:PACS100\_large\], right panel) and the bright and extended mid-infrared (mid-IR) emission in this region [@Telesco1993; @Galliano2005]. In this paper we present new far-infrared (far-IR) imaging observations of NGC 1365 performed with the ESA Herschel Space Observatory [@Pilbratt2010] and new mid-IR imaging and spectroscopy obtained with the camera/spectrograph Thermal-Region Camera Spectrograph [T-ReCS; @Telesco1998] instrument on the Gemini-South telescope. The Herschel images were obtained using the Photodetector Array Camera and Spectrometer [PACS; @Poglitsch2010] and the Spectral and Photometric Imaging REceiver [SPIRE; @Griffin2010] instruments. We also use archival Spitzer data taken with the Infrared Array Camera [IRAC; @Fazio2004], the Multi-Band Imaging Photometer for Spitzer [MIPS; @Rieke2004], and the InfraRed Spectrograph [IRS; @Houck2004] instruments. Using IR observations to study the nuclear and circumnuclear activity in the ILR region of NGC 1365 is crucial because the central region is crossed by a prominent dust lane that obscures from our view in the optical a significant fraction of emission produced there. This paper is organized as follows. We describe the observations in Section 2. We analyze the AGN IR emission and the spatially resolved IR emission in Sections 3 and 4, respectively, and we summarize our conclusions in Section 5. Observations ============ Herschel/PACS and SPIRE imaging ------------------------------- We obtained Herschel far-IR imaging observations of NGC 1365 using PACS at 70, 100, and $160\,\mu$m and SPIRE at 250, 350, and $500\,\mu$m. The data are part of the guaranteed time program entitled “Herschel imaging photometry of nearby Seyfert galaxies: testing the coexistence of AGN and starburst activity and the nature of the dusty torus” (PI: M. Sánchez-Portal). The PACS observations were carried out using the standard scan map mode that takes two concatenated scan line maps at 45$^{\circ}$ and 135$^{\circ}$ (in array coordinates), at a speed of 20arcsec/sec, each one with 26 lines of $9\arcmin$ length and cross-scan step of $20\arcsec$. This mode produces a rather homogeneous exposure map within a square region of about $7\arcmin \times 7\arcmin$. The set of maps were duplicated to observe through both the 70  (“blue”) and 100  (“green”) filters. Therefore the galaxy was observed twice through the 160  (“red”) filter. With the SPIRE photometer we observed the three available bands simultaneously using the “large map” mode, with two nearly orthogonal scan maps (2 scan lines each), at a scan speed of 30arcsec/sec, and three repetitions. The homogeneous exposure area for scientific use is approximately $8\arcmin \times 8\arcmin$. Table \[tab:her\_obs\] gives the summary of the observations. ------------ ------------------- ------------ ---------- ----------------- ----------------- -- Instrument $\lambda_{\rm c}$ Pixel size FWHM $f_\nu$ (r=15”) $f_\nu$ (r=30”) $(\mu{\rm m})$ (arcsec) (arcsec) (Jy) (Jy) MIPS 24 2.45 5.9 7.2 8.7 PACS 70 1.4 5.6 85.5 102.7 PACS 100 1.7 6.8 110.1 141.7 PACS 160 2.8 11.3 87.4 123.7 SPIRE 250 6.0 18.1 33.4 53.3 SPIRE 350 10.0 25.2 10.4 21.0 SPIRE 500 14.0 36.9 – 5.6 ------------ ------------------- ------------ ---------- ----------------- ----------------- -- Note.— The reported values of the FWHM are the nominal values. The errors in the aperture photometry are dominated by the photometric calibration uncertainty of the instruments that is typically 10%. We reduced the data with the Herschel Interactive Processing Environment (HIPE) v8.0.1 and Scanamorphos [@Roussel2012] v15. For the PACS instrument, we used HIPE and the Calibration Database V32 to build Level 1 products. These included detecting and flagging bad pixels, converting the ADU readings to flux units (Jy/pixel), and adding the pointing information. We did not attempt to perform deglitching at this stage to prevent the bright AGN nucleus to be affected by the MMT deglitching process. The final maps were built from the Level 1 products using Scanamorphos, which performs a baseline subtraction, correction of the striping effect due to the scan process, removal of global and individual pixel drifts, and finally the map assembly using all the nominal and cross-direction scans. For SPIRE we used the standard (small) map script and Calibration Database v8.1. The processing included detection of thermistor jumps in the time line, frame deglitching, low pass filter correction, conversion of readings to flux units (Jy/beam), temperature drift and bolometer time response corrections, and addition of pointing information. We built the final maps using the “naïve” scan mapper task. Colour corrections (for PACS, see @Poglitsch2010; please refer to the *Observer’s Manual* for the SPIRE ones) are small for blackbodies at the expected temperatures (e.g., @PerezGarcia2001) and have been neglected. More details on the processing of Herschel data are given in Sánchez-Portal et al. (in preparation). Figure \[fig:PACS100\_large\] shows the fully reduced PACS $100\,\mu$m image of NGC 1365 together with the optical BVR image from [@Elmegreen2009]. Fig. \[fig:IRcentralimages\] shows the PACS images together with the SPIRE $250\,\mu$m with a field of view (FoV) covering the approximate extent of the ILR region of NGC 1365 (see also Section \[sec:alignment\]). We performed aperture photometry on all the Herschel images using two different radii, $r=15\arcsec$ and $r=30\arcsec$. The latter encompasses the ILR region, whereas the former includes mostly the ring of star formation. Table \[tab:photometry\] lists the measured flux densities. Gemini/T-ReCS imaging and spectroscopy {#sec:trecs} -------------------------------------- We obtained mid-IR imaging of NGC 1365 using T-ReCS on the Gemini-South telescope on September 8 and 9, 2011 as part of proposal GS-2011B-Q-20 (PI: N. Levenson). We used the Si-2 ($\lambda_{\rm c}=8.74\,\mu$m) and the Qa ($\lambda_{\rm c}=18.3\,\mu$m) filters and mid-IR standard observation techniques. The plate scale of the T-ReCS imaging observations is 0.089/pixel with a FoV of $28.5\arcsec \times 21.4\arcsec$. The total integration times (on-source) were 145s and 521s in the Si-2 and Qa filters, respectively. The Qa filter observations were split between the two nights, whereas those in the Si-2 filter were done on the second night. We observed standard stars immediately before or after the science observations in the same filters, to both flux-calibrate the galaxy observations and to estimate the unresolved nuclear emission (see below). The observations were diffraction limited, with a full width half maximum (FWHM) of $0.34\arcsec$ in the Si-2 filter and $0.55-0.58\arcsec$ in the Qa filter, as measured from the standard star observations. We reduced the imaging data using the CanariCam data reduction package RedCan (González-Martín et al., in preparation). We refer the interested reader to this work and [@RamosAlmeida2011AGN] for details. Fig. \[fig:trecs\] shows the fully reduced T-ReCS Qa image resulting from the combination of the data taken during the two observing nights. The T-ReCS $8.7\,\mu$m image (not shown here) shows a similar morphology. We also retrieved archival T-ReCS spectroscopic observations in the $N$-band ($\sim 8-13\,\mu$m) using a $0.35\arcsec$ slit width as part of proposal GS-2009B-Q-19 (PI: M. Pastoriza). The total on-source integration time was 600s. We reduced the galaxy and corresponding standard star observations using the RedCan package following the steps described in [@AAH11AGN]. Finally we extracted the nuclear spectrum as a point source. To estimate the nuclear unresolved emission from the mid-IR imaging data we followed the point spread function (PSF) scaling technique implemented by [@RamosAlmeida2009; @RamosAlmeida2011AGN]. This unresolved emission which is assumed to represent the torus emission (see Section \[sec:torusfit\]). To do so, we scaled the observation of the corresponding standard star to the peak of the nuclear emission of the galaxy in each of the two filters. This represents the maximum contribution of the unresolved source (100%), whereas the residual of the total emission minus the scaled PSF accounts for the extended emission. In both filters we found that a 90% PSF scaling provided a realistic estimate of the extended emission. The estimated errors in the T-ReCS unresolved flux densities reported in Table \[tab:AGNfluxes\] are 15% and 25% in the Si-2 and Qa filters, respectively, and account for both the flux calibration and PSF subtraction uncertainties [see @RamosAlmeida2009 for more details]. The unresolved $8.7\,\mu$m emission computed this way is in good agreement with the flux density at the same wavelength obtained from the T-ReCS nuclear spectrum. Archival Spitzer/IRAC and MIPS imaging -------------------------------------- We retrieved from the Spitzer archive imaging data of NGC 1365 observed with all four IRAC bands (3.6, 4.5, 5.8, and $8\,\mu$m) and with MIPS at $24\,\mu$m (Program ID: 3672, PI: J. Mazzarella). These observations were part of The Great Observatories All-Sky LIRG Survey [GOALS, see @Armus2009]. We retrieved the basic calibrated data (BCD) from the Spitzer archive. The BCD processing includes corrections for the instrumental response (pixel response linearization, etc.), flagging of cosmic rays and saturated pixels, dark and flat fielding corrections, and flux calibration based on standard stars. We combined the BCD images into mosaics using the MOsaicker and Point source EXtractor (MOPEX) software provided by the Spitzer Science Center using the standard parameters. The final mosaics were repixeled to half of the original pixel size of the images, that is, the IRAC mosaics have $0.6\arcsec$/pixel, whereas the MIPS $24\,\mu$m mosaic has $1.225\arcsec$/pixel. In Fig. \[fig:IRcentralimages\] we show the Spitzer images with a FoV covering the approximate extent of the ILR region, as done with the PACS images and the SPIRE $250\,\mu$m image. The angular resolutions of the IRAC images are between 1.7 and $2\arcsec$ (FWHM) and that of the MIPS $24\,\mu$m is $5.9\arcsec$. Archival optical ground-based imaging ------------------------------------- We retrieved from the ESO archive optical images obtained with the WFI instrument on the MPG/ESO 2.2m telescope using the narrow-band Halpha/7 filter ($\lambda_{\rm c}=6588.3$Å, FWHM=70Å) obtained as part of proposal 065.N-0076 (PI: F. Bresolin). We combined a total of 6 images, each of them with a 350s exposure. The plate scale of the images is $0.238\arcsec$/pixel. The filter contains the H$\alpha$ and \[N[ ii]{}\] emission lines plus adjacent continuum. Since we use this image for morphological purposes only, we did not attempt to either subtract the continuum or calibrate it photometrically. The positions of the H$\alpha$ hot spots identified by [@Alloin1981] in the central region of NGC 1365 are displayed in Fig. \[fig:closeups\] (upper panel). Alignment of the images {#sec:alignment} ----------------------- The alignment of the Spitzer/IRAC and the Gemini/T-ReCS images is straightforward because the AGN and the mid-IR clusters detected by [@Galliano2005] are clearly identified in all these images. The AGN is also bright in the optical image and therefore we used it as our reference. In Fig. \[fig:closeups\] we present a close-up of the ILR region in the IRAC $8\,\mu$m band and the optical narrow-band H$\alpha$ image. We marked the positions of the bright (designated as M4, M5, and M6) and faint mid-IR (designated as M2, M3, M7, and M8) clusters and the nucleus using the relative positions given by [@Galliano2005]. We also indicated the positions of radio sources and H$\alpha$ sources (see Section \[sec:morphology\_detailed\]). Although the Spitzer/MIPS $24\,\mu$m image has a poorer angular resolution when compared to that of IRAC, the AGN is still sufficiently bright at $24\,\mu$m that can be distinguished from the mid-IR clusters. The alignment of the Herschel/PACS images is not as simple because the AGN does not appear to be a bright source in the far-IR. We used the astrometry information in the Herschel image headers and the optical position of the nucleus of NGC 1365 given by [@Sandqvist1995]: RA(J2000)=$03^{\rm h}33^{\rm m}36.37^{\rm s}$ and Dec(J2000)=$-36^{\circ}08\arcmin25.5\arcsec$, for the initial alignment. These coordinates placed the AGN to the southwest of the bright source detected in the PACS images and in the MIPS $24\,\mu$m image that appears to coincide with the region containing clusters M4, M5, and M6. Additionally we compared the morphologies of the PACS $70\,\mu$m and the MIPS $24\,\mu$m images as they have similar angular resolutions (see Table \[tab:photometry\]). We used the mid-IR source located $\sim 17\arcsec$ N, $16\arcsec$ E of the AGN identified in the IRAC and MIPS images and also seen in the PACS $70\,\mu$m and $100\,\mu$m images for a finer alignment. This source appears to be coincident with the L4 H[ii]{} region or H$\alpha$ hot spot (see Fig. \[fig:closeups\] and Table \[tab:sources\]) identified by [@Alloin1981] and also seen in our archival H$\alpha$ image. The PACS $160\,\mu$m image was aligned relative to the other PACS images with the astrometry information in the headers. Finally since the bright mid-IR clusters cannot be resolved in the SPIRE $250\,\mu$m image, we placed the center of image at the position of the AGN. We note that the alignment of the PACS images in Figs. \[fig:IRcentralimages\] and \[fig:closeups\] is only good to within 1 pixel in each band. Since the main goal of this work is to study the processes giving rise to the IR emission within the ILR region of NGC 1365, in Fig. \[fig:IRcentralimages\] we show the aligned IR images with a FoV covering the approximate extent of this region. We do not show the SPIRE 350 and $500\,\mu$m images because of the small number of pixels covering the ILR region of NGC 1365. ------------ ---------------------- ---------------------- Wavelength $f_\nu$ Method ($\mu$m) (mJy) 8.7 $203\pm30$ Imaging (unresolved) 13.0 $400\pm60$ Spectroscopy 18.3 $818\pm205$ Imaging (unresolved) 24 $1255^{+783}_{-500}$ BC fit 70 $734^{+1482}_{-422}$ BC fit 100 $271^{+632}_{-163}$ BC fit 160 $<78$ BC fit ------------ ---------------------- ---------------------- : AGN emission[]{data-label="tab:AGNfluxes"} References.— The quoted uncertainties for the BC fit fluxes are the $\pm 1\sigma$ uncertainties, as discussed in Sections \[sec:torusfit\] and \[sec:farIRAGNemission\]. Spitzer/IRS spectral mapping ---------------------------- We retrieved from the Spitzer archive low spectral resolution ($R \sim 60-126$) observations (Program ID: 3269, PI: J. Gallimore) of NGC 1365 obtained with the spectral mapping capability of IRS. These observations were part of the Spitzer study of the spectral energy distributions (SED) of the $12\,\mu$m sample of active galaxies [@Gallimore2010]. The observations were obtained with the Short-Low (SL1; $7.5-14.3\,\mu$m and SL2; $5.1-7.6\,\mu$m) and Long-Low (LL1; $19.9-39.9\,\mu$m and LL2; $13.9-21.3\,\mu$m) modules. The plate scales of the SL and LL modules are $1.8\arcsec$/pixel and $5.1\arcsec$/pixel, respectively. ------------ -------------------------- ---------------- --------- -- -- -- -- Name Spectral Rel. Position Ref. Range arcsec, arcsec M2, L2 mid-IR, H$\alpha$ -4.7, -5.1 1, 2 M3, L3 mid-IR, H$\alpha$ -5.4, -2.6 1, 2, 3 M4, D mid-IR, radio 0.4, 7.1 1, 4 M5, E, L12 mid-IR, radio, H$\alpha$ 2.8, 10.0 1, 4, 3 M6, G mid-IR, radio 4.8, 7.0 1, 4 M7, H mid-IR, radio $^*$ 1, 4 M8, L1 mid-IR, H$\alpha$ $^*$ 1, 2 L11 H$\alpha$ 10, 15 3 L4 H$\alpha$ 17, 16 3 A radio -4.1, -2.4 4 ------------ -------------------------- ---------------- --------- -- -- -- -- : Summary of sources in the circumnuclear region[]{data-label="tab:sources"} The positions are relative to that of the AGN and correspond to those of the first listed source. $^*$The positions shown in Figs. 3 and 4 are estimated from the mid-IR images of Galliano et al. (2005). References.— 1. Galliano et al. (2005). 2. Kristen et al. (1997). 3. Alloin et al. (1981). 4. Sandqvist et al. (1995). The data were processed using the Spitzer IRS pipeline. The IRS data cubes were assembled using [cubism]{} [the CUbe Builder for IRS Spectra Maps, @Smith2007CUBISM] v1.7 from the individual BCD spectral images obtained from the Spitzer archive. [cubism]{} also provides error data cubes built by standard error propagation, using, for the input uncertainty, the BCD-level uncertainty estimates produced by the IRS pipeline from deviations of the fitted ramp slope fits for each pixel. We used these uncertainties to provide error estimates for the extracted spectra, and the line and continuum maps [see @Smith2007CUBISM for full details] discussed in the next two sections. ### Extraction of the 1D spectra {#sec:1Dspectra} We used [cubism]{} to extract spectra of regions of interest using small apertures taking advantage of the angular resolution of the spectral mapping observations obtained with the SL1+SL2 modules [$\sim 4\arcsec$ FWHM, see @Pereira2010IRSmapping]. We used square or rectangular apertures in the original orientation of the SL data cubes with sizes of two or three pixels (see Table \[tab:spectroscopy\] for the extraction apertures used for each region). The selected regions include the AGN, and the regions containing the M2+M3, M4, and M5+M6 mid-IR clusters (see Fig. \[fig:closeups\] for the positions) identified by [@Galliano2005]. We note that we did not attempt to apply a point source correction to the SL $3.7\arcsec \times 3.7\arcsec$ spectrum of the AGN because we are mostly interested in the extended features, that is, the polycyclic aromatic hydrocarbon (PAH) features and the \[Ne[ii]{}\]$12.81\,\mu$m fine structure line. Fig. \[fig:SLspectra\] shows the SL1+SL2 spectra of the selected regions normalized at $13\,\mu$m. Finally we extracted the integrated spectrum of the region covered by the observations, i.e. the central $27.8\arcsec \times 24.0\arcsec$. We measured the fluxes and the equivalent widths (EW) of the \[Ne[ii]{}\]$12.81\,\mu$m emission line and the 6.2 and $11.3\,\mu$m PAH features fitting Gaussian profiles to the lines and lines to the local continuum. Our measurements of the \[Ne[ii]{}\] flux for clusters M4, M5, M6 are in good agreement with those reported by [@Galliano2008] from ground-based high angular resolution observations. Since the PAH features are broad, it has been noted in the literature that Gaussian profiles might not be appropriate to measure their flux because a large fraction of the energy in these bands in radiated in the wings. A Lorentzian profile might be a better approximation [see @Galliano2008PAH] to measure their flux, and therefore we repeated the line fits with this method. Table \[tab:spectroscopy\] lists the measurements for the extracted spectra for the Gaussian profiles. To illustrate the effect of using different profiles, in this table we give the measured 6.2 to $11.3\,\mu$m PAH ratio for the two profiles and the selected regions. ------------ ---------------------------------- ------- ------ --------------------------- --------- ------ --------- ----- ----- Region Extraction \[Ne[ii]{}\]$12.81\,\mu$m $S_{\rm Si}$ Aperture flux EW flux EW flux G L AGN $3.7\arcsec \times 3.7\arcsec$ 8.6 0.09 5.3 0.13 1.7 $-0.10$ 1.6 1.2 M2+M3 $5.6\arcsec \times 3.7\arcsec$ 21.0 0.50 12.7 0.61 4.6 $-0.24$ 1.7 1.1 M4 $3.7\arcsec \times 3.7\arcsec$ 10.7 0.45 5.8 0.63 2.3 $-0.85$ 1.8 1.7 M5+M6 $5.6\arcsec \times 3.7\arcsec$ 20.0 0.45 10.2 0.40 6.5 $-0.81$ 2.0 1.5 Integrated $27.8\arcsec \times 24.0\arcsec$ 265.0 0.42 161.0 0.49 60.5 $-0.45$ 1.6 1.3 ------------ ---------------------------------- ------- ------ --------------------------- --------- ------ --------- ----- ----- Notes.— Fluxes (observed, not corrected for extinction) are in units of $\times 10^{-13}\,{\rm erg \, cm}^{-2}\,{\rm s}^{-1}$ and EW in units of $\mu$m for measurements done with Gaussian profiles. The typical errors are 10% for the fluxes and $0.05\,\mu$m for the EW. $S_{\rm Si}$ is the strength of the silicate feature (see Section \[sec:spectralmaps\] for the definition). The ratio of the $6.2\,\mu$m PAH flux to the $11.3\,\mu$m PAH feature flux is given for fits to the features done with Gaussians (G) and Lorentzian (L) profiles (see Section \[sec:1Dspectra\] for more details). ### Spectral Maps {#sec:spectralmaps} We used [cubism]{} to construct spectral maps of the most prominent features in the SL data cubes, namely, the $6.2\,\mu$m and $11.3\,\mu$m PAH features, and the \[Ne[ii]{}\]$12.81\,\mu$m fine structure line. The technique used here was very similar to that of [@AAH09] and involves integrating the line flux over the user-defined emission line regions. Note that, unlike the line measurements in the previous section, the features are not actually fitted with Gaussian or Lorentzian profiles, and therefore these maps are only used for morphological purposes. Since [cubism]{} does not fit or deblend emission lines, the SL \[Ne[ii]{}\]$12.81\,\mu$m map includes some contribution from the $12.7\,\mu$m PAH feature. We also built a continuum map at $5.5\,\mu$m. The Spitzer/IRS SL spectral maps shown in Fig. \[fig:spectralmaps\] were trimmed and rotated to the usual orientation of north up, east to the left. We used the associated uncertainty maps produced by [cubism]{} to compute the relative errors of the spectral maps. Finally we constructed the map of the silicate feature, which is shown in Fig. \[fig:silicatesPAHratio\], following the technique of [@Pereira2010IRSmapping]. Briefly, it involves fitting the silicate feature from 1D spectra extracted in $2{\rm pixel} \times 2{\rm pixel}$ boxes from the IRS SL data cubes. The map is then constructed by moving by 1 pixel in the x and y directions until the FoV of the IRS SL data cubes is completely covered. We measured the apparent strength of the silicate feature in the IRS spectra following [@Spoon2007]: $S_{\rm Si} = \ln f_\nu({\rm obs})/f_\nu({\rm cont})$, where $f_\nu({\rm obs})$ is the flux density at the feature and $f_\nu({\rm cont})$ is the flux density of the underlying continuum. The latter was fitted as a power law between 5.5 and $14\,\mu$m. We evaluated the strength of the silicate feature at $10\,\mu$m. When the silicate strength is negative, the silicate feature is in absorption, whereas a positive silicate strength indicates that the feature is seen in emission. The uncertainties of the measured strengths in the spectral map and the extracted 1D spectra of previous section are $\pm 0.05$. AGN Infrared Emission {#sec:spectraldecomposition} ===================== A number of works have studied in detail the mid-IR emission of the AGN hosted by NGC 1365 using high angular resolution observations, including imaging, spectroscopy, and interferometry. Our T-ReCS nuclear ($\sim 0.35\arcsec$) mid-IR spectrum of NGC 1365 (see Fig. \[fig:torusmodel\]) is an almost featureless continuum with no clear evidence of the presence of the $9.7\,\mu$m silicate feature or PAH features [see also @Siebenmorgen2004; @Tristram2009]. By contrast, the spectrum integrated over several arcseconds is rich in spectral features related to star formation activity, such as, PAH features and the \[Ne[ii]{}\] line (see Fig. \[fig:SLspectra\]). [@Tristram2009] modelled mid-IR interferometric observations of the nuclear region and derived a size (diameter) of $1.1-2.7\,$pc for the $12\,\mu$m emitter. Up until now, there has been no estimates of the far-IR emission arising from the AGN of NGC 1365, and thus the main goal of this section is to provide such an estimate. On scales of a few arcseconds the AGN of NGC 1365 is surrounded by a number of mid-IR bright clusters and extended and diffuse emission (see Figs. \[fig:IRcentralimages\] and \[fig:closeups\], and Section \[sec:spatiallyresolved\]). Moreover, in the far-IR AGN emission becomes less luminous for increasing wavelengths when compared to that of the clusters (see Fig. \[fig:IRcentralimages\]). This together with the limited angular resolution of the PACS images (i.e., the best angular resolution is FWHM$\sim 5.6\arcsec$ at $70\,\mu$m) prevents us from doing aperture photometry on the PACS images to derive the far-IR flux densities of the AGN. Instead, in this section we estimate the AGN emission by both fitting the nuclear near- and mid-IR SED with torus models and doing a spectral decomposition of the Spitzer/IRS nuclear spectrum. Parameter Interval ------------------------------------------------ ------------ Torus radial thickness $Y$ \[5, 100\] Torus angular width $\sigma_{\rm torus}$ (deg) \[15, 70\] Number clouds along an equatorial ray $N_0$ \[1, 15\] Index of the radial density profile $q$ \[0, 3\] Viewing angle $i$ (deg) \[0, 90\] Optical depth per single cloud $\tau_V$ \[5, 150\] : Parameters of the *CLUMPY* Torus Models[]{data-label="tab:torusmodels"} Fit to the nuclear SED and mid-IR spectrum using clumpy torus models {#sec:torusfit} -------------------------------------------------------------------- Clumpy torus models [e.g., @Honig2006; @Schartmann2008] and in particular the [@Nenkova2008a; @Nenkova2008b] models (also known as [clumpy]{} models) have been shown to fit well the near and mid-IR SEDs and spectroscopy of Seyfert galaxies [@RamosAlmeida2009; @RamosAlmeida2011AGN; @Mason2009; @Nikutta2009; @AAH11AGN; @Sales2011; @Lira2012]. The [clumpy]{} torus models are described by six parameters to account for the torus geometry and the properties of the dusty clouds (see Table \[tab:torusmodels\]). The geometry of the torus is defined with the torus radial thickness $Y$, which is the ratio between the outer and inner radii of the torus $Y = R_{\rm o}/R_{\rm d}$[^2], and the width of the angular distribution of the clouds $\sigma_{\rm torus}$, that is, the angular size of the torus as measured from its equator. The clouds have an optical depth $\tau_V$ and are arranged with a radial distribution expressed as a declining power law with index $q$ ($\propto r^{−q}$), with $N_0$ giving the mean number of clouds along a radial equatorial ray. The last parameter is the viewing angle to the torus $i$. We refer the reader to the sketch of the [clumpy]{} model geometry in Figure 1 of [@Nenkova2008a]. We updated the fit of [@RamosAlmeida2011AGN] with two main differences. First, we used the higher angular resolution mid-IR unresolved flux densities inferred in Section \[sec:trecs\] and also included the T-ReCS mid-IR spectrum for the fit. Second, this galaxy is classified as a Seyfert 1.5 [it shows a broad component in H$\beta$, @Schulz1999] and therefore there is an unobscured view to the AGN. To account for it, we added an AGN component to the resulting torus model (see below) and allowed for a small amount of foreground extinction [$A_V <5\,$mag, see @Alloin1981], to redden the fitted models. We also included in the nuclear SED the HST/NICMOS $1.6\,\mu$m unresolved flux [@Carollo2002]. Similarly to our previous work [@RamosAlmeida2011AGN; @AAH11AGN], we took a Bayesian approach to both fit the torus models to the data and derive meaningful confidence levels for the fitted parameters. To this end, we used the BayesClumpy (BC) fitting routine, an interpolated version of the [clumpy]{} models [@AsensioRamos2009]. The new version of BC allows to fit both photometric points and spectra. For the modelling we used the whole range of parameters probed by the [clumpy]{} models. We list them in Table \[tab:torusmodels\]. Unlike in some of our previous work, here we allowed for the torus extent to vary in the range $Y=5-100$, instead of restricting the models to small tori. This is because larger tori may be required to reproduce the far-IR emission of AGN [@RamosAlmeida2011Herschel]. In addition to the six torus model parameters to be fitted plus the foreground extinction, there is an extra parameter to account for the vertical displacement needed to match the fluxes of a given model to the observations. This vertical shift, which we allow to vary freely, scales with the AGN bolometric luminosity [see @Nenkova2008b]. We assumed uniform distributions in the ranges given in Table \[tab:torusmodels\] for the prior distributions of the torus model parameters. The BC fitting routine translates the probability distributions of the fitted torus model parameters into two model spectra. The first corresponds to the maximum-a-posteriori (MAP) values that represent the best fit to the data, while the second is the model produced with the median value of the probability distribution of each parameter. Fig. \[fig:torusmodel\] shows the best fit to the nuclear near and mid-IR SED and the mid-IR spectrum of the NGC 1365 AGN. We also plot in this figure the range of acceptable torus models at the $68\%$ confidence level. As can be seen from this figure, the fit to the photometric and spectroscopic data is very good. Specifically, the [clumpy]{} models reproduce well the flat silicate feature at around $9.7\,\mu$m [see @Nenkova2008b their figures 16 and 17]. We can now compare the inferred torus model parameters with observations. We start with the AGN bolometric luminosity. As explained above, we chose to keep the scaling of the torus models to the observations as a free parameter. Because the scaling is proportional to the AGN bolometric luminosity, when compared with the observations it provides information about the goodness of the fit [see @AAH11AGN for more details]. The derived AGN bolometric luminosity from our fit is $L_{\rm bol}({\rm AGN}) = 2.6 \pm 0.5 \times 10^{43}\,{\rm erg \, s}^{-1}$. @Risaliti2005 infered a $2-10\,$keV luminosity in the range $0.8-1.4\times 10^{42}\,{\rm erg \, s}^{-1}$ (corrected for the assumed distance in this work). Our derived AGN bolometric luminosity is consistent with the values from X-rays after applying a bolometric correction. We can also compare the torus radius $R_{\rm o}=5^{+0.5}_{-1}\,$pc and torus angular width $\sigma_{\rm torus}=36^{+14}_{-6}\deg$ from the modelling with observations. The width of the angular distribution of the clouds in the torus is related to the opening angle of the ionization cones, with some dependence with the geometry of the gas [see discussion in @AAH11AGN]. The half opening angle of the NGC 1365 ionization cone modeled by [@Lindblad1999] is $\Theta_{\rm cone}=50$deg (measured from the pole), which is compatible with the derived angular size of the torus. The size (diameter) of the $12\,\mu$m emitting source inferred from the modelling of the mid-IR interferometric observations of this galaxy is $s_{12\mu{\rm m}}=1.1-2.7\,$pc [@Tristram2009]. We note that the size of the $12\,\mu$m emitting source, which traces the warm dust within the torus, is expected to be smaller than the [*true*]{} size of the torus. This is because the $12\,\mu$m emission is rather insensitive to cooler material further from the nucleus [see discussion in @AAH11AGN]. Given the good agreement with the observations, we conclude that the fitted torus model provides a good representation of the AGN torus emission. AGN torus emission in the far-IR {#sec:farIRAGNemission} -------------------------------- To estimate the AGN far-IR flux density in the PACS bands, we extrapolated the fitted torus model beyond $\sim 20\,\mu$m. As can be seen from Fig. \[fig:torusmodel\], the torus model flux density peaks at $40-50\,\mu$m and falls off rapidly at longer wavelengths. This is similar to results for other AGN [see e.g., @Mullaney2011]. In Table \[tab:AGNfluxes\] we list the extrapolated AGN far-IR flux densities together with those in the mid-IR from the T-ReCS imaging and spectroscopy. The ranges given in this table for the far-IR fluxes from the BC fit of the NGC 1365 take into account the $\pm 1\sigma$ confidence region of the fitted [clumpy]{} torus models. A quick comparison between the extrapolated AGN far-IR fluxes and the aperture photometry of the central regions (Table \[tab:photometry\]) clearly shows that the AGN contribution at these wavelengths is small and decreases for increasing wavelengths. Within the central 5.4kpc ($r=30\arcsec$) the AGN is predicted to contribute $15\%$ ($+8\%$, $-6\%$) of the total $24\,\mu$m emission. At $70\,\mu$m the AGN accounts for at most 2% of the emission in the same region and less than 1% at 100 and $160\,\mu$m. Using the flux densities measured for $r=30\arcsec$ and the following equation from [@Dale2002], $$\begin{aligned} L_{3-1000\mu{\rm m}} = 1.559\,\nu L_{\nu}(24\mu{\rm m}) + 0.7686\, \nu L_{\nu}(70\mu{\rm m}) + \nonumber \\ 1.347 \,\nu L_{\nu}(160\mu{\rm m}) \end{aligned}$$ we estimate a $3-1000\,\mu$m luminosity within the ILR region of $\sim 9 \times 10^{10}\,{\rm L}_\odot$. The ratio between the AGN bolometric luminosity and the IR luminosity of this region is then $0.05\pm0.01$. This is in good agreement with findings for the majority of local LIRGs hosting an AGN [@AAH11LIRGs]. Spatially-resolved IR emission in the ILR region {#sec:spatiallyresolved} ================================================ Morphology {#sec:morphology_detailed} ---------- The central region of NGC 1365 is a bright source of mid- and far-IR emission [see e.g., @Telesco1993; @Galliano2005; @Wiebe2009], although its contribution to the integrated emission of the galaxy decreases significantly at longer wavelengths. The mid-IR morphology, especially in the IRAC 5.8 and $8\,\mu$m and in the T-ReCS images, resembles that of a circumnuclear ring of star formation with a diameter of approximately 1.8kpc (see Figs. \[fig:IRcentralimages\] and \[fig:trecs\]). The AGN and the bright mid-IR clusters M4, M5, and M6 identified (see Table \[tab:sources\]) by [@Galliano2005] out to $\lambda \sim 12\,\mu$m are also clearly identified in the Gemini/T-ReCS image at $18.3\,\mu$m (see Fig. \[fig:trecs\]). The region to the southwest of the AGN is more diffuse and similar to the H$\alpha$ morphology (see Fig. \[fig:closeups\], upper panel). We do not, however, identify any bright compact sources there. This diffuse emission is reminiscent of the optical morphology of the region containing superstar clusters SSC3 and SSC6 along with a number of less luminous optical clusters [@Kristen1997; @Lindblad1999]. In the Spitzer/MIPS $24\,\mu$m image the emission from the AGN is still differentiated from that of the mid-IR clusters, although the latter are not resolved individually because of the decreased angular resolution. The Herschel observations of NGC 1365 show that the far-IR AGN emission becomes faint when compared to that of the clusters. This is not only because the clusters in the central region of NGC 1365 are still intrinsically bright in the far-IR, but also because in general the AGN emission is observed to fall steeply in the far-IR [see e.g., @Netzer2007; @Mullaney2011 and references therein]. Moreover, in general the far-IR emission of AGN appears to be dominated by a starburst component [@Hatziminaoglou2010]. It is also apparent from Fig. \[fig:IRcentralimages\] that the group of clusters M5...M8 are brighter than the group of M2 and M3 in the far-IR, as is also the case in the mid-IR. This implies, as will be discussed further in Section \[sec:sfr\], that the obscured star formation rate (SFR) is higher in the former group than in the latter. Moreover, the morphology of the SFR tracers (e.g., the $8$, $24$, and $70\,\mu$m emissions) is similar to that of the molecular gas, especially that of the rare C$^{18}$O(J=2-1) isotope [@Sakamoto2007], where a “twin-peak” morphology is observed. Such molecular gas twin peaks are observed in barred galaxies [@Kenney1992] and are formed where dust lanes intersect nuclear rings of star formation, as is the case of NGC 1365. We can also compare the IR morphology, in particular at 8, 24, and $70\,\mu$m, of the ILR region with that of H$\alpha$[^3]. Since all these emissions probe the SFR in galaxies [@Kennicutt1998; @Kennicutt2009; @AAH06; @Calzetti2007; @Rieke2009; @Li2010], a good morphological correspondence on scales of hundreds of parsecs is expected. Fig. \[fig:closeups\] indeed shows that the H$\alpha$ hot spots in the ILR region [Table 4, and also @Alloin1981; @Kristen1997] are clearly associated with IR emitting regions (compare with the IRAC $8\,\mu$m image as it has the best angular resolution). This seems to be the case out to $\lambda=100\,\mu$m. At longer wavelengths the decreased angular resolution of the Herschel images does not allow us to compare the morphologies. The opposite is not necessary true because the H$\alpha$ emission is strongly affected by extinction caused by the prominent dust lane crossing the ILR region. Mid-IR clusters M4 and M5 appear to be detected in the H$\alpha$ image, but they are very faint. This is because they are in one of the regions with the highest extinction, as we shall see in more detail in Section \[sec:silicatefeature\]. The Spitzer/IRS map of the \[Ne[ii]{}\]$12.81\,\mu$m emission covering the central 2.2kpc of NGC 1365 is shown in Fig.\[fig:spectralmaps\]. This emission line is a good tracer of the current SFR because its luminosity is proportional to the number of ionizing photons [@Roche1991; @HoKeto2007]. The \[Ne[ii]{}\] emission appears as a partial ring and it is enhanced in the mid-IR clusters. Its morphology is similar to the PACS $70\,\mu$m morphology and the CO “twin peaks” discussed by @Sakamoto2007. The AGN of NGC 1365, on the other hand is not a bright \[Ne[ii]{}\]$12.81\,\mu$m source, as found for other active galaxies [see e.g., @PereiraSantaella2010]. Analogously to other SFR indicators, namely the 24 and $70\,\mu$m emissions, the region to the northeast of the AGN (clusters M4, M5, and M6) is brighter in \[Ne[ii]{}\] than the region of clusters M2 and M3. The regions containing these bright mid-IR clusters account for approximately $20-25\%$ of the total \[Ne[ ii]{}\]$12.81\,\mu$m emission in the central 2.2kpc (see Table \[tab:spectroscopy\]). Note, however that the emission from the clusters is unresolved at the Spitzer/IRS SL angular resolution and we did not apply a correction for unresolved emission. Therefore this contribution from the clusters should be taken as a lower limit. Figure \[fig:spectralmaps\] also shows the maps of the 6.2 and the $11.3\,\mu$m PAH features. Both features probe mostly the emission from B stars rather than that from massive on-going star formation [O stars, see @Peeters2004]. The $6.2\,\mu$m PAH morphology is very similar to \[Ne[ii]{}\], whereas the $11.3\,\mu$m PAH map appears to have a deficit of emission in the region to the northeast of the AGN. The map of the ratio between these two features is in Fig. \[fig:silicatesPAHratio\]. Although several works found variations in the PAH ratios of galaxies not explained by changes in the extinction [see e.g., @Galliano2008; @Pereira2010IRSmapping], in the case of the central region of NGC 1365 it might be entirely due to extinction. The 6.2 to $11.3\,\mu$m PAH ratio is strongly increased in the region enclosing clusters M4, M5, M6, which is also the region suffering the highest extinction (see Section \[sec:silicatefeature\] and Fig. \[fig:silicatesPAHratio\]). Since the relative absorption at $11.3\,\mu$m is higher than at $6.2\,\mu$m [see e.g., table 5 in @Brandl2006], correcting the $11.3\,\mu$m PAH map for extinction would produce a morphology similar to that of the $6.2\,\mu$m PAH map. The measured EW and PAH ratios of the regions not affected by strong extinction are similar to those measured in other starburst galaxies [see e.g., @Brandl2006; @Spoon2007; @Sales2010]. We conclude that the star formation activity inside the ILR region of NGC 1365 has been taking place for at least a few tens of million years, as the PAH features trace the emission from B stars [@Peeters2004], and is taken place currently because of the bright \[Ne[ii]{}\]$12.81\,\mu$m emission requires the presence of young ionizing stars. Extinction: Silicate Feature {#sec:silicatefeature} ---------------------------- The circumnuclear region of NGC 1365 is crossed by a prominent dust lane entering the ring [see @Lindblad1999 and references therein and also Figs. \[fig:PACS100\_large\] and \[fig:closeups\]]. This dust lane is still apparent in the shortest wavelength IRAC bands and can be seen passing between clusters M5 and M6 at $3.6\,\mu$m (see Fig. \[fig:IRcentralimages\]). The brightest mid-IR clusters M4, M5, and M6 are located at the edge of the dust features where the bar dust lane enters the ILR region to the northeast of the AGN [@Elmegreen2009 and also Figs. \[fig:IRcentralimages\] and \[fig:closeups\]]. Mid-IR clusters M2 and M3 to the southwest of the nucleus are in a region less affected by extinction and are likely to be associated with optical super star clusters and optically-detected H[ii]{} regions [@Kristen1997; @Sakamoto2007]. The broad feature at $\sim 10\,\mu$m is believed to be produced by amorphous silicate grains and in [*normal*]{} star-forming galaxies is observed in mild absorption [@Roche1991; @Smith2007SINGS]. In the simplest dust geometry of a purely absorbing foreground screen the observed apparent depth of this feature is proportional to the optical extinction. The general variation of the obscuration inside the ILR region of NGC 1365 revealed by the optical imaging can be traced with the spectral map of the silicate feature. As can be seen from Fig. \[fig:silicatesPAHratio\], the apparent strength of the silicate feature is high (in absorption) in the region encompassing clusters M4, M5, and M6 to the northeast of the AGN, almost zero at the location of the AGN, and intermediate (also in absorption) around clusters M2 and M3. In Table \[tab:spectroscopy\] we list the strength of the silicate feature measured in the SL 1D spectra of selected regions. Assuming a foreground screen of dust and adopting the [@Rieke1985] extinction law ($A_V/S_{\rm Si} = 16.6$), the observed apparent depths of the star-forming regions in the circumnuclear region of NGC 1365 imply values of the visual extinction of between $A_V \sim 4\pm1\,$mag and $A_V \sim 14\pm1\,$mag for the regions of clusters M2+M3,and M4+M5+M6, respectively. The high value of the extinction derived for the latter region is entirely consistent with the values derived by [@Galliano2008] for the individual clusters using hydrogen recombination lines. Our value of the extinction in the region containing M2 and M3 is higher than the optical estimate for the L2 and L3 H[ii]{} regions from [@Kristen1997]. The apparent strength of the silicate feature of the AGN measured from the Spitzer/IRS SL spectrum is $S_{\rm Si}=-0.10\pm 0.05$ (Table \[tab:spectroscopy\]) indicating that the feature is present slightly in absorption. This value, however, is contaminated by extended emission as is apparent from the $8\,\mu$m image (see Fig. \[fig:closeups\]) and the presence of PAH features in the IRS nuclear spectrum (Fig. \[fig:SLspectra\]). Indeed, the Gemini/T-ReCS high angular resolution spectrum of the AGN is mostly a featureless continuum with no evidence of PAH feature emission (Fig. \[fig:torusmodel\]). Finally, the relatively moderate strength of the silicate feature of the integrated $\sim 30\arcsec \sim 2.7\,$kpc central region ($S_{\rm Si}=-0.45\pm0.05$, Table \[tab:spectroscopy\]) of NGC 1365[^4] is typical of the observed nuclear silicate strengths measured in other local LIRGs [@Pereira2010IRSmapping; @AAH11LIRGs] and indicates an average visual extinction in this region of $A_V\sim 7\pm1\,$mag. Dust Color Temperatures {#sec:colortemperature} ----------------------- The Herschel images can be used to trace the spatial variations of the temperature of the dust inside the ILR region of NGC 1365. If we assume that the far-IR emission of a galaxy can be approximated with a modified blackbody, then in the case of optically thin emission the flux density can be expressed as $$\label{eq:modifiedblackbody} f_{\nu} \propto \nu^{\beta}\,B(\nu, T_{\rm dust})$$ where $B(\nu,T_{\rm dust})$ is the blackbody function for a dust temperature of $T_{\rm dust}$ and $\beta$ is the dust emissivity. In the simplest approximation we can calculate the color temperature $T_{\rm c}$ of the dust using the ratio of the surface brightness at two wavelengths with the following equation expressed in terms of wavelengths $$\label{eq:colortemperature} \frac{f_{\nu}(\lambda_1)}{f_{\nu}(\lambda_2)}= \left(\frac{\lambda_2}{\lambda_1}\right)^{3+\beta}\, \left(\frac{{\rm e}^{hc/\lambda_2KT_{\rm c}}-1} {{\rm e}^{hc/\lambda_1KT_{\rm c}}-1}\right)$$ We chose to construct a map of the PACS $100\,\mu$m to the PACS $70\,\mu$m ratio as these two wavelengths provide the best angular resolutions with Herschel. We rebinned the PACS $70\,\mu$m image to the pixel size of the PACS $100\,\mu$m image. We matched the PSF of the two images by smoothing the $70\,\mu$m with a Gaussian function, although this may introduce some artifacts [see e.g., @Bendo2010; @Bendo2012]. We then calculated the dust color temperature by solving Equation \[eq:colortemperature\]. We fixed the value of the dust emissivity $\beta=2$, as it is found to fit well the integrated SEDs of local IR galaxies [@Dunne2001]. The map of the PACS $100\,\mu$m to $70\,\mu$m ratio of the inner $\sim 60\arcsec$ is shown in Fig \[fig:colortemperature\]. The observed range of $f_\nu(100\,\mu{\rm m})/f_\nu (70\,\mu{\rm m})$ in the central region translates into values of the dust color temperature of between $T_{\rm c} (100\mu{\rm m}/70\mu{\rm m})\sim 32\,$K and $T_{\rm c}(100\mu{\rm m}/70\mu{\rm m})\sim 22\,$K. Within the ILR region of NGC 1365 the highest $T_{\rm c} (100\mu{\rm m}/70\mu{\rm m})$ color temperatures correspond to regions actively forming stars, that is, the regions containing the mid-IR star clusters and the L4 H$\alpha$ hot spot. These bright star forming regions have color temperatures similar to, although slightly higher than, those of H[ii]{} regions in the spiral arms. The regions with the coldest $T_{\rm c} (100\mu{\rm m}/70\mu{\rm m})$ in the ILR region might be associated with some of the foreground dust features seen in the optical images (see Fig. \[fig:PACS100\_large\]). Finally, the region of the AGN does not appear different in terms of the color temperature when compared to the bright star forming regions. This is probably because the AGN is faint in the far-IR (see next section) coupled with the relatively coarse angular resolution of this map (i.e., that of the PACS $100\,\mu$m). Dust Properties {#sec:dustproperties} --------------- In this section we study the properties of the dust in the ILR region of NGC 1365, particularly the dust heated by processes other than the AGN. To do so, we first subtracted the AGN emission (see Table \[tab:AGNfluxes\]) from the observed fluxes in the Spitzer/MIPS $24\,\mu$m band and all the Herschel bands. We then fitted the observed non-AGN mid- and far-IR SED using a combination of two modified blackbody following [@Dunne2001] and [@Clements2010]: $$f_\nu = N_{\rm w} \nu^\beta B(\nu, T_{\rm w}) + N_{\rm c} \nu^\beta B(\nu, T_{\rm c})$$ where $N_{\rm w}$ and $N_{\rm c}$ are the relative masses of the warm and cold dust components and $T_{\rm c}$ and $T_{\rm w}$ are their temperatures. As done in Section \[sec:colortemperature\], we fixed the dust emissivity $\beta = 2$ and then used a standard $\chi^2$ minimization technique to fit the two dust temperatures and the relative masses of the two dust components. For the fit we used the fluxes measured through an $r=30\arcsec$ aperture, and normalized them to that at $100\,\mu$m. For the Spitzer/MIPS $24\,\mu$m flux density, we added in quadrature the photometric error and the uncertainty associated with subtracting the AGN component at this wavelength to a total error budget of $\sim 20\%$. For the Herschel data points because the AGN emission in the far-IR is very small compared to the total emission inside the ILR region, the dominant source of error is that associated with the photometric calibration of the data: $\sim 10\%$. Fig. \[fig:SEDfit\] shows the best fit to the AGN-subtracted SED of the ILR region of NGC 1365, which was obtained with $T_{\rm c}=24\,$K and $T_{\rm w}=54\,$K, and relative masses of cold to warm dust of $N_{\rm c}/N_{\rm w}\sim 120$. It is worth noting that the fit using two modified blackbodies is formally better than a fit to the far-IR data (i.e., excluding the $24\,\mu$m data point) using a single modified blackbody. This can be explained because there is a non-negligible contribution from the warm dust component to the Herschel PACS $70\,\mu$m flux density, as found for the integrated emission of other nearby galaxies [see e.g., @Bendo2010; @Smith2010]. However, the temperature of the warm component is not tightly constrained since varying the warm dust temperature by as much as $\pm 4\,$K produces statistically similar good fits ($\chi^2 \le 2\,\chi^2_{\rm min}$). This is because the peak of this component at around $50\,\mu$m is not well sampled with the present data and the mass contribution of the warm component is small. The temperature of the cold component, on the other hand is well constrained. To get an estimate of the uncertainties on the fitted cold dust temperature, we changed it while fixing the temperature of the warm component. As can be seen from Fig. \[fig:SEDfit\], cold dust temperatures in the range $T_{\rm c}=24\pm 1\,$K provide acceptable fits to the SED. Previous works showed that the integrated mid-IR and far-IR SEDs of Seyfert galaxies could be fitted with a combination of various dust temperatures probing different heating mechanisms. These include warm dust associated with the AGN torus, cold dust similar to that observed in starburst galaxies, and very cold dust at temperatures characteristic of dust heated by the interstellar medium [@PerezGarcia2001; @Spinoglio2002]. With the superior angular resolution of the Herschel observatory we can now obtain spatially resolved observations of nearby Seyfert galaxies and star forming galaxies. The dust temperature of the ring of star formation in NGC 1365 is similar to that of other nuclear and circumnuclear starbursts with or without an AGN with similar IR luminosities and physical sizes, for instance, the ring of NGC 3081 [@RamosAlmeida2011Herschel] and the nuclear region of M83 [@Foyle2012]. On the other hand, the dust of Mrk 938, which is part of our survey of Seyfert galaxies, has a considerably higher dust temperature of $T=36\pm 4\,$K, probably due to the smaller size of the IR emitting region of this galaxy and higher IR luminosity [@Esquej2011]. We calculated the dust mass ($M_{\rm dust}$) within the ILR region of NGC 1365 using the following equation from [@Clements2010] [adapted from @Hildebrand1983]: $$M_{\rm dust} = \frac{f_\nu\,D^2}{\kappa_{\rm dust}(\nu)} \times \left( \frac{N_{\rm c}}{B(\nu,T_{\rm c})} + \frac{N_{\rm w}}{B(\nu,T_{\rm w})}\right)$$ where $f_\nu$ is the observed flux density, $D$ is the luminosity distance, $B(\nu,T)$ is the blackbody emission for the best fitting dust temperatures, and $\kappa_{\rm d}(\nu)$ is the dust absorption coefficient. As done by [@Esquej2011], we evaluated this expression at $250\,\mu$m using an absorption coefficient of $\kappa_{\rm dust}(250\mu{\rm m}) = 4.99\,{\rm cm}^2\,{\rm g}^{-1}$, as interpolated from the dust model of [@Li2001]. We derived a dust mass in the ILR region of $M_{\rm dust}({\rm ILR})=6.9\times 10^7\,{\rm M}_\odot$, which accounts for approximately 25% of the total dust mass of this galaxy $M_{\rm dust} ({\rm total}) = 3\times10^8\,{\rm M}_\odot$. The latter estimate is from [@Wiebe2009] but recalculated for the absorption coefficient used in this work. Star Formation Rate {#sec:sfr} ------------------- The ages and masses of the mid-IR clusters [$\sim 6-8\,$Myr and $\sim 10^7\,{\rm M}_\odot$, @Galliano2008] and the large amount of molecular gas available in the central regions of this galaxy [@Sakamoto2007] indicate that there is intense on-going star formation activity there. We can use the IR observations of NGC 1365 to estimate the obscured SFR in the ILR region, and compare it with the unobscured SFR. As discussed by [@Kennicutt2009], a combination of the observed (not corrected for extinction) H$\alpha$ luminosity and a mid-IR monochromatic luminosity (preferably $24\,\mu$m) will provide the best estimate of the total SFR in a moderately obscured environment (see Section \[sec:silicatefeature\]), such as the ILR region of NGC 1365. We note that this empirically calibrated recipe includes contributions from dust heating from all stars and not only the youngest [see @Kennicutt2009 for a full discussion]. Using the H$\alpha$ flux of [@Forster2004] for a 40-diameter aperture and our $24\,\mu$m flux (after subtracting the AGN component) we estimated a total SFR within the ILR region (inner $\sim 5.4\,$kpc, $r=30\arcsec$) of NGC 1365 of SFR$=7.3\,{\rm M}_\odot\,{\rm yr}^{-1}$ for a Salpeter initial mass function (IMF). Approximately 85% of this SFR is contributed by the $24\,\mu$m emission, and thus it originates from dust-obscured star forming regions. Most of the IR emission in the ILR region comes from the star formation ring containing the bright mid-IR clusters identified by [@Galliano2005]. Indeed, we estimate that within the inner $2.7\,{\rm kpc}$ the total SFR is $5.6\,{\rm M}_\odot\,{\rm yr}^{-1}$. This has been calculated using the AGN-subtracted $24\,\mu$m flux density ($r=15\arcsec$) and the 24-diameter aperture H$\alpha$ flux from [@Forster2004]. This corresponds to a SFR surface density in the ring of $\Sigma_{\rm SFR} =2.2 \,{\rm M}_\odot\,{\rm yr}^{-1}\,{\rm kpc}^{-2}$, for a ring radius of 900pc. This value of the SFR density is similar to those of other circumnuclear starbursts [@Kennicutt1998] and is expected to be larger by factors of $100-1000$ compared to the disk of the galaxy [@Elmegreen1994]. Summary and Conclusions ======================= In this paper we have studied the IR emission associated with the star formation activity in the ILR region of NGC 1365, as well as the IR emission of the AGN. To this end we have analyzed new far-IR ($70-500\,\mu$m) Herschel/PACS and SPIRE imaging, and high angular resolution ($\sim 0.4\arcsec$) Gemini/T-ReCS mid-IR imaging and spectroscopy of this galaxy. We have also made use of archival Spitzer/IRAC and MIPS imaging and IRS spectral mapping data. Our main findings for the inner $D \sim 5\,$kpc region of NGC 1365 are: - The new Herschel PACS imaging data at 70, 100, and $160\,\mu$m reveal that the ring of star formation in the ILR region is bright in the far-IR. The AGN is the brightest mid-IR source in the inner 2kpc up to $\lambda\simeq 24\,\mu$m, but it becomes increasingly fainter in the far-IR when compared with the mid-IR clusters or groups of them in the ring. - The 24 and $70\,\mu$m emissions as well as the \[Ne[ ii]{}\]$12.81\,\mu$m line and PAH features trace the star-forming ring in the ILR region and have morphologies similar to the CO “twin-peaks”. This all indicates that there is intense on-going star formation taking place in the inner few kpc of NGC 1365. - The unresolved near and mid-IR nuclear emission and mid-IR spectrum (i.e., AGN-dominated emission) of NGC 1365 are well reproduced with a relatively compact torus (outer radius of $R_{\rm o}=5^{+0.5}_{-1}\,$pc) with an opening angle of $\sigma_{\rm torus}=36^{+14}_{-6}$deg, and an AGN bolometric luminosity $L_{\rm bol}({\rm AGN})=2.6\pm0.5\times 10^{43}\,{\rm erg \, s}^{-1}$ using the [clumpy]{} torus models. These parameters are in good agreement with independent estimates in the literature. - Using the fitted torus model we quantified the AGN emission in the far-IR. The AGN only contributes at most 1% of the $70\,\mu$m emission within the inner 5.4kpc ($r=30\arcsec$), and less than 1% at longer wavelengths. At $24\,\mu$m the AGN accounts for $\sim 15\%$ of the emission in the same region. We estimated that the AGN bolometric contribution to the $3-1000\,\mu$m luminosity in the inner 5.4kpc is approximately 5%. - The non-AGN 24 to $500\,\mu$m SED of the ILR region (inner 5.4kpc) of NGC 1365 is well fitted with a combination of two modified blackbodies with warm and cold temperatures of 54K and 24K, respectively. However, the cold dust component accounts for most of total dust mass inferred in this region ($M_{\rm dust}({\rm ILR})= 7\times 10^7\,{\rm M}_\odot$) and has a temperature similar to that of other nuclear and circumnuclear starbursts of similar sizes and IR luminosities. - From the comparison between the SFR from H$\alpha$ (unobscured) and the SFR from $24\,\mu$m (obscured) we infer that up to $\sim 85\%$ of the on-going SFR inside the ILR region of NGC 1365 is taking place in dust-obscured regions in the ring of star formation. Acknowledgments {#acknowledgments .unnumbered} =============== We are grateful to B. Elmegreen and E. Galliano for providing us with the BVR map of NGC 1365 shown in the right panel of Figure 1. We also thank Andrés Asensio Ramos for developing the BayesClumpy fitting routine. We thank an anonymous referee for comments that helped improve the paper. A.A.-H., M.P.-S., and P.E. acknowledge support from the Spanish Plan Nacional de Astronomía y Astrofísica under grant AYA2009-05705-E. A.A.-H. also acknowledges support from the Universidad de Cantabria through the Augusto González Linares Program and AYA2010-21161-C02-01. C.R.A. acknowledges the Spanish Ministry of Science and Innovation (MICINN) through project Consolider-Ingenio 2010 Program grant CSD2006-00070: First Science with the GTC (http://www.iac.es/consolider-ingenio-gtc/) and the Spanish Plan Nacional grant AYA2010-21887-C04.04. M.P. acknowledges the Junta de Andalucía and the Spanish Ministry of Science and Innovation through projects PO8-TIC-03531 and AYA2010-15169, respectively. [*Herschel*]{} is an ESA space observatory with science instruments provided by European-led Principal Investigator consortia and with important participation from NASA. PACS has been developed by a consortium of institutes led by MPE (Germany) and including UVIE (Austria); KU Leuven, CSL, IMEC (Belgium); CEA, LAM (France); MPIA (Germany); INAF-IFSI/OAA/OAP/OAT, LENS, SISSA (Italy); IAC (Spain). This development has been supported by the funding agencies BMVIT (Austria), ESA-PRODEX (Belgium), CEA/CNES (France), DLR (Germany), ASI/INAF (Italy), and CICYT/MCYT (Spain). SPIRE has been developed by a consortium of institutes led by Cardiff Univ. (UK) and including: Univ. Lethbridge (Canada); NAOC (China); CEA, LAM (France); IFSI, Univ. Padua (Italy); IAC (Spain); Stockholm Observatory (Sweden); Imperial College London, RAL, UCL-MSSL, UKATC, Univ. Sussex (UK); and Caltech, JPL, NHSC, Univ. Colorado (USA). This development has been supported by national funding agencies: CSA (Canada); NAOC (China); CEA, CNES, CNRS (France); ASI (Italy); MCINN (Spain); SNSB (Sweden); STFC, UKSA (UK); and NASA (USA). Based on observations obtained at the Gemini Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under a cooperative agreement with the NSF on behalf of the Gemini partnership: the National Science Foundation (United States), the Science and Technology Facilities Council (United Kingdom), the National Research Council (Canada), CONICYT (Chile), the Australian Research Council (Australia), Ministério da Ciéncia, Tecnologia e Inovação (Brazil) and Ministerio de Ciencia, Tecnología e Innovación Productiva (Argentina). 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. [99]{} Alloin D., Edmunds M. G., Lindblad P. O., Pagel B. E. J., 1981, A&A, 101, 377 Alonso-Herrero A., Pereira-Santaella M., Rieke G. H., Rigopoulou D., 2012, ApJ, 744, 2 Alonso-Herrero A., Ramos Almeida C., Mason R. et al., 2011, ApJ, 736, 82 Alonso-Herrero A. et al., 2009, ApJ, 697, 660 Alonso-Herrero A. et al., 2006, ApJ, 650, 835 Armus L. et al., 2009, PASP, 121, 559 Asensio Ramos A., Ramos Almeida C., 2009, ApJ, 696, 2075 Bendo G. J. et al., 2012, MNRAS, 419, 1833 Bendo G. J. et al., 2010, A&A, 518, L65 Brandl B. et al., 2006, ApJ, 653, 1129 Buta R., Combes, F., 1996, Fundam. Cosm. Phys., 17, 95 Calzetti D. et al., 2007, ApJ, 666, 870 Carollo C. M., Stiavelli M., Seigar M., de Zeeuw P. T., Dejonghe H., 2002, AJ, 123,159 Clements D. L., Dunne L., Eales S., 2010, MNRAS, 403, 274 Dale D., Helou G., 2002, ApJ, 576, 159 Dunne L., Eales S. A., 2001, MNRAS, 327, 697 Elmegreen B. G., Galliano E., Alloin D., 2009, ApJ, 703, 1297 Elmegreen B. G., 1994, ApJ, 425, L73 Esquej P., Alonso-Herrero A., Pérez-García A. M. et al., 2012, MNRAS, in press (astro-ph/1202.4577) Fazio G. G. et al., 2004, ApJS, 154, 10 Forbes D. A., Norris R. P., 1998, MNRAS, 300, 757 Förster Schreiber N. M., Roussel H., Sauvage M., Charmandaris V., 2004, A&A, 419, 501 Foyle K., et al., 2012, MNRAS, 421, 2917 Galliano E., Alloin D., Pantin E., Lagage P. O., & Marco O., 2005, A&A, 438, 803 Galliano E., Alloin D., Pantin E., Granato G. L., Delva P., Silva L., Lagage P. O., Panuzzo P., 2008a, A&A, 492, 3 Galliano F., Madden S. C., Tielens A. G. G. M., Peeters E., Jones A. P., 2008b, ApJ, 679, 310 Gallimore J. F. et al., 2010, ApJS, 187, 172 García-Burillo S., Combes F., Schinnerer E., Boone F., Hunt L. K., 2005, A&A, 441, 1011 Garc[í]{}a-Burillo S., Fern[á]{}ndez-Garc[í]{}a S., Combes F. et al., 2009, A&A, 496, 85 Griffin M.J., Abergel A., Abreu A. et al., 2010, A&A, 518, L3 Hatziminaoglou E. et al., 2010, A&A, 518, L33 Hildebrand R. H., 1983, QJRAS, 24, 267 Ho L. C., Keto E., 2007, ApJ, 658, 314 Hönig S. F., Beckert T., Ohnaka K., Weigelt G., 2006, A&A, 452, 459 Hopkins P. F., Quataert E., 2011, MNRAS, 415, 1027 Houck J. R. et al., 2004, ApJS, 154, 18 Hunt L. K., Combes F., Garc[í]{}a-Burillo S. et al., 2008, A&A, 482, 133 Jogee S., 2006, in Alloin D., ed., Lecture Notes in Physics Vol. 693, Physics of Active Galactic Nuclei at all Scales. Springer-Verlag, Berlin , p. 143 Jungwiert B., Combes F., Axon D. J., 1997, A&AS, 125, 479 Kenney J. D. P., Wilson C. D., Scoville N. Z., Devereux N. A., Young, Y. S., 1992, ApJ, 395, L79 Kennicutt R. C. Jr., 1998, ARA&A, 36, 189 Kennicutt R. C. Jr. et al., 2009, ApJ, 703, 1672 Kristen H., Jörsäter S., Lindblad P. O., Boksenberg A., 1997, A&A, 328, 483 Li Y. et al., 2010, ApJ, 725, 677 Li A., Draine B. T., 2001, ApJ, 554, 778 Lindblad P. A. B., Lindblad P. O., Athanassoula E., 1996, A&A, 313, 65 Lindblad P. O., 1999, A&ARv, 9, 221 Lira P., Videla L., Wu Y., Alonso-Herrero A., Alexander D. M., Ward M., 2012, ApJ, submitted Mason R. E., Levenson N. A., Shi,Y., Packham C., Gorjian V., Cleary K., Rhee J., Werner M., 2009, ApJ, 693, L136 Mullaney J. R., Alexander D. M., Goulding A. D., Hickox R. C., 2011, MNRAS, 414, 1082 Nenkova M., Sirocky M. M., Ivezić Z., Elitzur M., 2008a, ApJ, 685, 145 Nenkova, M., Sirocky M. M., Nikkuta R., Ivezić Z., Elitzur M., 2008b, ApJ, 685, 160 Netzer H. et al., 2007, ApJ, 666, 806 Nikutta R., Elitzur M., Lacy M., 2009, ApJ, 707, 1550 Peeters E., Spoon H. W. W., Tielens A. G. G. M., 2004, ApJ, 613, 986 Pereira-Santaella M., Alonso-Herrero A., Rieke G. H., Colina L. et al., 2010b, ApJS, 188, 447 Pereira-Santaella M., Diamond-Stanik A. M., Alonso-Herrero A., Rieke G. H., 2010a, ApJ, 725, 2270 Pérez García A. M., Rodríguez Espinosa J. M., 2001, ApJ, 557, 39 Pilbratt G.L., Riedinger J.R., Passvogel T. et al., 2010, A&A, 518, L1 Poglitsch A., Waelkens C., Geis N. et al., 2010, A&A, 518, L2 Ramos Almeida C., Sánchez-Portal M., Pérez-García A. M. et al., 2011a, MNRAS, 417, L46 Ramos Almeida C., Levenson N. A., Alonso-Herrero A., et al., 2011b, ApJ, 731, 92 Ramos Almeida C., Levenson N. A., Rodríguez Espinosa J. M. et al., 2009, ApJ, 702, 1127 Rieke G. H., Lebofsky M. J., 1985, ApJ, 288, 618 Rieke G. H. et al., 2004, ApJS, 154, 25 Rieke G. H., Alonso-Herrero A., Weiner B. J., Pérez-González P. G., Blaylock M., Donley J. L., Marcillac D., 2009, ApJ, 692, 556 Risaliti G., Elvis M., Fabbiano G., Baldi A., Zezas, A., 2005, ApJ, 623, L93 Roche P. F., Aitken D. K., Smith C. H., Ward M. J., 1991, MNRAS, 248, 606 Roussel H. 2012, in press (astro-ph/1205.2576) Sakamoto K., Ho P. T. P., Mao R.-Q., Matsushita S., Peck A. B., 2007, ApJ, 654, 782 Sales D. A., Pastoriza M. G., Riffel R., Winge C., 2010, ApJ, 725, 605 Sales D. A., Pastoriza M. G., Riffel R., Winge C., Rodríguez-Ardila A., Carciofi A. C., 2011, ApJ, 738, 109 Sanders D. B., Mazzarella J. M., Kim D.-C., Surace J. A., Soifer B. T., 2003, AJ, 126, 1607 Sandqvist Aa., Jörsäter S., Lindblad P. O., 1995, A&A, 295, 585 Schartmann M., Meisenheimer K., Camenzind M., Wolf S., Tristram K. R. W., Henning T., 2008, A&A, 482, 67 Schulz H., Komossa S., Schmitz C., Mücke A., 1999, A&A, 346, 764 Sérsic J. L., Pastoriza, M., 1965, PASP, 77, 287 Shlosman I., Frank J., Begelman M. C., 1989, Nature, 338, 45 Siebenmorgen R., Krügel E., Spoon H. W. W., 2004, A&A, 414, 123 Smith M. W. L. et al., 2010, A&A, 518, L51 Smith J. D. T. et al., 2007a, PASP, 119, 1133 Smith J. D. T. et al., 2007b, ApJ, 656, 770 Spinoglio L., Andreani P., Malkan M. A., 2002, ApJ, 572, 105 Spoon H. W. W. et al., 2007, ApJ, 654, L49 Telesco C. M., Dressler L. L., Wolstencroft R. D., 1993, ApJ, 414, 120 Telesco C. M., Piña R. K., Hanna K. T., Julian J. A., Hon D. B., Kisko T. M., 1998, Proc. SPIE, 3354, 534 Tristram K. R. W., et al. 2009, A&A, 502, 67 Wang J., Fabbiano G., Elvis M., Risaliti G., Mazzarella J. M., Howell J. H., Lord S., 2009, ApJ, 694, 718 Wiebe D. W., et al. 2009, ApJ, 707, 1809 \[lastpage\] [^1]: Herschel is an ESA space observatory with science instruments provided by European-led Principal Investigator consortia and with important participation from NASA. [^2]: The inner radius of the torus in these models is set by the dust sublimation temperature, which is assumed to be $T_{\rm sub} \sim 1500\,$K, and the AGN bolometric luminosity $L_{\rm bol} ({\rm AGN})$. Then the dust sublimation radius in pc is $R_{\rm d}=0.4\,L_{\rm bol} ({\rm AGN})^{0.5}$, where the AGN bolometric luminosity is in units of $10^{45}\,{\rm erg \,s}^{-1}$. [^3]: Note that the H$\alpha$+\[N[ii]{}\] image shown in Fig. \[fig:closeups\] has not been continuum subtracted, and thus we restrict our discussion to the H$\alpha$ bright hot spots identified by other works [e.g., @Alloin1981; @Kristen1997]. [^4]: The $8-1000\,\mu$m IR luminosity of this galaxy is $\log (L_{\rm IR}/{\rm L}_\odot) = 11.03$ using the IRAS flux densities from [@Sanders2003].

--- abstract: 'We have performed a high resolution search for galactic halo axions in cold flows using a microwave cavity detector. The analysis procedure and other details of this search are described. No axion signal was found in the mass range 1.98–2.17 $\mu$eV. We place upper limits on the density of axions in local discrete flows based on this result.' author: - 'L. D. Duffy' - 'P. Sikivie' - 'D. B. Tanner' - 'S. J. Asztalos' - 'C. Hagmann' - 'D. Kinion' - 'L. J Rosenberg' - 'K. van Bibber' - 'D. B. Yu' - 'R. F. Bradley' date: 'March 3, 2006' title: 'A High Resolution Search for Dark-Matter Axions' --- Introduction ============ In the current concordance cosmology, 23% of the universe’s total energy density is contributed by exotic dark matter [@WMAP]. The axion, arising from the Peccei-Quinn solution to the strong CP problem [@PQWW], satisfies the two criteria necessary for cold dark matter [@realign]: (1) a very cold population of axions could be present in our universe in sufficient quantities to provide the required dark matter energy density and (2) axions are effectively collisionless; i.e., their only significant long-range interaction is gravitational. The Axion Dark Matter eXperiment (ADMX) [@ADMX] uses a Sikivie microwave cavity detector [@detect] to search for axions in our galactic halo. The power in an axion signal observed by a microwave cavity detector is proportional to the local axion density. The signal width is caused by the velocity dispersion of dark-matter axions. Therefore, in searching for axions it is necessary to make some assumptions about their velocity distribution in our galactic halo. A variety of galactic halo models have been put forward: the isothermal model, results from N-body simulations [@nbody] and the caustic ring model [@cau1; @cau2]. The predictions of these models are used to guide ADMX’s search. In the isothermal model it is expected that a significant fraction of the dark matter halo will have an isothermal velocity distribution resulting from a period of “violent relaxation” of the early galaxy [@violentrelax]. This component of the halo will have velocities described by a Maxwell-Boltzmann distribution. The “medium resolution” (MR) channel [@MR] searches for these axions, assuming that the velocity dispersion is $\mathcal{O}(10^{-3}c)$ or less. (The escape velocity from our galaxy for axions is approximately $2\times10^{-3}c$.) N-body simulations and the caustic ring model both predict substructure within halos. Numerical simulations indicate that hundreds of smaller clumps, or subhalos, exist within the larger halo [@nbody]. Tidal disruption of these subhalos leads to flows in the form of “tidal tails” or “streams”. The Earth may currently be in a stream of dark matter from the Sagittarius A dwarf galaxy [@streams]. Non-thermalized flows from late infall of dark matter onto the halo are also expected [@ips]. Insufficient time has elapsed for dark matter that has fallen into the gravitational potential of the galaxy after violent relaxation to thermalize with the rest of the halo. Matter which has fallen onto the galaxy only recently will be present in the halo in the form of discrete flows. There will be one flow of particles falling into the gravitational potential for the first time, one flow of particles falling out for the first time, one due to particles falling in for the second time, etc. Furthermore, where the gradient of the particle velocity diverges, particles “pile up” and form caustics. In the limit of zero flow velocity dispersion, caustics have infinite particle density. The velocity dispersion of cold axions at a time, $t$, prior to galaxy formation is approximately $\delta v_a \sim 3\times10^{-17}(10^{-5}\;\mathrm{eV}/m_a) (t_0/t)^{2/3}$ [@cau2], where $t_0$ is the present age of the universe and $m_a$ is the axion mass, constrained to lie between $10^{-6}$ and $10^{-2}$ eV by cosmology and astrophysical processes [@axrev]. Thus, a flow of dark matter axions will have a small velocity dispersion, leading to large, but finite density at the location of a caustic. The caustic ring model predicts that the Earth is located near a caustic feature [@MW]. Fitting the model to bumps in the Milky Way rotation curve and a triangular feature seen in the IRAS maps predicts that the flows falling in and out of the halo for the fifth time contain a significant fraction of the local halo density. The predicted densities are $1.7\times10^{-24}$ g/cm$^3$ and $1.5\times10^{-25}$ g/cm$^3$ [@MW], comparable to the local dark matter density of $9.2\times10^{-25}$ g/cm$^3$ predicted in [@density]. The flow of the greatest density is referred to as the “Big Flow”. The possible existence of discrete flows, or streams, provides an opportunity to increase the ADMX discovery potential. A discrete axion flow produces a narrow peak in the spectrum of microwave photons in the experiment and such a peak can be searched for with higher signal-to-noise than a signal from axions in an isothermal model halo. The “high resolution” (HR) channel was built to take advantage of this opportunity. Furthermore, if a signal is found, the HR channel will provide us with detailed information on the structure of the Milky Way halo. The HR channel is the most recent addition to ADMX, implemented as a simple addition to the receiver chain, running in parallel with the MR channel. This channel and the possible existence of discrete flows can improve ADMX’s sensitivity by a factor of 3 [@HR], significantly enhancing its discovery potential. The full ADMX detector is described in Section \[sec:experiment\]. Each discrete flow of cold axions with small velocity dispersion will be seen as a narrow peak in the detector’s output spectrum. Our expectations for a signal are discussed in Section \[sec:analysis\], which also contains the details of the HR analysis, the primary topic of this paper. After a full search of the frequency range 478–525 MHz, no axion signals were found and we place limits on the density of cold axions in discrete flows in Section \[sec:results\]. This limit is compared to our previous results for the MR channel and halo substructure predictions in Section \[sec:discuss\]. Axion Dark Matter eXperiment {#sec:experiment} ============================ ADMX uses a microwave cavity detector to search for axions in our galactic halo. We outline the principle of the detector and briefly describe ADMX. Further details of the experiment can be found in [@ADMX; @Peng:2000hd]. The microwave cavity detector utilizes the axion-electromagnetic coupling to induce resonant conversion of axions to photons. The relevant interaction is $$\mathcal{L}_{a\gamma\gamma}=g_{\gamma} \frac{\alpha}{\pi} \frac{a(x)}{f_a} \mathbf{E}\cdot\mathbf{B} \; ,$$ where $\mathbf{E}$ and $\mathbf{B}$ are the electric and magnetic fields, $\alpha$ is the fine structure constant, $f_a$ is the axion decay constant, $a(x)$ is the axion field and $g_{\gamma}$ is a model-dependent coupling, of order one. In the Kim-Shifman-Vainshtein-Zakharov (KSVZ) model [@KSVZ], $g_{\gamma} = -0.97$, and in the Dine-Fischler-Srednicki-Zhitnitsky (DFSZ) model [@DFSZ], $g_{\gamma} = 0.36$. Axions in the galactic halo are non-relativistic, i.e. the energy of a single axion with mass $m_a$ and velocity $v$ is $$E_a=m_ac^2+\frac12m_a v^2 \; , \label{eq:NRE}$$ where $c$ is the speed of light. The axion-to-photon conversion process conserves energy, i.e. an axion of energy $E_{a}$ converts to a photon of frequency $\nu=E_{a}/h$. When $\nu$ falls within the bandwidth of a cavity mode, the conversion process is resonantly enhanced. The signal is a peak in the spectrum output by the detector. The power, $P$, developed in the cavity due to resonant axion-photon conversion is [@detect] $$\label{eqn-convpower} P=\left(\frac{\alpha g_{\gamma}}{\pi f_{a}}\right)^{2} \frac{VB_{0}^{2}\rho_{a}C}{m_{a}}\min(Q,Q_{a})\; ,$$ where $V$ is the cavity volume, $B_0$ is the magnetic field strength, $\rho_{a}$ is the local density of axions with energy corresponding to the cavity frequency, $Q$ is the loaded quality factor of the cavity, $Q_{a}$ is the ratio of the energy of the halo axions to their energy spread, equivalent to a “quality factor” for the halo axion signal, and $C$ is a mode dependent form factor which is largest for the fundamental transverse magnetic mode, $TM_{010}$. The quantity $C$ is given by $$C=\frac{\left|\int_{V}d^{3}x\mathbf{E_{\omega}}\cdot\mathbf{B_{0}}\right|^{2}} {B_{0}^{2}V\int_{V}d^{3}x\epsilon|\mathbf{E_{\omega}}|^2}\mathrm{,}$$ in which $\mathbf{E_{\omega}}(\mathbf{x})e^{i\omega t}$ is the time dependent electric field of the mode under consideration, $\mathbf{B_{0}}(\mathbf{x})$ is the static magnetic field in the cavity and $\epsilon$ is the dielectric constant of the medium inside the cavity. The frequency-dependent form factor is evaluated numerically. Eq. (\[eqn-convpower\]) can be recast in the convenient form $$\begin{aligned} \label{e:power} P=&0.5\times10^{-21}\;\mathrm{W}\left({V\over 500\;\mathrm{L}}\right) \left({B_{0}\over 7\;\mathrm{T}}\right)^{2}C \left({g_{\gamma}\over 0.36}\right)^{2} \nonumber \\ &\times\left({\rho_{a}\over 0.5\times10^{-24}\; \mathrm{g.cm}^{-3}}\right) \left({\nu_{a}\over 1 \mathrm{GHz}} \right)\left({\mathrm{min}[Q,Q_{a}]\over 10^5}\right)\, ,\end{aligned}$$ where $\nu_a$ is the axion mass frequency. As the experiment operates with the cavity at critical coupling, half the power developed in the cavity is lost to its walls and half is passed to the receiver chain. The HR channel became fully operational in August, 2002. A schematic of ADMX, showing both the MR and HR channels, is given in Fig. \[fig-receiver\]. A more detailed illustration of the magnet, cavity and cryogenic components is shown in Fig. \[fig-detector\]. The microwave cavity has an inner volume, $V$, of 189 L. The frequency of the $TM_{010}$ mode can be tuned by moving a pair of metal or dielectric rods inside the cavity. The cavity is located in the bore of a superconducting solenoid, which generates a static magnetic field, $B_0$, of 7.8 T. The voltage developed across a probe coupled to the electromagnetic field inside the cavity is passed to the receiver chain. During typical operation, $Q$ is approximately $7\times10^{4}$ and the total noise temperature for the experiment, $T_{n}$, is conservatively estimated to be 3.7 K, including contributions from both the cavity and the receiver chain. The first segment of the receiver chain is common to both the MR and HR channels. It consists of a cryogenic GaAs HFET amplifier built by NRAO, a crystal bandpass filter and mixers. At the end of this segment, the signal is centered at 35 kHz, with a 50 kHz span. The MR signal is sampled directly after this part of the receiver chain. The HR channel contains an additional bandpass filter and mixer, resulting in a spectrum centered at 5 kHz with a 6 kHz span. Time traces of the voltage output from the HR channel, consisting of $2^{20}$ data points, are taken with a sampling frequency of $20$ kHz. This results in a data stream of 52.4 s in length, corresponding to 0.019 Hz resolution in the frequency spectrum. The data were primarily taken in parallel with the operations of the MR channel over a period beginning in November, 2002 and ending May, 2004. Continuous HR coverage has been obtained and candidate peak elimination performed for the frequency range 478–525 MHz. Data with $Q$ less than 40 000 and/or cavity temperature above 5 K were discarded. In these cases, additional data were taken to ensure coverage of the entire range. High resolution analysis {#sec:analysis} ======================== We use the HR channel to search for narrow peaks caused by flows of cold axions through the detector. The background is thermal and electronic noise plus narrow lines from the rf environment of the experiment. These environmental peaks are signals from nearby emitters (e.g.  computer clocks) which leak into the cavity by a variety of means. When placing limits on cold flows of axions, we assume that the flows are steady, i.e. the rates of change of velocity, velocity dispersion and density of the flows are slow compared to the time scale of the experiment. The assumption of a steady flow implies that the signal we are searching for is always present. Even so, the kinetic energy term in Eq.(\[eq:NRE\]) and the corresponding signal frequency change over time due to the Earth’s rotational and orbital motions. In addition to a signal frequency shift in data taken at different times, apparent broadening of the signal occurs because its frequency shifts while the data are being taken. In this section, we first describe the signal frequency shifts due to the Earth’s rotation and orbital motion, and the associated signal broadening (subsection A). Next, we describe the properties of the noise in the HR channel (subsection B). In subsection C, we describe how we correct each spectrum for systematic effects introduced by the receiver chain. Finally, in subsection D, we describe how candidate peaks are selected and shown not to be caused by cold flows of dark matter axions. Axion signal properties {#ss:sigprop} ----------------------- An axion signal will undergo diurnal and annual modulations due to the Earth’s rotation and orbital motion, respectively [@Ling:2004aj]. Thus, the frequency at which axions are resonantly converted to photons will shift. We show here that this frequency modulation cannot move a signal by an amount which is larger than the detector bandwidth. As the energy of an axion in the ADMX detector is non-relativistic (Eq. (\[eq:NRE\])), the shift in frequency of the signal, $\Delta f$, due to a change in velocity of the axion flow relative to the detector, $\Delta v$, is $$\Delta f = \frac{f v \Delta v}{c^2} \; . \label{eq:sigmove}$$ We have investigated the magnitude of both annual and diurnal signal modulation at $f=$ 500 MHz. The velocity of a dark matter flow relative to the Earth will be in the range 100–1000 km/s. We chose 600 km/s as a representative value for the purpose of estimation. For the daily modulation, we have assumed that the detector is located at the Earth’s equator and that, in the frame in which the axis of rotation of the Earth is stationary, the flow velocity is first aligned and then anti-aligned (or vice versa) with the detector’s motion due to the Earth’s rotation over the course of a day. These assumptions result in the largest possible change in relative velocity between the detector and the flow due to the Earth’s rotation. The Earth’s rotational velocity is 0.4 km/s at the equator. The resulting daily signal modulation is of order 1 Hz. For the annual modulation, to again maximize the change in relative velocity, we have also considered the case of extreme flow velocity alignment with the Earth’s orbital motion. The Earth’s orbital velocity is 30 km/s. The annual modulation produces the larger frequency shift, of order 100 Hz within a year. The HR channel has a 6 kHz bandwidth, ensuring that an axion peak will appear in spectra taken with center frequency equal to a previously observed axion signal frequency. The signal broadening, $\delta f$, due to a change $\delta v$ in flow velocity while data are being taken is $$\delta f = \frac{f v \delta v}{c^2} \; . \label{eq:sigbroad}$$ The most significant signal broadening is due to the Earth’s rotation. Using the same assumptions as before, we find that the broadening is at most $4\times10^{-3}$ Hz during the 52 s taken to acquire a single time trace. This is less than the spectral resolution of 0.019 Hz. The broadening due to the Earth’s orbital motion is only of order $10^{-4}$ Hz in this same time interval. As the signal broadening due to the Earth’s rotation and orbital motion is negligible, we can use Eq. (\[eq:sigbroad\]) to relate the width, $\delta f$, of a signal peak to the velocity dispersion, $\delta v$, of the axion flow that causes it. In general, we do not know the velocity dispersion of the cold axion flows which we search for, although we note that [@MW] claims an upper limit of 53 m/s on the velocity dispersion of the Big Flow. Subsequently, we do not know the signal width. To compensate, we perform our search at multiple resolutions by combining 0.019 Hz wide bins. These searches are referred to as $n$-bin searches, where $n =$ 1, 2, 4, 8, 64, 512 and 4096. For $f=500$ MHz and $v=300$ km/s, the corresponding flow velocity dispersions are $$\delta v_n= 12\,n \;\mathrm{m}/\mathrm{s}\left(\frac{300 \mathrm{km}/ \mathrm{s}}{v}\right) \; . \label{eq:vdisp}$$ Further details on the $n$-bin searhces are given in Section \[ss:sproc\]. Noise in the HR Channel {#ss:hrnoise} ----------------------- The power output from the HR channel is expressed in units of $\sigma$, the rms noise power. This noise power is related to the noise temperature, $T_n$, via $$\sigma=k_B T_n \sqrt{\frac{b}{t}} \; ,$$ where $k_B$ is Boltzmann’s constant, $b$ is the frequency resolution and $t$ is the acquisition time. The total noise temperature $T_n=T_{C}+T_{el}$, where $T_{C}$ is the physical cavity temperature and $T_{el}$ is the electronic noise contribution from the receiver chain. As no averaging is performed in HR sampling, $b=1/t$. Thus, the rms noise power is $$\sigma=k_B b T_n \; . \label{eq:Pnoise}$$ Output power is normalized to $\sigma$ and $T_n$ is used to determine this power. We verified Eq. (\[eq:Pnoise\]) experimentally by allowing the cavity to warm and observing that $\sigma$ is proportional to $T_{C}$. As this is our calibration of the power output from the cavity, it is important that we understand the noise in the HR channel. The noise in the HR channel is observed to have an exponential distribution. We now explain why this is expected. The noise in a single bin is the sum of independent sine and cosine components, as no averaging occurs. We expect that the noise amplitude, $a$, for a single component (i.e. sine or cosine) has a Gaussian probability distribution, $$\label{eqn-gauss} \frac{dP}{da}=\frac{1}{\sqrt{2\pi}\sigma_{a}}\exp\left(-\frac{a^{2}}{2 \sigma_{a}^{2}}\right) \, ,$$ where $\sigma_{a}$ is the standard deviation. Indeed, the energy distribution should be proportional to a Boltzmann factor, $\exp(-E/kT)$, and non-relativistic and classical energies, such as $E=mv^{2}/2$ or $E=kx^{2}/2$ are proportional to squares of the amplitude. As there are two components per bin, the addition of $n$ bins is that of $2n$ independent contributions. The sum of $2n$ independent normal-distributed components is described by a chi-square distribution with $2n$ degrees of freedom. Thus, the probability distribution for an $n$-bin is a $\chi^2(2n)$ distribution. We demonstrate this explicitly in the following. The probability distribution, $dP/dp_n$, of observing noise power $p_n$ in an $n$-bin is $$\frac{dP}{dp_n}=\left(\prod_{i=1}^{2n}\int^{\infty}_{-\infty}da_i\right) \frac{\exp(-\frac{1}{2\sigma_a^2}\sum_{j=1}^{2n} a_j^2)}{(\sqrt{2\pi} \sigma_a)^{2n}}\,\delta(p_n-\sum_{k=1}^{2n}\frac{a_k^2}2 )\, .$$ Evaluating the above expression, $$\frac{dP}{dp_{n}}=\frac{{p_{n}}^{n-1}}{(n-1)!\sigma_{a}^{2n}} \exp\left(-\frac{p_{n}}{\sigma_{a}^2}\right) \; . \label{eq:nprob}$$ For $n=1$, $$\label{eqn-sigaprob} \frac{dP}{dp_{1}}=\frac{1}{\sigma_{a}^{2}}\exp\left(-\frac{p_{1}}{\sigma_{a}^{2}}\right) \mathrm{,}$$ which is indeed a simple exponential, as expected. Using this noise distribution, we can easily see that the average (rms) noise power in the one bin search, $\sigma$, is $\sigma={\sigma_{a}}^2$. Substituting this in Eq. (\[eqn-sigaprob\]), the noise power distribution function becomes $$\label{eqn-npprob} \frac{dP}{dp_{1}}=\frac{1}{\sigma} \exp\left(-\frac{p_{1}}{\sigma}\right) \; .$$ For each individual spectrum, the baseline noise level, $\sigma$, is determined by plotting the number of frequency bins, $N_{p}$, with power between $p$ and $p+\Delta p$ against $p$. According to Eq. (\[eqn-npprob\]), $$N_{p}=\frac{N\Delta p}{\sigma}\exp\left(-\frac{p}{\sigma}\right) \; ,$$ where $N$ is the total number of data points. As $$\ln N_{p}=-\frac{p}{\sigma}+\ln\left(\frac{N\Delta p}{\sigma}\right) \; , \label{eqn-1bindist}$$ $\sigma$ is the inverse of the slope of the $\ln N_{p}$ versus $p$ plot. Fig. \[fig:stat1\] demonstrates that the data is in good agreement with this relation for $p$ less than $20\sigma$. The deviation of the data from Eq. (\[eqn-1bindist\]) for $p$ greater than $20\sigma$ is due to the fact that our background is not pure noise, but also contains environmental signals of a non-statistical nature. ![Power distribution for a large sample of 1-bin data.[]{data-label="fig:stat1"}](stat1small.eps) As we combine an increasing number of bins, the noise power probability distribution approaches a Gaussian, in accordance with the central limit theorem. The right-hand side of Eq. (\[eq:nprob\]) approaches a Gaussian in the limit of large $n$. We have examined a large sample of noise in each $n$-bin search and verified that it is distributed according to Eq. (\[eq:nprob\]). Figure \[fig:8stats\] illustrates statistics for the 8-bin search and Fig. \[fig:nstats\] shows statistics for the 4096-bin search, a near Gaussian curve. Figures \[fig:stat1\] through \[fig:nstats\] illustrates the progression from exponential to Gaussian noise power probability distribution. ![Power distribution for a large sample of 8-bin data.[]{data-label="fig:8stats"}](Figure4small.eps) ![Power distribution for a large sample of 4096-bin data.[]{data-label="fig:nstats"}](Figure5small.eps) In addition to examining the behavior of the noise statistics, we have performed a cross-calibration between the HR and MR channels. The signal power of an environmental peak, observed at 480 MHz and shown in Fig. \[fig:crosscal\], was examined in both the HR and MR channels. The observed HR signal power was $(1.8 \pm 0.1)\times10^{-22}$ W, where the error quoted is the statistical uncertainty. The MR channel observed signal power $1.7\times10^{-22}$ W, in agreement with the HR channel. Note that the MR signal was acquired with a much longer integration time than that of the HR signal (2000 s for MR versus 52 s for HR). ![An environmental peak as it appears in the MR search (top) and the 64–bin HR search. The unit for the vertical axis is the rms power fluctuation in each case.[]{data-label="fig:crosscal"}](crosscsmall.eps) The combination of the calibration of the noise power with cavity temperature, the consistency between expected and observed noise statistics, and the agreement of signal power observed in both the HR and MR channels, makes us confident that the signal power is accurately determined in the HR channel. Removal of systematic effects {#ss:system} ----------------------------- There are two systematic effects introduced in the receiver chain shown in Fig. \[fig-receiver\]. Two passband filters are present on the HR receiver chain: one with bandwidth 35 kHz on the shared MR-HR section and a passive LC filter of bandwidth 6 kHz, seen by the HR channel only. The combined response of both these filters has been analyzed and removed from the data. The second systematic effect is due to the frequency-dependent response of the coupling between the cavity and the first cryogenic amplifier. This effect is removed using the equivalent circuit model described later. The combined passband filter response was determined by taking data with a white noise source at the rf input of the receiver chain. A total of 872 time traces were recorded over a two day period. In order to achieve a reasonably smooth calibration curve, 512 bins in the frequency spectrum for each time trace were averaged giving 9.77 Hz resolution. The combined average of all data is shown in Fig. \[fig-filtercal\]. This measured response was removed from all data used in the HR search, as follows. The raw power spectra have frequency 0–10 kHz, where the center frequency of 5 kHz has been mixed down from the cavity frequency. Each raw power spectrum is cropped to the region 2–8 kHz to remove the frequencies not within the LC filter bandwidth. Each remaining frequency bin is then weighted by a factor equal to the receiver chain response at the given frequency divided by the maximum receiver chain response. Interpolation for frequency points not specifically included in the calibration curve is performed by assuming that each point on the calibration curve was representative of 512 bins centered on that frequency, so all power corresponding to frequencies within that range is normalized by the same factor. As the calibration curve varies slowly with frequency within the window to which each spectra is cropped, this is an adequate treatment of the normalization. ![HR filter response calibration data (512 bin average). The power has been normalized to the maximum power output.[]{data-label="fig-filtercal"}](Figure7small.eps) In the MR channel, the effect of the cavity-amplifier coupling is described using an equivalent-circuit model [@DawPhD]. This model has been adapted for use in the HR channel. The frequency dependent response of the cavity amplifier coupling is most evident in the 4096-bin search, thus this is the data used to apply the equivalent circuit model. A sample spectrum before correction is shown in Fig. \[fig-4096before\]. ![Sample 4096-bin spectrum before correction for the cavity-amplifier coupling. The line is the fit obtained using the equivalent circuit model.[]{data-label="fig-4096before"}](NewFigure8.eps){width="45.00000%"} In the equivalent-circuit model, each frequency is given by $\Delta$, the number of bins it is offset from the bin of the center frequency, measured in units of the 4096-bin resolution, i.e. $b_{4096}=78.1$ Hz. The equivalent-circuit model predicts that the power (in units of the rms noise) at the NRAO amplifier output (the point labelled “RF” in Fig. \[fig-receiver\]) in the 4096–bin search at the frequency offset $\Delta$ is $$P(\Delta)=\frac{a_{1}+8a_{3}\left(\frac{\Delta-a_{5}}{a_{2}}\right)^{2} +4a_{4}\left(\frac{\Delta-a_{5}}{a_{2}}\right)}{1+4\left(\frac{\Delta-a_{5}}{a_{2}}\right)^{2}} \mathrm{,} \label{eq:5fit}$$ where the parameters $a_{1}$ through $a_{5}$ are $$\begin{aligned} a_{1} & = & (b_{4096}/b)(T_{C} + T_{I} + T_{V})/T_{n}\; , \\ a_{2} & = & f_{0}/(b_{4096}\,Q)\; , \\ a_{3} & = & (b_{4096}/b)(T_{I} + T_{V} + (T_{I} - T_{V})\cos(2kL))/T_{n} \; , \\ a_{4} & = & (b_{4096}/b)((T_{I} - T_{V})\sin(2kL))/T_{n}\;\mathrm{and} \\ a_{5} & = & (f_{0} - f_{cen})/b_{4096} \; . \end{aligned}$$ In the above expressions, $T_{C}$ is the physical temperature of the microwave cavity, $T_{I}$ and $T_{V}$ are the current and voltage noise, respectively, contributed by the amplifier, $T_{n}$ is the noise temperature contributed from all components, $b$ is the frequency resolution of the HR channel, i.e. 0.019 Hz, $L$ is the electrical (cable) length from the cavity to the HFET amplifer, $f_{0}$ is the cavity resonant frequency, $f_{cen}$ is the center frequency of the spectrum and $k$ is the wavenumber corresponding to frequency $f_{cen}+b\Delta$. The factor $b_{4096}/b$ appears in the parameters $a_{1}$, $a_{3}$ and $a_{4}$ as it is an overall factor which results from normalizing the power to the single bin noise baseline. In practice, the parameters $a_{1}$ through $a_{5}$ are established by fitting. The line in Fig. \[fig-4096before\] shows the fit obtained using the equivalent circuit model. Large peaks in the data, e.g. an axion signal or environmental peak, are removed before fitting to prevent bias. The 4096-bin spectrum is used to perform the fit and then the original 1-bin spectrum is corrected to remove the systematic effect. The weighting factors are calculated using Eq. (\[eq:5fit\]) and the fitted parameters, $a_1$ through $a_5$, at the center of each bin of width $b_{4096}$. These factors are the ratio of the fit at a given point to the maximum value of the fit. Each 1-bin is multiplied by the factor calculated for the bin of width $b_{4096}$ within which it falls. The removal of the cavity-amplifier coupling and the passband filter response using the techniques described above has been demonstrated to result in flat HR spectra. Axion signal search procedure {#ss:sproc} ----------------------------- We now describe the search for an axion signal and summarize the analysis performed on each time trace. The width of an axion signal is determined by the signal frequency, axion velocity and flow velocity dispersion (Eq. (\[eq:sigbroad\])), the latter being the most uncertain variable. $n$-bin searches, where $n$ is the number of adjacent 1-bins added together ($n=$ 1, 2, 4, 8, 64, 512 and 4096), are conducted to allow for various velocity dispersions. For searches with $n>1$, there is an overlap between successive $n$-bins such that each $n$-bin overlaps with the last half of the previous and first half of the following $n$-bin. This scheme is illustrated for the 2, 4 and 8-bin searches in Fig. \[f:coadd\]. 1–bin search: --------------- -- 2–bin search: 4–bin search: 8–bin search: The search for an axion signal is performed by scanning each spectrum for peaks above a certain threshold. All such peaks are considered candidate axion signals. The thresholds are set at a level where there is only a small probability that a pure noise peak will occur and such that the number of frequencies considered as candidate axion peaks is manageable. The candidate thresholds used were 20, 25, 30, 40, 120, 650 and 4500 $\sigma$, in increasing order of $n$. All time traces are analyzed in the same manner. A fast Fourier transform is performed and an initial estimate of $\sigma$ is obtained by fitting the 1-bin noise distribution to Eq. (\[eqn-1bindist\]). Systematic effects are then removed, i.e.  the corrections described in Section \[ss:system\] for the filter passband response and cavity–amplifier coupling are performed. “Large” peaks not included in the equivalent circuit model fit for the cavity-amplifier response are defined to be those greater than 120% of the search threshold for each $n$–bin search. After the removal of systematic effects, the 1-bin noise distribution is again fitted to Eq. (\[eqn-1bindist\]) to obtain the true value of $\sigma$ and the search for peaks above the thresholds takes place. The axion mass is not known, requiring that a range of frequencies must be examined. Full HR coverage has been obtained for the region 478–525 MHz, corresponding to axion masses between 1.98 and 2.17 $\mu$eV. The selected frequency range is examined in three stages for axion peaks, as follows: *Stage 1:* Data for the entire selected frequency range is taken. The frequency step between successive spectra is approximately 1 kHz, i.e. the center frequency of each spectrum differs from the previous spectrum by 1 kHz. Frequencies at which candidate axion peaks occur are recorded for further examination during stage 2. *Stage 2:* Multiple time traces are taken at each candidate frequency from stage 1. The steady flow assumption described in Section \[ss:sigprop\] means that a peak will appear in spectra taken with center frequency equal to the candidate frequency from stage 1 if such a peak is an axion signal. The frequencies of persistent peaks, i.e. peaks that appear during both stage 1 and 2 are examined further in stage 3. *Stage 3:* Frequencies of persistent peaks undergo a three-part examination. The first step is to repeat stage 2, to ensure the peaks still persist. Secondly, the warm port attenuator is removed from the cavity and multiple time traces taken. If the peak is due to external radio signals entering the cavity (an environmental peak), the signal power will increase dramatically. If the signal originates in the cavity due to axion-photon conversion, the power developed in the cavity will remain the same as that for the normal configuration. The third step is to use an external antenna probe as a further confirmation that the signal is environmental. Some difficulties were encountered with the antenna probe, due to polarization of environmental signals. However, the second step is adequate to confirm that peaks are environmental. If a persistent peak is determined to not be environmental, a final test will confirm that it is an axion signal. The power in such a signal must grow proportionally with the square of the magnetic field ($B_0$ in Eq. (\[e:power\])) and disappear when the magnetic field is switched off. No axion peaks were found in the range 478–525 MHz using this approach. The exclusion limit calculated from this data is discussed in the following section. Results {#sec:results} ======= Over the frequency range 478–525 MHz, we derive an upper limit on the density of individual flows of axion dark matter as a function of the velocity dispersion of the flow. The corresponding axion mass range is 1.97–2.17 $\mu$eV. Each $n$-bin search places an upper limit on the density of a flow with maximum velocity dispersion, $\delta v_n$, as given by Eq. (\[eq:vdisp\]). Several factors reduce the power developed in an axion peak from that given in Eq. (\[e:power\]). The experiment is operated near critical coupling of the cavity to the preamplifier, so that half this power is observed when the cavity resonance frequency, $f_0$, is precisely tuned to the axion energy. If $f_0$ is not at the center of a $1$-bin, the power is spread into adjacent bins, as discussed below. When the axion energy is off-resonance, but still within the cavity bandwidth at a frequency $f$, the Lorentzian cavity response reduces the power developed by an additional factor of $$h(f)=\frac{1}{1+4Q^2\left(\frac{f}{f_{0}}-1\right)^2}\; .$$ To be conservative, we calculate the limits at points where successive spectra overlap, i.e. at the frequency offset from $f_0$ that minimizes $h(f)$. If a narrow axion peak falls at the center of a 1-bin, all power is deposited in that 1-bin. However, if such a peak does not fall at the center of a 1-bin, the power will be spread over several 1-bins. We now calculate the minimum power in a single $n$-bin caused by a randomly situated, infinitely narrow axion line. The data recorded is the voltage output from the cavity as a function of time. The voltage as a function of frequency is obtained by Fourier transformation and then squared to obtain a raw “power” spectrum. The actual power is obtained by comparison to the rms noise power. The data are sampled for a finite amount of time and thus, the Fourier transformation of the output, $\mathcal{F}(f)$, will be of the voltage multiplied by a windowing function, i.e. $$\mathcal{F}(f)=\int_{-\infty}^{\infty}v(t)w(t)\exp(i2\pi ft)dt \; , \label{eq:ft1}$$ where $v(t)$ is the measured output voltage and $w(t)$ is the windowing function for a sampling period $T$, $$w(t)=\left\{ \begin{array}{ll} 1 & \mathrm{if }-T/2\leq t \leq T/2 \; ,\\ 0 & \mathrm{otherwise} \; . \end{array} \right.$$ Eq. (\[eq:ft1\]) is equivalent to $$\mathcal{F}(f)=\int_{-\infty}^{\infty}V(k)W(f-k)dk \; , \label{eq:ft2}$$ where $V(f)$ and $W(f)$ are the Fourier transforms of the output voltage, $v(t)$, and the windowing function, $w(t)$, i.e. $\mathcal{F}(f)$ is the convolution of $V(f)$ and $W(f)$, given by $$W(f)=\frac{\sin(\pi f T)}{\pi f} \; . \label{eq:sinc}$$ Discretizing Eq. (\[eq:ft2\]) and inserting Eq. (\[eq:sinc\]), we have $$\mathcal{F}(f)=\sum_{m=0}^N V((m+\frac12)b) \frac{\sin(\pi(\frac{f}{b}-(m+\frac12)))} {\pi(\frac{f}{b}-(m+\frac12))} \; , \label{eq:ft3}$$ where $b$ is the frequency resolution of the HR channel, $2N$ points are taken in the original time trace, and the center frequency of the $j$th 1-bin is $(j+1/2)b$. Thus, for an axion signal of frequency $f$ falling in 1-bin $j$, a fraction of the power $$g(m)=\left(\frac{\sin(m\pi+\delta)}{m\pi+\delta}\right)^2\; , \label{eq:sinc2}$$ is lost to the $m$th 1-bin from 1-bin $j$, where $\delta = \pi(m+1/2 - f/b)$. If $\delta =0$, i.e the axion signal frequency is exactly equal to a 1-bin center frequency, all the power is deposited in a single 1-bin. However, if this is not the case, power is lost to other 1-bins. In setting limits, we assume that the power loss is maximal. The maximum power loss occurs when a signal in the 1-bin search falls exactly between the center frequency of two adjacent 1-bins. In this case, when $\delta=\pi/2$, Eq. (\[eq:sinc2\]) shows that 40.5% of the power will be deposited in each of two 1-bins. In $n$-bin seaches with $n\geq 2$, not as much power is lost to other $n$-bins, due to the overlap between successive $n$-bins. The minimum power deposited in an $n$-bin is 81% for $n=2$, 87% for $n=4$ and 93% for $n=8$. For $n=$ 64, 512 and 4096, the amount of power not deposited in a single $n$-bin is negligible. For the $n$-bin searches with $n=$ 64, 512 and 4096, a background noise subtraction was performed which will lead to exclusion limits at the 97.7% confidence level. These limits are derived using the power at which the sum of the signal power and background noise power have a 97.7% probability to exceed the candidate thresholds. We call this power the “effective” threshold for each search. The effective thresholds are obtained by integrating the noise probability distribution, Eq. (\[eq:nprob\]), numerically solving for the background noise power corresponding to the 97.7% confidence level for each $n$ and subtracting these values from the original candidate thresholds. For $n=$ 64, 512 and 4096, the effective thresholds are 71, 182 and 531 $\sigma$, respectively. For smaller values of $n$, background noise subtraction does not significantly improve the limits and the effective threshold was taken to be the candidate threshold. Table \[t:resultinfo\] summarizes this information and shows the frequency resolution of each search with the corresponding maximum flow velocity dispersion from Eq. (\[eq:vdisp\]) for $v=600$ km/s. [rrdr]{} $n$ & Effective & b\_n & $\delta v_n$\ &threshold ($\sigma$) & & (m/s)\ 1 & 20 & 0.019 & 6\ 2 & 25 & 0.038 & 10\ 4 & 30 & 0.076 & 20\ 8 & 40 & 0.15 & 50\ 64 & 71& 1.2 & 400\ 512 & 182& 9.8 & 3000\ 4096 & 531& 78 & 20000\ Our exclusion limits were calculated for an axion signal with power above the effective threshold reduced by the appropriate factors. These factors arise from the critical coupling, the Lorentzian cavity response and the maximum power loss due to the peak not falling in the center of an $n$-bin, as outlined above. Equations (\[e:power\]) and (\[eq:Pnoise\]) were used, for both KSVZ and DFSZ axion couplings. The cavity volume, $V$, is 189 L. Measured values of the quality factor, $Q$, the magnetic field, $B_0$, and the cavity temperature, $T_C$, are recorded in each data file. Numerically determined values of the form factor, $C$ are given in Table \[t:formfactor\]. The electronic noise temperature, $T_{el}$, was conservatively taken from the specifications of the NRAO amplifier, the dominant source of noise in the receiver chain, although our measurements indicate that $T_{el}$ is less than specified. These values are also given in Table \[t:formfactor\]. Linear interpolation between values at the frequencies specified was used to obtain values of $C$ and $T_{el}$ at all frequencies. [cdd]{} Frequency (MHz) & & T\_[el]{}\ 450&0.43&1.9\ 475&0.42&1.9\ 500&0.41&1.9\ 520&0.38&1.9\ 550&0.36&2.0\ The 2-bin search density exclusion limit obtained using these values is shown in Fig. \[f:limits2\]. For values of $n$ other than $n=2$, the exclusion limits differ by only constant factors. The constant factors are 1.60, 1.00, 1.12, 1.39, 2.53, 5.90 and 17.2 for $n=$ 1, 2, 4, 8, 64, 512 and 4096, respectively. Discussion {#sec:discuss} ========== We have obtained exclusion limits on the density in local flows of cold axions over a wide range of velocity dispersions. The most stringent limit, shown in Fig. \[f:limits2\], is from the 2-bin search. For a flow velocity of 600 km/s relative to the detector, the 2-bin search corresponds to a maximum flow velocity dispersion of 10 m/s. The 1-bin search limit is less general, in that the corresponding flow velocity dispersion is half that of the 2-bin limit. It is also less stringent; much more power may be lost due to a signal occurring away from the center of a bin than in the $n=2$ case. For $n>2$, the limits are more general, but the larger power threshold of the searches make them less stringent. The largest flow predicted by the caustic ring model has density $1.7\times10^{-24}$ g/cm$^3$ ($0.95$ GeV/cm$^3$), velocity of approximately $300$ km/s relative to the detector, and velocity dispersion less than 53 m/s [@MW]. Using Eq. (\[eq:sigbroad\]) with Table \[t:resultinfo\] and the information displayed in Fig. \[f:limits2\] multiplied by the appropriate factors of 1.12 to obtain the 4-bin limit, it can be seen that the 4-bin search, corresponding to maximum velocity 50 m/s for $v=300$ km/s, would detect this flow if it consisted of KSVZ axions. For DFSZ axions, this flow would be detected for approximately half the search range. Figure \[f:limits2\] demonstrates that the high resolution analysis improves the detection capabilities of ADMX when a significant fraction of the local dark matter density is due to flows from the incomplete thermalization of matter that has only recently fallen onto the halo. The addition of this channel to ADMX provides an improvement of a factor of 3 over our previous medium resolution analysis.\ Acknowledgements {#acknowledgements .unnumbered} ================ This research is supported in part by the U.S. Department of Energy under Contract W-7405-ENG-48 at Lawrence Livermore National Laboratory, under grant DE-FG02-97ER41029 at the University of Florida, and by an IBM Einstein Endowed Fellowship at the Institute for Advanced Study. [longenoughnow]{} C. L. Bennett [*et al.*]{}, Astrophys. J. Suppl.  [**148**]{}, 1 (2003). R. D. Peccei and H. R. Quinn, Phys. Rev. D [**16**]{}, 1791 (1977); R. D. Peccei and H. R. Quinn, Phys. Rev. Lett.  [**38**]{}, 1440 (1977); S. Weinberg, Phys. Rev. Lett.  [**40**]{}, 223 (1978); F. Wilczek, Phys. Rev. Lett.  [**40**]{}, 279 (1978). L. F. Abbott and P. Sikivie, Phys. Lett. B [**120**]{}, 133 (1983). J. Preskill, M. B. Wise and F. Wilczek, Phys. Lett. B [**120**]{}, 127 (1983). M. Dine and W. Fischler, Phys. Lett. B [**120**]{}, 137 (1983). S. Asztalos et al., Phys. Rev. [**D64**]{} 092003 (2001). P. Sikivie, Phys. Rev. Lett. [**51**]{}, 1415 (1983) \[Erratum-ibid. [**52**]{}, 695 (1984)\]. J. F. Navarro, C. S. Frenk and S. D. M. White, Astrophys. J.  [**462**]{}, 563 (1996); B. Moore, F. Governato, T. Quinn, J. Stadel and G. Lake, Astrophys. J.  [**499**]{}, L5 (1998). P. Sikivie, Phys. Lett. [**B432**]{} 139 (1998). P. Sikivie, Phys. Rev. D [**60**]{}, 063501 (1999). D. Lynden-Bell, Mon. Not. Roy. Astron. Soc.  [**136**]{}, 101 (1967). S. J. Asztalos [*et al.*]{}, Phys. Rev. D [**69**]{}, 011101(R) (2004). K. Freese, P. Gondolo, H. J. Newberg and M. Lewis, Phys. Rev. Lett.  [**92**]{}, 111301 (2004); K. Freese, P. Gondolo and H. J. Newberg, Phys. Rev. D [**71**]{}, 043516 (2005). P. Sikivie and J. Ipser, Phys. Lett. [**B291**]{} 288 (1992); A. Natarajan and P. Sikivie, Phys. Rev. D [**72**]{}, 083513 (2005). J.E. Kim, Phys. Rep. [**150**]{} 1 (1987); H.-Y. Cheng, Phys. Rep. [**158**]{} 1 (1988); M.S. Turner, Phys. Rep. [**197**]{} 67 (1990); G.G. Raffelt, Phys. Rep. [**198**]{} 1 (1990). P. Sikivie, Phys. Lett. B [**567**]{}, 1 (2003). E. I. Gates, G. Gyuk and M. S. Turner, Astrophys. J.  [**449**]{}, L123 (1995). L. Duffy [*et al.*]{}, Phys. Rev. Lett. [**95**]{}, 091304 (2005). J. E. Kim, Phys. Rev. Lett. [**43**]{}, 103 (1979); M. A. Shifman, A. I. Vainshtein and V. I. Zakharov, Nucl. Phys. B [**166**]{}, 493 (1980). M. Dine, W. Fischler and M. Srednicki, Phys. Lett. B [**104**]{}, 199 (1981). A. R. Zhitnitsky, Sov. J. Nucl. Phys.  [**31**]{}, 260 (1980) \[Yad. Fiz.  [**31**]{}, 497 (1980)\]. S. De Panfilis [*et al.*]{}, Phys. Rev. Lett.  [**59**]{}, 839 (1987); C. Hagmann, P. Sikivie, N. S. Sullivan and D. B. Tanner, Phys. Rev. D [**42**]{}, R1297 (1990); C. Hagmann [*et al.*]{}, Phys. Rev. Lett.  [**80**]{}, 2043 (1998); H. Peng [*et al.*]{}, Nucl. Instrum. Meth. A [**444**]{}, 569 (2000); R. Bradley [*et al.*]{}, Rev. Mod. Phys.  [**75**]{}, 777 (2003). M. S. Turner, Phys. Rev. D [**42**]{}, 3572 (1990); F. S. Ling, P. Sikivie and S. Wick, Phys. Rev. D [**70**]{}, 123503 (2004). E. J. Daw, PhD thesis, Massachusetts Institute of Technology, 1998.

--- abstract: 'We present Deep Shape-from-Template (DeepSfT), a novel Deep Neural Network (DNN) method for solving real-time automatic registration and 3D reconstruction of a deformable object viewed in a single monocular image. DeepSfT advances the state-of-the-art in various aspects. Compared to existing DNN SfT methods, it is the first fully convolutional real-time approach that handles an arbitrary object geometry, topology and surface representation. It also does not require ground truth registration with real data and scales well to very complex object models with large numbers of elements. Compared to previous non-DNN SfT methods, it does not involve numerical optimization at run-time, and is a dense, wide-baseline solution that does not demand, and does not suffer from, feature-based matching. It is able to process a single image with significant deformation and viewpoint changes, and handles well the core challenges of occlusions, weak texture and blur. DeepSfT is based on residual encoder-decoder structures and refining blocks. It is trained end-to-end with a novel combination of supervised learning from simulated renderings of the object model and semi-supervised automatic fine-tuning using real data captured with a standard RGB-D camera. The cameras used for fine-tuning and run-time can be different, making DeepSfT practical for real-world use. We show that DeepSfT significantly outperforms state-of-the-art wide-baseline approaches for non-trivial templates, with quantitative and qualitative evaluation.' author: - | David Fuentes-Jimenez, David Casillas-Perez, Daniel Pizarro\ Universidad de Alcalá\ [{d.fuentes,david.casillas}@edu.uah.es,daniel.pizarro@uah.com]{} - | Toby Collins\ Institut de Recherche contre les Cancers de l’Appareil Digestif\ [toby.collins@gmail.com]{} - | Adrien Bartoli\ Université Clermont-Auvergne\ [adrien.bartoli@gmail.com]{} bibliography: - 'egbib.bib' title: | Deep Shape-from-Template: Wide-Baseline, Dense and Fast\ Registration and Deformable Reconstruction from a Single Image --- Introduction ============ The joint task of registration and 3D reconstruction of deformable objects from RGB videos and images is a major objective in computer vision, with numerous potential applications, for instance in augmented reality. In comparison with other mature 3D reconstruction problems, such as Structure-from-Motion (SfM) where rigidity is imposed on the scene [@Hartley2003], deformable registration and 3D reconstruction present significant unsolved problems. Two main scenarios exist in this task: Non-Rigid SfM (NRSfM) [@Bregler2000; @Torresani2008; @Dai2012; @Chhatkuli2017a] and Shape-from-Template (SfT) [@Salzmann2008; @Bartoli2015; @Ngo2016; @Chhatkuli2017]. NRSfM reconstructs the 3D shape of a deformable object from multiple RGB images. In contrast, SfT reconstructs the 3D shape from a single RGB image using an object *template*. The template includes knowledge about the object’s appearance, shape and permissible deformations. These are typically represented by a texture-map, a 3D mesh and a simple mechanical model. SfT is suitable for many applications where the template is known or can be acquired, using for instance SfM or any available 3D scanning solution. SfT solves two fundamental and intimately related problems: *i)* template-image [*registration*]{}, which associates pixels in the image to their corresponding locations in the template, and *ii)* [*shape inference*]{}, which recovers the observed 3D shape or equivalently the template’s 3D deformation. The majority of SfT methods focus on solving shape inference assuming that registration is independently obtained with existing feature-based or dense methods [@Pizarro2012; @Gay-Bellile2010; @Collins2014]. In all other cases, both problems are solved simultaneously using tracking with iterative optimization [@Ngo2016; @Collins2016; @Agudo2016]. To date there exists no non-DNN wide-baseline SfT method capable of solving both problems densely and in real-time. DNN SfT methods have been very recently proposed [@Pumarola2018; @hdm_net], following the success of the DNN methodology in related problems such as 3D human pose estimation [@Martinez2017; @Gueler2018], depth [@Eigen2015; @Garg2016; @Liu2016] and surface normal reconstruction with rigid objects [@Bansal2016; @Wang2015]. The general idea is to learn the function that maps an input image to the template’s 3D deformation parameters from training data. This has the potential to jointly solve registration and shape inference and eliminates the need for iterative optimization at run-time. These two recent methods are promising but bear important limitations. First, they are limited to flat templates described by regular meshes with very small vertex counts. Second, they require ground-truth registration for training, which is practically impossible to obtain for real data. We propose DeepSfT, the first DNN SfT method based on a fully-convolutional network without the above limitations. DeepSfT has the following desirable characteristics. *1)* It is dense and provides registration and 3D reconstruction at the pixel level. *2)* It does not require temporal continuity and handles large deformations and pose changes between the template and the object. *3)* It runs in real-time using conventional GPU hardware. *4)* It is applicable for templates with arbitrary geometry, topology and surface representation, including meshes, implicit and explicit functions such as NURBS. *5)* It is highly robust and handles well the major challenges of SfT, including self and external occlusions, illumination changes and blur. *6)* Training involves a novel combination of supervised learning with synthetic data and semi-supervised learning with RGB-D real data. Crucially, we do not require ground-truth registration for the real training data but only RGB-D. Compared to previous approaches, this makes it feasible to acquire the real training data *automatically*, and therefore feasible to deploy it in real settings. *7)* The network complexity, training cost and running cost are independent of the template representation, for instance of the mesh vertex count. It therefore scales very well to highly complex templates with detailed geometry that were, until now, not solvable in real-time. There exists no previous method in the literature with the above characteristics. Our method thus pushes SfT significantly forward. We present quantitative and qualitative experimental results showing that our method concretely outperforms in accuracy, robustness and computation time. Previous Work ============= We first review the non-DNN SfT methods, forming the vast majority of existing work. We start with the shape inference methods and then the integrated methods combining shape inference and registration. We finally review the recent DNN SfT methods. #### Shape inference methods. The shape inference methods assume that the registration between the template and the image is given, which is a fundamental limiting factor of applicability. We classify them according to the deformation model. The most popular deformation model is isometry, which attempts to approximately preserve the geodesic distance, and has been shown to be widely applicable. [*Isometric*]{} methods follow three main strategies: *i)* Using a convex relaxation of isometry called inextensibility [@Salzmann2009; @Salzmann2008; @Perriollat2011; @Brunet2014], *ii)* using local differential geometry [@Bartoli2015; @Chhatkuli2017] and *iii)* minimizing a global non-convex cost [@Brunet2014; @Oezguer2017]. Methods in *iii)* are the most accurate but also the most expensive. They require an initial solution found using a method from *i)* or *ii)*. There also exist [*non-isometric*]{} methods, with the angle preserving conformal model [@Bartoli2015] or simple mechanical models with linear [@Malti2013; @Malti2015] and non-linear elasticity [@Haouchine2017; @Haouchine2014; @Agudo2015]. These models all require boundary conditions in the form of known 3D points, which is another fundamental limiting factor of applicability. Their well-posedness remain open research questions. #### Integrated methods. The integrated methods compute both registration and shape inference. We classify them according to their ability to handle wide-baseline cases. [*Short-baseline*]{} methods are restricted to video data and may work in real time [@Ngo2016; @Collins2016; @Liu-Yin2017]. They are based on the iterative minimization of a non-convex cost and use keypoint correspondences [@Ngo2016] or optic flow [@Collins2016; @Liu-Yin2017]. The latter supports dense solutions and resolve complex, high-frequency deformations. Their main limitations are two-fold. First, they break down when there is fast deformation or camera motion. Second, at run-time, they must solve an optimization process that is highly computationally demanding, requiring careful hand-crafted design and balancing of data and deformation constraints. In contrast, [*wide-baseline*]{} SfT methods can deal with individual images showing the object with strong deformation without priors on the camera viewpoint [@Ngo2016; @collins14b]. These methods solve registration sparsely using keypoints such as SIFT [@sift] with filtering to reduce the mismatches [@Pizarro2012; @pilet08a]. The main limitations of these methods are two-fold. First, they are fundamentally limited by the feature-based registration, which fails due to a weak or repetitive texture, low image resolution, blur or viewpoint distortion. second, they require to solve a highly demanding optimization problem at run-time. Because of these limitations, the existing wide-baseline methods have only been shown to work for simple objects with simple deformations, such as bending sheets of paper. #### DNN SfT methods. Two DNN SfT methods [@Pumarola2018; @hdm_net] have been recently proposed. They address isometric SfT by learning the mapping between the input image and the 3D vertex coordinates of a regular mesh. Both methods use regression with a fully-convolutional encoder. They require the template to be flat and to contain a smaller number of regular elements. In [@Pumarola2018] belief maps are obtained for the 2D position of the vertices which are then combined with depth estimation and reprojection constraints to recover their 3D positions. This considerably limits the size of the mesh, as shown by the reported examples with fewer than $10\times10$ vertex counts. Both methods were trained and tested with synthetically generated images. Only [@Pumarola2018] provides results on a real video of a bending paper sheet, but required ground-truth registration and shape to fine-tune the network on part of the video. These two methods thus form a preliminary step toward applying DNN to SfT, but are strongly limited by the low template complexity and requirement for ground truth registration. Indeed, even if depth may be relatively easy to obtain for training, ground truth registration is extremely difficult to measure for real data. Problem Formulation {#sec:problemForm} =================== Figure \[fig:scheme\] shows the geometrical model of SfT. ![Differential Geometrical Model[]{data-label="fig:scheme"}](diagrama_volumetrico_2.jpg){width="\linewidth"} The *template* is known and represented by a 3D surface $\mathcal{T}\subset\mathbb{R}^3$ jointly with its appearance, described as a *texture map* $\mathcal{A}$. The texture map is standard and represented as a collection of flattened *texture charts* $\mathcal{U}_i$ whose union cover the appearance of the whole *template*, as seen in Figure \[fig:scheme\]. In our approach *templates* are not restricted to a specific topology, modelling both *thin-shell* and *volumetric* objects. They are also not restricted by a specific representation. In our experimental section we use mesh representations because of their generality, but this is not a requirement of the method. The bijective map between $\mathcal{A}$ and $\mathcal{T}$ is known and denoted by $\Delta:\mathcal{A} \longmapsto \mathcal{T}$. We assume that the *template* surface $\mathcal{T}$ is deformed with an unknown *quasi-isometric* map $\Psi:\mathcal{T} \longmapsto \mathcal{S}$, where $\mathcal{S}\subset\mathbb{R}^3$ denotes the unknown deformed surface. *Quasi-isometric* maps permit localized extension/compression, common with real world deforming objects. The input image is modeled as the colour intensity function $I:\mathbb{R}^2 \longmapsto (r,g,b)$, which is discretized into a regular grid of pixels in the retinal plane. The visible part $\mathcal{S}_v\subset\mathcal{S}$ of the surface is projected on an unknown subset of the image plane $\mathcal{I}\subset\mathbb{R}^2$. We assume the perspective camera for projection: $$\begin{aligned} \label{eq:cameraPerspectiva} \Pi_{p}:\qquad\mathbb{R}^{3} &\longmapsto \mathbb{R}^2\nonumber\\ \left(x,y,z\right) &\longmapsto \left(\frac{x}{z},\frac{y}{z}\right),\end{aligned}$$ where $\mathcal{S}_v$ is represented with a perspective embedding $X_v:\mathcal{I} \longmapsto \mathcal{S}_v$ with $X_{v}(u,v) = \rho(u,v) \left(u,v,1\right)$. We assume $X_v$ is known and any lens distortion is either negligible or has been corrected. The depth function $\rho:\mathcal{I} \longmapsto \mathcal{S}_v$ represents the depth coordinate of $\mathcal{S}_v$ from the camera’s coordinate system. In the absence of self-occlusions, $\mathcal{S}_v=\mathcal{S}$. Volumetric *templates* always induce self-occlusions in the image. The unknown registration function, or *warp*, $\eta:\mathcal{I} \longmapsto \mathcal{A}$ is an injective map that relates each point of $\mathcal{I}$ to its corresponding point in $\mathcal{A}$. Network architecture {#sec:arch} ==================== We propose a DNN, hereinafter *DeepSfT*, that estimates $\rho(u,v)$ and $\eta(u,v)$ directly from the input image $I(u,v)$: $$(\hat{\rho},\hat{\eta}) = \mathcal{D}(I,\theta_\mathcal{T}), \label{eq:dnnFunc}$$ where $\hat{\rho}$ and $\hat{\eta}$ are normalized ($\hat{\rho}\in [-1,1]$, $\hat{\eta}\in [-1,1]\times[-1,1]$) and discretized versions of $\rho$ and $\eta$. Our method also recovers $\mathcal{I}$ as both $\rho$ and $\eta$ are equal to $-1$ outside the domain $\mathcal{I}$ of the image. In this sense DeepSfT performs object segmentation at a pixel level. $\theta_\mathcal{T}$ are the network weights, that depend on the *template* $\mathcal{T}$, and are learned with training (see Section \[sec:training\]). *DeepSfT* has been trained to recognize a specific *template* so a large amount of deformations are required as described in Section \[sec:TrainDB\]. Figure \[fig:arquitectura\] shows the proposed network architecture. The complete architecture receives an RGB input image $I_i$ with a resolution of $h_i\times w_i= 270\times480$ pixels and returns the estimated depth map $\hat{\rho}_i$ and the registration maps $\hat{\eta}_i$. Both $\hat{\eta}_i$ and $\hat{\rho}_i$ have the same size as the input image. {width="\linewidth"} *DeepSfT* is divided into two main blocks: the *main block* is modelled on an *encoder-decoder* architecture, very similar to those used in semantic segmentation [@segnet]. This gives a first depth map estimation and the proposed registration function $\hat{\eta_i}$. The second one is a *domain adaptation block* that uses the RGB input image $I_i$ together with the output of the previous block to refine the depth map estimation $\hat{\rho_i}$. This cascade topology where the input image is feed into refinement blocks has proven to improve the results obtained using single stages in methods for 3D depth estimation [@cascade_3dreconstruction]. This block is also crucial to adapt the network to real data as described in Section \[sec:training\]. Both the *main* and *domain adaptation blocks* use identity, convolutional and deconvolutional residual feed-forwarding structures based on the ResNet50 [@resnet] (see Figure \[fig:blocks\]). ![Identity, convolutional and deconvolutional residual blocks.[]{data-label="fig:blocks"}](diagramaBloques.png){width="0.65\linewidth"} Each block is composed of two unbalanced parallel branches with convolutional layers that propagate the information forward into deeper layers, preserving the high frequencies of the data. Table \[tab:capas\_main\_block\] shows the layered decomposition of the *main block*. It first receives the RGB input image and performs a first reduction of the input size. Then, a sequence of three *convolutional and identity blocks* encode texture and the depth information as deep features. [max width=]{} Layer num Type Output size Kernels/Activation ----------- ------------------------------ ----------------------------- -------------------- 1 Input (270,480,3) – 2 Convolution 2D (135,240,64) (7,7) 3 Batch Normalization (135,240,64) – 4 Activation (135,240,64) Relu 5 MaxPooling 2D (45,80,64) (3,3) 6 Encoding Convolutional Block (45,80,\[64, 64, 256\]) (3,3) 7-8 Encoding identity Block x 2 (45,80,\[64, 64, 256\]) (3,3) 9 Encoding Convolutional Block (23,40,\[128, 128, 512\]) (3,3) 10-12 Encoding identity Block x 3 (23,40,\[128, 128, 512\]) (3,3) 13 Encoding Convolutional Block (12,20,\[256, 256, 1024\]) (3,3) 14-16 Encoding identity Block x 3 (12,20,\[256, 256, 1024\]) (3,3) 17-20 Encoding identity Block x 3 (12,20,\[1024, 1024, 256\]) (3,3) 21 Decoding Convolutional Block (24,40,\[512, 512, 128\]) (3,3) 22 Cropping 2D (23,39,128) (1,1) 23-25 Encoding identity Block x 3 (23,39,\[512, 512, 128\]) (3,3) 26 Decoding Convolutional Block (46,78,\[256, 256, 64\]) (3,3) 27 Zero Padding (46,80,64) (0,1) 28-29 Encoding identity Block x 2 (46,80,\[256, 256, 64\]) (3,3) 30 Upsampling (138,240,64) (3,3) 31 Cropping 2D (136,240,64) (2,0) 32 Convolution 2D (136,240,64) (7,7) 33 Batch Normalization (135,240,64) – 34 Activation (136,240,64) Relu 35 Upsampling (272,480,64) (3,3) 36 Cropping 2D (270,480,64) (2,0) 37 Convolution 2D (272,480,3) (3,3) 38 Activation (270,480,1) Linear : Main block architecture.[]{data-label="tab:capas_main_block"} Information is reduced to a compressed feature vector in a representation space of dimension $12\times20\times1024$. Information related with the deformable surface is coded in this vector per each RGB input image. Decoding is performed with *decoding blocks*. These require upsampling layers to increase the dimensions of the input tensors before passing through the convolution layers, as shown in Figure \[fig:blocks\].c. Finally, the last layers consist of CNNs and cropping layers that adapt the output of the decoding block to the size of the output maps ($270\times480\times3$). The first channel provides the depth estimate and the last two channels provide the registration warp. Table \[tab:capas\_refinement\_block\] shows the layered decomposition of the *domain adaptation block*. It is a reduced version of the *main block* where only the first two *encoding and decoding blocks* are included. The *domain adaptation block* take as input the concatenation of the input image and the output from the *main block* (6 channels) and it outputs a new refined depth map. [max width=]{} Layer num Type Output size Kernels/Activation ----------- ------------------------------ --------------------------- -------------------- 1 Input (270,480,3) – 2 Convolution 2D (135,240,64) (7,7) 3 Batch Normalization (135,240,64) – 4 Activation (135,240,64) Relu 5 MaxPooling 2D (45,80,64) (3,3) 6 Encoding Convolutional Block (45,80,\[64, 64, 256\]) (3,3) 7-8 Encoding identity Block x 2 (45,80,\[64, 64, 256\]) (3,3) 9 Encoding Convolutional Block (23,40,\[128, 128, 512\]) (3,3) 10-13 Encoding identity Block x 4 (23,40,\[128, 128, 512\]) (3,3) 14 Decoding Convolutional Block (46,80,\[512, 512, 128\]) (3,3) 15-16 Encoding identity Block x 2 (46,80,\[512, 512, 128\]) (3,3) 17 Upsampling (92,160,128) (2,2) 18 Cropping 2D (92,160,128) (2,0) 19 Convolution 2D (90,160,64) (3,3) 20 Batch Normalization (90,160,64) – 21 Activation (90,160,64) Relu 22 Upsampling (270, 480, 64) (3,3) 23 Convolution 2D (270, 480, 32) (3,3) 24 Activation (270, 480, 32) Relu 25 Convolution 2D (272,480,1) (3,3) 26 Activation (270,480,1) Linear : Domain adaptation block architecture. The block achieves to adapt the network to real data domain[]{data-label="tab:capas_refinement_block"} Training {#sec:TrainDB} ======== We create a quasi-photorealistic SfT synthetic database using simulation software. Synthetic data allows us to easily train our DNN end-to-end. We then follow by re-training the *domain adaption* block using a much smaller dataset collected using a standard RGB-D sensor. We recall that there are no public training datasets of this kind. Synthetic Data {#sec:synDB} -------------- This process involves randomized sampling from the object’s deformation space, generating the resulting deformation, and rendering from randomized viewpoints. We now describe the process for generating these training datasets for the templates used in the experiential section below (two thin-shell and two volumetric templates, see Table \[tb:synDeformations\]). *DB1* corresponds to a *DIN A4* piece of poorly-texture paper. *DB2* has the same shape as *DB1* but with a richer texture. *DB3* is a soft child’s toy and *DB4* is an adult sneaker. We emphasize that no previous work has been able to solve SfT for these last two objects in wide-baseline. The rest shape surfaces for *DB3* and *DB4* are obtained with triangulated meshes built using SfM (Agisoft Photoscan [@photoscan]). [max width=]{} [|m[3cm]{}|m[3cm]{}|m[3cm]{}|m[3cm]{}|]{} DB1&DB2&DB3&DB4\ {width="30mm"} & {width="30mm"} & {width="30mm"} & {width="30mm"}\ Mesh Faces=1521& Mesh Faces=1521&Mesh Faces=36256&Mesh Faces=5212\ \ {width="30mm"} & {width="30mm"} & {width="30mm"} & {width="30mm"}\ \ {width="30mm"} & {width="30mm"} & {width="30mm"} & {width="30mm"}\ {width="30mm"} & {width="30mm"} & {width="30mm"} & {width="30mm"}\ \ {width="30mm"} & {width="30mm"} & {width="30mm"} & {width="30mm"}\ {width="30mm"} & {width="30mm"} & {width="30mm"} & {width="30mm"}\ We use *Blender* [@Blender] to sample the deformation spaces and to create quasi-photorealistic renderings. It includes a physics simulation engine to simulate deformations with different degrees of stiffness using position based dynamics. For the paper templates we used Blender’s cloth simulator using a high stiffness term to model the stiffness of paper, with contour conditions and tensile and compressive forces in randomized 3D directions. This generates continuous deformation videos. For the other two templates we used rig-based deformation with hand-crafted rigs. This generates non-continuous deformation instances, using randomized joint angle configurations. For each deformation we generate random viewpoint variations with random rotations and translations of the camera, lighting variations using different specular light models and random backgrounds obtained with [@nyudepth]. In total, each dataset consists of $60.000$ RGB images, depth maps, and registrations (2-channel optical flow maps between the image and the template’s texturemap). All images have a resolution of $480\times270$ to fit the input/output of the network. We refer the reader to the supplementary material for a copy of these datasets with rigs and simulation parameters. Real Data {#sec:realDB} --------- We used *Microsoft Kinect v2* to record a total of $5.000$ RGB-D frames of the four objects while undergoing deformations induced by hand manipulation, and viewpoint changes (see Table \[tb:synDeformations\]). Image resolution was downsized to $480\times270$ to fit the input shape of the network. Training Procedure {#sec:training} ================== The training procedure is divided in two main steps: *1)* training with synthetic data followed by *2)* semi-supervised fine-tuning with real data. In step *1)* both *main* and *domain adaptation blocks* are trained end-to-end as a single block. We use ADAM [@adam] optimization with $10^{-3}$ learning rate and parameters $\beta_1=\beta_2 = 0.9$. We train for $40$ epochs with a batch size of $bs=7$. We initialize *DeepSfT* with uniform random weights [@xavier_glorot]. The loss function is defined as follows: $$\begin{split} & \mathcal{L}_s(I_i,\rho_i,\eta_i,\theta_\mathcal{T}) = \mathcal{L}_1(I_i,\eta_i,\theta_\mathcal{T}) + \mathcal{L}_2(I_i,\rho_i,\theta_\mathcal{T}) = \\ & %\frac{1}{N_1}\sum_{i=1}^{N_1} (\mathcal{D}_1(p_i,\theta_\mathcal{T}) - q_{i})^2 + \frac{1}{N_2}\sum_{i=1}^{N_2}(\mathcal{D}_2(p_i,\theta_\mathcal{T}) - \rho_{i})^2 \frac{1}{N_1}\sum_{i=1}^{N_1} \| \hat{\eta}_i - \eta_i \|^2 + \frac{1}{N_2}\sum_{i=1}^{N_2} \| \hat{\rho}_i - \rho_i \|^2 \end{split}$$ where $\hat{\rho}_i$ and $\hat{\eta}_i$ are the output depth map and warp estimates given by *DeepSfT* respectively. The terms $\rho_i$ and $\eta_i$ are the respective ground truths, and $N_1=2 h_i w_i bs$ and $N_2=h_i w_i bs$ are constants. The symbol $\|.\|$ denotes the Euclidean norm. Observe that $\hat{\rho}$ and $\hat{\eta}$ inherently depend on the network weights $\theta_\mathcal{T}$ and on the input image $I_i$, see Eq. . In step *2)* we train the *domain adaptation block* using real data while freezing the weights of the *main block*. This step is crucial to adapt the network to handle the ‘render gap’ and include the appearance characteristics of real data, such as the complex illumination, camera response and color balance. Also crucial is the fact that this can be done automatically, without the need for ground truth registration. We use stochastic gradient descend (SGD) with a small and fixed learning rate of $10^{-5}$. We train the network during $10$ epochs with a batch size of $bs=7$. Having both a low learning rate and a reduced number of epochs allows us to adapt our network to real data while avoiding overfitting. In this step a different loss function $\mathcal{L}_r$ is used, which only includes the depth information given by the depth sensor as the target of the *domain adaptation block*:$$\mathcal{L}_r(I_i,\rho_i,\theta_\mathcal{T}) = \mathcal{L}_2(I_i,\rho_i,\theta_\mathcal{T}) = \frac{1}{N_2}\sum_{i=1}^{N_2} \| \hat{\rho}_i - \rho_i \|^2,$$ where $N_2 = h_i w_i bs$. Experimental Results {#sec:Experiments} ==================== We evaluate *DeepSfT* in terms of 3D reconstruction and registration error with synthetic and real test data (described in §\[sec:realDB\]). Synthetic test data was generated using the same process as the synthetic training data, using new randomized configurations not present in the training data. Real test data was generated using the same process as the real training data, using new video sequences, consisting of new viewpoints and object manipulations not present in the training data. We also generated new test data using two new cameras, as described below in §\[sec:otherCams\]. We compare *DeepSfT* against a state-of-the-art isometric SfT method [@Chhatkuli2017] refereed as *CH17*. We provide this method with two types of registration: *CH17+GTR* uses the ground truth registration (indicting its best possible performance independent of the registration method) and *CH17+DOF* using the output of a state-of-the-art dense optical flow method [@optical_flow]. In the latter case we generate registration for image sequences using frame-to-frame tracking. We also compare these two variants using *a posteriori* deformation refinement using Levenberg–Marquardt, which is standard practice for improving the output of closed-form SfT methods. We refer to these improvements as *CH17R+GTR* and *CH17R+DOF*. We compare *DeepSfT* with two DNN-based methods: The first is a na[ï]{}ve application of the popular Resnet architecture [@resnet] to SfT, referred as  *R50F*. We performed this by removing the final two layers of Resnet and introducing one dense layer with 200 neurons and a final dense layer with a 3-channel output (for depth and warp maps) of the same size as the input image. We trained  *R50F* with exactly the same training data as *DeepSfT* and real-data fine tuning. Fine-tuning was implemented by optimizing the depth loss while forcing the the warp outputs to be unchanged, using the same optimizer and learning rate as we used for *DeepSfT*. The second DNN method is [@hdm_net], applicable only for DB1 and DB2. Because public code is not available, we carefully re-implemented it, requiring an adaption of the image input size and the mesh size, so that it matched the size of meshes for DB1 and DB2. We refer to this as *HDM-net*. We evaluate reconstruction error using the root mean square error (RMSE) between the 3D reconstruction and the ground truth in millimeters. We also use RMSE to evaluate the registration accuracy in pixels. The evaluation of registration accuracy is notoriously difficult for real data, because there is no way to reliably obtain ground truth. We propose to use as a proxy for the ground truth the output from a state-of-the-art dense trajectory optical flow method *DOF*. We only make this evaluation for video sequence data, for which DOF can reliably estimate optical flow over the sequence. [max width=0.8]{} --------------- --------- ----- ------ --------- ---------- ---------- ----------- ----------- --------- ------ --------- Sequence Type Samples DOF R50F DeepSfT CH17+GTR CH17+DOF CH17R+GTR CH17R+DOF HDM-net R50F DeepSfT - --------------- --------- ----- ------ --------- ---------- ---------- ----------- ----------- --------- ------ --------- Experiments with thin-shell objects and continuous test sequences ------------------------------------------------------------------ We show in Tables \[tb:exp\_sintetic\] and \[tb:examples\] the quantitative and qualitative results obtained with the thin-shell templates *DB1* and *DB2* with synthetic test datasets, denoted by *DB1S* and *DB2S*, and real test datasets, denoted by *DB1R* and *DB2R*. In terms of reconstruction error *DeepSfT* is considerably better than other methods, both in synthetic data, where the error remains below 2mm, and for real data, where the error is below 10mm. The Kinect V2 have an uncertainty of about 10mm at a distance of one meter, which partially explains the higher error for real data. The second and third best methods are *R50F* and *HDM-net*, also based on deep learning. However their results are far from those of *DeepSfT*. The method *CH17* obtains reasonable results when it is provided with ground truth registration (*CH17-GTR* and *CH17R-GTR*). However, the performance is considerably worse when real registration is provided using dense optical flow (CH17-DOF and CH17R-DOF). In terms of registration error, *DeepSfT* also has the best results both for synthetic test data, where ground-truth registration is available, and in real test data, where *DOF* is used as the proxy. In all cases *DeepSfT* has a mean registration error approximately 2 pixels. The performance of *R50F* is competitive with *DOF*, with registration errors approximately 5 pixels. We note that *DOF* exploits temporal coherence while *RF50* and *DeepSfT* do not and process each frame independently. Experiments with volumetric objects and non-continuous test images ------------------------------------------------------------------- The quantitative and qualitative results of the experiments for volumetric templates *DB3* and *DB4* are provided in Tables \[tb:exp\_sintetic2\] and \[tb:examples\] with both synthetic test data, denoted by *DB3S* and *DB4S*, and real test data, denoted by *DB4R* and *DB4R*. In this case we only provide registration error with synthetic data, because reliable registration using *DOF* is impossible with non-continuous test images. The method *CH17+GTR* and *CH17R+GTR* is tested only in the case of DB4S, because this is the only case that it can work (requiring a continuous texture map and a registration). [max width=]{} --------------- --------- ------ --------- ---------- ----------- ------ --------- Sequence Type Samples R50F DeepSFT CH17+GTR CH17R+GTR R50F DeepSfT - - --------------- --------- ------ --------- ---------- ----------- ------ --------- : Evaluation on synthetic and real databases *DB3S*, *DB4S*, *DB3R* and *DB4R*[]{data-label="tb:exp_sintetic2"} We observe a similar trend as with the thin-shell objects. *DeepSfT* is the best method both in terms of 3D reconstruction, with errors of the order of millimeters, and in registration with errors close to 2 pixels. The second best method is *R50F* although its results are significantly worse than those obtained by *DeepSfT*. The results of *CH17* and its variants are very poor. This may be due to the fact that *CH17* is not a method well adapted for volumetric objects with non-negligible deformation strain. [max width=]{} [|m[3cm]{}|m[3cm]{}|m[3cm]{}|m[3cm]{}|c|]{} Input Image & Depth Output& Warp-U Output& Warp-V Output & Depth Error (mm)\ {width="30mm"} & {width="30mm"} & {width="30mm"} & {width="30mm"}& 3.21\ {width="30mm"} & {width="30mm"} & {width="30mm"} & {width="30mm"}& 4.69\ {width="30mm"} & {width="30mm"} & {width="30mm"} & {width="30mm"}& 11.26\ {width="30mm"} & {width="30mm"} & {width="30mm"} & {width="30mm"}&8.96\ {width="30mm"} & {width="30mm"} & {width="30mm"} & {width="30mm"}&9.08\ {width="30mm"} & {width="30mm"} & {width="30mm"} & {width="30mm"}&7.49\ We show in Table \[tb:texture\_mapped\] qualitative reconstruction results obtained with DB1, DB3 and DB4 with real images. [max width=]{} [|m[3cm]{}|m[3cm]{}|m[3cm]{}|m[3cm]{}|m[3cm]{}|]{} Input Image & Output vs GTH& Textured Output& Textured Groundtruth & Error Map\ {width="30mm"} & {width="30mm"} & {width="30mm"} &{width="30mm"}& {width="30mm"}\ {width="30mm"} & {width="30mm"} & {width="30mm"} &{width="30mm"}& {width="30mm"}\ {width="30mm"} & {width="30mm"} & {width="30mm"} & {width="30mm"}& {width="30mm"}\ We observe that shapes recovered with *DeepSfT* are similar to ground-truth obtained with the RGB-D camera. We can observe that the error is larger near self-occlusion boundaries. Errors for DB1 are qualitatively smaller than for volumetric objects, which is consistent with Tables \[tb:exp\_sintetic\] and \[tb:exp\_sintetic2\]. Experiments with other cameras {#sec:otherCams} ------------------------------ We now present experiments showing the ability of *DeepSfT* to be used with a different camera at run-time, without any fine tuning with the new camera. The different cameras are an Intel Realsense D435[@realsense] (an RGB-D camera that we use for quantitative evaluation) and a Gopro Hero V3[@gopro] (an RGB camera for qualitative evaluation). Table \[tb:qualitative\_cameras\] shows their respective camera intrinsics. [max width=]{} Cameras Image Resolution $f_u$ $f_v$ $c_u$ $c_v$ ---------------------- ------------------ --------- --------- --------- --------- Kinect V2 1920x1080 1057.8 1064 947.64 530.38 Intel Realsense D435 1270x720 915.457 915.457 645.511 366.344 Gopro Hero V3 1920x1080 1686.8 1694.2 952.8 563.5 : Camera description table[]{data-label="tb:qualitative_cameras"} We have trained *DeepSfT* with a source RGB-D camera (Kinect V2), which has different intrinsics to the new cameras. We cannot immediately use images from the new camera because the network weights are specific to the intrinsics of the source camera. We propose to handle this by adapting the new camera’s effective intrinsics to match the source camera. Because the object’s depth within the training set varies (and so the perspective effects vary), we cam emulate training with the new camera’s intrinsics simply by an affine transform of the new camera image. This eliminates the need to retrain the network. We assume lens distortion is either negligible or has been corrected *a priori* using *e.g. OpenCV*. The affine transform is given by $A=diag(f_{u}/f'_u,f_{v}/f'_v)$ and displacement $t=(c_u - c'_u f_{u}/f'_u, c_v - c'_v f_{v}/f'_v )^\top$, where $f'_u,f'_v,c'_u,c'_v$ are the intrinsics of the new camera and $f_u,f_v,c_u,c_v$ are the intrinsics of source camera divided by $4$. The corrected image is then clipped about its optical centre and zero padded (if necessary), to obtain the resolution of $480\times270$ (the input image size of *DeepSfT*. Table \[tb:diferent\_cameras\] gives 3D reconstruction error for Intel Realsense D435[@realsense]. For the Gopro Hero V3[@gopro] we show qualitative result. [max width=]{} [|m[1.3cm]{}|m[2cm]{}|m[5cm]{}|c|]{} Camera & Converted Images & Results {width="38mm"} & Depth error\ Kinect V2 & {width="20mm"} & {width="50mm"} & 7.12\ Realsense D435 &{width="20mm"} &{width="50mm"} & 12.34\ Gopro Hero V3 & {width="20mm"} & {width="50mm"} & –\ Quantitatively the 3D reconstruction error of the original camera and the Intel Realsense D435[@realsense] are quite similar. This clearly demonstrates the ability of *DeepSfT* to generalize well to images taken with a different camera. *DeepSfT* is able to cope with images from other cameras even if the focals are quite different as it is the case with the GoPro camera. Light and Oclusion Resistance ----------------------------- We show that *DeepSfT* is resistant to light changes and significant occlusions. The first two rows of Table \[tb:resistance\] show representative examples of scenes with external and self occlusions for the thin-shell and volumetric objects. *DeepSfT* is able to cope with them, accurately detecting the occlusion boundaries. [max width=]{} [|m[1cm]{}|m[2cm]{}|m[2cm]{}|m[2cm]{}|m[2cm]{}|m[2cm]{}|]{} \ Image & {width="20mm"} & {width="20mm"} & {width="20mm"} & {width="20mm"} & {width="20mm"}\ Output & {width="20mm"} & {width="20mm"} & {width="20mm"} & {width="20mm"} & {width="20mm"}\ \ Input & {width="20mm"} & {width="20mm"} & {width="20mm"} & {width="20mm"} & {width="20mm"}\ Output & {width="20mm"} & {width="20mm"} & {width="20mm"} & {width="20mm"} & {width="20mm"}\ \ Input & {width="20mm"} & {width="20mm"} & {width="20mm"} & {width="20mm"} & {width="20mm"}\ Points with information & {width="20mm"} & {width="20mm"} & {width="20mm"} & {width="20mm"} & {width="20mm"}\ The third and fourth rows of Table \[tb:resistance\] show examples of scenes with light changes that produce significant changes in shading. *DeepSfT* shows resistance to those changes. Failure Modes ------------- There are some instances where *DeepSfT* fails, shown in the final two rows of Table \[tb:resistance\]. There are general failure modes of SfT (very strong occlusions and illumination changes), for which all methods will fail at some point. We also have failure modes specific to a learning-based approach (excessive deformations that are not represented in the training set). Timing Experiments ------------------ Table \[tb:speed\] shows the average frame rates of the compared methods, benchmarked on a conventional Linux desktop PC with a single NVIDIA GTX-1080 GPU. [max width=0.8]{} DeepSfT R50F CH17 CH17R DOF ----------- --------- ------ ------ ------- ------ Time(fps) 20.4 37 0.75 0.193 8.84 : Frame rates of the evaluated methods.[]{data-label="tb:speed"} The DNN-based methods are considerably faster than the other methods, with frame rates close to real time (*DeepSfT*). Solutions based on *CH17* are far from real-time. Conclusions =========== We have presented *DeepSfT*, the first dense, real-time solution for wide-baseline SfT with generic templates. This has been a long-standing computer vision problem for over a decade. *DeepSfT* will enable many real-world applications that require dense registration and 3D reconstruction of deformable objects, in particular augmented reality with deforming objects. We also expect it to be an important component for dense NRSfM in the wild. In the future we aim to improve results by incorporating temporal context information with recurrant neural networks, and to extend *DeepSfT* for unsupervised learning.

--- abstract: 'Extending the Kruppa’s prescription for the continuum level density, we have recently improved the BCS method with seniority-type pairing force in such a way that the effects of discretized unbound states are properly taken into account for finite depth single-particle potentials. In this paper, it is further shown, by employing the Woods-Saxon potential, that the calculation of spatial observables like nuclear radius converges as increasing the basis size in the harmonic oscillator expansion. Namely the disastrous problem of a “particle gas” surrounding nucleus in the BCS treatment can be circumvented. In spite of its simplicity, the new treatment gives similar results to those by more elaborate Hartree-Fock-Bogoliubov calculations; e.g., it even mimics the pairing anti-halo effect. The obtained results as well as the reason of convergence in the new method are investigated by a variant of the Thomas-Fermi approximation within the limited phase space which corresponds to the harmonic oscillator basis truncation.' author: - 'Toshiya Ono and Yoshifumi R. Shimizu' - Naoki Tajima - Satoshi Takahara title: | Method to circumvent the neutron gas problem\ in the BCS treatment for nuclei far from stability --- Introduction {#sec:intro} ============ The advent of new radioactive beam facilities makes it increasingly interesting to study unstable nuclei near the neutron or proton drip line. One of their specific features is weak binding of constituent nucleons, which leads to spatially extended nuclear profiles and a striking phenomenon of the neutron halo [@Tani96]. Among many issues expected in researches of nuclei far from the stability [@DobNaz98; @Dob99], the pairing correlation plays a special role because virtual scatterings into continuum states occur more easily. The basic quantity concerning the spatial distribution of nucleons is the nuclear radius, which is also a prerequisite for the analysis of reaction cross sections [@Tani96]. However, its theoretical evaluation in the presence of pairing correlation is not straightforward due to the fact that the continuum states, into which weakly bound nucleons virtually scatter, occupy infinite volume. In fact, the calculated radius is unreasonably large or divergent because of finite occupation probabilities of unbound states in the simple-minded BCS treatment; the so-called “neutron gas” problem, the solution of which requires the more sophisticated Hartree-Fock-Bogoliubov (HFB) theory [@DFT84; @DNW96; @DNW96a]. A similar problem exists for the Strutinsky shell correction calculation with finite depth potential [@BFN72], where the smoothed part of binding energy does not converge when increasing the size of the single-particle basis: The finite occupation probabilities of continuum states are required for extracting the smooth part, but the level density of unbound states is infinite, which leads to divergence of the Strutinsky smoothed energy. An efficient method to avoid this problem was introduced by Kruppa [@Kru98], and was used in the shell correction method [@VKN00]. The idea of the method is to calculate the so-called continuum level density as a difference between the level densities obtained by diagonalizing the full Hamiltonian including a finite depth potential and the free Hamiltonian. By employing the harmonic oscillator expansion, it was shown for the Woods-Saxon potential that the smoothed energy calculated by the Kruppa method well converges as increasing the size of basis [@VKN00]. Recently, three of the authors of this paper have proposed a new method of BCS calculation [@TST10], which is free from the divergence as increasing the model space, by utilizing the similar idea to Kruppa’s. In Ref. [@TST10] not only the treatment of pairing correlation but all the other procedures in the microscopic-macroscopic method for the calculation of nuclear binding energy are reexamined and improved. In this paper, we further show that the new method is capable of calculating the spatial observables like the nuclear radius without the problem of particle gas surrounding nucleus. Kruppa prescription for expectation values of one-body operators {#sec:Kruppa} ================================================================== Basic idea {#sec:basic} ---------- In the Kruppa’s prescription [@Kru98], the level density is replaced as $$g(\epsilon) \quad\Rightarrow\quad g{^{\rm K}}(\epsilon) =g(\epsilon)-g_0(\epsilon)\equiv \sum_{i=1}^{M} \delta \left( \epsilon - \epsilon_i \right) - \sum_{j=1}^{M} \delta ( \epsilon - {\epsilon^{0}_{j}} ), \label{eq:KruppaDensity}$$ where $\epsilon_i$ and ${\epsilon^{0}_{j}}$ are the eigenvalues of the full and the free Hamiltonians, respectively. In the following, we are mainly concerned with neutrons, but it should be reminded for protons that the Coulomb potential is included in the free Hamiltonian, i.e., “free” here means that the nuclear potential is left out. The eigenvalues in Eq. (\[eq:KruppaDensity\]) are calculated by diagonalizations with the harmonic oscillator basis, and $M$ is the total number of the basis states commonly used in the two diagonalizations. Both the full and free level densities, $g(\epsilon)$ and $g_0(\epsilon)$, are divergent as increasing the basis size ($M\rightarrow\infty$) for the positive single-particle energy, $\epsilon > 0$. It is shown for the Woods-Saxon potential [@Kru98] that their difference $g{^{\rm K}}(\epsilon)$ remains finite, and converges, in the spherical case, to the well-known expression in terms of the scattering phase shift $\delta_{lj}(\epsilon)$, $$\frac{1}{\pi}\sum_{lj} (2j+1)\frac{d\delta_{lj}(\epsilon)}{d\epsilon}. \label{eq:gphaseshift}$$ The great merit of Eq. (\[eq:KruppaDensity\]) is that it can be easily applied to the deformed cases (see Ref. [@Kru98] for the corresponding expression to Eq. (\[eq:gphaseshift\]) in terms of the scattering S-matrix in such general cases). Extending the idea of subtracting the free contribution in Eq. (\[eq:KruppaDensity\]), we propose to calculate the expectation value of an arbitrary one-body observable $O$ by a similar replacement as $$\langle O \rangle \quad\Rightarrow\quad \langle O \rangle{^{\rm K}}= \langle O \rangle - \langle O \rangle_0, \label{eq:OKruppa}$$ where the first (second) term is a summation of the expectation values with respect to the wave functions calculated by diagonalizing the full (free) Hamiltonian multiplied with the occupation numbers. Note that, in the independent particle approximation, e.g., the Hartree-Fock (HF) theory, the second term does not contribute for bound systems, i.e., if the Fermi energy is below the particle threshold. The free contribution manifests itself when the occupation probabilities of continuum states are non-zero due to the residual interactions, e.g., in the case of BCS treatment for pairing correlations. Although generalizations are possible, we consider in this work the following seniority-type separable pairing interaction, $$V{_{\rm pair}}=-G\,\hat{P}^\dagger \hat{P},\qquad \hat{P}^\dagger =\sum_{i>0} {f_{\rm c}}(\epsilon_i)\,c^\dagger_i c^\dagger_{\bar{i}}, \label{eq:SPairV}$$ where $\bar{i}$ represents the time-reversed state of $i$, forming a time reversal pair $(i\bar{i})$ (they are degenerate in energy, $\epsilon_i=\epsilon_{\bar{i}}$), and $\sum_{i>0}$ means the sum is taken over all the pairs. A cutoff of the model space is necessary for such a pairing interaction, and it is realized by introducing a smooth cutoff function [@BFH85], ${f_{\rm c}}(\epsilon)$ (we use a different form from that of Ref. [@BFH85]), defined by $${f_{\rm c}}(\epsilon) = \frac{1}{2} \left[1+ {\rm erf}\left(\frac{\epsilon-\tilde{\lambda}+\Lambda{_{\rm l}}}{d{_{\rm cut}}}\right) \right]^{1/2} \left[1+ {\rm erf}\left(\frac{-\epsilon+\tilde{\lambda}+\Lambda{_{\rm u}}}{d{_{\rm cut}}}\right) \right]^{1/2}, \label{eq:fe}$$ where ${\displaystyle {\rm erf}(x)\equiv\frac{2}{\sqrt{\pi}} \int_{0}^{x}e^{-t^2}dt }$ is the error function. We use the cutoff parameters of the pairing model space, $\Lambda{_{\rm u}}=\Lambda{_{\rm l}}=1.2 \, \hbar \omega$ and $d{_{\rm cut}}=0.2 \,\hbar \omega$, with $\hbar\omega=41/A^{1/3}$ MeV. The predefined parameter $\tilde{\lambda}$ in the cutoff function (\[eq:fe\]) is chosen to be the smoothed Fermi energy obtained by the Strutinsky smoothing procedure, see Ref. [@TST10] for details. The gap equation can be derived from the variational principle with the number constraint, $$\Delta=G\langle \hat{P}^\dagger \rangle,\qquad\mbox{and}\qquad N=\langle \hat{N} \rangle, \label{eq:BCSeqFromMean}$$ where $N$ is the neutron or proton number to be fixed. Applying the prescription (\[eq:OKruppa\]), the modified gap and number constraint equations are obtained: $$\begin{aligned} \frac{2}{G} &= & \sum_{i>0} \left[ \frac{{f_{\rm c}}(\epsilon_i)^2}{E(\epsilon_i)} -\frac{{f_{\rm c}}({\epsilon^{0}_{i}})^2}{E({\epsilon^{0}_{i}})} \right], \label{eq:KruppaBCSGapEq} \\ N &=& \sum_{i>0} \left[ 2v^2(\epsilon_i) - 2v^2({\epsilon^{0}_{i}}) \right]. \label{eq:KruppaBCSNumberEq}\end{aligned}$$ Here the BCS quasiparticle energy and occupation probability are given, as usual, by $$E(\epsilon)={\sqrt{(\epsilon - \lambda)^2+{f_{\rm c}}(\epsilon)^2 \Delta^2}}, \qquad v^2(\epsilon) = \frac{1}{2} \left( 1 -\frac{\epsilon-\lambda}{E(\epsilon)}\right). \label{eq:occBCS}$$ From these equations, the pairing gap and the chemical potential, ($\Delta,\lambda$), are determined for a given pairing strength $G$. The second terms in the square brackets in Eqs. (\[eq:KruppaBCSGapEq\]) and (\[eq:KruppaBCSNumberEq\]) are the extra (negative) contributions form the free spectra. We call this new way of the BCS treatment Kruppa-BCS method, and the resultant equation Kruppa-BCS equation. The consequences of this Kruppa-BCS equation have been investigated in detail in Ref. [@TST10]. As long as the chemical potential is well below the free particle threshold (minimum of $\{{\epsilon^{0}_{i}},i=1,2,...\}$), the Kruppa-BCS equation has a unique solution and can be solved in the same way as the ordinary BCS equation. It is shown that the calculated pairing gaps with this new method converge to reasonable values for large basis size in contrast to those with the ordinary BCS equation, which strongly depend on the size of basis and hardly applicable to the calculations with large basis size [@NWD94]. See Ref. [@TST10] for a comprehensive discussion. Root mean square radii and deformation parameters {#sec:rmsdef} ------------------------------------------------- ![ Neutron’s rms radii as functions of the maximum harmonic oscillator quantum number ${N_{\rm osc}^{\rm max}}$, which specifies the size of basis. Those calculated by the ordinary prescription, $\langle r^2 \rangle{_{\rm n}}$, and the Kruppa prescription, $\langle r^2 \rangle{^{\rm K}}{_{\rm n}}$ in Eq. (\[eq:rmsK\]), are included for a near drip-line nucleus $^{160}$Sn, the panel (a), and for a $\beta$-stable nucleus $^{120}$Sn, (b). Both nuclei are spherical. The neutron pairing gap is fixed to be either $\Delta{_{\rm n}}=0.5$, 1.0, or 1.5 MeV. []{data-label="fig:SnRMS"}](SnRMS.eps){width="75mm"} Once the pairing gap and the chemical potential, $(\Delta,\lambda)$, are obtained by the Kruppa-BCS equation (\[eq:KruppaBCSGapEq\]) and (\[eq:KruppaBCSNumberEq\]), the expectation value of one-body observables can be calculated within the BCS treatment; for example, the root mean square (rms) radius by $$\langle r^2 \rangle{^{\rm K}} = \langle r^2 \rangle- \langle r^2 \rangle_0 = \frac{1}{N} \sum_{i>0}\left[ \langle i |r^2| i \rangle\, 2v^2(\epsilon_i) - \langle i |r^2| i \rangle_0\, 2v^2({\epsilon^{0}_{i}}) \right], \label{eq:rmsK}$$ where the diagonal matrix elements $\langle i |r^2| i \rangle$ and $\langle i |r^2| i \rangle_0$ are with respect to the Woods-Saxon and free spectra, respectively. Figure \[fig:SnRMS\] shows the calculated neutron radii for a stable nucleus $^{120}$Sn and an unstable nucleus $^{160}$Sn in the lower and upper panels, respectively. They are depicted as functions of the maximum value of the harmonic oscillator quantum number, ${N_{\rm osc}^{\rm max}}$, specifying the model space. As for the Woods-Saxon potential, we use the parameter set recently developed by Ramon Wyss [@RWyss], which gives similar density distributions, both for neutrons and protons, to those obtained by Skyrme and Gogny HF calculations. In order to see the impact of the pairing correlation on the radius, the pairing gap in this calculation is kept constant to the values $\Delta{_{\rm n}}=0.5$, 1.0, and 1.5 MeV, and only the number constraint equation (\[eq:KruppaBCSNumberEq\]) is solved for determining the chemical potential $\lambda{_{\rm n}}$. The Kruppa mean value $\langle r^2 \rangle{^{\rm K}}{_{\rm n}}$ in Eq. (\[eq:rmsK\]) and the ordinary mean value $\langle r^2 \rangle{_{\rm n}}$ are compared. Note that for the calculation of the latter, the ordinary number equation is solved so that the chemical potentials in the calculations of two radii, $\langle r^2 \rangle{^{\rm K}}{_{\rm n}}$ and $\langle r^2 \rangle{_{\rm n}}$, are generally different. As is quite evident from Fig. \[fig:SnRMS\], the calculated radii by the ordinary method diverge as increasing the basis size ${N_{\rm osc}^{\rm max}}$. The divergence is more rapid for larger pairing gaps. There is no way to obtain meaningful results even for the stable nucleus $^{120}$Sn. This is nothing but the problem of the “neutron gas” surrounding the nucleus [@DFT84; @DNW96; @DNW96a]. In contrast, the radii calculated by the Kruppa method converge to definite values. Comparing with Fig. 15 in Ref. [@DFT84], our Kruppa results nicely corresponds to those of the HFB calculation, although the coordinate space is utilized in Ref. [@DFT84] and the box size instead of ${N_{\rm osc}^{\rm max}}$ is changed to demonstrate the convergence. It is sometimes recommended to be content with a small model space [@BFN72; @MNM95], like ${N_{\rm osc}^{\rm max}}\approx 12$, for finite depth potentials, because the plateau condition for the shell correction energy is not met with larger model spaces. Such a backward idea may work for stable nuclei, but it is clear from Fig. \[fig:SnRMS\] (a) that the space ${N_{\rm osc}^{\rm max}}\approx 12$ is not enough to calculate accurately the rms radius of $^{160}$Sn, and using a larger model space requires inevitably the Kruppa method. Another interesting feature of the Kruppa calculations seen in Fig. \[fig:SnRMS\] is that the radius of stable nucleus stretches for larger pairing gaps while that of the nucleus near drip-line shrinks, although the absolute amount of changes is very small. The stretching of the radius due to the pairing correlation is well-known: The pairing induces the couplings to particle states above the Fermi surface, whose rms radii are larger in average than those of hole states, and increased occupation probabilities of particles and decreased occupation probabilities of holes make the expectation value of radius larger (see Fig. \[fig:OBTFr\] (c) and discussions in Sec.\[sec:convergence\]). In spite of this effect, the radius of near drip-line nucleus $^{160}$Sn shrinks. We believe that this is related to the interesting “pairing anti-halo” effect [@BDP00] in the HFB theory, and will come back to this point later in Sec. \[sec:antihalo\]. ![ Neutron’s deformation parameter (\[eq:defbeta\]) as functions of the maximum harmonic oscillator quantum number ${N_{\rm osc}^{\rm max}}$. Those calculated by the ordinary and the Kruppa prescriptions are included for a near drip-line nucleus $^{208}$Dy, the panel (a), and a $\beta$-stable nucleus $^{160}$Dy, (b). Both nuclei are axially deformed. The pairing gap, $\Delta{_{\rm n}}=0.5$, 1.0, and 1.5 MeV, are employed. []{data-label="fig:DyBeta"}](DyBeta.eps){width="75mm"} In order to further illustrate the importance of the Kruppa prescription of Eq. (\[eq:OKruppa\]), presented in Fig. \[fig:DyBeta\] are examples of calculation for the quadrupole deformation parameter defined as a ratio of the quadrupole moment to the mean square radius, $$\beta = \frac{4\pi}{5}\, \frac{\langle r^2Y_{20} \rangle}{\langle r^2 \rangle},\qquad \beta{^{\rm K}} = \frac{4\pi}{5}\, \frac{\langle r^2Y_{20} \rangle{^{\rm K}}}{\langle r^2 \rangle{^{\rm K}}} \equiv \frac{4\pi}{5}\, \frac{\langle r^2Y_{20} \rangle - \langle r^2Y_{20} \rangle_0} {\langle r^2 \rangle - \langle r^2 \rangle_0}. \label{eq:defbeta}$$ In the upper and lower panels ((a) and (b)) the results for $^{208}$Dy and $^{160}$Dy are depicted, respectively, as functions of the basis size. Both nuclei are axially deformed, and the deformation parameters used in the Woods-Saxon potential are $(\beta_2,\beta_4)=(0.244,0.053)$ and $(0.269,0.035)$, respectively, which are obtained by the improved microscopic-macroscopic method of Ref. [@TST10]. As in Fig. \[fig:SnRMS\], the fixed pairing gaps of $\Delta{_{\rm n}}=0.5$, 1.0, and 1.5 MeV are used. Apparently, the quadrupole moment as well as the radius diverges as increasing the basis size due to the neutron gas surrounding the nucleus, if being calculated by the ordinary BCS treatment. One might expect that their ratio would converge to the correct value. It is not the case, however, even in stable nuclei, and the error is larger in neutron rich nuclei, amounting to about 10% for $\Delta{_{\rm n}}=1.5$ MeV and ${N_{\rm osc}^{\rm max}}=30$ in $^{208}$Dy. If the Kruppa prescription is used, the results with ${N_{\rm osc}^{\rm max}}\approx 12$ are already sufficiently accurate. Namely, compared with the radius or the quadrupole moment itself, the convergence of the deformation parameter as their ratio against the increase of the model space is much faster. Therefore it is better to use the Kruppa method even for calculations with ${N_{\rm osc}^{\rm max}}\approx 12$. Semiclassical consideration {#sec:semiclassical} --------------------------- In the case of the single-particle level density, the convergence of the Kruppa density (\[eq:KruppaDensity\]) to the exact density (\[eq:gphaseshift\]) in the limit of infinite model space, $M\rightarrow\infty$ (continuum limit), can be proved rigorously [@Kru98; @Shl92; @TO75]. Recently, we have shown the convergence more pictorially by employing a variant of the semiclassical (Thomas-Fermi) approximation. We call it the oscillator-basis Thomas-Fermi (OBTF) approximation [@TST10], which is suitable to treat the problem with the truncation in terms of the harmonic oscillator basis. In the following, we briefly sketch this OBTF approximation, and show that the Kruppa prescription (\[eq:OKruppa\]) for the expectation value of any spatial observables like the radius is convergent as increasing the basis size, ${N_{\rm osc}^{\rm max}}\rightarrow\infty$. The basic quantity in the Thomas-Fermi approximation is the phase space distribution function [@RS80]. Because the momentum distribution is isotropic in the nuclear ground state [@RS80], the distribution function depends only on the magnitude of the momentum variable, which one can transform to the energy ${\epsilon}$. Thus the distribution function for the single-particle Hamiltonian, [${\displaystyle H=\frac{{\mbox{\boldmath $p$}}^2}{2m}+V({\mbox{\boldmath $r$}}) }$]{}, is written, using the Heaviside step function $\theta(x)$, as $${f_{\!\!\phantom{.}_{\mbox{\tiny TF}}}}({\mbox{\boldmath $r$}},{\epsilon})= \frac{(2m)^{3/2}}{2\pi^2 \hbar^3} \left\vert {\epsilon}- V({\mbox{\boldmath $r$}}) \right\vert^{1/2} \theta\left( {\epsilon}- V({\mbox{\boldmath $r$}}) \right), \label{eq:lldensTF}$$ with which the Thomas-Fermi level density can be obtained as [${\displaystyle {g_{\!\!\phantom{.}_{\mbox{\tiny TF}}}}({\epsilon})= \int {f_{\!\!\phantom{.}_{\mbox{\tiny TF}}}}({\mbox{\boldmath $r$}},{\epsilon})d^3 r }$]{}. In the OBTF approximation, the distribution function is modified to incorporate the effect of the limited phase space corresponding to the truncated harmonic oscillator (HO), i.e., $${f_{\!\!\phantom{.}_{\mbox{\tiny OB}}}}({\mbox{\boldmath $r$}},{\epsilon})= {f_{\!\!\phantom{.}_{\mbox{\tiny TF}}}}({\mbox{\boldmath $r$}},{\epsilon})\, \theta\left({\epsilon_{\rm kin}^{\rm max}}(r)+ V({\mbox{\boldmath $r$}}) -{\epsilon}\right). \label{eq:lldensOBTF}$$ Here the quantity ${\epsilon_{\rm kin}^{\rm max}}(r)$ is the local maximum kinetic energy of the isotropic HO potential, [${\displaystyle V{_{\rm HO}}({\mbox{\boldmath $r$}})=\frac{1}{2}m\omega^2 r^2 }$]{}, $${\epsilon_{\rm kin}^{\rm max}}(r)= \frac{1}{2}m\omega^2\left( {R_{\rm max}}^2-r^2\right) \theta\left( {R_{\rm max}}-r\right), \label{eq:ekmax}$$ where the maximum radius ${R_{\rm max}}$ allowed in the HO potential is specified by the cutoff energy ${\epsilon_{\rm cut}}$, related to the maximum HO quantum number ${N_{\rm osc}^{\rm max}}$ of the basis truncation; $${R_{\rm max}}= \sqrt{\frac{2{\epsilon_{\rm cut}}}{m\omega^2}},\qquad {\epsilon_{\rm cut}}= \hbar \omega \left[ ({N_{\rm osc}^{\rm max}}+1)({N_{\rm osc}^{\rm max}}+2)({N_{\rm osc}^{\rm max}}+3) \right]^{1/3}. \label{eq:rmaxecut}$$ (Although the anisotropic HO basis is utilized in the actual calculation, it is enough to consider the isotropic HO potential for proving the convergence.) Thus, the infinite model space limit ${N_{\rm osc}^{\rm max}}\rightarrow\infty$ is realized by ${R_{\rm max}}\rightarrow\infty$ in $f_{\mbox{\tiny OB}}({\mbox{\boldmath $r$}},{\epsilon})$, and the OBTF level density [${\displaystyle {g_{\!\!\phantom{.}_{\mbox{\tiny OB}}}}({\epsilon})= \int {f_{\!\!\phantom{.}_{\mbox{\tiny OB}}}}({\mbox{\boldmath $r$}},{\epsilon})d^3 r }$]{} is finite and calculable as long as ${R_{\rm max}}$ (or ${N_{\rm osc}^{\rm max}}$) is kept finite. Note that this is not the case for the usual Thomas-Fermi quantities; ${g_{\!\!\phantom{.}_{\mbox{\tiny TF}}}}({\epsilon})$ and ${f_{\!\!\phantom{.}_{\mbox{\tiny TF}}}}({\mbox{\boldmath $r$}},{\epsilon})$ are both infinite for the energy ${\epsilon}$ above the particle threshold due to the infinite spatial volume of the phase space. In the same way, the OBTF distribution function for the free Hamiltonian, $f^0_{\mbox{\tiny OB}}({\mbox{\boldmath $r$}},{\epsilon})$, is defined by dropping the potential $V({\mbox{\boldmath $r$}})$ (or replacing it with the Coulomb potential for protons). The Kruppa level density in the OBTF approximation, $g{^{\rm K}}_{\mbox{\tiny OB}}({\epsilon}) =g_{\mbox{\tiny OB}}({\epsilon})- g^0_{\mbox{\tiny OB}}({\epsilon})$, is obtained by the spatial integral of $f{^{\rm K}}_{\mbox{\tiny OB}}({\mbox{\boldmath $r$}},{\epsilon}) =f_{\mbox{\tiny OB}}({\mbox{\boldmath $r$}},{\epsilon})-f^0_{\mbox{\tiny OB}}({\mbox{\boldmath $r$}},{\epsilon})$. Both level densities, ${g_{\!\!\phantom{.}_{\mbox{\tiny OB}}}}({\epsilon})$ and ${g_{\!\!\phantom{.}_{\mbox{\tiny OB}}}}^0({\epsilon})$, diverges as ${R_{\rm max}}\rightarrow\infty$, but the Kruppa density ${g_{\!\!\phantom{.}_{\mbox{\tiny OB}}}}{^{\rm K}}({\epsilon})$ remains finite for the potential $V({\mbox{\boldmath $r$}})$ that vanishes rapidly enough as $|{\mbox{\boldmath $r$}}|\rightarrow\infty$ like the Woods-Saxon potential. It has been shown [@TST10] that, for single-particle energies satisfying $0 < {\epsilon}< {\epsilon_{\rm kin}^{\rm max}}({\mbox{\boldmath $r$}})+V({\mbox{\boldmath $r$}})$ (assuming $V({\mbox{\boldmath $r$}})\le 0$) everywhere, $${g_{\!\!\phantom{.}_{\mbox{\tiny OB}}}}{^{\rm K}}({\epsilon}) \approx \frac{(2m)^{3/2}}{2\pi^2 \hbar^3} \int \left[\left({\epsilon}-V({\mbox{\boldmath $r$}})\right)^{1/2}-{\epsilon}^{1/2}\right]\,d^3 r, \label{eq:gobasympt}$$ asymptotically in the limit of ${R_{\rm max}}\rightarrow \infty$. It is straightforward to apply the OBTF approximation to the calculation of physical observables with pairing correlation (see, e.g., Ref. [@RS80] for the semiclassical treatment of the BCS theory). For a spatial one-body observable, $O({\mbox{\boldmath $r$}})$, its Kruppa OBTF expectation value can be expressed as $${\langle O \rangle}{^{\rm K}}_{\mbox{\tiny OB}}= \int_{-\infty}^\infty {g_{\!\!\phantom{.}_{\mbox{\tiny OB}}}}{^{\rm K}}({\langle O \rangle};{\epsilon})\, 2v^2({\epsilon})\,d{\epsilon}, \label{eq:OBTFKruppa}$$ where the BCS occupation probability $v^2({\epsilon})$ is given in Eq. (\[eq:occBCS\]), and the distribution of ${\langle O \rangle}$ with respect to the energy is defined by $${g_{\!\!\phantom{.}_{\mbox{\tiny OB}}}}{^{\rm K}}({\langle O \rangle};{\epsilon})\equiv \int O({\mbox{\boldmath $r$}}){f_{\!\!\phantom{.}_{\mbox{\tiny OB}}}}{^{\rm K}}({\mbox{\boldmath $r$}},{\epsilon})\,d^3 r, \label{eq:OBTFOden}$$ which reduces to the level density ${g_{\!\!\phantom{.}_{\mbox{\tiny OB}}}}{^{\rm K}}({\epsilon})$ for the case $O=1$. The problem of “particle gas” occurs at energies above the particle threshold, ${\epsilon}>0$, as in the same way as the level density. Then, similarly to Eq. (\[eq:gobasympt\]), the quantity (\[eq:OBTFOden\]) converges asymptotically as ${R_{\rm max}}\rightarrow \infty$ to $${g_{\!\!\phantom{.}_{\mbox{\tiny OB}}}}{^{\rm K}}({\langle O \rangle};{\epsilon}) \approx \frac{(2m)^{3/2}}{2\pi^2 \hbar^3} \int \left[\left({\epsilon}-V({\mbox{\boldmath $r$}})\right)^{1/2}-{\epsilon}^{1/2}\right]O({\mbox{\boldmath $r$}})\,d^3 r, \label{eq:Oasympt}$$ for the spatial observable $O({\mbox{\boldmath $r$}})$ like the radius or the quadrupole moment (assuming the upper cutoff energy of the pairing model space by the function (\[eq:fe\]), $\approx \tilde{\lambda}+\Lambda_u$, is smaller than ${\epsilon_{\rm kin}^{\rm max}}({\mbox{\boldmath $r$}})+V({\mbox{\boldmath $r$}})$ everywhere). This guarantees the convergence of the Kruppa expectation value $\langle O \rangle{^{\rm K}}_{\mbox{\tiny OB}}$ in the BCS treatment. Convergence in the Kruppa prescription {#sec:convergence} -------------------------------------- In order to see how the calculated rms radii converge as functions of ${N_{\rm osc}^{\rm max}}$, we have done many test calculations [@Ono10]. Generally, the rate of convergence is fast for stable nuclei as is already shown in Fig. \[fig:SnRMS\] (b), while it is slow for neutron rich nuclei depending on what kind of orbits exist near the Fermi surface. Figure \[fig:Rconv\] includes four selected examples of nuclei near the neutron drip-line (note the extension of the abscissa to ${N_{\rm osc}^{\rm max}}=40$ and the enlarged scale of ordinate compared with Fig. \[fig:SnRMS\]). All nuclei except $^{208}$Dy are spherical. The radius has converged within accuracy of about 0.5% already at ${N_{\rm osc}^{\rm max}}\approx 20$ in most cases where the pairing correlation is effective ($\Delta \ge 1$ MeV). The very slow convergence in the weak pairing correlations, e.g., the case of $\Delta{_{\rm n}}=0.5$ MeV in $^{88}$Ni (panel (a)) is due to the filling of the “halo” orbit near the Fermi surface, e.g., $\nu 3s_{1/2}$ (and $\nu 2d_{3/2}$) in $^{88}$Ni. The description of such spatially extended wave functions requires many oscillator basis states. Another striking feature shown in Fig. \[fig:Rconv\] is that the calculated radii with the largest basis size (${N_{\rm osc}^{\rm max}}=40$) are always smaller for larger pairing gaps: It is found that this calculated feature is rather general for nuclei near the drip-line, and mimics the pairing anti-halo effect in the HFB theory. We discuss this point in more detail in Sec. \[sec:antihalo\]. {width="150mm"} ![ The distribution of ${\langle r^2 \rangle}$ with respect to the single-particle energy $\epsilon$ (c.f. Eqs. (\[eq:OBTFOden\]), (\[eq:r2den\])), panel (a), the level density, panel (b), and the rms radius of orbits at ${\epsilon}$ defined in Eq. (\[eq:rmsorb\]), panel (c), as functions of $\epsilon$ for the spherical drip-line nucleus $^{126}$Zr. The OBTF and Strutinsky smoothed ones with the smoothing parameter $\gamma=0.8\,\hbar\omega$ and $1.8\,\hbar\omega$ and with the six order of the curvature correction polynomial ($p=3$ in Ref. [@TST10]) are included. ${N_{\rm osc}^{\rm max}}=40$ is used for the basis size. []{data-label="fig:OBTFr"}](OBTFr.eps){width="75mm"} In Ref. [@TST10] the OBTF level density is compared with the Strutinsky smoothed level density to clarify the convergence mechanism by the Kruppa prescription, i.e., by subtracting the contributions of free spectra. It is instructive to perform a similar analysis for the mean square radius. The microscopic quantity corresponding to the distribution of ${\langle O \rangle}={\langle r^2 \rangle}$ in Eq. (\[eq:OBTFOden\]) is $$g{^{\rm K}}({\langle r^2 \rangle};{\epsilon})=\sum_{i>0} \left[ \langle i |r^2| i \rangle\,\delta({\epsilon}-{\epsilon}_i) - \langle i |r^2| i \rangle_0\,\delta({\epsilon}-{\epsilon^{0}_{i}}) \right], \label{eq:r2den}$$ with which the mean square radius in Eq. (\[eq:rmsK\]) is calculated by integral $${\langle r^2 \rangle}{^{\rm K}}=\frac{1}{N}\int_{-\infty}^{\infty} g{^{\rm K}}({\langle r^2 \rangle};{\epsilon})\,2v^2({\epsilon})\,d{\epsilon}. \label{eq:rmsKden}$$ In Fig. \[fig:OBTFr\], we depict the Strutinsky smoothed quantity of Eq. (\[eq:r2den\]) and its OBTF approximation (panel (a)), as well as the corresponding level densities (panel (b)). The large model space with ${N_{\rm osc}^{\rm max}}=40$ is used. As for the OBTF approximation, we use Eq. (\[eq:Oasympt\]) with $O=r^2$ for ${\epsilon}>0$ and the conventional Thomas-Fermi expression, i.e., $\left({\epsilon}-V(r)\right)^{1/2}-{\epsilon}^{1/2}$ in the square brackets in Eq. (\[eq:Oasympt\]) being replaced with $\left|{\epsilon}-V(r)\right|^{1/2}\theta({\epsilon}-V(r))$, for ${\epsilon}<0$, considering only the central part of the Woods-Saxon potential with spherical symmetry for $V(r)$ [@TST10]. The sharp cusp behaviors at ${\epsilon}=0$ seen in all panels of Fig. \[fig:OBTFr\] are due to the threshold effect characteristic in the OBTF approximation [@TST10]. As it was already discussed in detail in Ref. [@TST10], apart from the oscillation and the cusp at ${\epsilon}=0$, the OBTF approximation reproduces nicely the microscopic level density smoothed with $\gamma=1.8\,\hbar\omega$. It also reproduces, but less nicely, the quantity $g({\langle r^2 \rangle};{\epsilon})$. The result of the smoothing with $\gamma=1.8\,\hbar\omega$ is larger than the OBTF approximation in $-20$ MeV $\le {\epsilon}\le -2$ MeV (where the error becomes largest) and ${\epsilon}\ge 2$ MeV (from where there is little contribution because of the vanishing occupation probability $v^2({\epsilon})$). It is helpful to introduce a physically more meaningful quantity, “rms radius of orbits at the single-particle energy ${\epsilon}$”, defined by $${\langle r^2 \rangle}^{1/2}({\epsilon})\equiv \sqrt{g{^{\rm K}}({\langle r^2 \rangle};{\epsilon})/g{^{\rm K}}({\epsilon})}, \label{eq:rmsorb}$$ which is also depicted in Fig. \[fig:OBTFr\] (c). The corresponding quantity in the HFB theory is the rms radius of the canonical basis states plotted versus the expectation value of the mean field Hamiltonian, whose gradually increasing tendency above the particle threshold is shown in Ref. [@Taj04]. From Eqs. (\[eq:gobasympt\]) and (\[eq:Oasympt\]) (and corresponding ones in the conventional Thomas-Fermi approximation for ${\epsilon}<0$), the Kruppa OBTF radius ${\langle r^2 \rangle}^{1/2}({\epsilon})$ in Eq. (\[eq:rmsorb\]) monotonically increases in ${\epsilon}< 0$, takes a maximum value ${\langle r^2 \rangle}^{1/2}_{\sqrt{V}}$ at ${\epsilon}=0$ and goes to an asymptotic value ${\langle r^2 \rangle}^{1/2}_V$ as ${\epsilon}\rightarrow\infty$, where (assuming the spherical symmetry) $$\begin{aligned} {\langle r^2 \rangle}_{\sqrt{V}}&\equiv& \int \sqrt{-V(r)}\,r^4dr/ \int \sqrt{-V(r)}\,r^2dr, \label{eq:rmspotsq} \\ {\langle r^2 \rangle}_V&\equiv& \int (-V(r))\,r^4dr/ \int (-V(r))\,r^2dr. \label{eq:rmspot}\end{aligned}$$ The microscopic Strutinsky smoothed radii in Fig. \[fig:OBTFr\] (c) behave similarly to the OBTF expression except a complete smearing out of the cusp (with $\gamma=1.8\,\hbar\omega$) and an increasing overestimation at positive energies. The monotonically increasing trend of the quantity ${\langle r^2 \rangle}^{1/2}({\epsilon})$ in ${\epsilon}< 0$ (smoothing width $\gamma=0.8\,\hbar\omega$) clearly indicates that the orbit with larger energy has larger radius in average, which leads to the increase of the total nuclear radius for larger pairing gap. The analyses in this subsection have shown how finite results for rms radii are obtained with the Kruppa prescription in the OBTF approximation and in which ways the quantum mechanical results deviate from the results of this approximation. Pairing anti-halo like effect {#sec:antihalo} ----------------------------- It is well-known that the occupation of the weakly bound orbit with low orbital angular momentum, typically $s$-orbits, leads to the nuclear halo and many examples are experimentally observed in light nuclei [@Tani96]. However, it is speculated that the increased pairing correlation binds these halo orbits more tightly and prevents the appearance of halo phenomena in heavier nuclei [@BDP00]. To confirm this speculation is one of the most interesting subjects both theoretically and experimentally. ![ Neutron’s rms radii calculated by the Kruppa method with ${N_{\rm osc}^{\rm max}}=40$ for the Ni–isotopes near the drip-line. The Kruppa-BCS equation is solved to obtain the selfconsistent pairing gap, where the pairing force strength is determined by the improved average gap method of Ref. [@TST10]. The results without including the pairing correlation are also included. []{data-label="fig:NiRMS"}](NiRMS.eps){width="75mm"} In this respect, the Ni–isotopes are very interesting, whose drip-line may possibly be reached in near future. The coordinate-space HFB calculation in Ref. [@GSG01], as an example, showed that indeed an abrupt increase of radii in the HF calculation is diminished by including the pairing correlation. We have calculated the same quantity with the Kruppa method in the following way. The calculated neutron drip-line based on the original Woods-Saxon potential [@RWyss] is slightly inside of the expected position. Therefore, the modification of its depth has been done according to a new systematic method developed in Ref. [@TST10]; then the drip-line isotope is the $A=92$ nucleus (the $A=94$ nucleus is unbound), while it is the $A=90$ nucleus in the calculation of Ref. [@GSG01]. The Kruppa-BCS equation is solved for each nucleus, and the full microscopic-macroscopic (Strutinsky shell correction) method is applied with the results that all the isotopes are spherical. The average pairing gap method with $\Delta=13/\sqrt{A}$ MeV is used to fix the strength $G$, and the calculated pairing gaps are $\Delta{_{\rm n}} \approx$ 1.2 to 1.7 MeV, monotonically increasing from the $A=80$ to $A=92$ isotopes. We depict the result of our Kruppa calculations for radii of Ni–isotopes in Fig. \[fig:NiRMS\]. Although the absolute values of radii are slightly larger in our calculations, the behavior of how the radii increase is very similar to the calculation shown in Fig. 7 of Ref. [@GSG01]. It may be worthwhile mentioning that a BCS treatment with including only the resonance orbits (the resonance-BCS) is tested in Ref. [@GSG01]: The calculated radii are not good approximation to those of HFB (no pairing anti-halo effect is reproduced), although the obtained binding energy is very good. This clearly indicates the subtlety of the neutron gas problem in the conventional BCS treatment, and we should be very careful to calculate such quantities for weakly bound systems [@DobNaz98]. The pairing anti-halo effect in the HFB theory [@BDP00] is a result of real shrinkage of the hole component of the quasiparticle orbits due to the selfconsistent modification of the potentials. Apparently, such effect is absent in the Kruppa prescription because the BCS treatment of quasiparticle does not change the spatial distribution of each single-particle orbit. Then it is surprising that a similar shrinkage of the rms radius comes out from the Kruppa-BCS calculation. The key to understand the reason is the rms radius ${\langle r^2 \rangle}^{1/2}({\epsilon})$ defined in Eq. (\[eq:rmsorb\]) and depicted in Fig. \[fig:OBTFr\] (c). Note that the total mean square radius is calculated as an integral (\[eq:rmsKden\]) with the weight of the BCS occupation probability $v^2({\epsilon})$, which is a step function at ${\epsilon}=\lambda$ smeared with the pairing gap $\Delta$. Therefore, if the quantity ${\langle r^2 \rangle}^{1/2}({\epsilon})$ is larger for ${\epsilon}< \lambda$ and smaller for ${\epsilon}> \lambda$, the increase of $\Delta$ leads to the effective shrinkage of the total mean square radius. In fact, this behavior, i.e., the local decrease near the particle threshold ${\epsilon}\approx 0$, is observed in the result of smaller smoothing width $\gamma=0.8\,\hbar\omega$ in Fig. \[fig:OBTFr\] (c). ![ The same as Fig. \[fig:OBTFr\] (c) but in an enlarged scale, and for other unstable nuclei, $^{88}$Ni (a) and $^{208}$Dy (b). The Strutinsky smoothed results with the smoothing parameter $\gamma=0.3\,\hbar\omega$ and $0.8\,\hbar\omega$ are included. []{data-label="fig:OBTFrex"}](OBTFrex.eps){width="75mm"} We show in Fig. \[fig:OBTFrex\] the same quantities for $^{88}$Ni and $^{208}$Dy, but are enlarged to see them more closely in the region near the particle threshold, which is most influential for pairing correlation in drip-line nuclei. To show further details, the Strutinsky smoothed results with the smaller averaging parameter $\gamma=0.3\,\hbar\omega \approx 2\Delta$ are also included. As is depicted in Fig. \[fig:OBTFr\] (c) and Fig. \[fig:OBTFrex\], we have found it quite general that the rms radius ${\langle r^2 \rangle}^{1/2}({\epsilon})$ in Eq. (\[eq:rmsorb\]) is a decreasing function near the particle threshold ${\epsilon}\approx 0$. This trend comes from two factors: One is that, just below the threshold, the rms radius is larger than the classical OBTF values due to the quantum mechanical effect of the weak binding. Another is that the rms radii of the free spectra are larger than those of full spectra just above the threshold, and the subtraction leads a considerably smaller total radius than the OBTF values given in Eq. (\[eq:rmspot\]). The former factor is natural but we do not understand the physical reason of the latter factor. Note that even if the (bound) halo orbit exists above the Fermi level, $\lambda < {\epsilon}(\mbox{halo}) < 0$, the subtraction effect of the free spectra is larger, and consequently the total radius shrinks near the drip line, $-\lambda \ltsim \Delta$. Thus, the subtraction of the free contributions is essential for the reduction of radius in the Kruppa method. The actual amount of decrease strongly depends on the nature of orbits just below the threshold; if the orbits are of halo nature, the amount is larger so that the pairing shrinks the rms radius more strongly. As is shown in Fig. \[fig:Rconv\], the rms radius of $^{88}$Ni having a halo orbit $\nu 3s_{1/2}$ more strongly shrinks than that of $^{208}$Dy when the pairing is increased, which clearly corresponds to the stronger decrease of the quantity ${\langle r^2 \rangle}^{1/2}({\epsilon})$ shown in Fig. \[fig:OBTFrex\]; the deformation in $^{208}$Dy also prevents the appearance of strong halo orbits due to the mixing of orbital angular momenta. It is now clear that the mechanism of the shrinking rms radius when increasing the pairing correlation is quite different in our Kruppa-BCS method from that in the HFB theory. From the discussion above, however, the essential ingredient for the conspicuous shrinking effect is the presence of the halo-like extended orbits below the Fermi energy, which is common to the situation where the strong pairing anti-halo effect is expected in the HFB approach. In this way, the Kruppa-BCS calculation well mimics the pairing anti-halo effect. Neutron skin {#sec:skin} ------------ Since the radius can be accurately calculated by employing the Kruppa prescription, it seems also meaningful to apply it to the calculation of the neutron skin. We compare the calculated neutron skin thickness of the Sn–isotopes with recent experimental data [@KAB04] in Fig. \[fig:SnSkin\]. Here the neutron skin thickness $\Delta r{_{\rm np}}$ is a difference of the neutron and proton rms radii, $$\Delta r{_{\rm np}} =\langle r^2 \rangle{_{\rm n}}^{1/2}-\langle r^2 \rangle{_{\rm p}}^{1/2}, \label{eq:nskin}$$ and they are calculated by the Kruppa method with ${N_{\rm osc}^{\rm max}}=20$, which is large enough for the (not very neutron rich) isotopes shown in the figure. ![ Neutron skin thickness calculated by the Kruppa method with ${N_{\rm osc}^{\rm max}}=20$ for the Sn–isotopes. Experimental data [@KAB04] are also included. The same calculational procedure is used as that in Fig. \[fig:NiRMS\]. []{data-label="fig:SnSkin"}](SnSkin.eps){width="75mm"} The procedure of the calculation is the same as in Fig. \[fig:NiRMS\] for the Ni–isotopes, see Ref. [@TST10] for details. It is interesting that the experimental data are nicely reproduced although the Woods-Saxon potential used in this work [@RWyss] is not particularly aimed to fit the quantity like the neutron skin thickness. (However, there are considerable ambiguities in the experimental neutron skin thickness depending on which types of experiments are used to extract it [@KFA99; @TJL01].) For the Sn–isotopes, a recent Skyrme HF+BCS calculation with the SLy4 functional gives slightly smaller neutron skin thickness, see, e.g., Ref. [@SGG07]. It is pointed out that the neutron skin has strong correlations with the coefficient of the asymmetry energy in the mass formula and with the equation of state of the asymmetric nuclear matter, see, e.g., Refs. [@Fur02; @WVR09]. The determination of skin thickness is therefore very important both theoretically and experimentally. We hope that one can utilize the Kruppa method used in this work as an equal alternative for such an investigation if one optimizes the parameters of the Woods-Saxon potential as well as the liquid drop model. Other observables {#sec:otherO} ----------------- As an example of different kinds of observables, we consider the moment of inertia about the $x$-axis (perpendicular to the symmetry axis) employing the Kruppa prescription, $${{\cal J}}{^{\rm K}}={{\cal J}}-{{\cal J}}_0. \label{eq:cJKruppa}$$ Here the free contribution, for example, is given as [@RS80] $${{\cal J}}_0=2\sum_{i,j>0} \frac{\left|\langle i|J_x|j\rangle_0\right|^2}{E({\epsilon^{0}_{i}})+E({\epsilon^{0}_{j}})} \left(u({\epsilon^{0}_{i}})v({\epsilon^{0}_{j}})-u({\epsilon^{0}_{j}})v({\epsilon^{0}_{i}})\right)^2, \label{eq:cJfree}$$ with $\langle i|J_x|j\rangle_0$ being the matrix element of the operator $J_x$ with respect to the free spectra, and $u(\epsilon)=\sqrt{1-v(\epsilon)^2}$. Eqs. (\[eq:cJKruppa\]) and (\[eq:cJfree\]) are obtained by the cranking method with taking the limit of zero rotational frequency: $${{\cal J}}{^{\rm K}}= \mathop{\mbox{lim}}_{\omega{_{\rm rot}}\rightarrow 0} \frac{\langle J_x \rangle{^{\rm K}}}{\omega{_{\rm rot}}} =\mathop{\mbox{lim}}_{\omega{_{\rm rot}}\rightarrow 0} \left[ \frac{\langle\omega{_{\rm rot}}|J_x|\omega{_{\rm rot}}\rangle}{\omega{_{\rm rot}}} -\frac{\langle\omega{_{\rm rot}}|J_x|\omega{_{\rm rot}}\rangle_0}{\omega{_{\rm rot}}} \right], \label{eq:crankKruppa}$$ where the state $|\omega{_{\rm rot}}\rangle$ is the cranked mean field associated with $H'=H-\omega{_{\rm rot}}J_x$, and $|\omega{_{\rm rot}}\rangle_0$ is the same but for the free Hamiltonian. Thus, the moment of inertia is essentially the expectation value of the angular momentum operator, which is composed not only of the coordinate but also of the momentum variables. ![ Neutron’s contributions to moment of inertia as functions of the neutron pairing gap calculated by the Kruppa and ordinary methods for the drip-line nucleus $^{40}$Mg. The size of model space is specified by ${N_{\rm osc}^{\rm max}}=30$, with which the results well converge. In the same way as Ref. [@YS08], are included two types of calculations, one for a weakly bound situation, $\lambda{_{\rm n}}\approx -0.3$ MeV, and another for a deeply bound situation, $\lambda{_{\rm n}}\approx -9.0$ MeV. The original depth of the Woods-Saxon potential is modified in each case. See text for details. []{data-label="fig:JmMg"}](JmMg.eps){width="75mm"} The effect of weak binding of constituent nucleons on the moment of inertia in neutron rich nuclei has been investigated in Ref. [@YS08] by the coordinate-space HFB method (the particle-hole channel is approximated by using the Woods-Saxon potential). It has been reported that the dependence of moment of inertia on the pairing gap is much stronger in drip-line nuclei, e.g., in $^{40}$Mg. We have done similar calculations but by using the simple BCS with the Kruppa prescription (\[eq:cJKruppa\]) and without it. The results of neutron moment of inertia, ${{\cal J}}{_{\rm n}}$, are depicted in Fig. \[fig:JmMg\]. The used deformation parameters are $\beta_2=0.3$ ($\beta_4=0.0$) [@YS08]. Two calculations are presented in Ref. [@YS08] with the neutron chemical potential being $\lambda{_{\rm n}}\approx -0.3$ and $-9.0$ MeV at $\Delta{_{\rm n}}\approx 0$, the latter of which corresponds to a stable nucleus and is artificially realized by deepening the potential. We have done the same adjustment to match with these calculations; the depth of the original Woods-Saxon potential [@RWyss] is modified by multiplying a factor 0.986 (1.365) for obtaining the desired value $\lambda{_{\rm n}}\approx -0.3$ ($-9.0$) MeV. Our result is very similar to that shown in Fig. 11 of Ref. [@YS08]; it reproduces the feature of stronger dependence of moment of inertia on pairing correlation for weakly bound systems. However, the free contribution ${{\cal J}}_0$ is rather small. In fact, the result ${{\cal J}}{_{\rm n}}{^{\rm K}}>{{\cal J}}{_{\rm n}}$ for the weak binding situation ($\lambda{_{\rm n}}\approx -0.3$ MeV) might seem strange at first sight because ${{\cal J}}_0$ in Eq. (\[eq:cJKruppa\]) is always positive. Note that the Kruppa-BCS equation is solved for the calculation of ${{\cal J}}{^{\rm K}}$ instead of the ordinary BCS equation for ${{\cal J}}$. Then the chemical potential obtained by the Kruppa-BCS equation is nearer to the particle threshold, and the moment of inertia ${{\cal J}}$ becomes larger. The difference, ${{\cal J}}$($\lambda$(Kruppa-BCS))$-{{\cal J}}$($\lambda$(ordinary-BCS)), is larger than ${{\cal J}}_0$, leading to the final result, ${{\cal J}}{^{\rm K}}>{{\cal J}}$. This clearly shows that even if ${{{\cal J}}}_0>0$ its contribution is minor. In fact, for the stable (deeply bound) situation ($\lambda{_{\rm n}}\approx -9.0$ MeV) the difference between the Kruppa and ordinary calculations is negligible as is seen in Fig. \[fig:JmMg\]. We have done systematic calculations for many heavier nuclei [@Ono10] and this is generally the case. Namely, the neutron gas problem is not so harmful as far as the calculation of moment of inertia is concerned; subtraction of the free contribution in Eq. (\[eq:cJKruppa\]) is not necessarily required. This is mainly because the free matrix elements of the $J_x$ operator in Eq. (\[eq:cJfree\]) are small, which reflects that the momentum operators contained in the angular momentum (${\mbox{\boldmath $l$}} = {\mbox{\boldmath $r$}} \times {\mbox{\boldmath $p$}}$) do not favor spatially-extended free wave functions in contrast to the spatial observable like the radius (${\mbox{\boldmath $r$}}^2$). Summary {#sec:summary} ======= We have extended the original Kruppa prescription for the single-particle level density to the calculation of one-body spatial observables like nuclear radius in the presence of pairing correlations. By simply subtracting the contribution of the free spectra, the effects of continuum states can be properly taken into account. The convergence property as increasing the basis size is carefully examined both numerically and analytically, and it is confirmed that reasonable convergence is attained for radius with manageable basis sizes except for the case of the halo-like situation with weak pairing correlation. The results of a few applications of the Kruppa method are presented for the basic quantities like nuclear radii, quadrupole moments, the neutron skin thickness, and the moment of inertia in very neutron rich nuclei. As for the moment of inertia, which is essentially the expectation value of the angular momentum operator at finite rotational frequency and is considered to be an example of operators consisting of both the coordinate and momentum, the effect of subtracting the free contributions is small. It has been found that the pairing anti-halo like effect naturally emerges, although the underlying mechanism is quite different. In this way, it has been shown that the so-called neutron gas problem is circumvented, and the new Kruppa-BCS method gives reliable results similar to those of more sophisticated HFB calculations. Combined with the improved microscopic-macroscopic method developed in Ref. [@TST10], we hope that the Kruppa prescription provides a simple, yet useful, new method for investigating nuclei far from stability. For the continuum single-particle level density, the Kruppa prescription in Eq. (\[eq:KruppaDensity\]) has a sound basis [@Kru98; @Shl92; @TO75]. However, the replacement in Eq. (\[eq:OKruppa\]) proposed in this work is missing such a solid foundation. It is an interesting future subject to clarify what is the meaning of this replacement, or on what kind of approximations it is based, from more fundamental many-body theory. ACKNOWLEDGEMENTS {#acknowledgements .unnumbered} ================ Useful discussions with Masayuki Yamagami are greatly appreciated. This work is supported in part by the JSPS Core-to-Core Program, International Research Network for Exotic Femto Systems (EFES), and by Grant-in-Aid for Scientific Research (C) No. 18540258 from Japan Society for the Promotion of Science. [99]{} I. Tanihata, J. Phys. G [**22**]{}, 157 (1996). J. Dobaczewski and W. Nazarewicz, Phill. Trans. R. Soc. Lond. A [**356**]{}, 2007 (1998). J. Dobaczewski, Act. Phys. Pol. B [**30**]{}, 1647 (1999). J. Dobaczewski, H. Flocard, and J. Treiner, Nucl. Phys. [**A422**]{}, 103 (1984). J. Dobaczewski, W. Nazarewicz, T. R. Werner, J. F. Berger, C. R. Chinn, and J. Decharg[é]{}, Phys. Rev. C [**53**]{}, 2809 (1996). J. Dobaczewski, W. Nazarewicz, and T. R. Werner, Z. Phys. A [**354**]{}, 27 (1996). M. Bolsterli, E. O. Fiset, J. R. Nix, and J. L. Norton, Phys. Rev. C [**5**]{}, 1050 (1972). A. T. Kruppa, Phys. Lett. [B431]{}, 237 (1998). T. Vertse, A. T. Kruppa, W. Nazarewicz, Phys. Rev. C [**61**]{}, 064317 (2000). N. Tajima, Y. R. Shimizu and S. Takahara, [*Improved microscopic-macroscopic approach incorporating the effects of continuum states*]{}, preprint arXiv:1006.2186. P. Bonche, H. Flocard, P.-H. Heenen, S.J. Krieger, and M.S. Weiss, Nucl. Phys. [**A443**]{}, 39 (1985). W. Nazarewicz, T. R. Werner, J. Dobaczewski, Phys. Rev. C [**50**]{}, 2860, (1994). R. Wyss, private communication. The parameter set of this new Woods-Saxon potential is tabulated in Ref. [@SS09]. T. Shoji and Y. R. Shimizu, Prog. Theor. [**121**]{}, 319 (2009). P. M[ö]{}ller, J.R. Nix, W.D. Myers, and W.J. Swiatecki, At. Data Nucl. Data Tables [**59**]{}, 185 (1995). K. Bennaceur,l J. Dobaczewski, and M. Ploszajczak, Phys. Lett. [**B496**]{}, 154 (2000). S. Shlomo, Nucl. Phys. [**A539**]{}, 17 (1992). T. Y. Tsang and T. A. Osborn, Nucl. Phys. [**A247**]{}, 566 (1975). P. Ring and P. Schuck, [*The nuclear many-body problem*]{}, Springer, New York (1980). T. Ono, Master Thesis (in Japanese), Department of Physics, Kyushu University, March, 2010. N. Tajima, Phys. Rev. C [**69**]{}, 034305 (2004). M. Grasso, N. Sandulescu, Nguyen Van Giai, and R. J. Liotta, Phys. Rev. C [**64**]{}, 064321 (2001). A. Krasznahorkay et al., Nucl. Phys. [**A731**]{}, 224 (2004). A. Krasznahorkay et al., Phys. Rev. Lett. [**82**]{}, 3216 (1999). A. Trzcińska et al., Phys. Rev. Lett. [**87**]{}, 082501 (2001). P. Sarriguren, M. K. Gaidarov, E. Moya de Guerra, and A. N. Antonov, Phys. Rev. C [**76**]{}, 044322 (2007). R. J. Furnstahl, Nucl. Phys. [**A706**]{}, 85 (2002). M. Warda, X. Viñas, X. Roca-Maza, M. Centelles, Phys. Rev. C [**80**]{}, 024316 (2009). M. Yamagami and Y. R. Shimizu, Phys. Rev. C [**77**]{}, 064319 (2008).

--- abstract: 'We discuss, based on first principles calculations, the possibility to tune the magnetism of oxygen vacancies at the (001) surface of strontium titanate ([[$\rm SrTiO_3$]{}]{}). The magnetic moment of single and clustered vacancies stemming from ${\ensuremath{\rm Ti-O}\xspace}$ broken bonds can be both quenched and stabilized controllably by chemical potential adjustment associated with doping the system with electrons or holes. We discuss to what extent this route to magnetization state control is robust against other external influences like chemical doping, mechanical action and electric field. Such control of vacancy state and magnetization can conceivably be achieved experimentally by using local probe tips.' author: - 'Oleg O. Brovko' - Erio Tosatti title: 'Controlling the magnetism of oxygen surface vacancies in [[$\rm SrTiO_3$]{}]{}through charging' --- Introduction {#introduction .unnumbered} ============ Throughout decades strontium titanate ([[$\rm SrTiO_3$]{}]{}) has continuously been in the spotlight of solid state research as a material with rich and varied physics. More recently it also became the substrate of choice for research and applications in the field of oxide electronics. [@McKee1998; @Goodenough2004; @Marshall2015] [[$\rm SrTiO_3$]{}]{}interfaces were shown to support a 2D electron gas with extremely high electron mobility values and diverse superconducting and relativistic physics. [@Ohtomo2004; @Ohta2007; @Santander-Syro2014; @Plumb2014; @Taniuchi2016] Strain-controlled room temperature ferroelectricity, [@Haeni2004] a useful trait for device applications, has been shown to exist in [[$\rm SrTiO_3$]{}]{}alongside quantum paraelectricity at low temperatures. [@Muller1979; @Muller1991] With the advent of spintronics [[$\rm SrTiO_3$]{}]{}has emerged as a wide band gap insulator material capable of itinerant, impurity and vacancy based magnetism. [@Marshall2015] It has been long known, that bulk impurities [@Blazey1983; @Bannikov2008; @Liu2015a] in [[$\rm SrTiO_3$]{}]{}can be used to tailor the oxide’s optical and electronic properties. Even in the absence of foreign atoms, pristine [[$\rm SrTiO_3$]{}]{}is prone to forming oxygen vacancies ([[$\rm V_{\!O}$]{}]{}) if annealed at higher temperatures under oxygen-poor conditions, [@Klie2000; @Janousch2007] when bombarded with noble gas ions [@Chang2015] or under intense laser or ultraviolet irradiation. [@Rao2014; @Zhang2015] Oxygen vacancies in [[$\rm SrTiO_3$]{}]{}bulk were studied extensively both theoretically [@Selme1983; @Shanthi1998; @Ricci2003; @Shein2007a; @Kotomin2008; @Kim2009; @Hou2010; @Liao2012; @Lin2012b; @Lin2013a; @Pavlenko2012; @Gryaznov2013; @Lopez-Bezanilla2015; @Lopez-Bezanilla2015a; @Zhang2015a] and experimentally. [@Klie2000; @Muller2004a; @Kalabukhov2007; @Janousch2007; @Kim2009; @Jiang2011; @Middey2012; @DeSouza2012; @Rao2014; @Rice2014; @Rice2014a; @Chang2015; @Zhang2015; @Trabelsi2016] They were shown not only to represent the key to metalization and control over the carrier density and mobility in [[$\rm SrTiO_3$]{}]{}, but also to be inherently magnetic. [@Shanthi1998; @Ricci2003; @Shein2007a; @Hou2010; @Liao2012; @Middey2012; @Lin2013a; @Rao2014; @Rice2014; @Rice2014a; @Lopez-Bezanilla2015; @Lopez-Bezanilla2015a; @Zhang2015a; @Trabelsi2016] Importantly for spintronic applications and nanoscale surface studies, similar trends for oxygen vacancies were found at [[$\rm SrTiO_3$]{}]{}surfaces. [@Crandles2010; @Pavlenko2013; @Xu2013; @Li2015; @Garcia-Castro2016; @Taniuchi2016; @Altmeyer2016; @Kimura1995; @Stashans2002; @Cai2006; @Alexandrov2009; @Alexandrov2009a; @Zhukovskii2009; @Shen2012a; @Choi2013] Depending on the concentration [@Pavlenko2013; @Lopez-Bezanilla2015] and clustering patterns [@Cuong2007; @Hou2010; @Liao2012; @Lopez-Bezanilla2015a; @Li2015] [[$\rm V_{\!O}$]{}]{}were shown to exhibit either local uncorrelated [@Lin2013a; @Garcia-Castro2016] magnetic signatures or, when sufficiently abundant, a long-range and stable magnetic order. [@Liao2012; @Pavlenko2013; @Rice2014; @Rice2014a; @Taniuchi2016; @Altmeyer2016; @Trabelsi2016] As abundant as the existing pool of literature on surface [[$\rm V_{\!O}$]{}]{}in [[$\rm SrTiO_3$]{}]{}is, it is also equally controversial, especially its theoretical component. Partially it is due to the limitations of the supercell computational approach, mainly used for crystal surface calculations. As our study has shown relatively large cells (typically at least 3-4 units of [[$\rm SrTiO_3$]{}]{}laterally) are required to quench the spurious interaction between the impurities/vacancies. In view of this we do not attempt to survey all the existing claims and assess their veracity referring the reader instead to the abundant above list of citations and concentrating instead on a different issue, namely the possibility to tailor the magnetic properties of the [[$\rm V_{\!O}$]{}]{}(the magnetic nature thereof is almost unanimously accepted). Some degree of control over [[$\rm V_{\!O}$]{}]{}-related magnetism was shown to be achievable through external or interface stress [@Zhang2015a], but the quest for an effective magnetization tuning mechanism is still on. Non-magnetic investigations of the [[$\rm V_{\!O}$]{}]{}have repeatedly demonstrated that the charge state of the vacancy has, as it is natural, a pronounced effect on its electronic and structural properties. The aim of the present work is to address the related important questions, namely: (i) how does charging of oxygen vacancies at the (100) surfaces of [[$\rm SrTiO_3$]{}]{}affect in detail its magnetic properties?; (ii) under what conditions would this charging and control be achievable?; and lastly (iii) how robust can this effect be against external influences, such as mechanical action, doping or electric field exposure? We present calculations and arguments showing that oxygen vacancies at, and close to, the (100) surface of [[$\rm SrTiO_3$]{}]{}can be indeed inherently magnetic depending on the charging state, and in addition the extent to which this magnetism is robust against structural changes in the atomic arrangement, doping and electric fields. Model results also highlight the way in which the magnetic state of vacancies and vacancy clusters can be either turned or quenched by externally induced charging. Methods and Geometries {#methods-and-geometries .unnumbered} ====================== First principles calculations were carried out in the framework of the density functional theory (DFT), based on the projector-augmented-wave method, [@Blochl1994] and a plane-wave basis set [@Kresse1996] as implemented in the Vienna Ab-initio Simulation Package (VASP). [@Kresse1993; @Kresse1996] Exchange and correlation were treated with the gradient-corrected functional as formalized by Perdew, Burke and Ernzerhof. [@Perdew1996] On-site Coulomb interaction corrections were accounted for in the framework of the LSDA+U formalism as introduced by Dudarev *et al.* [@Dudarev1998] The values for the Hubbard $U$ and $J$ parameters for the Ti $d$-orbitals were taken to be $5~{\ensuremath{\mathrm{eV}}\xspace}$ and $0.64~{\ensuremath{\mathrm{eV}}\xspace}$ respectively after Ref.  and checked against instability with $U=4~{\ensuremath{\mathrm{eV}}\xspace}$ after Ref. . For bulk and lattice constant calculations an energy cutoff of $600~{\ensuremath{\mathrm{eV}}\xspace}$ for the plane wave expansion and a Monkhorst-Pack $k$-point mesh [@Monkhorst1976] with $29\!\times\!29\!\times\!29$ points (before symmetry operations application) were used. The repeated-cell geometry for bulk vacancy calculations was taken to comprise $4\!\times\!3\!\times\!3$ [[$\rm SrTiO_3$]{}]{}unit cells. For surface calculations a slab of 4 [[$\rm SrTiO_3$]{}]{}unit layers was used where two units were fixed at bulk geometry and the top two were allowed to fully relax. In the $z$ direction the slabs were separated from their periodic images by $15{\ifmmode{\mathrm{\AAt}}\else{\AAt}\fi\xspace}$ of vacuum. For vacancy calculation atoms within two atomic shells from the vacancy site were allowed to relax. Obtained relaxations are in line with those found in similar studies, [@Zhukovskii2009; @Alexandrov2009; @Alexandrov2009a; @Janotti2014] namely generally outward vertical and inward (towards the impurity) relaxation of surface oxygen atoms surrounding the vacancy and outward in-plane relaxation of Ti atoms (depending on the charge state). The relaxations of Ti atoms around a neutral vacancy is a controversial issue and the reported relaxations vary among the above publications. We find an outward relaxation to be the energetic ground state. Moreover we find that the answer is sensitive to the size of the calculation cell, wich in our case was larger than in any of the previous studies. For each vacancy state, relevant quantities are calculated including among others, the formation energy of a $\nu$-atom oxygen vacancy with $q$ electrons ($q=0$ being the neutral vacancy) $$E_{\mathrm form}(\nu, q) = E_{\nu}^q + \nu \cdot E_{{\ensuremath{\rm O_2}\xspace}}/2 - E_{0}^q,\label{eq:eform}$$ as well al the work function $$\Phi = \epsilon_0 - E_{\mathrm F}$$ where $E_{\mathrm F}$ is the one-electron Fermi level position within Kohn-Sham eigenvalues, $\epsilon_0$ is the vacuum zero extrapolated as far above the surface as possible, and $E_{\nu}^q$, $E_0^q$ and $E_{{\ensuremath{\rm O_2}\xspace}}$ are the total energies of a $\nu$-atom oxygen vacancy, a clean surface (with $q$ electrons) and an diatomic oxygen molecule in gas phase respectively. The stable charge state of a vacancy or vacancy cluster will correspond to the lowest value of the grand potential $E_{\mathrm form}(\nu, q) - \mu \cdot q$. Here $\mu$ is the chemical potential, whose difference from vacuum zero coincides with the work function $\Phi$ in the absence of external fields, becoming in our context a free variable controlled doping, or by the external potential of a tip, etc. For the calculation of a reconstructed surface and surface-based vacancies the supercell of the calculation consisted of $3\!\times\!3$ and $4\!\times\!5$ [[$\rm SrTiO_3$]{}]{}unit cells in plane of the surface. The $k$-point mesh used in this case was $3\!\times\!3\!\times\!1$ and $\Gamma$-point-only respectively. Most of the conclusions derived in the present paper are based on calculations carried out for vacancies residing on, or close to, non-reconstructed [$\rm TiO$]{} and [$\rm SrO$]{} terminated [[$\rm SrTiO_3$]{}]{}surfaces. We deliberately chose to neglect the tetragonal antiferrodistortive phase of [[$\rm SrTiO_3$]{}]{}below the transition temperature of $\sim 105\!-\!110\degree\mathrm{C}$, [@Muller1969] since the latter is known to have a limited effect on the electronic properties of [[$\rm SrTiO_3$]{}]{} [@Choi2013] while neglecting it allows for a higher degree of generality and transferability in first principle calculations. From numerous experimental and theoretical studies it is known that the [$\rm TiO$]{} termination is prevalent and energetically slightly more stable under ambient conditions, [@Padilla1998; @Piskunov2005] though SrO terminated surfaces can be easily produced by growth in [$\rm Sr$]{}-rich atmosphere or controlled hydroxylation. [@Koster1998] Moreover, pristine surfaces of [[$\rm SrTiO_3$]{}]{}often reconstruct exhibiting a broad range of geometries, $2\!\times\!1$, $2\!\times\!2$, $c(4\!\times\!3)$, $c(6\!\times\!2)$, $\sqrt{5}\!\times\!\sqrt{5}-R26.6\degree$ and $\sqrt{13}\!\times\!\sqrt{13}-R37.7\degree$ being the the most common ones. [@Newell2007] To test the validity of our results for realistic reconstructed surfaces we chose, guided by discussions with M. Kisiel [^1], to investigate oxygen vacancies at several proposed realizations of one particular reconstruction pattern, namely a $2\!\times\!2$ one. [@Shiraki2010; @Lin2011a] Results and Discussion {#results-and-discussion .unnumbered} ====================== Single surface vacancy {#single-surface-vacancy .unnumbered} ---------------------- We start our investigation with examining a single oxygen vacancy at a [$\rm TiO$]{} surface of ${{\ensuremath{\rm SrTiO_3}\xspace}}(001)$. [^2] Our calculation yields, in accord with the extensive existing literature pool, a ground state with excess charge of the vacancy localized at the [[$\rm Ti$]{}]{}atoms neighboring the [[$\rm V_{\!O}$]{}]{}site \[see the geometry sketch in Fig. \[fig:dos:v1\](b)\]. The $d$-orbitals accommodating the charge form an impurity level deep inside the electronic band gap of the [[$\rm SrTiO_3$]{}]{}surface as is illustrated by the partial density of states \[PDOS, Fig. \[fig:dos:v1\](a)\] of one of the [[$\rm Ti$]{}]{}atoms neighboring the vacancy \[shaded dark blue in Fig. \[fig:dos:v1\](b)\]. The two excess electrons left behind by the departed [$\rm O$]{} atom cause the [[$\rm Ti$]{}]{}atoms to acquire magnetic moments of $1{\ensuremath{\rm \mu_B}\xspace}$ each, localized in their $d_z$ [[$\rm Ti$]{}]{}-orbital. The two [[$\rm Ti$]{}]{}spins facing each other across the vacancy are antiferromagnetically coupled with an exchange energy of about $750~{\ensuremath{\rm meV}\xspace}$, (as measured by the energy difference $E_{\rm tot}(m=2{\ensuremath{\rm \mu_B}\xspace}) - E_{\rm tot}(m=0{\ensuremath{\rm \mu_B}\xspace})$) so that the total vacancy magnetization is zero. To visualize the strength of localization of the charge trapped by the vacancy at the neighboring [[$\rm Ti$]{}]{}atoms we examine the spatial charge density of the Kohn-Sham states corresponding to the [[$\rm V_{\!O}$]{}]{}level. Fig. \[fig:chredist\](a) shows a cut of the above charge density by a plane normal to the surface and passing through the vacancy site and the neighboring [[$\rm Ti$]{}]{}atoms. It is apparent that the majority of the electron density of this state is concentrated in the Ti-$d$ orbitals with some of it spilling over to the neighboring oxygens. The antiferromagnetic alignment of [[$\rm Ti$]{}]{}spins can be clearly observed if we visualize the electron spin density by plotting the spin asymmety $P=\rho^{\uparrow}-\rho^{\downarrow}$ ($\rho^{\uparrow,\downarrow}$ are the densities of majority and minority electrons respectively) of electrons in the same plane as shown in Fig. \[fig:chredist\](a). The antisymmetry of the map is a usual signature of antiferromagnetism. The atoms in the first neigbor shell of the magnetic [[$\rm Ti$]{}]{}atoms acquire a small induced magnetic moment, but note that the color scale here is non-linear, so that the absolute spin polarization value on those sites is almost negligibly small. Another way of visualizing the charge contents of the oxygen vacancy is looking at the charge density redistribution caused by its creation. In Fig. \[fig:chredist\](c) we plot the charge redistribution $\Delta\rho = \rho_{{{\ensuremath{\rm V_{\!O}}\xspace}}} + \rho_{{\ensuremath{\rm O}\xspace}} - \rho_{{\ensuremath{\rm clean}\xspace}}$, where $\rho_{{{\ensuremath{\rm V_{\!O}}\xspace}}}$, $\rho_{{\ensuremath{\rm clean}\xspace}}$ and $\rho_{{\ensuremath{\rm O}\xspace}}$ are the charge density distributions of the surface slab with and without a vacancy and a free-standing oxygen atom respectively \[the cross-section plane is the same as for Fig. \[fig:chredist\](a)\]. It represents visually the transfer of the electrons formerly attached to oxygen (nominally two) from the site now vacant to the $d$-orbitals of the neighboring [[$\rm Ti$]{}]{}atoms. Since the magnetization of the [[$\rm Ti$]{}]{}atoms is directly linked to the excess charge conveyed to them by creation of the vacancy electrons it stands to reason that the magnetization state should be highly susceptible to charge doping and/or depletion in the system. To verify that we calculate the ground state of the system (the largest supercell considered here, *i.e.* the $4\!\times\!5$ slab) with the charge $q$ of the cell increased or reduced by an integer number of electrons $n_{{\ensuremath{\mathrm{\overline{e}}}\xspace}}$. On account of the nonmagnetic and insulating character of bulk [[$\rm SrTiO_3$]{}]{}the total magnetization $m$ is set to $1{\ensuremath{\rm \mu_B}\xspace}$ for odd $n_{{\ensuremath{\mathrm{\overline{e}}}\xspace}}$, and $m= 0$ otherwise. Higher magnetizations were also explored and found to be energetically unfavorable with respect to the lower ones. The vacancy electronic structure of Fig. \[fig:v1\](a) makes it clear why. There are only two midgap [[$\rm Ti$]{}]{}-broken bond levels in the insulating gap; magnetizations larger than $1~{\ensuremath{\rm \mu_B}\xspace}$ involve additional promotion of electrons/holes from midgap to the conduction or valence-derived bands and are thus energetically more costly. [^3] Fig. \[fig:v1\](a) shows the evolution of the magnetic moments’ magnitude $|m_{{{\ensuremath{\rm Ti}\xspace}}}|$ (blue squares) and the valence[^4] Bader charge [@Bader1985; @Tang2009] $q_{{{\ensuremath{\rm Ti}\xspace}}}$ (red circles) of each of the two [[$\rm Ti$]{}]{}atoms neighboring the [[$\rm V_{\!O}$]{}]{}as we vary the number of electrons. While lowering of the chemical potential (increasing number of electrons) barely has an effect on the already occupied impurity level localized at [[$\rm Ti$]{}]{}atoms (their Bader charge remains unchanged as the cell is negatively charged), an increase in chemical potential causes the impurity level to gradually deplete, resulting concurrently in a reduction of $m_{{{\ensuremath{\rm Ti}\xspace}}}$. The corresponding change of the work function in [$\rm eV$]{}and the vacancy formation energy change are plotted in Fig. \[fig:v1\](b) (green triangles and blue squares respectively). Note that in the numerical approach used in the present study the chemical potential is altered by constraining the number of electrons in the calculation cell, the actual change of the Bader charge of the near-vacancy [[$\rm Ti$]{}]{}atoms associated with integer electron addition or depletion is relatively small (fractions of an electron), which is nonetheless sufficient to completely quench their magnetic moment. The remaining charge is detracted from the surrounding atoms with non-zero occupation of the vacancy level (visually the extent of impurity level can be estimated from Fig. \[fig:chredist\](a)). This demonstrates that control over magnetism in oxygen vacancies at [[$\rm SrTiO_3$]{}]{}surfaces can be achieved through chemical potential tailoring (charge injection/depletion). The magnetic coupling between the near-vacancy [[$\rm Ti$]{}]{}is found to be antiferromagnetic as shown by the total moment $m$ of the impurity alternating between zero for an uncharged vacancy and $1~{\ensuremath{\rm \mu_B}\xspace}$ for $|n_{{\ensuremath{\mathrm{\overline{e}}}\xspace}}|=1{\ensuremath{\mathrm{\overline{e}}}\xspace}$. The next important thing to consider, however, is that while our constrained density functional calculation does yield a ground state solution for each charge state of the finite periodic system, it does not automatically imply that the state shall be the ground state of an isolated impurity in a real-life [[$\rm SrTiO_3$]{}]{}sample, where the result would be determined by the position of the chemical potential. To examine the landscape of achievable charge states we plot in Fig. \[fig:v1\](c) the formation energy (Eq. \[eq:eform\]) [^5] diagram for different charge states of a single oxygen vacancy at the [$\rm TiO$]{}-terminated ${{\ensuremath{\rm SrTiO_3}\xspace}}(001)$ unreconstructed surface as a function of the chemical potential (given with respect to the valence band maximum (VBM). The markers on the lines represent the constrained-charge ground state given by DFT. The green circles represent the transition points and are annotated to mark the charge states between which the system switches (with “$+$” and “$-$” denoting the number of holes in the system with respect to the neutral state “0”). At any given chemical potential the lowest line in the diagram defines the preferred charge state of the oxygen vacancy. Somewhat unexpectedly, our calculations indicate that the neutral state of the vacancy is not achieved at any chemical potential. In its place, a direct transition from the singly positively charged to singly negatively charged state is preferred. In the real world this would amount to one electron of the vacancy in a neutral state being donated to the host matrix. While this is in line with numerous other theoretical predictions of oxygen vacancies being responsible for the formation of an itinerant electron gas at the surface of [[$\rm SrTiO_3$]{}]{}, it has also to be considered, that the range of chemical potentials where a neutral vacancy is closest to becoming the ground state is very close to the width of the band gap, [@Zhukovskii2009] which is underestimated by the DFT by almost an [$\rm eV$]{}. Correcting the size of the gap in the calculation might yield a slightly different formation energy diagram from that depicted in Fig. \[fig:v1\](c), perhaps with a range of chemical potentials allowing for an uncharged ground state of the [[$\rm V_{\!O}$]{}]{}. However, the above mentioned gap correction in DFT would require employing hybrid exchange and correlation functionals which unfortunately is computationally too heavy for the system sizes considered here. Nonetheless, the qualitative statement that a magnetic state of an oxygen vacancy (as a quantum dot) is susceptible to manipulation through chemical potential engineering still stands. To underscore and confirm the generality of the above conclusion we examine first of all several other configurations of single oxygen vacancies at [[$\rm SrTiO_3$]{}]{}surfaces. We calculate single vacancies in the second and third layers of the [$\rm TiO_2$]{}-terminated surface, as well as [[$\rm V_{\!O}$]{}]{}in the first and second layers of the [$\rm SrO$]{}-terminated [[$\rm SrTiO_3$]{}]{}crystal. All cases exhibit similar traits, *i.e.* magnetic [[$\rm Ti$]{}]{}ion pairs at the vacancy site in its neutral state and responsiveness to charge manipulation leading to magnetic moment reduction and quenching following the drop of chemical potential. A valid question to address at this point would be the scope of mechanisms available to achieve the chemical potential tuning, *i.e.* external agents capable of changing the local chemistry in the vicinity of the oxygen vacancy. Addressing all of them would spring the limits of the present study, but we would like to specifically mention one straightforward way of injecting charge (both electrons and holes) into a surface vacancy which is particularly relevant in the framework of contemporary spintronic applications, namely to approach the [[$\rm V_{\!O}$]{}]{}with a local probe (such as an STM or AFM tip) and possibly create a potential gradient (electric field) giving rise the transfer/tunnelling of electrons between the tip and the vacancy. While it is also conceivable [@Li2013c; @Brovko2014; @Shimizu2015; @Coey2016] to use the effect of the electric field, doping or mechanical action to tune the charge and magnetic properties of surface defects, our test calculations involving exposure of a [[$\rm V_{\!O}$]{}]{}at a [[$\rm Ti$]{}]{}-terminated [[$\rm SrTiO_3$]{}]{}surface to electric fields up to $1~{\ensuremath{\rm eV}\xspace}/{\ifmmode{\mathrm{\AAt}}\else{\AAt}\fi\xspace}$ and mechanical lift-off of the topmost layer with forces in excess of those achievable by Van der Waals forces in junction geometry have shown that neither of the latter has a significant effect on the charge state or indeed the magnetic properties of an oxygen vacancy. This means, on one hand, that the choice of tools for harnessing oxygen vacancies as magnetic quantum dots it limited, but on the other hand it indicates that directly addressing [[$\rm V_{\!O}$]{}]{}-s with a local probe tip is a viable channel of spin manipulation not susceptible to environmental instabilities. Note, however, that in the present study the change in electron number, as attainable in real life by the external potential exerted by a tip, has been forced onto a system devoid of free charge carriers so that no other effects of the external potential are expected apart from the local chemical potential change. In real [[$\rm SrTiO_3$]{}]{}samples the accumulation of impurities at the surface can cause a non-zero population of free or semi-free surface electrons and so can the surface state. [@Santander-Syro2011; @Meevasana2011; @Shen2012a; @Marshall2015; @Coey2016] Those electrons can be redistributed by the application of a bias via a local probe tip leading to a population or depletion of a vacancy site under the tip and thus a change of its magnetic moment. This scenario can actually be seen as the main practical instrument to alter the chemical potential locally leading to the alteration of the impurities’ spin state as described above. Vacancy clusters {#vacancy-clusters .unnumbered} ---------------- So far we described charge and magnetization switching in a single, isolated oxygen vacancy. Both experimental and theoretical evidence however point towards vacancy migration to surfaces and interfaces and their aggregation into clusters. [@Alexandrov2009; @Marshall2015; @Li2015] We therefore proceed to carry out calculations for oxygen vacancy clusters. In view of controversial claims in the literature concerning the nature of in-plane vacancy clustering, see e.g. , we seek to further increase the generality of our conclusions by studying several representative two-dimensional vacancy clusters residing in the topmost surface layer (see Supplemental Material Fig. S2 for the list of cluster configurations studied) for the signatures of magnetism and charge state transitions. We find that all the vacancy clusters studied exhibit ground state magnetization of [[$\rm Ti$]{}]{}atoms neighboring the [[$\rm V_{\!O}$]{}]{}sites. Moreover, for most non-linear vacancy cluster configurations the coupling between the [[$\rm Ti$]{}]{}atom spins is antiferromagnetic, yielding either $m=0$ ($m=1{\ensuremath{\rm \mu_B}\xspace}$) net magnetic moment of the vacancy cluster as a whole for even (odd) electron number $n_{{\ensuremath{\mathrm{\overline{e}}}\xspace}}$. Similarly to the case of a single vacancy, electron depletion causes the vacancy cluster to lose the charge localized therein and the moment of the [[$\rm Ti$]{}]{}atoms to be gradually quenched. We consider however that previous studies, also confirmed by our calculations, predict the formation of stable linear vacancy chains [@Cuong2007] as an important mode of vacancy clustering. We address this possibility by studying as an illustrative example the case of a linear vacancy chain consisting of 3 contiguous oxygen vacancies. Our calculations show that similar to other linear vacancy arrays spins on near-vacancy [[$\rm Ti$]{}]{}atoms are in this case ferromagnetically coupled therefore stabilizing the maximum $m$ value at any given charge state $n_{{\ensuremath{\mathrm{\overline{e}}}\xspace}}$. The neutral state of the triple linear vacancy has therefore $m= 6{\ensuremath{\rm \mu_B}\xspace}$. The sketch of the system is presented in Fig. \[fig:dos:v3b\](b) and the partial projected DOS of one of the [[$\rm Ti$]{}]{}atoms neighboring the middle oxygen vacancy is shown in panel (a) of the same figure. The presence can be noted of several [[$\rm Ti$]{}]{}-connected levels in the vacancy cluster gap. The apparent broadening caused by their crowding in the, as well as by the unphysical overlap due to periodic boundary conditions, preserves the magnetic character thereof in some cases raising the vacancy magnetization from lowest to highest. The charging dependence of significant physical variables similar to that given above for a single vacancy are presented in Fig. \[fig:v3b\]. Panel (a) shows the evolution of the valence Bader charge (red circles for average charge $q_{\rm av}$ of the [[$\rm Ti$]{}]{}atoms neighboring the chain \[shaded dark blue in Fig. \[fig:dos:v3b\](b)\] and pale red rhombs for the charge $q_{\rm c}$ on one of the [[$\rm Ti$]{}]{}atoms neighboring the middle oxygen vacancy) and the cumulative magnetic moment $m$ (blue squares) of the [[$\rm V_{\!O}$]{}]{}chain as the number of electrons in the calculation supercell is varied. Here we observe a similar tendency as in the case of a single vacancy – electron depletion (or hole doping) causes a decrease in the localized charge and an ensuing reduction of the magnetic moment. The surplus charge is again predominantly localized at the [[$\rm Ti$]{}]{}atoms surrounding the vacancy chain. Note also how the depletion of vacancy charge affects the edge atoms stronger than it does the central ones \[compare the rhombs and the circles in Fig. \[fig:v3b\](a)\]. The total magnetic moment is, as said above $m= (6-|n_{{\ensuremath{\mathrm{\overline{e}}}\xspace}}|){\ensuremath{\rm \mu_B}\xspace}$. Fig. \[fig:v3b\](c) shows the formation energy diagram (similar to that shown in Fig. \[fig:v1\](c) for a single oxygen vacancy) of a triple linear oxygen vacancy cluster at the [$\rm TiO$]{}-terminated ${{\ensuremath{\rm SrTiO_3}\xspace}}(001)$ surface as a function of the chemical potential (given with respect to the valence band maximum (VBM). Here, even clearer than in the case of a single [[$\rm V_{\!O}$]{}]{}, one can observe a sequence of charge state transitions, starting with a $3{\ensuremath{\mathrm{\mathbf{h}}}\xspace}\rightarrow 2{\ensuremath{\mathrm{\mathbf{h}}}\xspace}$ at a chemical potential of $1.25~{\ensuremath{\rm eV}\xspace}$ above the valence band maximum and followed by a sequence of transition taking the system through the neutrality to electron doped states. We stress again, that since each state change constitutes a dissipation channel for the external agent responsible for the change, the charge and spin state transitions should be observable experimentally, f.e. within the scope of atomic force microscopy at low temperatures by monitoring the damping of the AFM cantilever oscillations directly related to the dissipation due to the charge and magnetization state changes. For longer linear vacancy chains the magnetization behavior with charging is expected to be very similar to that presented above. The limiting case of an infinite linear oxygen vacancy chain is shown in the Supplemental Material Fig. S3 (sketch and PDOS) and Fig. S4 (formation energy diagram). At this point we would like to mention a possibility to experimentally sense or measure the presence of charge and magnetic moment transitions in the vacancy quantum dots. To achieve that, we would argue, that, since any change of state or level crossing (with subsequent relaxation of the system) implies a dissipation channel for the external agent provoking the change, the charge and magnetization state changes can be tested experimentally, f.e. within the scope of such dissipation sensitive techniques as atomic force microscopy (AFM) at low temperatures. [@Stomp2005; @Cockins2010; @Gysin2011; @Kisiel2015; @Miyahara2017] The relevant quantity hereby would be the damping of the AFM cantilever oscillations directly linked to the magnetic-state-transition induced dissipation in the system. Finally we note that motivated by a discussion with experimental colleagues we studied oxygen vacancies at one of the many known reconstruction patterns of a [$\rm TiO_2$]{}-terminated [[$\rm SrTiO_3$]{}]{}(001) surface, namely the $2\!\times\!2$ reconstruction (see Supplemental Material Fig. S5 and its caption). We find that including the reconstruction into consideration does not alter the main conclusions of the present study – the lack of an oxygen atom inevitably leads to an excess charge localization on the neighboring Titania and results in a spontaneous magnetization of the latter. Depleting the localization by removing electrons from the surface (f.e. by locally altering the chemical potential) results in a reduction of the magnetic moment of the Titania with its subsequent complete quenching. Conclusions {#conclusions .unnumbered} =========== Oxygen vacancies at (001) surfaces of [[$\rm SrTiO_3$]{}]{}can be regarded as externally accessible magnetic quantum dots. Their electronic states are determined by [[$\rm Ti$]{}]{}broken bonds that give rise to very localized $d$ states in the gap of the insulating host. More or less like for transition metal impurities, electron-electron interactions give rise to a multiplicity of charge and spin states whose energies are relatively close. The direct exchange coupling of two Ti broken bonds facing each other across the missing [$\rm O$]{} surface site is strong and antiferromagnetic, therefore the single vacancy only stabilizes states with low or zero total magnetization. Multiple vacancies offer an even richer scenario, depending on the relative positions of the missing [$\rm O$]{} atoms. In this case, an overall state of highest magnetization can also be achieved, as in the linear case which we studied in detail. Our study of the impurity cluster properties as a function of chemical potential indicates that as in a quantum dot the different surface vacancy states can be tuned and switched by adjusting the local chemical potential, suggesting their investigation with a local probe tip. The tip-induced transition between different charge and spin states should be traceable in dissipation-sensitive experiments such as the atomic force microscopy. Acknowledgements {#acknowledgements .unnumbered} ================ We are very grateful to Marcin Kisiel and collaborators who provided our initial motivation for this research. We gratefully acknowledge the financial support of the ERC Grant No. 320796, MODPHYSFRICT, as well as that of the COST Action MP1303 project, and thank Valeri S. Stepanyuk for the support in terms of calculational resources. Supplemental Information {#supplemental-information .unnumbered} ======================== See Supplemental Material for reference calculations of [[$\rm SrTiO_3$]{}]{}bulk, list of extended vacancies and surface reconstruction configurations studied as well as the asymptotic example of an extended vacancy - an infinite linear chain - for which a charging manipulation analysis (similar to that presented above for a single and tripple linear vacancy) is given. [92]{}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 10.1103/PhysRevLett.81.3014) [****,  ()](\doibase 10.1088/0034-4885/67/11/R01) in [**](\doibase 10.1007/978-3-319-14367-5_11), Vol. ,  (, , ) pp.  [****,  ()](\doibase 10.1038/nature04773) [****,  ()](\doibase 10.1038/nmat1821) [****,  ()](\doibase 10.1038/nmat4107) [****,  ()](\doibase 10.1103/PhysRevLett.113.086801),  [****,  ()](\doibase 10.1038/ncomms11781) [****,  ()](\doibase 10.1038/nature02773) [****, ()](\doibase 10.1103/PhysRevB.19.3593) [****,  ()](\doibase 10.1007/BF01313549) [****,  ()](\doibase 10.1016/0038-1098(83)90332-0) [****,  ()](\doibase 10.1016/j.jmmm.2007.09.012) [****, ()](\doibase 10.1016/j.matchemphys.2015.04.009) [****,  ()](\doibase 10.1557/PROC-654-AA1.7.1) [****,  ()](\doibase 10.1002/adma.200602915),  [****,  ()](\doibase 10.7567/APEX.8.055701) [****,  ()](\doibase 10.1063/1.4891184) [****,  ()](\doibase 10.1016/j.apsusc.2015.02.035) [****,  ()](\doibase 10.1088/0022-3719/16/13/018) [****, ()](\doibase 10.1103/PhysRevB.57.2153) [****,  ()](\doibase 10.1103/PhysRevB.68.224105) [****,  ()](\doibase 10.1016/j.physleta.2007.06.013) [****,  ()](\doibase 10.1088/1742-6596/117/1/012019) [****,  ()](\doibase 10.1063/1.3139767), [****,  ()](\doibase 10.1143/JPSJ.79.114704) [****,  ()](\doibase 10.1111/jace.12072) [****,  ()](\doibase 10.1103/PhysRevB.86.161102) [****,  ()](\doibase 10.1103/PhysRevLett.111.217601) [****,  ()](\doibase 10.1103/PhysRevB.86.064431),  [****,  ()](\doibase 10.1021/jp400609e) [****,  ()](\doibase 10.1103/PhysRevB.92.115112) [****, ()](\doibase 10.1063/1.4932347),  [****,  ()](\doibase 10.1039/C5CP04310G) [****,  ()](\doibase 10.1038/nature02756) [****,  ()](\doibase 10.1103/PhysRevB.75.121404) [****,  ()](\doibase 10.1063/1.3622623) [****,  ()](\doibase 10.1063/1.4738785) [****,  ()](\doibase 10.1103/PhysRevB.85.174109) [****,  ()](\doibase 10.1038/nmat3914),  [****,  ()](\doibase 10.1116/1.4871691) [****, ()](\doibase 10.1016/j.jallcom.2016.04.145) [**** (), 10.1063/1.3481344](\doibase 10.1063/1.3481344),  [****,  ()](\doibase 10.1007/s10948-012-2045-8) [****,  ()](\doibase 10.1063/1.4825257) [****,  ()](\doibase 10.1016/j.susc.2015.04.020) [****,  ()](\doibase 10.1103/PhysRevB.93.045405),  [****,  ()](\doibase 10.1103/PhysRevLett.116.157203),  [****,  ()](\doibase 10.1103/PhysRevB.51.11049) [****,  ()](\doibase 10.1016/S0039-6028(01)01656-9) [****,  ()](\doibase 10.1063/1.2189225) **, @noop [Ph.D. thesis]{},  () [****,  ()](\doibase 10.1140/epjb/e2009-00339-4) [****, ()](\doibase 10.1016/j.ssc.2009.05.023) [****,  ()](\doibase 10.1103/PhysRevB.86.195119),  [****,  ()](\doibase 10.1002/adma.201203580) [****,  ()](\doibase 10.1103/PhysRevLett.98.115503) [****,  ()](\doibase 10.1103/PhysRevB.50.17953) [****,  ()](\doibase 10.1103/PhysRevB.54.11169) [****, ()](\doibase 10.1103/PhysRevB.47.558) [****,  ()](http://www.ncbi.nlm.nih.gov/pubmed/10062328) [****,  ()](\doibase 10.1103/PhysRevB.57.1505) [****,  ()](\doibase 10.1038/nature04773) [****,  ()](\doibase 10.1103/PhysRevB.51.12880) [****, ()](\doibase 10.1103/PhysRevB.13.5188) [****,  ()](\doibase 10.1103/PhysRevB.90.085202),  [****, ()](\doibase 10.1103/PhysRev.186.361) [****, ()](\doibase 10.1016/S0039-6028(98)00670-0),  [****,  ()](\doibase 10.1016/j.susc.2004.11.008) [****,  ()](\doibase 10.1063/1.122630) [****,  ()](\doibase 10.1103/PhysRevB.75.205429) [****,  ()](\doibase 10.1063/1.3447796) [****, ()](\doibase 10.1016/j.susc.2011.06.001) [**](http://pubs.acs.org/doi/abs/10.1021/ar00109a003) (, , ) p. [****,  ()](\doibase 10.1088/0953-8984/21/8/084204) [****,  ()](\doibase 10.1021/nl402088f) [****,  ()](\doibase 10.1088/0953-8984/26/9/093001) [****,  ()](\doibase 10.1103/PhysRevB.92.165304) [****,  ()](\doibase 10.1088/0953-8984/28/48/485001) [****,  ()](\doibase 10.1038/nature09720),  [****,  ()](\doibase 10.1038/nmat2943),  [**** (), 10.1088/1367-2630/17/2/023034](\doibase 10.1088/1367-2630/17/2/023034),  [****,  ()](\doibase 10.1103/PhysRevLett.94.056802),  [****,  ()](\doibase 10.1073/pnas.0912716107),  [****,  ()](\doibase 10.1063/1.3551603) [****,  ()](\doibase 10.1103/PhysRevLett.115.046101) [****, ()](\doibase 10.1088/1361-6528/aa5261),  Bulk $\mathbf{SrTiO_3(001)}$ and single vacancies therein ========================================================= Bulk [[$\rm SrTiO_3$]{}]{}\[Fig. \[fig:dos:bulk\](a)\] calculations were consistent with the state-of-the-art literature. The equilibrium lattice constant is found to be $a_{\rm lat} = 3.99~{\ifmmode{\mathrm{\AAt}}\else{\AAt}\fi\xspace}$, the gap is $2.25~{\ensuremath{\rm eV}\xspace}$. It is reduced with respect to the experimental value of $3.25~{\ensuremath{\rm eV}\xspace}$ which however does not impact the qualitative message of the present work. The partial atom and orbital projected DOS of the system is shown in Fig. \[fig:dos:bulk\](b). A single oxygen vacancy calculated including full relaxation of the surrounding atoms yields a ground state with two electrons trapped in the $d$-orbitals of Titania atoms neighboring the vacancy site. The excess charge leads to a spontaneous magnetization of the Titania leaving them with a magnetic moment of $\sim 0.5~{\ensuremath{\rm \mu_B}\xspace}$ each, yet the net magnetic moment of the vacancy site is zero, since according to our calculations the two [[$\rm Ti$]{}]{}spins are coupled antiferromagnetically with an exchange energy of about $180~{\ensuremath{\rm meV}\xspace}$. The PDOS of the system is shown in Fig. \[fig:dos:bulk\](c). Extended vacancies at $\mathbf{SrTiO_3(001)}$ ============================================= To sample a larger set of potential experimental realizations of surface vacancies we have studied a set of typical configurations of single, double, triple and quadruple 2D oxygen vacancies residing in the topmost layer of the [[$\rm SrTiO_3$]{}]{}(001) surface. The studied geometries are shown in Fig. \[fig:cfg:ext\]. All the studied vacancy clusters exhibit ground state magnetization of [[$\rm Ti$]{}]{}atoms neighboring to [[$\rm V_{\!O}$]{}]{}sites. Moreover, for most vacancy cluster configurations the coupling between the [[$\rm Ti$]{}]{}atom spins is antiferromagnetic, resulting in either a zero net magnetic moment of the vacancy cluster as a whole or, where the symmetry does not allow for full compensation of the magnetic moment, small met magnetization values. Similarly to the case of a single vacancy, electron depletion causes the vacancy cluster to loose the charge localized therein and the moment of the [[$\rm Ti$]{}]{}atoms to be gradually quenched. A limit case of a infinite linear chain of impurities depicted in Fig. \[fig:cfg:ext\](d, top panel) is in its properties very close the case of a tripple impurity discussed in the text. It exhibits ferromagnetic alignment of the [[$\rm Ti$]{}]{}atom spins with about $2{\ensuremath{\rm \mu_B}\xspace}$ per vacancy site in the neutral state, which additively accumulates with increasing chain length. The sketch of the system is presented in Fig. \[fig:dos:vlin\](b) and the partial projected DOS of one of the [[$\rm Ti$]{}]{}atoms neighboring the oxygen vacancy chain is shown in Fig. \[fig:dos:vlin\](a). The vacancy gap state is slightly broadened also here, similar to the case of a tripple [[$\rm V_{\!O}$]{}]{}discussed in the main text. Panel (a) of Fig. \[fig:vlin\] shows the evolution of the valence Bader charge (red circles) and the cumulative magnetic moment (blue squares) of the two [[$\rm Ti$]{}]{}atoms surrounding a unit (monomer) of the vacancy chain \[note the difference to Fig. 3(a) where the magnetic moment of only one [[$\rm Ti$]{}]{}was presented\] as the number of electrons in the calculation supercell is varied. Here we observe a similar tendency as in the case of a single vacancy – electron depletion (or hole doping) causes a decrease in the localized charge and an ensuing reduction of the magnetic moment. Fig. \[fig:vlin\](c) shows the formation energy (per oxygen unit) diagram (similar to that shown in Fig. \[fig:v1\](c) for a single oxygen vacancy) of a linear oxygen vacancy chain at the [$\rm TiO$]{}-terminated ${{\ensuremath{\rm SrTiO_3}\xspace}}(001)$ surface as a function of the chemical potential (given with respect to the valence band maximum (VBM). A sequence of charge state transitions can be observed starting with a $3{\ensuremath{\mathrm{\mathbf{h}}}\xspace}\rightarrow 2{\ensuremath{\mathrm{\mathbf{h}}}\xspace}$ at a chemical potential of $0.75~{\ensuremath{\rm eV}\xspace}$ above the valence band maximum and followed by a rapid sequence of transition around $2.2~{\ensuremath{\rm eV}\xspace}$ taking the system through the neutrality to electron doping states. Correcting for the reduced size of the gap in our calculations would likely space out the charge-state transitions making them more distinguishable and better defined. Vacancies at the $\mathbf{2 \times 2}$ reconstructed (001) surface of $\mathbf{SrTiO_3}$ ======================================================================================== Motivated by a discussion with experimental colleagues we seek to further increase the applicability of the conclusions attained here by studying the behavior of single oxygen vacancies at a reconstructed [[$\rm SrTiO_3$]{}]{}(001) surface, choosing as a test subject the example of tree known $2\!\times\!2$ reconstruction patterns. [@Shiraki2010; @Lin2011a] We find that including the reconstruction into consideration does not alter the conclusions of the present study – the lack of an oxygen atom inevitably leads to a excess charge localization on the neighboring Titania and results in a spontaneous magnetization of the latter. Depleting the localization by removing electron from the surface (f.e. by locally altering the chemical potential) results in a reduction of the magnetic moment of the Titania with its subsequent complete quenching. [^1]: private communication [^2]: A brief summary of a benchmark calculation of the bulk vacancy can be found in the Supplemental Material Fig. S1 [^3]: Note, however, that prevalence of low magnetizations is not universal and does not generally hold for extended or interacting oxygen vacancies in [[$\rm SrTiO_3$]{}]{}. [^4]: The “core” electrons are excluded from the summation. [^5]: The formation energy of the vacancy in a given charge state is calculated as the total energy difference between the cell with an oxygen vacancy and a sum of clean [[$\rm SrTiO_3$]{}]{}surface energy and the chemical potential of an oxygen atom in a free-standing oxygen molecule.

--- abstract: 'We examine three possible implementations of non-deterministic linear optical <span style="font-variant:small-caps;">cnot</span> gates with a view to an in-principle demonstration in the near future. To this end we consider demonstrating the gates using currently available sources such as spontaneous parametric down conversion and coherent states, and current detectors only able to distinguish between zero or many photons. The demonstration is possible in the co-incidence basis and the errors introduced by the non-optimal input states and detectors are analysed.' author: - 'A. Gilchrist' - 'W.J. Munro' - 'A.G. White' title: Input states for quantum gates --- Introduction ============ Optics is a natural candidate for implementing a variety of quantum information protocols. Photons make beguiling qubits: at optical frequencies the qubits are largely decoupled from the environment and so experience little decoherence, and single qubit gates are easily realised via passive optical elements. Some protocols, notably quantum computation, also require two-qubit gates. Until recently this was regarded as optically infeasible, since the required nonlinear interaction is much greater than that available with extant materials. However, it is now widely recognised that the necessary nonlinearity can be realised non-deterministically via measurement, and that deterministic gates can be achieved by combining such non-deterministic gates and teleportation [@01klm46]. There are a number of proposals for implementing a non-deterministic <span style="font-variant:small-caps;">cnot</span> gate with linear optics and photodetectors [@01klm46; @02ralph012314; @0111092; @ndCNOT2; @ndCNOT3; @0107091]. The proposals require deterministic, or heralded, single photon sources, and/or *selective* detectors, that can distinguish with very high efficiency between zero, one and multiple photons. Current commercial optical sources and detectors fall well short of these capabilities. Although there are a number of active research programs aimed at producing both efficient selective detectors [@02imamoglu163602; @02james183601], and deterministic photon sources [@qDOT1; @qDOT2; @diamond], nonselective avalanche photodiodes, spontaneous parametric downconversion (SPDC) and coherent states remain the best accessible laboratory options. While we could side-step the single photon source problem by using an SPDC source conditioned on the detection of a photon in one arm if we had selective detectors, demonstrating a four-photon <span style="font-variant:small-caps;">cnot</span> gate without quantum memory would be frustratingly slow. In this paper we examine three proposals which allow a <span style="font-variant:small-caps;">cnot</span> to be implemented non-destructively on the control and target modes, to ascertain under what conditions it is possible to demonstrate and characterise the gates operation using SPDC sources, coherent states and *non-selective* detectors (detectors only able to resolve zero and multiple photons). The aim is to identify a scheme that allows a scalable <span style="font-variant:small-caps;">cnot</span> implementation to be initially examined with current sources and detectors, and into which we can easily incorporate single photon sources and selective detectors as they become available. Typically the gates involve four photons with the qubit states are encoded in the polarisation state of the control and target modes $c$ and $t$, and the <span style="font-variant:small-caps;">cnot</span> operation is implemented with the aid of some ancillary modes $a$, $b$ etc. We will consider starting with the control and target modes each in a general superposition (we could also consider initially entangled states though these may be more difficult experimentally) $$\label{eq:init-fock} {\mbox{$|\psi_\mathrm{in}\rangle$}}_{ct}=(A_h{{{\mbox{\boldmath $\hat{c}$}}_{h}^{\dagger }}}+A_v{{{\mbox{\boldmath $\hat{c}$}}_{v}^{\dagger }}})(B_h{{{\mbox{\boldmath $\hat{t}$}}_{h}^{\dagger }}}+B_v{{{\mbox{\boldmath $\hat{t}$}}_{v}^{\dagger }}}){{\mbox{$|\mathbf{0}\rangle$}}}$$ with $|A_h|^2+|A_v|^2=|B_h|^2+|B_v|^2=1$, and where ${{{\mbox{\boldmath $\hat{c}$}}_{h,v}^{\dagger }}}$ and ${{{\mbox{\boldmath $\hat{t}$}}_{h,v}^{\dagger }}}$ are bosonic creation operators for mode $c_{h,v}$ and $t_{h,v}$ etc. In the interest of brevity we will use the notation above where we write the state in terms of creation operators acting on the vacuum state. The modes are first entangled with a linear optics network $U_\textsc{cnot}$ comprised of beamsplitters , phase shifters, waveplates, and polarising beam splitters. Finally the gate is conditioned on detecting the ancillary modes in some appropriate state, which leaves the state of the control and target modes as if a <span style="font-variant:small-caps;">cnot</span> had been applied. The key simplification for our purposes is to detect in the ‘coincidence basis’ — where we detect the output of the ancillary modes and also of the target and control modes and postselect out those events that do not simultaneously register a photon in all four modes. The advantage of this configuration is that now we can use *non-selective* detectors, since if we get a “click” on all four detectors we’ve accounted for all the photons in the system. This is a much less stringent requirement on the detectors and in particular can be fulfilled by existing avalanche photodiodes. We model the non-selective detectors with a positive-operator-valued measure (POVM), with the POVM elements associated with detecting no photons or photons (one or more) simply being $\Pi_0={\mbox{$|0\rangle\langle 0|$}}$ and $\Pi_m= \sum_{n=1}^\infty {\mbox{$|n\rangle\langle n|$}}$ respectively. The output state of a type-I SPDC can be described as $$\begin{aligned} {\mbox{$|\lambda\rangle$}}&=&\mathcal{M}_\lambda({\mbox{$|00\rangle$}}+\lambda{\mbox{$|11\rangle$}}+\lambda^2{\mbox{$|22\rangle$}}+\cdots)\\ &=&\mathcal{M}_\lambda \sum_{\small\begin{array}{c}n=0\\ \mathrm{(even)}\end{array}}^\infty \frac{(\lambda{{{\mbox{\boldmath $\hat{a}$}}_{}^{\dagger }}}{{{\mbox{\boldmath $\hat{b}$}}_{}^{\dagger }}})^{\frac{n}{2}}}{\frac{n}{2}!}{{\mbox{$|\mathbf{0}\rangle$}}}\end{aligned}$$ where $\mathcal{M}_\lambda=(1-\lambda^2)^{-1}$ and the sum is over even $n$ where $n$ is the number of photons in each term. Now suppose that our input state to the optical circuit is some initial pure state ${\mbox{$|\psi_\mathrm{in}\rangle$}}$, and that after passing through the linear optical elements we are left in the state ${\mbox{$|\psi_\mathrm{out}\rangle$}}=U_\textsc{cnot}{\mbox{$|\psi_\mathrm{in}\rangle$}}$. The probability that we get a count simultaneously in modes $c$, $t$, $a$ and $b$ with non-selective detectors is $$P = {\mbox{$\langle \psi_\mathrm{out}|$}}\Pi_m^{(c)}\otimes\Pi_m^{(t)}\otimes \Pi_m^{(a)}\otimes\Pi_m^{(b)}{\mbox{$|\psi_\mathrm{out}\rangle$}}$$ For the ideal case where we had single photon inputs to the gate, we will label this probability as $P_1$. We can now introduce the “single photon visibility” as a figure of merit for how close the gate operates to the ideal: $$\label{eq:visibility} \mathcal{V} = \frac{1}{2}\left(\frac{s-e}{s+e}+1\right)$$ where $s$ is the product of the probability of obtaining the single photon terms from the source, with $P_1$ the probability of the gate functioning. The “error” $e=\max (s-P)$ where $P$ is the actual probability of obtaining a count on the detectors. The maximisation is over all qubit input states to the gate. Hence if the error totally dominates the visibility is close to zero, if the noise is small the visibility is close to one. As a guide a visibility of $0.8$ corresponds to an error a quarter of the size of the single photon “signal” $s$. Simplified KLM CNOT =================== In the originally proposed non-deterministic <span style="font-variant:small-caps;">cnot</span> gate [@01klm46] the nonlinear sign shift elements were interferometric: these elements can be replaced by sequential beamsplitters to make a simplified <span style="font-variant:small-caps;">cnot</span> gate [@02ralph012314], one example of which is $$\begin{aligned} \label{eq:sKLM} {\mbox{\boldmath $\hat{U}$}}_{\textsc{sklm}}&=&{\mbox{\boldmath $\hat{B}$}}_{t_ht_v}(\frac{\pi}{4}){\mbox{\boldmath $\hat{B}$}}_{c_vt_h}(\frac{\pi}{4}) {\mbox{\boldmath $\hat{B}$}}_{bt_h}(\theta_2){\mbox{\boldmath $\hat{B}$}}_{ac_v}(\theta_2)\nonumber\\ &&{}{\mbox{\boldmath $\hat{B}$}}_{c_vt_h}(\frac{\pi}{4}){\mbox{\boldmath $\hat{B}$}}_{t_hv_2}(\theta_1){\mbox{\boldmath $\hat{B}$}}_{t_ht_v}(\frac{\pi}{4}) {\mbox{\boldmath $\hat{B}$}}_{v_1c_v}(\theta_1)\end{aligned}$$ where $ {\mbox{\boldmath $\hat{B}$}}_{ab}$ represents a beam splitter with the following action $$\begin{aligned} {\mbox{\boldmath $\hat{B}$}}_{ab}(\theta){\mbox{\boldmath $\hat{a}$}}{\mbox{\boldmath $\hat{B}$}}^\dagger_{ab}(\theta) &=& {\mbox{\boldmath $\hat{a}$}}\cos\theta+{\mbox{\boldmath $\hat{b}$}}\sin\theta\\ {\mbox{\boldmath $\hat{B}$}}_{ab}(\theta){\mbox{\boldmath $\hat{b}$}}{\mbox{\boldmath $\hat{B}$}}^\dagger_{ab}(\theta) &=& {\mbox{\boldmath $\hat{a}$}}\sin\theta -{\mbox{\boldmath $\hat{b}$}}\cos\theta\end{aligned}$$ and $\cos^2\theta$ is the reflectivity. The angle choices for the gate are given by $\theta_1=\cos^{-1}\sqrt{5-3\sqrt{2}}$ and $\theta_2=\cos^{-1}\sqrt{(3-\sqrt{2})/7}$; $c$ and $t$ are the control and target modes and $a$, $b$, $v_1$ and $v_2$ are independent ancillary modes. The gate is conditioned on detecting a single photon in the modes $a$ and $b$ and detecting no photons in the modes $v_1$ and $v_2$. Consider the case where both the control, target and ancillary photons are supplied by two independent SPDC sources. The input state is ${\mbox{$|\lambda\rangle$}}_{ct}{\mbox{$|\epsilon\rangle$}}_{ab}$ which can be written as a sum over total photon number $$\begin{aligned} \label{eq:2SPDC} {\mbox{$|\phi_\mathrm{in}\rangle$}}&=& \mathcal{M}_\lambda\mathcal{M}_\epsilon \sum_{\small \begin{array}{c}n=0\\\mathrm{(even)}\end{array}}^\infty{\mbox{\boldmath $\hat{Q}$}}_n{{\mbox{$|\mathbf{0}\rangle$}}}\\ {\mbox{\boldmath $\hat{Q}$}}_n&=& \sum_{m=0}^{\frac{n}{2}} \frac{\epsilon^m\lambda^{\frac{n}{2}-m}}{m!(\frac{n}{2}-m)!}({{{\mbox{\boldmath $\hat{a}$}}_{}^{\dagger }}}{{{\mbox{\boldmath $\hat{b}$}}_{}^{\dagger }}})^m({{{\mbox{\boldmath $\hat{c}$}}_{}^{\dagger }}}{{{\mbox{\boldmath $\hat{t}$}}_{}^{\dagger }}})^{\frac{n}{2}-m}\nonumber\end{aligned}$$ The control and target horizontal and vertical polarisation modes are then each mixed on a beamsplitter so that we achieve the input state for those modes. Since we are postselecting on getting a ‘click’ at four detectors then the terms with $n<4$ will always get postselected out. Similarly, the terms with $n>4$ will get postselected out if we used selective detectors otherwise they represent error terms. In the latter case, so long as $\epsilon,\lambda\ll1$ these terms will be small. For the case were $n=4$, three input terms contribute: $$\label{eq:in4} {\mbox{$|\psi_\mathrm{in}^{(4)}\rangle$}}=(\lambda\epsilon{{{\mbox{\boldmath $\hat{a}$}}_{}^{\dagger }}}{{{\mbox{\boldmath $\hat{b}$}}_{}^{\dagger }}}{{{\mbox{\boldmath $\hat{c}$}}_{}^{\dagger }}}{{{\mbox{\boldmath $\hat{t}$}}_{}^{\dagger }}} +\frac{\lambda^2}{2!}{{{\mbox{\boldmath $\hat{c}$}}_{}^{\dagger 2}}}{{{\mbox{\boldmath $\hat{t}$}}_{}^{\dagger 2}}}+\frac{\epsilon^2}{2!}{{{\mbox{\boldmath $\hat{a}$}}_{}^{\dagger 2}}}{{{\mbox{\boldmath $\hat{b}$}}_{}^{\dagger 2}}}){{\mbox{$|\mathbf{0}\rangle$}}}$$ While the first of these terms is equivalent to having four initial Fock states, the remaining two terms have the possibility of surviving the postselection criteria and skewing the statistics observed. Fortunately these last two terms lead to output terms which *all* get postselected out in the coincidence basis (e.g. two photons in the control mode). This means that with selective detectors we could in principle postselect out all terms that do not correspond to single photon inputs from the output statistics. With non-selective detectors the error terms will scale at least as $\lambda^3$ in amplitude (due to the $n>4$ terms) so the figure of merit will scale with $\lambda$ (taking $\epsilon=\lambda$) as $\mathcal{V}\sim1/(1+\lambda^2)$ and $\lambda$ is typically very small. Now consider the situation where a SPDC supplies the two photons for the control and target modes and weak coherent states are used for the ancillary modes. The input state is then ${\mbox{$|\phi_\mathrm{in}\rangle$}}={\mbox{$|\lambda,\alpha,\beta\rangle$}}$ where ${{{\mbox{\boldmath $\hat{a}$}}_{}^{\dagger }}}$ and ${{{\mbox{\boldmath $\hat{b}$}}_{}^{\dagger }}}$ will be the creation operators for the coherent states. After rearranging the state as a primary sum over photon number we get $${\mbox{$|\phi_\mathrm{in}\rangle$}}=\sum_{n=0}^\infty \sum_{\small\begin{array}{c}p=0\\ \mathrm{(even)}\end{array}}^n \sum_{q=0}^{n-p} \frac{(\lambda{{{\mbox{\boldmath $\hat{c}$}}_{}^{\dagger }}}{{{\mbox{\boldmath $\hat{t}$}}_{}^{\dagger }}})^{\frac{p}{2}}}{\frac{p}{2}!} \frac{(\alpha{{{\mbox{\boldmath $\hat{a}$}}_{}^{\dagger }}})^q}{q!} \frac{(\beta{{{\mbox{\boldmath $\hat{b}$}}_{}^{\dagger }}})^{n-p-q}}{(n-p-q)!}{{\mbox{$|\mathbf{0}\rangle$}}}$$ Again, terms with $n<4$ will get postselected out and terms with $n>4$ will be weak error terms. The extra freedom from two independent coherent states means that now there will be nine terms with $n=4$ and only one of these is equivalent to using single photon inputs. The terms were a single coherent state supplies all the photons always gets postselected out. By setting $\beta=i\alpha$ the two terms where a single coherent state supplies two photons and the paramp supplies two will cancel each other due to the symmetry in the circuit. Finally the term where the paramp supplies all the photons is postselected out as before. This means that we will still get errors arising from the input terms: $$\frac{i\alpha^4}{6}({{{\mbox{\boldmath $\hat{a}$}}_{}^{\dagger 3}}}{{{\mbox{\boldmath $\hat{b}$}}_{}^{\dagger }}} - {{{\mbox{\boldmath $\hat{a}$}}_{}^{\dagger }}}{{{\mbox{\boldmath $\hat{b}$}}_{}^{\dagger 3}}}){{\mbox{$|\mathbf{0}\rangle$}}}$$ Note that these do not depend on the input state that is encoded on the control and target modes and by setting $\alpha\ll\lambda$ we can scale away these terms relative the single photon terms. Unfortunately this means that we cannot beat the photon collection rate that could be achieved using two independent SPDC sources. It should be noted that all the observations made for the simplified KLM <span style="font-variant:small-caps;">cnot</span> also hold for the full KLM <span style="font-variant:small-caps;">cnot</span> in the coincidence basis. However from the perspective of an initial demonstration of the gate the simplified version is more desirable. In the following two sections we will compare these results against two other implementations of optical <span style="font-variant:small-caps;">cnot</span> gates. Entangled ancilla CNOT ====================== ![The single photon visibility with non-selective detectors as a function of the strengths of the SPDC sources. (a) the entangled ancilla gate, (b) the Knill gate. In both cases the input state was truncated at six photon terms, and the maximisation of the error was performed numerically.[]{data-label="fig:combined"}](combined_vert.eps){width=".7\columnwidth"} In a recent paper, Pittman, Jacobs and Franson [@0107091] proposed using entangled ancilla to further simplify implementation of the <span style="font-variant:small-caps;">cnot</span>. Consider that we have at our disposal an entangled state ${\mbox{$|\phi\rangle$}}=({\mbox{\boldmath $\hat{a}$}}_h{\mbox{\boldmath $\hat{b}$}}_h+{\mbox{\boldmath $\hat{a}$}}_v{\mbox{\boldmath $\hat{b}$}}_v)/\sqrt{2}{{\mbox{$|\mathbf{0}\rangle$}}}$, then we can implemented the <span style="font-variant:small-caps;">cnot</span> between modes $c$ and $t$ by first applying the unitary $${\mbox{\boldmath $\hat{U}$}}_{\mathrm{ent}} = {\mbox{\boldmath $\hat{P}$}}_{bd} {\mbox{\boldmath $\hat{P}$}}_{ae} {\mbox{\boldmath $\hat{W}$}}_a {\mbox{\boldmath $\hat{W}$}}_t {\mbox{\boldmath $\hat{W}$}}_b {\mbox{\boldmath $\hat{P}$}}_{bt} {\mbox{\boldmath $\hat{W}$}}_t {\mbox{\boldmath $\hat{W}$}}_b {\mbox{\boldmath $\hat{P}$}}_{ac}$$ where ${\mbox{\boldmath $\hat{W}$}}_a$ represents a half-wave plate on mode $a$ and ${\mbox{\boldmath $\hat{P}$}}_{ab}$ is a polarising beam splitter in modes $a$ and $b$ with the effect that $a_h\rightarrow a_h$, $b_h\rightarrow b_h$, $a_v\rightarrow b_v$, and $b_v\rightarrow a_v$. Finally the resulting state is then conditioned on detecting a single photon in modes $a$ and $b$. The raw success probability of this gate is $1/16$ which rises to $1/4$ if fast feed-forward and correction is used. Consider that the entangled pair in modes $a$ and $b$ are provided by two type-I parametric downconverting crystals sandwiched together. We’ll fix the relative phase to get a particular Bell pair for the two photon term: $$\begin{aligned} {\mbox{$|\epsilon_2\rangle$}}&=&\mathcal{M}_\epsilon^2({\mbox{$|00\rangle$}}+\epsilon{\mbox{$|11\rangle$}}+\cdots)({\mbox{$|00\rangle$}}+\epsilon{\mbox{$|11\rangle$}}+\cdots)\nonumber\\ &=&\mathcal{M}_\epsilon^2[\cdots+\epsilon({\mbox{$|0011\rangle$}}+{\mbox{$|1100\rangle$}})+\cdots]\end{aligned}$$ where the modes are $a_h$, $b_h$, $a_v$, and $b_v$ respectively. Such sources have been previously built and provide a relatively bright source of polarisation entangled photons [@99kwwae773; @99wjek3103]. We can write this source succinctly as $$\begin{aligned} {\mbox{$|\epsilon_2\rangle$}}&=&\mathcal{M}_\epsilon^2\sum_{\small \begin{array}{c}n=0\\ \mathrm{(even)}\end{array}}^\infty {\mbox{\boldmath $\hat{L}$}}_n{{\mbox{$|\mathbf{0}\rangle$}}}\\ {\mbox{\boldmath $\hat{L}$}}_n&=&\sum_{m=0}^{n/2}\frac{\epsilon^\frac{n}{2}({{{\mbox{\boldmath $\hat{a}$}}_{h}^{\dagger }}}{{{\mbox{\boldmath $\hat{b}$}}_{h}^{\dagger }}})^m ({{{\mbox{\boldmath $\hat{a}$}}_{v}^{\dagger }}}{{{\mbox{\boldmath $\hat{b}$}}_{v}^{\dagger }}})^{\frac{n}{2}-m}}{m!(\frac{n}{2}-m)!}\end{aligned}$$ With another independent paramp, ${\mbox{$|\lambda\rangle$}}$, supplying the photons for the control and target modes, the input state becomes $${\mbox{$|\phi_\mathrm{in}\rangle$}}\equiv\mathcal{M}_\epsilon^2 \mathcal{M}_\lambda \sum_{\small\begin{array}{c}n=0\\ \mathrm{(even)}\end{array}}^\infty \sum_{\small\begin{array}{c}q=0\\ \mathrm{(even)}\end{array}}^n {\mbox{\boldmath $\hat{L}$}}_q \frac{\lambda^\frac{n-q}{2}({{{\mbox{\boldmath $\hat{c}$}}_{}^{\dagger }}}{{{\mbox{\boldmath $\hat{t}$}}_{}^{\dagger }}})^{\frac{n-q}{2}}} {\frac{n-q}{2}!}$$ where we will encode the qubits in the polarisation state of the control and target modes, as in . Again all terms with $n<4$ will get postselected out. There are six terms with $n=4$ of which two terms represents our single photon input terms, the rest are error terms due to the sources. With non-selective detectors terms with $n>4$ will also contribute to the error. The four photon terms in the output state that do not get postselected out are $$\begin{aligned} {\mbox{$|out\rangle$}}&=&\frac{1}{2\sqrt{2}}\lambda {{{\mbox{\boldmath $\hat{a}$}}_{}^{\dagger }}}{{{\mbox{\boldmath $\hat{b}$}}_{}^{\dagger }}}( A_vB_h \epsilon {{{\mbox{\boldmath $\hat{c}$}}_{v}^{\dagger }}}{{{\mbox{\boldmath $\hat{t}$}}_{v}^{\dagger }}} + A_vB_v \epsilon{{{\mbox{\boldmath $\hat{c}$}}_{v}^{\dagger }}}{{{\mbox{\boldmath $\hat{t}$}}_{h}^{\dagger }}}\nonumber\\ &&{}+A_h [A_vB_h^2\lambda-A_vB_v^2\lambda+B_v\epsilon]{{{\mbox{\boldmath $\hat{c}$}}_{h}^{\dagger }}}{{{\mbox{\boldmath $\hat{t}$}}_{v}^{\dagger }}}\nonumber\\ &&{}+A_h [A_vB_h^2\lambda-A_vB_v^2\lambda+B_h\epsilon]{{{\mbox{\boldmath $\hat{c}$}}_{h}^{\dagger }}}{{{\mbox{\boldmath $\hat{t}$}}_{h}^{\dagger }}} ){{\mbox{$|\mathbf{0}\rangle$}}}\end{aligned}$$ and by making $\lambda\ll\epsilon$ we can recover the single photon terms and the action of the <span style="font-variant:small-caps;">cnot</span> with selective detectors. This of course means that the count rate with this gate would be considerably less than with the simplified KLM gate. With non-selective detectors, if we make $\lambda$ too small the error due to the six photon input terms will dominate, so there is an optimum $\lambda$ for a given $\epsilon$ see figure \[fig:combined\] (a). There does not appear to be a way of using two coherent states to replace one of the SPDC sources. If we replace either the control or target mode then it is hard to see how the ${\mbox{$|02\rangle$}}$ and ${\mbox{$|20\rangle$}}$ terms could cancel as with the simplified KLM <span style="font-variant:small-caps;">cnot</span> since these terms will have factors that depend on the encoded qubit. Similarly replacing the source of entangled photons would then mean we would have to entangle the single photon components which is difficult. Knill CNOT ========== A recent numerical search for optical gates by Knill yielded a <span style="font-variant:small-caps;">cnot</span> gate [@0110144] which operates with a probability of $2/27$ and is described by the following unitary, $$\begin{aligned} {\mbox{\boldmath $\hat{U}$}}_{\mathrm{Knill}} &=& {\mbox{\boldmath $\hat{B}$}}_{t_vt_h}(\frac{\pi}{4}){\mbox{\boldmath $\hat{B}$}}_{ab}(\theta_3){\mbox{\boldmath $\hat{B}$}}_{c_vt_v}(\theta_2) {\mbox{\boldmath $\hat{B}$}}_{t_vb}(\theta_1)\nonumber\\ &&{}{\mbox{\boldmath $\hat{B}$}}_{c_va}(\theta_1){\mbox{\boldmath $\hat{B}$}}_{t_vt_h}(\frac{\pi}{4}){\mbox{\boldmath $\hat{F}$}}_a(\pi)\end{aligned}$$ where ${\mbox{\boldmath $\hat{F}$}}_a(\theta)$ is a phaseshift of $\theta$ on mode $a$ and the reflectivities are given by $\theta_1=\cos^{-1}\sqrt{1/3}$, $\theta_2=-\theta_1$ and $\theta_3=\cos^{-1}\sqrt{1/2+1/\sqrt{6}}$. The gate requires two ancillary modes $a$ and $b$ initially in Fock states to be finally detected also in single Fock states. Consider the case where both the control, target and ancillary photons are supplied by two independent SPDC sources. The input state is given by with the usual qubit encoding as in equation . We will again get the three terms possibly contributing to the error for $n=4$. The last term again leads to output terms which all get postselected out in the coincidence basis. Unfortunately the output terms produced by the second term do not get postselected out leading to inherent errors in the statistics we will observe. Notice however that all these terms will be proportional to $\lambda^2$ so again by making $\lambda\ll\epsilon$ we can scale these terms away with selective detectors at the expense of the count rate. With non-selective detectors there will again be an optimum $\lambda$, see figure \[fig:combined\] (b), which is very similar to the previous gate. Conclusion ========== We have examined three possible implementations for linear optics <span style="font-variant:small-caps;">cnot</span> gates with a view to experimentally demonstrating their operation in the near future. In considering demonstrating the gates with SPDC and coherent state sources and non-selective detectors there is a clear advantage to the simplified KLM <span style="font-variant:small-caps;">cnot</span> gate, where the inherent symmetries in the gate allow the use of two independent SPDC sources to supply the control, target and ancillary photons, with errors from the use of non-Fock states making little contribution. The other two implementations suffer from errors introduced by the non-Fock state inputs which cannot be postselected out. While the situation may be mitigated somewhat by using a weak SPDC source this would occur at the expense of the count rate of valid events that may be collected from the gate. The conclusion we arrive at is that an experimental program focusing on the simplified KLM <span style="font-variant:small-caps;">cnot</span> gate would then allow immediate characterisation of the gate with current sources and detectors, with the operation of the gate in a non-destructive fashion becoming possible when single photon sources and selective detectors become available. We would like to acknowledge support from the the Australian Research Council and the US Army Research Office. AG was supported by the New Zealand Foundation for Research, Science and Technology under grant UQSL0001. WJM acknowledges support for the EU project RAMBOQ. We would also like to thank Michael Nielsen, Jennifer Dodd, Nathan Langford, Tim Ralph and Gerard Milburn for helpful discussions. [10]{} , [R. Laflamme]{}, and [G. Milburn]{}, Nature [**409**]{}, 46 (2001). T. C. Ralph, A. G. White, W. J. Munro, and G. J. Milburn, Phys. Rev. A [ **65**]{}, 012314 (2002). H. F. Hofmann and S. Takeuchi, Phys. Rev. A [**66**]{}, 024308 (2002). T. C. Ralph, N. K. Langford, T. B. Bell, and A. G. White, Phys. Rev. A [ **65**]{}, 062324 (2002). K. Sanaka, K. Kawahara, and T. Kuga, quant-ph/0108001, 2001. T. Pittman, B. C. Jacobs, and J. D. Franson, Phys. Rev. A [**64**]{}, 062311 (2001). A. Imamoglu, Phys. Rev. Lett. [**89**]{}, 163602 (2002). D. F. James and P. G. Kwiat, Phys. Rev. Lett. [**89**]{}, 183601 (2002). P. Michler [*et al.*]{}, Science [**290**]{}, 2282 (2000). M. Pelton [*et al.*]{}, quant-ph/0208054, 2002. A. Beveratos [*et al.*]{}, Eur. Phys. D [**18**]{}, 191 (2002). P. G. Kwiat [*et al.*]{}, Phys. Rev. A [**60**]{}, R773 (1999). A. G. White, D. F. V. James, P. H. Eberhard, and P. G. Kwiat, Phys. Rev. Lett. [**83**]{}, 3103 (1999). E. Knill, quant-ph/0110144, 2001.

\#1\#2\#3 \#1 \#2 \#3 0.8cm [ **A note on string solutions in $AdS_{3}$**]{} 4ex 1.0cm [ Kazuhiro  Sakai, ]{}\ -1ex [*Department of Physics, Keio University*]{} -2ex [*Hiyoshi, Yokohama 223-8521, Japan*]{}\ 2ex [ Yuji  Satoh]{}\ -1ex [*Institute of Physics, University of Tsukuba*]{}\ -2ex [*Tsukuba, Ibaraki 305-8571, Japan*]{} 10ex =3.5ex [**Abstract**]{}  We systematically search for classical open string solutions in $AdS_{3}$ within the general class expressed by elliptic functions (i.e., the genus-one finite-gap solutions). By explicitly solving the reality and Virasoro conditions, we give a classification of the allowed solutions. When the elliptic modulus degenerates, we find a class of solutions with six null boundaries, among which two pairs are collinear. By adding the $S^{1}$ sector, we also find four-cusp solutions with null boundaries expressed by the elliptic functions. July 2009 = 3.3ex Classical open string solutions in the anti-de Sitter (AdS) space with null boundaries give the scattering amplitudes in planar $\calN =4$ super Yang-Mills theory at strong coupling [@Alday:2007hr] (for a review, see for example, [@Alday:2008yw]). Because of this, the problem of finding such solutions have attracted much attention [@Ryang:2007bc]-[@Dorn:2009kq]. Though the solution with four null boundaries and cusps is found in [@Kruczenski:2002fb; @Alday:2007hr], finding the solutions with more than four cusps is still challenging. Recently, a prescription to construct multi-cusp solutions is provided in [@AM2]. There, it is also discussed how to compute the scattering amplitudes without using explicit form of the solutions, and this is demonstrated in the case of the eight-cusp solutions. Regarding the numerical multi-cusp solutions, see [@Dobashi:2008ia]. With applications to the scattering amplitudes in mind, we discuss the classical open string solutions in $AdS_3$. For this purpose, a good starting point would be a general construction of the classical string solutions in $dS_{2n+1}$ [@Krichever], where the solutions are expressed by theta functions and integrals defined over the underlying spectral curve. This construction can also be applied to $AdS_{2n+1}$.[^1] However, for constructing relatively simple solutions, it may be easier to make an ansatz of the finite-gap form (i.e., the general form implied by [@Krichever]), where one regards the periods and the integrals as free paramters, and search for particular solutions which satisfy definite reality, Virasoro and boundary conditions. In this paper, we take this approach for the genus-one finite-gap solutions (elliptic solutions). We determine the parameters of the solutions by explicitly solving the equations of motion, and the reality and Virasoro conditions. As a result, we give a classification of the allowed genus-one finite-gap solutions. When the elliptic modulus degenerates, we also find a class of solutions with six null boundaries, among which two pairs are collinear. The solutions are expressed simply by hyperbolic and exponential functions, and describe non-flat minimal surfaces in $AdS_{3}$. The analysis can be generalized to the classical string solutions in $AdS_{5} \times S^{5}$. By adding $S^{1}$, as a simple example, we find four-cusp solutions with null boundaries expressed by elliptic functions. The rest of this paper is organized as follows. In section 2, starting with the genus-one finite-gap form, we solve the equations of motion and the normalization condition. We then summarize the Virasoro condition and the reality condition. By solving these conditions, we determine the allowed solutions and give a classification in section 3. In section 4, we discuss examples of the solutions. In particular, we present a class of solutions with six null boundaries. In section 5, we analyze the case of the strings in $AdS_{3} \times S^{1}$, and find four-cusp solutions expressed by the elliptic functions. We conclude with a discussion in section 6. The appendix includes our conventions and some formulas of the elliptic theta functions. We begin with parametrizing the $AdS_3$ target space by the embedding coordinates in $R^{2,2}$, namely, $Y_a(\sigma_+,\sigma_-)$, $a=-1,0,1,2$, with a constraint := -Y\_[-1]{}\^2-Y\_0\^2+Y\_1\^2+Y\_2\^2=-1 \[Ynorm\] They satisfy the equations of motion \_+\_- -(\_+\_-)=0 \[Yeom\] and the Virasoro constraints (\_)\^2=0 \[YVirasoro\] The solutions span minimal surfaces in $AdS_3$. In the following, we concentrate on the Euclidean world-sheet with $(\sigma_{+})^{*} = \sigma_{-}$. The case of the Lorentzian world-sheet can be discussed similarly. To find the solutions, we introduce the vector $\vec{\varphi} = (\varphi_{1}, \phiinv_{1}, \varphi_{2}, \phiinv_{2})$ which satisfies 1 [&=&]{} \_[j=1]{}\^2\_j\^\_j\[varphinorm\]\ 0 [&=&]{}(\_[+]{} \_[-]{} + u) \[varphieom\]\ 0 [&=&]{} \_[j=1]{}\^2\_ \_j\_\^\_j \[varphiVirasoro\] with the self-consistent potential u= \_[j=1]{}\^2( \_[+]{} \_j\_[-]{} \^\_j +\_[-]{}\_j\_[+]{}\^\_j ). \[potu\] The equations (\[varphinorm\])-(\[varphiVirasoro\]) are equivalent to (\[Ynorm\])-(\[YVirasoro\]) under the identification $\varphi = Y$, where := ( [cc]{} \_[1]{} & \_[2]{}\ -\_[2]{} & \_[1]{} ) Y := ( [cc]{} Y\_[-1]{}+Y\_[2]{} & Y\_[1]{} + Y\_[0]{}\ Y\_[1]{} - Y\_[0]{} & Y\_[-1]{} -Y\_[2]{} ) As discussed shortly, more general identifications between $\varphi$ and $Y$ are possible. In the genus-one case, the finite-gap solution to (\[varphinorm\])-(\[varphiVirasoro\]) takes the form [@Krichever] \_j[&=&]{}r\_je\^[p\_j\^+\_+ +p\_j\^-\_-]{},\ \^\_j[&=&]{} \_[j]{} e\^[-(p\_j\^+\_+ +p\_j\^-\_-)]{} Here, X [&=&]{} U\^+\_+ + U\^-\_- + X\_0 - K(k) \[X0X\] and $K(k)$ is the complete elliptic integral of the first kind with $k$ the elliptic modulus. ${\vartheta}_{a}(z)$ are the elliptic theta functions which have the quasi-periods $(2K(k),2iK'(k))$ with $K'(k) = K(k')$ and $(k')^{2} = 1-k^{2}$. Compared with the standard notation, we have rescaled the argument of the theta functions by $2K$. With this convention, for example, ${\vartheta}_{0}(z+K) = {\vartheta}_{3}(z)$ and $\sn\,z={\vartheta}_3(0){\vartheta}_1(z)/{\vartheta}_2(0){\vartheta}_0(z)$. To make the following expressions simpler, we have shifted $X$ by $K$ as in (\[X0X\]), which results in the combination of ${\vartheta}_3$ and ${\vartheta}_0$ in ${\varphi}$. Our conventions of the elliptic theta functions are summarized in the appendix. Other parameters should be determined by imposing appropriate conditions. First, one finds that the normalization condition (\[varphinorm\]) gives r\_1 \_[1]{}=, r\_2 \_[2]{}= \[rr\] for $A_{1} \neq A_{2}$. The case of $A_1 = A_2 $ is discussed later in section 3.3.[^2] In deriving this, we have used \_j\_j\^= r\_j \_j = r\_j \_j \[phiphi\] which follow from product identities of ${\vartheta}_a$. Next, to consider the equations of motion, we introduce \_j\^:=Z(A\_j)+, \[betadef\] where $Z(z):=\partial_z\ln{\vartheta}_0(z)$. With the help of the formula (\[addZ\]), one then obtains [&=&]{}U\^+U\^- and similar equations for $\phiinv_{j}$ with $A_{j}, p^{\pm}_{j}$ replaced by $-A_{j}, -p^{\pm}_{j}$. For these to be equated with $-u$, the $X$-dependence should be common to all $\varphi_{j},\phiinv_{j}$. This requirement fixes $\beta_j^\pm$ as \_j\^+ + \_j\^- [&=&]{} -,\ \_j\^+ \_j\^- [&=&]{} k\^2\^2A\_j+u\_0, \[betaeom\] where $u_{0}$ is a constant. Substituting these, one obtains =U\^+U\^-(2k\^2\^2X+u\_0). \[-u\] On the other hand, the potential $u$ in (\[potu\]) is evaluated using the equations for $\varphi_j \phiinv_j $ in (\[phiphi\]). Some computations show that $-u$ is indeed given by the right-hand side of (\[-u\]), which verifies the equations of motion. Third, let us turn to the Virasoro condition. Again, after some algebra, one finds that the constraint \_[j=1]{}\^2 ( \_+\_j\_+\^\_j + \_-\_j\_-\^\_j ) =0 determines the constant $u_0=-u(X = 0)/U^{+}U^{-}$ to be u\_0=2(+-1-k\^2) whereas the other constraint reads 0[&=&]{} \_[j=1]{}\^2 ( \_+\_j\_+\^\_j - \_-\_j\_-\^\_j )\ [&=&]{} 2U\^+U\^- \_[j=1]{}\^2(-)\^[j+1]{} (\^[+]{}\_[j]{} - \^[-]{}\_[j]{}) In terms of $a := \sn^2A_1,b := \sn^2A_2$, this is equivalent to (a+b-ab)(a+b-k\^2ab)(a+b-(1+k\^2)ab)=0 the solutions of which are  A\_1A\_2=1, A\_1A\_2=1, [cd]{}A\_1[cd]{}A\_2=1. \[Virasoro\] In each of these three cases, one finds that  u\_0=-2k\^2, u\_0=-2, u\_0=0, \[u0\] and  u=2k\^2U\^+U\^-\^2 X, u=2U\^+U\^-\^2 X, u=-2k\^2U\^+U\^-\^2 X. The final condition to be imposed is the reality condition, for which we need to know the allowed identifications between $ \varphi $ and $Y$. In order to analyze these, we note that, from $\det \varphi = \det Y = 1$ and the equations of motion, the two matrices should be related by constant $ SL(2,{{\mathbb C}})$ matrices $U,V $ as $U \varphi V = Y$. This implies that $Y^{-1} dY = V^{-1} \varphi^{-1} d\varphi V$, and that the tangent spaces of $Y$ and $\varphi$ are isomorphic. Since $Y$ is an $SL(2,{{\mathbb R}})$ matrix, $\varphi$ should generically be an element of $SL(2,{{\mathbb R}})$ or $SU(1,1)$. Therefore, there are two cases for the reality condition: && ([I]{})  \_[j]{}\^\* = \_[j]{}   (\_[j]{})\^\* = \_[j]{}   [for ]{} SL(2,[[R]{}]{})\ && ([II]{}) \_[1]{}\^\* = \_[1]{} \_[2]{}\^\* = - \_[2]{} SU(1,1) \[realcond\] (up to the exchange of $\varphi_{1}, \phiinv_{1}$ and $\varphi_{2}, \phiinv_{2}$). In each case, the AdS solution $Y$ is identified with $\varphi$ as ([I]{}) = Y ([II]{}) = M\^[-1]{} Y M = ( [cc]{}Y\_[-1]{} + iY\_[0]{} & Y\_[1]{}+iY\_[2]{}\ Y\_[1]{}-iY\_[2]{} & Y\_[-1]{}-iY\_[0]{} ) up to $SL(2,{{\mathbb R}})$ and $SU(1,1)$ transformations, respectively, where $ M = \frac{1}{\sqrt{2}} \biggl(\begin{array}{cc} 1& i \\i & 1\end{array}\biggr)$. In both cases, the potential takes the form $u = -\del_{+}\vec{Y} \cdot \del_{-}\vec{Y}$, from which one can read off the conformal factor of the induced metric and hence the curvature of the surface described by the solution. In the following, we set $q = e^{\pi i \tau}$ ($ \tau = iK'/K$) in the theta functions to be real, which implies $ 0 \leq k^{2} \leq 1$. When $q$ is complex, the reality conditions may not be satisfied. In this section, we solve the reality and Virasoro conditions which are listed in the previous section. First, we concentrate on the case where $\varphi \in SL(2,{{\mathbb R}})$. The other case with $\varphi \in SU(1,1)$ is discussed later. When $\varphi \in SL(2,{{\mathbb R}})$, the reality condition is (I) in (\[realcond\]). For this to be satisfied for arbitrary $\sigma_{\pm}$, the theta functions ${\vartheta}_{0}(X), {\vartheta}_{0}(X+A_{j}) $ should be real or purely imaginary (after extracting the exponential factors due to possible shifts in $(X,A_{j})$ by $iK')$. This implies that $(X, A_{j})$ are real or purely imaginary up to the shifts $n K + i m K'$, under which ${\vartheta}_{a}$ transform as ${\vartheta}_{a}(u + n K + i m K') = $ (factor)$\times {\vartheta}_{b}(u)$ according to (\[thetashift\]). Furthermore, since ${\vartheta}_{a}(u+2K) = \pm {\vartheta}_{a}(u)$, ${\vartheta}_{a}(u +2 i K') = \pm e^{-\pi i (\frac{u}{K}+ \tau)} {\vartheta}_{a}(u)$, we have only to consider $0, K, iK', K + iK'$ as the shifts: other cases reduce to these cases by absorbing the factors into $r_{j}$ and $p^{\pm}_{j}$. Consequently, it is enough to assume that $A_{j}$ are in the fundamental region spanned by $(0,2K,2iK', 2K+2iK')$ with segments $[2K, 2K+2iK'], [2iK', 2K + 2iK']$ removed. Therefore, we have four cases of $(X,A_j)$: && (1) X [[R]{}]{}   [or]{}    [[R]{}]{}+ iK’ A\_j=a\_j  (a\_j [[R]{}]{})\ && (2) X [[R]{}]{}   [or]{}    [[R]{}]{}+ iK’ A\_j=a\_j +iK’  (a\_j [[R]{}]{})\ && (3) X i[[R]{}]{}   [or]{}    i[[R]{}]{}+ K A\_j= ia\_j  (a\_j [[R]{}]{}) \[XA\]\ && (4) X i[[R]{}]{}   [or]{}    i[[R]{}]{}+ K A\_j=ia\_j +K  (a\_j [[R]{}]{}) In addition, after the possible shifts of $iK'$ in $(X,A_j)$ are taken into account, real solutions for $\varphi \in SL(2,{{\mathbb R}})$ must be transformed into the canonical form where $r_j,\rinv_j$ and exponentials in $\varphi$ are real. These impose restrictions on $\beta^\pm_{j}$. Let us discuss these conditions in more detail, e.g., in case (2). In this case, $U^+ \sigma_+ + U^- \sigma_-$ is real, which implies $(U^{+})^{*} = U^{-}$ and $X_{0} -K \in {{\mathbb R}}$ or ${{\mathbb R}}+ iK'$. When $X_0 -K \in {{\mathbb R}}+ iK'$, the shift of $iK'$ results in a constant factor to the ratio of the theta functions, which we absorb into $r_j, \rinv_j$. As for the shift $iK'$ in $A_j$, extracting it from ${\vartheta}_0$ gives &&\_j \~ e\^[q\^[+]{}\_j\_+ + q\^[-]{}\_j \_-]{}  q\^\_j = p\^\_j - U\^\ && Z(a\_j + iK’) = - + Z\_[1]{}(a\_j) Z\_1(z) :=\_z \_1(z) and similarly for $\phiinv_j$ with the signs of $a_j, q_j^\pm$ flipped. The exponent after the shift should be real and thus $(q_j^{+})^{*} = q_j^{-}$. Note that $(q_j^{+}/U^{+})^{*} = q_j^{-}/U^{-}$, $Z_{1}(a_j) \in {{\mathbb R}}$, and the conditions from the equations of motion (\[betaeom\]) read \_j\^[+]{} + \_j\^[-]{} = 2 [[R]{}]{} \_j\^[+]{} \_j\^[-]{} = \^[2]{}a\_j + u\_[0]{} [[R]{}]{}. \[betapm1\] On the other hand, from the definition of $\beta_j^\pm$, (\[betadef\]), it follows that \_j\^ = Z\_[1]{}(a\_j) + \[betapm2\] and hence $(\beta_j^+)^*= \beta_j^-$. For given $k, a_j,U^\pm$, the real part of $\beta_j^{\pm}$ (or $p_j^{\pm}/U^{\pm}$) is determined by the first equation in (\[betapm1\]), whereas the imaginary part is consistently determined by the second, if \_[\_j]{}:=(\_j\^[+]{} - \_j\^[-]{})\^[2]{} = -(1+k\^[2]{} + u\_[0]{}) 0 with $1/\sn^2 A_j = k^2 \sn^2 a_j$. Similarly analyzing other cases, we find that the reality condition imposes \_[\_j]{} 0 \_[\_j]{} 0 Applying the value of $u_{0}$ in (\[u0\]), these are solved in each case, which imposes the following conditions: [cccccc]{} $A_j = a_j \ (a_{j} \in {{\mathbb R}})$\ (i) $k' = \sn^{2}a_j = 1$;  (ii) (no solutions);  (iii) $\sn^{2}a_j \geq \frac{1}{1+k^{2}}$. $A_j = a_j +iK'\ (a_j \in {{\mathbb R}})$\ (i) $\sn^{2} a_j \leq \frac{(k')^{2}}{k^{2}}$;  (ii) (no solutions);  (iii) (automatic). $A_j = ia_j\ (a_j \in {{\mathbb R}})$\ (i) (no solutions);  (ii) $\sn^{2}(a_j,k') \geq \frac{1}{1+(k')^{2}}$;  (iii) (no solutions). $A_j = ia_j +K \ (a_j \in {{\mathbb R}})$\ (i) $\sn^{2}(a_j,k') \leq \frac{k^{2}}{(k')^{2}}$;  (ii) (automatic) ;  (iii) $ k= a_j =0$. In the table, “no solutions” indicates the cases where the solutions do not exit, whereas “automatic" indicates the cases where the reality condition is automatically satisfied and imposes no restrictions. The cases where $\beta_j^\pm$ are diverging have also been excluded. We have also omitted the values of $X$ in the above. We remark that, when considering both $\varphi_1,\phiinv_1$ and $\varphi_2,\phiinv_2$, $X$ is common and only the combinations among cases (1) and (2), or (3) and (4) are allowed. From the discussion in the previous section, we find that there are six cases of the combinations of $(A_{1},A_{2})$. In each combination, there are three possibilities of satisfying the Virasoro condition as in (\[Virasoro\]). It is straightforward to write down the explicit from of the condition in each case and check whether it has solutions or not. For example, when $A_{1}= ia_{1}, A_{2} = ia_{2} $ ($a_{1}, a_{2} \in {{\mathbb R}}$), the condition of case (i) in (\[Virasoro\]) reads $ \nc(a_{1},k') \, \nc(a_{2},k') = \pm 1$. Since $\nc^{2} u \geq 1$ for real $u$, the condition is satisfied only when $a_{1} = a_{2} = 0$. When $A_{1,2} = a_{1,2} + iK'$ ($a_{1,2} \in {{\mathbb R}}$), the condition of case (i) in (\[Virasoro\]) reads $-k^{-2}\ds\, a_{1} \, \ds\, a_{2} =\pm 1$. Since $\ds^{2} u \geq (k')^{2}$ for real $u$, the condition has solutions when $ 1/2 \leq k^{2}$. Repeating similar analysis for all cases, one finds that the Virasoro constraints impose the following conditions: [cccccc]{} $A_{1} = a_{1}, \ A_{2} = a_{2}$  $(a_{j} \in {{\mathbb R}})$\ (i) $ a_{1,2} = 0 $;  (ii) $a_{1,2} = 0 $ or $k=0$;  (iii) $ a_{1,2} = 0 $ or $k=1$. $A_{1} = a_{1} + iK', \ A_{2} = a_{2} + iK' $  $(a_{j} \in {{\mathbb R}})$\ (i) $k^2 \sd\, a_{1} \, \sd\, a_{2} =\pm 1$ and $ \half \leq k^{2}$;  (ii) $\sc\, a_{1} \, \sc \, a_{2} = \pm 1$;  (iii) $k=1$. $A_{1} = a_{1}, \ A_{2} = a_{2} + iK' $  $(a_{j} \in {{\mathbb R}})$\ (i) (no solutions);  (ii) (no solutions);   (iii) $ k \, \dc \, a_{1} \, \cd \, a_{2} = \pm 1$. $A_{1} = ia_{1}, \ A_{2} = ia_{2} $  $(a_{j} \in {{\mathbb R}})$\ (i) $a_{1,2} = 0$;  (ii) $ a_{1,2} = 0 $ or $k=0$;  (iii) $ a_{1,2} = 0 $ or $k=1$. $A_{1} = ia_{1}+K, \ A_{2} = ia_{2}+K$  $(a_{j} \in {{\mathbb R}})$\ (i) $ (k')^{2}\sd(a_{1},k') \, \sd(a_{2},k') =\pm 1$ and $ \half \geq k^{2}$;  (ii) $k=0$;   (iii) $\sc(a_{1},k') \, \sc(a_{2},k') = \pm 1$. $A_{1} = ia_{1}, \ A_{2} = ia_{2}+K$  $(a_{j} \in {{\mathbb R}})$\ (i) (no solutions);  (ii) $ k' \, \dc(a_{1},k') \, \cd(a_{2},k') = \pm 1$;  (iii) (no solutions). Combining the tables in the previous two subsections, we can determine the allowed cases and their conditions: [cccccc]{} $A_{1} = a_{1}, \ A_{2} = a_{2}$  $(a_{j} \in {{\mathbb R}})$\ (i) (no solutions);  (ii) (no solutions);  (iii) $k=1$ and $\sn^{2}a_{1,2} \geq \half$. $A_{1} = a_{1} + iK', \ A_{2} = a_{2} + iK' $  $(a_{j} \in {{\mathbb R}})$\ (i) $k^{2}= \half$ and $\sn^{2}a_{1,2} = 1$;  (ii) (no solutions);   (iii) $k=1$. $A_{1} = a_{1}, \ A_{2} = a_{2} + iK' $  $(a_{j} \in {{\mathbb R}})$\ (i) (no solutions);  (ii) (no solutions);  (iii) $ k \, \dc \, a_{1} \, \cd \, a_{2} = \pm 1$ and $\sn^{2}a_{1} \geq \frac{1}{1+k^{2}}$. $A_{1} = ia_{1}, \ A_{2} = ia_{2} $  $(a_{j} \in {{\mathbb R}})$\ (i) (no solutions);  (ii) $k=0$ and $\sn^{2}(a_{1,2},k') \geq \half$;   (iii) (no solutions). $A_{1} = ia_{1}+K, \ A_{2} = ia_{2}+K$  $(a_{j} \in {{\mathbb R}})$\ (i) $k^{2} = \half$ and $\sn^{2}(a_{1,2},k') = 1$;  (ii) $k=0$;  (iii) (no solutions). $A_{1} = ia_{1}, \ A_{2} = ia_{2}+K$  $(a_{j} \in {{\mathbb R}})$\ (i) (no solutions);   (ii) $ k' \, \dc(a_{1},k') \, \cd(a_{2},k') = \pm 1$ and $\sn^{2}(a_{1},k') \geq \frac{1}{1+(k')^{2}}$;\ (iii) (no solutions). We note that the result is symmetric between the real and the imaginary $A_j$. This is a consequence of the modular transformation $\tau \to -1/\tau$ with purely imaginary $\tau$. In fact, one can check that $\varphi$ in the first three cases are mapped to the last three, up to certain factors which can be absorbed into the exponential factors and the normalization constants of $\varphi$. Some asymmetries in the intermediate stage of the analysis are due to having started with the fixed exponents in $\varphi$. From this result, one finds that the allowed solutions fall into three types. One is the solution with $k \neq 0,1$ and $A_1 \neq A_2$ as in 1-2 (iii) and 3-4 (ii). Such solutions are expressed by the elliptic functions, as we initially intended. We call this type of solutions “elliptic solution”. The second one is the solution with $k=0$ or $1$ and $A_1 \neq A_2 $. In this case, the elliptic functions degenerate and the solutions become simpler. One has to be a little careful in taking $k=0,1$, since $K',K$ are singular, respectively. For $k=0$ as in 3-3 (ii) and 4-4 (ii), $q = e^{i \pi \tau}$ is vanishing and ${\vartheta}_0(X)$ reduces to a constant for finite $X$. However, if we take $k\to 0$ after shifting $X$, which is imaginary in these cases, by $iK'(\to i\infty)$, the ratio of ${\vartheta}_0$’s becomes a ratio of the hyperbolic functions. For $k=1$ with $q \to 1$ as in 1-1 (iii) and 2-2 (iii), by making use of the modular transformation $\tau \to -1/\tau $, one finds that for finite $X$ the ratio of ${\vartheta}_0$’s becmes a ratio of the hyperbolic functions. However, if we take $k\to 1$ after shifting $X$ by $K (\to \infty)$, which is allowed in these cases, the ratio of ${\vartheta}_0$’s becomes a constant. Thus, depending on the way to take the limit, the degenerate solutions reduce to (a) the known solutions with $\varphi_j,\phiinv_j \sim$ const.$\times $(exponentials), or (b) the solutions with $\varphi_j,\phiinv_j \sim$ (ratio of hyperbolic functions)$\times $(exponentials). In the latter case, the potential $u$ is not constant, and the minimal surface spanned by $Y$ is not flat. We call the former type “exponential solution”, and the latter “hyperbolic solution”. The third type is the solution with $A_1 = A_2$, which we have not considered so far, since the normalization condition (\[rr\]) becomes singular. In this case, (\[phiphi\]) implies that the only possibilities to satisfy the normalization condition of $\varphi$ is $k=0,1$ or $A_{1,2} =0$, since the $X$-dependence has to be canceled. Thus, cases 2-2 (i), 4-4 (i) are excluded, though we left them in the table taking into account a possibility that they could be regarded as limiting cases. When $A_{1,2}= 0$, the solutions reduce to the exponential type. When $k=0,1$, as discussed above, the solutions become of the exponential or the hyperbolic/trigonometric type. In the latter case, it turns out that one has to set $A_1 = A_2 =0$ to satisfy the normalization condition. In sum, if $A_1=A_2$, only the solutions of the exponential type are allowed. For $k=0,1$ or $A_{1} = A_{2}$, one may take appropriate limits from the generic cases to write down the solutions. However, it is more straightforward to start with the generic form of the solutions in these cases, and determine them as in section 2. So far, we have considered the case where $\varphi \in SL(2,{{\mathbb R}})$. As discussed in section 2, there is another case with $\varphi \in SU(1,1)$. The reality and Virasoro conditions are analyzed similarly. First, from the reality condition (II) in (\[realcond\]), one finds that there are four cases of $(X,A_j)$: && (1’) X i [[R]{}]{}   [or]{}    i [[R]{}]{}+ K A\_j=a\_j  (a\_j [[R]{}]{})\ && (2’) X i [[R]{}]{}   [or]{}    i [[R]{}]{}+ K A\_j=a\_j +iK’  (a\_j [[R]{}]{})\ && (3’) X [[R]{}]{}   [or]{}    [[R]{}]{}+ iK’ A\_j= ia\_j  (a\_j [[R]{}]{})\ && (4’) X [[R]{}]{}   [or]{}    [[R]{}]{}+ iK’ A\_j=ia\_j +K  (a\_j [[R]{}]{}) In addition, the normalization constants should satisfy $ r_{j}^{*} = \pm \rinv_{j}$ and the exponentials in $\varphi_{j}$ and $\phiinv_{j}$ should be complex conjugate to each other (after the possible shifts of $iK'$ in $(X,A_{j})$). In any of these cases, the combinations of $p_j^{\pm}$ and $U^{\pm}$ satisfy the same relations as the corresponding ones in the $SL(2,{{\mathbb R}})$ case. For example, in case $(1')$, $(p_j^{+}/U^{+})^{*} = p_j^{-}/U^{-}$, though $p_j^\pm, U^\pm$ have different relations $(p_j^{+})^{*} = -p_j^{-}$ and $(U^{+})^{*} = -U^{-}$. Thus, the constraints from the reality condition are the same. The Virasoro condition is irrelevant of which embedding we use, $SL(2,{{\mathbb R}})$ or $SU(1,1)$. Therefore, the allowed cases are read off from the same table as in the $SL(2,{{\mathbb R}})$ case in section 3.3. In the $SU(1,1)$ case, the condition on $r_{j},\rinv_{j}$ implies r\_[1]{}\_[1]{} r\_[2]{} \_[2]{} 0 This may give further constraints on the parameters, e.g., on $X_{0}$. When $r_{1}\rinv_{1} < 0$ and $r_{2}\rinv_{2} > 0$, we need to exchange $\varphi_{1}, \phiinv_{1}$ and $\varphi_{2},\phiinv_{2}$. For the Euclidean world-sheet, which results in space-like surfaces, one finds no solutions eventually. Our main motivation to studying the AdS string solutions is the application to the scattering amplitudes in the super Yang-Mills theory. With this in mind, we discuss the obtained solutions. Before going into details, let us summarize some general properties of the solutions in relation to the cusp solutions with null boundaries. First, when $\varphi \in SU(1,1)$, the exponential part of $\varphi$ is complex and, since, e.g., $Y_{-1} =$ Re$\, \varphi_1$, the solutions are generally rapidly oscillating near the world-sheet boundary $|\sigma_\pm| \gg 1 $. Thus, to search for the cusp solutions, one should look into the case with $\varphi \in SL(2,{{\mathbb R}})$ (unless the world-sheet is consistently restricted). This case also includes oscillating solutions. For example, in case (2) with real $X$ in (\[XA\]), the solution has a factor ${\vartheta}_2(X+a)$ and this is oscillating. In case (1) with real $X$, the solution has an oscillating factor ${\vartheta}_0(X+a)$ but, since this does not change the sign, the oscillation is harmless (as can be checked by the modular transformation). For imaginary $X$, if the solutions contain the factors of ${\vartheta}_{0,1}$, they oscillate, whereas if the factors are ${\vartheta}_{2,3}$, they do not. The limiting cases with $k=0,1$ are similarly considered. Once one finds the solutions which grow large without harmful oscillation near the world-sheet boundary, they are good candidates of the cusp solutions. Though some of the cusps are generally at the infinity in the boundary Poincaré coordinates, $x_{\pm} = (Y_{1}\pm Y_{0})/(Y_{-1}+Y_{2})$, they can be brought to finite points by an $SL(2,{{\mathbb R}})$ transformation: Y = U V U = ( [cc]{}a & b\ c & d )   V = ( [cc]{} & -\ -& ) with $\det U = \det V = 1$. To see this, we trace a contour with a large radius in the world-sheet. Supposed that one of $\varphi_{j}, \phiinv_{j}$ alternatively becomes dominant along the contour, one finds that the contour is mapped to a rectangular in the $ x_{\pm}$-plane which has null boundaries and four cusps at $(x_{+},x_{-}) = (\frac{c}{a}, -\frac{\beta}{\delta}), (\frac{c}{a}, -\frac{\alpha}{\gamma}),(\frac{d}{b}, -\frac{\alpha}{\gamma}),(\frac{d}{b}, -\frac{\beta}{\delta})$. If $\varphi$ shows a more intricate behavior, more cusps and null boundaries may appear. We note that one should choose the $SL(2,{{\mathbb R}})$ transformation so that the Poincaré radial coordinate $1/(Y_{-1}+Y_{2})$ is non-negative: otherwise, the interpretation of the solution in the Poincaré corrdinates may be subtle. From the discussion in the above, we find that the elliptic solutions in 1-2 (iii) and 3-4 (ii) are harmfully oscillating solutions. However, in the limit $k \to 1$ for 1-2 (iii) and $k \to 0$ for 3-4 (ii), they become the exponential or the hyperbolic solutions in which the oscillation disappears. This is because, e.g., for 1-2 (iii), the period of the oscillation is of order $K$, and this is diverging as $k \to 1$; only the strip with the width of order $K$ survives after taking the limit. Conversely, this shows that unwanted oscillation might be eliminated by restricting the world-sheet in some region. In our case, simply taking the world-sheet to be the strip does not give cusp solutions with null boundaries, since the two sides of the strip are mapped to the boundary of the surface in $AdS$ which is not null. However, this may be a useful tip to further search for the cusp solutions. These elliptic solutions are regarded as elliptic generalizations of the known exponential solutions. The degenerate solutions of the hyperbolic/trigonometic type with $k=0,1$ are new solutions. As mentioned at the end of section 3.3, instead of taking appropriate limits from the generic cases, one may start with the generic form of the solutions in this case, \[degensol\] \_j[&=&]{} r\_j e\^[p\_[j]{}\^[+]{} \_[+]{} + p\_[j]{}\^[-]{} \_[-]{}]{},\ \_j [&=&]{} \_j e\^[-(p\_[j]{}\^[+]{} \_[+]{} + p\_[j]{}\^[-]{} \_[-]{})]{}, and determine them as in section 2. Indeed, one finds that the normalization condition and the equations of motion give 1 = r\_[1]{} \_[1]{} + r\_[2]{} \_[2]{} 0 = r\_[1]{}\_[1]{} \^[2]{}\_[1]{} + r\_[2]{}\_[2]{} \^[2]{}\_[2]{} \[degnorm\] and p\_[1]{}\^[+]{} p\_[1]{}\^[-]{} = p\_[2]{}\^[+]{} p\_[2]{}\^[-]{} 0= (p\_[j]{}\^[+]{} \^- + p\_[j]{}\^[-]{} \^+) \_[j]{} + 2 \^+ \^- (j=1,2) \[degeom\] respectively, whereas the Virasoro condition imposes && 0 = (p\_[1]{}\^)\^[2]{} r\_[1]{} \_1 + (p\_[2]{}\^)\^[2]{} r\_[2]{} \_2\ && 0 = (p\_[1]{}\^)\^[2]{} - (p\_[2]{}\^)\^[2]{} + 2\^ ( - ) \[degvirasoro\] Since we are interested in the cusp solutions with real $\varphi$, we impose the reality condition $(p_j^\pm)^* = p_j^\mp$, $(\mu^\pm)^* = \mu^\mp$. It turns out that the solutions to the constraints (\[degnorm\])-(\[degvirasoro\]) are essentially unique (up to conformal transformations of $\sigma_\pm$ etc.), and given by && \^= e\^[i]{} p\^\_1 = 1 p\^\_2 = i r\_1 \_1 = r\_2 \_2 =\ && \_1 =- \_2 = These give a class of real and non-oscillating solutions of the form,[^3] \_1[&=&]{} (- B) e\^[t]{},\ \_1\^[&=&]{} (+ B) e\^[-t]{},\ \_2[&=&]{} (+ B) e\^[s]{}, \[degeneratesol\]\ \_2\^[&=&]{} (- B) e\^[-s]{}, where $ B =\frac{\cos\theta}{\sqrt{2}} t - \frac{\sin\theta}{\sqrt{2}}s $, $ \sigma_{\pm}=(t \pm i s)/2 $, and we have assumed $\cos 2\theta > 0$. (The case with $\cos 2 \theta < 0$ is similar.) The potential $u$ reads u = - \^2 B \[degenpot\] ![ Boundary of minimal surface described by (\[degeneratesol\]) in $(x_{+}, x_{-})$-plane. A contour with a large radius in the world-sheet is mapped to the $(x_{+}, x_{-})$-plane along the arrows. ](6null_sol.eps){width="7cm"} This class of solutions includes solutions which have four cusps, two horns and six null boundaries, among which two pairs are collinear. We remark that all the boundaries are null. To check these properties, we first restrict to the case where $\cos \theta > 1/\sqrt{2}$ so that the Poincaré radial coordinate $r = 1/\varphi_1 $ is non-negative. (In the other case with $\cos \theta < -1/\sqrt{2}$, we have only to flip the signs of $r_j, \rinv_j$.) Next, we note that the AdS boundary is given by $|Y_{-1} + iY_0 | \to \infty$. Plugging the solution (\[degeneratesol\]) into $|Y_{-1} + iY_0 | $, we then find that the world-sheet boundary where $| t | $ or $ | s | \to \infty $ is mapped to the AdS boundary unless $ \cos \theta, \sin \theta = 0, \pm \frac{1}{\sqrt{2}} $. When $\theta$ takes such a generic value, similarly to the discussion in section 4.1 we find that the image of the world-sheet boundary traces six null segments in the $(x_+,x_-)$-plane. Concretely, the contour $t = \rho \sin \omega, \, s = \rho \cos \omega$ with $\rho \to \infty$ is mapped to $(-\infty,0) \to (\infty, 0) \to (0,0) \to (0,-\infty) \to (0,\infty) \to (-\infty, \infty) \to (-\infty, 0)$ as $\omega$ varies from $0 $ to $2\pi$. The resultant boundary of the surface is not convex, but crossed and folded as in Fig.1. Among the six end-points of the segments, $(0,0), (0, -\infty), (-\infty, \infty), (-\infty, 0)$ are the cusps and $(\infty, 0),(0, \infty)$ are the tips of the two horns. The essence in producing the six null boundaries is that the change of the sign of $\tanh B$ “splits” the cusps, which, in the $x_\pm$-plane, is observed as the transitions $(-\infty,0) \to (+\infty, 0)$ and $(0,-\infty) \to (0,+\infty)$. In these transitions, the surface boundary has to keep touching the AdS boundary. Since the solution has two pairs of collinear null boundaries, one expects that it gives the scattering amplitudes at strong coupling in a collinear limit. When $ \sin \theta =0$, the surface boundaries mapped from $ t = 0$ do not reach the AdS boundary, and they form two boundaries inside $AdS_{3}$. Consequently, the solution describes a surface which has four null boundaries at the AdS boundary, and two boundaries inside AdS. The surface pinches at a point where these two boundaries intersect each other. The shape of the surface is obtained by diagonally cutting a four-cusp surface and twisting it. From the potential in (\[degenpot\]), one finds that the surface has non-trivial curvature. This shows a clear difference from the four-cusp solution in [@Kruczenski:2002fb; @Alday:2007hr], where the potential is constant and hence the corresponding surface is flat. In fact, the two solutions are not related to each other by simple transformations: First, they cannot be related by an $SO(2,2)$ transformation, since the potential $u = -\del_{+}\vec{Y}\cdot \del_{-}\vec{Y}$ is invariant. Second, as long as we work with the Euclidean world-sheet, the allowed world-sheet analytic continuation is the continuation of both $t$ and $s$, which results in $\del_{\pm} \to \pm i \del_{\pm}$. Thus, $u$ is invariant up to a sign and renaming the world-sheet coordinates. Third, one may generate a new solution by a target-space analytic continuation such as $Y_{a} \to i Y_{a}$ together with the world-sheet analytic continuation as in [@Kruczenski:2007cy]. However, the potential should again be invariant (up to a sign and renaming of the world-sheet coordinates) in order to keep the equations of motion invariant. Finally, if the potential $u$ has a factorized form $f(\sigma_{+})g(\sigma_{-})$, it may be brought to a constant by a world-sheet conformal transformation, but this is not possible for the degenerate solution. The analysis so far can be generalized to the case of the strings in $AdS_{5} \times S^{5}$. Here, for simplicity, we consider the case of $AdS_{3} \times S^{1}$. In an appropriate gauge, the $S^{1}$ field is set to be $W = \kappa_{+} \sigma_{+} + \kappa_{-} \sigma_{-} $. The reality of $W$ requires $(\kappa_{+})^{*} = \kappa_{-}$. Adding $S^{1}$ does not change the reality condition on $Y$, but that changes the Virasoro constraints to \_[j=1]{}\^2\_ \_j\_\^\_j = \_\^[2]{} Similarly to the case without $S^{1}$, linear combinations of these give && u\_[0]{} = 2(+-1-k\^2) - ( \_[+]{}\^[2]{} + \_[-]{}\^[2]{})\ && 2U\^+U\^- \_[j=1]{}\^2(-)\^[j+1]{} (\^[+]{}\_[j]{} - \^[-]{}\_[j]{}) = ( \_[+]{}\^[2]{} - \_[-]{}\^[2]{}) Because of the change of the Virasoro constraints, the allowed solutions for $Y$ also change. Though they can be classified as in section 3, we do not go into details. However, we know that, in order to find cusp solutions, we have only to look into the cases without harmful oscillation. Among the elliptic cases, they are 1-1 or 3-3 in section 3.2, which are related to each other by the modular transformation. In the following, we take 3-3. It turns out that this case indeed gives four-cusp solutions with null boundaries which are expressed by the elliptic functions. For example, for $k = 0.7, U^\pm = i, A_{1} = i K'/2, \kappa_{\pm} = \sqrt{12/5}(1\pm i)$, the Virasoro condition gives $A_{2} = 1.277...$ and $u_{0}=-4.781... \,$. Further setting $X_0 = 0$, the theta function takes the form ${\vartheta}_0(X +ia)= {\vartheta}_3(i t+ ia)$. By repeating the shifts in the imaginary direction as in (\[thetashift\]), one then finds that the ratio of the theta functions shows an exponential behavior ${\vartheta}_0(X +A)/{\vartheta}_0(X) \sim e^{\pi a t/(2KK')}$. Thus, along a contour with a large radius in the world-sheet, one of $\varphi_j,\phiinv_j$ alternatively becomes dominant. Since this shows that the mechanism in section 4.1 works in this case, the solution describes a surface with four null boundaries and four cusps. The points of the cusps in the $x_{\pm}$-plane can be brought to finite points as in section 4.1. Since the Virasoro constraints are changed, the surface spanned by the solution is not necessarily space-like anymore. This can be checked by considering the normal vector to the surface $N_{a} := u^{-1} \ep_{abcd} Y^{b} \del_{+} Y^{c} \del_{-} Y^{d} $, the norm of which is $ N^{2} = 1 - {\kappa_{+}^{2} \kappa_{-}^{2}}/{u^{2}} \period $ Evaluating $u = 2k^{2} \sn^{2} X + u_{0} $ in this example shows that $N^{2} < 0$ and hence the surface is time-like. We have systematically searched for the classical open string solutions in $AdS_{3}$ within the genus-one finite-gap solutions, and given a classification of the allowed solutions. When the elliptic modulus degenerates, we have found a class of solutions with six null boundaries, among which two pairs are collinear. Adding $S^{1}$ to $AdS_{3}$, we have also found solutions expressed by the elliptic functions, which have four cusps and four null boundaries. The analysis in this paper can straightforwardly be applied to the case with the Lorentzian world-sheet. It may also be useful for studying the classical solutions describing the Wilson loops in the super Yang-Mills theory at strong coupling [@Maldacena:1998im; @Rey:1998ik]. The classical open string solutions in $AdS_{5} \times S^{5}$ are similarly discussed. In particular, for the strings in $AdS_{5}$, we have only to add another pair of $\varphi_{3}, \phiinv_{3}$. In this case, these are identified with a complex combination of the embedding coordinates in $AdS_{5}$ as $\varphi_{3}, \phiinv_{3} = Y_{3}\pm i Y_{4}$, and thus the solutions are generally (harmfully) oscillating. In such oscillating cases, a way to remove the unwanted oscillation is to restrict the world-sheet, as mentioned in section 4.2. Though it is still non-trivial to find desired solutions with cusps and null boundaries, the prescription in [@AM2] suggests that effectively restricting the world-sheet by conformal transformations deserves further consideration. The essence of the solution with six null boundaries in section 4.3 is the change of the sign of $\tanh B$ in front of the exponentials, which “splits” the cusps. Similarly, more intricate behavior of the corresponding factors in the higher-genus cases may produce solutions with more null boundaries. It is interesting to consider the relation to the mechanism provided in [@AM2]. Most of the end points of the null segments in our solutions with six null boundaries are located at the infinity of the $AdS_{3}$ boundary. Since the surface is space-like, this is inevitable in the $AdS_{3}$ boundary. However, it is desirable to bring them to finite points in the $AdS_{5} $ boundary by some transformations, as discussed in [@AM2]. This may be a first step toward applications to the scattering amplitudes. We would like to report progress in the analysis of the higher-genus finite-gap solutions, multi-cusp solutions and the applications to the scattering amplitudes, elsewhere. Our conventions of the elliptic theta functions are: \_[ab]{}(w,) := \_[n=-]{}\^ and \_[0]{}(z) := \_[01]{}(w,) \_[1]{}(z) := -\_[11]{}(w,) \_[2]{}(z) := \_[10]{}(w,) \_[3]{}(z) := \_[00]{}(w,) where $w = z/(2K) $ and $K(k)$ is the complete elliptic integral of the first kind. In the main text, we use the formulas Z(u+v)=Z(u)+Z(v)-k\^2uv(u+v) \[addZ\] where $Z(z) := \del_{z} \ln {\vartheta}_{0}(z)$, and \_[0]{}(u K) &=& \_[3]{}(u)\ \_[0]{}(u iK’) &=& i e\^[-i ( + )]{}\_[1]{}(u) \[thetashift\]\ \_[0]{}(u (K+iK’)) &=& e\^[-i ( + )]{}\_[2]{}(u) [**Acknowledgments**]{} We would like to thank D. Bak, S. Hirano, N. Ishibashi, K. Ito, H. Itoyama, C. Kalousios, T. Matsuo and K. Mohri for useful conversations. The work of K.S. and Y.S. is supported in part by Grant-in-Aid for Scientific Research from the Japan Ministry of Education, Culture, Sports, Science and Technology. \#1[ 40004000 ‘=1000]{} = [**References**]{} [999]{} =-2.5pt L. F. Alday and J. M. Maldacena, JHEP [**0706**]{} (2007) 064 \[arXiv:0705.0303 \[hep-th\]\]. L. F. Alday and R. Roiban, Phys. Rept.  [**468**]{} (2008) 153 \[arXiv:0807.1889 \[hep-th\]\]. S. Ryang, Phys. Lett.  B [**659**]{} (2008) 894 \[arXiv:0710.1673 \[hep-th\]\]. A. Jevicki, K. Jin, C. Kalousios and A. Volovich, JHEP [**0803**]{} (2008) 032 \[arXiv:0712.1193 \[hep-th\]\]. D. Astefanesei, S. Dobashi, K. Ito and H. Nastase, JHEP [**0712**]{} (2007) 077 \[arXiv:0710.1684 \[hep-th\]\]. S. Dobashi, K. Ito and K. Iwasaki, JHEP [**0807**]{} (2008) 088 \[arXiv:0805.3594 \[hep-th\]\]. A. Mironov, A. Morozov and T. N. Tomaras, JHEP [**0711**]{} (2007) 021 \[arXiv:0708.1625 \[hep-th\]\]; Phys. Lett.  B [**659**]{} (2008) 723 \[arXiv:0711.0192 \[hep-th\]\]. H. Itoyama, A. Mironov and A. Morozov, Nucl. Phys.  B [**808**]{} (2009) 365 \[arXiv:0712.0159 \[hep-th\]\]. H. Itoyama and A. Morozov, Prog. Theor. Phys.  [**120**]{} (2008) 231 \[arXiv:0712.2316 \[hep-th\]\]. C. M. Sommerfield and C. B. Thorn, Phys. Rev.  D [**78**]{} (2008) 046005 \[arXiv:0805.0388 \[hep-th\]\]. H. Dorn, G. Jorjadze and S. Wuttke, arXiv:0903.0977 \[hep-th\]. M. Kruczenski, JHEP [**0212**]{} (2002) 024 \[arXiv:hep-th/0210115\]. L. F. Alday and J. Maldacena, arXiv:0904.0663 \[hep-th\]; arXiv:0903.4707 \[hep-th\]. I. M. Krichever, Func. An. $\&$ Apps. [**28**]{} (1994) No. 1, 26. V. A. Kazakov and K. Zarembo, JHEP [**0410**]{} (2004) 060 \[arXiv:hep-th/0410105\]. H. Hayashi, K. Okamura, R. Suzuki and B. Vicedo, JHEP [**0711**]{} (2007) 033 \[arXiv:0709.4033 \[hep-th\]\]. M. Kruczenski, R. Roiban, A. Tirziu and A. A. Tseytlin, Nucl. Phys.  B [**791**]{} (2008) 93 \[arXiv:0707.4254 \[hep-th\]\]. J. M. Maldacena, Phys. Rev. Lett.  [**80**]{} (1998) 4859 \[arXiv:hep-th/9803002\]. S. J. Rey and J. T. Yee, Eur. Phys. J.  C [**22**]{} (2001) 379 \[arXiv:hep-th/9803001\]. [^1]: For the closed strings in $AdS_{3} \times S^{1}$, another general construction is given in [@Kazakov:2004nh]. Explicit genus-one finite-gap solutions are discussed in [@Hayashi:2007bq]. [^2]: In addition, when $A_{j} = 0, iK'$, some of the expressions below become singular. In the case of $A_{j} = 0$, the solution becomes of the exponential type without the theta functions. The case with $A_{j} = iK'$ is treated as a limiting case from $A_{j} \neq iK'$. [^3]: Shifting the argument of $\cosh$ in (\[degensol\]) by $\pi i/2$ gives solutions with $\coth B$.

--- abstract: | The theory of quantum jump trajectories provides a new framework for understanding dynamical phase transitions in open systems. A candidate for such transitions is the atom maser, which for certain parameters exhibits strong intermittency in the atom detection counts, and has a bistable stationary state. Although previous numerical results suggested that the “free energy” may not be a smooth function, we show that the atom detection counts satisfy a large deviations principle, and therefore we deal with a phase cross-over rather than a genuine phase transition. We argue however that the latter occurs in the limit of infinite pumping rate. As a corollary, we obtain the Central Limit Theorem for the counting process. The proof relies on the analysis of a certain deformed generator whose spectral bound is the limiting cumulant generating function. The latter is shown to be smooth, so that a large deviations principle holds by the Gärtner-Ellis Theorem. One of the main ingredients is the Krein-Rutman Theorem which extends the Perron-Frobenius theory to a general class of positive compact semigroups. author: - | Merlijn van Horssen and M[ă]{}d[ă]{}lin Gu[ţ]{}[ă]{}\ University of Nottingham, School of Mathematical Sciences\ University Park, NG7 2RD Nottingham, UK bibliography: - 'bibliography.bib' title: 'Large Deviations, Central Limit and dynamical phase transitions in the atom maser' --- \[thm\][Definition]{} \[thm\][Lemma]{} \[thm\][Corollary]{} Introduction ============ The last couple of decades have witnessed a revolution in the experimental realisation of quantum systems . Ultracold atomic gases are created and used for the study of complex many body phenomena such as quantum phase transitions [@Sachdev] shedding light on open problems in condensed-matter physics . Real quantum systems are “open” in the sense that they interact with their environment, which leads to an irreversible loss of coherence and to energy dissipation. In many cases, the dynamics can be well described by the Markov approximation in which the environment possesses no memory and interacts weakly with the system. The joint unitary evolution of the system and environment can be described through the input-output formalism [@Gardiner2004] using quantum stochastic calculus [@Parthasarathy1992]. In this framework, the Markov semigroup can be seen as the average of a stochastic quantum trajectories arising from continuous-time measurements performed in the environment. Since in many experiments the system is not directly accessible, its (conditional) evolution is inferred from the detection trajectories via stochastic Schrödinger (or filtering) equations [@Belavkin; @Plenio1998]. In [@Garrahan2009] a new perspective was put forward, which looks at quantum jumps from the viewpoint of non-equilibrium statistical mechanics . Detection trajectories are seen as “configurations” of a stochastic system, and large deviations theory [@Dembo2010; @Ellis1995] is employed to study the *dynamical phase transitions* arising in this way. Consider for simplicity the case of a counting measurement, which is directly relevant for the model considered in this paper. The interesting scenarios are that of a *phase cross-over* in which the counting trajectories show intermittency between long active periods (many counts) and passive ones (few counts), and that of *phase coexistence* where the counting process exhibit a mixture of infinitely long trajectories of either type. In the latter case, the asymptotic cumulant generating function (or “free energy”) of the total counts process $\Lambda_t$ is singular at the origin, and the total counts do not obey a large deviations principle (LDP). In contrast, in a phase cross-over an LDP may hold but numerically and practically there would be a strong resemblance to an actual phase transition. For finite dimensional systems the counting process $\Lambda_t$ satisfies an LDP when the Markov dynamics is mixing, i.e. irreducible and aperiodic [@Hiai2007]. The proof uses the Gärtner-Ellis theorem according to which it suffices to prove the convergence of the cumulant generating function to a smooth limit. By the Markov property, the former can be expressed in terms of a certain “deformed generator” $\mathcal{L}_s$, and the existence of the limit $$\lim_{t\to\infty}\frac{1}{t} \log \mathbb{E}(e^{s\Lambda_t}) = \lim_{t\to\infty}\frac{1}{t} \log {\rm Tr}(\rho_{in} e^{t\mathcal{L}_s} (\mathbf{1}))=\lambda(s)$$ follows from the spectral gap property of $\mathcal{L}_s$, with $\lambda(s)$ being the spectral bound of $\mathcal{L}_s$. In this paper we investigate the existence of dynamical phase transitions for the atom maser, a well known quantum open system exhibiting interesting properties such as bistability and sub-Poissonian statistics . The maser consists of a beam of excited atoms passing through a cavity with which they interact according to the Jaynes-Cummings model. After the interaction the atoms are measured in the standard basis and the trajectory of measurement outcomes is recorded. For certain values of the interaction strength, the stationary mean photon number changes abruptly (cf. Figure \[fig:stationary\]), and the distribution is bistable, having a low and a high energy “phase”. The measurement trajectories alternate between periods of low and high ground state atoms counts (cf. Figure \[fig:trajectories\]), and its limiting moment generating function exhibits characteristic phase separation lines (cf. Figure \[fig:grids\]). Our main result (Theorem \[th.main\]) is that the counts process satisfies an LDP, and therefore the atom maser does not have the non-analytic properties characteristic of phase transitions, although it exhibits clear phase cross-over(s) which become sharper with increasing pumping rate. As a corollary, we obtain the Central Limit Theorem for the counting process, using a result of [@Bryc]. The proof follows the line of [@Hiai2007], but the novelty here is the treatment of an infinite dimensional system in continuous time dynamics. We use an $L^2$-representation [@Fagnola1994; @Carbone2000] of the semigroup generated by $\mathcal{L}_s$ and show that the corresponding semigroup is compact. We then use the Krein-Rutman Theorem to establish the uniqueness and strict positivity of the eigevector of $\lambda(s)$, and hence the existence of the spectral gap. Some steps of the proof rely on a special feature of the maser dynamics which allows us to restrict the attention to the commutative invariant algebra of diagonal operators. However the line of the proof is applicable to general infinite dimensional quantum Markov dynamics. For recent work on quantum dynamical phase transitions we refer to [@Garrahan2009; @Garrahan2007; @Hedges2009; @Garrahan2008; @Budini2010]. In particular, our investigation was motivated by the numerical results of [@Garrahan2011] indicating a possible non-analytic behaviour of $\lambda(s)$. In [@Hiai2007] (see also [@Ogata2010]) a large deviation principle is shown to hold for correlated states on quantum spin chains; large deviations for quantum Markov semigroups are studied in [@Comman2008]. Metastable behaviour in a different atom maser has been investigated in [@Bruneau2008]. More broadly, there is a large body of large deviations work in quantum systems . In Section 2 we introduce the background of our problem: the atom maser and its Markov semigroup, the counting processes associated to the jump terms in the Lindblad generator, the static and dynamical phase transitions and the interplay between them, and the general setup of large deviations theory. In Section 3 we formulate the large deviations results and give a point by point outline of the proof. The details of the proof found in Section 4. The results of a detailed numerical analysis are presented in Section 5, where we argue that “phase transitions” does occur in the limit of very large pumping rate, at $\alpha\approx 1$ (second order), at $\alpha\approx 6.66$ and further points (first order). Background ========== In this section we introduce the setup of the atom maser dynamics, investigate the counting process associated to the measurement of outgoing atoms, and describe the basic elements of large deviations theory used in the paper. Atom maser ---------- In the atom maser, two-level atoms pass successively through a cavity and interact resonantly with the electromagnetic field inside the cavity. The two-level atoms are identically and independently prepared in the excited state, and for simplicity we assume that only a single atom passes through the cavity at any time. In addition, the cavity is also coupled to a thermal bath which represents the interaction between the (non-ideal) cavity and the environment. The combined effects of the interactions with the atoms and the environment changes the state of the cavity whose time evolution is described by a *quantum Markov semigroup*, in a certain coarse grained approximation described below; see Refs. [@Carbone2000] and  [@Fagnola1994] for a mathematical overview, and [@Englert2002] for the physical derivation of the master equation. In this section we give an intuitive description of the dynamics starting with a simplified discrete time model, with an emphasis on the statistics of measurements performed on the atoms. The cavity is described by a one mode continuous variable system with Hilbert space $\mathfrak{h} = \ell^{2}(\mathbb{N})$ whose canonical basis vectors $(|e_{n}\rangle)_{n \geq 0}$ represent pure states of fixed number of photons. Therefore, if $|\psi\rangle \in \mathfrak{h}$ is a pure state, the *photon number distribution* of the cavity is given by ${\lvert {\left\langle e_{n}, \psi \right\rangle} \rvert}^{2}$. Mixed states are described by density operators, i.e. trace-class operators $\rho \in L^1(\mathfrak{h})$ which are positive and normalised to have unit trace, and the observables are represented by self-adjoint elements of the von Neumann algebra of bounded operators $\mathcal{B}(\mathfrak{h})$ whose predual is $L^1(\mathfrak{h})$. Recall that the *annihilation operator* $a$ on $\mathfrak{h}$ is defined by $$a |e_{n}\rangle = \begin{cases} \sqrt{n} |e_{n-1}\rangle &\text{if } n > 0 \\0 & \text{if } n = 0 \end{cases};$$ its adjoint is the *creation operator* $a^{*}$, and $N = a^{*}a$ is the photon number operator such that $N|e_n\rangle = n |e_n\rangle $. The atom is modelled by a two-dimensional Hilbert space $\mathbb{C}^{2}$ with standard orthonormal basis ${ \left\lbrace {\vert 0 \rangle},{\vert 1 \rangle} \right\rbrace }$ consisting of the “ground” and “excited” states. We denote by $\sigma^{*}$ and $\sigma$ the corresponding raising and lowering operators (i.e. $\sigma^{*} {\vert 0 \rangle} = {\vert 1 \rangle}$ etc.). The interaction between an atom and the cavity is described by the Jaynes-Cummings hamiltonian on $\mathbb{C}^2\otimes \mathfrak{h}$ $$H_{\mbox{int}} = - g ( \sigma \otimes a^{*} + \sigma^{*} \otimes a),$$ where $g$ is the coupling constant which, for simplicity, is considered to be constant across the cavity. The free hamiltonian is $$H_{\mbox{free}} = \omega \mathbf{1}\otimes a^{*} a + \omega \sigma^{*} \sigma \otimes \mathbf{1} ,$$ where $\omega$ is the frequency of the resonant mode; however by passing to the interaction picture the effect of the free evolution can be ignored. Therefore if the interaction lasts for a time $t_{0}$, the joint evolution is described by the unitary operator $U:= \exp(it_{0}H_{\mbox{int}})$ whose action on a product initial state is $$U : {\vert k \rangle} \otimes {\vert 1 \rangle} \mapsto \cos(\phi\sqrt{k+1}){\vert k \rangle}\otimes{\vert 1 \rangle} + \sin(\phi\sqrt{k+1}){\vert k+1 \rangle}\otimes{\vert 0 \rangle} ,$$ where $\phi:= t_{0} g$ is the *accumulated Rabi angle*. If a measurement is performed on the outgoing atom in the standard basis, then the cavity remains in state ${\vert k \rangle}$ with probability $\cos^{2}(\phi\sqrt{k+1})$ or gains an excitation with probability $\sin^{2}(\phi\sqrt{k+1})$. If we average over the outcomes, we obtain the cavity transfer operator $\mathcal{T}_{*}: L^1(\mathfrak{h})\to L^1(\mathfrak{h})$ $$\label{eq:transitionoperator} \mathcal{T}_{*}(\rho) = K_{1} \rho K_{1}^{\ast} + K_{2} \rho K_{2}^{\ast} = \mathcal{K}_1(\rho) + \mathcal{K}_2(\rho)$$ where the *Kraus operators* $K_{i}$ are given by $$\quad K_{1} = a^{*}\frac{\sin(\phi\sqrt{a a^{*}})}{\sqrt{a a^{*}}}, \quad K_{2} = cos(\phi\sqrt{a a^{*}}),$$ and $\mathcal{K}_{i}$ are the corresponding jump operators on the level of density matrices. Since each atom interacts with the cavity only once, the state of the cavity after $n$ such interactions is given by $\rho(n)= \mathcal{T}_{*}^{n}(\rho)$, which can be interpreted as a *discrete time* quantum Markov dynamics. Let us imagine that after the interaction, each atom is measured in the standard basis and found to be either in the excited or the ground state. The master dynamics can be unravelled according to these events as follows $$\label{eq:discrete.unravel} \mathcal{T}_{*}^{n}(\rho) = \sum_{{\bf i} = (i_1,\dots ,i_n)} \mathcal{K}_{i_n} \dots \mathcal{K}_{i_1} (\rho)$$ where each term of the sum represents the (unnormalised) state of the cavity after a certain sequence ${\bf i}=(i_1,\dots ,i_n)\in\{e,g\}^n$ of measurement outcomes, whose probability is $$\mathbb{P}( i_1, \dots , i_n) = {\rm Tr} (\mathcal{K}_{i_n} \dots \mathcal{K}_{i_1} (\rho) ).$$ If $\Lambda_n({\bf i}):= \#\{ j: i_j= g\}$ denotes the number of ground state atoms detected up to time $n$, we can use the previous relation to compute its moment generating function $$\label{eq:mgfdiscrete} \mathbb{E}\left(e^{s \Lambda_{n}}\right) = \sum_{k \geq 0} \mathbb{P}\left(\Lambda_{n} = k\right) e^{sk} = \sum_{\bf i} e^{s\Lambda_n({\bf i})} {\rm Tr} ( \mathcal{K}_{i_n} \dots \mathcal{K}_{i_1} (\rho) ) = {\rm Tr} ( \mathcal{T}_{*s}^n (\rho) )$$ where $$\mathcal{T}_{*s}(\rho) =e^{s} \mathcal{K}_1 (\rho) +\mathcal{K}_2 (\rho)$$ is a “deformed” transfer operator, i.e. a completely positive but not trace preserving map on $L^1(\mathfrak{h})$. The relation and its continuous time analogue will be the key to analysing the large deviations properties of the counting process in terms of spectral properties of operators such as $\mathcal{T}_s$ and $\mathcal{L}_s$ below. To make the model more realistic we will pass to a continuous time description in which the incoming atoms are Poisson distributed in time with intensity $N_{\text{ex}}$, and the cavity is in contact with a thermal bath. If one ignores the details of short term cavity evolution, the discrete time dynamics can be replaced by coarse grained *continuous time* Lindblad (master) equation [@Englert2002a] $$\begin{aligned} \frac{{\text{d}}}{{\text{d}}t} \rho(t) &= \mathcal{L}_{*}(\rho(t)),\nonumber\\ \mathcal{L}_{*}(\rho) &= \sum_{i=1}^{4} \left( L_{i} \rho L_{i}^{\ast} - \frac{1}{2}\lbrace L_{i}^{\ast} L_{i}, \rho \rbrace \right)\nonumber\\ &=\sum_{i=1}^{4} L_{i} \rho L_{i}^{\ast} + \mathcal{L}^{(0)}_{*} (\rho) = \sum_{i=1}^{4} \mathcal{J}_{i} (\rho) + \mathcal{L}^{(0)}_{*} (\rho) \label{eq:lindblad} \end{aligned}$$ with jump operators $L_{i}$ defined by $$\begin{aligned} L_{1} &= \sqrt{N_{\text{ex}}} a^{*} \frac{\sin(\phi \sqrt{a a^{*}})}{\sqrt{a a^{*}}},\label{eq:L2}\\ L_{2} &= \sqrt{N_{\text{ex}}} \cos(\phi \sqrt{a a^{*}}),\\ L_{3} &= \sqrt{\nu+1}a,\\ L_{4} &= \sqrt{\nu} a^{*}.\label{eq:L4} \end{aligned}$$ As before, the operators $L_{1}$ and $L_{2}$ are associated to the detection of an atom in the ground and excited state, respectively. The emission and absorption of photons due to contact with the bath is represented by operators $L_{3}$ and $L_{4}$, respectively. The Heisenberg picture Lindbladian $\mathcal{L}$ is the generator of a strongly continuous semigroup on $\mathcal{B}(\mathfrak{h})$ (cf. [@Engel2001] for an introduction to the general theory.) This means that there exists a family $(\mathcal{T}(t))_{t \geq 0}$ of maps on $\mathcal{B}(\mathfrak{h})$ which satisfy the *semigroup property* $$\begin{aligned} \mathcal{T}(t)\mathcal{T}(s) &= \mathcal{T}(t+s)\quad\text{for all }t,s \geq 0,\nonumber\\ \mathcal{T}(0) &= I,\nonumber \end{aligned}$$ such that $t \mapsto \mathcal{T}(t) (X)$ is norm continuous for all $ X\in \mathcal{B}(\mathfrak{h})$. Moreover, $\mathcal{T}(t)$ is completely positive and unit preserving for all $t\geq 0$. The generator $\mathcal{L}$ can be recovered by $$\mathcal{L}(X) = \lim_{h \downarrow 0} \tfrac{1}{h}\left( \mathcal{T}(h) (X) - X\right),$$ for all $X$ in the domain of $\mathcal{L}$. Although no simple expression exists for the operators $\mathcal{T}(t)$ in terms of the generator $\mathcal{L}$, it is helpful to think of $\mathcal{T}(t)$ as the exponential of the generator $$\label{eq:semigroup} \mathcal{T}(t) (X) = e^{t \mathcal{L}} (X),$$ especially from the point of view of relating spectral properties of $\mathcal{L}$ to those of $\mathcal{T}(t)$, e.g. *spectral mapping theorems*. By definition, equation does hold for [analytic vectors]{} of $\mathcal{L}$ which form a core of its domain (for details, see proof of Lemma \[lemma:transfer.mgf\]). The Markov semigroup $\left(\mathcal{T}(t)\right)_{t \geq 0}$ has a unique stationary state $\rho_{\text{ss}}$ which has diagonal density matrix in the Fock (photon number) basis with entries $$\label{eq:stationary} \rho_{\text{ss}}(n) := \rho_{\text{ss}}(0) \prod_{k=1}^{n}\left( \frac{\nu}{\nu +1} + \frac{N_{\text{ex}}}{\nu +1} \frac{\sin^{2}(\phi \sqrt{k})}{k} \right)$$ equal to the probabilities of finding $n$ photons in the cavity, with $\rho_{\text{ss}}(0)$ taken such that ${\text{Tr}}(\rho_{\text{ss}}) =1$. Moreover, the Markov semigroup is ergodic, in the sense that any initial states $\rho$ converges to the stationary state [@Fagnola2001] $$\lim_{t\to\infty} \mathcal{T}_{*}(t)(\rho) = \rho_{ss}.$$ ![Mean photon number (black line) and photon number distribution (background) in the stationary state $\rho_{ss}$ as function of $\alpha = \sqrt{N_{\mbox{ex}}} \phi$[]{data-label="fig:stationary"}](fig1ssgrid){width="90.00000%"} The dependence of the stationary mean photon number and photon number distribution on the “pumping parameter” $\alpha:= \sqrt{N_{\text{ex} }}\phi$ is shown in Fig. \[fig:stationary\], for $\nu = 0.15$ and $N_{\text{ex}} = 150$. We note two interesting features in this figure: first, there is a sharp change in the mean photon number at $\alpha \approx 1$ followed by less pronounced jumps near $\alpha = 6.66 $ and $\alpha = 12$. The other, related, feature to note is that the photon number distribution has a single peak for most values of $\alpha$ except in certain regions such as around the critical point $\alpha \approx 6.66$, where the stationary state has two local maxima. We will come back to these aspects in the next section. The counting process -------------------- To better understand the behaviour of the stationary state illustrated in Figure \[fig:stationary\], we unravel the Markov semigroup $\mathcal{T}_*(t)$ with respect to the four counting processes associated to the jump terms (\[eq:L2\] - \[eq:L4\]), each of them corresponding to a counting measurement of the quantum output process. If $\rho$ is the initial state of the cavity, then $\rho(t):=\mathcal{T}_*(t)$ is the evolved state at time $t$ which (in analogy to Eq.) can be seen as an average over all possible counting events in the environment $$\label{eq:unravel} \rho(t):=\mathcal{T}_*(t) = \sum_{k\geq 0} \, \sum_{i_1,\dots i_k =1}^4 \underset{0\leq t_1\leq \dots \leq t_k\leq t}{\int \dots\int } \rho(t; t_1,i_1\dots ,t_k, i_k) dt_1 \dots dt_k$$ where the integrand $$\rho(t; t_1,i_1,\dots, t_k, i_k):= e^{(t-t_k) \mathcal{L}^{(0)}_{*}} \mathcal{J}_{i_{k}} \dots e^{(t_2- t_1) \mathcal{L}^{(0)}_{*}} \mathcal{J}_{i_1} e^{t_1 \mathcal{L}^{(0)}_{*}} (\rho),$$ is the unnormalised state of the cavity given that detections of type $i_1,\dots , i_k\in\{1,2,3,4\}$ have occurred at times $0\leq t_1\leq \dots \leq t_k\leq t$, and no other counting events happened in the meantime. From the four counting processes we focus on the first one associated with the detection of an atom in the ground state and simultaneous absorption of a photon by the cavity. We denote by $\Lambda_t$ the total number of such atoms detected up to time $t$. Similarly to the discrete case, by using the above unravelling we can show that the moment generating function of $\Lambda_t$ is given by $$\label{eq:mgf} \mathbb{E}\left(e^{s \Lambda_{t}}\right) = {\text{Tr}}\left( \mathcal{T}_{*s}(t)(\rho)\right)= {\text{Tr}}\left( \rho \mathcal{T}_s(t)(\mathbf{1})\right).$$ where $\left(\mathcal{T}_{*s}(t)\right)_{t \geq 0}$ is the completely positive semigroup on $L^1(\mathfrak{h})$ with generator $$\mathcal{L}_{*s}(\rho) = e^s \mathcal{J}_1(\rho) + \sum_{i=2}^{4} \mathcal{J}_j(\rho)+ \mathcal{L}^{(0)}_*(\rho)= (e^s-1) \mathcal{J}_1(\rho)+ \mathcal{L}_{*}(\rho) ,\label{eq:perturbed}$$ and $\left(\mathcal{T}_s(t)\right)_{t \geq 0}$ is the dual semigroup on $\mathcal{B}(\mathcal{H})$. Equation plays a central role in this paper; we will use it to formulate a large deviations principle for the counting process $\Lambda_{t}$, and in particular, to relate the moment generating function of $\Lambda_{t}$ to the spectral properties of $\mathcal{L}_{s}$. Note that $\mathcal{L}_{s}$ differs from the Lindblad generator by the factor $e^s$ multiplying the jump term associated to the detection of a ground state atom. It is still the generator of a completely positive semigroup, but it is no longer identity preserving, and therefore does not represent a physical evolution except for $s=0$. ![The birth (blue) and death rates as functions of $\vartheta$ for different values of $\alpha$. The intersection points correspond to minima and maxima of the stationary distribution.[]{data-label="fig:b.d.rates"}](fig2rates){width="70.00000%"} The unravelling allows for a classical interpretation of the cavity dynamics. Indeed, the semigroup generated by $\mathcal{L}$ (and $\mathcal{L}_s$) leaves invariant the commutative subalgebra $\mathcal{N}\subset \mathcal{B}(\mathfrak{h})$ generated by the number operator $N$, and the restriction of $\left(\mathcal{T}(t)\right)_{t \geq 0}$ to the diagonal algebra is the dynamical semigroup of a classical *birth-death process* on the state space $\{0, 1,2, \dots\}$, with rates $$\begin{aligned} && \lambda_{k}^2:= N_{ex} \sin(\phi \sqrt{k+1})^{2} + \nu (k+1) ,\quad k\geq 0\nonumber \\ && \mu_{k}^2 := (\nu+1)k, \quad k\geq 1\label{eq.birth-death}\end{aligned}$$ and stationary distribution $\pi_{\text{ss}} (n)= \rho_{\text{ss}}(n)$. Figure \[fig:b.d.rates\] shows the birth and death rates (minus the common factor $\nu k$) as functions of the parameter $\vartheta:= \sqrt{(k+1)/N_{ex}} \alpha$ in the limit $N_{ex} \rightarrow \infty$. The intersection points correspond to minima and maxima of the stationary distribution [@Englert2002a] as suggested by the following argument. For $\alpha<1$ the death rate is always larger than the birth rate and the distribution is maximum at the vacuum state. For $1< \alpha<4.6$ there is a single non-trivial intersection point such that the birth rate is larger to its left and smaller to its right, and therefore corresponds to the maximum of the stationary distribution. Similarly, when $4.6<\alpha< 7.8$ the rates intersect in three points, the first and last are located at local maxima while the middle point is a local minimum, so we deal with a bimodal distribution. However, while this analysis clarifies the emergence of multimodal distributions, it does not explain the sudden jump of the mean photon number at $\alpha\approx 6.66$. This feature can be intuitively understood by appealing to the effective potential model . If we think of the photon number as a continuous variable and introduce a fictitious potential $U$ defined by $$\label{eq.potential} \rho_{ss}(n) = \rho_{ss}(0) e^{-U(n)},$$ then the photon number distribution appears as the thermal equilibrium distribution of a particle moving in the potential $U$ (with $k_B \cdot T=1 $), see Figure \[fig:potentials\]. When the potential has a single local minimum (for $0<\alpha<4.6$), the stationary distribution is unimodal and concentrates around this point. The cavity state fluctuates around the mean, and $\Lambda_t$ increases steadily with average rate. When there are two (or more) local minima of different height, the higher minimum corresponds to a metastable phase from which the system eventually escapes due to thermal fluctuations. The rate of return to the metastable phase is typically much lower due to the larger potential barrier that needs to be climbed. The point $\alpha \approx 6.66$ where the two local minima are equal plays the role of a “phase transition”, and corresponds roughly to the point where the mean photon number changes abruptly. Here the cavity spends long periods of time around the two local maxima with rare but quick transitions between them. The change from the low energy to the high energy mode is accompanied by a clear change in the slope of the counting process $\Lambda_t$ [0.95]{} [0.3]{} ![Rescaled potentials $U(n)/N_{\text{ex}}$ as function of $n/N_{\text{ex}}$, for various finite $N_{\text{ex}}$ converge to a limit potential for $N_{\text{ex}} \rightarrow \infty$. For $\alpha < 1$ the potential is minimum at zero; for $1<\alpha <4.6$ it has a unique minimum away from $n=0$; for $4.6<\alpha< 7.8$ there are two local minima which become equal at $\alpha \approx 6.66$.[]{data-label="fig:potentials"}](fig3potential1 "fig:"){width="\textwidth"} [0.3]{} ![Rescaled potentials $U(n)/N_{\text{ex}}$ as function of $n/N_{\text{ex}}$, for various finite $N_{\text{ex}}$ converge to a limit potential for $N_{\text{ex}} \rightarrow \infty$. For $\alpha < 1$ the potential is minimum at zero; for $1<\alpha <4.6$ it has a unique minimum away from $n=0$; for $4.6<\alpha< 7.8$ there are two local minima which become equal at $\alpha \approx 6.66$.[]{data-label="fig:potentials"}](fig3potential2 "fig:"){width="\textwidth"} [0.3]{} ![Rescaled potentials $U(n)/N_{\text{ex}}$ as function of $n/N_{\text{ex}}$, for various finite $N_{\text{ex}}$ converge to a limit potential for $N_{\text{ex}} \rightarrow \infty$. For $\alpha < 1$ the potential is minimum at zero; for $1<\alpha <4.6$ it has a unique minimum away from $n=0$; for $4.6<\alpha< 7.8$ there are two local minima which become equal at $\alpha \approx 6.66$.[]{data-label="fig:potentials"}](fig3potential3 "fig:"){width="\textwidth"} $$\frac{\mathbb{E} ( \Lambda_t) }{t} =N_{ex} \sum_n \rho_{\text{ss}}(n) \sin^2(\phi \sqrt{n+1})= N_{ex} \left( \sum_n n \rho_{\text{ss}}(n) - \nu\right).$$ Unlike the “first order transition occurring at $\alpha=6.66$, a ”second order transition occurs at $\alpha\approx 1$. Here the first derivative of the mean photon number has a jump in the limit of $N_{\text{ex}}\to\infty$. This and the scaling of the potential $U$ with $N_{\text{ex}}$ will be discussed in section \[sec.numerics\]. The statistics of the trajectories are therefore closely related to the dynamics of the cavity and consequently with its stationary state. The next step is to think of the time trajectories as “configurations” of stochastic system draw from ideas in non-equilibrium statistical mechanics and large deviations theory to study their phases and phase transitions. [0.95]{} [0.31]{} ![Sample trajectories for cavity state (top) and output paths $\Lambda_{t}$ (bottom) with $N_{\text{ex}}=50$, corresponding to stationary state distribution (center) showing large variance at $\alpha \approx 1$ (red) and bistability at $\alpha \approx 6.66$ (green)[]{data-label="fig:trajectories"}](fig4traj1 "fig:"){width="\textwidth"} [0.31]{} ![Sample trajectories for cavity state (top) and output paths $\Lambda_{t}$ (bottom) with $N_{\text{ex}}=50$, corresponding to stationary state distribution (center) showing large variance at $\alpha \approx 1$ (red) and bistability at $\alpha \approx 6.66$ (green)[]{data-label="fig:trajectories"}](fig4amssmeans "fig:"){width="\textwidth"} [0.31]{} ![Sample trajectories for cavity state (top) and output paths $\Lambda_{t}$ (bottom) with $N_{\text{ex}}=50$, corresponding to stationary state distribution (center) showing large variance at $\alpha \approx 1$ (red) and bistability at $\alpha \approx 6.66$ (green)[]{data-label="fig:trajectories"}](fig4traj2 "fig:"){width="\textwidth"} Large deviations ---------------- The main result of this paper is the existence of a *large deviations principle* for the counting process $\Lambda_{t}$ introduced above. Such results have already been obtained in the context of discrete time quantum Markov chains with finite dimensional systems [@Hiai2007], but the novelty here is that we consider a continuous time Markov process with an infinite-dimensional system. The physical motivation lies in the new approach to the study of phase transitions for open systems developed in [@Garrahan2009; @Garrahan2011]. Here the idea is to identify *dynamical phase transitions* of the open system, by analysing the statistics of jump trajectories in the long time (stationary) regime. The trajectories play an analogous role to the configurations of a statistical mechanics model at equilibrium. In this analogy, the parameter $s$ of the moment generating function can be seen as a “field” which biases the distribution of trajectories in the direction of active or passive trajectories by effectively changing the probability of a trajectory $\omega:= (i_1, t_1, \dots, i_n, t_n)$ by a factor $\exp(s \Lambda_t(\omega))$. When $\alpha$ is such that the stationary distribution is unimodal, the trajectories’ distribution changes smoothly from passive ones for $s<0 $ to active ones for $s>0 $. However, near $\alpha\approx 6.66$ (corresponding to the jump in the mean photon number) there is a steep change in the counting rates around $s=0$. The active trajectories are associated to periods when the cavity is in the higher, excited phase while the passive trajectories are connected to the lower phase. Since the cavity makes very rare transitions between the phases, any trajectory – when followed for long but finite periods of time – falls typically into one of the two distinct categories (see Figure \[fig:trajectories\]). Our goal is to investigate whether this distinction survives the infinite time limit, in which case we would deal with a dynamical phase transition characterised by the non-analyticity of a certain large deviations rate function. We will show that this is not the case, but rather we deal with a *cross-over* behaviour; that is, the count rate does not jump but has a very steep change around $s=0$, which appears to become a jump in the limit of infinite pumping rate $N_{ex} \to\infty$ (see Section 5). We will now briefly review some basic notions of large deviations needed in the paper. We refer the reader to [@Dembo2010] for a complete reference; see [@Ellis1995] for a comprehensive overview and [@Touchette2009] for an introduction to large deviations in the context of statistical mechanics. Large deviations is a framework for studying rare events, more precisely events whose probabilities decay exponentially for a sequence of probability distributions. A key result is the *Gärtner-Ellis theorem*, which relates the rate of the exponential decay to the limiting behaviour of the moment generating functions associated to the random variables. Informally, a sequence $(\mu_{n})_{n \in \mathbb{N}}$ of probability distributions on a topological space $(\mathcal{X}, \mathcal{B})$ satisfies a *large deviation principle* (LDP) if there exists a function $I : \mathcal{X} \rightarrow [0,\infty]$ such that $$\mu_{n}({\text{d}}x) \approx e^{-n I(x)} {\text{d}}x.$$ More rigorously, the function $I$ is called a *rate function* if it is lower semicontinuous (that is, its level sets ${ \left\lbrace x \in \mathcal{X} : I(x) \leq \alpha \right\rbrace }$ are closed); if in addition its level sets are compact, we call it a *good rate function*. The domain of $I$ is the set of points in $\mathcal{X}$ for which $I$ is finite. The limiting behaviour of the probability measures ${ \left\lbrace \mu_{n} \right\rbrace }$ is characterised in terms of asymptotic upper and lower bounds on the values that $\mu_{n}$ assigns to measurable subsets $\Gamma \in \mathcal{B}$. The sequence of probability measures ${ \left\lbrace \mu_{n} \right\rbrace }$ satisfies a large deviation principle with a rate function $I$ (or shortly, satisfies an LDP) if for all $\Gamma \in \mathcal{B}$, $$\label{eq:ldp.def} -\inf_{x \in \Gamma^{0}} I(x) \leq \liminf_{n \rightarrow \infty} \tfrac{1}{n} \log \mu_{n}(\Gamma) \leq \limsup_{n \rightarrow \infty} \tfrac{1}{n} \log \mu_{n}(\Gamma) \leq -\inf_{x \in \bar{\Gamma}} I(x).$$ Our goal is to prove an LDP for the counting process $\Lambda_{t}$ of the atom maser; we will do this not by showing that $\Lambda_{t}$ satisfies the above definition directly, but by applying the Gärtner-Ellis theorem, which gives sufficient conditions on the sequence of probability measures in order to satisfy an LDP. Let $(Z_{n})_{n\in \mathbb{N}}$ be a sequence of random variables in $\mathbb{R}^{d}$ with laws $\mu_{n}$. Suppose that the (limiting) logarithmic moment generating function $$\lambda({\bf s}) = \lim_{n \rightarrow \infty} \tfrac{1}{n} \log \mathbb{E}\left[ e^{\langle n {\bf s}, Z_{n} \rangle} \right],\quad {\bf s} \in \mathbb{R}^{d}$$ exists as an extended real number and is finite in a neighbourhood of the origin, and let $\Lambda^{\ast}$ denote the Fenchel-Legendre transform of $\Lambda$, given by $$\lambda^{\ast}(x) = \sup_{\lambda \in \mathbb{R}^{d}} { \left\lbrace \langle {\bf s}, x\rangle - \lambda({\bf s}) \right\rbrace }.$$ If $\Lambda$ is an essentially smooth, lower semicontinuous function (e.g. $\Lambda$ is differentiable on $\mathbb{R}^{d}$) then $(Z_{n})_{n\in \mathbb{N}}$ satisfies a LDP with good rate function $\Lambda^{\ast}$. The discrete index in the Gärtner-Ellis Theorem can be replaced by a continuous one with the obvious modifications in . By the Gärtner-Ellis Theorem, $\Lambda_t$ satisfies an LDP if the following limit exists and is a differentiable function, $$\label{eq:mfgconvergence} \lambda(s) := \lim_{t \rightarrow \infty} \frac{1}{t} \log \mathbb{E}\left(e^{s \Lambda_{t}}\right)=\lim_{t \rightarrow \infty} \frac{1}{t} \log {\text{Tr}}\left( \rho \mathcal{T}_s(t)(\mathbf{1})\right).$$ We will show that this is indeed true and $\lambda(s)$ is *spectral bound* (i.e. the eigenvalue with the largest real part) of a certain generator $L_{s}$ which is closely related to $\mathcal{L}_s$. An essential ingredient is the Krein-Rutman theorem, a generalisation of the Perron-Frobenius Theorem to compact positive semigroups which ensures that $\lambda(s)$ is real, and under additional conditions, non-degenerate. In particular our analysis shows that $\lambda(s)$ is smooth and its derivatives at $s=0$ are the limiting cumulants of $\Lambda_t$ $$\lim_{t\to\infty} \frac{1}{t} C_k(\Lambda_t) = \left. \frac{d^k \lambda(s)}{ds^k} \right|_{s=0}, \quad k\geq 1,$$ the first two being the mean and the variance. Moreover the generator $\mathcal{L}_s$ has a non-zero spectral gap; this spectral analysis is illustrated in Figure \[fig:grids\]. The main results ================ Our main results are the following Large Deviations and Central Limit theorems. For reader’s convenience we outline the key steps of the proofs below, followed by more technical details in the next section. \[th.main\] The limit $\lambda(s)$ in exists, and is a smooth function of $s$ equal to the spectral bound of a certain semigroup generator $L^{(d)}_s$ defined below. Therefore the counting process $\Lambda_{t}$ satisfies the large deviations principle with rate function equal to the Legendre transform of $\lambda(s)$. In particular the atom maser does not exhibit dynamical phase transitions, but rather cross-over transitions which become sharper as $N_{ex}$ increases. The counting process $\Lambda_{t}$ satisfies the Central Limit Theorem $$\frac{1}{t} ( \Lambda_t - t \cdot m) \overset{\mathcal{D}}{\longrightarrow} N(0, V),$$ where $\mathcal{D}$ denotes convergence in distribution and $m$ and $V$ are the mean and variance $$m = \left. \frac{d \lambda(s) }{ds}\right|_{s=0}, \qquad V = \left. \frac{d^2 \lambda(s) }{ds^2}\right|_{s=0}.$$ *Proof.* - The operator $$\mathcal{L}_{s}(X) = \Delta_s(X)+ \mathcal{L}(X) = (e^{s} -1) L_{1}^{\ast} X L_{1} +\sum_{i=1}^{4} \left( L_{i}^{\ast} X L_{i} - \frac{1}{2}\lbrace L_{i}^{\ast} L_{i}, X \rbrace \right).$$ is the generator of a strongly continuous semigroup $\left(\mathcal{T}_s(t)\right)_{t \geq 0}$ of $*$-weakly continuous CP maps $\mathcal{T}_s(t)$ on $\mathcal{B}(\mathfrak{h})$, and has the same domain $D(\mathcal{L})$ as $\mathcal{L}$. For $s=0$ the semigroup reduces to the “physical” (master) semigroup $\left(\mathcal{T}(t)\right)_{t \geq 0}$ with generator $\mathcal{L}$. The moment generating function of $\Lambda_t$ is given by $$\mathbb{E}\left(e^{s \Lambda_{t}}\right) = {\text{Tr}}\left( \rho_{\text{in}} \mathcal{T}_s(t)(\mathbf{1})\right).$$ where $\rho_{\text{in}}$ is the initial state of the cavity. - Define the following “representation” of $\left(\mathcal{T}(t)\right)_{t \geq 0}$ on $L^2(\mathfrak{h})$. Let $i$ be the symmetric embedding $$\begin{aligned} i : \mathcal{B}(\mathfrak{h}) &\rightarrow L^{2}(\mathfrak{h})\\ X &\mapsto \rho_{ss}^{1/4} X \rho_{ss}^{1/4}.\end{aligned}$$ Then ([@Carbone2000], Thm. 2.3) there exists a unique strongly continuous contraction semigroup $(T(t))_{t \geq 0}$ on $L^{2}(\mathfrak{h})$ such that $$T(t)(i(X)) = i(\mathcal{T}(t)(X)),\quad X \in \mathcal{B}(\mathfrak{h}).$$ For every $X \in \mathcal{D}(\mathcal{L})$ the Hilbert-Schmidt operator $i(X)$ belongs to the domain of the generator $L$ of $(T(t))_{t \geq 0}$ ([@Carbone2000], Prop. 3.2) and $$L(i(X)) = i(\mathcal{L}(X)),\quad X \in \mathcal{D}(\mathcal{L}).$$ Because we deal with a *reversible* quantum Markov process $L$ is a self-adjoint operator on the Hilbert space $L^2(\mathfrak{h})$. Moreover the set $\mathcal{M}(\mathfrak{h})$ of finite rank operators given by finite matrices with respect to the Fock basis form a core ([@Carbone2000], Thm. 3.3) for $L$. - Similarly, we define the embedded version $\left(T_s(t)\right)_{t \geq 0}$ of the semigroup $\left(\mathcal{T}_s(t)\right)_{t \geq 0}$. This is a semigroup with generator $L_s = L+ \delta_s$, with $\delta_s$ a bounded perturbation. The domain of $L_s$ coincides with that of $L$ and $$L_s (i(X)) = i(\mathcal{L}_s(X)).$$ - The moment generating function of $\Lambda_t$ can be expressed in terms of the embedded semigroup as (see Lemma \[lemma:transfer.mgf\]) $$\mathbb{E}(e^{s \Lambda_{t}}) = {\text{Tr}}( \rho_{\text{in}} \mathcal{T}_{s}(t)(\mathbf{1})) ={\text{Tr}}( \tilde{\rho}_{\text{in}}T_s(t)(\rho_{ss}^{1/2}) ) =\langle \tilde{\rho}_{\text{in}}, T_{s}(t)(\rho_{ss}^{1/2}) \rangle_{\text{HS}},$$ where $\tilde{\rho}_{\text{in}}:= \rho_{ss}^{-1/4} \rho_{\text{in}} \rho_{ss}^{-1/4}$ is assumed to belong to $L^2(\mathfrak{h})$. This holds for instance if $\rho_{\text{in}}$ has a finite number of photons. - The semigroup $\left(T_s(t)\right)_{t \geq 0}$ leaves invariant the subspace $L^2_d(\mathfrak{h})\subset L^2(\mathfrak{h})$ of Hilbert-Schmidt operators which are diagonal in the Fock basis, and we denote its restriction to this subspace by $\left(T_s^{(d)}(t)\right)_{t \geq 0}$, and similarly for the generator. Since $\rho_{ss}\in L^2_d(\mathfrak{h}) $ the moment generating function can be expresses as $$\mathbb{E}(e^{s \Lambda_{t}}) = \langle \tilde{\rho}^{(d)}_{\text{in}}, T^{(d)}_{s}(t)(\rho_{ss}^{1/2}) \rangle_{\text{HS}}$$ with $ \tilde{\rho}^{(d)}_{\text{in}}$ denoting the diagonal of $ \tilde{\rho}_{\text{in}}$. The analogous restriction of $\left(\mathcal{T}(t)\right)_{t \geq 0}$ to the diagonal sub-algebra of $\mathcal{B}(\mathfrak{h})$ can be identified with the transition semigroup of the birth-death process on $\mathbb{N}$ with rates . - The restricted generator $L^{(d)}$ has compact resolvent, and $\left(T^{(d)}(t)\right)_{t \geq 0}$ is immediately compact (i.e. $T^{(d)}(t)$ is compact for all $t>0$). Moreover, the semigroup $\left(T^{(d)}_s(t)\right)_{t \geq 0}$ is also immediately compact. - The semigroup $(T^{(d)}_{s}(t))_{t \geq 0}$ is strictly positive, that is $T^{(d)}_{s}(t)(D)>0$ for all operators $D\geq 0$ in $L^2_d(\mathfrak{h})$ and $t\geq 0$. - Since $\left(T_{s}(t)\right)_{t \geq 0}$ is compact and strictly positive, the Krein-Rutman theorem implies that the spectral radius of $T_{s}(t)$ is an eigenvalue with strictly positive right and left eigenvectors $r(s)$ and $l(s)$ . The spectral radius is equal to $e^{t\lambda(s)}$ where $\lambda(s)$ is the spectral bound of $L_s$, i.e. the eigenvalue with the largest real part. Using point 4. this implies that $$\begin{aligned} \mathbb{E} (e^{s \Lambda_t} ) &=& \langle \tilde{\rho}_{\text{in}}, T_{s}(t)(\rho_{ss}^{1/2}) \rangle_{\text{HS}} \nonumber\\ &=& e^{t\lambda(s)} \left( \langle \tilde{\rho}_{\text{in}}, r(s) \rangle_{HS}\langle l(s), \rho_{ss}^{1/2} \rangle_{HS} + o(t)\right). \label{eq.dominant.term}\end{aligned}$$ Since $r(s),l(s) >0$ and $\tilde{\rho}_{\text{in}}, \rho_{ss}^{1/2} \geq 0$ the inner products are non-zero and we obtain the limiting cumulant generating function $$\lim_{t\to\infty} \frac{1}{t} \log \mathbb{E} (e^{s \Lambda_t} ) = \lambda(s).$$ - Using analytic perturbation theory, the spectral bound $\lambda(s)$ is shown to be a smooth function of $s$. - Using points 8. and 9., we apply the Gärtner-Ellis theorem to conclude that $\Lambda_t$ satisfies the LD principle with rate function equal to the Legendre transform of $\lambda(s)$. In particular, the limiting cumulants of $\Lambda_t$ can be computed as derivatives of $\lambda(s)$ at $s=0$, $$\lim_{t\to\infty} \frac{1}{t} C_k(\Lambda_t) = \left. \frac{d^k \lambda(s)}{ds^k}\right|_{s=0}.$$ - Again by analytic perturbation theory, the spectral bound is analytic in the neighbourhood of the origin of the complex plane. By the result of [@Bryc], it follows that $\Lambda_t$ satisfies the Central Limit Theorem. Details of proof ================ Here we give point by point details on the steps of the proof. 1\. The strongly continuous semigroup $(\mathcal{T}(t))_{t \geq 0}$ of $w^{\ast}$-continuous, identity preserving CP maps on $\mathcal{B}(\mathfrak{h})$ was analysed in [@Fagnola1994]. Since $\Delta_s$ is bounded, by the bounded perturbation theorem (cf. [@Engel2001], Theorem III 1.3) the operator $$\mathcal{L}_{s}(X) = \Delta_s (X)+ \mathcal{L}(X)$$ is the generator of a strongly continuous semigroup $(\mathcal{T}_s (t))_{t \geq 0}$, and has the same domain $\mathcal{D}(\mathcal{L}_{s}) = \mathcal{D}(\mathcal{L})$. 2\. Denote $E_{i,j}:= |e_i\rangle \langle e_j|$ and $E_i:= |e_i\rangle \langle e_i|$ the rank one “matrix elements”. For any $X= \sum_{jk} x_{j,k}E_{j,k}$ in the domain, the explicit action of generator is [@Carbone2000] $$\begin{aligned} L(X) &= -\frac{1}{2}\sum_{j,k \geq 0} \left( \lambda_{j}^{2} + \lambda_{k}^{2} + \mu_{j}^{2} + \mu_{k}^{2} \right) x_{j,k} E_{j,k}\\ &+\sum_{j,k \geq 0} \sqrt{\lambda_{j} \lambda_{k} \mu_{j+1} \mu_{k+1}} \left( x_{j+1,k+1} E_{j,k}+ x_{j,k} E_{j+1,k+1}\right).\end{aligned}$$ where the coefficients $\lambda_{n}, \mu_{n}$ are the rates of the associated birth-and-death process . It can be directly checked that $L$ is symmetric operator but proving its selfadjointness is non-trivial and holds only under certain assumption about $\lambda_{n}, \mu_{n}$ [@Carbone2000]. *Point 4.* is proved in the following lemma. \[lemma:transfer.mgf\] The moment generating function of $\Lambda_s$ can be expressed in terms of the embedded semigroup $\left(T_s(t)\right)_{t \geq 0}$ as follows $$\mathbb{E}(e^{s \Lambda_{t}}) = {\text{Tr}}( \rho_{\text{in}} \mathcal{T}_{s}(t)(\mathbf{1})) ={\text{Tr}}( \tilde{\rho}_{\text{in}} e^{t L_{s}}(\rho_{ss}^{1/2}) ) =\langle \tilde{\rho}_{\text{in}}, T_{s}(t)(\rho_{ss}^{1/2}) \rangle_{\text{H-S}},$$ where $\tilde{\rho}_{\text{in}}:= \rho_{ss}^{-1/4} \rho_{\text{in}} \rho_{ss}^{-1/4}\in L^2(\mathfrak{h})$. We first note that the linear span $\mathcal{M}$ of the matrix units $E_{j,k}$ are analytic vectors for $\mathcal{L}_s$, i.e. there exists a time $T>0$ such that for all $0 \leq t < T$, the series $$\sum_{k \geq 0} \frac{t^{k} \mathcal{L}_{s}^{k}(x)}{k!} , \qquad x\in \mathcal{M}.$$ converges in norm in $\mathcal{B}(\mathfrak{h})$ and the limit is $\mathcal{T}_s(t)(x)$. The proof is similar to that of Lemma 5.4 in [@Cipriani2000] to which we refer for details. A similar statement holds for the generator $L_s$ on $L^2(\mathfrak{h})$. We now define the truncation $$X_{n} = P_{n} X P_{n},$$ of an arbitrary operator $X\in \mathcal{B}(\mathfrak{h})$ to the finite dimensional space spanned by the first $n$ Fock basis vectors, whose orthogonal projection is $P_{n}$. Recall that the action of $T_s(t)= e^{t \mathcal{L}_{s}}$ and its predual semigroup are related by $${\text{Tr}}(e^{t \mathcal{L}_{s \ast}}(\rho) X) = {\text{Tr}}( \rho e^{t \mathcal{L}_{s}}(X) ), \qquad \rho \in L^{1}(\mathfrak{h}), X \in \mathcal{B}(\mathfrak{h})$$ where $\mathcal{L}_{s \ast}$ is the generator of $\left(T_{s*}(t)\right)_{t \geq 0}$. We want to show that $$\label{eq:transfer.mgf} {\text{Tr}}( \rho_{\text{in}} e^{t \mathcal{L}_{s}}(\mathbf{1})) = {\text{Tr}}( \tilde{\rho}_{\text{in}} e^{t L_{s}} (\rho_{\text{ss}}^{1/2}) )$$ for all $t>0$. Partition the time $t$ by writing $t = t_{1} + \ldots + t_{l}$ where $t_{i} < T$. The main idea behind showing the above equality is by applying sequentially the analyticity of finite-rank operators and weak-$\ast$ continuity to ‘move over’ each element of the semigroup $e^{t_{i} \mathcal{L}}$ to the trace-class operators and then the Hilbert-Schmidt operators, and repeat this until the entire semigroup lives on the other side. Using the projections $P_{n}$ introduced above, we find that by weak-$\ast$ continuity and the semigroup property we have $${\text{Tr}}\left( \rho_{\text{in}} e^{(t_{1} + \ldots + t_{l}) \mathcal{L}_{s}}(\mathbf{1})\right) = \lim_{n \rightarrow \infty} {\text{Tr}}\left( \rho_{\text{in}} e^{t_{1} \mathcal{L}_{s}}\left( P_{n} e^{(t_{2} + \ldots + t_{l}) \mathcal{L}_{s}}(\mathbf{1})P_{n}\right) \right).$$ Let $A:= e^{(t_{2} + \ldots + t_{l}) \mathcal{L}_{s}}(\mathbf{1})\in \mathcal{B}(\mathfrak{h}) $ and $ \tilde{A}:= \rho_{\text{ss}}^{1/4} A\rho_{\text{ss}}^{1/4}\in L^2(\mathfrak{h})$ and define the finite-rank truncations $A_{n} = P_{n} AP_{n}$, and $\tilde{A}_n = P_{n} \tilde{A}P_{n}$. Since $A_n$ is an analytic vector for $\mathcal{L}_s$, we may express the RHS as a power series and thus we obtain $$\lim_{n \rightarrow \infty} {\text{Tr}}\left( \rho_{\text{in}} \sum_{k \geq 0} \frac{t_{1}^{k} \mathcal{L}_{s}^{k}}{k!}(A_{n}) \right) = \lim_{n \rightarrow \infty} {\text{Tr}}\left( \rho_{ss}^{-1/4} \rho_{\text{in}} \rho_{ss}^{-1/4} \sum_{k \geq 0} \frac{t_{1}^{k} L_{s}^{k}}{k!}(\tilde{A}_{n} )\right)$$ by definition of the embedded generator $L_{s}$. Now $\tilde{A}_{n} $ is analytic for $L_s$ so the inner sum is equal to $ e^{t_1L_s} (\tilde{A}_{n} ) $ and the limit may be written as $$\begin{aligned} \lim_{n \rightarrow \infty} {\text{Tr}}\left( \tilde{\rho}_{\text{in}} e^{t_{1} L_{s}}\left( \tilde{A}_n \right) \right) &= \lim_{n \rightarrow \infty} \langle \tilde{\rho}_{\text{in}}, e^{t_{1} L_{s}} (\tilde{A}_{n}) \rangle_{HS}\\ &= \lim_{n \rightarrow \infty} \langle e^{t_{1} L_{s}^*}(\tilde{\rho}_{\text{in}}), \tilde{A}_{n} \rangle_{HS}\\ &= \langle e^{t_{1} L_{s}^*}(\tilde{\rho}_{\text{in}}), \tilde{A} \rangle_{HS}. \end{aligned}$$ The last equality follows from the fact that $\tilde{A}_{n}$ converges weakly to $\tilde{A}$ in $L^2(\mathfrak{h})$. Indeed for any positive $\tau \in L^2(\mathfrak{h})$ we have $$\begin{aligned} \langle \tau, P_{n} \rho_{ss}^{1/4} A \rho_{ss}^{1/4} P_{n} \rangle_{HS} &= {\text{Tr}}\left( \tau P_{n} \rho_{ss}^{1/4} A \rho_{ss}^{1/4} P_{n} \right)\\ &= {\text{Tr}}\left( \rho_{ss}^{1/4} \tau \rho_{ss}^{1/4} P_{n} A P_{n}\right)\\ &\rightarrow {\text{Tr}}\left( \rho_{ss}^{1/4} \tau \rho_{ss}^{1/4} A \right); \end{aligned}$$ taking into account that $\rho_{ss}^{1/4} \tau \rho_{ss}^{1/4} \in L^1(\mathfrak{h})$ by $${\text{Tr}}\left( \rho_{ss}^{1/4} \tau \rho_{ss}^{1/4} \right) = {\text{Tr}}\left( \rho_{ss}^{1/2} \tau \right) = \langle \rho_{ss}^{1/2}, \tau \rangle_{HS} < \infty.$$ We now repeat the same argument for the term $e^{t_2\mathcal{L}_s}$, by defining $B = e^{(t_{3} + \ldots + t_{l}) \mathcal{L}_{s}}(\mathbf{1})$, and $\tilde{B}, B_n, \tilde{B}_n$ as before. Then $$\begin{aligned} {\text{Tr}}( \rho_{\text{in}} e^{t \mathcal{L}_{s}}(\mathbf{1})) &={\text{Tr}}\left( \tilde{\rho}_{\text{in}} e^{t_{1} L_{s}} \left( \rho_{ss}^{1/4} e^{(t_{2} + \ldots + t_{l}) \mathcal{L}_{s}}(\mathbf{1}) \rho_{ss}^{1/4} \right) \right)\\ &= \lim_{n \rightarrow \infty} {\text{Tr}}\left( e^{t_{1} L_{s \ast}}(\tilde{\rho}_{\text{in}}) \rho_{ss}^{1/4} e^{t_{2} \mathcal{L}_{s}}\left(P_{n} B P_{n}\right) \rho_{ss}^{1/4}\right)\\ &= \langle e^{(t_{1}+t_2) L_{s \ast}}(\tilde{\rho}_{\text{in}}) , \tilde{B} \rangle_{HS} \end{aligned}$$ and after a finite number of steps we arrive at . This shows that the semigroup $L_{s}$ on the Hilbert-Schmidt space gives rise to the desired expectation values. *Point 5.* follows immediately from the definition of the generator $\mathcal{L}_s$. *Point 6.* is shown in the following lemma. \[lemma:compactness\] The restricted generator $L^{(d)}$ has compact resolvent, and $\left(T^{(d)}(t)\right)_{t \geq 0}$ is immediately compact, i.e. $T^{(d)}(t)$ is compact for all $t>0$. Moreover, the semigroup $\left(T^{(d)}_s(t)\right)_{t \geq 0}$ is also immediately compact. Let $L^{(d)}$ be the restriction of $L$ to the subspace $L^2_d(\mathfrak{h})$ of diagonal Hilbert-Schmidt operators with respect to the Fock basis. Its concrete action on a diagonal operator $ D = \sum_{j \geq 0} d_{j}E_j$ is $$\begin{aligned} L(D) &= -\sum_{j \geq 0} \left( \lambda_{j}^{2} + \mu_{j}^{2}\right) d_{j} E_j +\sum_{j \geq 0} \lambda_{j} \mu_{j+1} \left( d_{j+1} E_j+ d_{j} E_{j+1}\right)\\ & :=A(D)+ B(D).\end{aligned}$$ In order to establish that $L^{(d)}$ has compact resolvent, we extend a similar argument used in [@Janas2001] to our setting. Note that $A$ is a selfadjoint operator with (point) spectrum $$\sigma_{p}(A) = \lbrace a_j:= -( \lambda_{j}^{2} + \mu_{j}^{2}) : j \in \mathbb{N}\rbrace.$$ The resolvent operator $R(z,A) = (A - z\text{Id})^{-1}$ is well defined whenever $z \notin \sigma_{p}(A)$ and is given by $$(A - z \text{Id})^{-1}(X) = \sum_{j \geq 0}(a_{j}-z)^{-1}x_{j}E_j $$ Since ${\lvert a_{j} \rvert} \rightarrow \infty$ when $j \rightarrow \infty$, $R(z,A)(X)$ may be approximated in the $L^{2}$-norm by its finite-rank truncations $$R(z,A)(X) = \lim_{N \rightarrow \infty} \sum_{j=0}^{N}(a_{j} -z)^{-1}x_{j} E_j , \quad X \in \mathcal{N}$$ and so $A$ has compact resolvent. To show that $L^{(d)}=A+B$ itself has compact resolvent it is enough to show that $B$ is a *relatively bounded* perturbation of $A$ ([@Kato1976] Thm. 3.17, p. 214). This means showing that there exists a $b >0$ such that $$\text{Tr} \left( B(X)^{\ast} B(X)\right) \leq b \text{Tr}\left(A(X)^{\ast}A(X)\right)\quad\text{for all }X \in L^2_d(\mathfrak{h})$$ Using the Cauchy-Schwarz inequality in the Hilbert space $L^{2}(\mathfrak{h})$ such a bound is readily found, and we may conclude that $L$ has compact resolvent. We now show that $T^{(d)}(t)$ is immediately compact. Since $L^{(d)}$ is self-adjoint, we find that its resolvent operator $R(z,L)$ satisfies the bound ([@Kato1976] Thm. 3.16, p. 271) $${\lVert R(iz,L) \rVert}\leq {\lvert \text{Im} (z) \rvert}^{-1}\quad\text{for all Im}(z) \neq 0.$$ Therefore ([@Engel2001] Thm. II 4.20 p. 115) the semigroup $(T^{(d)}(t))_{t \geq 0}$ is *immediately norm continuous*. An immediately norm continuous semigroup whose generator has compact resolvent is immediately compact ([@Engel2001] Thm. 4.29, p. 119), therefore the semigroup $(T^{(d)}(t))_{t \geq 0}$ is immediately compact. Similarly, the restriction $L^{(d)}_{s}$ is the perturbation of the generator $L^{(d)}$ $$\begin{aligned} L^{(d)}_{s}(D) &= L^{(d)}(D) + \delta_{s}(D),\\ \delta_{s}(D) &= (e^{s}-1) N_{\text{ex}} \sum_{j \geq 0} \sin^2(\phi \sqrt{j+1}) \frac{\mu_{j+1}}{\lambda_{j}} d_{j+1}E_j $$ Since $\delta_{s}$ is bounded, the semigroup $\left(T^{(d)}_s(t)\right)_{t \geq 0}$ is also immediately compact, cf. [@Engel2001] Thm. III.1.16. *Point 7.* We show first that the unperturbed semigroup $(T(t))_{t \geq 0}$ on $L^2_d(\mathfrak{h})$ is strictly positive. Since any positive $D\in L^2_d(\mathfrak{h})$ is of the form $D= \sum_k d_k E_k$ with $d_k \geq 0$, it is enough to show that $T^{(d)}(t) (E_i) >0$. This is equivalent to $$\langle E_j , T^{(d)}(t) (E_i) \rangle_{HS} >0, \qquad i,j\in \mathbb{N}.$$ By using the technique of Lemma \[lemma:transfer.mgf\] we find $$\langle E_j , T^{(d)}(t) (E_i) \rangle_{HS} = {\rm Tr} (E_j \mathcal{T}^{(d)}(t) (E_i) ) = {\rm Tr} (E_i \mathcal{T}^{(d)}_*(t) (E_j) )$$ where the right side is the probability $\mathbb{P}_{j,i}(t)$ of going from state $j$ to $i$ in time $t$, for the associated birth and death process. This probability can be unravelled as $$\begin{aligned} \mathbb{P}_t(j,i) &=& \sum_{k\geq 0} \, \sum_{i_1,\dots i_k =1}^4 \underset{0\leq t_1\leq \dots \leq t_k\leq t}{\int \dots\int } p_{j,i}(t; t_1,i_1\dots ,t_k, i_k)dt_1 \dots dt_k $$ where $$p_{j,i}(t; t_1,i_1,\dots, t_k, i_k):= {\rm Tr} \left( E_j \cdot e^{(t-t_k) \mathcal{L}_0} \mathcal{J}_{i_{k}} \dots \mathcal{J}_{i_1} e^{t_1 \mathcal{L}_0} (E_i)\right),$$ is the probability density for the trajectory consisting of jumps of type $i_1, \dots i_k$ occurring at times $0\leq t_1\dots \leq t_k\leq t$, respectively. To show strict positivity, we can restrict our attention to trajectories which exhibit only jumps of type $3$ and $4$ and connect the states $j$ and $i$. Since the jump rates are strictly positive, the probability of such a trajectory is strictly positive and therefore $\mathcal{T}^{(d)}(t)$ and $T^{(d)}(t)$ are strictly positive. The same argument can be repeated for the semigroup $\left(T^{(d)}_s(t)\right)_{t \geq 0}$ whose unravelling is $$\begin{aligned} \langle E_j , T^{(d)}_s(t) (E_i) \rangle_{HS} &=& \sum_{k\geq 0} \, \sum_{i_1,\dots i_k =1}^4 \underset{0\leq t \dots \leq t_k\leq t}{\int \dots\int } p^{(s)}_{j,i}(t; t_1,i_1\dots ,t_k,i_k) dt_1 \dots dt_k\\ $$ where $$p^{(s)}_{j,i}(t; t_1,i_1\dots ,t_k,i_k) = e^{s n(1)} p_{j,i}(t; t_1,i_1,\dots, t_k, i_k),$$ with $n(1)$ equal to the number of jumps of type 1. *Point 8.* Recall that to establish the LDP for the counting process $\Lambda_{t}$ it is enough to show that the limit $$\lim_{t \rightarrow \infty} \frac{1}{t} \log \mathbb{E}\left( e^{s \Lambda_{t}} \right) = \lim_{t \rightarrow \infty} \frac{1}{t} \log {\text{Tr}}( \tilde{\rho}_{\text{in}} e^{t L_{s}} (\rho_{\text{ss}}^{1/2}) )$$ exists and is a smooth function of $s$. Since $(T^{(d)}_{s}(t))$ is an immediately compact semigroup, we have ([@Engel2001], Col. IV.3.12) a spectral mapping theorem of the form $$e^{t \sigma(L^{(d)}_{s})} = \sigma(T^{(d)}_{s}(t)) \setminus { \left\lbrace 0 \right\rbrace },\quad t > 0;$$ in particular, the *spectral radius* of $T^{(d)}_{s}(t)$ is given by $$r_s(t):= r(T^{(d)}_{s}(t)) = e^{t \lambda(s)}.$$ where $\lambda(s)$ is the *spectral bound* of $L^{(d)}_{s}$,i.e. the real part of the eigenvalue with the largest real part. Since $T^{(d)}_s(t)$ is compact and strictly positive, the Krein-Rutman Theorem implies that $\lambda(s)$ is a real eigenvalue with unique strictly positive right and left eigenvectors $r(s)$ and $l(s)$ such that $\langle l(s) | r(s)\rangle=1$ . In particular $L_s$ has a spectral gap $g(s) = \lambda(s)- {\rm Re} \lambda_1(s)$ and $$T^{(d)}_s(t) (D) = e^{t\lambda(s)} |r(s) \rangle \langle l(s)| + R(t)$$ where the reminder term satisfies $\| R(t)\| \leq C e^{t(\lambda(s)-g+ \epsilon)}$ for some constant $C$ and $\epsilon< g$. Therefore holds. *Point 9.* To complete the proof we need to show that $\lambda(s)$ is a differentiable function of $s$. This follows from analytic perturbation theory for the generator $L^{(d)}_{s}$, cf. [@Chatelin1983] (Prop. 3.25, p. 141); any isolated eigenvalue (of finite multiplicity) and its associated eigenprojection are analytic functions of $s$ in some disc around $s=0$. Applied to the family of perturbations $$L^{(d)}_{s} = L^{(d)} + \delta_{s},$$ we find that the spectral bound of $\lambda_(s)$ is an analytic function of $s$, and remains isolated as a function of $s$. Numerical analysis {#sec.numerics} ================== [0.45]{} ![$\lambda'(s)$ and the spectral gap $g(s)$ of $L_{s}$ as functions of $s$ and $\alpha = \phi /\sqrt{N_{\text{ex}}}$ (after Fig. 3 in [@Garrahan2011]).[]{data-label="fig:grids"}](fig5activity "fig:"){width="\textwidth"} [0.45]{} ![$\lambda'(s)$ and the spectral gap $g(s)$ of $L_{s}$ as functions of $s$ and $\alpha = \phi /\sqrt{N_{\text{ex}}}$ (after Fig. 3 in [@Garrahan2011]).[]{data-label="fig:grids"}](fig5gap "fig:"){width="97.50000%"} The existence of a “phase transition” in the atom maser has been discussed in several theoretical physics papers [@Briegel1994; @Englert2002a; @Benson1994; @Rempe1990; @Garrahan2011]. There is a general agreement that if $N_{ex}$ is sufficiently large (for instance $N_{ex}\approx 150$ ), then “for all practical purposes” we can consider that the mean photon number of the stationary state has a jump at $\alpha\approx 6.66$ (see Figure \[fig:stationary\]) which matches up with a jump between the left and right derivatives of $\lambda(s)$ at $s=0$, in the dynamical scenario (see Figure \[fig:grids\]). However, the question whether we are dealing with a “true” (dynamical) phase transition or rather a steep but smooth cross-over was left open, and motivated this investigation. Having proved that the latter is the case, we would like to briefly put the result in the context of a numerical analysis. As the proof shows, dynamical phase transitions are intimately connected with the closing of the spectral gap of the semigroup generator. Figure \[fig:grids\] shows the close match between the behaviour of the first derivative of $\lambda(s)$ and the spectral gap $g(s):=\lambda(s)- {\rm Re} \lambda_1(s)$. In particular, at first sight it would appear that for $\alpha\geq 4.6$ (the point where the stationary state becomes bistable), the entire $s=0$ line is a phase separation line. However, by zooming in a vertical strip of size $10^{-7}$ in this region (see Figure \[fig:phaseboundaries\]), we find that the line separating the phases is not perfectly vertical but crosses $s=0$ at $\alpha\approx 6.6$ which corresponds roughly to the transition point for the stationary state. Moreover, on this scale it is clear that we deal with a steep but smooth transition between phases. Figure \[fig:phaseboundaries\] shows that the phase separation lines become sharper with larger $N_{ex}$, and a “true” phase transition emerges in infinite pumping rate limit. A similar conclusion can be drawn by plotting the rescaled stationary mean $\langle N\rangle / N_{ex}$, cf. Figure \[fig:rescaledmeans\]. This can be intuitively understood by appealing to the effective potential . As $N_{ex}$ increases the potential barrier becomes larger and two stable phases emerge at the point where the local minima are equal. Indeed, Figure \[fig:potentials\] shows the plot of the rescaled potential $U/N_{ex}$ as a function of the rescaled variable $x=n/N_{ex}$, which approaches the ($N_{ex}$ independent) limit $$v(x) = - \int_0^x \log [ ( \nu+ \sin^2(\alpha \sqrt{y} ) )/(\nu+1)] dy$$ as it can be deduced from the formulas and . Therefore, in the limit of large pumping rate we deal with a particle in a fixed potential $v(x)$ at inverse temperature $1/kT= N_{ex}$. At $\alpha\approx 1$ the dependence of the mean on $N_{\text{ex}}$ switches from constant to linear behaviour as the minimum of the potential $v(x)$ moves away from zero. When the two minima are at different heights, the lower one becomes the stable and other one is metastable. Communication between the phases becomes increasingly unlikely, with probability decreasing exponentially with $N_{ex}$. When the two minima are equal, we have two stable states, and the corresponding value of $\alpha$ is the phase transition point for the mean photon number. [0.95]{} [0.31]{} ![Phase boundaries at the $s=0, \alpha \approx 6.66$ crossover with $N_{ex} = 75,100$ and $125$.[]{data-label="fig:phaseboundaries"}](fig7crossing1 "fig:"){width="\textwidth"} [0.31]{} ![Phase boundaries at the $s=0, \alpha \approx 6.66$ crossover with $N_{ex} = 75,100$ and $125$.[]{data-label="fig:phaseboundaries"}](fig7crossing2 "fig:"){width="\textwidth"} [0.31]{} ![Phase boundaries at the $s=0, \alpha \approx 6.66$ crossover with $N_{ex} = 75,100$ and $125$.[]{data-label="fig:phaseboundaries"}](fig7crossing3 "fig:"){width="\textwidth"} [0.65]{} ![Rescaled stationary state mean photon numbers, $\langle \rho_{\text{ss}} \rangle /N_{\text{ex}}$ for increasing $N_{\text{ex}}$, showing phase transition becomes sharp as $N_{\text{ex}} \rightarrow \infty$.[]{data-label="fig:rescaledmeans"}](fig6rescaledmeans "fig:"){width="95.00000%"} [0.95]{} [0.54]{} ![Spectrum of semigroup generator for $N_{ex} = 150$, and behaviour of ‘spectral density’ (blue) and spectral gap (red) at the critical points $\alpha=1$ (top) and $\alpha=6.66$ (bottom) as $N_{ex}$ increases.[]{data-label="fig:spectrum"}](fig8spectrum "fig:"){width="\textwidth"} [0.4]{} ![Spectrum of semigroup generator for $N_{ex} = 150$, and behaviour of ‘spectral density’ (blue) and spectral gap (red) at the critical points $\alpha=1$ (top) and $\alpha=6.66$ (bottom) as $N_{ex}$ increases.[]{data-label="fig:spectrum"}](fig8amspectraldensity "fig:"){width="\textwidth"} More information about the dynamical phase transitions may be obtained from the rest of the spectrum of the semigroup generator, shown in Figure \[fig:spectrum\]. We note several points of interest: firstly, at all “transition points” the spectral gap closes; but at the second order transition point $\alpha\approx 1$ the gap closes considerably more slowly than at the other transition points $\alpha \approx 6.66, \ldots$. Secondly, at $\alpha\approx 1$ the spectrum becomes increasingly dense, in the sense that a fixed interval contains an increasing number of eigenvalues as $N_{ex}$ increases. Although the mean of the counting process $\Lambda_{t}$ coincides with the mean of the stationary state photon number distribution, as we show in Figure \[fig:variance\], this is not the case with the respective variances. The critical point $\alpha\approx 1$, associated with an increasing density of the spectrum of the generator and closing of the spectral gap, also exhibits a change in the scaling of the cumulants with $N_{ex}$ (for a more in-depth treatment of these numerical aspects, we refer the reader to [@vanHorssen]). [0.95]{} [0.47]{} ![Variance of stationary state (left) and variance of counting process (right), rescaled by factors of $N_{ex}^{-1}$ and $N_{ex}^{-1.6}$, respectively[]{data-label="fig:variance"}](fig9amssvariance "fig:"){width="\textwidth"} [0.47]{} ![Variance of stationary state (left) and variance of counting process (right), rescaled by factors of $N_{ex}^{-1}$ and $N_{ex}^{-1.6}$, respectively[]{data-label="fig:variance"}](fig9amcountsvariance "fig:"){width="\textwidth"} Conclusions and outlook ======================= We have studied the counting process associated to the measurement of the outgoing atoms in the atom maser, and shown that this process satisfies the large deviations principle. In particular, this means that the cross-over behaviour observed in numerical simulations is not associated with the non-analyticity of the limiting log-moment generating function, as one would expect for a genuine phase transition. The rescaled counting process $\Lambda_{t}/N_{\text ex}$ does exhibit such a transition in the limit of infinite rate $N_{\text ex}$, as argued in the previous section using the potential model, and illustrated in Figures \[fig:potentials\], \[fig:phaseboundaries\], and \[fig:rescaledmeans\]. In particular, the transitions occurring at $\alpha\approx 1$ and $\alpha \approx 6.66$ are of different types, as seen in the scaling of the moments as well as the behaviour of the spectrum at these points. As a corollary, we have showed that the counting process satisfies the central limit theorem, which can be used to develop the statistical estimation theory of local asymptotic normality [@Guta2012]. The model we have investigated has the property that the stationary state is diagonal in the Fock basis and all the jump operators leave the set of diagonal states invariant. The large deviations problem could then be considered in the framework of “classical” probability, as a property of the birth-death process process associated to the cavity dynamics. Indeed, at one point our proof relies on the restriction to the diagonal algebra for proving the strict positivity of the dynamical semigroup. However, the steps of the proof are formulated in the language of non-commutative probability theory and offer a general recipe for other settings where no classical reduction is possible. An example would be the atom maser where the outgoing atoms are measured in a different basis than the standard one, thus breaking the invariance of the diagonal algebra. In this case, using analytic perturbation theory one can show that the limiting log-moment generating function $\lambda(s)$ is smooth in a neighbourhood of $s=0$ but we were not able to extend this to all $s$. The compactness of the Markov semigroup makes our model tractable as it becomes essentialy finite dimensional, as the bath decay dominates the absorption due to the atom interaction. An interesting problem would be to explore more general classes of infinite dimensional systems (e.g. continuous variables or infinite spin chains) where a similar phenomenon holds. Another issue is the general relation between the “static” transitions which refer to non-analytic properties of the stationary state, and dynamic transitions which characterise properties of the measurement process. As shown in [@LesanovskyvanHorrsenGutaGarrahan] one can construct examples where the stationary state does not change while the system undergoes a dynamical phase transition. Finally, a more general large deviations setup can be considered which takes into account the correlations between the detection events rather than the total number of counts [@vanHorrsenLesanovskyGarrahanGuta]. *Acknowledgements.* We thank J. P. Garrahan and I. Lesanovsky for numerous fruitful discussions. This work was supported by EPSRC Grant no. EP/J009776/1 and Fellowship EP/E052290/1.

--- abstract: 'Deployment of high-penetration photovoltaic (PV) power is expected to have a range of effects – both positive and negative – on the distribution grid. The magnitude of these effects may vary greatly depending upon feeder topology, climate, PV penetration level, and other factors. In this paper we present a simulation study of eight representative distribution feeders in three California climates at PV penetration levels up to 100%, supported by a unique database of distributed PV generation data that enables us to capture the impact of PV variability on feeder voltage and voltage regulating equipment. When comparing the influence of feeder location (i.e. climate) versus feeder type on outcomes, we find that location more strongly influences the incidence of reverse power flow, reductions in peak loading and the presence of voltage excursions. On the other hand, we find that feeder characteristics more strongly influence the magnitude of loss reduction and changes in voltage regulator operations. We find that secondary distribution transformer aging is negligibly affected in almost all scenarios.' author: - 'Michael A. Cohen and Duncan S. Callaway[^1][^2]' bibliography: - 'IEEEabrv.bib' - 'cohen\_callaway\_journal\_physical.bib' title: 'Physical Effects of Distributed PV Generation on California’s Distribution System' --- power distribution, photovoltaic systems, power system simulation. Introduction ============ As the deployment of distributed photovoltaics (PV) accelerates, researchers and power industry professionals have increasingly attended to the impacts – both positive and negative – that PV might have on the distribution system. Areas of concern include PV’s effect on [@katiraei2011solar]: - System losses - Peak load (which impacts capacity investments) - Transformer aging - Voltage regulator mechanical wear - Power quality, particularly voltage magnitude - Reverse power flow and its effect on protection systems Prior work in this area consists largely of case studies that use simulations to examine a selection of these issues in detail for a single feeder or a single climate, e.g. [@quezada2006assessment; @shugar1990photovoltaics; @woyte2006voltage; @thomson2007impact; @navarro2013monte; @widen2010impacts]. Results in these papers range from finding that distributed PV can cause resistive losses to *increase* at relatively low penetrations to finding that resistive losses continue to decline up to very high penetrations. Of those papers that examine the impact of PV on voltage excursions, results range from very positive (i.e. acceptable voltages at all penetration levels [@widen2010impacts]) to negative (i.e. unacceptable voltages at high penetration levels [@navarro2013monte]). However, because distribution systems are highly heterogeneous in terms of topology, climate and loads served, it can be difficult to draw useful generalizations from these case studies. We are aware of only two existing studies that examine a diversity of climates and feeder architectures [@paatero2007effects; @hoke2013steady]. In both cases, however, the simulations are driven with hourly solar irradiance data from a single location for each feeder. Therefore these studies cannot provide insight into how cloud transients and geographic diversity of distributed PV systems will influence distribution system operation. The aim of this paper is to evaluate some of distributed PV’s impacts across a diversity of conditions and to inform policy makers and utility decision-makers regarding how extensive these impacts might be at penetrations that are rare today but could be prevalent in the future. The key points of distinction from earlier studies are that and examine a larger number of impacts, climates and feeder types. In addition to studying voltage excursions, resistive losses, incidence of reverse flow and impact on peak loading – as have the aforementioned papers, to varying degrees – we report on voltage regulator operation and loss of life in secondary transformers. The PV data set comprises highly distributed production from residential and small commercial PV systems recorded over a full year at time intervals as small as one minute. These data allow us study the impacts caused by PV variability on feeder voltage and operation of voltage regulation equipment. By looking at all these factors together across different climates, feeder types and PV penetrations, we gain insight into what drives both negative and positive effects of distributed PV in distribution systems. This article is based on a prior conference paper [@cohen2013modeling], and expands it by covering more climates, adding a detailed comparison of simulated load shapes to actual load shapes, and presenting new observations about the importance of geographic diversity. Our central findings have to do with resistive losses and voltage regulation. As one might expect, feeder type – rather than location – has the strongest influence on the total reduction in resistive losses. We also find that percent peak load reduction and incidence of reverse power flow depend more on location (climate) than on feeder type. However, the most severe voltage problems appear to be a function of location, rather than feeder type. As we will describe, impacts on voltage regulators are small and can either increase or decrease relative to a no PV baseline, depending on feeder type (and independent of location). Though we investigate a very large range of impacts in this paper, we acknowledge that there are other impacts that are outside of our scope. For example, we did not investigate the impact of the harmonic content of PV inverters on power quality and transformer aging. We also limit our investigation of protection equipment impact to assessing the prevalence of reverse flow conditions. Furthermore, though our simulations captured the effect of phase imbalances that might occur from random placement of single phase PV on a three phase network, we did not investigate scenarios where we deliberately loaded one phase with more or less PV than others. These omissions and others are due to space, data and modeling limitations, and they merit further systematic investigation in future research. [l l l r r r r r r r r r r]{} & & & & & & & &\ & & & & & & & **Berk.** & **L.A.** & **Sac.** & **Berk.** & **L.A.** & **Sac.**\ & R1-12.47-1 & mod. suburban & rural & 7.15 & 618 & 93% & 5.5 & 5.56 & 5.38 & 7.59 & 21 & 38 & 26\ & R1-12.47-2 & mod. suburban & lt. rural & 2.83 & 264 & 84% & 10.3 & 2.00 & 2.04 & 2.82 & 30 & 30 & 30\ & R1-12.47-3 & moderate urban & 1.35 & 22 & 13% & 1.9 & 1.27 & 1.25 & 1.60 & 10 & 10 & 8\ & R1-12.47-4 & heavy suburban & 5.30 & 50 & 57% & 2.3 & 4.31 & 4.09 & 5.65 & 12 & 17 & 12\ & R1-25.00-1 & light rural & 2.10 & 115 & 2% & 52.5 & 2.35 & 2.23 & 3.00 & 28 & 23 & 30\ & R3-12.47-1 & heavy urban & 8.40 & 472 & 32% & 4.0 & 6.64 & 6.30 & 8.70 & 20 & 31 & 25\ & R3-12.47-2 & moderate urban & 4.30 & 62 & 0% & 5.7 & 3.45 & 3.27 & 4.40 & 13 & 22 & 18\ & R3-12.47-3 & heavy suburban & 7.80 & 1,733 & 84% & 10.4 & 7.54 & 7.00 & 9.67 & 56 & 48 & 55\ \ \ \ The Distribution Feeder Models {#sec:models} ============================== We used version 2.3 (with the forward-backward sweep power flow solver) to model distribution circuits due to its integration of power flow analysis and time-varying load models, availability of representative feeder models, and open-source license. Feeder Topologies ----------------- Pacific Northwest National Lab (PNNL) has compiled a set of representative “taxonomy” feeders drawn from utilities throughout the United States[@schneider2008modern]. As summarized in , the feeders vary along a number of important dimensions such as loads served (urban vs. rural), peak loading, and physical length. The feeders are organized by climate region. For this work, we selected the eight feeders originating from region 1 (temperate west coast) and region 3 (desert southwest) as these climates dominate California. Locations and Timeframe ----------------------- We simulated each of the eight feeders in three locations – Berkeley, Los Angeles and Sacramento – during the 366 days between September 25, 2011 and September 24, 2012, inclusive. We chose these locations and time span due to the availability of high-resolution PV generation and weather data. See Sections \[sec:pv\_data\] to \[sec:geographic\] for more on this data and feeder placement. Note that the California peak demand during the selected year was fairly typical relative to the past decade, with a peak load of in 2012 versus a high of in 2006 [@caiso2013peak]. This means that the simulations do not include extreme conditions that may affect PV’s overall value in important ways in the long run. Feeder Loads and Power Factors {#sec:loads} ------------------------------ Because the taxonomy feeders specify only static planning (i.e. peak) loads, PNNL provides a script to populate the feeders with time-varying residential and commercial loads[@pnnl2012population]. Details of the loading process are discussed in detail in Sections 2.2-2.4 of [@schneider2010evaluation]; we limit the discussion here to a few points of relevance. The PNNL method models end-use loads with “house” objects that have a weather-dependent HVAC component and schedules for other types of loads such as appliances. The schedules for each house are scaled and time-shifted to provide heterogeneity among loads. Commercial loads are modeled as groups of “houses” with a different set of load schedules corresponding to commercial activities. The PNNL script applies a different distribution of load types depending on the climate region selected; e.g. air conditioning is more common in region 3 than in region 1. In this study, we applied region 3 loads to Los Angeles and Sacramento simulations and used region 1 loads in Berkeley, in keeping with the actual climate zone location of these cities. Referring to the literature[@schneider2010evaluation; @shoults2012power; @bravo2012dynamic], we adjusted the script-default load power factors as summarized in Table \[table:pf\]. We also reduced a capacitor bank on one feeder from to after noticing that it was overcompensating for reactive power, possibly because it is a rural feeder and is meant to handle more pumping load. [l r | l r | l r]{} **HVAC** & & **Residential** & & **Commercial** &\ Base HVAC & 0.97 & Water heater & 1.0 & Int. lights\* & 0.90\ Fans & 0.96 & Pool pump\* & 0.87 & Ext. lights\* & 0.95\ Motor losses & 0.125 & Other res.\* & 0.95 & Plug loads\* & 0.95\ & & & & Street lights & 1.0\ \ PV Generation Data {#sec:pv_data} ------------------ The PV integrator SolarCity provided us with a database of instantaneous power at each inverter they monitor (roughly 7,000 systems, mostly in California) under the terms of a non-disclosure agreement. All the inverters are single phase and provide data on the quarter hour; We performed data quality filtering to ensure we used only complete and credible profiles in the models. To address remaining missing readings in the selected profiles, we chose a very complete profile (with at least 365.8 days of non-zero readings between 8:00 and 16:00) from near the center of each location. We used readings from these “filler” profiles to fill gaps longer than one hour in other profiles from that location, scaling the filler readings by the ratio of the two profiles’ rated capacity. Any shorter gaps we allowed to be handled internally by , which uses the last-seen generation value until the model clock reaches the timestamp of the next reading. -------------- --------- ---------- --------------------- --------- ---------- ---------- ---------- ---- [**PV Profiles**]{} **Location** **Low** **Mean** **High** **Low** **Mean** **High** **Used** Berkeley 0 13 35 32 56 94 97 () Los Angeles 4 17 34 39 62 94 99 () Sacramento -4 16 43 25 61 109 101 () -------------- --------- ---------- --------------------- --------- ---------- ---------- ---------- ---- Weather Data {#sec:weather_data} ------------ Table \[table:locations\] summarizes the weather data we used in this study. We obtained one-minute temperature, humidity, and solar irradiance data for Berkeley from Lawrence Berkeley National Laboratoray [@fernandes2012personal] and for Los Angeles and Sacramento from SOLRMAP at Loyola Marymount University and Sacramento Municipal Utility District[@nrel2012midc]. The Los Angeles and Sacramento data, having been quality controlled at the source, appeared to be quite complete and reliable and was used with only minor reformatting. The Berkeley data required the following edits: We calculated direct solar irradiance from global and diffuse irradiance using the solar zenith angle. Also, when irradiance data were missing or zero during the daytime, if less than an hour of data were missing we interpolated between adjacent values (for a total of 30 hours). For longer gaps (totaling 37.4 days) we copied in data from nearby days with similar cloud conditions as measured at Oakland Airport, () south [@noaa2013quality]. We also filled sub-hourly gaps in temperature data (totaling 5.5 days) by interpolation and longer gaps (totaling 25.6 days) directly with hourly measurements from Oakland Airport. The temperature, humidity and irradiance data determined HVAC load in but were not used to simulate PV generation, which was instead extracted from the SolarCity database. By using generation data sources located not far from the weather stations we preserved some (if not all) of the correlation between air conditioning load and PV generation. Given that buildings have significant thermal mass (resulting in a lagged and smoothed response to weather) and our goal was to preserve broad correlations between PV output and building load, we believe that the necessary corrections to the Berkeley weather data are acceptable and do not substantially affect the results. Geographic Assignment of PV Profiles {#sec:geographic} ------------------------------------ We sought to attach PV profiles to houses in a way that reflects the diversity of solar generation over the area of a distribution feeder. This geographic diversity is driven in part by variations in cloud cover, but also by differences in PV system orientation, technology and shading – all of which are reflected in the SolarCity data set. The taxonomy feeders are anonymized and therefore we do not know their physical layout. However, the models do contain electrical connectivity for all components and lengths for each overhead and underground line segment. We used this information and the graph layout utility Graphviz to create a geographic layout for each feeder subject to these constraints. These layouts are available online[@cohen2013taxonomy]. We then used ArcGIS to superimpose the resulting feeder layouts on the SolarCity profile sources. We manually placed the feeders in locations with high densities of generation profiles to capture as much spatial diversity as possible. We then ran a “nearest neighbor” query to assign each distribution transformer to the closest SolarCity profile with acceptable data quality. As Table \[table:locations\] shows, at each location roughly 100 profiles were used (that is, matched with a transformer) with at least one feeder. Table \[table:feeders\] breaks down the number of profiles used in each individual scenario. Penetration Levels and PV Placement {#sec:penetration} ----------------------------------- For each run, we populated only a portion of the houses with PV, to test various levels of penetration. To define “penetration” we first needed to establish a baseline loading for each feeder. To this end, we executed a baseline run for each feeder (with no PV) in each location and recorded its peak load. We then defined penetration as: $$\mbox{PV penetration} = \frac{\sum \mbox{(PV system ratings)}}{\mbox{Peak feeder load from baseline run}}$$ We tested PV penetration levels of 0%, 7.5% 15%, 30%, 50%, 75% and 100%. We placed PV randomly across the available house models and used the same random number seed for all scenarios. Using the same seed ensured that PV was placed at houses in the same order for each climate (Berkeley, Sacramento, Los Angeles), and that all systems populated in lower penetration runs were also populated in higher penetration runs. This allowed us to make comparisons across climates and penetration levels. We modeled the PV as a unity power factor “negative load”. Each house’s PV generation followed the time-varying load profile associated with its distribution transformer (as described in Section \[sec:geographic\]), scaled to an appropriate size for the building as described in Section \[sec:pv\_scaling\]. Because GridLAB-D simulates three phase power flow and we randomly assigned PV systems to single phase points in the system, we are naturally capturing any phase imbalances that would occur from distributed PV in the specific case of random placement. To the extent these imbalances influence voltage magnitudes, they will influence our results in Sections \[sec:regulators\] and \[sec:voltage\_quality\]. All penetration levels should be treated as approximate for two reasons. First, our denominator for penetration was the baseline peak load during the test year, rather than the long-run feeder peak load which would typically be used in situations where more data was available. Second, due to transformer scaling (see Section \[sec:xfmr\_scaling\]) and other minor adjustments, the peak loads from the final 0% penetration runs differ slightly from the peak loads of our baseline runs. In general this difference is small, with the 0% penetration runs having peak load ranging between 3.9% lower and 2.9% higher than the baseline runs. However, in one scenario () the final peak load was 8.0% lower than the baseline peak load. So in this worst case scenario the nominal 100% penetration might more accurately be read as a 108.7% penetration. PV Generation Profile Scaling {#sec:pv_scaling} ----------------------------- All of the selected PV generation profiles appear to be residential-scale, with system ratings ranging from to . To establish a reasonable installation capacity for each building, we first used the following formula from PNNL’s load population script[@pnnl2012population]: $$\mbox{building PV rating estimate} = A \times 0.2 \times 92.902$$ where $A$ is the floor area of the building in square feet, 0.2 is a rough estimate of the rated efficiency of the installations, and is the “standard test conditions” insolation. We scaled up all commercial PV generation profiles so that their ratings matched this rating estimate. For residential installations, we scaled down the generation profile if its rating was *higher* than the rating estimate for the house. We did not scale up residential profiles with ratings smaller than the rating estimate since it is common for residential installations not to occupy the entire roof space. We note that we did not simulate the effect of even larger standalone “utility scale” (multi-MW) PV systems. Had we done so, we expect that voltage and reverse flow problems would be more severe than those we present in Sections \[sec:voltage\_quality\] and \[sec:backflow\]. Transformer Scaling {#sec:xfmr_scaling} ------------------- Transformer aging is one of our outcomes of interest, and it depends not on absolute loading of the transformer but loading relative to the transformer’s rating [@ieee1996guide]. While the simulated loads are roughly scaled to the planning load value listed at each transformer in the taxonomy feeders, the loads may be somewhat larger or smaller than the planning loads due, for instance, to our use of different weather data at the three locations. This means that, unless corrected, some transformers would be sized inappropriately for the loads attached to them. To address this issue, we assembled a “menu” of distribution transformers in standard sizes based on the units present in the taxonomy feeders and manufacturers’ data[@ge1972distribution; @abb2001distribution]. We then replaced each transformer with the smallest transformer from the menu with a rating greater than the observed peak apparent power for that transformer from the baseline run. This is a conservative size estimate for distribution transformers given that in practice many carry power over their ratings during peak periods[@ieee1996guide]. Note that to some extent the concern about transformer sizing also applies to conductor sizing; some taxonomy feeder line conductors may not be sized appropriately for the simulated loads. Because conductor sizing was not a focus of this work, we did not undertake to resize the conductors in the way we did the transformers, and indeed when we run we occasionally observe warnings that conductors are modestly overloaded. This may slightly distort the absolute results for line losses. To address this we instead report the percent change in losses between penetration scenarios. The percent change should not be affected significantly by conductor size since line resistance is a linear scaling factor on line losses and all penetration levels use the same conductors. GridLAB-D Configuration ----------------------- All of the taxonomy feeders have an on-load tap changer (LTC) at the substation, and two of them feature additional line voltage regulators. During the baseline runs, we observed that the upper bound of the LTC and regulator deadbands were set at approximately , right at the edge of ANSI standards for end-use voltages. This contributed to a significant number of voltage violations due to time lag in regulator response when voltages rose outside the deadband. We therefore lowered the top of the LTC and regulator deadbands to (maintaining the bandwidth) for our production model runs. Results ======= Our results must be interpreted with several important caveats in mind. First, the simulation covers one particular year that was chosen primarily for PV data availability. It may not include extreme weather or other events that would drive true system peaks in the long term. Second, though the load models are physically-based and the taxonomy feeders are based on real feeders, we did not model the actual feeders and loads in the study locations. Third, the prototypical feeders are “typical”, meaning they do not have special problems such as poor voltage regulation or capacity constraints that would require special attention when integrating PV. Finally, because of the transformer and conductor sizing issues discussed in Section \[sec:xfmr\_scaling\] and later in Section \[sec:transformer\_aging\], we do not consider thermal overloading of transformers and lines; this prevents us from being able to make specific claims about the hosting capacity of each network. The base quantity for all normalized results is the value of the metric in question for the feeder at 0% penetration. System Losses ------------- We measured instantaneous system losses (including transformer and line losses) every fifteen minutes. As shown in Figure \[fig:losses\], we found that increasing PV penetration decreased system losses, with diminishing effects at high penertations. The impact of PV on losses was similar across the three locations, but varied considerably by topology, with losses reduced by anywhere from 7% () to 28% () at 100% penetration. In particular, feeders with higher nominal peak loads (see Table \[table:feeders\]) tended to have less loss reduction with increasing PV, though this trend was not universal. We also found, unsurprisingly, that the feeder that experienced the largest reduction in percent losses was also the longest. For reference, the total annual losses in the 0% penetration case for each feeder are: 810 MWh (R1-1247-1), 310 MWh (R1-1247-2), 40 MWh (R1-1247-3), 290 MWh (R1-1247-4), 150 MWh (R1-2500-1), 540 MWh (R3-1247-1), 130 MWh (R3-1247-2), 1440 MWh (R3-1247-3). Note that to produce these statistics we averaged across each of the three locations for each feeder and we rounded to the nearest 10 MWh. The actual values for each feeder varied on the order of $\pm$ 5% across locations. We attribute the reduced marginal effect of PV at high penetrations to the fact that losses are proportional to current squared; the more PV reduces power (and thus current) flow on the lines, the less effect further reductions will have on losses. For some feeders (mainly in Sacramento) losses *increased* as penetration rose from 75% to 100%, presumably because the losses associated with high “backflow” currents at certain times began to exceed the losses “saved” at other times when net current flow was lower. Other studies have found that resistive losses increase with penetration [@quezada2006assessment; @widen2010impacts; @navarro2013monte; @thomson2007impact]. However in contrast to other work, our finding is that on *most* feeders we study, losses continue to decline up to 100% penetration. We note that in the feeder / location pairs here, location seems to determine whether or not losses begin to increase in the range of penetrations we examined, but that the total magnitude of losses is much more strongly influenced by the feeder type. Figure \[fig:losses\] shows that losses as a percentage of energy consumed by loads from the grid (i.e. as a percentage of utility wholesale power purchases) generally increase with PV penetration. This is likely because most of the load reduction happens off-peak, when system losses are lower than on-peak. Peak Loading {#sec:peak_load} ------------ We measured peak load as the maximum fifteen-minute rolling average of one-minute measurements at the substation. The extent to which PV reduces feeder peak load depends largely on the timing of the peaks. Clearly, peak load reduction will be greatest if peak load is coincident with peak PV production. In California, however, load typically peaks later in the day than PV production, and therefore peak loads are reduced by only a fraction of the PV’s rating. As shown in Figure \[fig:peak\_loads\], we observed that PV generally reduced peak loads by much less than the penetration percentage. In contrast to system losses, location (i.e. climate) had a strong effect on the peak load reduction impact of PV, with Sacramento and Berkeley showing more significant reductions than Los Angeles. Figure \[fig:peak\_load\] shows the normalized peak load as a function of PV penetration, whereas Figure \[fig:solar\_effectiveness\] shows the peak reduction as a percentage of the solar penetration. Figure \[fig:solar\_effectiveness\] illustrates that low penetrations of PV can be quite effective at reducing peak loads, although this is not true in all cases. Peak load reduction effectiveness diminishes as penetration increases because early increments of PV tend to reduce daytime peaks, causing the new peak to be in the evening when PV contributes less power. Figure \[fig:peak\_date\_time\] illustrates trends in the timing of peaks as PV penetration increases. Without PV, peak loads arrived in August 2012 for most Sacramento feeders and half of the Los Angeles feeders, while Berkeley feeders generally peaked in fall 2011 or June 2012. Peak times were widely dispersed between 14:22 and 17:18. However, a 7.5% penetration of PV was sufficient to eliminate August peaks for all but one Los Angeles feeder, shifting their peaks to the later afternoon during a relatively warm spell in October 2011. Berkeley peaks, while initially shifting towards the summer, were ultimately also moved to the fall by high penetrations of PV. Meanwhile the Sacramento peaks, driven by larger air conditioning loads, remained in the summer at all levels of penetration, although moving noticeably later in the afternoon. In all locations, peaks were moved later in the day as PV reduced daytime usage. {width="5in"} When interpreting the peak load reduction results, it is important to consider how well the simulated feeder load shapes align with feeder load shapes actually found in California. We do not have access to a large enough corpus of load shapes to do a rigorous analysis of this issue, but a high-level comparison will suffice to contextualize our findings. Figure \[fig:load\_profiles\] shows the average hourly load and PV generation for each of the simulated feeders on August 13, 2012, which was the day CAISO recorded its peak demand for 2012 [@caiso2013peak]. It is also the peak demand day for five simulated Sacramento feeders, though not for any Los Angeles or Berkeley feeders. Each individual profile is normalized against the peak hour for that profile. As in the other figures, the locational means are straight averages of the eight normalized feeder simulations, i.e. the feeders are not weighted by their size or expected frequency of occurrence in the field. The load plot also shows normalized CAISO system load (larger green circles) and PG&E system load (larger blue circles). From this figure we can see that the simulated peaks match well with the PG&E and CAISO peaks in the 15:00-16:00 range. However, the simulated feeders universally drop in demand more quickly than the CAISO system. Note from the bottom panel in Figure \[fig:load\_profiles\] that PV production goes to zero after the simulated load drops, but before any significant drop in CAISO load. This suggests the possibility that peak demand might be relatively unaffected by PV in the CAISO system, but strongly affected in our simulations. This simple one-day comparison ignores several factors that are important when calculating annual peak demand reduction, such as load variation within each hour and the fact that PV often shifts the peak to a different day, rather than a different time on the same day. Also, the comparison to an overall system load profile greatly obscures the wide variation of individual feeder profiles that comprise it. For instance, SCADA data provided by PG&E under the terms of a nondisclosure agreement indicates that on August 13, 2012 the most common hours for feeders to peak were 16:00 and 17:00, but each of these hours only accounted for about 16% of feeders, with 37% peaking earlier (including 10% before noon) and 31% later in the evening [@carruthers2013personal]. Thus, it is likely that the simulated load shapes are a good match to some subset of California feeders and therefore the reported peak load reduction is achievable in some locations. However, the fact that the simulated feeder profiles are not a good match for the general system profile in the evening indicates that it would be optimistic to expect the simulated peak load reduction to occur universally across California. Transformer Aging {#sec:transformer_aging} ----------------- 2.3 implements the IEEE Standard C57.91 [@ieee1996guide] method for estimating transformer insulation aging under various loading conditions. implements the method for single phase center tapped transformers only. This is the most common type of transformer on the taxonomy feeders, but one feeder () did not have any so it was excluded from the aging analysis. In the model, a “normal” year of aging corresponds to the amount of insulation degradation expected if the transformer hot spot were at a constant throughout the year. A transformer that is often overloaded will age more than in a year, and thus may need to be taken out of service due to insulation degradation before its rated lifetime. On the other hand, one that is loaded below its rating will age less than per year, and will be unlikely to have its insulation fail prematurely. In general, we observed minimal aging in all scenarios and penetration levels, with a mean equivalent aging of up to in one scenario (, Sac.) and all other scenarios having mean aging less than . We attribute this slow aging to the fact that the transformers were conservatively sized at or above their baseline peak load (see Section \[sec:xfmr\_scaling\]). However, in (Sac.) at PV penetrations of 30% and above we did observe a small number of transformers aging quite rapidly, up to during the simulated year (all other scenarios had maximum individual transformer aging less than per year). These few rapidly aging transformers are likely at a location where net PV generation is often higher than the load they were sized to handle, and in reality they would need to be upgraded to handle this backflow. Voltage Regulators {#sec:regulators} ------------------ Tap-changing voltage regulator wear and tear is driven primarily by the number of tap changes the device must perform and the current that it handles during operation. In our simulations, tap changes at the substation LTC were on the order of 20 per day. However the count was not affected by topology, climate or PV penetration, varying between 7,166 and 7,243 changes across all model runs over the year of simulation – a difference of only 1%. This small difference is because the models did not include a transmission impedance component, with the transmission voltage instead following a fixed schedule of values recorded from an actual substation in the U.S. Western Interconnection (WECC). The substation LTC operates to maintain voltage immediately downstream within the deadband despite fluctuations in the WECC schedule, and is insensitive to downstream changes in load. Due to the lack of a transmission model, our simulations do not provide reliable insight on LTC response to PV. The two mid-feeder regulators in the simulation (at and ) do have simulated impedances and varying loads both upstream and downstream and thus exhibit more variation. Figure \[fig:tap\_changes\] shows that PV has little effect at until 50% penetration, at which point tap changes begin rising noticeably. This result echoes other work[@mather2012quasi] and concerns from utilities that PV variability will increase regulator maintenance needs. We examined two sensitivity scenarios to study the impact that the PV data had on the regulator results. To produce the dotted lines in Figure \[fig:regulators\] we used the single PV profile with the most one-minute data available (82% of days) at all PV sites. The dashed line shows the same scenario with the one-minute data downsampled to fifteen-minute resolution; this intermediate scenario helps us to distinguish the effect of the one-minute data from the effect of eliminating geographic diversity. We limited the sensitivities to Los Angeles because this was our source of one-minute data. Figure \[fig:tap\_changes\] suggests that geographic diversity reduces tap change frequency (because the solid lines which include geographic diversity fall well below their corresponding single-profile dotted and dashed lines) and that fifteen-minute PV data is a reasonable proxy for one-minute data when studying regulator behavior (because the dashed lines track their corresponding dotted lines closely). The effect of PV on regulator current duty was more consistent than the effect on tap changes, as illustrated by Figure \[fig:avg\_current\]. With PV reducing the downstream load, current through the regulator declines steadily as penetration increases. This suggests that even in cases where PV increases a regulator’s activity, its expected lifetime may stay the same or even increase because each tap change is less destructive under lighter current duty. Our sensitivity runs suggest that neither geographic diversity nor the use of one-minute resolution data has a substantial effect on regulator current duty. Voltage Quality {#sec:voltage_quality} --------------- We recorded voltage at all point-of-use meters at fifteen minute intervals and tabulated in Figure \[fig:bad\_voltages\] the proportion of readings falling outside of the ANSI standard range of . In general, voltages appear to be well-controlled, with most runs having less than 0.002% of readings out of range, and the worst case (, Sac.) having 0.32% of readings out of range. [This is consistent with prior work suggesting that many feeders can support high penetrations of PV without voltage violations [@hoke2013steady], however it may be counter-intuitive that feeders designed for one-way power flow can host so much PV capacity without more negative voltage impacts. There are several explanations for this. First, the feeders we investigated had relatively good voltage control and voltage regulators rarely saturated; it is plausible that there are feeders in operation whose control is more likely to saturate. Second, we did not model scenarios with PV heavily concentrated in part of a feeder – this would exacerbate local reverse power flow and voltage rise. Finally, though the maximum penetration we investigated is relatively high, penetrations could be on the order of 200% if systems were sized to produce as much energy over the course of a year as each building consumes. We expect that voltage excursions would be much more significant at those penetrations.]{} In general, the voltage violations that did occur took place on rural and suburban feeders (see Table \[table:feeders\]) with violations being very rare on urban feeders at all penetration levels. Except at feeder , almost all out-of-range voltages observed were greater than . As expected these high-side excursions generally become more frequent as penetration increased and the power injection from PV raised some voltages locally. At the out of range voltages were predominantly less than , with a small amount greater than . Under these conditions, increasing PV penetration improved voltage quality on the feeder by boosting some local voltages that would otherwise be low. As noted in Section \[sec:regulators\], it is possible that more brief voltage excursions would be observed with higher resolution PV generation data. Reverse Power Flow {#sec:backflow} ------------------ We also investigated the incidence of negative real power flow (“backflow”) through the substation, which can be a proxy for protection issues and higher interconnection costs. At 50% penetration, twelve of the 24 scenarios exhibited occasional backflow, up to 1% of the time each. At 100% penetration, all scenarios experienced backflow at least 4% of the time. In general, backflow was more prevalent in Sacramento because PV penetration in Sacramento was measured against a higher peak air conditioning load. This led to a larger absolute quantity of PV generation in Sacramento but with similar low loads to Los Angeles and Berkeley on cooler days. Observations Regarding Geographic Diversity {#sec:diversity} ------------------------------------------- We ran our sensitivity scenarios primarily to assess the effect of PV profile time resolution and geographic diversity on voltage regulator operation (see Section \[sec:regulators\]). However, these scenarios enable us to observe how other outcomes vary with the input data as well. These observations are necessarily tentative because the sensitivities were run for only two feeders (R1-25.00-1 and R3-12.47-3) in one location (Los Angeles). First, we note that for all outcomes observed, differences between the single-profile one-minute input and that input downsampled to fifteen-minute resolution were minimal. This implies that fifteen-minute PV data is “good enough” for a reliable study of PV’s effects on the distribution system. Second, for two metrics we did observe changes in outcomes when switching from the full geographic diversity of profiles to the single profile for all PV installations. First, peak load reduction was larger with geographic diversity than without it. We attribute this to the fact that the diverse set of profiles includes west-facing installations that are more effective at reducing peak load. We also noticed substantially less backflow at high penetrations with geographic diversity. This is expected because with a single profile periods of high generation will be completely coincident, whereas with a diverse set of profiles they will be spread out somewhat – by system orientation if not by cloud cover differences – reducing the overall “peakiness” of PV generation and thus backflow. Taken together, these observations suggest that studies that do not account for the geographic diversity of PV – even on a distribution feeder scale – may underestimate some of its benefits and/or overstate its drawbacks. Concluding remarks ================== We studied how distributed PV impacts distribution systems across a variety of feeder architectures and climates within California over a full year of operation. In contrast to earlier studies, we ran simulations with real PV data (either 1-minute or 15-minute resolution), which allowed us to uniquely address issues of voltage regulation on the time scale of cloud transients. In addition to studying voltage excursions, resistive losses, reverse flow and impact on peak loading – as have researchers before us – we examined voltage regulator operation and loss of life in secondary transformers. We used unique PV data that captured the impacts of fast cloud transients, array shading and spatial diversity. It is worth emphasizing that, while this paper is extensive in terms of its combination of geographic scope, number of feeder types and high resolution PV data, it is not an exhaustive assessment of all possible outcomes. Overall, we observe that undesirable impacts are relatively small in most cases but large in some; we expect that a similar pattern of observations would hold across an even larger range of California scenarios than we consider in this paper. Acknowledgment {#acknowledgment .unnumbered} ============== We wish to thank John Carruthers, Donovan Currey & Matt Heling of PG&E; Jason Fuller and the team at PNNL; Eric Carlson, Justin Chebahtah & Karthik Varadarajan of SolarCity; Dan Arnold, Lloyd Cibulka, Josiah Johnston, Paul Kauzmann, Kevin Koy, James Nelson, Ciaran Roberts, Michaelangelo Tabone, Alexandra von Meier and Shuyu (Simon) Yang of UC Berkeley; Luis Fernandes and Emma Stewart of LBNL; the UCB CITRIS computing cluster for their invaluable assistance; and the editorial staff and three anonymous reviewers of this manuscript, whose comments greatly improved the clarity of the paper. [^1]: The authors are with the Energy and Resources Group, University of California, Berkeley, CA, 94720-3050 USA e-mail: [macohen,dcal]{}@berkeley.edu [^2]: This work was supported by the California Solar Initiative RD&D program and Robert Bosch LLC through its Bosch Energy Research Network program.

--- abstract: 'This paper surveys some recent developments in fundamental limits and optimal algorithms for network analysis. We focus on minimax optimal rates in three fundamental problems of network analysis: graphon estimation, community detection, and hypothesis testing. For each problem, we review state-of-the-art results in the literature followed by general principles behind the optimal procedures that lead to minimax estimation and testing. This allows us to connect problems in network analysis to other statistical inference problems from a general perspective.' author: - Chao Gao - Zongming Ma bibliography: - 'reference.bib' date:   title: 'Minimax Rates in Network Analysis: Graphon Estimation, Community Detection and Hypothesis Testing ' --- Introduction ============ Network analysis [@goldenberg2010survey] has gained considerable research interests in both theory [@bickel09] and applications [@girvan2002community; @wasserman1994social]. In this survey, we review recent developments that establish the fundamental limits and lead to optimal algorithms in some of the most important statistical inference tasks. Consider a stochastic network represented by an adjacency matrix $A\in\{0,1\}^{n\times n}$. In this paper, we restrict ourselves to the setting where the network is an undirected graph without self loops. To be specific, we assume that $A_{ij}=A_{ji}\sim \text{Bernoulli}(\theta_{ij})$ for all $i<j$. The symmetric matrix $\theta\in[0,1]^{n\times n}$ models the connectivity pattern of a social network and fully characterizes the data generating process. The statistical problems we are interested is to learn structural information of the network coded in the matrix $\theta$. We focus on the following three problems: 1. *Graphon estimation.* The celebrated Aldous-Hoover theorem [@aldous81; @hoover79] asserts that the exchangeability of $\{A_{ij}\}$ implies the representation that $\theta_{ij}=f(\xi_i,\xi_j)$ with some nonparametric function $f(\cdot,\cdot)$. Here, $\xi_i$’s are i.i.d. random variables uniformly distributed in the unit interval $[0,1]$. The function $f$ is coined as the graphon of the network. The problem of graphon estimation is to estimate $f$ with the observed adjacency matrix. 2. *Community detection.* Many social networks such as collaboration networks and political networks exhibit clustering structure. This means that the connectivity pattern is determined by the clustering labels of the network nodes. In general, for an assortative network, one expects that two network nodes are more likely to be connected if they are from the same cluster. For a disassortative network, the opposite pattern is expected. The task of community detection is to learn the clustering structure, and is also referred to as the problem of graph partition or network cluster analysis. 3. *Hypothesis testing.* Perhaps the most fundamental question for network analysis is whether a network has some structure. For example, an [[Erdős–Rényi]{}]{} graph has a constant connectivity probability for all edges, and is regarded to have no interesting structure. In comparison, a stochastic block model has a clustering structure that governs the connectivity pattern. Therefore, before conducting any specific network analysis, one should first test whether a network has some structure or not. The test between an [[Erdős–Rényi]{}]{} graph and a stochastic block model is one of the simplest examples. This survey will emphasize the developments of the minimax rates of the problems. The state-of-the-art of the three problems listed above will be reviewed in Section \[sec:graphon\], Section \[sec:community-detection\], and Section \[sec:test\], respectively. In each section, we will introduce critical mathematical techniques that we use to derive optimal solutions. When appropriate, we will also discuss the general principles behind the problems. This allows us to connect the results of the network analysis to some other interesting statistical inference problems. Real social networks are often sparse, which means that the number of edges are of a smaller order compared with the number of nodes squared. How to model sparse networks is a longstanding topic full of debate [@lloyd2012random; @bickel09; @crane2016edge; @caron2017sparse]. In this paper, we adopt the notion of network sparsity $\max_{1\leq i<j\leq n}\theta_{ij}=o(1)$, which is proposed by [@bickel09]. Theoretical foundations of this sparsity notion were investigated by [@borgs2014p1; @borgs2014p2]. There are other, perhaps more natural, notions of network sparsity, and we will discuss potential open problems in Section \[sec:disc\]. We close this section by introducing some notation that will be used in the paper. For an integer $d$, we use $[d]$ to denote the set $\{1,2,...,d\}$. Given two numbers $a,b\in\mathbb{R}$, we use $a\vee b=\max(a,b)$ and $a\wedge b=\min(a,b)$. For two positive sequences $\{a_n\},\{b_n\}$, $a_n\lesssim b_n$ means $a_n\leq Cb_n$ for some constant $C>0$ independent of $n$, and $a_n\asymp b_n$ means $a_n\lesssim b_n$ and $b_n\lesssim a_n$. We write $a_n\ll b_n$ if $a_n/b_n\rightarrow 0$. For a set $S$, we use ${{\mathbf{1}_{\left\{{S}\right\}}}}$ to denote its indicator function and $|S|$ to denote its cardinality. For a vector $v\in\mathbb{R}^d$, its norms are defined by ${\|{v} \|}_1=\sum_{i=1}^n|v_i|$, ${\|{v} \|}^2=\sum_{i=1}^nv_i^2$ and ${\|{v} \|}_{\infty}=\max_{1\leq i\leq n}|v_i|$. For two matrices $A,B\in\mathbb{R}^{d_1\times d_2}$, their trace inner product is defined as ${\left \langle A, B \right\rangle}=\sum_{i=1}^{d_1}\sum_{j=1}^{d_2}A_{ij}B_{ij}$. The Frobenius norm and the operator norm of $A$ are defined by ${\|A\|_{\rm F}}=\sqrt{{\left \langle A, A \right\rangle}}$ and ${\|A\|_{\rm op}}=s_{\max}(A)$, where $s_{\max}(\cdot)$ denotes the largest singular value. Graphon estimation {#sec:graphon} ================== Problem settings ---------------- Graphon is a nonparametric object that determines the data generating process of a random network. The concept is from the literature of exchangeable arrays [@aldous81; @hoover79; @kallenberg89] and graph limits [@lovasz12; @diaconis07]. We consider a random graph with adjacency matrix $\{A_{ij}\}\in\{0,1\}^{n\times n}$, whose sampling procedure is determined by $$(\xi_1,...,\xi_n)\sim\mathbb{P}_{\xi},\quad A_{ij}|(\xi_i,\xi_j)\sim\text{Bernoulli}(\theta_{ij}),\quad\text{where }\theta_{ij}=f(\xi_i,\xi_j).\label{eq:graphon}$$ For $i\in[n]$, $A_{ii}=\theta_{ii}=0$. Conditioning on $(\xi_1,...,\xi_n)$, the $A_{ij}$’s are mutually independent across all $i<j$. The function $f$ on $[0,1]^2$, which is assumed to be symmetric, is called graphon. The graphon offers a flexible nonparametric way of modeling stochastic networks. We note that exchangeability leads to independent random variables $(\xi_1,...,\xi_n)$ sampled from $\text{Uniform}[0,1]$, but for the purpose of estimating $f$, we do not require this assumption. We point out an interesting connection between graphon estimation and nonparametric regression. In the formulation of (\[eq:graphon\]), suppose we observe both the adjacency matrix $\{A_{ij}\}$ and the latent variables $\{(\xi_i,\xi_j)\}$, then $f$ can simply be regarded as a regression function that maps $(\xi_i,\xi_j)$ to the mean of $A_{ij}$. However, in the setting of network analysis, we only observe the adjacency matrix $\{A_{ij}\}$. The latent variables are usually used to model latent features of the network nodes [@hoff2002latent; @ma2017exploration], and are not always available in practice. Therefore, graphon estimation is essentially a nonparametric regression problem without observing the covariates, which leads to a new phenomenon in the minimax rate that we will present below. In the literature, various estimators have been proposed. For example, a singular value threshold method is analyzed by [@chatterjee12], later improved by [@xu2017rates]. The paper [@lloyd2012random] considers a Bayesian nonparametric approach. Another popular procedure is to estimate the graphon via histogram or stochastic block model approximation [@wolfe13; @chan14; @airoldi13; @olhede13; @borgs2015private; @borgs2015consistent]. Minimax rates of graphon estimation are investigated by [@gao2015rate; @gao2016optimal; @klopp2017structured]. Optimal rates ------------- Before discussing the minimax rate of estimating a nonparametric graphon, we first consider graphons that are block-wise constant functions. This is equivalently recognized as stochastic block models (SBMs) [@holland83; @nowicki01]. Consider $A_{ij}\sim \text{Bernoulli}(\theta_{ij})$ for all $1\leq i<j\leq n$. The class of SBMs with $k$ clusters is defined as $$\begin{aligned} \label{eq:SBM-k} \Theta_k &=& \Bigg\{\{\theta_{ij}\}\in [0,1]^{n\times n}: \theta_{ii}=0,\text{ }\theta_{ij}=B_{uv}=B_{vu}\\ \nonumber && \text{ for }(i,j)\in z^{-1}(u)\times z^{-1}(v)\text{ with some }B_{uv}\in[0,1]\text{ and }z\in[k]^n\Bigg\}.\end{aligned}$$ In other words, the network nodes are divided into $k$ clusters that are determined by the cluster labels $z$. The subsets $\{\mathcal{C}_u(z)\}_{z\in[k]}$ with $\mathcal{C}_u(z)=\{i\in[n]:z(i)=u\}$ form a partition of $[n]$. The mean matrix $\theta\in[0,1]^{n\times n}$ is a piecewise constant with respect to the blocks $\{\mathcal{C}_u(z)\times \mathcal{C}_v(z):u,v\in [k] \}$. In this setting, graphon estimation is the same as estimating the mean matrix $\theta$. If we know the clustering labels $z$, then we can simply calculate the sample averages of $\{A_{ij}\}$ in each block $\mathcal{C}_u(z)\times \mathcal{C}_v(z)$. Without the knowledge of $z$, a least-squares estimator proposed by [@gao2015rate] is $${\widehat}{\theta}={\mathop{\rm argmin}}_{\theta\in\Theta_k}{\|A-\theta\|_{\rm F}}^2,\label{eq:least-squares}$$ which can be understood as the sample averages of $\{A_{ij}\}$ over the estimated blocks $\{\mathcal{C}_u({\widehat}{z})\times \mathcal{C}_v({\widehat}{z})\}$. To study the performance of the least-squares estimator ${\widehat}{\theta}$, we need to introduce some additional notation. Since ${\widehat}{\theta}\in\Theta_k$, the estimator can be written as ${\widehat}{\theta}_{ij}={\widehat}{B}_{{\widehat}{z}(i){\widehat}{z}(j)}$ for some ${\widehat}{B}\in[0,1]^{k\times k}$ and some ${\widehat}{z}\in[k]^n$. The true matrix that generates $A$ is denoted by $\theta^*$. Then, we define $${\widetilde}{\theta}={\mathop{\rm argmin}}_{\theta\in\Theta_k({\widehat}{z})}{\|\theta^*-\theta\|_{\rm F}}^2.$$ Here, the class $\Theta_k({\widehat}{z})\subset\Theta_k$ consists of all SBMs with clustering structures determined by ${\widehat}{z}$. Then, we immediately have the Pythagorean identity $${\|{\widehat}{\theta}-\theta^*\|_{\rm F}}^2 = {\|{\widehat}{\theta}-{\widetilde}{\theta}\|_{\rm F}}^2 + {\|{\widetilde}{\theta}-\theta^*\|_{\rm F}}^2.\label{eq:Pythagorean}$$ By the definition of ${\widehat}{\theta}$, we have the basic inequality ${\|{\widehat}{\theta}-A\|_{\rm F}}^2\leq {\|\theta^*-A\|_{\rm F}}^2$. After a simple rearrangement, we have $$\begin{aligned} {\|{\widehat}{\theta}-\theta^*\|_{\rm F}}^2 &\leq& 2\left|{\left \langle {\widehat}{\theta}-\theta^*, A-\theta^* \right\rangle}\right| \\ &\leq& 2{\|{\widehat}{\theta}-{\widetilde}{\theta}\|_{\rm F}}\left|{\left \langle \frac{{\widehat}{\theta}-{\widetilde}{\theta}}{{\|{\widehat}{\theta}-{\widetilde}{\theta}\|_{\rm F}}}, A-\theta^* \right\rangle}\right| + 2{\|{\widetilde}{\theta}-\theta^*\|_{\rm F}}\left|{\left \langle \frac{{\widetilde}{\theta}-\theta^*}{{\|{\widetilde}{\theta}-\theta^*\|_{\rm F}}}, A-\theta^* \right\rangle}\right| \\ &\leq& {\|{\widehat}{\theta}-\theta^*\|_{\rm F}}\sqrt{\left|{\left \langle \frac{{\widehat}{\theta}-{\widetilde}{\theta}}{{\|{\widehat}{\theta}-{\widetilde}{\theta}\|_{\rm F}}}, A-\theta^* \right\rangle}\right|^2+\left|{\left \langle \frac{{\widetilde}{\theta}-\theta^*}{{\|{\widetilde}{\theta}-\theta^*\|_{\rm F}}}, A-\theta^* \right\rangle}\right|^2},\end{aligned}$$ where the last inequality is by Cauchy-Schwarz and (\[eq:Pythagorean\]). Therefore, we have $$\begin{aligned} {\|{\widehat}{\theta}-\theta^*\|_{\rm F}}^2 &\leq& \left|{\left \langle \frac{{\widehat}{\theta}-{\widetilde}{\theta}}{{\|{\widehat}{\theta}-{\widetilde}{\theta}\|_{\rm F}}}, A-\theta^* \right\rangle}\right|^2+\left|{\left \langle \frac{{\widetilde}{\theta}-\theta^*}{{\|{\widetilde}{\theta}-\theta^*\|_{\rm F}}}, A-\theta^* \right\rangle}\right|^2 \\ &\leq& \sup_{\{v\in\Theta_k:{\|v\|_{\rm F}}=1\}}|{\left \langle v, A-\theta^* \right\rangle}|^2 + \max_{1\leq j\leq k^n}|{\left \langle v_j, A-\theta^* \right\rangle}|^2,\end{aligned}$$ where $\{v_j\}_{1\leq j\leq k^n}$ are $k^n$ fixed matrices with Frobenius norm $1$. To understand the last inequality above, observe that $\frac{{\widehat}{\theta}-{\widetilde}{\theta}}{{\|{\widehat}{\theta}-{\widetilde}{\theta}\|_{\rm F}}}$ belongs to $\Theta_k$ and has Frobenius norm $1$, and the matrix $\frac{{\widetilde}{\theta}-\theta^*}{{\|{\widetilde}{\theta}-\theta^*\|_{\rm F}}}$ takes at most $k^n$ different values. Finally, an empirical process argument and a union bound leads to the inequalities $$\begin{aligned} \mathbb{E}\left[\sup_{\{v\in\Theta_k:{\|v\|_{\rm F}}=1\}}|{\left \langle v, A-\theta^* \right\rangle}|^2\right] &\lesssim& k^2+n\log k, \\ \mathbb{E}\left[\max_{1\leq j\leq k^n}|{\left \langle v_j, A-\theta^* \right\rangle}|^2\right] &\lesssim& n\log k,\end{aligned}$$ which then implies the bound $$\mathbb{E}{\|{\widehat}{\theta}-\theta^*\|_{\rm F}}^2 \lesssim k^2 + n\log k. \label{eq:SBM-upper}$$ The upper bound (\[eq:SBM-upper\]) consists of two terms. The first term $k^2$ corresponds to the number of parameters we need to estimate in an SBM with $k$ clusters. The second term results from not knowing the exact clustering structure. Since there are in total $k^n$ possible clustering configurations, the complexity $\log(k^n)=n\log k$ enters the error bound. Even though the bound (\[eq:SBM-upper\]) is achieved by an estimator that knows the value of $k$, a penalized version of the least-squares estimator with the penalty $\lambda(k^2 + n\log k)$ can achieve the same bound (\[eq:SBM-upper\]) without the knowledge of $k$. The paper [@gao2015rate] also shows that the upper bound (\[eq:SBM-upper\]) is sharp by proving a matching minimax lower bound. While it is easy to see that the first term $k^2$ cannot be avoided by a classical lower bound argument of parametric estimation, the necessity of the second term $n\log k$ requires a very delicate lower bound construction. It was proved by [@gao2015rate] that it is possible to construct a $B\in[0,1]^{k\times k}$, such that the set $\{B_{z(i)z(j)}:z\in[k]^n\}$ has a packing number bounded below by $e^{cn\log k}$ with respect to the norm ${\|\cdot\|_{\rm F}}$ and the radius at the order of $\sqrt{n\log k}$. This fact, together with a standard Fano inequality argument, leads to the desired minimax lower bound. We summarize the above discussion into the following theorem. \[thm:sbm-minimax\] For the loss function $L({\widehat}{\theta},\theta)={{n\choose 2}}^{-1}\sum_{1\leq i<j\leq n}({\widehat}{\theta}_{ij}-\theta_{ij})^2$, we have $$\inf_{{\widehat}{\theta}}\sup_{\theta\in\Theta_k}\mathbb{E}L({\widehat}{\theta},\theta)\asymp \frac{k^2}{n^2} + \frac{\log k}{n},$$ for all $1\leq k\leq n$. Having understood minimax rates of estimating mean matrices of SBMs, we are ready to discuss minimax rates of estimating general nonparametric graphons. We consider the following loss function that is widely used in the literature of nonparametric regression, $$L({\widehat}{f},f)=\frac{1}{{n\choose 2}}\sum_{1\leq i<j\leq n}\left({\widehat}{f}(\xi_i,\xi_j)-f(\xi_i,\xi_j)\right)^2.$$ Note that $L({\widehat}{f},f)=L({\widehat}{\theta},\theta)$ if we let ${\widehat}{\theta}_{ij}={\widehat}{f}(\xi_i,\xi_j)$ and $\theta_{ij}=f(\xi_i,\xi_j)$. Then, the minimax risk is defined as $$\inf_{{\widehat}{f}}\sup_{f\in\mathcal{H}_{\alpha}(M)}\sup_{\mathbb{P}_{\xi}}\mathbb{E}L({\widehat}{f},f).$$ Here, the supreme is over both the function class $\mathcal{H}_{\alpha}(M)$ and the distribution $\mathbb{P}_{\xi}$ that the latent variables $(\xi_1,...,\xi_n)$ are sampled from. While $\mathbb{P}_{\xi}$ is allowed to range from the class of all distributions, the Hölder class $\mathcal{H}_{\alpha}(M)$ is defined as $$\mathcal{H}_{\alpha}(M)=\left\{\|f\|_{\mathcal{H}_{\alpha}}\leq M: f(x,y)=f(y,x)\text{ for }x\geq y\right\},$$ where $\alpha>0$ is the smoothness parameter and $M>0$ is the size of the class. Both are assumed to be constants. In the above definition, $\|f\|_{\mathcal{H}_{\alpha}}$ is the Hölder norm of the function $f$ (see [@gao2015rate] for the details). The following theorem gives the minimax rate of the problem. \[thm:graphon-minimax\] We have $$\inf_{{\widehat}{f}}\sup_{f\in\mathcal{H}_{\alpha}(M)}\sup_{\mathbb{P}_{\xi}}\mathbb{E}L({\widehat}{f},f) \asymp \begin{cases} n^{-\frac{2\alpha}{\alpha+1}}, & 0<\alpha<1,\\ \frac{\log n}{n}, & \alpha\geq 1, \end{cases}$$ where the expectation is jointly over $\{A_{ij}\}$ and $\{\xi_i\}$. The minimax rate in Theorem \[thm:graphon-minimax\] exhibits different behaviors in the two regimes depending on whether $\alpha\geq 1$ or not. For $\alpha\in(0,1)$, we obtain the classical minimax rate for nonparametric regression. To see this, one can related the graphon estimation problem to a two-dimensional nonparametric regression problem with sample size $N=\frac{n(n-1)}{2}$, and then it is easy to see that $N^{-\frac{2\alpha}{2\alpha+d}}\asymp n^{-\frac{2\alpha}{\alpha+1}}$ for $d=2$. This means for a nonparametric graphon that is not so smooth, whether or not the latent variables $\{(\xi_i,\xi_j)\}$ are observed does not affect the minimax rate. In contrast, when $\alpha\geq 1$, the minimax rate scales as $\frac{\log n}{n}$, which does not depend on the value of $\alpha$ anymore. In this regime, there is a significant difference between the graphon estimation problem and the regression problem. Both the upper and lower bounds in Theorem \[thm:graphon-minimax\] can be derived by an SBM approximation. The minimax rate given by Theorem \[thm:graphon-minimax\] can be equivalently written as $$\min_{1\leq k\leq n}\left\{\frac{k^2}{n^2}+\frac{\log k}{n}+k^{-2(\alpha\wedge 1)}\right\},$$ where $\frac{k^2}{n^2}+\frac{\log k}{n}$ is the optimal rate of estimating a $k$-cluster SBM in Theorem \[thm:sbm-minimax\], and $k^{-2(\alpha\wedge 1)}$ is the approximation error for an $\alpha$-smooth graphon by a $k$-cluster SBM. As a consequence, the least-squares estimator (\[eq:least-squares\]) is rate-optimal with $k\asymp n^{\frac{1}{\alpha\wedge 1+1}}$. The result justifies the strategies of estimating a nonparametric graphon by network histograms in the literature [@wolfe13; @chan14; @airoldi13; @olhede13]. Despite its rate-optimality, an disadvantage of the least-squares estimator (\[eq:least-squares\]) is its computational intractability. A naive algorithm requires an exhaustive search over all $k^n$ possible clustering structures. Although a two-way $k$-means algorithm in [@gao2016optimal] works well in practice, there is no theoretical guarantee that the algorithm can find the global optimum in polynomial time. An alternative strategy is to relax the constraint in the least-squares optimization. For instance, let ${\widetilde}{\Theta}_k$ be the set of all symmetric matrices $\theta\in[0,1]^{n\times n}$ that have at most $k$ ranks. It is easy to see $\Theta_k\subset{\widetilde}{\Theta}_k$. Moreover, the relaxed estimator ${\widehat}{\theta}={\mathop{\rm argmin}}_{\theta\in{\widetilde}{\Theta}_k}{\|A-\theta\|_{\rm F}}^2$ can be computed efficiently through a simple eigenvalue decomposition. This is closely related to the procedures discussed in [@chatterjee12]. However, such an estimator can only achieve the rate $\frac{k}{n}$, which can be much slower than the minimax rate $\frac{k^2}{n^2}+\frac{\log k}{n}$. To the best of our knowledge, $\frac{k}{n}$ is the best known rate that can be achieved by a polynomial-time algorithm so far. We refer the readers to [@xu2017rates] for more details on this topic. Extensions to sparse networks ----------------------------- In many practical situations, sparse networks are more useful. A network is sparse if the maximum probability of $\{A_{ij}=1\}$ tends to zero as $n$ tends to infinity. A sparse graphon $f$ is a symmetric nonnegative function on $[0,1]$ that satisfies $\sup_{x,y}f(x,y)\leq \rho=o(1)$ [@bickel09; @borgs2014p1; @borgs2014p2]. Analogously, a sparse SBM is characterized by the space $\Theta_k(\rho)=\{\theta\in\Theta_k:\max_{ij}\theta_{ij}\leq \rho\}$. An extension of Theorem \[thm:sbm-minimax\] is given by the following result. \[thm:minimax-sparse-graphon\] We have $$\inf_{{\widehat}{\theta}}\sup_{\theta\in\Theta_k}\mathbb{E}L({\widehat}{\theta},\theta)\asymp \min\left\{\rho\left(\frac{k^2}{n^2} + \frac{\log k}{n}\right),\rho^2\right\},$$ for all $1\leq k\leq n$. Theorem \[thm:minimax-sparse-graphon\] recovers the minimax rate of Theorem \[thm:sbm-minimax\] if we set $\rho\asymp 1$. The result was obtained independently by [@klopp2017oracle] and [@gao2016optimal] around the same time. Besides the the loss function $L(\cdot,\cdot)$ on the probability matrix, the paper [@klopp2017oracle] also considered integrated loss for the graphon function. To achieve the minimax rate, one can consider the least-squares estimator $${\widehat}{\theta}={\mathop{\rm argmin}}_{\theta\in\Theta_k(\rho)}{\|A-\theta\|_{\rm F}}^2\label{eq:sparse-least-squares}$$ when $\frac{k^2}{n^2} + \frac{\log k}{n}\geq \rho$. In the situation when $\frac{k^2}{n^2} + \frac{\log k}{n}< \rho$, the minimax rate is $\rho^2$ and can be trivially achieved by ${\widehat}{\theta}=0$. Theorem \[thm:minimax-sparse-graphon\] also leads to optimal rates of nonparametric sparse graphon estimation in a Hölder space [@klopp2017oracle; @gao2016optimal]. In addition, sparse graphon estimation in a privacy-aware setting [@borgs2015private] and a heavy-tailed setting [@borgs2015consistent] have also been considered in the literature. Biclustering and related problems --------------------------------- SBM can be understood as a special case of biclustering. A matrix has a biclustering structure if it is block-wise constant with respect to both row and column clustering structures. The biclustering model was first proposed by [@hartigan1972direct], and has been widely used in modern gene expression data analysis [@cheng2000biclustering; @madeira2004biclustering]. Mathematically, we consider the following parameter space $$\begin{aligned} \Theta_{k,l} &=& \Bigg\{\{\theta_{ij}\}\in \mathbb{R}^{n\times m}: \theta_{ij}=B_{z_1(i)z_2(j)}, B\in\mathbb{R}^{k\times l}, z_1\in[k]^n, z_2\in[l]^m\Bigg\}.\end{aligned}$$ Then, for the loss function $L({\widehat}{\theta},\theta)=\frac{1}{nm}\sum_{i=1}^n\sum_{j=1}^m({\widehat}{\theta}_{ij}-\theta_{ij})^2$, it has been shown in [@gao2015rate; @gao2016optimal] that $$\inf_{{\widehat}{\theta}}\sup_{\theta\in\Theta_{k,l}}\mathbb{E}L({\widehat}{\theta},\theta)\asymp \frac{kl}{mn} + \frac{\log k}{m} + \frac{\log l}{n}, \label{eq:biclustering-minimax}$$ as long as $\log k\asymp \log l$. The minimax rate (\[eq:biclustering-minimax\]) holds under both Bernoulli and Gaussian observations. When $k=l$ and $m=n$, the result (\[eq:biclustering-minimax\]) recovers Theorem \[thm:sbm-minimax\]. The minimax rate (\[eq:biclustering-minimax\]) reveals a very important principle of sample complexity. In fact, for a large collection of popular problems in high-dimensional statistics, the minimax rate is often in the form of $$\frac{(\#\text{parameters}) + \log(\#\text{models})}{\text{\#samples}}.\label{eq:general}$$ For the biclustering problem, $nm$ is the sample size and $kl$ is the number of parameters. Since the number of biclustering structures is $k^nl^m$, the formula (\[eq:general\]) gives (\[eq:biclustering-minimax\]). To understand the general principle (\[eq:general\]), we need to discuss the structured linear model introduced by [@gao2015general]. In the framework of structured linear models, the data can be written as $$Y=\X_Z(B)+W\in\mathbb{R}^N,$$ where $\X_Z(B)$ is the signal to be recovered and $W$ is a mean-zero noise. The signal part $\X_Z(B)$ consists of a linear operator $\X_Z(\cdot)$ indexed by the model/structure $Z$ and parameters that are organized as $B$. The structure $Z$ is in some discrete space $\Z_{\tau}$, which is further indexed by $\tau\in\T$ for some finite set $\T$. We introduce a function $\ell(\Z_{\tau})$ that determines the dimension of $B$. In other words, we have $B\in\mathbb{R}^{\ell(\Z_{\tau})}$. Then, the optimal rate that recovers the signal $\theta=\X_Z(B)$ with respect to the loss function $L({\widehat}{\theta},\theta)=\frac{1}{N}\sum_{i=1}^N({\widehat}{\theta}_i-\theta_i)^2$ is given by $$\frac{\ell(\Z_{\tau})+\log|\Z_{\tau}|}{N}.\label{eq:general-rigorous}$$ We note that (\[eq:general-rigorous\]) is a mathematically rigorous version of (\[eq:general\]). In [@gao2015general], a Bayesian nonparametric procedure was proposed to achieve the rate (\[eq:general-rigorous\]). Minimax lower bounds in the form of (\[eq:general-rigorous\]) have been investigated by [@klopp2017structured] under a slightly different framework. Below we present a few important examples of the structured linear models. *Biclustering.* In this model, it is convenient to organize $\X_Z(B)$ as a matrix in $\mathbb{R}^{n\times m}$ and then $N=nm$. The linear operator $\X_Z(\cdot)$ is determined by $[\X_Z(B)]_{ij}=B_{z_1(i)z_2(j)}$ with $Z=(z_1,z_2)$. With the relations $\tau=(k,l)$, $\T=[n]\times[m]$, $\Z_{k,l}=[k]^n\times[l]^m$, we get $\ell(Z_{k,l})=kl$ and $\log|\Z_{k,l}|=n\log k+m\log l$, and the rate (\[eq:biclustering-minimax\]) can be derived from (\[eq:general-rigorous\]). *Sparse linear regression.* The linear model $X\beta$ with a sparse $\beta\in\mathbb{R}^p$ can also be written as $\X_Z(B)$. To do this, note that a sparse $\beta$ implies a representation $\beta^T=(\beta_S^T,0_{S^c}^T)$ for some subset $S\subset[p]$. Then, $X\beta=X_{*S}\beta_S=\X_Z(B)$, with the relations $Z=S$, $\tau=s$, $\T=[p]$, $\Z_s=\{S\subset[p]:|S|=s\}$, $\ell(\Z_s)=s$ and $B=\beta_S$. Since $|\Z_s|={p\choose s}$, the numerator of (\[eq:general-rigorous\]) becomes $s+\log{p\choose s}\asymp s\log\left(\frac{ep}{s}\right)$, which is the well-known minimax rate of sparse linear regression [@donoho1994minimax; @ye2010rate; @raskutti2011minimax]. The principle (\[eq:general-rigorous\]) also applies to a more general row and column sparsity structure in matrix denoising [@ma2015volume]. *Dictionary learning.* Consider the model $\X_Z(B)=BZ\in\mathbb{R}^{n\times d}$ for some $Z\{-1,0,1\}^{p\times d}$ and $Q\in\mathbb{R}^{n\times p}$. Each column of $Z$ is assumed to be sparse. Therefore, dictionary learning can be viewed as sparse linear regression without knowing the design matrix. With the relations $\tau=(p,s)$, $\T=\{(p,s)\in[n\wedge d]\times[n]:s\leq p\}$ and $\Z_{p,s}=\{Z\in\{-1,0,1\}^{p\times d}:\max_{j\in[d]}|{{\rm supp}}(Z_{*j})|\leq s\}$, we have $\ell(\Z_{p,s})+\log|\Z_{p,s}|\asymp np+ds\log\frac{ep}{s}$, which is the minimax rate of the problem [@klopp2017structured]. The principle (\[eq:general\]) or (\[eq:general-rigorous\]) actually holds beyond the framework of structured linear models. We give an example of sparse principal component analysis (PCA). Consider i.i.d. observations $X_1,...,X_n\sim N(0,\Sigma)$, where $\Sigma=V\Lambda V^T+I_p$ belongs to the following space of covariance matrices $$\begin{aligned} \mathcal{F}(s,p,r,\lambda) &=& \Bigg\{\Sigma=V\Lambda V^T+I_p: 0<\lambda\leq \lambda_r\leq...\leq\lambda_1\leq\kappa\lambda, \\ &&\qquad V\in O(p,r), |{{\rm rowsupp}}(V)|\leq s\Bigg\},\end{aligned}$$ where $\kappa$ is a fixed constant. The goal of sparse PCA is to estimate the subspace spanned by the leading $r$ eigenvectors $V$. Here, the notation $O(p,r)$ means the set of orthonormal matrices of size $p\times r$, ${{\rm rowsupp}}(V)$ is the set of nonzero rows of $V$, and $\Lambda$ is a diagonal matrix with entries $\lambda_1,...,\lambda_r$. It is clear that sparse PCA is a covariance model and does not belong to the class of structured linear models. Despite that, it has been proved in [@cai2013sparse] that the minimax rate of the problem is given by $$\inf_{{\widehat}{V}}\sup_{\Sigma\in\mathcal{F}(s,p,r,\lambda)}\mathbb{E}{\|{\widehat}{V}{\widehat}{V}^T-VV^T\|_{\rm F}}^2\asymp\frac{\lambda+1}{\lambda} \frac{r(s-r)+s\log\frac{ep}{s}}{n}. \label{eq:minimax-SPCA}$$ The minimax rate (\[eq:minimax-SPCA\]) can be understood as the product of $\frac{\lambda+1}{\lambda}$ and $\frac{r(s-r)+s\log\frac{ep}{s}}{n}$. The second term $\frac{r(s-r)+s\log\frac{ep}{s}}{n}$ is clearly a special case of (\[eq:general\]). The first term $\frac{\lambda+1}{\lambda}$ can be understood as the modulus of continuity between the squared subspace distance used in (\[eq:minimax-SPCA\]) and the intrinsic loss function of the problem (e.g. Kullback-Leibler), because the principle (\[eq:general\]) generally holds for an intrinsic loss function. In addition to the sparse PCA problem, the minimax rate that exhibits the form of (\[eq:general\]) or (\[eq:general-rigorous\]) can also be found in sparse canonical correlation analysis (sparse CCA) [@gao2015minimax; @gao2017sparse]. Community detection {#sec:community-detection} =================== Problem settings ---------------- The problem of community detection is to recover the clustering labels $\{z(i)\}_{i\in[n]}$ from the observed adjacency matrix $\{A_{ij}\}$ in the setting of SBM (\[eq:SBM-k\]). It has wide applications in various scientific areas. Community detection has received growing interests in past several decades. Early contributions to this area focused on various cost functions to find graph clusters, in particular those based on graph cuts or modularity [@girvan2002community; @newman2002random; @newman2010networks]. Recent research has put more emphases on fundamental limits and provably efficient algorithms. In order for the clustering labels to be identifiable, we impose the following conditions in addition to (\[eq:SBM-k\]), $$\min_{1\leq u\leq k}B_{uu}\geq p,\quad \max_{1\leq u<v\leq k}B_{uv}\leq q.\label{eq:assortative}$$ This is referred to as the assortative condition, which implies that it is more likely for two nodes in the same cluster to share an edge compared with the situation where they are from two different clusters. Relaxation of the condition (\[eq:assortative\]) is possible, but will not be discussed in this survey. Given an estimator ${\widehat}{z}$, we consider the following loss function $$\ell({\widehat}{z},z)=\min_{\pi\in S_k}\frac{1}{n}\sum_{i=1}^n{{\mathbf{1}_{\left\{{{\widehat}{z}(i)\neq \pi\circ z(i)}\right\}}}}.$$ The loss function measures the misclassification proportion of ${\widehat}{z}$. Since permutations of labels correspond to the same clustering structure, it is necessary to take infimum over $S_k$ in the definition of $\ell({\widehat}{z},z)$. In ground-breaking works by [@mossel2015reconstruction; @mossel2013proof; @massoulie2014community], it is shown that the necessary and sufficient condition to find a ${\widehat}{z}$ that is positively correlated with $z$ (i.e. $\ell({\widehat}{z},z)\leq\frac{1}{2}-\delta$) when $k=2$ is $\frac{n(p-q)^2}{2(p+q)}>1$. Moreover, the necessary and sufficient condition for weak consistency ($\ell({\widehat}{z},z)\rightarrow 0$) when $k=O(1)$ is $\frac{n(p-q)^2}{2(p+q)}\rightarrow\infty$ [@mossel2014consistency]. Optimal conditions for strong consistency ($\ell({\widehat}{z},z)=0$) were studied by [@mossel2014consistency; @abbe2016exact]. When $k=2$, it is possible to construct a strongly consistent ${\widehat}{z}$ if and only if $n(\sqrt{p}-\sqrt{q})^2>2\log n$, and extensions to more general SBM settings were investigated in [@abbe2015community]. We refer the readers to a thorough and comprehensive review by [@abbe2017community] for those modern developments. Here we will concentrate on the minimax rates and algorithms that can achieve them. We favor the framework of statistical decision theory to derive minimax rates of the problem because the results automatically imply optimal thresholds in both weak and strong consistency. To be specific, the necessary and sufficient condition for weak consistency is that the minimax rate converges to zero, and the necessary and sufficient condition for strong consistency is that the minimax rate is smaller than $1/n$, because of the equivalence between $\ell({\widehat}{z},z)<1/n$ and $\ell({\widehat}{z},z)=0$. In addition, the minimax framework is very flexible and it allows us to naturally extend our results to more general degree corrected block models (DCBMs). Results for SBMs {#sec:SBM} ---------------- We first formally define the parameter space that we will work with, $$\Theta_k(p,q,\beta)=\left\{\theta=\{B_{z(i)z(j)}\}\in\Theta_k: n_u(z)\in\left[\frac{n}{\beta k},\frac{\beta n}{k}\right], B\text{ satisfies (\ref{eq:assortative})}\right\},$$ where the notation $n_u(z)$ stands for the size of the $u$th cluster, defined as $n_u(z)=\sum_{i=1}^n{{\mathbf{1}_{\left\{{z(i)=u}\right\}}}}$. We introduce a fundamental quantity that determines the signal-to-noise ratio of the community detection problem, $$I=-2\log\left(\sqrt{pq}+\sqrt{(1-p)(1-q)}\right).$$ This is the Rényi divergence of order $1/2$ between $\text{Bernoulli}(p)$ and $\text{Bernoulli}(q)$. The next theorem gives the minimax rate for $\Theta_k(p,q,\beta)$ under the loss function $\ell({\widehat}{z},z)$. \[thm:minimax-community-detection\] Assume $\frac{nI}{k\log k}\rightarrow\infty$, and then $$\inf_{{\widehat}{z}}\sup_{\Theta_k(p,q,\beta)}\mathbb{E}\ell({\widehat}{z},z)=\begin{cases} \exp\left(-(1+o(1))\frac{nI}{2}\right), & k=2, \\ \exp\left(-(1+o(1))\frac{nI}{\beta k}\right), & k\geq 3, \end{cases}\label{eq:minimax-community-detection}$$ where $1+Ck/n\leq \beta <\sqrt{5/3}$ with some large constant $C>1$ and $0<q<p<(1-c_0)$ with some small constant $c_0\in(0,1)$. In addition, if $nI/k=O(1)$, then we have $\inf_{{\widehat}{z}}\sup_{\Theta_k(p,q,\beta)}\mathbb{E}\ell({\widehat}{z},z)\asymp 1$. Theorem \[thm:minimax-community-detection\] recovers some of the optimal thresholds for weak and strong consistency results in the literature. When $k=O(1)$, weak consistency is possible if and only if $\inf_{{\widehat}{z}}\sup_{\Theta_k(p,q,\beta)}\mathbb{E}\ell({\widehat}{z},z)=o(1)$, which is equivalently the condition $nI\rightarrow\infty$ [@mossel2014consistency]. Similarly, strong consistency is possible if and only if $\frac{nI}{2}>\log n$ when $k=2$ and $\frac{nI}{\beta k}>\log n$ when $k$ is not growing too fast [@mossel2014consistency; @abbe2016exact]. Between the weak and strong consistency regimes, the minimax misclassification proportion converges to zero with an exponential rate. To understand why Theorem \[thm:minimax-community-detection\] gives a minimax rate in an exponential form, we start with a simple argument that relates the minimax lower bound to a hypothesis testing problem. We only consider the case where $3\leq k=O(1)$ and $nI\rightarrow\infty$ are satisfied, and refer the readers to [@zhang2016minimax] for the more general argument. We choose a sequence $\delta=\delta_n$ that satisfies $\delta=o(1)$ and $\log\delta^{-1}=o(nI)$. Then, we choose a $z^*\in[k]^n$ such that $n_u(z^*)\in\left[\frac{n}{\beta k}+\frac{\delta n}{k},\frac{\beta n}{k}-\frac{\delta n}{k}\right]$ for any $u\in[k]$ and $n_1(z^*)=n_2(z^*)=\ceil{\frac{n}{\beta k}+\frac{\delta n}{k}}$. Recall the notation $\mathcal{C}_u(z^*)=\{i\in[n]:z^*(i)=u\}$. Then, we choose some ${\widetilde}{\mathcal{C}}_1\subset\mathcal{C}_1(z^*)$ and ${\widetilde}{\mathcal{C}}_1\subset\mathcal{C}_1(z^*)$ such that $|{\widetilde}{\mathcal{C}}_1|=|{\widetilde}{\mathcal{C}}_2|=\ceil{n_1(z^*)-\frac{\delta n}{k}}$. Define $$T={\widetilde}{\mathcal{C}}_1\cup {\widetilde}{\mathcal{C}}_2 \cup \left(\cup_{u=3}^k \mathcal{C}_u(z^*)\right)\quad\text{and}\quad\mathcal{Z}_T=\left\{z\in[k]^n:z(i)=z^*(i)\text{ for all }i\in T\right\}.$$ The set $\mathcal{Z}_T$ corresponds to a sub-problem that we only need to estimate the clustering labels $\{z(i)\}_{i\in T^c}$. Given any $z\in\mathcal{Z}_T$, the values of $\{z(i)\}_{i\in T}$ are known, and for each $i\in T^c$, there are only two possibilities that $z(i)=1$ or $z(i)=2$. The idea is that this sub-problem is simple enough to analyze but it still captures the hardness of the original community detection problem. Now, we define the subspace $$\Theta_k^0(p,q,\beta)=\left\{\theta\in\{B_{z(i)z(j)}\}\in\Theta_k: z\in \mathcal{Z}_T, B_{uu}=p, B_{uv}=q,\text{ for all }1\leq u<v\leq k\right\}.$$ We have $\Theta_k^0(p,q,\beta)\subset\Theta_k(p,q,\beta)$ by the construction of $\mathcal{Z}_T$. This gives the lower bound $$\inf_{{\widehat}{z}}\sup_{\Theta_k(p,q,\beta)}\mathbb{E}\ell({\widehat}{z},z)\geq \inf_{{\widehat}{z}}\sup_{\Theta_k^0(p,q,\beta)}\mathbb{E}\ell({\widehat}{z},z)=\inf_{{\widehat}{z}}\sup_{z\in\mathcal{Z}_T}\frac{1}{n}\sum_{i=1}^n\mathbb{P}\{{\widehat}{z}(i)\neq z(i)\}.\label{eq:reduction}$$ The last inequality above holds because for any $z_1,z_2\in\mathcal{Z}_T$, we have $\frac{1}{n}\sum_{i=1}^n{{\mathbf{1}_{\left\{{z_1(i)\neq z_2(i)}\right\}}}}=O(\frac{\delta k}{n})$ so that $\ell(z_1,z_2)=\frac{1}{n}\sum_{i=1}^n{{\mathbf{1}_{\left\{{z_1(i)\neq z_2(i)}\right\}}}}$. Continuing from (\[eq:reduction\]), we have $$\begin{aligned} \nonumber \inf_{{\widehat}{z}}\sup_{z\in\mathcal{Z}_T}\frac{1}{n}\sum_{i=1}^n\mathbb{P}\{{\widehat}{z}(i)\neq z(i)\} &\geq& \frac{|T^c|}{n}\inf_{{\widehat}{z}}\sup_{z\in\mathcal{Z}_T}\frac{1}{|T^c|}\sum_{i\in T^c}\mathbb{P}\{{\widehat}{z}(i)\neq z(i)\} \\ \label{eq:ratio} &\geq& \frac{|T^c|}{n}\frac{1}{|T^c|}\sum_{i\in T^c}\inf_{{\widehat}{z}(i)}\text{ave}_{z\in\mathcal{Z}_T}\mathbb{P}\{{\widehat}{z}(i)\neq z(i)\}.\end{aligned}$$ Note that for each $i\in T^c$, $$\begin{aligned} \nonumber && \inf_{{\widehat}{z}(i)}\text{ave}_{z\in\mathcal{Z}_T}\mathbb{P}\{{\widehat}{z}(i)\neq z(i)\} \\ \label{eq:to-test} &\geq& \text{ave}_{z_{-i}}\inf_{{\widehat}{z}(i)}\left(\frac{1}{2}\mathbb{P}_{(z_{-i},z(i)=1)}\left({\widehat}{z}(i)\neq 1\right)+\frac{1}{2}\mathbb{P}_{(z_{-i},z(i)=2)}\left({\widehat}{z}(i)\neq 2\right)\right).\end{aligned}$$ Thus, it is sufficient to lower bound the testing error between each pair $\left(\mathbb{P}_{(z_{-i},z(i)=1)},\mathbb{P}_{(z_{-i},z(i)=2)}\right)$ by the desired minimax rate in (\[eq:minimax-community-detection\]). Note that $|T^c|\gtrsim \frac{\delta n}{k}$ with a $\delta$ that satisfies $\log\delta^{-1}=o(nI)$. So the ratio $|T^c|/n$ in (\[eq:ratio\]) can be absorbed into the $o(1)$ in the exponent of the minimax rate. The above argument leading to (\[eq:to-test\]) implies that we need to study the fundamental testing problem between the pair $\left(\mathbb{P}_{(z_{-i},z(i)=1)},\mathbb{P}_{(z_{-i},z(i)=2)}\right)$. That is, given the whole vector $z$ but its $i$th entry, we need to test whether $z(i)=1$ or $z(i)=2$. This simple vs simple testing problem can be equivalently written as $$\label{eq:hypo} \begin{aligned} & H_1: X\sim \bigotimes_{i=1}^{m_1}\text{Bern}\left(p\right) \otimes \bigotimes_{i=m_1+1}^{m_1+m_2}\text{Bern}\left(q\right) \\ & \quad \quad \quad \mbox{vs.}\quad H_2: X\sim \bigotimes_{i=1}^{m_1}\text{Bern}\left(q\right) \otimes \bigotimes_{i=m_1+1}^{m_1+m_2}\text{Bern}\left(p\right). \end{aligned}$$ The optimal testing error of (\[eq:hypo\]) is given by the following lemma. \[lem:testing-error\] Suppose that as $m_1\rightarrow\infty$, $1<p/q=O(1)$, $p=o(1)$, $\left|m_1/m_2-1\right|=o(1)$ and $m_1I\rightarrow\infty$, we have $$\inf_{\phi}\left(\mathbb{P}_{H_1}\phi+\mathbb{P}_{H_2}(1-\phi)\right)=\exp\left(-(1+o(1))m_1I\right).$$ Lemma \[lem:testing-error\] is an extension of the classical Chernoff–Stein theory of hypothesis testing for constant $p$ and $q$ (see Chapter 11 of [@cover2012elements]). The error exponent $m_1I$ is a consequence of calculating the Chernoff information between the two hypotheses in (\[eq:hypo\]). In the setting of (\[eq:to-test\]), we have $m_1=(1+o(1))m_2=(1+o(1))\frac{nI}{\beta k}$, which implies the desired minimax lower bound for $k\geq 3$ in (\[eq:minimax-community-detection\]). For $k=2$, we can slightly modify the result of Lemma \[lem:testing-error\] with asymptotically different $m_1$ and $m_2$ but of the same order. In this case, one obtains $\exp\left(-(1+o(1))\frac{m_1+m_2}{2}I\right)$ as the optimal testing error, which explains why the minimax rate in (\[eq:minimax-community-detection\]) for $k=2$ does not depend on $\beta$. The testing problem between the pair $\left(\mathbb{P}_{(z_{-i},z(i)=1)},\mathbb{P}_{(z_{-i},z(i)=2)}\right)$ is also the key that leads to the minimax upper bound. Given the knowledge of $z$ but its $i$th entry, one can use the likelihood ratio test to recover $z(i)$ with the optimal error given by Lemma \[lem:testing-error\]. Inspired by this fact, @gao2017achieving considered a two-stage procedure to achieve the minimax rate (\[eq:minimax-community-detection\]). In the first stage, one uses a reasonable initial label estimator ${\widehat}{z}^0$. This serves as an surrogate for the true $z$. Then in the second stage, one needs to solve the estimated hypothesis testing problem $$\left(\mathbb{P}_{(z_{-i}={\widehat}{z}^0_{-i},z(i)=1)},\mathbb{P}_{(z_{-i}={\widehat}{z}^0_{-i},z(i)=2)},...,\mathbb{P}_{(z_{-i}={\widehat}{z}^0_{-i},z(i)=k)}\right).\label{eq:essential-testing}$$ Since this is an upper bound procedure, we need to select from the $k$ possible hypotheses. The solution, derived by [@gao2017achieving], is given by the formula $${\widehat}{z}(i) = {\mathop{\rm argmax}}_{u\in[k]}\left(\sum_{\{j:{\widehat}{z}^0(j)=u\}}A_{ij} - {\widehat}{\rho}n_u({\widehat}{z}^0)\right). \label{eq:refinement}$$ The number ${\widehat}{\rho}$ is a data-driven tuning parameter that has an explicit formula given ${\widehat}{z}^0$ (see [@gao2017achieving] for details). The formula (\[eq:refinement\]) is intuitive. For the $i$th node, its clustering label is given by the one that has the most connections with the $i$th node, offset by the size of that cluster multiplied by ${\widehat}{\rho}$. This one-step refinement procedure enjoys good theoretical properties. As was shown in [@gao2017achieving], the minimax rate (\[eq:minimax-community-detection\]) can be achieved given a reasonable initialization such as regularized spectral clustering. In practice, after updating all $i\in[n]$ according to (\[eq:refinement\]), one can regard the current ${\widehat}{z}$ as the new ${\widehat}{z}^0$, and refine the estimator using (\[eq:refinement\]) for a second round. From our experience, this will improve the performance, and usually less than ten steps of refinement is more than sufficient. The “refinement after initialization" method is a commonly used idea in community detection to achieve exponentially small misclassification proportion. Comparable results as [@gao2017achieving] are also obtained by [@yun2014accurate; @yun2016optimal]. In addition to the likelihood-ratio-test type of refinement, the paper [@zhang2017theoretical] shows that a coordinate ascent variational algorithm also converges to the minimax rate (\[eq:minimax-community-detection\]) given a good initialization. Results for DCBMs ----------------- As was observed in [@bickel09; @zhao2012consistency], SBM is not a satisfactory model for many real data sets. An interesting generalization of SBM that captures degree heterogeneity was proposed by [@dasgupta2004spectral; @karrer2011stochastic], called degree corrected block model (DCBM). DCBM assumes that $A_{ij}\sim \text{Bernoulli}(d_id_jB_{z(i)z(j)})$. The extra sequence of parameters $(d_1,...,d_n)$ models individual sociability of network nodes. This extra flexibility is important in real-world network data analysis. However, the extra nuisance parameters $(d_1,...,d_n)$ impose new challenges for community detection. There are not many papers that extend the results of SBM in [@mossel2015reconstruction; @mossel2013proof; @massoulie2014community; @mossel2014consistency; @abbe2016exact; @abbe2015community] to DCBM. A few notable exceptions are [@zhao2012consistency; @chen2015convexified; @gulikers2015impossibility; @gulikers2017spectral]. On the other hand, the decision theoretic framework can be naturally extended from SBM to DCBM, and the results automatically imply optimal thresholds for both weak and strong consistency. We first define the parameter space of DCBMs as $$\Theta_k(p,q,\beta,d;\delta)=\left\{\theta=\{d_id_jB_{z(i)z(j)}\}: \{B_{z(i)z(j)}\}\in \Theta_k(p,q,\beta), \left|\frac{\sum_{i\in\mathcal{C}_u(z)}d_i}{n_u(z)}-1\right|\leq\delta\right\}.$$ Note that the space $\Theta_k(p,q,\beta,d;\delta)$ is defined for a given $d\in\mathbb{R}^n$. This allows us to characterize the minimax rate of community detection for each specific degree heterogeneity vector. The inequality $\left|\frac{\sum_{i\in\mathcal{C}_u(z)}d_i}{n_u(z)}-1\right|\leq\delta$ is a condition for $z$. This means that the average value of $\{d_i\}_{i\in\mathcal{C}_u(z)}$ in each cluster is roughly $1$, which implies approximate identifiability of $d,B,z$ in the model. Before stating the minimax rate, we also introduce the quantity $J$, which is defined by the following equation, $$\exp(-J) = \begin{cases} \frac{1}{n}\sum_{i=1}^n\exp\left(-d_i\frac{nI}{2}\right), & k=2, \\ \frac{1}{n}\sum_{i=1}^n\exp\left(-d_i\frac{nI}{\beta k}\right), & k\geq 3. \end{cases}\label{eq:J-def}$$ When $d_i=1$ for all $i\in[n]$, $J$ is the exponent that appears in the minimax rate of SBM in (\[eq:minimax-community-detection\]). \[thm:minimax-DCBM\] Assume $\min(J,\log n)/\log k\rightarrow\infty$, the sequence $\delta=\delta_n$ satisfies $\delta=o(1)$ and $\log\delta^{-1}=o(J)$, and $(d_1,...,d_n)$ satisfies Condition N in [@gao2016community]. Then, we have $$\inf_{{\widehat}{z}}\sup_{\Theta_k(p,q,\beta,d;\delta)}\mathbb{E}\ell({\widehat}{z},z)=\exp(-(1+o(1))J),\label{eq:minimax-DCBM}$$ where $1+Ck/n\leq \beta <\sqrt{5/3}$ and $1<p/q\leq C$ with some large constant $C>1$. With slightly stronger conditions, Theorem \[thm:minimax-DCBM\] generalizes the result of Theorem \[thm:minimax-community-detection\]. By (\[eq:J-def\]) and (\[eq:minimax-DCBM\]), the minimax rate of community detection for DCMB is an average of $\exp(-d_i\frac{nI}{2})$ or $\exp(-d_i\frac{nI}{\beta k})$, depending on whether $k=2$ or $k\geq 3$. A node with a larger value of $d_i$ will be more likely clustered correctly. Similar to SBM, the minimax rate of DCBM is also characterized by a fundamental testing problem. With the presence of degree heterogeneity, the corresponding testing problem is $$\label{eq:hypo-DCBM} \begin{aligned} & H_1: X\sim \bigotimes_{i=1}^{m_1}\text{Bern}\left(d_0d_ip\right) \otimes \bigotimes_{i=m_1+1}^{m_1+m_2}\text{Bern}\left(d_0d_iq\right) \\ & \quad \quad \quad \mbox{vs.}\quad H_2: X\sim \bigotimes_{i=1}^{m_1}\text{Bern}\left(d_0d_i q\right) \otimes \bigotimes_{i=m_1+1}^{m_1+m_2}\text{Bern}\left(d_0d_i p\right). \end{aligned}$$ Here, we use $0$ as the index of node whose clustering label is to be estimated. The optimal testing error of (\[eq:hypo-DCBM\]) is given by the following lemma. \[lem:testing-error-DCBM\] Suppose that as $m_1\rightarrow \infty$, $1<p/q=O(1)$, $p\max_{0\leq i\leq m_1+m_2}d_i^2=o(1)$, $|m_1/m_2-1|=o(1)$ and $\left|\frac{1}{m_1}\sum_{i=1}^{m_1}d_1-1\right|\vee\left|\frac{1}{m_2}\sum_{i=m_1+1}^{m_2}d_i-1\right|=o(1)$. Then, whenever $d_0m_1I\rightarrow\infty$, we have $$\inf_{\phi}\left(\mathbb{P}_{H_1}\phi+\mathbb{P}_{H_2}(1-\phi)\right)=\exp\left(-(1+o(1))d_0m_1I\right).$$ A comparison between Lemma \[lem:testing-error-DCBM\] and Theorem \[thm:minimax-DCBM\] reveals the principle that the minimax clustering error rate can be viewed as the average minimax testing error rate. Finally, we remark that the minimax rate (\[eq:minimax-DCBM\]) can be achieved by a similar “refinement after initialization" procedure to that in Section \[sec:SBM\]. Some slight modification is necessary for the method to be applicable in the DCBM setting, and we refer the readers to [@gao2016community] for more details. Initialization procedures ------------------------- In this section, we briefly discuss consistent initialization strategies so that we can apply the refinement step (\[eq:refinement\]) afterwards to achieve the minimax rate. The discussion will focus on the SBM setting. The goal is to construct an estimator ${\widehat}{z}^0$ that satisfies $\ell({\widehat}{z}^0,z)=o_{\mathbb{P}}(1)$ under a minimal signal-to-noise ratio requirement. We focus our discussion on the case $k=O(1)$. Then, we need a ${\widehat}{z}^0$ that is weakly consistent whenever $nI\rightarrow\infty$. The requirement for ${\widehat}{z}^0$ when $k\rightarrow\infty$ was given in [@gao2017achieving]. A very popular computationally efficient network clustering algorithm is spectral clustering [@shi2000normalized; @mcsherry2001spectral; @von2007tutorial; @von2008consistency; @rohe2011spectral; @chaudhuri2012spectral; @coja2010graph]. There are many variations of spectral clustering algorithms. In what follows, we present a version proposed by [@gao2016community] that avoids the assumption of eigengap. The algorithm consists of the following three steps: 1. Construct an estimator ${\widehat}{\theta}$ of $\theta=\mathbb{E}A$. 2. Compute ${\widetilde}{\theta}={\mathop{\rm argmin}}_{{\mathop{\sf rank}}(\theta)\leq k}{\|\theta-{\widehat}{\theta}\|_{\rm F}}^2$. 3. Apply $k$-means algorithm on the rows of ${\widetilde}{\theta}$, and record the clustering result by ${\widehat}{z}^0$. The three steps are highly modular and each one can be replaced by a different modification, which leads to different versions of spectral clustering algorithms [@zhou2018analysis]. The vanilla spectral clustering algorithm either chooses the adjacency matrix or the normalized graph Laplacian as ${\widehat}{\theta}$ in Step 1. Then, the $k$-means algorithm will be applied on the rows of ${\widehat}{U}$ instead of those of ${\widetilde}{\theta}$ in Step 2, where ${\widehat}{U}\in\mathbb{R}^{n\times k}$ is the matrix that consists of the $k$ leading eigenvectors. In comparison, the choice of ${\widetilde}{\theta}$ in Step 2 can be written as ${\widetilde}{\theta}={\widehat}{U}{\widehat}{\Lambda}{\widehat}{U}^T$, where ${\widehat}{\Lambda}$ is a diagonal matrix that consists of the $k$ leading eigenvalues of ${\widehat}{\theta}$. Our modified Step 1 and Step 2 make the algorithm consistent when $nI\rightarrow\infty$ without an eigengap assumption. The choice of ${\widehat}{\theta}$ in Step 1 is very important. Before discussing the requirement we need for ${\widehat}{\theta}$, we need to understand the requirement for ${\widetilde}{\theta}$. According to a standard analysis of the $k$-means algorithm (see, for example, [@lei2015consistency; @gao2016community]), a smaller $\mathbb{E}{\|{\widetilde}{\theta}-\theta\|_{\rm F}}^2$ leads to a smaller clustering error of ${\widehat}{z}^0$. Therefore, the least-squares estimator (\[eq:sparse-least-squares\]) will be the best option for ${\widetilde}{\theta}$ because it is minimax optimal (Theorem \[thm:minimax-sparse-graphon\]). However, there is no known polynomial-time algorithm to compute (\[eq:sparse-least-squares\]). On the other hand, the low-rank approximation in Step 2 can be computed efficiently through eigenvalue decomposition, and it enjoys the risk bound $$\mathbb{E}{\|{\widetilde}{\theta}-\theta\|_{\rm F}}^2=O\left(\mathbb{E}{\|{\widehat}{\theta}-\theta\|_{\rm op}}^2\right).$$ This means it is sufficient to find an ${\widehat}{\theta}$ that achieves the minimal risk in terms of the squared operator norm loss. In other words, we seek an optimal sparse graphon estimator ${\widehat}{\theta}$ with respect to the loss function ${\|\cdot\|_{\rm op}}^2$. The fundamental limit of the problem is given by the following theorem. \[thm:graphon-operator\] For $n^{-1}\leq \rho\leq 1$ and $k\geq 2$, we have $$\inf_{{\widehat}{\theta}}\sup_{\theta\in\Theta_k(\rho)}\mathbb{E}{\|{\widehat}{\theta}-\theta\|_{\rm op}}^2\asymp \rho n.$$ The minimax rate can be achieved by the estimator proposed by [@chin2015stochastic]. Define the trimming operator $T_{\tau}:A\mapsto T_{\tau}(A)$ by replacing the $i$th row and the $i$th column of $A$ with $0$ whenever $\sum_{j=1}^nA_{ij}\geq\tau$. Then, we set ${\widehat}{\theta}=T_{\tau}(A)$ with $\tau=C\frac{1}{n-1}\sum_{i=1}^n\sum_{j=1}^nA_{ij}$ for some large constant $C>0$. This estimator can be shown to achieve the optimal rate given by Theorem \[thm:graphon-operator\] [@chin2015stochastic; @gao2017achieving]. When $\rho \gtrsim \frac{\log n}{n}$ or the graph is dense, the native estimator ${\widehat}{\theta}=A$ also achieves the minimax rate. This justifies the optimality of the results in [@lei2015consistency] in the dense regime. With ${\widehat}{\theta}$ described in the last paragraph, the three steps in the algorithm are fully specified. It can be shown that $\ell({\widehat}{z}^0,z)=o(1)$ with high probability as long as $nI\rightarrow\infty$ [@gao2017achieving; @gao2016community]. Another popular version of spectral clustering is to apply $k$-means on the leading eigenvectors of the normalized graph Laplacian $L=D^{-1/2}AD^{-1/2}$, where $D$ is a diagonal degree matrix. The advantage of using graph Laplacian in spectral clustering is discussed in [@von2008consistency; @sarkar2015role]. When the graph is sparse, it is important to use regularized version of graph Laplacian [@amini2013pseudo; @qin2013regularized; @joseph2016impact], defined as $L_{\tau}=D_{\tau}^{-1/2}A_{\tau}D_{\tau}^{-1/2}$, where $(A_{\tau})_{ij}=A_{ij}+\tau/n$ and $D_{\tau}$ is the degree matrix of $A_{\tau}$. The regularization parameter plays a similar role as the $\tau$ in $T_{\tau}(A)$. Performance of regularized spectral clustering is rigorously studied by [@le2015sparse; @gao2017achieving; @le2017concentration]. For DCBM, it is necessary to apply a normalization for each row of ${\widetilde}{\theta}$ in Step 2. Instead of applying $k$-means directly on the rows of ${\widetilde}{\theta}$, it is applied on $\left\{{\widetilde}{\theta}_{1*}/\|{\widetilde}{\theta}_{1*}\|,...,{\widetilde}{\theta}_{n*}/\|{\widetilde}{\theta}_{n*}\|\right\}$. Then, the dependence on the nuisance parameter $d_i$ will be eliminated at each ratio ${\widetilde}{\theta}_{i*}/\|{\widetilde}{\theta}_{i*}\|$. The idea of normalization is proposed by [@jin2015fast] and is further developed by [@lei2015consistency; @qin2013regularized; @gao2016community]. Besides spectral clustering algorithms, another popular class of methods is semi-definite programming (SDP) [@cai2015robust; @chen2015convexified; @amini2018semidefinite]. It has been shown that SDP can achieve strong consistency with the optimal threshold of signal-to-noise ratio [@hajek2016achieving; @hajek2016achieving2]. Moreover, unlike spectral clustering, the error rate of SDP is exponential rather than polynomial [@fei2018exponential]. Some related problems --------------------- The minimax rates of community detection for both SBM and DCBM are exponential. A fundamental principle for such discrete learning problems is the connection between minimax rates and optimal testing errors. In this section, we review several other problems in the literature that share this connection. *Crowdsourcing.* In many machine learning problems such as image classification and speech recognition, we need a large amount of labeled data. Crowdsourcing provides an efficient while inexpensive way to collect labels through online platforms such as Amazon Mechanical Turk [@turk2010url]. Though massive in amount, the crowdsourced labels are usually fairly noisy. The low quality is partially due to the lack of domain expertise from the workers and presence of spammers. Let $\{X_{ij}\}_{i\in[m],j\in[n]}$ be the matrix of labels given by the $i$th worker to the $j$th item. The classical Dawid and Skene model [@dawid1979maximum] characterizes the $i$th worker’s ability by a confusion matrix $$\pi_{gh}^{(i)}=\mathbb{P}(X_{ij}=h|y_j=g),\label{eq:dawid-skene}$$ which satisfies the probabilistic constraint $\sum_{h=1}^k\pi_{gh}^{(i)}=1$. Here, $y_j$ stands for the label of the $j$th item, and it takes value in $[k]$. Given $y_j=g$, $X_{ij}$ is generated by a categorical distribution with parameter $\pi_{g*}^{(i)}=(\pi_{g1}^{(i)},...,\pi_{gk}^{(i)})$. The goal is to estimate the true labels $y=(y_1,...,y_n)$ using the observed noisy labels $\{X_{ij}\}$. With the loss function $\ell({\widehat}{y},y)=\frac{1}{n}\sum_{j=1}^n{{\mathbf{1}_{\left\{{{\widehat}{y}_j\neq y_j}\right\}}}}$, it is proved by [@gao2016exact] that under certain regularity conditions, the minimax rate of the problem is $$\inf_{{\widehat}{y}}\sup_{y\in[k]^n}\mathbb{E}\ell({\widehat}{y},y)=\exp\left(-(1+o(1))mI(\pi)\right),\label{eq:minimax-crowdsourcing}$$ where $$I(\pi)=-\max_{g\neq h}\min_{0\leq t\leq 1}\frac{1}{m}\sum_{i=1}^m\log\left(\sum_{l=1}^k\left(\pi_{gl}^{(i)}\right)^{1-t}\left(\pi_{hl}^{(i)}\right)^t\right)$$ is a quantity that characterizes the collective wisdom of a crowd. The fact that (\[eq:minimax-crowdsourcing\]) takes a similar form as (\[eq:minimax-community-detection\]) is not a coincidence. The crowdsourcing problem is essentially a hypothesis testing problem. For each $j\in[n]$, one needs to select from the $k$ hypotheses $\{H_g\}_{g\in[k]}$, with the data generating process associated with $H_g$ given by (\[eq:dawid-skene\]). *Variable selection.* Consider the problem of variable selection in the Gaussian sequence model $X_j\sim N(\theta_j,\sigma^2)$ independently for $j=1,...,d$. The parameter space of interest is defined as $$\Theta_d(s,a)=\left\{\theta\in\mathbb{R}^d: \theta_j=0\text{ if }z_j=0, \theta_j\geq a\text{ if }z_j=1, z\in\{0,1\}^d\text{ and }\sum_{j=1}^dz_j=s\right\}.$$ For any $\theta\in\Theta_d(s,a)$, there are exactly $s$ nonzero coordinates whose values are greater than or equal to $a$. With the loss function $\ell({\widehat}{z},z)=\frac{1}{s}\sum_{j=1}^d{{\mathbf{1}_{\left\{{{\widehat}{z}_j\neq z_j}\right\}}}}$, the minimax risk of variable selection derived by [@butucea2015variable] is $$\inf_{{\widehat}{z}}\sup_{\Theta_d(s,a)}\mathbb{E}\ell({\widehat}{z},z)=\Psi_+(d,s,a),\label{eq:minimax-variable-selection}$$ where the quantity $\Psi_+(d,s,a)$ is given by $$\Psi_+(d,s,a)=\left(\frac{d}{s}-1\right)\Phi\left(-\frac{a}{2\sigma}-\frac{\sigma}{a}\log\left(\frac{d}{s}-1\right)\right)+\Phi\left(-\frac{a}{2\sigma}+\frac{\sigma}{a}\log\left(\frac{d}{s}-1\right)\right).$$ The notation $\Phi(\cdot)$ is the cumulative distribution function of $N(0,1)$. Obviously, the optimal estimator that achieves the above minimax risk is the likelihood ratio test between $N(0,\sigma^2)$ and $N(a,\sigma^2)$ weighted by the knowledge of sparsity $s$. Despite the connection between variable selection and hypothesis testing, the minimax risk (\[eq:minimax-variable-selection\]) has two distinct features. First of all, the loss function is the number of wrong labels divided by $s$ instead of the overall dimension $d$. This is because the problem has an explicit sparsity constraint, which is not present in community detection or crowdsourcing. Second, given the Gaussian error, one can evaluate the minimax risk (\[eq:minimax-variable-selection\]) exactly instead of just the asymptotic error exponent. The result (\[eq:minimax-variable-selection\]) can be extended to more general settings and we refer the readers to [@butucea2015variable]. *Ranking.* Consider $n$ objects with ranks $r(1),r(2),...,r(n)\in[n]$. We observe pairwise interaction data $\{X_{ij}\}_{1\leq i\neq j\leq n}$ that follow the generating process $X_{ij}=\mu_{r(i)r(j)}+W_{ij}$. The goal is to estimate the ranks $r=(r(1),...,r(n))$ from the data matrix $\{X_{ij}\}_{1\leq i\neq j\leq n}$. A natural loss function for the problem is $\ell_0({\widehat}{r},r)=\frac{1}{n}\sum_{i=1}^n{{\mathbf{1}_{\left\{{{\widehat}{r}(i)\neq r(i)}\right\}}}}$. However, since ranks have a natural order, we can also measure the difference $|{\widehat}{r}(i)-r(i)|$ in addition to the indicator whether or not ${\widehat}{r}(i)=r(i)$. This motivates a more general $\ell_q$ loss function $\ell_q({\widehat}{r},r)=\frac{1}{n}\sum_{i=1}|{\widehat}{r}(i)-r(i)|^q$ for some $q\in[0,2]$ by adopting the convention that $0^0=0$. In particular, $\ell_1({\widehat}{r},r)$ is equivalent to the well known Kendall tau distance [@diaconis1977spearman] that is commonly used for a ranking problem. Rather than discussing the general framework in [@gao2017phase], we consider a special model with $X_{ij}\sim N(\beta(r(i)-r(j)),\sigma^2)$. Then, the minimax rate of the problem in [@gao2017phase] is given by $$\inf_{{\widehat}{r}}\sup_{r\in\mathcal{R}}\mathbb{E}\ell_q({\widehat}{r},r)\asymp\begin{cases} \exp\left(-(1+o(1))\frac{n\beta^2}{4\sigma^2}\right), & \frac{n\beta^2}{4\sigma^2}>1, \\ \left[\left(\frac{n\beta^2}{4\sigma^2}\right)^{-1}\wedge n^2\right]^{q/2}, & \frac{n\beta^2}{4\sigma^2}\leq 1. \end{cases}\label{eq:minimax-ranking}$$ The set $\mathcal{R}$ is a general class of ranks that allow ties (approximate ranking). The detailed definition is referred to [@gao2017phase]. If $\mathcal{R}$ is replaced by the set of all permutations (exact ranking without tie), then the minimax rate (\[eq:minimax-ranking\]) will still hold after $\frac{n\beta^2}{4\sigma^2}$ being replaced by $\frac{n\beta^2}{2\sigma^2}$ [@chen-minimax-ranking]. The rate (\[eq:minimax-ranking\]) exhibits an interesting phase transition phenomenon. When the signal-to-noise ratio $\frac{n\beta^2}{4\sigma^2}>1$, the minimax rate of ranking has an exponential form, much like the minimax rate of community detection in (\[eq:minimax-community-detection\]). In contrast, when $\frac{n\beta^2}{4\sigma^2}\leq 1$, the minimax rate becomes a polynomial of the inverse signal-to-noise ratio. When $\frac{n\beta^2}{4\sigma^2}>1$, the difficulty of ranking is determined by selecting among the following $n$ hypotheses $$\left(\mathbb{P}_{r_{-i},r(i)=1},\mathbb{P}_{r_{-i},r(i)=2},...,\mathbb{P}_{r_{-i},r(i)=n}\right),$$ for each $i\in[n]$. Since the error of the above testing problem is dominated by the neighboring hypotheses, i.e., we need to decide whether $\mathbb{P}_{r_{-i},r(i)=j}$ or $\mathbb{P}_{r_{-i},r(i)=j+1}$ is more likely, the minimax ranking is then given by the exponential form in (\[eq:minimax-ranking\]), where the exponent is essentially the Chernoff information between $\mathbb{P}_{r_{-i},r(i)=j}$ and $\mathbb{P}_{r_{-i},r(i)=j+1}$. Discussion {#sec:disc} ========== There are a number of open problems in the research topics we have discussed. For graphon estimation, the best rate achievable by polynomial-time algorithms has not been well understood. For community detection, even in the simple SBM setting, dependence of the tightest separation condition on the number of clusters $k$ is an interesting problem worth further investigation. Furthermore, a challenging next step in network testing is to test a composite null of a $k$-community model against a composite alternative of models with more than $k$ communities. The paper is focused on the notion of network sparsity introduced by [@bickel09; @borgs2014p1; @borgs2014p2]. Mathematically speaking, a network is sparse if $\max_{1\leq i<j\leq n}\theta_{ij}=o(1)$. However, this notion of sparsity contradicts the property of exchangeability [@lloyd2012random], which is crucial for the inferential results to be able to generalize to the entire population [@mccullagh2002statistical]. Recently, two alternative notions of network sparsity have been developed in the literature. One of the proposals considers sparse networks induced by exchangeable random measures [@caron2017sparse; @veitch2015class; @borgs2016sparse], and the other considers the notion of edge exchangeability [@crane2016edge; @crane2018probabilistic]. Unlike the framework of [@bickel09; @borgs2014p1; @borgs2014p2], these two alternative notions of sparsity allow well-defined sparse network models on the entire population, which implies a valid out-of-sample inference. However, rigorous and optimal statistical estimation and inference under these two frameworks are not well developed, except for only a few recent efforts [@todeschini2016exchangeable; @herlau2016completely]. It is natural to ask whether the current state-of-the-art techniques of network analysis discussed in this paper can be modified or generalized to analyze sparse networks in these two alternative exchangeability frameworks. This question is of obvious significance and deserves extensive efforts of future research.

--- abstract: 'We introduce a scheme that exploits laser cooling and phonon-mediated spin-spin interactions in crystals of trapped atomic ions to explore the transport of energy through a quantum magnet. We show how to implement an effective transport window to control the flow of energy through the magnet even in the absence of fermionic statistics for the carriers. This is achieved by shaping the density of states of the effective thermal reservoirs that arise from the interaction with the external bath of the modes of the electromagnetic field, and can be experimentally controlled by tuning the laser frequencies and intensities appropriately. The interplay of this transport window with the spin-spin interactions is exploited to build an analogue of the Coulomb-blockade effect in nano-scale electronic devices, and opens new possibilities to study quantum effects in energy transport.' author: - Alejandro Bermudez - Tobias Schaetz title: Quantum Transport of Energy in Controlled Synthetic Quantum Magnets --- **Introduction** ================ Richard Feynman’s visionary character has served as a source of inspiration for physicists of different fields. In his famous talk “There’s plenty of room at the bottom” [@feynman_porb], he identified a wide and diverse number of technical applications based on manipulating and controlling devices at very small scales, ultimately, [*atom by atom*]{}. Independently of their technological prospects, these small devices have the potential of displaying utterly different phenomena from their large-scale counterparts, the understanding of which lies at the forefront of current fundamental research. This situation is clearly exemplified by experiments in nanoscale electronic devices [@roukes_porb], such as the quantisation of electrical conductance [@electric_transport_quantization], and the control of currents at the level of single electrons [@coulomb_blockade]. These non-equilibrium processes are ruled by the laws of quantum mechanics, and can be considered as paradigms in the field of [*quantum transport*]{} [@nazarov_blanter]. Controlling energy (heat) transport in small-scale devices, and understanding its similarities and differences with respect to semiclassical Boltzmann-type theories in macroscopic materials [@rmp_heat_transport_phonons], is also a topic of fundamental and technological importance. Especially in the context of the shrinking computer hardware, where understanding and mitigating the mechanisms for the increasing energy dissipation is of the utmost importance. Even if the experimental study of energy transport in nano-devices is more challenging than its electronic counterpart [@roukes_porb; @rmp_heat_transport], landmarks of quantum transport have also been achieved, such as the measurement of the quantum of thermal conductance [@thermal_transport_quantization]. However, to the best of our knowledge, controlling heat transport at the level of single energy quanta, in analogy to the above electron Coulomb-blockade experiments [@coulomb_blockade], has not been considered so far. The purpose of this work is to study the occurrence of such phenomenon by considering the energy transport through an insulating quantum magnet, and to draw interesting analogies with the electronic transport in metallic nano-scale devices. Although energy flow is generally dominated by phonon transport [@rmp_heat_transport_phonons], other carriers can also have large contributions, such as electrons in nano-devices as a consequence of the large electron densities rather than a reduced temperature [@rmp_heat_transport]. In insulating magnetic materials, the contribution of spin excitations (i.e. magnons) to the energy transport has been identified by measuring thermal conductances under different temperatures and magnetic fields [@heat_tarnsport_spin_waves; @review_magnon_heat_transport]. Low-dimensional cuprate compounds provide a well-suited testbed to study spin-mediated heat transport [@spinon_heat_transport_1d], as the anisotropy allows to distinguish the contributions to the thermal conductivity of the spin excitations (i.e. spinons) from that of the phonons [@review_thermal_transport_spins_exp]. Moreover, these experiments link to a large body of theoretical studies that explore the relation of anomalous transport with the integrability of paradigmatic Heisenberg-type spin models [@review_hermal_transport_spins_theory]. Nonetheless, these studies focus on energy transport through macroscopically large magnets, whereas we are interested in microscopic devices which can be controlled at a much smaller scale, ultimately, [*spin by spin*]{}. Just as the spin degrees of freedom in single-molecule magnets can be exploited to control electronic currents [@molecular_spintronics] for molecular electronics [@molecular_elec_book], we aim at controlling and studying the quantum transport of energy at the single quantum level by exploiting the spin degrees of freedom of an insulating quantum magnet. To analyse a situation of experimental relevance, we shall rely on another of Feynman’s visions. In his lecture “Simulating physics with computers” [@feynman_qs], he put forth the possibility of mimicking the behaviour of a complex quantum model by a different, exquisitely controlled, quantum device. This idea, which has the potential of solving long-standing open questions in the field of quantum many-body physics [@qs_goals], has already found several applications within the realm of atomic, molecular and optical physics [@QS_cold_atoms; @QS_trapped_ions]. These experiments can be considered as performed in some sort of artificial quantum matter, synthesised to behave according to the models that are supposed to capture the essence of the complex phenomena observed in the condensed-matter systems. In particular, Coulomb crystals of trapped atomic ions [@wineland_review] offer a versatile playground to implement a range of synthetic one-dimensional quantum magnets [@porras_spin_models_ions] that can be designed spin by spin [@Ising_exp_ions]. We will introduce a particular scheme to study energy transport in such synthetic quantum magnets, and discuss the appearance of genuine quantum effects through analogues of the well-known electron Coulomb-blockade physics. Although we shall focus on the one-dimensional case, the connectivity of the effective spin models can be changed by considering the two-dimensional magnets realised in Penning traps [@penning_exp], the generalisation of the separate-well quantum magnets [@qs_separate_wells] to two-dimensional surface traps [@qs_surface_traps], or the flexibility of digital approaches to quantum simulations in linear traps [@digital_qs]. This article is organised as follows. In Sec. \[energy\_transport\], we describe a general microscopic model to focus on energy transport through magnetic materials. In Sec. \[sec:trapped\_ions\_energy\_tranport\], we discuss in detail how all the required ingredients of such a microscopic model can be implemented using Coulomb crystals of trapped atomic ions, and exploiting available tools that have been developed for high-precision metrology and quantum information processing. We also derive a master equation for quantum transport that describes the energy flow through a synthetic trapped-ion magnet, which is used to explore the energy-transport analogue of Coulomb blockade and single-electron transport in Sec. \[sec:blockade\]. In Sec. \[sec:measurement\], we discuss a possible strategy to detect the relevant observables and to control and probe the energy transport in the trapped-ion setup, and we present our conclusions and outlook in Sec. \[sec:outlook\]. **Energy transport in quantum magnets** {#energy_transport} ======================================= A typical experiment to study energy transport in magnetic materials [@review_thermal_transport_spins_exp] in general relies on introducing a heat source that acts as a reservoir of lattice vibrations, whose energy is then transferred onto the magnetic degrees of freedom by the inherent crystalline spin-phonon coupling (e.g. scattering of phonons by paramagnetic ions). In analogy to the theory of electron transport, we consider a pair of macroscopically-large collections of harmonic oscillators represented by the phonon modes of the crystal that act as source and drain reservoirs. Therefore, the reservoirs are described by the Hamiltonian \[reservoirs\] H\_[p]{}=H\_[p,S]{}+H\_[p,D]{}=\_[n]{}\_[n,[S]{}]{}a\^\_[n,[S]{}]{}a\^\_[n,[S]{}]{}+\_[n]{}\_[n,[D]{}]{}a\^\_[n,[D]{}]{}a\^\_[n,[D]{}]{}, where $a^{{\dagger}}_{n,r},a^{\phantom{\dagger}}_{n,r}$ are the creation annihilation operators of phonons with frequency $\omega_{n,r}$ in the $r\in\{\rm S,D\}$ source/drain reservoir labelled by a quantum number $n$ (e.g. momentum in translationally-invariant systems), and we set $\hbar=1$. The source and drain oscillators shall induce a non-equilibrium energy current through the magnet under biased thermal conditions $\rho_{\rm p}^{\rm eq}\propto{{\rm e}}^{-H_{\rm p,S}/k_{\rm B}T_{\rm S}}\otimes{{\rm e}}^{-H_{\rm p,D}/k_{\rm B}T_{\rm D}}$, where $T_{\rm S}>T_{\rm D}$, and $k_{\rm B}$ is Boltzmann’s constant. Moreover, to qualify as reservoirs, their state should not be modified by the coupling to the magnetic system $\rho_{\rm p}(t)={\rm tr}_{\rm m}\{\rho(t)\}=\rho_{\rm p}(0)$ $\forall t$, which is typically justified by considering the macroscopically-large number of degrees of freedom of the reservoirs  in comparison to those of a smaller quantum system they are connected to. The insulating magnetic system will be described by some microscopic spin-chain model \[spin\_chain\_model\] H\_[m]{}=\_ih\_i+\_[i,j]{}h\_[i,j]{}, which represents the spin-1/2 particles by Pauli matrices $\boldsymbol{\sigma}_i=(\sigma_i^x,\sigma_i^y,\sigma_i^z)$, subjected to local terms (e.g. a transverse magnetic field $h_i=-h\sigma_i^x$ ), and pairwise interactions (e.g. Heisenberg $h_{i,j}=J_{ij}\boldsymbol{\sigma}_i\cdot \boldsymbol{\sigma}_j$, or Ising $h_{i,j}=J_{ij}\sigma_i^z {\sigma}^z_j$ couplings). ![ [**Energy transport in quantum magnets:**]{} Biased source/drain reservoirs represented by a macroscopic number of harmonic oscillators in a thermal state with temperatures $T_{\rm S}>T_{\rm D}$, such that an energy current through a quantum magnet is established. The quantum magnet is depicted as a chain of interacting spins that can exchange energy with the reservoirs, such that transport occurs longitudinally along the chain [**(a)**]{}, or transversely across it [**(b)**]{}. []{data-label="fig_scheme"}](fig_1.pdf){width="0.9\columnwidth"} As stated above, it is crucial that some microscopic mechanism provides a spin-phonon coupling that allows energy to be transferred between the reservoirs and the magnetic system. Although there might be spin-Peierls-type couplings responsible for the energy exchange [@review_thermal_transport_spins_exp], which could also be synthesised in the trapped-ion magnet [@spin_peierls_ions], we shall rely on a generic spin-phonon scattering mechanism that can be described again as a pairwise spin-phonon coupling \[energy\_exchange\] H\_[mp]{}=\_[n,i]{}g\_[n,i,[[S]{}]{}]{}S\^\_ia\^\_[n,[S]{}]{}+\_[n,i]{}g\_[n,i,[[D]{}]{}]{}S\^\_ia\^\_[n,[D]{}]{}+[H.c.]{}, where we have introduced the source-drain spin-phonon couplings $g_{n,i,r}$, and a general spin operator $S^{\phantom{\dagger}}_i$ that should induce a transition between two eigenstates $\ket{\epsilon_{\ell}}\to\ket{\epsilon_{\ell'}}$ of the above spin-chain model , such that energy can be exchanged between the vibrational and magnetic degrees of freedom. Within the so-called orthodox theory of quantum electronic transport [@orthodox_theory; @single_electron_review], the corresponding transition rates, $\Gamma(\ell,\ell')$, are obtained applying Fermi’s golden rule. One can thus calculate such rates for our current problem, and use them later to describe the energy transport across the quantum magnet. In the following, we shall use this section as a guiding principle, and discuss the particular trapped-ion realisation of the above ingredients, namely [*(a)*]{} the source-drain biased reservoirs, [*(b)*]{} the synthetic quantum magnet, [*(c)*]{} the engineered spin-phonon coupling for the energy exchange mechanism, and [*(d)*]{} the analogue of the Fermi golden rule transition rates used to study the transport. According to the dependence of the spin-phonon couplings $g_{n,i,r}$ on the spin site index $i$, we can model either the longitudinal energy transport \[cf. Fig. \[fig\_scheme\][**(a)**]{}\], or transverse energy transport \[cf. Fig. \[fig\_scheme\][**(b)**]{}\]. **Coulomb crystals for energy transport** {#sec:trapped_ions_energy_tranport} ========================================= Before embarking on the above goal, let us start by reviewing the studies on energy transport through trapped-ion Coulomb crystals that have already appeared in the literature. In the case of lattice vibrations, one possibility is to consider local quenches, namely pump-probe experiments where an initial excited state with an inhomogeneous energy density is prepared by some local perturbation, and its evolution under the microscopic vibrational Hamiltonian is probed at different instants of time [@prop_vib_excitations_th_ions; @prop_vib_excitations_exp_ions]. Pump-probe experiments have also been performed for the synthetic quantum magnets [@global_quench_spins_trapped_ions; @dynamics_spins_trapped_ions]. Interestingly, for the local quenches [@dynamics_spins_trapped_ions], by measuring the spread of certain variances for a particular local perturbation [@real_time_conductivities], these experiments could address the effect of long-range terms, or additional perturbations, on the anomalous energy transport predicted for the integrable nearest-neighbour XY model [@review_hermal_transport_spins_theory]. However, finite-size effects associated to the small number of spins may obscure the results. The other possibility relies on the more standard transport scenario, where a pair of temperature-biased reservoirs is connected to the system . Unfortunately, the ion Coulomb crystals typically considered in this context are rather small, and the number of phonon modes differs markedly from the required macroscopic number of degrees of freedom of a reservoir . On the other hand, more fundamentally, the key property of an energy reservoir is that it should be capable of supplying/absorbing arbitrary amounts of energy into/from the system without being modified. Provided that the vibrational modes of the ions display this property, which can be controlled by means of laser cooling and heating, one may consider them as effective thermal reservoirs despite their finite number. In some sense, the laser cooling couples the vibrational modes to the infinite number of photonic modes in the electromagnetic bath, such that one obtains an effective reservoir. So far, this has only been considered for phonon-mediated energy transport in trapped-ion crystals, where the heat reservoirs correspond either to individually-addressed laser-cooled ions [@prop_vib_excitations_th_ions; @ion_chain_two_reservoirs_thermalization; @heat_transport_zz_ions; @heat_transport_hubard_ions], or to ions of a different species in sympathetically-cooled crystals [@heat_transport_sympathetic_toolbox_ions]. As analysed in [@heat_transport_sympathetic_toolbox_ions] (see the detailed [Supplemental Material](http://journals.aps.org/prl/supplemental/10.1103/PhysRevLett.111.040601) of that paper), there are some rather stringent conditions on the cooling rates that must be fulfilled for the laser-cooled ions to resemble a canonical transport reservoir. In particular, the cooling rates must be much larger than the vibrational couplings between distant ions. As a consequence, standard Doppler cooling by a travelling wave in a linear Paul trap does not suffice, and other cooling schemes must be adopted, such as standing-wave cooling [@laser_coolig_sw] or electromagnetic-induced-transparency (EIT) cooling [@eit_cooling]. Although EIT cooling has been demonstrated [@eit_cooling_exp], both schemes add on to the complexity of the transport setup, and it would be desirable to devise new protocols where travelling-wave Doppler cooling suffices. According to all the above discussions, we identify a two-fold interest in focusing on the energy transport through a trapped-ion synthetic magnet, rather than via lattice vibrations. As shown below, the energy transport setup shall only require travelling-wave Doppler cooling. Besides, the intrinsic non-linearities associated to the synthetic magnet shall yield a heat analog of Coulomb-blockade physics, paving the way to access a so far neglected, yet relevant and new, quantum effect in energy transport. Once this has been discussed, we can embark upon the description of the trapped-ion realisations of [*(a)*]{} the source-drain biased reservoirs, [*(b)*]{} the synthetic quantum magnet, [*(c)*]{} the engineered spin-phonon coupling for the energy exchange, and [*(d)*]{} the analogue of the Fermi golden rule transition rates used to study the energy transport. Collective transport reservoirs {#sec:reservoirs} ------------------------------- We consider a mixed Coulomb crystal of N atomic ions of two different species/isotopes (e.g. $^{25}{\rm Mg}^+$ and $^{26}{\rm Mg}^+$), confined in a linear Paul trap with frequencies $\omega_x\neq\omega_y\gg\omega_z$. The collective lattice vibrations of this crystal can be described in terms of three phonon branches H\_[p]{}=\_[n,]{}\_[n,]{}a\_[n,]{}\^a\_[n,]{}\^, where we have introduced the vibrational frequencies $\omega_{n,\alpha}$ for each normal mode $n\in\{1,\dots,N\}$ in each branch $\alpha\in\{x,y,z\}$, and the creation-annihilation operators $a_{n,\alpha}^{\dagger}, a_{n,\alpha}^{\phantom{\dagger}}$ of phonons for each of those frequencies. The ions are subjected to a laser beam tuned close to the resonance of a dipole-allowed transition of one of the atomic species, which shall be referred to as the ionic coolant (e.g. $^{26}{\rm Mg}^+$), responsible for sympathetic Doppler cooling to feature the characteristics of a energy-transport reservoir. However, we consider a regime opposite to our previous work [@heat_transport_sympathetic_toolbox_ions], by focusing on individual laser-cooling rates [@laser_coolig_sw] much smaller than the vibrational couplings between both species. In this limit, laser cooling pumps the collective vibrational modes, and not the local vibrations [@heat_transport_sympathetic_toolbox_ions], into a thermal steady state. This eases the requirements substantially on the cooling schemes discussed above, and allows us to consider the established travelling-wave cooling. If the travelling-wave beam is directed along the axis of the trap, it induces a damping of the longitudinal phonons described by a dissipation Lindblad-type [@oqs_book] super-operator, namely \[cooling\_dissipator\] \_[p]{}(\_[p]{})=\_[n]{}\_[s=+,-]{}([L\_[n,s]{}\_[p]{} L\^\_[n,s]{}-L\_[n,s]{}\^L\_[n,s]{}\_[p]{}]{})+[H.c.]{}, with the following jump operators \[jump\_op\] L\_[n,+]{}=a\_[n,z]{}\^, L\_[n,-]{}=a\_[n,z]{}\^, where the heating (cooling) rates $\Gamma_{n, +} (\Gamma_{n, -})$ can be obtained from the laser-cooling rates of a single ion $\Gamma_{\pm}(\omega_{\rm t})$ [@laser_coolig_sw] with a trap frequency corresponding to the normal mode frequency $\omega_{\rm t}=\omega_{n,z}$, after considering the normal-mode displacements $\mathcal{M}_{i,n}^z$ at the positions of the ionic coolants, \[heating\_cooling\] \_[n, ]{}=\_[i]{}(\_[i,n]{}\^z)\^2\_(\_[n,z]{}). Since the longitudinal modes are well separated in frequencies, we can focus on a couple of modes, which will play the role of the source and drain thermal reservoirs $n_{\rm S}, n_{\rm D}\in\{1,\cdots, N\}$. When the laser beam is red detuned from the dipole-allowed transition, $\Gamma_{n, +} <\Gamma_{n, -}$, the normal modes are Doppler cooled to the desired steady state, \[cooling\_steady\_state\] \_[p]{}\^[eq]{}=, where we have introduced the temperatures \[temperatures\] T\_[r]{}=,|[n]{}\_[r]{}=, and defined $\omega_r:=\omega_{n_{r},z}$ to simplify the notation. The equilibrium temperatures of the reservoirs, as well as the temperature bias $\delta T=T_{\rm S}-T_{\rm D}$, can be controlled to some extent by simply modifying the detuning of the cooling laser, as detailed below in Fig. \[fig\_Ts\]. Let us note that the range of the temperature bias could be extended by using two normal modes along different axes as the source/drain reservoirs. As far as the sympathetic Doppler cooling remains switched on during the transport experiment, and the spin-phonon exchange mechanism is weaker than the overall cooling rate, the vibrational state of the two modes remains frozen in the desired thermal state $\rho_{\rm p}(t)={\rm tr}_{\rm m}\{\rho(t)\}=\rho_{\rm p}^{\rm eq}$ $\forall t$, and the normal modes can be considered as an effective biased reservoir for transport. Here, the energy supply of the source reservoir comes from the laser driving, whereas the absorption of the excess energy in the drain reservoir is stored in the electromagnetic bath through the dipole-allowed transition. Synthetic quantum magnet ------------------------ Once we have discussed the scheme to synthesise the biased reservoirs by laser-cooled collective modes, let us describe how an interacting spin chain  can be implemented using an ion crystal [@porras_spin_models_ions]. We encode the spins $\ket{\uparrow},\ket{\downarrow}$ in two long-lived electronic states of the remaining species with energy difference $\omega_0$ (e.g. in two hyperfine states $\ket{F,M}$ of the $^{25}{\rm Mg}^+$ groundstate manifold, $\ket{\uparrow}=\ket{2,2},\ket{\downarrow}=\ket{3,3}$). Therefore, the sums in Eq.  must be restricted to the sites of the crystal where electronic degrees of freedom remain unaffected, namely to the sites of the non actively-cooled atomic ions. To induce the spin-spin interaction, we use state-dependent dipole forces that push the ions along the radial $x,y$ axes, and provide a spin-phonon coupling with the vibrational modes that are not affected by the continuously laser cooling of the longitudinal branch. While near-resonant forces are already established to implement two-qubit gates for quantum information processing [@molmer_sorensen; @didi_gate; @bermudez_gate], far-detuned forces  lead to interacting spin models where the radial phonons act as carriers of the spin-spin interactions that only get excited virtually, and can be adiabatically eliminated from the dynamics. The local terms in Eq.  correspond to ac-Stark shifts or carrier transitions [@wineland_review]. The paradigmatic synthetic spin chain , already implemented in a variety of experiments [@Ising_exp_ions], is the quantum Ising model \[Ising\] h\_i=-h\_i\^x,h\_[i,j]{}=J\_[ij]{}\_i\^z\_j\^z, which only requires a carrier for the transverse field, and a state-dependent force for the Ising interactions. Note that the relative magnitude of both terms, and the range of the antiferromagnetic interactions, can be experimentally controlled. Using a couple of state-dependent forces, each along a different axes, and exploiting the different trap frequencies $\omega_x\neq\omega_y$ [@porras_spin_models_ions], it is also possible to realise the anisotropic XY model in a tunable transverse field \[XY\] h\_i=-h\_i\^z,h\_[i,j]{}=J\^x\_[ij]{}\_i\^x\_j\^x+J\^y\_[ij]{}\_i\^y\_j\^y, where the transverse field now requires an ac-Stark shift, and all the relative magnitudes can be engineered experimentally. A particular limit, the isotropic XY model in a strong transverse field $h\gg J_{i,j}^x=J_{ij}^y$, can also be obtained by exploiting a single state-dependent dipole force under a strong transverse field, as demonstrated experimentally in [@global_quench_spins_trapped_ions; @dynamics_spins_trapped_ions]. To get access to the anisotropic regime with a single dipole force, one needs to rapidly modulate the transverse field periodically in time, as proposed in [@duality]. If we combine these ideas with another state-dependent force along the remaining axes in the same basis as the strong local term, it is possible to engineer the XXZ model \[XXZ\] h\_i=-h\_i\^z,h\_[i,j]{}=J\^\_[ij]{}(\_i\^x\_j\^x+\_i\^y\_j\^y)+J\_[ij]{}\^[||]{}\_i\^z\_j\^z, where the transverse field is stronger than the spin-spin couplings. The dynamics of such models in a strong transverse field [@global_quench_spins_trapped_ions; @dynamics_spins_trapped_ions], or the transport transport phenomena they can give rise to, can be indeed highly non-trivial and interesting. The last, and most involved possibility, would be consider three state-dependent dipole forces along all of the axes [@porras_spin_models_ions], which would lead to a XYZ model \[XYZ\] h\_i=-h\_i\^z,h\_[i,j]{}=J\^x\_[ij]{}\_i\^x\_j\^x+J\^y\_[ij]{}\_i\^y\_j\^y+J\_[ij]{}\^[z]{}\_i\^z\_j\^z, where the ratio of the transverse field and all the remaining couplings remains tunable. In order to use the laser-cooled longitudinal modes also as carriers of a spin-spin coupling, one must employ larger detunings, at the expense of obtaining weaker spin-spin couplings [@dissipation_assisted]. To study the energy transport through any of these synthetic quantum magnets, we now need to discuss the energy transfer between the spins and the laser-cooled phonons playing the role of reservoirs. Spin-phonon energy exchange --------------------------- Since the longitudinal vibrational modes of an ion chain are well-separated in frequency, we can consider a pair of laser excitations (e.g. each provided by two Raman beams for $^{25}{\rm Mg}^+$), with frequencies $\omega_{{\rm L},{\rm S}}\approx\omega_0-\omega_{\rm S}$, $\omega_{{\rm L},{\rm D}}\approx\omega_0-\omega_{\rm D}$, tuned close to the red sideband [@wineland_review] of two different normal modes $n_{\rm S}, n_{\rm D}\in\{1,\cdots, N\}$. In the resolved-sideband limit, these terms are analogous to the required energy-exchange mechanism in Eq. , namely \[energy\_exchange\] H\_[mp]{}(t)=\_[i]{}g\_[i,[[S]{}]{}]{}S\^\_ia\^\_[n\_[S]{},z]{}[[e]{}]{}\^[-[[i]{}]{}\_[[S]{}]{}t]{}+\_[i]{}g\_[i,[[D]{}]{}]{}S\^\_ia\^\_[n\_[D]{},z]{}[[e]{}]{}\^[-[[i]{}]{}\_[[D]{}]{}t]{}+[H.c.]{}, where $S_i=\sigma_i^+:=\ket{{\uparrow}_i}\bra{{\downarrow}_i}$. Here, the energy-exchange couplings $g_{i,r}$ can be obtained from the individual red-sideband couplings $g_i$ [@wineland_review] by considering the normal-mode frequencies and displacements $g_{i,r}=g_i\mathcal{M}^z_{i,n_{r}}\sqrt{\omega_z/\omega_{n_{r},z}}$, the laser detunings correspond to \[laser\_detunings\] \_[[r]{}]{}=\_[[L]{},[r]{}]{}-(\_0-\_[n\_[r]{},z]{}), and we are working in an interaction picture. Since each of the reservoirs is, in principle, coupled to all of the spins in the synthetic magnet, the trapped-ion transport scheme depicted in Fig. \[fig\_scheme\_ions\] resembles the transverse energy transport \[cf. Fig. \[fig\_scheme\][**(b)**]{}\]. Once the trapped-ion energy exchange mechanism has been introduced, and the expression of the energy-exchange couplings $g_{i,r}$ explicitly given, we can revisit the crucial point raised in Sec. \[sec:reservoirs\]: to consider the two vibrational modes as reservoirs, their state $\rho_{\rm p}(t)={\rm tr}_{\rm m}\{\rho(t)\}=\rho_{\rm p}^{\rm eq}$ $\forall t$ must remain unperturbed , regardless of their coupling with the magnetic system. This sets a limitation on the possible strengths of the energy-exchange couplings with respect to the cooling rates \[cond\_reservoir\] |g\_[i,r]{}|\_[r]{}:=2[Re]{}{\_[n\_[r]{},-]{}-\_[n\_[r]{},+]{}}, which implies that the resulting damping rates $\kappa_{r}$ from the interplay of the cooling and heating processes in Eq.  must be much stronger. In this case, the state of the laser-cooled modes is effectively frozen, such that they play the role of the thermal reservoirs. ![ [**Energy transport in an ion quantum magnet:**]{} The effective transport reservoirs correspond, in the trapped ion setup, to only a couple of normal modes that are laser cooled to different temperatures. The interacting spin model is composed of pseudo-spins corresponding to a couple of electronic levels, such that spin-spin interaction can be mediated by phonons. The energy exchange between the synthetic magnet and the reservoirs occurs transversally to the effective spin chain, however, best mediated via axial laser excitations in the resolved-sideband regime.[]{data-label="fig_scheme_ions"}](fig_3.pdf){width="1\columnwidth"} Transition rates and transport window ------------------------------------- Having introduced all the independently-controllable ingredients of the trapped-ion transport toolbox in the previous sections, we shall now describe how to calculate the analogue of the transition rates $\Gamma_{\rm i\to f}$ used in the conventional orthodox theory of quantum electronic transport [@single_electron_review]. In this theory, such rates are calculated by means of Fermi’s golden rule, and turn out to be proportional to the density of states and the Fermi-Dirac distribution of the metallic leads that act as the source/drain reservoirs. One typically assumes the wide-band limit, where the lead’s density of states $\mathfrak{D}_{r}(\epsilon)$ is featureless. On the other hand, the Fermi-Dirac distribution $\mathfrak{f}_r(\epsilon)$, which can be controlled by the external voltages biasing the leads, plays a fundamental role in determining the electric current in nano-scale devices: electrons can only propagate across the device provided that the source reservoir features electrons of the required energy to tunnel into the device, while the drain reservoir has to provide vacancies at the energy of the electron trying to enter from the device. Therefore, Pauli exclusion principle defines the electronic transport window that is crucial for the theory of quantum transport in nanostructures \[cf. Fig. \[fig\_transport\_window\][**(a)**]{}\]. In contrast, the standard setup of energy transport through quantum magnets of Sec. \[energy\_transport\], involves a pair of temperature-biased bosonic reservoirs, such that one cannot rely on the fermionic statistics to define a transport window. However, we can engineer the reservoir’s density of states in order to obtain a similar transport window, despite the Bose-Einstein statistics of the phonons \[cf. Fig. \[fig\_transport\_window\][**(b)**]{}\], while achieving a full control of its characteristics. In the following, we show how that is possible for the trapped-ion setups. ![ [**Transport window in electronic and energy currents:**]{} Transport in nano-structures can be described by the tunnelling of carriers from one reservoir onto the nano-device, and then onto the remaining reservoir. Such tunnelling events involve a transition between two energy levels $\ket{\epsilon_{\ell'}}\to\ket{\epsilon_{\ell}}$ of the nano-device, with a transition frequency $\omega_{\ell,\ell'}=\epsilon_{\ell}-\epsilon_{\ell'}$. [**(a)**]{} In electronic transport, current can flow through that particular channel if the transition frequency lies within the transport window, which is defined by the overlap of the Fermi-Dirac distributions for source electrons $\frak{f}_{\rm S}(\epsilon)$, and drain vacancies $1-\frak{f}_{\rm D},(\epsilon)$ respectively, and is represented by a shadowed grey area in the figure. This window is controlled by the bias voltages $V_{\rm S},V_{\rm D}$, and the leads temperatures $T_{\rm S}, T_{\rm D}$. [**(b)**]{} In energy transport, a similar transport window can be defined by the overlap of the density of states of the source $\mathfrak{D}_{\rm S}(\epsilon)$ and drain $\mathfrak{D}_{\rm D}(\epsilon)$ reservoirs, which can be controlled by the parameters that control the maximum $\delta_{\rm S},\delta_{\rm D}$ and width $\kappa_{\rm S},\kappa_{\rm D}$ of the source-drain density of states. []{data-label="fig_transport_window"}](fig_4.pdf){width="0.95\columnwidth"} The calculation of the transition rates is more involved in this case, as we have a mixture of coherent and dissipative dynamics, which must be treated using the formalism of quantum master equations [@oqs_book]. The use of quantum master equations in the theory of electronic transport is an alternative to more common approaches [@molecular_elec_book], such as the scattering formalism or non-equilibrium Green’s functions, and is gaining more attention recently [@timm_meq_transport]. In contrast to these other methods, and in addition to bringing up the possibility of describing transient effects, quantum master equations have the advantage that strong-correlation effects within the device need not be treated perturbatively. In our case, the mixture of coherent and dissipative dynamics forbids applying the general formalism [@timm_meq_transport] directly. However, we show that one can apply adiabatic elimination techniques in the theory of open quantum systems [@ad_elimination] in this context, and obtain a quantum master equation for energy transport through our synthetic magnet. To proceed further, let us consider the formal diagonalization of the spin model  corresponding to any of the possible synthetic realisations -, namely \[diag\_spin\] H\_[m]{}=\_\_, where $\ket{\epsilon_\ell}$ are the magnetic eigenstates, and $\epsilon_\ell$ the associated energies. Working in an interaction picture that also includes the spin model, the energy-exchange term  becomes \[mp\_spin\_basis\] H\_[mp]{}(t)=\_[,’]{}\_[r]{}\_[,’, r]{}J\_[,’]{}a\_[n\_r,z]{}[[e]{}]{}\^[[[i]{}]{}(\_[,’]{}-\_r)t]{}, where we have introduced the transition operators, transition frequencies, and transition coupling strengths J\_[,’]{}=,\_[,’]{}=\_-\_[’]{},\_[,’,r]{}=\_i[g]{}\_[i,r]{}\_i\^+. The complete coherent and dissipative elements of the dynamics can then be expressed as a master equation $\dot{\rho}(t)=(\mathcal{L}_0+\mathcal{L}_1)\rho(t)$, where $\mathcal{L}_0(\rho)=\mathcal{D}_{\rm p}(\rho)$ is the laser-cooling super-operator , and $\mathcal{L}_1(\rho)=-{{\rm i}}[H_{\rm mp}(t),\rho]$ is the Liouville super-operator associated to Eq. . Provided that the condition  is fulfilled, we can assume that the effective reservoirs remain unchanged during the whole evolution, and thus adiabatically eliminate them from the dynamics by using projection-operator techniques [@ad_elimination]. To lowest-order, the transport through the quantum magnet can be described by \[ad\_elim\] \_[m]{}(t)=[tr]{}\_[p]{}{\_0\^s\_1(t)[[e]{}]{}\^[\_0s]{}\_1(t-s)(t)}, where we have introduced the projector $\mathcal{P}O=\rho_{\rm p}^{\rm eq}\otimes{\rm tr}_{\rm p}\{O\}$ onto the reservoirs steady state , and the reduced density matrix of the quantum magnet ${\rho}_{\rm m}(t)={\rm tr}_{\rm p}\{\rho(t)\}$. The resulting master equation for the magnetic degrees of freedom, which can be obtained by applying the quantum regression theorem to evaluate the lesser/greater single-particle Green’s functions of the laser-cooled phonons, becomes \[transport\_meq\] \_[m]{}(t)=-[[i]{}]{}+\_[m]{}(\_[m]{}(t)), where we have introduced Lamb-type shifts of the quatum magnet energy levels  caused by their coupling to the source/drain $r\in\{\rm S,D\}$ reservoirs \_[,r]{}=-\_[’]{}. More relevant to the problem of quantum transport is the dissipative Lindblad-type super-operator \[transport\_dissipator\] \_[m]{}(\_[m]{})=\_[,’]{}\_[s=+,-]{}([\_[,’,s]{}\^]{}\_[,’,s]{}\^-\_[,’,s]{}\^[\_[,’,s]{}\^]{}\_[m]{})+[H.c.]{}, where we have introduced the jump operators \_[,’,+]{}=J\_[’,]{}\^,\ \_[,’,-]{}=J\_[,’]{}\^, which describe the quantum jumps $\ket{\epsilon_{\ell'}}\to\ket{\epsilon_{\ell}}$ by the transfer of an energy quantum from the reservoirs onto the magnet $\Gamma_{{\rm M}r}(\ell,\ell')$, or from the magnet onto the reservoirs $\Gamma_{r{\rm M}}(\ell,\ell')$. The transition rates can be expressed as follows \[rates\] \_[[M]{}r]{}(,’)&=2|\_[,’,r]{}|\^2\_r(\_[,’]{})\_r(\_r),\ \_[r[M]{}]{}(,’)&=2|\_[’,,r]{}|\^2\_r(\_[’,]{})(1+\_r(\_r)), which depend on the Bose-Einstein distribution of the thermal reservoirs $\frak{n}_r(\epsilon)=1/({{\rm e}}^{\epsilon/{k_{\rm B}T_r}}-1)$, and also on a Lorentzian density of states for each of the reservoirs \[lorentzian\_dos\] \_r()=. This effective density of states describes the broadening of the normal modes playing the role of the reservoirs caused by their coupling to the electromagnetic bath through the laser-cooling process, and also appears in the theory of spontaneous emission inside a leaky cavity [@lorentzian_cavity]. We have finally obtained and expression of the transition rates, which can be compared to the conventional orthodox theory of electronic quantum transport [@single_electron_review], and exploited to reavisit the discussion about the transport window at the beginning of this section. The rate $\Gamma_{r{\rm M}}(\ell,\ell')$, describing the transfer of an energy quantum from the magnet onto the reservoirs, displays a bosonic amplification $1+\frak{n}_r(\epsilon)$, which differs crucially from the fermionic suppression $1-\frak{f}_r(\epsilon)$ of the electronic case. This difference forbids the definition of an energy transport window similar to the electronic transport window of Fig. \[fig\_transport\_window\][**(a)**]{}, since the energy transfer does not require an empty level in the drain reservoir. The other crucial difference with respect to the orthodox theory of electronic transport, which assumes a featureless density of states of the metallic leads, is the appearance of a Lorenzian-shaped density of states $\mathfrak{D}_r(\epsilon)$ for the thermal reservoirs, whose centre and width can be controlled by tuning the parameters of the laser beams inducing the energy exchange , and the laser cooling , respectively. As depicted in Fig. \[fig\_transport\_window\][**(b)**]{}, one can envisage exploiting such density of states in order to engineer a similar transport window for the flow of energy quanta through the quantum magnet. This opens a vast amount of possibilities of observing interesting quantum effects in the transport of energy that had been restricted to electronic currents so far. In the context of trapped ions, this transport scheme adds onto the toolbox of [*sympathetic dissipative gadgets*]{}, where the ability to control the effective density of states, or equivalently the spectral density, has been exploited to propose schemes for dissipation-assisted two-qubit gates [@dissipation_assisted], and for the dissipative generation of multi-particle entangled states [@cooling_chain]. **Ising blockade for energy transport** {#sec:blockade} ======================================= To illustrate the prospects of the trapped-ion energy transport scheme, we apply the general formalism presented in Sec. \[sec:trapped\_ions\_energy\_tranport\] to a phenomenon of particular relevance already discussed in the introduction of this manuscript: controlling energy transport at the single quantum level by means of an analogue of the well-known Coulomb-blockade effect [@coulomb_blockade; @single_electron_review]. We discuss the energy transport through an Ising dimer connected to a pair of temperature-biased transport reservoirs. Moreover, its implementation with trapped-ion setups would be the cleanest possible experiment of the transport toolbox. ![ [**Temperatures of the laser-cooled effective reservoirs:** ]{} ([**main panel**]{}) Temperature of the laser-cooled center-of-mass mode acting as the source reservoir, and temperature difference between the center-of-mass and egyptian mode ([**inset**]{}) of a three-ion $^{25}{\rm Mg}^+$-$^{26}{\rm Mg}^+$-$^{25}{\rm Mg}^+$ crystal, as a function of the detuning of the laser with respect to a dipole allowed transition of $^{26}{\rm Mg}^+$. Parameters considered: trap frequency $\omega_z/2\pi=1{\rm MHz}$, natural decay rate $\Gamma/2\pi=41.4{\rm MHz}$, cooling laser in a travelling-wave configuration with Rabi frequency $\Omega_{\rm L}=\Gamma/2$, and various detunings $\Delta_{\rm L}$ with respect to the $3S_{1/2}$-$3P_{1/2}$ transition.[]{data-label="fig_Ts"}](fig_2.pdf){width="0.9\columnwidth"} For the sake of concreteness, let us consider a three-ion $^{25}{\rm Mg}^+$-$^{26}{\rm Mg}^+$-$^{25}{\rm Mg}^+$ crystal, where the inner ion is Doppler cooled by a travelling-wave laser beam with a frequency close to $^{26}{\rm Mg}^+$ dipole-allowed transition $3S_{1/2}$-$3P_{1/2}$. By the sympathetic cooling described in Sec. \[sec:reservoirs\], the collective longitudinal modes are continuously pumped into the thermal state . For our particular mixed-species crystal, these normal modes are only slightly perturbed with respect to those of a single-species crystal [@normal_modes]. In particular, the lowest- and highest-frequency modes, the so-called center-of-mass $n_{\rm S}=1$ and egyptian $n_{\rm D}=3$ modes, are separated in frequencies by a large energy gap. Thus, we can address them individually with the red-sideband laser excitations described in Eq. , such that these sympathetically-cooled modes play the role of the source-drain reservoirs of Eq. , whose temperature bias is depicted in Fig. \[fig\_Ts\]. To apply the general formalism of Sec. \[sec:trapped\_ions\_energy\_tranport\], let us start by diagonalising the Ising model  for a vanishing transverse field, which describes the antiferromagnetic spin-spin interactions between the two outer ions $^{25}{\rm Mg}^+$-$^{26}{\rm Mg}^+$- $^{25}{\rm Mg}^+$ mediated by the radial phonons. The eigenstates and energies are &\_=0,=, =,\ &\_[’]{}=2J,=, =. To simplify the analysis further, we exploit the configuration of the isotopes, which leads to symmetric displacements $\mathcal{M}^z_{1,n_r}=\mathcal{M}^z_{3,n_r}$ of the outer $^{25}{\rm Mg}^+$ ions in both normal modes. If the wavelengths of the Raman beams that yield the red-sideband excitations  are adjusted adequately with respect to the inter-ion distance, one finds that the energy-exchange couplings in that expression are homogeneous $g_{i,r}=g_r$. In this situation, the singlet state $\ket{\epsilon_{\rm S}}=(\ket{{\uparrow\downarrow}}-\ket{{\downarrow\uparrow}})/\sqrt{2}$ gets decoupled from the transport channels. This is the non-equilibrium transport analogue of the sub-radiant decay channel of two distant atoms coupled to the electromagnetic bath, and the associated singlet dark state in the spontaneous emission [@superadiance]. ![ [**Energy transport through an Ising dimer:**]{} The energy transport through a couple of interacting Ising spins can occur via two possible transport channels that involve the transitions $\ket{\epsilon_{\rm T}}\leftrightarrow\ket{\epsilon_{\uparrow\uparrow}}$, or $\ket{\epsilon_{\rm T}}\leftrightarrow\ket{\epsilon_{\downarrow\downarrow}}$ in the Ising magnet, induced by the absorption/emission of an energy quantum from/into the biased reservoirs. Controlling the Lorenztian density of states of the synthetic reservoirs, a single transport channel can be selected (e.g. $\ket{\epsilon_{\rm T}}\leftrightarrow\ket{\epsilon_{\downarrow\downarrow}}$, when the Lorentzians are centred at negative frequencies by using negative laser detunings). []{data-label="fig_ising_scheme"}](fig_5.pdf){width="0.8\columnwidth"} Accordingly, the energy transport gets simplified, as we only need to consider three states &\_=0,=(+)/,\ &\_[’]{}=2J,=, =. Due to the particular form of the energy-exchange term , the only transitions allowed form the ${\bf V}$-scheme depicted in Fig. \[fig\_ising\_scheme\], which corresponds to a couple of transport channels for the flow of energy through the Ising dimer. Using the general expressions , we find that the particular rates connecting $\ket{\epsilon_{\rm T}}\leftrightarrow\ket{\epsilon_{\uparrow\uparrow}}$ are governed by the effective density of states evaluated at positive frequencies \[channel\_1\] \_[r[M]{}]{}([T]{},)&=4|g\_r|\^2\_r(+2J)(1+\_r(\_r)),\ \_[[M]{}r]{}(,[T]{})&=4|g\_r|\^2\_r(+2J)\_r(\_r). Conversely, the rates connecting $\ket{\epsilon_{\rm T}}\leftrightarrow\ket{\epsilon_{\downarrow\downarrow}}$ depend on the effective density of states evaluated at negative frequencies \[channel\_2\] \_[r[M]{}]{}(,[T]{})&=4|g\_r|\^2\_r(-2J)(1+\_r(\_r)),\ \_[[M]{}r]{}([T]{},)&=4|g\_r|\^2\_r(-2J)\_r(\_r). Provided that [*(i)*]{} the centre of the Lorentzian densities , which are controlled by the red-sideband detunings $\delta_r$ , lie at negative frequencies $\delta _{r}\approx -2J$, and [*(ii)*]{} the width of the Lorentzian densities , which are controlled by the laser-cooling rate  given by the rates , fulfil $\kappa_r\ll J$, then one can ensure that only the transport channel $\ket{\epsilon_{\rm T}}\leftrightarrow\ket{\epsilon_{\downarrow\downarrow}}$ will be active \[cf. Fig. \[fig\_ising\_scheme\]\]. In this regime, the transport master equation  for the single-channel populations can be easily solved, leading to the steady state $\rho_{\rm m}^{\rm eq}=\sum_{\ell,\ell'}\rho^{\rm eq}_{\ell,\ell'}\ket{\ell}\bra{\ell'}$ that displays the following diagonal terms \[eq\_pop\] \_[,]{}\^[eq]{}=,\ \_[T,T]{}\^[eq]{}=, where we introduced $\Gamma_{\rm tot}=\sum_r\Gamma_{{\rm M}r}({\rm T},\downarrow\downarrow)+\sum_r\Gamma_{r{\rm M}}(\downarrow\downarrow,{\rm T})$, and assumed that the initial state of the quantum magnet has no population on the transport dark state (e.g. $\rho_{\rm m}(0)=\ket{\downarrow\downarrow}\bra{\downarrow\downarrow}$). Moreover, calculating the current of energy quanta flowing through the synthetic quantum magnet from the source reservoir, $ I_{\rm S}=\Gamma_{{\rm MS}}({\rm T},\downarrow\downarrow)\rho_{\downarrow\downarrow,\downarrow\downarrow}^{\rm eq}-\Gamma_{{\rm SM}}(\downarrow\downarrow,{\rm T})\rho_{\rm t,t}^{\rm eq} $, leads to the following expression \[current\_app\] I\_[S]{}=(\_[S]{}(\_[S]{})-\_[D]{}(\_[D]{})). Let us highlight the following physical predictions of this expression: [*(i)*]{} energy will flow through the synthetic magnet from the hotter source reservoir into the colder drain reservoir, since $\frak{n}_{\rm S}(\omega_{\rm S})>\frak{n}_{\rm D}(\omega_{\rm D})$. [*(ii)*]{} In general, the energy current violates Fourier’s law of heat conduction since the proportionality coefficient of $I_{\rm S}\propto\big(\frak{n}_{\rm S}\big(\omega_{\rm S})-\frak{n}_{\rm D}(\omega_{\rm D})\big)$ does also depend on the temperature of the reservoirs, and not only on features of the system. Note that violations of Fourier’s law abound at the microscopic level [@fourier_law], similarly to the violations of Ohm’s law in the electronic transport though nano-devices [@nazarov_blanter]. [*(iii)*]{} The energy current depends on the overlap of the Lorentzian densities of states of both reservoirs $I_{\rm S}\propto\frak{D}_{\rm S}(-2J)\frak{D}_{\rm D}(-2J)$ evaluated at an energy that depends on the Ising interaction strength. Unless the transport window, defined by the overlap of both Lorentzians, contains the region around $-2J$, energy transport will be blockaded $I_{\rm S}\approx0$ since the spins $\ket{\epsilon_{\downarrow\downarrow}}$ do not have the required energy to overcome the Ising gap to populate the Bell state $\ket{\epsilon_{\rm t}}$. This phenomenon is the energy-transport analogue of Coulomb blockade in electronic devices, and it can be understood as an Ising blockade mechanism of energy transport through the synthetic quantum magnet. ![ [**Ising blockade of energy transport :**]{} Asymptotic energy current through the quantum magnet as a function of the detunings of the energy-exchange mechanism. We consider the regime of a well-resolved single transport channel $\kappa_r\ll J$ in [**(a)**]{}, and study how the current is changed as one leaves this regime $\kappa_r\to J$ in [**(b)**]{}. Parameters considered: trap frequency $\omega_z/2\pi=1{\rm MHz}$, natural decay rate $\Gamma/2\pi=41.4{\rm MHz}$, cooling laser in a travelling-wave configuration with detuning $\Delta_{\rm L}=\Gamma_\tau/2$, and various cooling strengths $\Omega_{\rm L}/\gamma_\tau\in\{4\cdot 10^{-3}, 8\cdot 10^{-3},1.6\cdot 10^{-2},3.2\cdot 10^{-2}\}$, which lead to the cooling rates $\kappa_r$ of [**(b)**]{}. For the synthetic magnet, we consider far-detuned state-dependent dipole force with detuning $\delta_{\rm L}=0.1\omega_x$, for a radial trap frequency $\omega_x/2\pi=4{\rm MHz}$, and strength $Fx_0=0.1\delta_{\rm L}$, leading to antiferromagnetic Ising couplings $J/2\pi=0.16 $[kHz]{}. []{data-label="fig_ising_blockade"}](fig_6.pdf){width="0.8\columnwidth"} We now test the validity of Eq.  by a numerical comparison with the full transport master equation  containing the two transport channels with rates - for the $^{25}{\rm Mg}^+$-$^{26}{\rm Mg}^+$-$^{25}{\rm Mg}^+$ crystal. In Fig. \[fig\_ising\_blockade\][**(a)**]{}, we represent in a solid line the calculated asymptotic current after solving numerically the master equation as a function of the inverted laser detunings $-\delta_r$ of the red-sideband excitations, considering an initial state $\rho_{\rm m}(0)=\ket{\downarrow\downarrow}\bra{\downarrow\downarrow}$. We observe that the energy transport is Ising-blockaded, except for detunings $\delta_{r}\approx -2J$, where the transport window contains the channel , and energy is allowed to flow through the synthetic magnet. In the inset of this figure, we compare the Ising-blockaded oscillation with the theoretical prediction  based on the single-channel approximation, and we observe a very good agreement. As occurs in an electronic single-electron transistor [@single_electron_review], where electrons tunnel sequentially through a metallic island one by one when an applied gate voltage sets the Coulomb-blockaded channel within the transport window, the energy quanta also tunnel one by one through our synthetic magnet when the laser detunings fulfil $\delta_r\approx -2J$. Since these detunings depend on the transition frequency of the electronic levels conforming the spins $\omega_0$, and this frequency is shifted through the Zeeman effect, the role of the gate voltage in the single-electron transistor can be played by an external magnetic field in our setup, leading to an analogue single-quantum energy transistor. In Fig. \[fig\_ising\_blockade\][**(b)**]{}, we study how the Ising blockade is lifted as the transport window becomes wider and wider. We observe that, as one increases the laser-cooling rates corresponding to the Lorentzian widths $\kappa_r$, the Ising blockade peak broadens signalling that energy transport can also occur within a larger bandwidth. Moreover, the second transport channel gets activated as $\kappa_r\to J$, and we observe how a second peak in the energy current rises at negative detunings. **Energy current measurements** {#sec:measurement} =============================== The experimental study of energy transport through nano-structures is hampered by the lack of a controllable device that can measure energy/heat currents directly [@rmp_heat_transport]. This question was addressed in Ref. [@heat_transport_sympathetic_toolbox_ions], which considered a Ramsey scheme whereby an additional qubit serves as a quantum sensor to probe the relevant observables. The qubit gathers information about the mean energy current, and its fluctuation spectrum at zero frequency, while minimally perturbing the non-equilibrium steady state that supports the energy current. Unfortunately, this scheme is quite specific to the situation where energy flows via the vibrational excitations of the ion crystal. Moreover, the measurement scheme requires additional laser beams to implement a spin-dependent version of the photon-assisted tunnelling of vibrational excitations in the ion crystal [@pat_ions], which is essential to map the information of the energy current onto the phase of the qubit. In this section, we take a different approach, and devise a more direct measurement scheme for the energy current at the expense of destroying the non-equilibrium steady-state after each measurement. Therefore, the steady state of the biased quantum magnet must be prepared before each of the measurements. The main idea is that if the coupling of the synthetic magnet to the drain reservoir  is suddenly switched off, the evolution of the magnet populations at very short times encodes the expectation value of the energy current. In order to prove the above statement, let us consider the Ising-blockaded magnet of the previous section, which is described by the transport master equation  with the rates of the relevant channel . After the magnet has equilibrated to the state  during the interval $0<t<t_{\rm q}$, such that $t_{\rm q} \gg t_{\rm eq}$ and $t_{\rm eq}$ is the equilibration time, we quench the system by switching off its coupling to the drain reservoir at $t=t_{\rm q}$, namely $g_{\rm D}(t_{\rm q})=0$. The evolution of the populations for $t> t_{\rm q}$ is described by the following rate equations =\_[SM]{}(,[T]{}) \_[T,T]{}(t)-\_[MS]{}([T]{},) \_[,]{}(t),\ =\_[MS]{}([T]{},) \_[,]{}(t)-\_[SM]{}(,[T]{}) \_[T,T]{}(t). By formally integrating these equations, and introducing the solution iteratively a couple of times, one finds that \_[,]{}(t\_[q]{}+t))&=\_[,]{}(t\_[q]{})\ &+t (\_[[SM]{}]{}(,[T]{})\_[t,t]{}(t\_[q]{})-\_[[MS]{}]{}([T]{},)\_[,]{}(t\_[q]{}))\ &+(t\^2). Since the state $\rho_{\rm m}(t_{\rm q})$ at the beginning of the quench is the equilibrium state , we find that the time evolution at very short times fulfils \[en\_current\_obs\] I\_[S]{}=, t\_[[SM]{}]{}\^[-1]{},\_[[MS]{}]{}\^[-1]{}, where $I_{\rm S}$ is precisely the expectation value of the energy current in Eq. . The measurement scheme for the energy current can be thus described as follows: [*(i)*]{} Prepare the magnetic system in state $\rho_{\rm m}(0)=\ket{\downarrow\downarrow}\bra{\downarrow\downarrow}$ by optical pumping [@wineland_review], and let it equilibrate with the biased laser-cooled modes playing the role of the reservoirs for a time $t_{\rm q}$. [*(ii)* ]{} Switch off all the laser couplings, such that the populations of the quantum magnet get frozen at $t=t_{\rm q}$. [*(iii)*]{} Measure the spin population $\rho_{\downarrow\downarrow,\downarrow\downarrow}(t_{\rm q})=\bra{\downarrow\downarrow}\rho_{\rm m}(t_{\rm q})\ket{\downarrow\downarrow}$ through state-dependent fluorescence in a cycling transition of the ion crystal [@wineland_review]. [*(iv)*]{} Repeat the above steps $\it (i)-(iii)$, but quenching the coupling to the drain reservoir precisely at $t=t_{\rm q}$, and letting the system evolve for a very short additional time $\Delta t\ll\Gamma_{\rm SM}^{-1},\Gamma_{\rm MS}^{-1}$ before switching off all laser couplings to freeze the magnetic populations, thus obtaining $\rho_{\downarrow\downarrow,\downarrow\downarrow}(t_{\rm q}+\Delta t)=\bra{\downarrow\downarrow}\rho_{\rm m}(t_{\rm q}+\Delta t)\ket{\downarrow\downarrow}$. [*(iv)*]{} Infer the energy current from the ratio of Eq. . According to this scheme, one can validate our derivation of the energy current , and the quantum phenomenon of Ising blockade in the transport of energy. Let us note that it is also possible to measure how the system equilibrates with all the related transient phenomena, by sweeping over different values of $t_{\rm q}$. The available time resolution in switching on/off the laser couplings, which can easily reach sub-$\mu$s, is more than enough in the context of the slower transport dynamics of our scheme. We finally remark that, even if decoherence affects the internal degrees of freedom via external fluctuating magnetic fields, these fields are typically homogeneous along such small ion crystals, and will not modify the definition of the relevant transport channels. Moreover, since the dynamics of the populations in the transport is described by rate equations, such that coherences between the different transport channels are not relevant, the predicted Coulomb blockade in the energy transport should be insensitive to such sources of decoherence. Only the presence of fluctuating magnetic-field gradients can induce a coupling between the triplet and singlet levels, such that a more general description of the transport channels might be required. Anyhow, since the Coulomb-blockade effect relies on the energetic splitting of the different levels, its signatures should still be present in this situation. **Conclusions and Outlook** {#sec:outlook} =========================== We have proposed a scheme to study the quantum transport of energy through synthetic quantum magnets implemented with Coulomb crystals of trapped atomic ions. By exploiting sympathetic cooling in a mixed Coulomb crystal, a pair of temperature-biased thermal reservoirs can be mimicked via two laser-cooled longitudinal modes of the crystal. The analogue of the microscopic scattering mechanism that leads to the energy exchange between lattice vibrations and the magnetic moments required for energy transport in in solids can be realised by resolved laser-driven red-sideband couplings, whereas the interactions within the magnet can be designed at will by phonon-mediated spin-spin couplings induced by far-detuned state-dependent dipole forces. We have derived a general quantum master equation for the transport of energy, and applied it to the particular case of an Ising dimer, where we have found the energy-transport counterpart of the Coulomb blockade mechanism. This opens the possibility of exploring a variety of quantum effects in energy/heat transport that had been previously restricted to the realm of electronic transport. By quenching the coupling of the magnet to the drain reservoir, and measuring certain populations by state-dependent fluorescence of the ion crystal for very short times, we have shown that the energy current can be inferred, and the prediction of the blockade can be addressed in a trapped-ion experiment. The proposed toolbox can also address effects that go beyond the derived quantum master equation, and would require a systematic extension to higher orders of the adiabatic elimination in Eq. . For instance, since the ratio of the energy-exchange coupling and the Ising interaction strength can be modified at will, one can study the effects of energy co-tuneling in the Ising-blockaded regime or, ultimately, the possibility of observing the energy counterpart of the Kondo effect in quantum electronic transport [@bruus_flensberg]. By adding a transverse field to the Ising model, and considering larger ion chains, one could explore the consequences of quantum phase transition in the energy transport, and even search for the predicted signatures of quantum criticality in the observed transverse currents [@ising_chain_transverse_transport]. [****]{} A. B. acknowledges support from Spanish MINECO Project FIS2012-33022, and CAM regional research consortium QUITEMAD+. T.S. is supported by DFG (SCHA 973). A. B. thanks the hospitality of FRIAS (Freiburg Institute of Advanced Science) within the research focus on “Designed quantum transport in complex materials”, where parts of this work were developed. R. P. Feynman, [Engineering and Science, [**23**]{}, 22 (1960)](http://calteches.library.caltech.edu/47/). M. Roukes, [Sci. Am. [**285,**]{} 48 (2001)](http://www.nature.com/scientificamerican/journal/v285/n3/index.html). B. J. van Wees, H. van Houten, C. W. J. Beenakker, J. G. Williamson, L. P. Kouwenhoven, D. van der Marel, and C. T. Foxon, [Phys. Rev. Lett. [**60,**]{} 848 (1988)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.60.848); D. A. Wharam, T. J. Thornton, R. Newbury, M. Pepper, H. Ahmed, J. E. F. Frost, D. G. Hasko, D. C. Peacock, D. A. Ritchie and G. A. C. Jones, [J. Phys. C [**21,**]{} L209 (1988)](http://iopscience.iop.org/article/10.1088/0022-3719/21/8/002/meta). T. A. Fulton and G. J. Dolan, [Phys. Rev. Lett. [**59,**]{} 109 (1987)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.59.109). Y. V. Nazarov and Y. M. Blanter, [*Quantum Transport*]{} (Cambridge University Press, Cambridge, 2009). P. Carruthers, [Rev. Mod. Phys. [**33,**]{} 92 (1961)](http://journals.aps.org/rmp/abstract/10.1103/RevModPhys.33.92). Y. Dubi and M. Di Ventra, [Rev. Mod. Phys. [**83,**]{} 131 (2011)](http://journals.aps.org/rmp/abstract/10.1103/RevModPhys.83.131). K. Schwab, E. A. Henriksen, J. M. Worlock, and M. L. Roukes, [ Nature [**404,**]{} 974 (2000)](http://www.nature.com/nature/journal/v404/n6781/abs/404974a0.html). B. Lüthi, [J. Phys. Chem. Solids [**23,**]{} 35 (1962)](http://www.sciencedirect.com/science/article/pii/0022369762900549); R.L. Douglass, [Phys. Rev. [**129,**]{} 1132 (1963)](http://journals.aps.org/pr/abstract/10.1103/PhysRev.129.1132). D. J. Sanders and D. Walton, [Phys. Rev. B [**15,**]{} 1489 (1977)](http://journals.aps.org/prb/abstract/10.1103/PhysRevB.15.1489), [*and references therein*]{}. A. V. Sologubenko, K. Gianno, H. R. Ott, U. Ammerahl, and A. Revcolevschi, [Phys. Rev. Lett. [**84,**]{} 2714 (2000)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.84.2714); A. V. Sologubenko, K. Gianno, H. R. Ott, A. Vietkine, and A. Revcolevschi, [ Phys. Rev. B [**64,**]{} 054412 (2001)](http://journals.aps.org/prb/abstract/10.1103/PhysRevB.64.054412); C. Hess, C. Baumann, U. Ammerahl, B. Büchner, F. Heidrich-Meisner, W. Brenig, and A. Revcolevschi, [Phys. Rev. B [**64,**]{} 184305 (2001)](http://journals.aps.org/prb/abstract/10.1103/PhysRevB.64.184305). A. V. Sologubenko and H. R. Ott, in [*Strong Interactions in Low Dimensions*]{}, D. Baeriswyl, L. Degiorgi, Eds. (Kluwer Ac. Pub., Dordrech, 2004), [*and references therein*]{}. X. Zotos, and P. Prelovskev, in [*Strong Interactions in Low Dimensions*]{}, D. Baeriswyl, L. Degiorgi, Eds. (Kluwer Ac. Pub., Dordrech, 2004), [*and references therein*]{}. L. Bogani and W. Wernsdorfer, [Nat. Mat. [**7,**]{} 179 (2008)](http://www.nature.com/nmat/journal/v7/n3/full/nmat2133.html). J. C. Cuevas and E. Scheer, [*Molecular Electronics: An Introduction to Theory and Experiment*]{} (World Scientific, London, 2010). R. P. Feynman, [Int. J. Theor. Phys. [**21,**]{} 467 (1982)](http://link.springer.com/article/10.1007%2FBF02650179). J. I. Cirac and P. Zoller, [Nat. Phys. [**8,**]{} 264 (2012)](http://www.nature.com/nphys/journal/v8/n4/full/nphys2275.html). I. Bloch, J. Dalibard, and S. Nascimbène, [Nat. Phys. [**8,**]{} 267 (2012)](http://www.nature.com/nphys/journal/v8/n4/full/nphys2259.html), [*and references therein*]{}. R. Blatt and C. F. Roos, [Nat. Phys. [**8,**]{} 277 (2012)](http://www.nature.com/nphys/insight/quantum-simulation/index.html), [*and references therein.*]{} D. J. Wineland, C. Monroe, W. M. Itano, D. Leibfried, B. E. King, and D. M. Meekhof, [J. Res. Natl. I. St. Tech. [**103,**]{} 259 (1998)](http://nvlpubs.nist.gov/nistpubs/jres/103/3/cnt103-3.htm). D. Porras and J. I. Cirac, [Phys. Rev. Lett. [**92,**]{} 207901 (2004)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.92.207901); X.-L. Deng, D. Porras, and J. I. Cirac, [Phys. Rev. A [**72,**]{} 063407 (2005)](http://journals.aps.org/pra/abstract/10.1103/PhysRevA.72.063407). A. Friedenauer, H. Schmitz, J. T. Glueckert, D. Porras, and T. Schaetz, [Nat. Phys. [**4,**]{} 757 (2008)](http://www.nature.com/nphys/journal/v4/n10/full/nphys1032.html); K. Kim, M.-S. Chang, S. Korenblit, R. Islam, E. E. Edwards, J. K. Freericks, G.-D. Lin, L.-M. Duan, and C. Monroe, [Nature [**465,**]{} 590 (2010)](http://www.nature.com/nature/journal/v465/n7298/full/nature09071.html); R. Islam, E. E. Edwards, K. Kim, S. Korenblit, C. Noh, H. Carmichael, G.-D. Lin, L.-M. Duan, C.-C. Joseph Wang, J.K. Freericks, and C. Monroe, [Nat. Comm. [**2,**]{} 377 (2011)](http://www.nature.com/ncomms/journal/v2/n7/full/ncomms1374.html); A. Khromova, Ch. Piltz, B. Scharfenberger, T. F. Gloger, M. Johanning, A. F. Varón, and Ch. Wunderlich, [ Phys. Rev. Lett. [**108,**]{} 220502 (2012)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.108.220502) ; R. Islam, C. Senko, W. C. Campbell, S. Korenblit, J. Smith, A. Lee, E. E. Edwards, C.-C. J. Wang, J. K. Freericks, and C. Monroe, [Science [**340,**]{} 583 (2013)](http://www.sciencemag.org/content/340/6132/583.abstract); C. Senko, J. Smith, P. Richerme, A. Lee, W. C. Campbell, C. Monroe, [Science [**345,**]{} 430 (2014)](http://www.sciencemag.org/content/345/6195/430.abstract); Ch. Piltz, Th. Sriarunothai, S. Ivanov, S. Wölk, and Ch. Wunderlich, [arXiv:1509.01478 (2015)](http://arxiv.org/abs/1509.01478). J. W. Britton, B. C. Sawyer, A. C. Keith, C.-C. J. Wang, J. K. Freericks, H. Uys, M. J. Biercuk, and J. J. Bollinger, [ Nature [**484,**]{} 489 (2012)](http://www.nature.com/nature/journal/v484/n7395/full/nature10981.html). A. C. Wilson, Y. Colombe, K. R. Brown, E. Knill, D. Leibfried, and D. J. Wineland, [Nature [**512,**]{} 57 (2014).](http://www.nature.com/nature/journal/v512/n7512/full/nature13565.html) T. Schaetz, A. Friedenauer, H.Schmitz, L. Petersen, and S. Kahra, [J. Mod. Opt. [**54,**]{} 2317 (2007)](http://www.tandfonline.com/doi/abs/10.1080/09500340701639631#.VhU-qOl7F20); R. Schmied, J. H. Wesenberg, D. Leibfried, [Phys. Rev. Lett. [**102,**]{} 233002 (2009)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.102.233002). B. P. Lanyon, C. Hempel, D. Nigg, M. Müller, R. Gerritsma, F. Zähringer, P. Schindler, J. T. Barreiro, M. Rambach, G. Kirchmair, M. Hennrich, P. Zoller, R. Blatt, and C. F. Roos, [Science [**334,**]{} 57 (2011)](http://www.sciencemag.org/content/334/6052/57). A. Bermudez, and M. B. Plenio, [ Phys. Rev. Lett. [**109,**]{} 010501 (2012)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.109.010501) D. V. Averin, and K. K. Likharev, in [*Mesoscopic Phenomena in Solids* ]{}, Eds. B. L. Altshuler, P. A. Lee, and R. A. Webb, (North-Holland, Amsterdam, 1991). G . Schön, in [*Quantum Transport and Dissipation* ]{}, Eds. T. Dittrich, P. Hänggi, G.-L. Ingold, B. Kramer, G. Schön, and W. Zwerger (Wiley-VCH Verlag, Wienheim, 1998), [*and references therein.*]{} T. Pruttivarasin, M. Ramm, I. Talukdar, A. Kreuter, and H. Häffner, [New J. Phys. [**13,**]{} 075012 (2011)](http://iopscience.iop.org/article/10.1088/1367-2630/13/7/075012/meta). M. Ramm, T. Pruttivarasin, and H. Häffner, [New J. Phys. [**16,**]{} 063062, (2014)](http://iopscience.iop.org/article/10.1088/1367-2630/16/6/063062/meta). P. Richerme, Z.-X. Gong, A. Lee, C. Senko, J. Smith, M. Foss- Feig, S. Michalakis, A. V. Gorshkov, and C. Monroe, [Nature [**511,**]{} 198 (2014)](http://www.nature.com/nature/journal/v511/n7508/full/nature13450.html). P. Jurcevic, B. P. Lanyon, P. Hauke, C. Hempel, P. Zoller, R. Blatt, C. F. Roos, [Nature [**511,**]{} 202 (2014)](http://www.nature.com/nature/journal/v511/n7508/full/nature13461.html). S. Langer, M. Heyl, I. P. McCulloch, and F. Heidrich-Meisner, [Phys. Rev. B [**84,**]{} 205115 (2011)](http://journals.aps.org/prb/abstract/10.1103/PhysRevB.84.205115). G.-D. Lin and L.-M. Duan, [New J. Phys. [**13,**]{} 075015 (2011)](http://iopscience.iop.org/article/10.1088/1367-2630/13/7/075015/meta). A. Ruiz, D. Alonso, M. B. Plenio, and A. del Campo, [Phys. Rev. B [**89,**]{} 214305 (2014)](http://journals.aps.org/prb/abstract/10.1103/PhysRevB.89.214305). C. Guo, M. Mukherjee, and D. Poletti, [Phys. Rev. A [**92,**]{} 023637 (2015)](http://journals.aps.org/pra/abstract/10.1103/PhysRevA.92.023637). A. Bermudez, M. Bruderer, and M. B. Plenio, [Phys. Rev. Lett. 111, 040601 (2013)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.111.040601). J. I. Cirac, R. Blatt, P. Zoller, and W. D. Phillips, [Phys. Rev. A [**46,**]{} 2668 (1992)](http://journals.aps.org/pra/abstract/10.1103/PhysRevA.46.2668). G. Morigi, J. Eschner, and C. H. Keitel, [Phys. Rev. Lett. [**85,**]{} 4458 (2000)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.85.4458). C. F. Roos, D. Leibfried, A. Mundt, F. Schmidt-Kaler, J. Eschner, and R. Blatt, [Phys. Rev. Lett. [**85,**]{} 5547 (2000)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.85.5547); Y. Lin, J. P. Gaebler, T. R. Tan, R. Bowler, J. D. Jost, D. Leibfried, and D. J. Wineland, [Phys. Rev. Lett. [**110,**]{} 153002 (2013)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.110.153002). H. P. Breuer and F. Petruccione, [*The theory of open quantum systems*]{}, (Oxford University Press, Oxford, 2003). C. Timm, [Phys. Rev. B [**77,**]{} 195416 (2008)](http://journals.aps.org/prb/abstract/10.1103/PhysRevB.77.195416); [*ibid.*]{} [Phys. Rev. B [**83,**]{} 115416 (2011)](http://journals.aps.org/prb/abstract/10.1103/PhysRevB.83.115416); [*and references therein.*]{} S. Chaturvedi and F. Shibata, [Zeit. Physik B [**35,**]{} 297 (1979)](http://link.springer.com/article/10.1007/BF01319852). A. Sorensen and K. Molmer, [Phys. Rev. Lett. [**82,**]{} 1971 (1999)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.82.1971); A. Sorensen, and K. Molmer, [Phys. Rev. A [**62,**]{} 022311 (2000)](http://journals.aps.org/pra/abstract/10.1103/PhysRevA.62.022311) ; C. A. Sackett, D. Kielpinski, B. E. King, C. Langer, V. Meyer, C. J. Myatt, M. Rowe, Q. A. Turchette, W. M. Itano, D. J. Wineland, and C. Monroe, [Nature [**404,**]{} 256 (2000)](http://www.nature.com/nature/journal/v404/n6775/abs/404256a0.html). G. J. Milburn, S. Schneider, and D.F.V. James, [Fortschr. Phys. [**48,**]{} 801 (2000)](http://onlinelibrary.wiley.com/doi/10.1002/1521-3978(200009)48:9/11%3C801::AID-PROP801%3E3.0.CO;2-1/abstract;jsessionid=827087CD14FD217606F50B32375CDB96.d01t01); D. Leibfried, B. DeMarco, V. Meyer, D. Lucas, M. Barrett, J. Britton, W. M. Itano, B. Jelenkovic, C. Langer, T. Rosenband, and D. J. Wineland, [Nature [**422,**]{} 412 (2003)](http://www.nature.com/nature/journal/v422/n6930/abs/nature01492.html). A. Bermudez, P. O. Schmidt, M. B. Plenio, and A. Retzker, [ Phys. Rev. A [**85,**]{} 040302(R) (2012)](http://journals.aps.org/pra/abstract/10.1103/PhysRevA.85.040302); A. Lemmer, A. Bermudez, M. B. Plenio, [New J. Phys. [**15,**]{} 083001 (2013)](http://iopscience.iop.org/article/10.1088/1367-2630/15/8/083001/meta); T. R. Tan, J. P Gaebler, R. Bowler, Y. Lin, J. D. Jost, D. Leibfried, and D. J. Wineland, [Phys. Rev. Lett. [**110,**]{} 263002 (2013)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.110.263002). T. Graß, M. Lewenstein, and A. Bermudez, [ arXiv:1511.04323 (2015).](http://arxiv.org/abs/1511.04323) A. Bermudez, T. Schaetz, M. B. Plenio, [Phys. Rev. Lett. [**110,**]{} 110502 (2013)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.110.110502). A. Imamoglu, [Phys. Rev. A [**50,**]{} 3650 (1994)](http://journals.aps.org/pra/abstract/10.1103/PhysRevA.50.3650); B. M. Garraway, [Phys. Rev. A [**55,**]{} 2290 (1997)](http://journals.aps.org/pra/abstract/10.1103/PhysRevA.55.2290). C. Cormick, A. Bermudez, S. F. Huelga, and M. B. Plenio, [New J. Phys. [**15,**]{} 073027 (2013)](http://iopscience.iop.org/article/10.1088/1367-2630/15/7/073027/meta). D.F.V. James, [Appl. Phys. B [**66,**]{} 181 (1998)](http://link.springer.com/article/10.1007%2Fs003400050373#page-1). R. H. Dicke, [Phys. Rev. [**93,**]{} 99 (1954)](http://journals.aps.org/pr/abstract/10.1103/PhysRev.93.99); R. H. Lemberg, [Phys. Rev. A [**2,**]{} 889 (1970)](http://journals.aps.org/pra/abstract/10.1103/PhysRevA.2.889); Z. Ficek and R. Tanas, [Phys. Rep. [**372,**]{} 369 (2002)](http://www.sciencedirect.com/science/article/pii/S037015730200368X). S. Lepri, R. Livi, and A. Politi, [Phys. Rep. [**377,**]{} 1 (2003)](http://www.sciencedirect.com/science/article/pii/S0370157302005586), [*and references therein.*]{} A. Bermudez, T. Schaetz, and D. Porras, [Phys. Rev. Lett. [**107,**]{} 150501 (2011)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.107.150501); A. Bermudez, T. Schaetz, and D. Porras, [ New J. Phys. [**14,**]{} 053049 (2012)](http://iopscience.iop.org/article/10.1088/1367-2630/14/5/053049/meta). H. Bruus and K. Flensberg, [*Many-Body Quantum Theory in Condensed Matter Physics*]{} (Oxford University Press, Oxford, 2004). M. Vogl, G. Schaller, and T. Brandes, [Phys. Rev. Lett. [**109,**]{} 240402 (2012)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.109.240402).

--- abstract: 'Resilience is a rehashed concept in natural hazard management — resilience of cities to earthquakes, to floods, to fire, etc. In a word, a system is said to be resilient if there exists a strategy that can drive the system state back to “normal” after any perturbation. What formal flesh can we put on such a malleable notion? We propose to frame the concept of resilience in the mathematical garbs of control theory under uncertainty. Our setting covers dynamical systems both in discrete or continuous time, deterministic or subject to uncertainties. We will say that a system state is resilient if there exists an adaptive strategy such that the generated state and control paths, contingent on uncertainties, lay within an acceptable domain of random processes, called recovery regimes. We point out how such recovery regimes can be delineated thanks to so called risk measures, making the connection with resilience indicators. Our definition of resilience extends others, be they “à la Holling” or rooted in viability theory. Indeed, our definition of resilience is a form of controlability for whole random processes (regimes), whereas others require that the state values must belong to an acceptable subset of the state set.' author: - 'Michel De Lara[^1]' title: | A Mathematical Framework for Resilience:\ Dynamics, Uncertainties, Strategies\ and Recovery Regimes --- Introduction {#Introduction} ============ Consider a system whose state evolves with time, being subject to a dynamics driven both by controls and by external perturbations. The system is said to be resilient if there exists a strategy that can drive the system state towards a normal regime, whatever the perturbations. Basic references are [@Holling:73; @Martin:2004; @Martin-Deffuant-Calabrese:2011; @Rouge-Mathias-Deffuant:2013; @Arnoldi-Loreau-Haegeman:2016]. In the case of fisheries, the state can be a vector of abundances at ages of one or several species; the control can be fishing efforts; the external perturbations can affect mortality rates or birth functions appearing in the dynamics (an extreme perturbation could be an El Niño event, affecting the populations renewal). In the case of a city exposed to earthquakes, floods or other climatic events, the state can be a vector of capital stocks (energy reserves, energy production units, water treatment plants, health units, etc.); the controls would be the different investments in capital as well as current operations (flows in and out capital stocks); the dynamics would express the changes in the stocks due to investment and to day to day operations; external perturbations (rain, wind, climatic events, etc.) would affect the stocks by reducing them, possibly down to zero. In Sect. \[Ingredients\_for\_an\_abstract\_control\_system\_with\_uncertainties\], we introduce basic ingredients from the mathematical framework of control theory under uncertainty. Thus equipped, we frame the concept of resilience in mathematical garbs in Sect. \[Resilience:\_A\_Mathematical\_Framework\]. Then, in Sect. \[Illustrations\], we provide illustrations of the abstract general framework and compare our approach with others, “à la Holling” or the stochastic viability theory approach to resilience. In Sect. \[Resilience\_and\_risk\_control/minimization\], we sketch how concepts from risk measures (introduced initially in mathematical finance) can be imported to tackle resilience issues. Finally, we discuss pros and cons of our approach to resilience in Sect. \[sec:conclusions\]. Ingredients for an abstract control system with uncertainties {#Ingredients_for_an_abstract_control_system_with_uncertainties} ============================================================= We outline the mathematical formulations of time, controls, states, Nature (uncertainties), flow (dynamics) and strategies. As the reference [@Rouge-Mathias-Deffuant:2013] is the more mathematically driven paper on resilience, we will systematically emphasize in what our approach differs from that of Rougé, Mathias and Deffuant. Time, states, controls, Nature and flow --------------------------------------- We lay out the basic ingredients of control theory: time, states, controls, Nature (uncertainties) and flow (dynamics). ### Time The *time set* ${{\mathbb T}}$ is a (nonempty) subset of the real line ${{\mathbb R}}$. The set ${{\mathbb T}}$ holds a minimal element ${t_0}\in {{\mathbb T}}$ and an upper bound ${T}$, which is either a maximal element when ${T}<+\infty$ (that is, ${T}\in {{\mathbb T}}$) or not when ${T}=+\infty$ (that is, $+\infty \not\in {{\mathbb T}}$). For any couple ${(s,t)} \in {\big({{\mathbb R}}\cup \{ +\infty \} \big)}^2$, we use the notation $${s\;\!\!:\;\!\!t}= {\left\{ r \in {{\mathbb T}}\:\left|\: s \leq r \leq t \right.\right\}} $$ for the *segment* that joins $s$ to $t$ (when $s > t$, $ {s\;\!\!:\;\!\!t}= \emptyset $). #### Special cases of discrete and continuous time. This setting includes the discrete time case when ${{\mathbb T}}$ is a discrete set, be it infinite like ${{\mathbb T}}=\{{t_0}+k{\Delta\;\!\!t}, \quad k \in {{\mathbb N}}\}$ (with ${\Delta\;\!\!t}>0$), or finite like ${{\mathbb T}}=\{{t_0}+k{\Delta\;\!\!t}, \quad k =0,1,\ldots, K \}$ (with $K \in {{\mathbb N}}$). Of course, in the continuous time case, ${{\mathbb T}}$ is an interval of ${{\mathbb R}}$, like $[{t_0},{T}]$ when ${T}<+\infty$ or $[{t_0},+\infty[$ when ${T}=+\infty$. But the setting makes it possible to consider interval of continuous times separated by discrete times corresponding to jumps. For these reasons, our setting is more general than the one in [@Rouge-Mathias-Deffuant:2013], which considers discrete time systems. #### Environmental illustration. In fisheries, investment decisions (boats, equipment) are made at large time scale (years), regulations quotas are generally annual, boat operations are daily. By contrast, populations and external perturbations evolve in continuous time. Depending on the issues at hand, the modeler will choose the proper time scales, symbolized by the set ${{\mathbb T}}$. In an energy system, like a micro-grid with battery and solar panels, investment decisions in equipment (buying or renewal) occur at large time scale, whereas flows inside the system have to be decided at short time scales (minutes). ### States, controls, Nature At each time $t \in {{\mathbb T}}$, - the system under consideration can be described by an element ${x}_{t}$ of the *state set* ${{\mathbb X}}_{t}$, - the decision-maker (DM) makes a decision ${u}_{t}$, taken within a *control set* ${{\mathbb U}}_{t}$. A *state of Nature* $\omega$ affects the system, drawn within a *sample set* $\Omega$, also called *Nature*. No probabilistic structure is imposed on the set $\Omega$. #### Environmental illustration. In the case of dengue epidemics control at daily time steps, the state can be a vector of abundances of healthy and deseased individuals (possibly at ages), together with the same description for the mosquito vector; the control can be the daily fumigation effort, mosquito larva removal, quarantine measures, or the opening and closing of sanitary facilities; Nature represents unknown factors that affect the dengue dynamics, like rains, humidity, mosquito biting rates, individual susceptibilities, etc. Some of these factirs (like rain) can be progressively unfolded as times passes. #### Special case where the sample set is a set of scenarios. In many cases, at each time $t \in {{\mathbb T}}$, an uncertainty ${w}_{t}$ affects the system, drawn within an *uncertainty set* ${{\mathbb W}}_{t}$. Hence, a state of Nature has the form $\omega= {\left\{{w}_{t}\right\}_{t \in {{\mathbb T}}}}$ — and is called a *scenario* — drawn within a product [sample set]{} $\Omega= \prod_{t \in {{\mathbb T}}}{{\mathbb W}}_{t}$. #### Environmental illustration. The above definition of scenarios is in phase with the vocable of scenarios in climate change mitigation; it represents sequences of uncertainties that affect the climate evolution. In our framework, a scenario is not in the hands of the decision-maker; for instance, a scenario does not include investment decisions. #### Relevance for resilience. In the case of scenarios, as the [uncertainty sets]{} ${{\mathbb W}}_{t}$ depend on $t$, we cover the case where - an uncertainty ${w}_{t} \in {{\mathbb W}}_{t}$ affects the system at each time $t$, possibly progressively revealed to the DM, hence available when he makes decisions; - other uncertainties, that are present from the start (like parameters), hence are part of the set ${{\mathbb W}}_{{t_0}}$; such uncertainties are not necessarily revealed to the DM as times passes, and remain unknown. Our setting is more general than the one in [@Rouge-Mathias-Deffuant:2013]. First, we do not restrict the [sample set]{} to be made of scenarios as Rougé, Mathias and Deffuant do. Second, even in the case of scenarios, no probabilistic structure is imposed on the set $ \prod_{t \in {{\mathbb T}}}{{\mathbb W}}_{t} $ whereas Rougé, Mathias and Deffuant require that it be equipped with a probability distribution having a density (with respect to an, unspecified, measure, likely the Lebesgue measure on a Euclidian space). ### State and control paths {#State_and_control_paths,_scenarios_of_uncertainties,_states_of_Nature} With the basic set ${{\mathbb T}}$ and the basic families of sets ${\left\{{{\mathbb X}}_{t}\right\}_{t \in {{\mathbb T}}}}$ and ${\left\{{{\mathbb U}}_{t}\right\}_{t \in {{\mathbb T}}}}$, we define - the *set $\prod_{t \in {{\mathbb T}}}{{\mathbb X}}_{t}$ of state paths*, made of sequences ${\left\{{x}_{t}\right\}_{t \in {{\mathbb T}}}}$ where $ {x}_{t} \in {{\mathbb X}}_{t} $ for all $t \in {{\mathbb T}}$; *tail state paths* ${\left\{{x}_{r}\right\}_{r \in {s\;\!\!:\;\!\!t}}}$ (starting at time $s<t$) are elements of $ \prod_{r=s}^{t}{{\mathbb X}}_{r} $; - the *set $\prod_{t \in {{\mathbb T}}}{{\mathbb U}}_{t}$ of control paths*, made of sequences ${\left\{{u}_{t}\right\}_{t \in {{\mathbb T}}}}$ where $ {u}_{t} \in {{\mathbb U}}_{t} $ for all $t \in {{\mathbb T}}$; *tail control paths* ${\left\{{u}_{r}\right\}_{r \in {s\;\!\!:\;\!\!t}}}$ (starting at time $s<t$) are elements of $ \prod_{r=s}^{t}{{\mathbb U}}_{r} $. #### Relevance for resilience. We introduce paths because, as stated in the abstract, our (forthcoming) definition of resilience requires that, after any perturbation, the system returns to an acceptable “regime”, that is, that the state-control path as a whole must return to a set of acceptable paths (and not only the state values must belong to an acceptable subset of the state set). We introduce tail paths because resilience encapsulates the idea that recovery is possible after some time, and that the system remains “normal” after that time. ### Dynamics/flow {#Flow} We now introduce a dynamics under the form of a *flow* $ {\left\{\phi_{{s\;\!\!:\;\!\!t}}\right\}_{{(s,t)} \in {{\mathbb T}}^2}} $, that is, a family of mappings $$\phi_{{s\;\!\!:\;\!\!t}}: {{\mathbb X}}_{s} \times \prod_{r=s}^{t}{{\mathbb U}}_{r} \times \Omega \to \prod_{r=s}^{t}{{\mathbb X}}_{r} {\; .}\label{eq:flow}$$ When $s>t$, all these expressions are void because $ {s\;\!\!:\;\!\!t}= \emptyset $. The flow $\phi_{{s\;\!\!:\;\!\!t}}$ maps an initial state $\bar{x}_{s} \in {{\mathbb X}}_{s}$ at time $s$, a [tail control path]{} $ {\left\{{{u}}_{r}\right\}_{r\in{s\;\!\!:\;\!\!t}}} $ and a state of Nature $\omega$ towards a tail state path $${\left\{{{x}}_{r}\right\}_{r\in {s\;\!\!:\;\!\!t}}} = \phi_{{s\;\!\!:\;\!\!t}} {\big(\bar{x}_{s}, {\left\{{{u}}_{r}\right\}_{r\in {s\;\!\!:\;\!\!t}}}, \omega \big)} {\; ,}\label{eq:flow_path}$$ with the property that $ {{x}}_{s} = \bar{x}_{s} $. #### Relevance for resilience. Our setting is more general than the one in [@Rouge-Mathias-Deffuant:2013]: as illustrated below, we cover differential and stochastic differential systems, in addition to iterated dynamics in discrete time (which is the scope of Rougé, Mathias and Deffuant). Our approach thus allows for a general treatment of resilience. #### Cemetery point to take into account either analytical properties or bounds on the controls. In general, a state path cannot be determined by  for *any* state of Nature or for *any* control path, for analytical reasons (measurability, continuity) or because of bounds on the controls. To circumvent this difficulty, one can use a mathematical trick and add to any state set ${{\mathbb X}}_{t}$ a cemetery point $\partial$. Any time a state cannot be properly defined by the flow by , we attribute the value $\partial$. The vocable “cemetery” expresses the property that, once in the state $\partial$, the future state values, yielded by the flow, will all be $\partial$. Therefore, the stationary state path with value $\partial$ will be the image of those scenarios and control paths for which no state path can be determined by . #### Special case of an iterated dynamics in discrete time. In discrete time, when ${{\mathbb T}}={{\mathbb N}}$, the flow is generally produced by the iterations of a dynamic $${{x}}_{t}={x}{\; , \enspace}{{x}}_{s+1} = {F}_{s}{({{x}}_{s},{{u}}_{s},{{w}}_{s})} {\; , \enspace}t \geq s {\; .}\label{eq:iterated_dynamics_in__discrete_time}$$ How do we include control constraints in this setting? Suppose given a family of nonempty set-valued mappings $ \mathcal{{U}}_{s} : {{\mathbb X}}_{s} \rightrightarrows {{\mathbb U}}_{s} $, $s \in {{\mathbb T}}$. We want to express that only controls ${{u}}_{s}$ that belong to $ \mathcal{{U}}_{s}{({x}_{s})} $ are relevant. For this purpose, we add to all the state sets ${{\mathbb X}}_{s}$ a cemetery point $\partial$. Then, when $ {{u}}_{r} \not\in \mathcal{{U}}_{r}{({x}_{r})} $ in  for at least one $ r \in {s\;\!\!:\;\!\!t} $, we set $ \phi_{{s\;\!\!:\;\!\!t}} {\big(\bar{x}_{s}, {\left\{{{u}}_{r}\right\}_{r\in {s\;\!\!:\;\!\!t}}}, {\left\{{{w}}_{r}\right\}_{r\in {s\;\!\!:\;\!\!t}}},\gamma \big)} = {\left\{\partial\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} $ in . #### Environmental illustration. In natural resource management, many population models (anmal, plants) are given by discrete time abundance-at-age dynamical equations. Outside population models, many stock problems are also based upon discrete time dynamical equations. This is the case of dam management, where water stock balance equations are written at a daily scale (possibly less like every eight hours, or possibly more like months for long term planning); control constraints represent the properties that turbined water must be less than the current water stock and bounded by turbine capacity. #### Special case of differential systems. In continuous time, the mapping $ \phi_{{s\;\!\!:\;\!\!t}}$ in  generally cannot be defined over the whole set $ \prod_{r=s}^{t}{{\mathbb U}}_{r} \times \Omega $. Tail control paths and states of Nature need to be restricted to subsets of $\prod_{r=s}^{t}{{\mathbb U}}_{r}$ and $\Omega$, like the continuous ones for example when dealing with Euclidian spaces. For instance, when ${{\mathbb T}}={{\mathbb R}}_+$ and the flow is produced by a smooth dynamical system on a Euclidian space ${{\mathbb X}}={{\mathbb R}}^n$ $${{x}}_{t}={x}{\; , \enspace}\dot{{{x}}}_{s} ={f}_{s}{({{x}}_{s},{{u}}_{s})} {\; , \enspace}s \geq t {\; ,}$$ control paths $ {\left\{{{u}}_{s}\right\}_{s\in{t\;\!\!:\;\!\!{T}}}} $ are generally restricted to piecewise continuous ones for a solution to exist. #### Special case of stochastic differential equations. Under certain technical assumptions, a stochastic differential equation $$d{{\mathbf{{X}}}}_{s} ={f}_{s} {({\mathbf{{X}}}_{s},{\mathbf{{U}}}_{s},{\mathbf{{W}}}_{s})}ds + g{({\mathbf{{X}}}_{s},{\mathbf{{U}}}_{s},{\mathbf{{W}}}_{s})} d{\mathbf{{W}}}_{s} {\; ,}$$ where $ {\left\{{\mathbf{{W}}}_{s}\right\}_{s\in {{\mathbb R}}_+}} $ is a Brownian motion, gives rise to solutions in the strong sense. In that case, a flow can be defined (but not over the whole set $ \prod_{r=s}^{t}{{\mathbb U}}_{r} \times \Omega $). #### The case of the history flow. Any possible state derives from the so-called *history* $${h}_t={\big({\left\{{{u}}_{r}\right\}_{r\in {{t_0}\;\!\!:\;\!\!t}}}, \omega \big)} {\; .}$$ In that case, the flow  is trivially given by $ {\left\{{{h}}_{r}\right\}_{r\in {s\;\!\!:\;\!\!t}}} = {\big({h}_{s}, {\left\{{{u}}_{r}\right\}_{r\in {s\;\!\!:\;\!\!t}}} \big)} $. We will use the notion of history when we compare our approach with the viability approach to resilience. Adapted and admissible strategies --------------------------------- A control ${{u}}_{t}$ is an element of the [control set]{} ${{\mathbb U}}_{t}$. A *policy* (at time $t$) is a mapping $${{\lambda}}_{t} : {{\mathbb X}}_{t} \times \Omega \to {{\mathbb U}}_{t} \label{eq:policy}$$ with image in the [control set]{} ${{\mathbb U}}_{t}$. A *strategy* is a sequence $ {\left\{{{\lambda}}_{t}\right\}_{t\in {{\mathbb T}}}} $ of policies. #### Environmental illustration. In climate change mitigation, a strategy can be an investment policy in renewable energies as a function of the past observed temperatures. In epidemics control, a strategy can be quarantine measures or vector control as a function of observed infected individuals. ### Admissible strategies Suppose given a family of nonempty set-valued mappings $ \mathcal{{U}}_{t} : {{\mathbb X}}_{t} \times \Omega \rightrightarrows {{\mathbb U}}_{t} $, $t \in {{\mathbb T}}$. An *admissible strategy* is a strategy $ {\left\{{{\lambda}}_{t}\right\}_{t\in {{t_0}\;\!\!:\;\!\!{T}}}} $ such that control constraints are satisfied in the sense that, for all $t \in {{\mathbb T}}$, $${{\lambda}}_{t}{({x}_{t}, \omega )} \in \mathcal{{U}}_{t}{({x}_{t},\omega )} {\; , \enspace}\forall {( {x}_{t}, \omega )} \in {{\mathbb X}}_{t} \times \Omega {\; .}$$ ### Adapted strategies Suppose that the sample set $\Omega$ is equipped with a *filtration* $ {\left\{{\EuScript{F}}_{t}\right\}_{t\in {{\mathbb T}}}} $. Hence each ${\EuScript{F}}_{t}$ is a $\sigma$-field and the sequence $ t \mapsto {\EuScript{F}}_{t} $ is increasing (for the inclusion order). Suppose that each *state set* ${{\mathbb X}}_{t}$, is equipped with a $\sigma$-field ${\EuScript{{X}}}_{t}$. An *adapted policy* is a mapping  which is measurable with respect to the product $\sigma$-field ${\EuScript{{X}}}_{t} \otimes {\EuScript{F}}_{t} $. An *adapted strategy* is a family $ {\left\{{{\lambda}}_{t}\right\}_{t\in {{t_0}\;\!\!:\;\!\!{T}}}} $ of adapted policies. #### Special case where the sample set is a set of scenarios. Consider the case where $\Omega= \prod_{t \in {{\mathbb T}}}{{\mathbb W}}_{t}$ and where each set ${{\mathbb W}}_{t}$ is equipped with a $\sigma$-field ${\EuScript{{W}}}_{t}$ (supposed to contain the singletons). The natural [filtration]{} $ {\left\{{\EuScript{F}}_{t}\right\}_{t\in {{\mathbb T}}}} $ is given by $${\EuScript{F}}_{t} = \bigotimes_{r \leq t} {\EuScript{{W}}}_{r} \otimes \bigotimes_{s > t} \{ \emptyset, {{\mathbb W}}_{s} \} {\; .}$$ Then, in an [adapted strategy]{} $ {\left\{{{\lambda}}_{t}\right\}_{t\in {{t_0}\;\!\!:\;\!\!{T}}}} $, each policy can be identified with a mapping of the form [@Carpentier-Chancelier-Cohen-DeLara:2015] $${{\lambda}}_{t} : {{\mathbb X}}_{t} \times \prod_{r={t_0}}^{t}{{\mathbb W}}_{r} \to {{\mathbb U}}_{t} {\; .}$$ In that case, our definition of adapted strategy means that the DM can, at time $t$, use no more than time $t$, current state value ${x}_{t}$ and past scenario ${\left\{{{w}}_{s}\right\}_{s\in {{t_0}\;\!\!:\;\!\!t} }}$ to make his decision $ {u}_{t} = {{\lambda}}_{t}{( {x}_{t}, {\left\{ {{w}}_{s}\right\}_{s\in {{t_0}\;\!\!:\;\!\!t} }} )} $. #### Relevance for resilience. Though this is not the most general framework to handle information (see [@Carpentier-Chancelier-Cohen-DeLara:2015] for a more general treatment of information), we hope it can enlighten the notion of *adaptive response* often found in the resilience literature. Our setting is more general than the one in [@Rouge-Mathias-Deffuant:2013]: indeed, Rougé, Mathias and Deffuant only consider state feedbacks, that is, Markovian strategies as defined below. By contrast, our setting includes the case of corrupted and partially observed state feedback strategies, that is, the case where strategies have as input a partial observation of the state that is corrupted by noise. #### Special case of Markovian or state feedback strategies. Markovian or state feedback policies are of the form $${{\lambda}}_{t} : {{\mathbb X}}_{t} \to {{\mathbb U}}_{t} {\; .}$$ With this definition, we express that, at time $t$, the DM can only use time $t$ and current state value ${x}_{t}$ — but not the state of Nature $\omega$ — to make his decision $ {u}_{t} = {{\lambda}}_{t}{( {x}_{t} )} $. In some cases (when dynamic programming applies for instance), it is enough to restrict to Markovian strategies, much more economical than general strategies. Closed loop flow ---------------- From now on, when we say “strategy”, we mean “adapted and admissible strategy”. Given an initial state and a state of Nature, a strategy will induce a state path thanks to the flow: this gives the closed loop flow as follows. Let $s \in {{\mathbb T}}$ and $t \in {{\mathbb T}}$, with $s < t$. Let $ {\left\{{{\lambda}}_{t}\right\}_{t\in {{\mathbb T}}}} $ be a strategy. We suppose that, for any initial state $\bar{x}_{s} \in {{\mathbb X}}_{s}$ and any state of Nature $\omega$, the following system of (closed loop) equations $$\begin{aligned} {\left\{{{x}}_{r}\right\}_{r\in {s\;\!\!:\;\!\!t}}} &= \phi_{{s\;\!\!:\;\!\!t}} {\big(\bar{x}_{s}, {\left\{{{u}}_{r}\right\}_{r\in {s\;\!\!:\;\!\!t}}}, \omega \big)} \\ {{u}}_{r} &= {{\lambda}}_{r}{( {{x}}_{r} , \omega )} {\; , \enspace}\forall r \in {s\;\!\!:\;\!\!t} \end{aligned}$$ has a *unique* solution $ {\big( {\left\{{{x}}_{r}\right\}_{r \in {s\;\!\!:\;\!\!t}}} , {\left\{{{u}}_{r}\right\}_{r \in {s\;\!\!:\;\!\!t}}} \big)} $. Quite naturally, we define the *closed loop flow* $ \phi_{{s\;\!\!:\;\!\!t}}^{{{\lambda}}} $ by $$\phi_{{s\;\!\!:\;\!\!t}}^{{{\lambda}}} {\big(\bar{x}_{s},\omega \big)} = {\big( {\left\{{{x}}_{s}\right\}_{s\in {s\;\!\!:\;\!\!t}}} , {\left\{{{u}}_{s}\right\}_{s\in {s\;\!\!:\;\!\!t}}} \big)} {\; .}\label{eq:closed_loop_flow}$$ Resilience: a mathematical framework {#Resilience:_A_Mathematical_Framework} ==================================== Equipped with the material in Sect. \[Ingredients\_for\_an\_abstract\_control\_system\_with\_uncertainties\], we now frame the concept of resilience in mathematical garbs. For this purpose, we introduce the notion of *recovery regime*. Compared to other definitions of resilience [@Holling:73; @Martin:2004; @Martin-Deffuant-Calabrese:2011; @Rouge-Mathias-Deffuant:2013], our definition requires that, after any perturbation, the state-control path as a whole can be driven, by a proper strategy, to a set of acceptable paths (and not only the state values must belong to an acceptable subset of the state set, asymptotically or not). In addition, as state and control paths are contingent on uncertainties, we require that they lay within an acceptable domain of random processes, called recovery regimes. Once again, as the reference [@Rouge-Mathias-Deffuant:2013] is the more mathematically driven paper on resilience, we will systematically emphasize in what our approach differs from that of Rougé, Mathias and Deffuant. Robustness, resilience and random processes ------------------------------------------- The notion of robustness captures a form of stability to perturbations; it is a static notion, as no explicit reference to time is required. By contrast, the concept of resilience makes reference to time (dynamics), strategies and perturbations. This is why, to speak of resilience — a notion that mixes time and randomness — we find it convenient to use the framework of random processes, although this does not mean that we require any probability. From now on, we consider that the [sample space]{} $\Omega$ is a measurable set equipped with a $\sigma$-field ${\EuScript{F}}$ (but not necessarily equipped with a probability). When we consider a deterministic setting, $\Omega$ is reduced to a singleton (and ignored). The set of *measurable mappings* from $\Omega$ to any measurable set ${{\mathbb Y}}$ will be denoted by $ {{\mathbb L}^0{\big(\Omega,{{\mathbb Y}}\big)}} $. Elements of $ {{\mathbb L}^0{\big(\Omega,{{\mathbb Y}}\big)}} $ are called *random variables* or *random processes*, although this does not imply the existence of an underlying probability. Random variables are designated with bold capital letters like ${\mathbf{Z}}$. From now on, every [state set]{} ${{\mathbb X}}_{t}$ is a measurable set equipped with a $\sigma$-field ${\EuScript{{X}}}_{t}$, every [control set]{} ${{\mathbb U}}_{t}$ with a $\sigma$-field ${\EuScript{{U}}}_{t}$, and, when needed, every [uncertainty set]{} ${{\mathbb W}}_{t}$ with a $\sigma$-field ${\EuScript{{W}}}_{t}$. Fields are introduced when probabilities are needed. When they are not, as with the robust setting, it suffices to equip all sets with their complete $\sigma$-fields, made of all subsets. Then, [measurable mappings]{} $ {{\mathbb L}^0{\big(\Omega,{{\mathbb Y}}\big)}} $ from $\Omega$ to any set ${{\mathbb Y}}$ are all mappings. Recovery regimes ---------------- *Recovery regimes*, starting from $t \in {{\mathbb T}}$, are subsets of random processes of the form $${{\mathcal A}}^{t} \subset {{\mathbb L}^0{\big(\Omega,\prod_{s=t}^{{T}}{{\mathbb X}}_{s} \times \prod_{s=t}^{{T}}{{\mathbb U}}_{s} \big)}} {\; .}$$ When there are no uncertainties, $\Omega$ is reduced to a singleton, so that $ {{\mathcal A}}^{t} \subset \prod_{s=t}^{{T}}{{\mathbb X}}_{s} \times \prod_{s=t}^{{T}}{{\mathbb U}}_{s} $, as in the two first following examples. #### Example of recovery regimes converging to an equilibrium. Let ${{\mathbb T}}={{\mathbb R}}_+$, ${{\mathbb X}}_{t}={{\mathbb R}}^n$ and ${{\mathbb U}}_{t}={{\mathbb R}}^m$, for all $t \in {{\mathbb T}}$. Let $ \bar{x}$ be an equilibrium point of the dynamical system $ \dot{{x}}_{s} ={f}{({x}_{s},\bar{u})} $ when the control is stationary equal to $\bar{u}$, that is, $ 0 ={f}{(\bar{x},\bar{u})} $. The recovery regimes, starting from $t \in {{\mathbb T}}$, converging to the equilibrium $\bar{x}$ form the set $${{\mathcal A}}^{t} = {\left\{ {\big( {\left\{{x}_{s}\right\}_{s \geq t}}, {\left\{{u}_{s}\right\}_{s \geq t}} \big)} \in {{\mathbb X}}^{[t,+\infty[} \times {{\mathbb U}}^{[t,+\infty[} \:\left|\:{x}_{s} \to_{s\to +\infty} \bar{x}\right.\right\}} {\; .}$$ In general, the equilibrium $\bar{x}$ is supposed to be asymptotically stable, locally or globally. A more general definition would be $${{\mathcal A}}^{t} = {\left\{ {\big( {\left\{{x}_{s}\right\}_{s \geq t}}, {\left\{{u}_{s}\right\}_{s \geq t}} \big)} \:\left|\:\lim_{s\to +\infty}{x}_{s} \textrm{ exists} \right.\right\}} {\; ,}$$ and, to account for constraints on the values taken by the controls, we can consider $${{\mathcal A}}^{t} = {\left\{ {\big( {\left\{{x}_{s}\right\}_{s \geq t}}, {\left\{{u}_{s}\right\}_{s \geq t}} \big)} \:\left|\:\lim_{s\to +\infty}{x}_{s} \textrm{ exists and } {u}_{s} \in \mathcal{{U}}_{s}{({x}_{s})} {\; , \enspace}\forall s \geq t \right.\right\}} {\; ,}$$ where $ \mathcal{{U}}_{s} : {{\mathbb X}}_{s} \rightrightarrows {{\mathbb U}}_{s} $, for all $s \in {{\mathbb T}}$. #### Example of bounded recovery regimes. Let ${{\mathbb T}}={{\mathbb R}}_+$, ${{\mathbb X}}_{t}={{\mathbb R}}^n$ and ${{\mathbb U}}_{t}={{\mathbb R}}^m$, for all $t \in {{\mathbb T}}$. If $B$ is a bounded region of ${{\mathbb X}}={{\mathbb R}}^n$, we consider $${{\mathcal A}}^{t} = {\left\{ {\big( {\left\{{x}_{s}\right\}_{s \geq t}}, {\left\{{u}_{s}\right\}_{s \geq t}} \big)} \:\left|\:{x}_{s} \in B {\; , \enspace}\forall s \geq t \right.\right\}} {\; .}$$ When $B$ is a ball of small radius $\rho>0$ around the equilibrium $\bar{x}$, we obtain state paths that remain close to $\bar{x}$. #### Example of random recovery regimes. Suppose that the measurable [sample space]{} $(\Omega,{\EuScript{F}})$ is equipped with a probability ${{\mathbb P}}$. Let ${{\mathbb X}}_{t}={{\mathbb R}}^n$ and ${{\mathbb U}}_{t}={{\mathbb R}}^m$, for all $t \in {{\mathbb T}}$. Letting $B$ be a bounded region of ${{\mathbb X}}={{\mathbb R}}^n$ and $\beta \in ]0,1[ $, the set $$\begin{aligned} {{\mathcal A}}^{t} = & \{ {\big( {\left\{{\mathbf{{X}}}_{s}\right\}_{s \geq t }}, {\left\{{\mathbf{{U}}}_{s}\right\}_{s \geq t}} \big)} \in {{\mathbb L}^0{\big(\Omega,\prod_{s=t}^{{T}}{{\mathbb X}}_{s} \times \prod_{s=t}^{{T}}{{\mathbb U}}_{s} \big)}} \mid \nonumber \\ & {{\mathbb P}}{\big[ \exists s \geq t \mid {\mathbf{{X}}}_{s} \not \in B \big]} \leq \beta \} \end{aligned}$$ represents state paths that get at least once outside the bounded region $B$ with a probability less than $\beta$. If ${{\mathbb T}}$ is discrete, the set $$\begin{aligned} {{\mathcal A}}^{t} = & \{ {\big( {\left\{{\mathbf{{X}}}_{s}\right\}_{s \geq t }}, {\left\{{\mathbf{{U}}}_{s}\right\}_{s \geq t}} \big)} \in {{\mathbb L}^0{\big(\Omega,\prod_{s=t}^{{T}}{{\mathbb X}}_{s} \times \prod_{s=t}^{{T}}{{\mathbb U}}_{s} \big)}} \mid \nonumber \\ & {{\mathbb P}}{\big[ \exists s_1 \geq t {\; , \enspace}s_2 \geq t {\; , \enspace}s_3 \geq t \mid {\mathbf{{X}}}_{s_1} \not \in B {\; , \enspace}{\mathbf{{X}}}_{s_2} \not \in B {\; , \enspace}{\mathbf{{X}}}_{s_3} \not \in B \big]} = 0 \} \end{aligned}$$ represents state paths that get no more than two times outside the bounded region $B$. Resilient strategies and resilient states {#Resilient_strategies_and_resilient_states} ----------------------------------------- Consider a starting time $t\in{{\mathbb T}}$ and an initial state $\bar{x}_{t}\in{{\mathbb X}}_{t}$. We say that the strategy $ {{\lambda}}$ is a *resilient strategy* starting from time $t$ in state $\bar{x}_{t}$ if the random process $ {\Big( {\left\{{\mathbf{{X}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} , {\left\{{\mathbf{{U}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} \Big)} $ given by $$\begin{aligned} {\left\{{\mathbf{{X}}}_{s}{(\omega)}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} &= \phi_{{t\;\!\!:\;\!\!{T}}}^{{{\lambda}}}{\big({x}_{t},\omega \big)} \\ {\mathbf{{U}}}_{s}{(\omega)} &= {{\lambda}}_{s}{( {\mathbf{{X}}}_{s}{(\omega)} , \omega )} {\; , \enspace}\forall s\in {t\;\!\!:\;\!\!{T}} {\; ,}\end{aligned}$$ \[eq:output\] where the closed loop flow $\phi_{{t\;\!\!:\;\!\!{T}}}^{{{\lambda}}}$ is given in , is such that $${\Big( {\left\{{\mathbf{{X}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} , {\left\{{\mathbf{{U}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} \Big)} \in {{\mathcal A}}^{t} {\; .}$$ Notice that we do not use the part $ {\left\{{{\lambda}}_{r}\right\}_{r < t}} $ of the strategy $ {{\lambda}}= {\left\{{{\lambda}}_{r}\right\}_{r\in {{t_0}\;\!\!:\;\!\!{T}} }} $. With this definition, a resilient strategy is able to drive the state-control random process into an acceptable regime. As a resilient strategy is adapted, it can “adapt” to the past values of the randomness but no to its future values (hence, our notion of resilience does not require clairvoyance of the DM). Our definition of resilience is a form of controlability for whole random processes (regimes): a resilient strategy has the property to shape the closed loop flow $\phi_{{t\;\!\!:\;\!\!{T}}}^{{{\lambda}}}$ so that it belongs to a given subset of random processes. We denote by $ {{{{\Lambda}}}^{R}}_{t}{(\bar{x}_{t})} $ the set of *resilient strategies* at time $t$, starting from state $\bar{x}_{t}$. The set of *resilient states* at time $t$ is $${{{{\mathbb X}}}^{R}}_{t} = {\left\{ \bar{x}_{t} \in {{\mathbb X}}_{t} \:\left|\:{{{{\Lambda}}}^{R}}_{t}{(\bar{x}_{t})} \not = \emptyset \right.\right\}} {\; .}$$ Illustrations {#Illustrations} ============= In Sect. \[Resilience:\_A\_Mathematical\_Framework\], we have provided some illustrations in the course of the exposition. Now, we make the connection between the previous setting and two other settings, the resilience “à la Holling” [@Holling:73] in §\[Holling\] and the resilience-viability framework [@Martin:2004; @Martin-Deffuant-Calabrese:2011; @Rouge-Mathias-Deffuant:2013] in §\[resilience-viability\]. Deterministic control dynamical system with attractor {#Holling} ----------------------------------------------------- As the paper [@Holling:73] does not contain a single equation, it is bit risky to force the seminal Holling’s contribution into our setting. However, it is likely that it corresponds to ${{\mathbb T}}={{\mathbb R}}_+$ and to recovery regimes of the form $${{\mathcal A}}^{t} = {\left\{ {\big( {\left\{{x}_{s}\right\}_{s \geq t}}, {\left\{{u}_{s}\right\}_{s \geq t}} \big)} \:\left|\:{x}_{s} \textrm{ converges towards an attractor} \right.\right\}} {\; .}$$ Note that, as often in the ecological literature on resilience [@Arnoldi-Loreau-Haegeman:2016], the underlying dynamical system is not controlled. Resilience and viability {#resilience-viability} ------------------------ Some authors [@Martin:2004; @Martin-Deffuant-Calabrese:2011; @Rouge-Mathias-Deffuant:2013] propose to frame resilience within the mathematical theory of viability [@Aubin:1991]. Let ${{\mathbb X}}_{t}={{\mathbb X}}$ and ${{\mathbb U}}_{t}={{\mathbb U}}$, for all $t \in {{\mathbb T}}$. Let ${{\mathbb A}}\subset {{\mathbb X}}$ denote a set made of “acceptable states”. Let $ \mathcal{{U}}_{s} : {{\mathbb X}}\rightrightarrows {{\mathbb U}}$, $s \in {{\mathbb T}}$ be a family of set-valued mappings that represent control constraints. ### Deterministic viability Consider a starting time $t\in{{\mathbb T}}$ and the recovery regimes $$\begin{aligned} {{\mathcal A}}^{t} = \{ {\big( {\left\{{x}_{s}\right\}_{s \geq t}}, {\left\{{u}_{s}\right\}_{s \geq t}} \big)} \mid {x}_{s} \in {{\mathbb A}}{\; , \enspace}{u}_{s} \in \mathcal{{U}}_{s}{({x}_{s})} {\; , \enspace}\forall s \geq t \} {\; .}\end{aligned}$$ Then, a resilient strategy is one that is able to drive the state towards the set ${{\mathbb A}}$ of acceptable states. ### Robust viability {#Deterministic_and_robust_viability} When there are no uncertainties, we just established a connection between recovery regimes and viability. But, with uncertainties, as resilience requires a form of stability “whatever the perturbations”, we are in the realm of *robust viability* [@DeLara-Doyen:2008], as follows. Let $ \overline\Omega \subset \Omega $, corresponding to the (nonempty) subset of states of Nature with respect to which the DM expects the system to be robust. Consider a starting time $t\in{{\mathbb T}}$ and the recovery regimes $$\begin{aligned} {{\mathcal A}}^{t} = \{ & {\big( {\left\{{\mathbf{{X}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} {\left\{{\mathbf{{U}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} \big)} \in {{\mathbb L}^0{\big(\Omega,\prod_{s=t}^{{T}}{{\mathbb X}}_{s} \times \prod_{s=t}^{{T}}{{\mathbb U}}_{s} \big)}} \mid \nonumber \\ & \exists {\mathbf{\tau}} \in {{\mathbb L}^0{\big(\Omega,{{\mathbb T}}\big)}} {\; , \enspace}\forall \omega \in \overline\Omega {\; , \enspace}\nonumber \\ & {\mathbf{\tau}}{(\omega)} \geq t {\; , \enspace}{\mathbf{{X}}}_{s}{(\omega)} \in {{\mathbb A}}{\; , \enspace}{\mathbf{{U}}}_{s}{(\omega)} \in \mathcal{{U}}_{s}{({\mathbf{{X}}}_{s}{(\omega)})} {\; , \enspace}\forall s \geq {\mathbf{\tau}}{(\omega)} \} {\; .}\label{eq:recovery_regimes_robust_viability}\end{aligned}$$ Then, a resilient strategy is one that is able to drive the state towards the set ${{\mathbb A}}$ of acceptable states, after a random time ${\mathbf{\tau}}$, whatever the perturbations in $\overline\Omega$. ### Robust viability and recovery time attached to a resilient strategy {#recovery_time} Let $ \overline\Omega \subset \Omega $, whose elements can be interpreted as shocks. Consider a starting time $t\in{{\mathbb T}}$ and an initial state $\bar{x}_{t}\in{{\mathbb X}}_{t}$. If $ {{\lambda}}={\left\{{{\lambda}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} $ is a resilient strategy for the recovery regimes , the *recovery time* is the random time defined by $$\begin{aligned} {\mathbf{\tau}}{(\omega)} =\inf\{ r \in {t\;\!\!:\;\!\!{T}} \mid & {\left\{{\mathbf{{X}}}_{s}{(\omega)}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} = \phi_{{t\;\!\!:\;\!\!{T}}}^{{{\lambda}}} {\big(\bar{x}_{t}, \omega\big)} \nonumber \\ & {\mathbf{{U}}}_{s}{(\omega)} = {{\lambda}}_{s}{( {\mathbf{{X}}}_{s}{(\omega)} , \omega )} {\; , \enspace}\forall s\in {t\;\!\!:\;\!\!{T}} \nonumber \\ & {\mathbf{{X}}}_{s}{(\omega)} \in {{\mathbb A}}{\; , \enspace}{\mathbf{{U}}}_{s}{(\omega)} \in \mathcal{{U}}_{s}{( {\mathbf{{X}}}_{s}{(\omega)} )} {\; , \enspace}\forall s \geq r \} {\; ,}\end{aligned}$$ for all $ \omega \in \Omega $, with the convention that $\inf \emptyset = +\infty$. Thus, the resilient strategy drives the state towards the set ${{\mathbb A}}$ of acceptable states, after the random time ${\mathbf{\tau}}$, whatever the perturbations (shocks) in $\overline\Omega$. By contrast, the so-called *time of crisis* occurs before ${\mathbf{\tau}}$ [@Doyen-Saint-Pierre:1997]. ### Stochastic viability {#Stochastic_viability} Suppose that the measurable [sample space]{} $(\Omega,{\EuScript{F}})$ is equipped with a probability ${{\mathbb P}}$ and let $\beta \in [0,1] $, represent a probability level. Consider a starting time $t\in{{\mathbb T}}$ and the recovery regimes $$\begin{aligned} {{\mathcal A}}^{t} = \{ & {\big( {\left\{{\mathbf{{X}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}}, {\left\{{\mathbf{{U}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} \big)} \in {{\mathbb L}^0{\big(\Omega,\prod_{s=t}^{{T}}{{\mathbb X}}_{s} \times \prod_{s=t}^{{T}}{{\mathbb U}}_{s} \big)}} \mid \nonumber \\ & {{\mathbb P}}{\big[ {\mathbf{{X}}}_{s} \in {{\mathbb A}}{\; , \enspace}{\mathbf{{U}}}_{s} \in \mathcal{{U}}_{s}{( {\mathbf{{X}}}_{s} )} {\; , \enspace}\forall s \geq t \big]} \geq \beta \} {\; .}\end{aligned}$$ With these recovery regimes, we express that the probability to satisfy state and control constraints after time $t$ is at least $\beta$ [@Doyen-DeLara:2010]. ### Discussion and comparison with the viability theory approach for resilience. Our setting is more general than the viability theory approach for resilience as introduced in [@Martin:2004; @Martin-Deffuant-Calabrese:2011; @Rouge-Mathias-Deffuant:2013]. Indeed, the viability approach to resilience deals with constraints time by time; our approach does not. To illustrate our point, consider the deterministic case with discrete and finite time, and scalar controls, to make things easy. It is clear that the recovery regimes given by $$\begin{aligned} {{\mathcal A}}^{t} &= \{ {\big( {\left\{{x}_{s}\right\}_{s \geq t}}, {\left\{{u}_{s}\right\}_{s \geq t}} \big)} \mid \min_{s \geq t} {u}_{s} \leq 0 \} \\ &= \{ {\big( {\left\{{x}_{s}\right\}_{s \geq t}}, {\left\{{u}_{s}\right\}_{s \geq t}} \big)} \mid \exists s \geq t {\; ,}{u}_{s} \leq 0 \} {\; .}\end{aligned}$$ cannot be expressed as time by time constraints on the controls. Of course, the viability approach *could* handle such a case, but at the price of extending the state and the dynamics, to turn an intertemporal constraint into a time by time constraint. For instance, with the history state introduced at the end of §\[Flow\], we can always express any recovery regimes set as viability constraints. In the example above, we do not need the whole history to turn the set ${{\mathcal A}}^{t}$ into one described by time by time constraints: it suffices to introduce an additional component to the state like $ \sum_{s \geq t} {{\mathbf 1}}_{ \{ {u}_{s} \leq 0 \} } $ in discrete time and impose the final constraint that this new part of an extended state be non zero. To sum up, our approach to resilience covers more recovery regimes, described with the original states and controls, than those captured by the time by time constraints that make the specificity of the viability approach to resilience. Resilience and risk {#Resilience_and_risk_control/minimization} =================== We now sketch how concepts from risk measures (introduced initially in mathematical finance [@Follmer-Schied:2002]) can be imported to tackle resilience issues. This again is a novelty with respect to the stochastic viability theory approach for resilience as in [@Rouge-Mathias-Deffuant:2013]. Risk measures are potential candidates as *indicators of resilience*. Recovery regimes given by risk measures --------------------------------------- We start by a definition of recovery regimes given by risk measures, then we provide examples. ### Definition of recovery regimes given by extended risk measures Suppose that $ {{\mathbb T}}\subset {{\mathbb R}}$ is equipped with the trace ${\EuScript{T}}$ of the Borel field of ${{\mathbb R}}$. Then, $ {{\mathbb T}}\times \Omega $ is a measurable space when equipped with the product $\sigma$-field ${\EuScript{T}} \otimes {\EuScript{F}}$. Then, any random process in $ {{\mathbb L}^0{\big(\Omega,\prod_{s=t}^{{T}}{{\mathbb X}}_{s} \times \prod_{s=t}^{{T}}{{\mathbb U}}_{s} \big)}} $ can be identified with a random variable in $ {{\mathbb L}^0{\big({t\;\!\!:\;\!\!{T}} \times \Omega,\bigcup_{s=t}^{{T}}{{\mathbb X}}_{s} \bigcup \bigcup_{s=t}^{{T}}{{\mathbb U}}_{s} \big)}} $. We call *extended risk measure* any ${{\mathbb G}}_t$ that maps random variables in $ {{\mathbb L}^0{\big({t\;\!\!:\;\!\!{T}} \times \Omega,\bigcup_{s=t}^{{T}}{{\mathbb X}}_{s} \bigcup \bigcup_{s=t}^{{T}}{{\mathbb U}}_{s} \big)}} $ towards the real numbers [@Follmer-Schied:2002]. The lower the risk measure ${{\mathbb G}}_t$, the better. The basic example of a risk measure is the mathematical expectation under a given probability distribution. A celebrated risk measure in mathematical finance is the *tail/average/conditional value-at-risk*. With ${{\mathbb G}}_t$ an extended risk measure and $\alpha \in {{\mathbb R}}$ a given *risk level*, we define recovery regimes by $$\begin{aligned} {{\mathcal A}}^{t} = \{ & {\big( {\left\{{\mathbf{{X}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}}, {\left\{{\mathbf{{U}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} \big)} \in {{\mathbb L}^0{\big(\Omega,\prod_{s=t}^{{T}}{{\mathbb X}}_{s} \times \prod_{s=t}^{{T}}{{\mathbb U}}_{s} \big)}} \mid \nonumber \\ & {{\mathbb G}}_{t} {\big[ {\left\{{\mathbf{{X}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}}, {\left\{{\mathbf{{U}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} \big]} \leq \alpha \} {\; .}\label{eq:alpha}\end{aligned}$$ The quantity $ {{\mathbb G}}_{t} {\big[ {\left\{{\mathbf{{X}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}}, {\left\{{\mathbf{{U}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} \big]} $ measures the “risk” borne by the random process $ {\big( {\left\{{\mathbf{{X}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}}, {\left\{{\mathbf{{U}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} \big)} $. Therfore, recovery regimes like in  represent a form of “risk containment” under the level $\alpha$. ### Robust viability and the worst case risk measure The robust viability inspired definition of resilience in §\[Deterministic\_and\_robust\_viability\] corresponds to  with $ \alpha < 1 $ and the *worst case risk measure* $${{\mathbb G}}_{s}{\big( {\left\{{\mathbf{{X}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}}, {\left\{{\mathbf{{U}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} \big)} = \sup_{s\in {t\;\!\!:\;\!\!{T}}} \sup_{\omega \in \overline\Omega} {{\mathbf 1}}_{{{\mathbb A}}^c}{\big({\mathbf{{X}}}_{s}{(\omega)}\big)} {\; ,}$$ where $ \overline\Omega \subset \Omega $. Indeed, $ {{\mathbb G}}_{t} {\big[ {\left\{{\mathbf{{X}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}}, {\left\{{\mathbf{{U}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} \big]} \leq \alpha < 1 $ means that $ {{\mathbf 1}}_{{{\mathbb A}}^c}{\big({\mathbf{{X}}}_{s}{(\omega)}\big)} \equiv 0 $, that is, the state $ {\big({\mathbf{{X}}}_{s}{(\omega)}\big)} $ always belongs to ${{\mathbb A}}$ (as ${{\mathbb A}}^c$ is the complementary set of ${{\mathbb A}}$ in ${{\mathbb X}}$) for all $\omega \in \overline\Omega$. Here, the [worst case risk measure]{} captures that the state $ {\mathbf{{X}}}_{s}{(\omega)} $ belongs to ${{\mathbb A}}$ both for all times — the core of viability, here handled by the term $\sup_{s \geq t} $ — and for all states of Nature in $\overline\Omega$ — the core of robustness, here handled by the term $\sup_{\omega \in \overline\Omega}$. ### Stochastic viability and beyond: ambiguity The stochastic viability inspired definition of resilience in §\[Stochastic\_viability\] corresponds to  with $ \alpha=1-\beta $ and the risk measure $${{\mathbb G}}_{s}{\big( {\left\{{\mathbf{{X}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}}, {\left\{{\mathbf{{U}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} \big)} = {{\mathbb P}}{\big[ \exists s \geq t \mid {\mathbf{{X}}}_{s} \not\in {{\mathbb A}}{\,\mbox{ or }\,} {\mathbf{{U}}}_{s} \not\in \mathcal{{U}}_{s}{( {\mathbf{{X}}}_{s} )} \big]} {\; .}$$ Now, suppose that different risk-holders do not share the same beliefs and let $\mathcal{P}$ denote a set of probabilities on ${(\Omega, {\EuScript{F}})}$. We can arrive at an *ambiguity viability* inspired definition of resilience using the risk measure $${{\mathbb G}}_{s}{\big( {\left\{{\mathbf{{X}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}}, {\left\{{\mathbf{{U}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} \big)} = \sup_{{{\mathbb P}}\in \mathcal{P}} {{\mathbb P}}{\big[ \exists s \geq t \mid {\mathbf{{X}}}_{s} \not\in {{\mathbb A}}{\,\mbox{ or }\,} {\mathbf{{U}}}_{s} \not\in \mathcal{{U}}_{s}{( {\mathbf{{X}}}_{s} )} \big]} {\; .}$$ Here, $ {{\mathbb G}}_{t} {\big[ {\left\{{\mathbf{{X}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}}, {\left\{{\mathbf{{U}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} \big]} \leq \alpha=1-\beta $ means that $ {{\mathbb P}}{\big[ {\mathbf{{X}}}_{s} \in {{\mathbb A}}{\; , \enspace}{\mathbf{{U}}}_{s} \in \mathcal{{U}}_{s}{( {\mathbf{{X}}}_{s} )} {\; , \enspace}\forall s \geq t \big]} \geq \beta $, for all ${{\mathbb P}}\in\mathcal{P}$. ### Random exit time and viability Let $\mu$ be a measure on the time set ${{\mathbb T}}$ like, for instance, the counting measure when ${{\mathbb T}}={{\mathbb N}}$ or the Lebesgue measure when ${{\mathbb T}}={{\mathbb R}}_+$. Then, the random quantity $$\mu \{ s \geq t {\; , \enspace}{\mathbf{{X}}}_{s} \not\in {{\mathbb A}}{\,\mbox{ or }\,} {\mathbf{{U}}}_{s} \not\in \mathcal{{U}}_{s}{( {\mathbf{{X}}}_{s} )} \}$$ measures the number of times that the state-control path $ {\big( {\left\{{\mathbf{{X}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}}, {\left\{{\mathbf{{U}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} \big)} $ *exits* from the viability constraints. Using risk measures — like the *tail/average/conditional value-at-risk* [@Follmer-Schied:2002] — or stochastic orders [@Muller-Stoyan:2002; @Shaked-Shanthikumar:2007], we have differents ways to express that this random quantity remains “small”. ### The general umbrella of cost functions {#cost_functions} All the examples above, and many more [@Rouge-Mathias-Deffuant:2015], fall under the general umbrella of cost functions as follows. Consider a starting time $t\in{{\mathbb T}}$ and a measurable function $$\Psi_{t} : \prod_{s=t}^{{T}}{{\mathbb X}}_{s} \times \prod_{s=t}^{{T}}{{\mathbb U}}_{s} \times \Omega \to {{\mathbb R}}{\; , \enspace}\label{eq:utility}$$ that attachs a *disutility* or *cost* — the opposite of *value*, *utility*, *payoff* — to any tail state and control path, starting from time $t$, and to any state of Nature. Let ${{\mathbb F}}$ be a risk measure that maps random variables on $\Omega$ towards the real numbers. Then, an extended risk measure is given by $${{\mathbb G}}_{t} {\big[ {\left\{{\mathbf{{X}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}}, {\left\{{\mathbf{{U}}}_{s}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} \big]} = {{\mathbb F}}{\big[ \Psi_{t} {\big( {\left\{{\mathbf{{X}}}_{s}{(\cdot)}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}} , {\left\{{\mathbf{{U}}}_{s}{(\cdot)}\right\}_{s\in {t\;\!\!:\;\!\!{T}}}}, \cdot \big)} \big]} {\; .}$$ Resilience and risk minimization -------------------------------- When the set $ {{{{\Lambda}}}^{R}}_{t}{(\bar{x}_{t})} $ of [resilient strategies]{} at time $t$ in §\[Resilient\_strategies\_and\_resilient\_states\] is not empty, how can we select one among the many? Here is a possible way that makes use of risk measures for risk minimization purposes. ### Indicators of resilience Let ${{\mathbb G}}_{t}$ be an extended risk measure. We can look for resilient strategies that minimize risk, solution of $$\min_{ {{\lambda}}\in {{{{\Lambda}}}^{R}}_{t}{(\bar{x}_{t})} } {{\mathbb G}}_{t} {\big[ {\big( \phi_{{t\;\!\!:\;\!\!{T}}}^{{{\lambda}}} {(\bar{x}_{t}, \cdot)} , \cdot \big)} \big]} {\; .}$$ The minimum of the risk measure is a potential candidate as an *indicator of resilience*. Indeed, it is the best achievable measure of residual risk under a resilient strategy. ### Examples Using cost functions as in §\[cost\_functions\], we can look for resilient strategies that minimize *expected costs* $$\min_{ {{\lambda}}\in {{{{\Lambda}}}^{R}}_{t}{(\bar{x}_{t})} } {{\mathbb E}}{\big[ \Psi_{t} {\big( \phi_{{t\;\!\!:\;\!\!{T}}}^{{{\lambda}}} {(\bar{x}_{t}, \cdot)} , \cdot \big)} \big]} {\; , \enspace}$$ or that minimize *worst case costs*, where $ \overline\Omega \subset \Omega $, $$\min_{ {{\lambda}}\in {{{{\Lambda}}}^{R}}_{t}{(\bar{x}_{t})} } \sup_{ \omega \in \overline\Omega } \Psi_{t} {\big( \phi_{{t\;\!\!:\;\!\!{T}}}^{{{\lambda}}} {(\bar{x}_{t}, \omega )} , \omega \big)} {\; , \enspace}$$ or, more generally, that minimize $$\min_{ {{\lambda}}\in {{{{\Lambda}}}^{R}}_{t}{(\bar{x}_{t})} } {{\mathbb F}}{\big[ \Psi_{t} {\big( \phi_{{t\;\!\!:\;\!\!{T}}}^{{{\lambda}}} {(\bar{x}_{t}, \cdot)} , \cdot \big)} \big]} {\; , \enspace}$$ where ${{\mathbb F}}$ is a risk measure that maps random variables on $\Omega$ towards the real numbers [@Follmer-Schied:2002]. For instance, in the robust viability setting of §\[recovery\_time\], an [indicator of resilience]{} could be the minimum (over all resilient strategies) of the maximal (over all states of Nature in $\overline\Omega$) recovery time. Conclusion {#sec:conclusions} ========== Resilience is a rehashed concept in natural hazard management. Most of the formalizations of the concept require that, after any perturbation, the state of a system returns to an acceptable subset of the state set. Equipped with tools from control theory under uncertainty, we have proposed that resilience is the ability for the state-control random process as a whole to be driven to an acceptable “recovery regime” by a proper resilient strategy (adaptive). Our definition of resilience is a form of controlability: a resilient strategy has the property to shape the closed loop flow so that the resulting state and control random process belongs to a given subset of random processes, the acceptable recovery regimes. We have proposed to handle risk thanks to risk measures[^2], by defining recovery regimes that represent a form of “risk containment”. In addition, risk measures are potential candidates as [indicators of resilience]{} as they measure the residual risk under a resilient strategy. Our contribution is formal, with its pros and cons: by its generality, our approach covers a large scope of notions of resilience; however, such generality makes it difficult to propose resolution methods. For instance, the possibility to use dynamic programing in stochastic viability relies upon a white noise assumption that we have not supposed. Much would remain to be done regarding applications and numerical implementation. #### Acknowledgement. The author is indebted to the editor-in-chief, the advisory editor and two reviewers. They supplied detailed critique, comments and inputs which, ultimately, contributed to an improved version of the manuscript. \#1[\#1]{} [10]{} J.-F. Arnoldi, M. Loreau, and B. Haegeman. Resilience, reactivity and variability: A mathematical comparison of ecological stability measures. , 389:47 – 59, 2016. J.-P. Aubin. . Birkh[ä]{}user, Boston, 1991. P. Carpentier, J.-P. Chancelier, G. Cohen, and M. [De Lara]{}. . Springer-Verlag, Berlin, 2015. M. [De Lara]{} and L. Doyen. . Springer-Verlag, Berlin, 2008. L. Doyen and M. [De Lara]{}. Stochastic viability and dynamic programming. , 59(10):629–634, October 2010. L. Doyen and P. Saint-Pierre. Scale of viability and minimum time of crisis. , 5:227–246, 1997. H. F[ö]{}llmer and A. Schied. . Walter de Gruyter, Berlin, 2002. C. S. Holling. Resilience and stability of ecological systems. , 4:1–23, 1973. S. Martin. The cost of restoration as a way of defining resilience: a viability approach applied to a model of lake eutrophication. , 9(2), 2004. S. Martin, G. Deffuant, and J. Calabrese. , pages 15–36. Understanding complex systems. Springer, 2011. A. Muller and D. Stoyan. . John Wiley and Sons, New York, 2002. C. Rougé, J.-D. Mathias, and G. Deffuant. Extending the viability theory framework of resilience to uncertain dynamics, and application to lake eutrophication. , 29(Supplement C):420 – 433, 2013. C. Rougé, J.-D. Mathias, and G. Deffuant. Vulnerability: From the conceptual to the operational using a dynamical system perspective. , 73(Supplement C):218 – 230, 2015. M. Shaked and J. G. Shanthikumar. . Springer-Verlag, Berlin, 2007. [^1]: Université Paris-Est, Cermics (ENPC), F-77455 Marne-la-Vallée, [^2]: We also hinted at the possibility to use so-called stochastic orders.

--- abstract: 'We construct explicit counterexamples that show that it is impossible to get any remainder, other than the classical ones, in the Wiener-Ikehara theorem and the Ingham-Karamata theorem under just an additional analytic continuation hypothesis to a half-plane (or even to the whole complex plane).' address: | Department of Mathematics: Analysis, Logic and Discrete Mathematics\ Ghent University\ Krijgslaan 281\ 9000 Gent\ Belgium author: - Frederik Broucke - Gregory Debruyne - Jasson Vindas title: 'On the absence of remainders in the Wiener-Ikehara and Ingham-Karamata theorems: a constructional approach' --- [^1] [^2] [^3] Introduction ============ The Wiener-Ikehara theorem and the Ingham-Karamata theorem are two cornerstones of complex Tauberian theory. Both results have numerous applications in diverse areas such as number theory, operator theory, and partial differential equations. We refer to the monographs [@A-B-H-N; @korevaarbook; @Tenenbaumbook] for accounts on these theorems and related complex Tauberian theorems. The classical Wiener-Ikehara theorem states that if a function $S$ is non-decreasing on $[0,\infty)$ and has convergent Laplace-Stieltjes transform on the half-plane ${\operatorname{Re}}s>1$ such that $$\label{eq: Laplace extension W-I} \mathcal{L}\{\dif S; s\}-\frac{a}{s-1}= \int_{0}^{\infty} e^{-sx}\dif S(x) -\frac{a}{s-1}$$ admits an analytic extension beyond ${\operatorname{Re}}s=1$, then $S$ has asymptotic behavior $$\label{eq: asymp S W-I} S(x)=ae^{x}+o(e^x).$$ On the other hand, one version of the Ingham-Karamata theorem says that if a function $\tau$ is Lipschitz continuous on $[0,\infty)$ and if its Laplace transform $$\mathcal{L}\{\tau ; s\}=\int_{0}^{\infty} \tau(x)e^{-s x}\dif x$$ has a analytic continuation across the imaginary axis, then $$\label{eq: asymp tau I-K} \tau(x)=o(1).$$ We have stated here the simplest forms of these results, but we point out that both theorems have been extensively studied over the last century and have been generalized in a variety of ways. For instance, see [@B-B-T2016; @Chill-Seifert2016; @d-vW-I2016; @d-vOptIngham; @d-vCT; @revesz-roton; @S; @zhang2019] for recent contributions. In a recent article [@d-vAbsenceI] the last two named authors have proved that, in general, it is impossible to improve the error terms of the asymptotic formulas and in the Wiener-Ikehara theorem and the Ingham-Karamata theorem if one just augments the assumptions of these theorems by asking an additional analytic continuation hypothesis to a half-plane containing ${\operatorname{Re}}s>1$ or ${\operatorname{Re}}s>0$, respectively. In the case of the Wiener-Ikehara theorem, this disproves a conjecture by Müger [@muger], who had conjectured that a certain remainder could be obtained if can be analytically extended to ${\operatorname{Re}}s>\alpha$ with some $0<\alpha<1$. It has indeed been shown in [@d-vAbsenceI] that no stronger remainder than the one in can be achieved if this extra assumption is solely made together with the classical hypotheses. The proofs of the quoted results on the absence of remainders in the Wiener-Ikehara and Ingham-Karamata theorems given in [@d-vAbsenceI] are non-constructive as they rely on abstract functional analysis arguments (the open mapping theorem for Fréchet spaces). In particular, they do not deliver any concrete counterexample for specific remainders. One might still wonder how such counterexamples could explicitly be found. The goal of this article is to address the latter constructive problem. In fact, we shall construct explicit instances of functions that show the ensuing theorem. Note that Theorem \[th: absence in W-I and I-K\] improves [@d-vAbsenceI Theorem 3.1 and Theorem 4.2]. We recall that the notation $f(x)=\Omega_{\pm}(g(x))$ means that there is $c>0$ such that the inequalities $f(x)>cg(x)$ and $f(x)<-cg(x)$ hold infinitely often for arbitrary large values of $x$. \[th: absence in W-I and I-K\] Let $\rho$ be a positive function tending to $0$. 1. There is a non-decreasing function $S$ on $[0,\infty)$ whose Laplace-Stieltjes transform converges for ${\operatorname{Re}}s>1$ and $$\mathcal{L}\{\dif S; s\} - \frac{1}{s-1}$$ extends to the whole complex plane $\mathbb{C}$ as an entire function, but satisfying the oscillation estimate $$S(x)= e^{x}+\Omega_{\pm}(\rho(x)e^{x}).$$ 2. There is a smooth function $\tau$ on $(0,\infty)$ with bounded derivative whose Laplace transform $\mathcal{L}\{\tau; s\}$ can be analytically continued to the whole $\mathbb{C}$, but satisfying the oscillation estimate $$\tau(x)=\Omega_{\pm}(\rho(x)).$$ We end this introduction by mentioning that it is actually possible to obtain quantified error terms in complex Tauberian theorems for the Laplace transform, but, as e.g. Theorem \[th: absence in W-I and I-K\] shows, additional assumptions on the Laplace transform besides analytic continuation are required. Determining such conditions is a central problem in modern complex Tauberian theory and much progress on this question has been made in the last decade, see e.g. [@B-B-T2016; @B-T2010; @Chill-Seifert2016; @D-S2019; @R-S-S; @S]. Many of such results are motivated by the theory of operator semigroups and have applications in partial differential equations and dynamical systems. Main constructions {#Section: constructions examples W-I and I-K} ================== Our construction relies on three main lemmas, which are presented in this section. The first result allows one to regularize functions that increase to infinity slower than $\sqrt{x}$ . \[lem: reg\] Let $\omega$ be a positive non-decreasing function on $(0,\infty)$ satisfying $$\lim_{x\to \infty} \omega(x) = \infty \quad \text{and} \quad \omega(x) \ll \sqrt{x}.$$ Then there exists $W\in C^{\infty}(0,\infty)$ with the following properties: 1. \[itm: a\] $\omega(x) \ll W(x) \ll \omega(x^{2})$; 2. \[itm: b\] $W(ax) \geq aW(x)$ for every $a\leq 1$; 3. \[itm: c\] $W'(x) \geq 0$; 4. \[itm: d\] for any $n\geq 1$ and $x>0$, $$\abs[1]{W^{(n)}(x)} \leq 2^{n+1}n!\frac{W(x)}{x^n}.$$ Consider the Poisson kernel of the real line, $$P(x,y) = \frac{y}{y^{2}+x^{2}} = \frac{{\mathrm{i}}}{2}\biggl( \frac{1}{x+{\mathrm{i}}y} - \frac{1}{x-{\mathrm{i}}y}\biggr).$$ Differentiating the last expression with respect to $y$, we obtain the bounds $$\begin{aligned} \abs{\dpd[n]{P}{y}(x,y)} \leq \frac{2^{n+1}n! y^{n+1}}{(y^{2}+x^{2})^{n+1}}\max_{0\leq j \leq n+1}\abs{x/y}^{j} \leq \frac{2^{n+1}n!}{(y^{2}+x^{2})^{(n+1)/2}}, \quad \text{for all } n\geq 1.\end{aligned}$$ We set $$W(y) = \int_{0}^{\infty}\omega(x)P(x,y)\dif x = \int_{0}^{\infty}\omega(xy)P(x,1)\dif x.$$ We have $$W(y) \geq \int_{1}^{\infty}\omega(xy)P(x,1)\dif x \gg \omega(y);$$ and $$\begin{aligned} W(y) &= \int_{0}^{y}\omega(xy)P(x,1)\dif x + \int_{y}^{\infty}\omega(xy)P(x,1)\dif x \\ &\ll \omega(y^{2}) + \sqrt{y}\int_{y}^{\infty}\frac{\sqrt{x}}{1+x^{2}}\dif x \ll \omega(y^{2})+1.\end{aligned}$$ This proves \[itm: a\]. Property \[itm: b\] follows immediately from the definition of $W$. For \[itm: c\], $$\pd{P}{y}(x,y) = \frac{x^{2}-y^{2}}{(x^{2}+y^{2})^{2}},$$ so $$\begin{aligned} W'(y) &= \int_{0}^{y}\omega(x)\frac{x^{2}-y^{2}}{(x^{2}+y^{2})^{2}}\dif x + \int_{y}^{\infty}\omega(x)\frac{x^{2}-y^{2}}{(x^{2}+y^{2})^{2}}\dif x \\ &\geq \omega(y)\int_{0}^{\infty}\frac{x^{2}-y^{2}}{(x^{2}+y^{2})^{2}}\dif x = 0.\end{aligned}$$ Finally, $$\begin{aligned} \abs[1]{W^{(n)}(y)} &\leq 2^{n+1}n!\int_{0}^{\infty}\frac{\omega(x)}{(x^{2}+y^{2})^{(n+1)/2}}\dif x \\ &= 2^{n+1}n! y^{-n} \int_{0}^{\infty}\frac{\omega(xy)}{(x^{2}+1)^{(n+1)/2}}\dif x \\ &\leq 2^{n+1}n!y^{-n}W(y).\end{aligned}$$ The rest of this section is devoted to studying properties of various functions associated to the oscillatory function $\cos (x W(x))$, where $W$ satisfies some of the above properties \[itm: a\]-\[itm: d\]. \[lem: omega result\] Let $W$ be a differentiable function tending to $\infty$ which satisfies the properties \[itm: c\] and \[itm: d\] stated in Lemma \[lem: reg\]. Define $$\label{eq: def T} T(x) = \int_{0}^{x}e^{u}\cos \bigl(uW(u) \bigr) \dif u$$ and $$\label{eq: def V} V(x) = W(x) + xW'(x).$$ Then, $$\label{eq: asymptotics T} T(x) = \frac{e^{x}}{V(x)}\sin\bigl( xW(x)\bigr) + O\biggl( \frac{e^{x}}{V(x)^{2}} \biggr).$$ Integrating by parts, $$T(x) = \frac{e^{x}}{V(x)}\sin\bigl( xW(x)\bigr) + O(1) \\ - \int_{1}^{x}e^{u}\sin\bigl(uW(u)\bigr)\biggl( \frac{1}{V(u)} - \frac{V'(u)}{V(u)^{2}}\biggr)\dif u.$$ To estimate the remaining integral, we perform once more integration by parts and obtain that it equals $$\biggl(\frac{1}{V(x)^{2}} - \frac{V'(x)}{V(x)^{3}}\biggr)e^{x}\cos\bigl(xW(x)\bigr) + O(1) + O\Biggl( \int_{1}^{x}\abs[3]{ \biggl(\frac{e^{u}}{V(u)^{2}} - \frac{e^{u}V'(u)}{V(u)^{3}}\biggr)' } \dif u\Biggr).$$ The first term is of the desired order of growth in view of the regularity assumption \[itm: d\]. The derivative inside the integral equals $$e^{u}\left(\frac{1}{V(u)^{2}} - 3\frac{V'(u)}{V(u)^{3}} - \frac{V''(u)}{V(u)^{3}} + 3\frac{V'(u)^{2}}{V(u)^{4}}\right)= \frac{e^{u}}{V(u)^{2}}+O\left(\frac{e^{u}}{uV(u)^{2}}\right) ,$$ again by the regularity assumption \[itm: d\], and it is thus eventually positive. Hence the integral is bounded by $$O(1) + \frac{e^{x}}{V(x)^{2}} - \frac{e^{x}V'(x)}{V(x)^{3}}.$$ It remains to observe that property \[itm: d\] yields $W(x)\ll x^{4} $, which implies that the $O(1)$ terms above are in fact $O(x/V(x)^{2})$. This concludes the proof of the lemma. The last key ingredient in our argument is the analytic continuation property of the Laplace transform of $\cos(x W(x))$ that is obtained in the ensuing lemma. Before we state it, let us point out that if a locally bounded function $W$ satisfies property \[itm: b\], one must necessarily have $W(x)\ll x$. In fact, let $x\in[2^{n-1},2^{n}]$; upon iterating property \[itm: b\] with $a=1/2$, one obtains $W(x)\leq 2^{n}W(x/2^{n})\leq 2x \sup_{u\in[1/2,1]} W(u) $. \[lem: analytic continuation\] Suppose $W$ is a smooth function tending to $\infty$ and satisfying \[itm: b\]-\[itm: d\] from Lemma \[lem: reg\]. Then, the Laplace transform $$\mathcal{L} \{ \cos(xW(x)); s\} = \int_{0}^{\infty}\cos(xW(x))e^{-sx}\dif x$$ admits an analytic continuation to the whole complex plane. We shall prove the continuation of $$F(s) \coloneqq \int_{0}^{\infty}e^{{\mathrm{i}}xW(x)}e^{-sx}\dif x,$$ whence the lemma follows since $\mathcal{L} \{ \cos(xW(x)); s\} = \bigl( F(s) + \overline{F(\overline{s})} \bigr)/2$. Using property \[itm: d\], one sees that the $n$-th Taylor coefficient of $W$ at $x$, $c_{n,x}$, satisfies $\abs[0]{c_{n, x}} \leq 2^{n+1}x^{-n}W(x)$, so that its Taylor series at $x$ has radius of convergence at least $x/2$. This shows that $W$ has analytic continuation to the sector $\{ z \in {\mathbb{C}}: \abs{\arg z} < \pi/6 \}$. The idea of the proof is to shift the integration contour to one where the real part of ${\mathrm{i}}zW(z)$ is sufficiently negative, in order to obtain an integral which is convergent for any value of $s\in {\mathbb{C}}$. Consider $z = Re^{{\mathrm{i}}\theta}$ with $0\leq \theta \leq \arctan(1/3)$. First we deduce some bounds on $${\operatorname{Re}}({\mathrm{i}}zW(z))= -R\bigl(\sin \theta {\operatorname{Re}}W(z) + \cos \theta {\operatorname{Im}}W(z) \bigr).$$ Expanding $W$ in its Taylor series around $R\cos \theta$, we get $$\begin{aligned} W(Re^{{\mathrm{i}}\theta}) &= W(R\cos\theta) + \sum_{n=1}^{\infty}(-1)^{n}c_{2n, R\cos\theta}(R\sin\theta)^{2n} \\ &{}+ {\mathrm{i}}\sum_{n=0}^{\infty}(-1)^{n}c_{2n+1,R\cos\theta}(R\sin\theta)^{2n+1}.\end{aligned}$$ Employing the bounds on $c_{n, R\cos\theta}$ and property \[itm: c\], which implies $c_{1, R\cos\theta}\geq 0$, we get $$\begin{aligned} \label{eq: W real part} {\operatorname{Re}}W(Re^{{\mathrm{i}}\theta}) &\geq W(R\cos\theta) - W(R\cos\theta)(\tan\theta)^{2}\sum_{n=0}^{\infty}2^{2n+3}(\tan\theta)^{2n}, \\ \nonumber {\operatorname{Im}}W(Re^{{\mathrm{i}}\theta}) &\geq - W(R\cos\theta)(\tan\theta)^{3}\sum_{n=0}^{\infty}2^{2n+4}(\tan\theta)^{2n}.\end{aligned}$$ If we choose $\theta$ such that $(\tan\theta)^{2}\leq 1/W(R)$, we obtain $${\operatorname{Re}}({\mathrm{i}}zW(z)) \leq -R\bigl((\sin\theta) W(R\cos\theta) + O(1)\bigr).$$ Consider now the contours $$\Gamma_{R}\colon [R_{0}, R] \to {\mathbb{C}}\colon r \mapsto r\exp\biggl({\mathrm{i}}\arctan\frac{1}{\sqrt{W(r)}}\biggr)$$ for some $R_{0}$ sufficiently large (so that $\arctan(W(R_{0})^{-1/2})<\arctan(1/3)$), and $$C_{R}\colon \bigl[0, \arctan \bigl(W(R)^{-1/2}\bigr)\bigr] \to {\mathbb{C}}\colon \theta \mapsto Re^{{\mathrm{i}}\theta}.$$ Using , one verifies that for ${\operatorname{Re}}s \geq\sigma_{0}$, with sufficiently large $\sigma_{0}$, the integral over $C_{R}$ tends to 0 as $R\to\infty$. For the integral over $\Gamma_{\infty}$, we get $$\abs[3]{\int_{\Gamma_{\infty}}} \ll \int_{R_{0}}^{\infty}\exp\biggl(-r\frac{W(r)}{2\sqrt{1+W(r)}} + (C+ \abs{s})r\biggr)\dif r,$$ for some constant $C$, since $\sin\arctan\bigl(W(r)^{-1/2}\bigr) = (1+W(r))^{-1/2}$, $\dif z = O(1)\dif r$ by property \[itm: d\], and $W(r\cos\theta) \geq W(r)/2$ for $\theta\le \pi/3$ by property \[itm: b\]. Since $\sqrt{W(r)} \to \infty$ (by property \[itm: a\]), the integral over $\Gamma_{\infty}$ converges absolutely and uniformly for $s$ on any compact subset of ${\mathbb{C}}$, and hence represents an entire function. In conclusion, the formula $$F(s) = \int_{[0,R_{0}]\cup C_{R_{0}}\cup \Gamma_{\infty}}e^{{\mathrm{i}}zW(z)}e^{-sz}\dif z,$$ valid for in a certain right half-plane in view of Cauchy’s theorem, yields the analytic continuation of $F(s)$ to ${\mathbb{C}}$. The examples {#section: the examples W-I and I-K} ============= We have already done all the necessary work in order to establish Theorem \[th: absence in W-I and I-K\]. We set $$\tilde{\rho}(x) = \sup_{y\ge x}\rho(y), \quad \omega(x) = \min\bigl(\sqrt{x}, 1/\tilde{\rho}\bigl(\sqrt{x}\bigr)\bigr),$$ and let $W$ then be a function fulfilling the conditions \[itm: a\]-\[itm: d\] from Lemma \[lem: reg\]. \[counterexample 1 W-I\] We consider the non-decreasing function $$S(x) = \int_{0}^{x}e^{u}\bigl(1 + \cos\bigl(uW(u) \bigr)\bigr) \dif u, \qquad x\geq 0.$$ Since $$\begin{aligned} V(x) \leq 5 W(x) \ll \omega(x^{2}) \leq 1/\rho(x),\end{aligned}$$ Lemma \[lem: omega result\] tells us that $S(x) = e^{x} + \Omega_{\pm}(e^{x}\rho(x))$. On the other hand, by Lemma \[lem: analytic continuation\], its Laplace-Stieltjes transform $\mathcal{L}\{\dif S; s\}$ extends to a meromorphic function on ${\mathbb{C}}$ with a single simple pole with residue 1 at $s=1$. \[counterexample 2 I-K\] This time we define our example as $$\tau(x) = \int_{x}^{\infty} \cos\bigl(uW(u) \bigr) \dif u, \qquad x\geq 0.$$ Then, a similar calculation as in Lemma \[lem: omega result\] shows that $$\int_{x}^{y}\cos\bigl(uW(u) \bigr) \dif u= \frac{\sin (yW(y))}{V(y)}- \frac{\sin (xW(x))}{V(x)}+ O\left(\frac{1}{V(x)^{2}}\right),$$ so that the defining improper integral indeed converges and $\tau(x) = \Omega_{\pm}(1/V(x)) = \Omega_{\pm}(\rho(x))$. That the Laplace transform of $\tau$ has entire extension follows directly from Lemma \[lem: analytic continuation\]. [99]{} W. Arendt, C. J. K. Batty, M. Hieber, F. Neubrander, *Vector-valued Laplace transforms and Cauchy problems,* Birkhäuser/Springer, Basel, 2011. C. J. K. Batty, A. Borichev, Y. Tomilov, *$L^p$-tauberian theorems and $L^p$-rates for energy decay,* J. Funct. Anal. **270** (2016), 1153–1201. A. Borichev, Y. Tomilov, *Optimal polynomial decay of functions and operator semigroups,* Math. Ann. **347** (2010), 455–478. R. Chill, D. Seifert, *Quantified versions of Ingham’s theorem,* Bull. Lond. Math. Soc. **48** (2016), 519–532. G. Debruyne, D. Seifert, *An abstract approach to optimal decay of functions and operator semigroups*, Israel J. Math. **233** (2019), 439–451. G. Debruyne, J. Vindas, *Generalization of the Wiener-Ikehara theorem,* Illinois J. Math. **60** (2016), 613–624. G. Debruyne, J. Vindas, *Optimal Tauberian constant in Ingham’s theorem for Laplace transforms,* Israel J. Math. **228** (2018), 557–586. G. Debruyne, J. Vindas, *Note on the absence of remainders in the Wiener-Ikehara theorem,* Proc. Amer. Math. Soc. **146** (2018), 5097–5103. G. Debruyne, J. Vindas, *Complex Tauberian theorems for Laplace transforms with local pseudofunction boundary behavior,* J. Anal. Math. **138** (2019), 799–833. J. Korevaar, *Tauberian theory. A century of developments*, Springer-Verlag, Berlin, 2004. M. Müger, *On Ikehara type Tauberian theorems with $O(x^{\gamma})$ remainders*, Abh. Math. Semin. Univ. Hambg. **88** (2018), 209–216. Sz. Gy. Révész, A. de Roton, *Generalization of the effective Wiener-Ikehara theorem,* Int. J. Number Theory **9** (2013), 2091–2128. J. Rozendaal, D. Seifert, R. Stahn, *Optimal rates of decay for operator semigroups on Hilbert spaces,* Adv. Math. **346** (2019), 359–388. R. Stahn, *Decay of $C_0$-semigroups and local decay of waves on even (and odd) dimensional exterior domains*, J. Evol. Equ. **18** (2018), 1633–1674. G. Tenenbaum, *Introduction to analytic and probabilistic number theory,* American Mathematical Society, Providence, RI, 2015. W.-B. Zhang, *Exact Wiener-Ikehara theorems,* Acta Arith. **187** (2019), 357–380. [^1]: F. Broucke was supported by the Ghent University BOF-grant 01J04017 [^2]: G. Debruyne acknowledges support by Postdoctoral Research Fellowships of the Research Foundation–Flanders and the Belgian American Educational Foundation. The latter one allowed him to do part of this research at the University of Illinois at Urbana-Champaign. [^3]: J. Vindas was partly supported by Ghent University through the BOF-grant 01J04017 and by the Research Foundation–Flanders through the FWO-grant 1510119N

--- abstract: 'A rigorous Floquet mode analysis is proposed for a zero thickness space-time modulated Huygens’ metasurface to model and determine the strengths of the new harmonic components of the scattered fields. The proposed method is based on Generalized Sheet Transition Conditions (GSTCs) treating a metasurface as a spatial discontinuity. The metasurface is described in terms of Lorentzian electric and magnetic surface susceptibilities, $\chi_\text{ee}$ and $\chi_\text{mm}$, respectively, and its resonant frequencies are periodically modulated in both space and time. The unknown scattered fields are then expressed in terms of Floquet modes, which when used with the GSTCs, lead to a system of field matrix equations. The resulting set of linear equations are then solved numerically to determine the total scattered fields. Using a finite-difference time domain (FDTD) solver, the proposed method is validated and confirmed for several examples of modulation depths ($\Delta_p$) and frequencies ($\omega_p$). Finally, the computed steady-state scattered fields are Fourier propagated analytically, for visualization of refracted harmonics. The proposed method is simple and versatile and able to determine the steady-state response of a space-time modulated Huygen’s metasurface, for arbitrary modulation frequencies and depths.' author: - 'Shulabh Gupta, Tom. J. Smy and Scott A. Stewart [^1]' bibliography: - 'Gupta\_Analytical\_STModulated\_TAP\_2017.bib' title: | Floquet-Mode Solutions of\ Space-Time Modulated Huygens’ Metasurfaces --- Electromagnetic Metasurfaces, Electromagnetic Propagation, Floquet Analysis, Explicit Finite-Difference, Generalized Sheet Transition Conditions (GSTCs), Lorentz Dispersions, Parametric Systems. Introduction ============ Recently, there has been a strong growing interest in Huygen’s metasurfaces due to their impedance matching capabilities with free-space and their versatile applications in wavefront shaping [@Elliptical_DMS][@GeneralizedRefraction][@meta3]. They are constructed using a 2-D array of electrically small Huygen’s sources, exhibiting perfect cancellation of backscattered fields, due to optimal interactions of their electric and magnetic dipolar moments [@Kerker_Scattering]. Some efficient implementations of Huygens’ metasurfaces are based on all-dielectric resonators [@Kivshar_Alldielectric][@Elliptical_DMS][@AllDieelctricMTMS] and orthogonally collocated small electric and magnetic dipoles [@Grbic_Metasurfaces][@HuygenBook_Eleftheriades]. While the majority of the work on metasurfaces has been focussed on static (linear time invariant) metasurfaces, there is also a growing interest in dynamic metasurfaces, where the constitutive parameters of the metasurface unit cells are controlled in real-time. An important class of such dynamic electromagnetic structures is *space-time modulated metasurfaces*, where their constitutive parameters are periodically modulated in both space and time, at comparable frequency scales to the input excitations. This leads to a complex interaction of the incident wavefronts with the metasurfaces resulting in exotic effects such as generation of new harmonic components and Lorentz non-reciprocity [@STGradMetasurface][@ShaltoutSTMetasurface]. Space-time modulated metasurfaces fall within the general framework of space-time modulated mediums [@TamirST][@OlinerST], which have found important applications in acousto-optical systems for spectrum analysis, parametric oscillators and amplifiers, for instance [@Goodman_Fourier_Optics][@Saleh_Teich_FP][@TamirAcoustoDiffraction]. Consequently, a combination of the wave-shaping capabilities of Huygens’ metasurfaces with space-time modulation principles, is an interesting avenue to explore for advanced electromagnetic wave control, in both space and time. To investigate into the properties of space-time modulated Huygens’ metasurfaces, a finite-difference time-domain (FDTD) technique has recently been proposed to analyze a zero thickness model of Huygens’ metasurfaces [@Smy_Metasurface_Linear][@Stewart_Metasurface_STM], based on Generalized Sheet Transition Conditions (GSTCs) [@KuesterGSTC]. In contrast to the other existing techniques for analyzing static metasurfaces in frequency domain [@CalozFDTD][@Caloz_MetaModelling], the FDTD analysis is naturally applicable to the problem of space-time modulated metasurfaces, considered herein. While such numerical methods are useful to determine the time-evolution of scattered waves from space-time modulated metasurfaces for a given input wave, efficient determination of steady-state response of the metasurface is an equally important problem to solve. This issue is addressed in this work, whereby exploiting the periodic nature of the spatio-temporal perturbation on the metasurface, the scattered fields are expressed in terms of Floquet modes. The Huygens’ metasurface is modelled using surface susceptibilities following a physically motivated Lorentzian profile, whose resonant frequencies are parametrized to emulate a space-time modulation of the metasurface. Combined with GSTCs, the Floquet mode amplitudes are computed by solving a set of linear equations. The proposed method thus efficiently computes the steady-state response of a zero-thickness space-time modulated Huygens’ metasurface, consistent with the FDTD field solutions. Furthermore, integrating the proposed method with analytical Fourier methods [@Goodman_Fourier_Optics], the propagation of scattered fields are conveniently visualized in free-space. The paper is structured as follows. Section II describes the problem statement of this work, and develops the fields equations governing the scattering fields from a space-time modulated metasurface based on GSTCs and Lorentz surface susceptibilities. Section III presents the proposed method based on Floquet mode expansions, forming the set of linear equations to be solved numerically. Several results are then presented for both time-only and space-time modulated metasurfaces. Conclusions are provided in Sec. IV, and a brief summary of the FDTD method is provided in Appendix V, for the sake of self-consistency and completeness of the paper. Space-Time Modulated Metasurfaces ================================= Problem Statement ----------------- Metasurfaces are zero thickness electromagnetic structures that act as discontinuities in space. The exact zero thickness nature of electromagnetic structures was developed by Idemen in terms of Generalized Sheet Transition Conditions (GSTCs) [@IdemenDiscont], which were later applied to metasurfaces [@KuesterGSTC]. Consider a Huygens’ metasurface illustrated in Fig. \[Fig:Problem\], where the metasurface lies in the $x-y$ plane at $z=0$, with a wave incidence on the left, normal to the surface. The wave interacts with the metasurface and produces a transmitted and a reflected wave, along the forward and backward direction, respectively. This interaction of the metasurface with the electromagnetic waves is described by the GSTCs, using the transverse components of electric and magnetic surface polarizabilities $\mathbf{P}$ and $\mathbf{M}$, as [@Metasurface_Synthesis_Caloz] \[Eq:GSTC\] $$\hat{\mathbf{z}}\times\Delta \mathbf{H}(x,t) = \frac{d\mathbf{P}_{||}(x,t)}{dt}$$ $$\Delta \mathbf{E}(x,t)\times \hat{\mathbf{z}} = \mu_0\frac{d\mathbf{M}_{||}(x,t)}{dt},$$ where $$\begin{aligned} \Delta \mathbf{E} = (\mathbf{E}_t - \mathbf{E}_0 - \mathbf{E}_r),\quad \Delta \mathbf{H} = (\mathbf{H}_t - \mathbf{H}_0 - \mathbf{H}_r). \notag\end{aligned}$$ The surface polarizabilities on the Huygens’ metasurface are related to the average fields around the metasurface and can be described in terms of scalar electric and magnetic surface susceptibilities $\chi_\text{ee}$ and $\chi_\text{mm}$, respectively, as[^2]. $$\mathbf{\tilde{Q}}_{||}(\omega) = \tilde{\chi}_\text{ee} \mathbf{\tilde{E}}_\text{av}(\omega),$$ $$\mathbf{\tilde{M}}_{||}(\omega) = \tilde{\chi}_\text{mm} \mathbf{\tilde{H}}_\text{av}(\omega),$$ $$\text{where}~\mathbf{\tilde{E}}_\text{av} = \left[ \frac{\mathbf{\tilde{E}}_0 + \mathbf{\tilde{E}}_t + \mathbf{\tilde{E}}_r}{2}\right],~\mathbf{\tilde{H}}_\text{av} = \left[ \frac{\mathbf{\tilde{H}}_0 + \mathbf{\tilde{H}}_t + \mathbf{\tilde{H}}_r}{2}\right]\notag$$ and $\mathbf{Q} = \mathbf{P}/\epsilon_0$ is the normalized electric polarizability[^3]. \[c\]\[c\]\[0.8\][$z=0$]{} \[c\]\[c\]\[0.8\][$z$]{} \[c\]\[c\]\[0.8\][$x$]{} \[c\]\[c\]\[0.8\][$t$]{} \[c\]\[c\]\[0.8\][$\psi(\mathbf{r}, t) = \psi_0(\mathbf{r}, t) \sin(\omega_0t)$]{} \[c\]\[c\]\[0.6\][$\boxed{\chi_{ee}(x, t),\; \chi_{mm}(x, t)}$]{} \[c\]\[c\]\[0.6\][$\psi_2(\mathbf{r}, t) \sin\{(\omega_0 - \omega_p)t\}$]{} \[l\]\[c\]\[0.6\][$\psi_1(\mathbf{r}, t) \sin\{\omega_0t\}$]{} \[c\]\[c\]\[0.6\][$\psi_3(\mathbf{r}, t) \sin\{(\omega_0 + \omega_p)t\}$]{} ![A general Illustration of a space-time modulated Huygens’ metasurface under normally incident CW plane-wave resulting in generation of several frequency harmonics, refracted along different angles. $\omega_p$ is the pumping frequency.[]{data-label="Fig:Problem"}](TimeVaryingProblem "fig:"){width="0.75\columnwidth"} Next, consider a space-time modulated metasurface, whose electric and magnetic susceptibilities are both a function of space and time, i.e. $ \chi_\text{ee}(x, t)$ and $ \chi_\text{mm}(x, t)$. Let us restrict here to a periodic modulation only, with a pumping frequency $\omega_p$ and the spatial frequency $\beta_p$. Due to this periodic spatio-temporal perturbation on the metasurface, assumed to be infinite in size, the scattered transmission (and reflection) fields can be expressed in terms of Floquet series as $$\label{Eq:TotalTFields} E_t(x, z= 0_+, t) = \sum_{n=0}^\infty q_n e^{j\omega_n t} e^{j\beta_n x},$$ where $\omega_n = \omega_0 + n\omega_p$, and $\beta_n = \beta_{x0} + n\beta_p$ with $\beta_{x0}=0$ due to assumed normal input incidence. Each harmonic term of this expansion with a temporal frequency $\omega_n$, represents an oblique forward propagating plane-wave in the $+z$ direction, and making an angle $\theta_n$ measured from the normal of the metasurface, as illustrated in Fig. \[Fig:Problem\]. These refraction angles are given by $$\label{Eq:BetapRefract} \theta(\omega_n) = \sin^{-1}\left[\frac{n}{(1 + n\omega_p/\omega_0) }\frac{\beta_p}{k_0}\right],$$ where $k_0$ is the free-space wavenumber of the fundamental frequency. Based on this simple physical argument, the angle of refraction of these newly generated frequency components from a space-time modulated metasurface, can be determined. However, the strengths of these harmonics (weights $q_n$) are still unknown and must be determined taking the exact electromagnetic interaction of the waves with the metasurface into account using the GSTCs of . Field Equations --------------- The space-time modulated problem considered here is naturally treated in the time domain. Consequently, the surface susceptibilities must be defined in time domain as well. The most common, and causal, description of the susceptibilities is in terms of Lorentz dispersion, which is also typical of the Huygens’ sources used to construct the metasurfaces [@Kivshar_Alldielectric][@PDM_Gupta][@Huygen_Microwave_Achouri]. Consequently, the electric and magnetic susceptiblities can be expressed as \[Eq:Lorentz\] $$\tilde{\chi}_\text{ee}(\omega) = \frac{\omega_{ep}^2}{(\omega_{e0}^2 - \omega^2) + i\alpha_e \omega}$$ $$\tilde{\chi}_\text{mm}(\omega) = \frac{\omega_{mp}^2}{(\omega_{m0}^2 - \omega^2) + i\alpha_m \omega},$$ where $(\omega_{e0}$, $\omega_{m0}),\; (\omega_{ep}$, $\omega_{mp}),\; (\alpha_e, \alpha_m)$ are the electric and magnetic resonant frequencies, plasma frequencies and loss coefficients, respectively. The space-time modulation can now be introduced into the metasurface, by sinusoidally modulating, for instance, the resonant frequencies of the two Lorentzian susceptibilities, as [@Stewart_Metasurface_STM] \[Eq:w0Modulation\] $$\omega_{e0}(x, t) = \omega_{e0} \{1 + \Delta_e\cos(\omega_p t + \beta_px)\}$$ $$\omega_{m0}(x, t) = \omega_{m0} \{1 + \Delta_m\cos(\omega_pt + \beta_px)\},$$ where $\Delta_e$ and $\Delta_m$ are the modulation depths and the spatial frequencies of the perturbation, for electric and magnetic resonances, respectively. They can now be inserted into the GSTCs of , for a prescribed input fields, to determine the total scattered fields. To illustrate the procedure, let us assume a normally incident plane-wave ($|\mathbf{E_0}| =\text{const.}$) where the corresponding transmitted and reflected fields, given by $$\begin{aligned} \label{Eq:FieldConvention} &\mathbf{E}_0(z,t) = E_0e^{j(\omega_0t - k_0z)}~\mathbf{\hat{y}},~ \mathbf{H}_0(z,t) = \frac{\mathbf{\hat{z}} \times\mathbf{E}_0(z,t)}{\eta_0} \\ &\mathbf{E}_t(z,t) = E_t(x,t)e^{-j k_0z}~\mathbf{\hat{y}},~ \mathbf{H}_t(z,t) = \frac{\mathbf{\hat{z}} \times\mathbf{E}_t(z,t)}{\eta_0} \notag\\ &\mathbf{E}_r(z,t) = E_r(x,t)e^{+j k_0z}~\mathbf{\hat{y}},~ \mathbf{H}_r(z,t) = \frac{\mathbf{E}_r(z,t)\times \mathbf{\hat{z}} }{\eta_0}.\notag\end{aligned}$$ Each of these field components are related to their respective polarizabilities on the metasurface (at $z=0$) through the Lorentz susceptibilities of with , which can then be expressed in the time domain as[^4] [@Smy_Metasurface_Linear] \[Eq:TD\_Lorentz\] $$\frac{d^2Q_i}{dt^2} +\omega_{e0}^2\{1 + \Delta_e\cos(\omega_p t + \beta_px)\}^2Q_i = \omega_{ep}^2E_i(t)$$ $$\frac{d^2M_i}{dt^2} +\omega_{m0}^2\{1 + \Delta_m\cos(\omega_p t + \beta_px)\}^2M_i = \frac{\omega_{ep}^2}{\eta_0}E_i(t),\footnote{with proper sign of the right hand side term, depending on incident ($-$), reflected ($+$) or transmitted fields ($-$), according to \eqref{Eq:FieldConvention}.}$$ where the subscript $i = 0, t, r$ for incident, transmitted and reflected fields, respectively. Finally, all the scattered fields are related to their polarizabilities following the GSTCs of , as \[Eq:GSTCEquations\] $$\frac{dQ_0}{dt} + \frac{dQ_t}{dt} + \frac{dQ_r}{dt} = \frac{2}{\eta_0\epsilon_0}(E_0 - E_r - E_t),$$ $$\frac{dM_0}{dt} + \frac{dM_t}{dt} + \frac{dM_r}{dt} = \frac{2}{\mu_0}(E_t - E_r - E_0)$$ Equations and thus represent two sets of field equations, that must be solved to determine the transmitted and reflected fields, $E_t(x,z, t)$ and $E_r(x,z, t)$, in steady state, for a given input excitation $\mathbf{E_0} = E_0e^{j\omega_0t}~\mathbf{\hat{y}}$ at the input of the metasurface at $z=0_-$. $$\begin{aligned} \label{Eq:FieldEquations} &\sum_n \left[\Delta_{e1}a_{n-2} + \Delta_{e2} a_{n-1} + A_{e, n} a_n + \Delta_{e2} a_{n+1} + \Delta_{e1}a_{n+2} \right]e^{jn\Omega t} = E_0~\quad \text{(Incident E-Fields)} \\ % &\sum_n \left[\Delta_{e1}b_{n-2} + \Delta_{e2} b_{n-1} + A_{e, n}b_n + \Delta_{e2} b_{n+1} + \Delta_{e1}b_{n+2} -q_n \right]e^{jn\Omega t} = 0~\quad \text{(Transmitted E-Fields)} \notag\\ % &\sum_n \left[\Delta_{e1}c_{n-2} + \Delta_{e2} c_{n-1} + A_{e, n}c_n + \Delta_{e2} c_{n+1} + \Delta_{e1}c_{n+2} -p_n \right]e^{jn\Omega t} = 0~\quad \text{(Reflected E-Fields)} \notag\\ % &\sum_n \left[\Delta_{m1}d_{n-2} + \Delta_{m2} d_{n-1} + A_{m,n}d_n + \Delta_{m2} d_{n+1} + \Delta_{m1}d_{n+2} \right]e^{jn\Omega t} = -\frac{E_0}{\eta_0}~\quad \text{(Incident H-Fields)} \notag\\ % &\sum_n \left[\Delta_{m1}e_{n-2} + \Delta_{m2} e_{n-1} + A_{m,n}e_n + \Delta_{m2} e_{n+1} + \Delta_{m1} e_{n+2} -b_n \right]e^{jn\Omega t} = 0~\quad \text{(Transmitted H-Fields)} \notag\\ % &\sum_n \left[\Delta_{m1}f_{n-2} + \Delta_{m2} f_{n-1} + A_{m,n}f_n + \Delta_{m2} f_{n+1} + \Delta_{m1} f_{n+2} -f_n \right]e^{jn\Omega t} = 0~\quad \text{(Reflected H-Fields)} \notag\\ % &\sum_n \left[ 2cB_{n} a_n + 2cB_{n} b_n + 2cB_{n} c_n + p_n + q_n \right]e^{jn\Omega t} = E_0~\quad \text{(GSTC Equation)}\notag\\ % &\sum_n \left[ \frac{\mu_0B_{n}}{2} d_n + \frac{\mu_0B_{n}}{2} e_n + \frac{\mu_0B_{n}}{2} f_n + p_n - q_n \right]e^{jn\Omega t} = -E_0~\quad \text{(GSTC Equation)}\notag \end{aligned}$$ ------------------------------------------------------------------------ \ $$\begin{aligned} \label{Eq:FieldMatrix} &\overbrace{\left[ \begin{array}{cccc cccc} \mathbf{S_e} & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & \mathbf{S_e} & 0 & 0 & 0& 0 & 0 & -\mathbf{I} \\ 0 & 0 & \mathbf{S_e} & 0 & 0 & 0 & -\mathbf{I} & 0 \\ 0 & 0 & 0 & \mathbf{S_m} & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & \mathbf{S_m} & 0 & 0 & \mathbf{I}/\eta_0 \\ 0 & 0 & 0 & 0& 0 & \mathbf{S_m} & -\mathbf{I}/\eta_0 & 0 \\ \mathbf{S_g}/(2c_0) & \mathbf{S_g}/(2c_0) & \mathbf{S_g}/(2c_0) & 0& 0 & 0 & \mathbf{I} & \mathbf{I}\\ 0 & 0 & 0 & \mathbf{S_g} & \mathbf{S_g} & \mathbf{S_g} & \mathbf{I} & -\mathbf{I} \end{array} \right]}^{\mathbf{C_n}} % \overbrace{\left[ \begin{array}{c} \mathbf{a_n} \\ \mathbf{b_n}\\ \mathbf{c_n}\\ \mathbf{d_n}\\ \mathbf{e_n}\\ \mathbf{f_n}\\ \mathbf{p_n}\\ \mathbf{q_n} \end{array} \right]}^{\mathbf{V_n}} % = \overbrace{\left[ \begin{array}{c} \mathbf{I_0} \\ 0\\ 0\\ -\mathbf{I_0}/\eta_0\\ 0\\ 0\\ \mathbf{I_0} \\ -\mathbf{I_0} \end{array} \right] , \text{where}~\quad \mathbf{I_0} = \left[ \begin{array}{c} \vdots\\ 0\\ 0\\ 1\\ 0\\ 0\\ \vdots \end{array} \right]}^{\mathbf{E_n}} \notag\\ % &\mathbf{S_e} = \left[ \begin{array}{cccccc} A_{e, -N} & \Delta_{e2} & \Delta_{e1} & 0 & \cdots & 0 \\ \vdots & \vdots & \vdots& \vdots & \vdots & \cdots \\ \Delta_{e2} & A_{e, -1} & \Delta_2 & \Delta_{e1} & \cdots & 0 \\ \Delta_{e1} & \Delta_{e2} & A_{e,0} & \cdots & \Delta_{e2} & \Delta_{e1} \\ 0 & \cdots & \Delta_{e1} & \Delta_{e2} & A_{e, +1} & \Delta_{e2} \\ & \cdots \vdots & \vdots & \vdots& \vdots & \vdots \\ 0 & \cdots & 0 & \Delta_{e1} & \Delta_{e2} & A_{e, +N} \end{array} \right] % , \quad % % \mathbf{S_m} = \left[ \begin{array}{cccccc} A_{m, -N} & \Delta_{m2} & \Delta_{m1} & 0 & \cdots & 0 \\ \vdots & \vdots & \vdots& \vdots & \vdots & \cdots \\ \Delta_{m2} & A_{m, -1} & \Delta_2 & \Delta_{m1} & \cdots & 0 \\ \Delta_{m1} & \Delta_{m2} & A_{m, 0} & \cdots & \Delta_{m2} & \Delta_{m1} \\ 0 & \cdots & \Delta_{m1} & \Delta_{m2} & A_{m, +1} & \Delta_{m2} \\ & \cdots \vdots & \vdots & \vdots& \vdots & \vdots \\ 0 & \cdots & 0 & \Delta_{m1} & \Delta_{m2} & A_{m, +N} \end{array} \right]\notag\\ % &\mathbf{S_g} = \text{diag}\{B_{-N}, B_{-N+1},\cdots, B_{-1}, B_0, B_1, \cdots, B_{N-1}, B_N\}.\end{aligned}$$ ------------------------------------------------------------------------ Proposed Floquet Mode Expansion Solutions ========================================= Field Matrix Equations ---------------------- Let us consider a metasurface with a periodic space-time modulation of the resonant frequencies following , and expand the unknown reflected and transmitted fields using Floquet expansion as \[Eq:EtEt\_Expansion\] $$E_t(t) = \sum_n q_ne^{j\omega_0 t}e^{jn(\omega_pt + \beta_px)} = \sum_n q_ne^{j\omega_0 t}e^{jn\Omega},$$ $$E_r(t) = \sum_n p_ne^{j\omega_0 t}e^{jn(\omega_pt + \beta_px)} = \sum_n p_ne^{j\omega_0 t}e^{jn\Omega},$$ where a new variable $\Omega = (\omega_p t + \beta_p x)$ is introduced for compact notation. Similarly, the corresponding unknown polarizabilities, can also be expanded in Floquet series as $$\begin{aligned} \label{Eq:QM_Expansion} &Q_0(t) = \sum_n a_ne^{j\omega_0 t}e^{jn\Omega}, ~\quad M_0(t) = \sum_n d_ne^{j\omega_0 t}e^{jn\Omega}, \notag\\ &Q_t(t) = \sum_n b_ne^{j\omega_0 t}e^{jn\Omega}, ~\quad M_t(t) = \sum_n e_ne^{j\omega_0 t}e^{jn\Omega}, \notag\\ &Q_r(t) = \sum_n c_ne^{j\omega_0 t}e^{jn\Omega}, ~\quad M_r(t) = \sum_n f_ne^{j\omega_0 t}e^{jn\Omega}.\notag\end{aligned}$$ Next, substituting the above expressions in and and re-arranging the terms, for incident, transmitted and reflected E- and H-fields, we get , where $$\begin{aligned} &\Delta_{e1}= \frac{\Delta_e^2\omega_{e0}^2}{4\omega_{ep}^2},\quad \Delta_{e2}= \frac{\Delta_e\omega_{e0}^2}{\omega_{ep}^2}, \notag\\ & A_{e, n} = \left\{\frac{\omega_{e0}^2}{\omega_{ep}^2} - \frac{(\omega_0 + n\omega_p)^2}{\omega_{ep}^2} + \frac{\Delta_e^2\omega_{e0}^2}{2\omega_{ep}^2}\right\}\notag.\\ % &\Delta_{m1}= \frac{\Delta_m^2\omega_{m0}^2}{4\omega_{mp}^2},\quad \Delta_{m2}= \frac{\Delta_m\omega_{m0}^2}{\omega_{mp}^2}, \notag\\ & A_{m,n} = \left\{\frac{\omega_{m0}^2}{\omega_{mp}^2} - \frac{(\omega_0 + n\omega_p)^2}{\omega_{mp}^2} + \frac{\Delta_m^2\omega_{m0}^2}{2\omega_{mp}^2}\right\}\notag\\ & B_{n} = j(\omega_0 + jn\omega_p).\end{aligned}$$ These equations are obtained by expressing the cosine function using Euler’s form and then grouping the terms of common complex exponentials. Each of these series equations, can be truncated to $2N+1$ harmonic terms[^5], and be written in a matrix form as shown in . In compact form, the matrix equation can be written as $$[\mathbf{C_n}][\mathbf{V_n}] = [\mathbf{E_n}],$$ which represents $8\times (2N+1)$ linear equations with same number of unknowns. The sought field solutions can now be numerically computed using this matrix equation as $$[\mathbf{V_n}] = \text{inv}\{[\mathbf{C_n}]\} [\mathbf{E_n}],$$ from which the corresponding $p_n$’s and $q_n$’s can be extracted to construct the reflected and transmitted fields, respectively, following . Results ------- To validate the results of the proposed method, an FDTD solver is used, which has been recently proposed in [@Smy_Metasurface_Linear][@Stewart_Metasurface_STM] to directly simulate and , for an arbitrary time-domain input. A brief summary of the method is described in the appendix for self-consistency of the paper. In the results described next, the modulation depths of the electric and magnetic resonant frequencies are assumed to be equal for simplicity, i.e. $\Delta_e=\Delta_m= \Delta_p$. Furthermore, the output fields from the proposed method are constructed using a Gaussian envelopes[^6], for better conditioning of the numerical Fourier transforms and for a closer comparison with the FDTD method, given by $$\begin{aligned} \label{Eq:FieldConstruction} E_t(x,t) = \sum_{n= -N}^{+N} q_n \exp\left\{ - \left(\frac{t}{T_0}\right)^2\right\} \cos(\omega_n t + \beta_n x),\notag\\ E_r(x,t) = \sum_{n = -N}^{+N} p_n \exp\left\{ - \left(\frac{t}{T_0}\right)^2\right\} \cos(\omega_n t + \beta_n x).\end{aligned}$$ Figure \[Fig:TDvdFloquet\] shows several examples of the transmission and reflection spectrum for different modulation depths and frequencies. While the modulation frequency $\omega_p$ only fixes the location of the newly generated harmonic components, the modulation depth $\Delta_p$ controls the relative strengths of the harmonics, compared to the fundamental frequency of excitation. An excellent agreement is observed with FDTD solver, for all the cases considered, validating the proposed procedure. Minor discrepancies are observed between the two, which are attributed to several factors. Firstly, the Floquet solution assumed zero losses here[^7], while the FDTD time-domain results consider finite losses. Secondly, the Floquet solution is the steady state response of the metasurface, while the FDTD solver takes into account the dispersion based distortion of the input pulse, which is naturally not taken into account in the Floquet analysis in . \[c\]\[c\]\[0.8\][frequency, $\omega/\omega_0$]{} \[c\]\[c\]\[0.8\][$\mathcal{F}_t\{E_t(t)\}$]{} \[c\]\[c\]\[0.8\][$\mathcal{F}_t\{E_r(t)\}$]{} \[l\]\[c\]\[0.6\][FDTD]{} \[l\]\[c\]\[0.6\][Floquet]{} \[c\]\[c\]\[0.8\][$\Delta_p = 0.05$, $f_m = 0.05f_0$]{} \[c\]\[c\]\[0.8\][$\Delta_p = 0.1$, $f_m = 0.1f_0$]{} \[c\]\[c\]\[0.8\][$\Delta_p = 0.25$, $f_m = 0.25f_0$]{} ![Spectrum of transmission and reflection fields from an infinite time-only ($\beta_p=0$) modulated metasurface incident with a normally incident plane-wave, for the case of a) $\Delta_p = 0.05$, $\omega_p = 0.05\omega_0$, b) $\Delta_p = 0.1$, $\omega_p = 0.1\omega_0$ and c) $\Delta_p = 0.25$, $\omega_p = 0.25\omega_0$. Total number of harmonics $2N+1 = 61$ and $T_0 =100~$fs in . The metasurface parameters are: $\omega_{e0} = 2\pi(224.63~\text{THz})$, $\omega_{ep} = 0.36$ Trad/s, $\alpha_e = 500\times 10^9$, $\omega_{m0} = 2\pi(224.40~\text{THz})$, $\omega_{mp} = 0.29$ Trad/s, $\alpha_m = 100\times 10^9$, and excitation frequency $\omega_0 = 2\pi(230~\text{THz})$. The spectrum is normalized to the non-modulated case with $\Delta_p=0$. []{data-label="Fig:TDvdFloquet"}](results "fig:"){width="\columnwidth"} While Fig. \[Fig:TDvdFloquet\] showed results for few discrete points for $\Delta_p$ and $\Delta_m$, Fig. \[Fig:Scan\] shows a spectral map for a continuously varied modulation frequency $\omega_p$, for a given modulation depth $\Delta_p$. For simplicity, and considering the practical point of view, a perfectly matched metasurface with $\tilde{\chi}_\text{ee} = \tilde{\chi}_\text{mm}$ is assumed exhibiting zero reflections. The spectral map provides interesting and useful information about the interaction of the metasurface with the input fields. For example, at $\omega_p \approx 2\omega_0$, a substantial amplification of the fundamental frequency $\omega_0$ is observed, which is also the dominant spectral component in the transmitted fields. This is reminiscent of wave amplifications in diverse class of mechanical and electromagnetic parametric systems, where specific parameters of the system are modulated at twice the excitation frequency leading to wave instabilities [@TamirST][@OlinerST]. \[c\]\[c\]\[0.8\][frequency, $\omega/\omega_0$]{} \[c\]\[c\]\[0.8\][Modulation frequency $\omega_p/\omega_0$]{} \[l\]\[c\]\[0.8\][$20\log|E_t(z= 0_+,\omega)|$]{} \[c\]\[c\]\[0.8\][$20\log|E_t(\omega_0)|$]{} \[r\]\[c\]\[0.8\][$\boxed{\chi_\text{ee} = \chi_\text{mm},~\Delta_p = 0.1}$]{} ![Transmission spectrum of a time-only modulated metasurface for a varying modulation frequency and fixed modulation depth, $\Delta_p$. The metasurface is assumed to be matched with $\chi_\text{ee} = \chi_\text{mm}$.[]{data-label="Fig:Scan"}](fmScan "fig:"){width="\columnwidth"} Next, an example of a space-time modulated metasurface is considered. Due to the spatial dependence of the susceptibilities following , the metasurface is spatially discretized and the Floquet analysis is performed at each $x$ location to determine the transmitted and reflected fields, $\tilde{E}_t(x, z=0_+, \omega)$ and $\tilde{E}_r(x, z=0_+, \omega)$, respectively. An example is shown in Fig. \[Fig:Example\], for an input Gaussian beam, showing the spectrum of the transmitted fields at the output of the metasurface at each location, with the mismatched parameters of Fig. \[Fig:TDvdFloquet\]. The Gaussian waveform nature is clearly observed for each spectral component. More interestingly, the phase profile across the metasurface as a function of frequency is also shown in Fig. \[Fig:Example\]. A linear phase gradient with different spatial slopes, $d\angle E_t(x)/dx$, across the metasurface is clearly evident for each harmonic component, except the fundamental frequency. Therefore, each of these harmonics are expected to refract at different angles. Similar observations are also made for reflected fields (not shown here). \[c\]\[c\]\[0.8\][frequency, $\omega/\omega_0$]{} \[c\]\[c\]\[0.8\][$x~\mu$m]{} \[c\]\[c\]\[0.8\][$20\log|E_t(x, \omega)|$]{} \[c\]\[c\]\[0.8\][$\angle E_t(x, \omega)$ rad]{} \[c\]\[c\]\[0.8\][$\boxed{\Delta_p = 0.025$, $f_m = 0.1f_0}$]{} ![Transmission spectrum (amplitude and phase) of a space-time modulated metasurface, when incident with a normally incident Gaussian beam. The beam is given by $E_0(x) = \exp\{-(x/2w_x)^2\}$, with $w_x = 5~\mu$m. The metasurface size $\ell = 25~\mu$m and the excitation frequency $f_0 = 230$ THz. The pumping spatial frequency $\beta_p = 5\pi/\ell$.[]{data-label="Fig:Example"}](GaussianBeam "fig:"){width="\columnwidth"} \[c\]\[c\]\[0.8\][$x~\mu$m]{} \[c\]\[c\]\[0.8\][$z~\mu$m]{} \[c\]\[c\]\[0.8\][$20\log|E(\omega_0 - \omega_p)|$]{} \[c\]\[c\]\[0.8\][$20\log|E(\omega_0)|$]{} \[c\]\[c\]\[0.8\][$20\log|E(\omega_0 + \omega_p)|$]{} \[c\]\[c\]\[0.8\][spatial frequency $k_x\times 10^{-6}$]{} \[c\]\[c\]\[0.8\][$\mathcal{F}_x\{E(x, z= 0_{\pm}, \omega_0)\}^2$]{} \[c\]\[c\]\[0.8\][$k_{x,\text{in}}$]{} \[c\]\[c\]\[0.8\][$k_{x, -1}^\text{peak}$]{} \[c\]\[c\]\[0.8\][$k_{x, +1}^\text{peak}$]{} \[c\]\[c\]\[0.6\][$\boxed{\theta_{-1} = -8.67^\circ}$]{} \[c\]\[c\]\[0.6\][$\boxed{\theta_{0} = 0^\circ}$]{} \[l\]\[c\]\[0.6\][$\boxed{\theta_{+1} = +7.08^\circ}$]{} {width="1.75\columnwidth"} To confirm the refraction angles of the transmitted and reflected fields from the metasurface, the output fields at a specific harmonic frequency $\omega_n$, are then Fourier propagated in free-space as [@Goodman_Fourier_Optics] $$\tilde{E}_t(x, z, \omega_n) = \mathcal{F}_x^{-1}\left[\mathcal{F}_x\{ E_t(x, z_+, \omega_n)\}\exp\{-jk_{z,n}z\}\right]\notag$$ $$\tilde{E}_r(x, z, \omega_n) = \mathcal{F}_x^{-1}\left[\mathcal{F}_x\{ E_r(x, z_+, \omega_n)\}\exp\{+jk_{z,n}z\}\right]\notag$$ where $k_{z,n} = \sqrt{k_n^2 - k_x^2}$, is the free-space wave number in the $z$ direction. Fig. \[Fig:FourierFields\] shows the resulting field in the $x-z$ plane for the first down-converted $(\omega_0-\omega_p)$ and first up-converted harmonic $(\omega_0 + \omega_p)$, in addition to the fundamental frequency $\omega_0$. As expected from Fig. \[Fig:Example\] and consistent with , the fundamental frequency $\omega_0$ propagates without any refraction. On the other hand, the down-converted and up-converted harmonics are refracted in the lower and upper half of the $x-z$ plane, respectively. Their reflection angles using transverse wavenumbers are given by $$\theta(\omega_n = \omega_0 + n\omega_p) = \sin^{-1}\left[\frac{k_{x, n}^\text{peak}}{k_n}\right].$$ where $k_{x, n}^\text{peak}$ is the peak location of the spatial spectrum of the output fields, $k_n$ is the wavenumber of the $n^\text{th}$ harmonic, and the angle $\theta$ is measured from the normal of the metasurface. For the $n=\pm 1$ harmonics, the refraction angles are found to be $\theta = 8.67^\circ$ and $\theta = 7.08^\circ$, respectively, for both transmission and reflection, with a very good agreement with the theoretical values of $8.33^\circ$ and $6.81^\circ$ obtained from . Conclusions =========== A rigorous Floquet mode analysis has been proposed for a zero thickness space-time modulated Huygens’ metasurface to model and determine the strengths of the harmonics of the scattered fields. The proposed method is based on GSTCs treating the metasurface as a spatial discontinuity. The metasurface has been modelled using space-time varying resonant frequencies of the associated electric and magnetic surface susceptibilities, $\chi_\text{ee}$ and $\chi_\text{mm}$, respectively. The unknown scattered fields have been expressed in terms of Floquet modes, which when used with the GSTCs, led to a system of field matrix equations. The resulting set of linear equations were then solved to determine the transmitted and reflected scattered fields. Using an FDTD solver, the proposed method is validated and confirmed for several examples of modulation depths ($\Delta_p$) and frequencies ($\omega_p$). Finally, the computed steady-state scattered fields are Fourier propagated analytically, for convenient visualization of refracted harmonics. The proposed method is fast, simple and versatile, and is expected to be a useful tool in determining the steady-state scattered fields from a space-time modulated Huygen’s metasurface, for arbitrary modulation frequencies and depths. Appendix ======== Summary of Finite-Difference Time Domain (FDTD) Formulation of Space-Time Modulated Metasurface [@Smy_Metasurface_Linear][@Stewart_Metasurface_STM] {#summary-of-finite-difference-time-domain-fdtd-formulation-of-space-time-modulated-metasurface .unnumbered} --------------------------------------------------------------------------------------------------------------------------------------------------- Consider a space-time modulated metasurface of Fig. \[Fig:Problem\] with the plane-wave input and output fields described in . The field equations governing the transmitted and reflected fields are the time-domain Lorentz relations and the GSTC equations . Each Lorentz resonator equation of corresponding to incident, transmitted and reflected fields, is a second order differential equation. By introducing an auxiliary variable, each second-order differential equation can be decomposed into two first-order differential equations. For instance, the electric and magnetic polarizabilities corresponding to the incident fields are given by \[Eq:AuxilliaryLZ\] $$\omega_{e0}\bar{Q}_0 = \frac{dQ_0}{dt} + \alpha_e Q_0,$$ $$\frac{d\bar{Q}_0}{dt} + \bar{Q}_0\frac{1}{\omega_{e0}}\frac{d\omega_{e0}}{dt} + \omega_{e0} Q_0= \frac{ \omega_{ep}^2}{\omega_{e0}} E_0$$ $$\omega_{m0}\bar{M}_0 = \frac{dM_0}{dt} + \alpha_m M_0,$$ $$\frac{d\bar{M}_0}{dt} + \bar{M}_0\frac{1}{\omega_{m0} }\frac{d\omega_{m0}}{dt} + \omega_{m0} M_0 = -\frac{\omega_{mp}^2}{\eta_0\omega_{m0} } E_0,$$ where $\bar{Q}_0$ and $\bar{M}_0$ are the two unknown auxiliary variables in addition to $Q_0$ and $M_0$, for a specified input field $E_0$. Similar set of equations, can be developed for the transmitted and reflected fields in terms of unknown fields $E_t$ and $E_r$. They all represent a set of 12 equations, with 12 auxiliary unknowns and 2 primary unknowns. All the scattered fields and their respective polarizations are coupled to each other through the GSTC equations, which can be written in terms of new auxiliary variables as \[GSTCEquationsFDTD\] $$\begin{split} \omega_{m0}(\bar{M}_0 + \bar{M}_t+ \bar{M}_r) - \alpha_m ({M}_0 + {M}_t+ {M}_r) \\ = \frac{2}{\mu_0} (E_t - E_r- E_0), \end{split}$$ $$\begin{split} \omega_{e0}(\bar{Q}_0 + \bar{Q}_t+ \bar{Q}_r) - \alpha_e ({Q}_0 + {Q}_t+ {Q}_r) \\ = \frac{2}{\epsilon_0\eta_0} (E_0 - E_t - E_r). \end{split}$$ Equations  and , finally represents 14 equations with 14 unknowns, and can be expressed in a matrix form as $$\begin{aligned} \label{Eq:MatrixEquation} &\overbrace{\left[ \begin{array}{ccc} \mathbf{0} & \mathbf{0} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} & \mathbf{0}\\ \mathbf{W_1} & \mathbf{0} & \mathbf{0} \\ \mathbf{0} & \mathbf{T_1} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} & \mathbf{T_1} \\ \end{array} \right]}^{\mathbf{[C]}} % \frac{d[\mathbf{V}]}{dt}\notag\\ % &+ \overbrace{\left[ \begin{array}{ccc} \mathbf{A_1} & \mathbf{A_2} & \mathbf{A_3} \\ \mathbf{B_1} & \mathbf{B_2} & \mathbf{B_3}\\ \mathbf{W_2}(t) & \mathbf{0} & \mathbf{0} \\ \mathbf{0} & \mathbf{T_2}(t) & \mathbf{0} \\ \mathbf{0} & \mathbf{0} & \mathbf{T_2}(t) \end{array} \right]}^{\mathbf{[G(t)]}} % % [\mathbf{V}]=\mathbf{[E(t)]},\end{aligned}$$ where $[\mathbf{V}]$ is the solution vector containing the transmission and reflection field $E_t$ and $E_r$, and $[\mathbf{G(t)}]$ contains the exact description of the metasurface. For a space-time modulated metasurface considered here, the resonant frequencies of the Lorentzian susceptibilities are assumed to be a function of both space and time, i.e. $\omega_{e0}(x,t)$ and $\omega_{m0}(x,t)$. The above matrix equations can be re-written in a complex matrix equation form as $$\mathbf{[C]}\frac{d\mathbf{[V]}}{dt} + \mathbf{[G(t)]} \mathbf{[V]} = \mathbf{[E(t)]}, \label{Eq:TVMetasurface_Equation}$$ which now can be easily solved using standard finite-difference technique based on trapezoidal integration as $$\begin{aligned} \mathbf{[V]}_i = \left(\mathbf{[C]} + \frac{\Delta t}{2}\mathbf{[G]}_i \right )^{-1} &\left[\Delta t\frac{ \mathbf{[E]_i} + \mathbf{[E]_{i-1}}}{2} \right. \notag\\ & \left. + \left(\mathbf{[C]} - \frac{\Delta t}{2} \mathbf{[G]}_i\right) \mathbf{[V]}_{i-1}\right].\notag\end{aligned}$$ [^1]: S. Gupta, T. J. Smy and S. A. Stewart are with the Department of Electronics, Carleton University, Ottawa, Ontario, Canada. Email: shulabh.gupta@carleton.ca [^2]: Assuming no rotation of polarization and only one component of the fields along the principal axis. [^3]: The fields expressions with a tilde, $\tilde{\psi}$ denote the frequency domain quantities. [^4]: for the lossless case, for simplicity. [^5]: assuming that the harmonic amplitudes rapidly falls to zero with increasing index $n$, which is the case in all the forthcoming results. [^6]: This provides a finite bandwidth in the temporal FFTs as opposed to ideal delta functions which are difficult to capture in numerical computations. These Gaussian envelopes should not be seen as the input pulse shapes, but rather as windowing function. [^7]: only for simplicity, and clearer compact description of the associated matrices.

--- abstract: '$\ell_1$-minimization refers to finding the minimum $\ell_1$-norm solution to an underdetermined linear system $\bb=A\xx$. It has recently received much attention, mainly motivated by the new compressive sensing theory that shows under quite general conditions the minimum $\ell_1$-norm solution is also the sparsest solution to the system of linear equations. Although the underlying problem is a linear program, conventional algorithms such as interior-point methods suffer from poor scalability for large-scale real world problems. A number of accelerated algorithms have been recently proposed that take advantage of the special structure of the $\ell_1$-minimization problem. In this paper, we provide a comprehensive review of five representative approaches, namely, *Gradient Projection*, *Homotopy*, *Iterative Shrinkage-Thresholding*, *Proximal Gradient*, and *Augmented Lagrange Multiplier*. The work is intended to fill in a gap in the existing literature to systematically benchmark the performance of these algorithms using a consistent experimental setting. In particular, the paper will focus on a recently proposed face recognition algorithm, where a sparse representation framework has been used to recover human identities from facial images that may be affected by illumination, occlusion, and facial disguise. MATLAB implementations of the algorithms described in this paper have been made publicly available.' author: - | Allen Y. Yang, Arvind Ganesh, Zihan Zhou, S. Shankar Sastry,\ and Yi Ma [^1] bibliography: - 'paper.bib' title: 'A Review of Fast $\ell_1$-Minimization Algorithms for Robust Face Recognition [^2] ' --- Introduction {#sec:introduction} ============ Compressive sensing (CS) has been one of the hot topics in the signal processing and optimization communities in the last five years or so. In CS theory [@CandesE2006-ICM; @DonohoD2004; @BrucksteinA2007; @ChenS2001-SIAM], it has been shown that the minimum $\ell_1$-norm solution to an underdetermined system of linear equations is also the sparsest possible solution under quite general conditions. More specifically, suppose there exists an unknown signal $\xx_0\in{{\mathbb{R}}}^n$, a measurement vector $\bb\in{{\mathbb{R}}}^d$ ($d < n$), and a measurement matrix $A\in{{\mathbb{R}}}^{d \times n}$ such that $A$ is full rank and $\bb= A\xx_0$. Recovering $\xx_0$ given $A$ and $\bb$ constitutes a non-trivial linear inversion problem, since the number of measurements in $\bb$ is smaller than the number of unknowns in $\xx_0$. A conventional solution to this problem is the *linear least squares*, which finds the minimum $\ell_2$-norm solution (or the solution of least energy) to this system. However, if $\xx_0$ is sufficiently sparse[^3] and the sensing matrix $A$ is *incoherent* with the basis under which $\xx_0$ is sparse (i.e., the identity matrix in the standard form), then $\xx_0$ can be exactly recovered by computing the minimum $\ell_1$-norm solution, as given by the following optimization problem: $$(P_1): \quad \min_\xx \, \|\xx\|_1 \subj \bb= A\xx. \label{eq:l1}$$ We refer to the above problem as $\ell_1$-minimization or $\ell_1$-min. The sparsity-seeking property of $(P_1)$ has been shown to have applications in geophysics, data compression, image processing, sensor networks, and more recently, in computer vision. The interested reader is referred to [@CandesE2006-ICM; @BaronD2005; @BrucksteinA2007; @WrightJ2010-PIEEE; @YangA2010-PIEEE] for a comprehensive review of these applications. The $(P_1)$ problem can be recast as a linear program (LP) and can be solved by conventional methods such as interior-point methods. However, the computational complexity of these general-purpose algorithms is often too high for many real-world, large-scale applications. Alternatively, heuristic greedy algorithms have been developed to approximate $(P_1)$. *Orthogonal matching pursuit* (OMP) [@DavisG1997] and *least angle regression* (LARS) [@EfronB2004] are two examples in this category. Empirically, these greedy algorithms work better when $\xx_0$ is very sparse, but will deviate from the solution of $(P_1)$ when the number of non-zero entries in $\xx_0$ increases, as illustrated in [@PlumbleyM2006]. In other words, the greedy methods do not come with strong theoretical guarantees for global convergence. Besides scalability, another important requirement for real-world applications is robustness to noise, namely, the observation vector $\bb$ may be corrupted by data noise. To take into account of the noise, one can relax the equality constraint as follows: $$(P_{1,2}): \quad \min_\xx\,\|\xx\|_1 \subj \|\bb- A\xx\|_2 \le \epsilon, \label{eq:l1-l2}$$ where $\epsilon > 0$ is a pre-determined noise level. If the observation vector $\bb$ is assumed to be corrupted by white noise of magnitude up to $\epsilon$, then the ground-truth signal $\xx_0$ can be well approximated by $(P_{1,2})$, dubbed *basis pursuit denoising* (BPDN) in [@ChenS2001-SIAM; @CandesE2006-CPAM]. Based on the context of the data noise, the $\ell_2$-norm used in the penalty term can also be replaced by other $\ell_p$-norms. In this paper, we also focus on a special case of $(P_1)$ given by: $$(P_{1,1}): \quad \min_{\xx}\,\|\xx\|_1 \subj \|\bb - A\xx\|_1 \le \epsilon. \label{eq:l1-l1}$$ This formulation has been used in [@WrightJ2009-PAMI; @WrightJ2008-IT] for robust face recognition. Although $(P_{1,1})$ can be solved by any algorithm designed for $(P_1)$, as we will later explain, its special structure can be further exploited to develop more scalable methods. In light of the high interest in finding more efficient algorithms to solve these problems, many new algorithms have been proposed. Although it is impossible to summarize all existing algorithms in the literature, in this paper, we provide a comprehensive review of five representative methods, namely, *Gradient Projection* (GP) [@FigueiredoM2007; @KimS2007], *Homotopy* [@OsborneM2000; @MalioutovD2005; @DonohoD2006], *Iterative Shrinkage-Thresholding* (IST) [@DaubechiesI2004; @CombettesP2005; @HaleE2007; @WrightS2008], *Proximal Gradient* (PG) [@NesterovY1983-SMD; @NesterovY2007; @BeckA2009; @BeckerS2009], and *Augmented Lagrange Multiplier* (ALM) [@Bertsekas2003; @YangJ2009]. Although the paper is mainly focused on fast implementations of $\ell_1$-min, the reader may refer to [@BrucksteinA2007; @TroppJ2010] for a broader discussion about recovering sparse solutions via other approaches, such as greedy pursuit-type algorithms. This paper intends to fill in a gap in the existing literature to systematically benchmark the performance of these algorithms using a fair and consistent experimental setting. Due to the attention given to compressive sensing and $\ell_1$-minimization, other more sophisticated solutions continue to be developed at a rapid pace. More recent developments include *subspace pursuit* [@DaiW2009], CoSaMP [@NeedellD2008], *approximate message-passing* algorithm [@DonohoD2009-PNAS], and *Bregman iterative algorithm* [@YinW2008], to name a few. However, we do not believe there exists an overall winner that could achieve the best performance in terms of both speed and accuracy for all applications. Therefore, in addition to extensive simulations using synthetic data, our experiments will be focused on a specific application of robust face recognition proposed in [@WrightJ2009-PAMI], where a sparse representation framework has recently been developed to recognize human identities from facial images. To aid peer evaluation, all algorithms discussed in this paper have been made available on our website as a MATLAB toolbox: <http://www.eecs.berkeley.edu/~yang/software/l1benchmark/> Notation -------- For a vector $\xx\in{{\mathbb{R}}}^n$, we denote by $\xx_+$ and $\xx_-$ the vectors that collect the positive and negative coefficients of $\xx$, respectively: $$\xx = \xx_+ - \xx_-, \xx_+\ge 0, \xx_-\ge 0.$$ We also denote $$X = \mbox{diag}(x_1, x_2, \cdots, x_n)\in{{\mathbb{R}}}^{n\times n}$$ as a square matrix with the coefficients of $\xx$ as its diagonal and zero otherwise. The concatenation of two (column) vectors will be written following the MATLAB convention: $[\xx_1; \xx_2]\doteq \left[\begin{smallmatrix}\xx_1\\ \xx_2 \end{smallmatrix}\right]$; $[\xx_1, \xx_2]\doteq \left[\begin{smallmatrix}\xx_1 & \xx_2 \end{smallmatrix}\right]$. We denote by $\one$ a vector whose components are all one with dimension defined within the context. We represent the Euclidean or $\ell_2$-norm by $\|\cdot\|_2$ and the $\ell_1$-norm by $\|\cdot\|_1$. The notation $\|\cdot \|$ represents the $\ell_2$-norm for vectors and the spectral norm for matrices. For any real-valued differentiable function $f(\cdot)$, we denote its gradient by $\nabla f(\cdot)$. If a real-valued convex function $g(\cdot)$ is not differentiable everywhere, we represent its subgradient by $\partial g(\cdot)$, defined as follows: $$\partial g(\xx) = \{\eta\in{{\mathbb{R}}}^n: g(\bar\xx) - g(\xx) \ge \eta^T (\bar\xx - \xx), \forall \bar\xx\in{{\mathbb{R}}}^n \}.$$ Primal-Dual Interior-Point Methods {#sec:BP} ---------------------------------- We first discuss a classical solution to the $\ell_1$-min problem $(P_1)$, called the *primal-dual interior-point* method, which is usually attributed to the works of [@FrischK1955; @KarmarkarN1984; @MegiddoN1989; @MonteiroR1989-I; @KojimaM1993]. For the sake of simplicity, we assume here that the sparse solution $\xx$ is nonnegative.[^4] Under this assumption, it is easy to see that $(P_1)$ can be converted to the standard primal and dual forms in linear programming: $$\begin{array}{rllrl} & \mbox{\bf Primal (P)}& & & \mbox{\bf Dual (D)} \\ \min_\xx& \cc^T\xx & & \max_{\yy,\zz} & \bb^T\yy\\ \subj & A\xx = \bb & & \subj & A^T\yy + \zz = \cc\\ & \xx \ge 0 & & & \zz\ge 0, \end{array}$$ where for $\ell_1$-min, $\cc = \one$. The primal-dual algorithm simultaneously solves for the primal and dual optimal solutions [@BoydS2004]. It was proposed in [@FrischK1955] that $(P)$ can be converted to a family of logarithmic barrier problems[^5]: $$\begin{array}{lrl} (P_\mu): & \min_\xx & \cc^T\xx - \mu\sum_{i=1}^{n} \log x_i.\\ & \subj & A\xx = \bb, \xx > 0 \end{array}$$ Clearly, a feasible solution $\xx$ to $(P_\mu)$ cannot have zero coefficients. Therefore, we define the interiors of the solution domains for $(P)$ and $(D)$ as: $$\begin{array}{rcl} P_{++} &=& \{ \xx: A\xx = \bb, \xx > 0\},\\ D_{++} &=& \{(\yy, \zz): A^T\yy+ \zz = \cc, \zz>0 \}, \\ S_{++} &=& P_{++} \times D_{++}. \end{array}$$ Assuming that the above sets are non-empty, it can be shown that $(P_\mu)$ has a unique global optimal solution $\xx(\mu)$ for all $\mu>0$. As $\mu \rightarrow 0$, $\xx(\mu)$ and $(\yy(\mu), \zz(\mu))$ converge to optimal solutions of problems $(P)$ and $(D)$ respectively [@MegiddoN1989; @MonteiroR1989-I]. The primal-dual interior-point algorithm seeks the domain of the *central trajectory* for the problems $(P)$ and $(D)$ in $S_{++}$, where the central trajectory is defined as the set $S=\{(\xx(\mu), \yy(\mu), \zz(\mu)): \mu>0\}$ of solutions to the following system of equations: $$\begin{array}{c} XZ\one=\mu\one, \, A\xx = \bb, A^T\yy + \zz = \cc, \\ \xx\ge 0, \, \zz\ge 0. \end{array} \label{eq:central-trajectory}$$ The above condition is also known as the Karush-Kuhn-Tucker (KKT) conditions for the convex program $(P_\mu)$ [@MonteiroR1989-I; @KojimaM1993]. Hence, the update rule on the current value $(\xx^{(k)}, \yy^{(k)}, \zz^{(k)})$ is defined by the Newton direction $(\Delta\xx, \Delta\yy, \Delta\zz)$, which is computed as the solution to the following set of linear equations: $$\begin{matrix} Z^{(k)}\Delta\xx + X^{(k)}\Delta\zz &=& \hat{\mu}\one - X^{(k)}\zz^{(k)},\\ A \Delta\xx &=& 0,\\ A^T\Delta\yy + \Delta\zz &=& 0, \end{matrix} \label{eq:logbarrier-update}$$ where $\hat{\mu}$ is a penalty parameter that is generally different from $\mu$ in $(P_\mu)$. In addition to the update rule , an algorithm also needs to specify a stopping criterion when the solution is close to the optimum. For $\ell_1$-min, some simple rules can be easily evaluated: 1. The relative change of the sparse support set becomes small; 2. The relative change (in the sense of the $\ell_2$-norm) of the update of the estimate becomes small; 3. The relative change of the objective function becomes small. A more detailed discussion about choosing good stopping criteria in different applications is postponed to Section \[sec:simulation\]. Algorithm \[alg:interior-point\] summarizes a conceptual implementation of the interior-point methods.[^6] For more details about how to choose the initial values $(\xx^{(0)}, \yy^{(0)}, \zz^{(0)})$ and the penalty parameter $\hat{\mu}$, the reader is referred to [@KojimaM1993; @MonteiroR1989-I]. Algorithm \[alg:interior-point\] requires a total of $O(\sqrt{n})$ iterations, and each iteration can be executed in $O(n^3)$ operations for solving the linear system . [**Input:**]{} A full rank matrix $A\in{{\mathbb{R}}}^{d\times n}$, $d<n$, a vector $\bb\in{{\mathbb{R}}}^d$, initialization $(\xx^{(0)}, \yy^{(0)}, \zz^{(0)})$. Iteration $k\leftarrow 0$. Initial penalty $\mu$ and a decreasing factor $0<\delta<\sqrt{n}$. $k\leftarrow k+1$, $\mu\leftarrow \mu(1-\delta/\sqrt{n})$. Solve for $(\Delta\xx, \Delta\yy, \Delta\zz)$. $ \xx^{(k)} \leftarrow \xx^{(k-1)} + \Delta\xx, \yy^{(k)} \leftarrow \yy^{(k-1)} + \Delta\yy, \zz^{(k)} \leftarrow \zz^{(k-1)} + \Delta\zz. $ [**Output:**]{} $\xx^*\leftarrow \xx^{(k)}$. In one simulation shown in Figure \[fig:BP-comparison\], the computational complexity of Algorithm \[alg:interior-point\] with respect to (w.r.t.) the sensing dimension $d$ grows much faster than the other five algorithms in comparison. For example, at $d=1900$, the fastest algorithm in this simulation, i.e., DALM, only takes about 0.08 sec to complete one trial, which is more than 250 times faster than PDIPA. For this reason, basic BP algorithms should be used with caution in solving real-world applications. The details of the simulation are explained later in Section \[sec:simulation\]. ![Average run time of PDIPA in comparison with five other fast algorithms. The simulation setup: $n=2000$, $k=200$. The projection matrices are randomly generated based on the standard normal distribution with the dimension varies from 300 to 1900. The support of the ground truth $\xx_0$ is randomly selected at each trial, and the nonzero coefficients are sampled from the normal distribution.[]{data-label="fig:BP-comparison"}](varying-dimension-speed.png){width="2.4in"} Next, we will review the five fast $\ell_1$-min algorithms shown in Figure \[fig:BP-comparison\], namely, *Gradient Projection* in Section \[sec:GP\], *Homotopy* in Section \[sec:homotopy\], *Iterative Shrinkage-Thresholding* in Section \[sec:thresholding\], *Proximal Gradient* in Section \[sec:Nesterov\], and *Augmented Lagrange Multiplier* in Section \[sec:ADM\]. The algorithms of L1LS, Homotopy, and SpaRSA are provided by their respective authors. We have also provided our implementation of FISTA and DALM on our website. Fast $\ell_1$-Min Algorithms ============================ Gradient Projection Methods {#sec:GP} --------------------------- We first discuss *Gradient Projection* (GP) methods that seek a sparse representation $\xx$ along a certain gradient direction, which induces much faster convergence speed. The approach reformulates the $\ell_1$-min problem as a quadratic programming (QP) problem. We start with the $\ell_1$-min problem $(P_{1,2})$. It is equivalent to the so-called LASSO problem [@TibshiraniR1996]: $$(LASSO):\quad \min_{\xx}\|\bb-A\xx\|^2_2 \subj \|\xx\|_1\le \sigma, \label{eq:LASSO}$$ where $\sigma > 0$ is an appropriately chosen constant. Using a Lagrangian formulation, the problem $(LASSO)$ (and hence, $(P_{1,2})$) can be rewritten as an unconstrained optimization problem: $$\xx^* = \arg\min_{\xx} F(\xx) = \arg\min_{\xx}\frac{1}{2}\|\bb - A\xx\|_2^2 + \lambda \|\xx\|_1, \label{eq:Lagrangian}$$ where $\lambda$ is the Lagrangian multiplier. In the literature, there exist two slightly different methods that formulate as a quadratic programming problem, namely, *gradient projection sparse representation* (GPSR) [@FigueiredoM2007] and *truncated Newton interior-point method* (TNIPM) [@KimS2007].[^7] To formulate the GPSR algorithm, we separate the positive coefficients $\xx_+$ and the negative coefficients $\xx_-$ in $\xx$, and rewrite as $$\begin{array}{l} \min_\xx \, Q(\xx)=\frac{1}{2}\|\bb - [A, -A] [\xx_+; \xx_-]\|_2^2 + \lambda \one^T(\xx_+ +\xx_-)\\ \subj \: \xx_+\ge 0, \xx_-\ge 0. \end{array} \label{eq:bound-constrainted-quadratic}$$ Problem can be rewritten in the standard QP form as $$\begin{array}{rl} \min & Q(\zz) \doteq \cc^T\zz + \frac{1}{2}\zz^TB\zz\\ \subj & \zz\ge 0, \end{array} \label{eq:standard-quadratic}$$ where $\zz = [\xx_+;\xx_-]$, $\cc=\lambda\one + [-A^T\bb; A^T\bb]$, and $$B = \left[\begin{matrix}A^TA & -A^TA\\ -A^TA & A^TA \end{matrix}\right].$$ We note that the gradient of $Q(\zz)$ is defined as $$\nabla_\zz Q(\zz) = \cc + B\zz.$$ This leads to a steepest-descent algorithm that searches from each iterate $\zz^{(k)}$ along the negative gradient $-\nabla Q(\zz)$: $$\zz^{(k+1)} = \zz^{(k)} - \alpha^{(k)}\nabla Q(\zz^{(k)}), \label{eq:GPSR-update}$$ where $\alpha^{(k)}$ is the step size. This can be solved by a standard *line-search* process [@Hiriart-Urruty1996]. For example, in [@FigueiredoM2007], the direction vector $\g^{(k)}$ is defined as $$g_i^{(k)} = \left\{\begin{matrix} (\nabla Q(\zz^{(k)}))_i, &\mbox{if } z_i^{(k)} > 0 \mbox{ or } (\nabla Q(\zz^{(k)}))_i<0\\ 0, & \mbox{otherwise.}\end{matrix}\right.$$ Then the step size for the update is chosen to be $$\alpha^{(k)} = \arg\min_\alpha Q(\zz^{(k)} - \alpha \g^{(k)}),$$ which has a closed-form solution $$\alpha^{(k)} = \frac{(\g^{(k)})^T\g^{(k)}}{(\g^{(k)})^TB\g^{(k)}}.$$ The computational complexity and convergence of GPSR is difficult to estimate exactly [@FigueiredoM2007]. Another issue is that the formulation of doubles the dimension of the equations from . Therefore, the matrix operations involving $B$ must take into account its special structure w.r.t. $A$ and $A^T$. The second GP algorithm, which we will benchmark in Section \[sec:simulation\], is *truncated Newton interior-point method* (TNIPM) [@KimS2007]. It transforms the same objective function to a quadratic program but with inequality constraints: $$\begin{array}{rl} \min & \frac{1}{2}\|A\xx - \bb\|_2^2 + \lambda \sum_{i=1}^n u_i.\\ \subj & -u_i\le x_i \le u_i, \quad i=1,\cdots,n \end{array}$$ Then a *logarithmic barrier* function for the constraints $-u_i\le x_i \le u_i$ is constructed as follows [@FrischK1955]: $$\Phi(\xx, \uu) = - \sum_i\log(u_i + x_i) - \sum_i\log(u_i - x_i).$$ Over the domain of $(\xx,\uu)$, the central path consists of the unique minimizer $(\xx^*(t), \uu^*(t))$ of the convex function $$F_t(\xx,\uu) = t ( \|A\xx - \bb\|_2^2 + \lambda\sum_{i=1}^n u_i ) + \Phi(\xx, \uu),$$ where the parameter $t\in [0, \infty)$. Using the primal barrier method discussed in Section \[sec:BP\], the optimal search direction using Newton’s method is computed by $$\nabla^2 F_t(\xx, \uu) \cdot \left[\begin{matrix}\Delta\xx \\ \Delta\uu \end{matrix}\right] = -\nabla F_t(\xx, \uu) \in {{\mathbb{R}}}^{2n}. \label{eq:Newton-method}$$ Again, for large-scale problems, directly solving is computationally expensive. In [@KimS2007], the search step is accelerated by a *preconditioned conjugate gradients* (PCG) algorithm, where an efficient preconditioner is proposed to approximate the Hessian of $\frac{1}{2}\|A\xx - \bb\|_2^2$. The reader is referred to [@KelleyC1995; @NocedalJ2006] for more details about PCG. Homotopy Methods {#sec:homotopy} ---------------- One of the drawbacks of the PDIPA method is that they require the solution sequence $\xx(\mu)$ to be close to a “central path” as $\mu\rightarrow 0$, which sometimes is difficult to satisfy and computationally expensive in practice. In this section, we review an approach called *Homotopy methods* [@OsborneM2000; @MalioutovD2005; @DonohoD2006] that can mitigate these issues. We recall that $(P_{1,2})$ can be written as an unconstrained convex optimization problem: $$\begin{array}{rcl} \xx^* = \arg\min_{\xx} F(\xx) &=& \arg\min_{\xx}\frac{1}{2}\|\bb - A\xx\|_2^2 + \lambda \|\xx\|_1,\\ & \doteq & \arg\min_{\xx}f(\xx) + \lambda g(\xx) \end{array} \label{eq:Lagrangian-1}$$ where $f(\xx) = \frac{1}{2}\|\bb-A\xx\|_2^2$, $g(\xx) = \|\xx\|_1$, and $\lambda > 0$ is the Lagrange multiplier. On one hand, w.r.t. a fixed $\lambda$, the optimal solution is achieved when $\zero\in\partial F(\xx)$. On the other hand, similar to the interior-point algorithm, if we define $$\X\doteq \{\xx^*_\lambda: \lambda \in [0, \infty)\},$$ $\X$ identifies a solution path that follows the change in $\lambda$: when $\lambda\rightarrow \infty$, $\xx^*_\lambda = 0$; when $\lambda\rightarrow 0$, $\xx^*_\lambda$ converges to the solution of $(P_1)$. The Homotopy methods exploit the fact that the objective function $F(\xx)$ undergoes a homotopy from the $\ell_2$ constraint to the $\ell_1$ objective in as $\lambda$ decreases. One can further show that the solution path $\X$ is piece-wise constant as a function of $\lambda$ [@OsborneM2000; @EfronB2004; @DonohoD2006]. Therefore, in constructing a decreasing sequence of $\lambda$, it is only necessary to identify those “breakpoints” that lead to changes of the support set of $\xx_\lambda^*$, namely, either a new nonzero coefficient added or a previous nonzero coefficient removed. The algorithm operates in an iterative fashion with an initial value $\xx^{(0)}=0$. In each iteration, given a nonzero $\lambda$, we solve for $\xx$ satisfying $\partial F(\xx)=0$. The first summand $f$ in is differentiable: $\nabla f = A^T(A\xx-\bb)\doteq -\cc(\xx)$. The subgradient of $g(\xx)=\|\xx\|_1$ is given by: $$\uu(\xx) \doteq \partial\|\xx\|_1 = \left\{\uu\in{{\mathbb{R}}}^n : \begin{matrix}u_i=\mbox{sgn}(x_i), x_i\not=0\\ u_i\in [-1, 1], x_i=0 \end{matrix}\right\}. \label{eq:l1-subdifferential}$$ Thus, the solution to $\partial F(\xx)=0$ is also the solution to the following equation: $$\cc(\xx)=A^T\bb - A^TA\xx=\lambda \uu(\xx). \label{eq:homotopy-condition}$$ By the definition , the sparse support set at each iteration is given by $${\mathcal{I}}\doteq \{i: |\cc^{(l)}_i|=\lambda \}.$$ The algorithm then computes the update for $\xx^{(k)}$ in terms of the direction and the magnitude separately. Specifically, the update direction on the sparse support $\dd^{(k)}({\mathcal{I}})$ is the solution to the following system: $$A^T_{\mathcal{I}}A_{\mathcal{I}}\dd^{(k)}({\mathcal{I}}) = \mbox{sgn}(\cc^{(k)}({\mathcal{I}})), \label{eq:homotopy-LSE}$$ where $A_{\mathcal{I}}$ is a submatrix of $A$ that collects the column vectors of A w.r.t. ${\mathcal{I}}$, and $\cc^{(k)}({\mathcal{I}})$ is a vector that contains the coefficients of $\cc^{(k)}$ w.r.t. ${\mathcal{I}}$. For the coefficients whose indices are not in ${\mathcal{I}}$, their update directions are manually set to zero. Along the direction indicated by $\dd^{(k)}$, there are two scenarios when an update on $\xx$ may lead to a breakpoint where the condition is violated. The first scenario occurs when an element of $\cc$ not in the support set would increase in magnitude beyond $\lambda$: $$\gamma^+ = \min_{i\not\in{\mathcal{I}}}\left\{\frac{\lambda-c_i}{1-\aa_i^TA_{{\mathcal{I}}}\dd^{(k)}({\mathcal{I}})}, \frac{\lambda+c_i}{1+\aa_i^TA_{{\mathcal{I}}}\dd^{(k)}({\mathcal{I}})} \right\}. \label{eq:homotopy-add}$$ The index that achieves $\gamma^+$ is denoted as $i^+$. The second scenario occurs when an element of $\cc$ in the support set ${\mathcal{I}}$ crosses zero, violating the sign agreement: $$\gamma^- = \min_{i\in{\mathcal{I}}}\{-x_i/d_i \}. \label{eq:homotopy-remove}$$ The index that achieves $\gamma^-$ is denoted as $i^-$. Hence, the homotopy algorithm marches to the next breakpoint, and updates the sparse support set by either appending ${\mathcal{I}}$ with $i^+$ or removing $i^-$: $$\xx^{(k+1)} = \xx^{(k)} + \min\{\gamma^+, \gamma^- \}\dd^{(k)}. \label{eq:homotopy-update}$$ The algorithm terminates when the relative change in $\xx$ between consecutive iterations is sufficiently small. Algorithm \[alg:homotopy\] summarizes an implementation of the Homotopy methods.[^8] [**Input:**]{} A full rank matrix $A=[\vv_1, \cdots, \vv_n]\in{{\mathbb{R}}}^{d\times n}$, $d<n$, a vector $\bb\in{{\mathbb{R}}}^d$, initial Lagrangian parameter $\lambda = 2\|A^T\bb \|_{\infty}$. $k\leftarrow 0$. Find the first support index: $i=\arg\max_{j=1}^{n}\|\vv_j^T\bb \|$, ${\mathcal{I}}= \{i \}$. $k\leftarrow k+1$. Solve for the update direction $\dd^{(k)}$ in . Compute the sparse support updates and : $\gamma^* \leftarrow \min\{\gamma^+, \gamma^- \}$. Update $\xx^{(k)}$, ${\mathcal{I}}$, and $\lambda\leftarrow \lambda - \gamma^*$. [**Output:**]{} $\xx^* \leftarrow \xx^{(k)}$. Overall, solving using a Cholesky factorization and the addition/removal of the sparse support elements dominate the computation. Since one can keep track of the rank-1 update of $A^T_{\mathcal{I}}A_{\mathcal{I}}$ in solving using $O(d^2)$ operations in each iteration, the computational complexity of the homotopy algorithm is $O(kd^2 + kdn)$. Finally, we want to point out that Homotopy has been shown to share some connections with two greedy $\ell_1$-min approximations, namely, *least angle regression* (LARS) [@EfronB2004] and *polytope faces pursuit* (PFP) [@PlumbleyM2006]. For instance, if the ground-truth signal $\xx_0$ has at most $k$ non-zero components with $k << n$, all three algorithms can recover it in $k$ iterations. On the other hand, LARS never removes indices from the sparse support set during the iteration, while Homotopy and PFP have mechanisms to remove coefficients from the sparse support. More importantly, Homotopy provably solves $\ell_1$-min $(P_1)$, while LARS and PFP are only approximate solutions. A more detailed comparison between Homotopy, LARS, and PFP can be found in [@DonohoD2006]. Iterative Shrinkage-Thresholding Methods {#sec:thresholding} ---------------------------------------- Although Homotopy employs a more efficient iterative update rule that only involves operations on those submatrices of $A$ corresponding to the support sets of $\xx$, it may not be as efficient when the sparsity $k$ and the observation dimension $d$ grow proportionally with the signal dimension $n$. In such scenarios, one can show that the worst-case computational complexity is still bounded by $O(n^3)$. In this section, we discuss *Iterative Shrinkage-Thresholding* (IST) methods [@DaubechiesI2004; @CombettesP2005; @HaleE2007; @WrightS2008], whose implementation mainly involves simple operations such as vector algebra and matrix-vector multiplications. This is in contrast to most past methods that all involve expensive operations such as matrix factorization and solving linear least squares problems. A short survey on the applications of IST can be found in [@YinW2008]. In a nutshell, IST considers solving $(P_{1,2})$ as a special case of the following *composite objective function*: $$\min_x F(\xx)\doteq f(\xx) + \lambda g(\xx), \label{eq:IST}$$ where $f:{{\mathbb{R}}}^n\rightarrow {{\mathbb{R}}}$ is a smooth and convex function, and $g:{{\mathbb{R}}}^n\rightarrow {{\mathbb{R}}}$ as the regularization term is bounded from below but not necessarily smooth nor convex. For $\ell_1$-min in particular, $g$ is also separable, that is, $$g(\xx) = \sum_{i=1}^n g_i(x_i).$$ Clearly, let $f(\xx) = \frac{1}{2}\|\bb - A\xx\|_2^2$ and $g(\xx)=\|\xx \|_1$. Then the objective function becomes the unconstrained BPDN problem. The update rule to minimize is computed using a second-order approximation of $f$ [@WrightS2008; @BeckA2009]: $$\begin{array}{rcl} \xx^{(k+1)} & = & \arg\min_{\xx}\{f(\xx^{(k)}) + (\xx - \xx^{(k)})^T\nabla f(\xx^{(k)}) \\ & & + \frac{1}{2}\|\xx - \xx^{(k)} \|_2^2\cdot\nabla^2f(\xx^{(k)}) + \lambda g(\xx) \}\\ & \approx & \arg\min_{\xx}\{(\xx - \xx^{(k)})^T\nabla f(\xx^{(k)}) \\ & & + \frac{\alpha^{(k)}}{2}\|\xx - \xx^{(k)} \|_2^2 + \lambda g(\xx) \}\\ & = & \arg\min_{\xx}\{\frac{1}{2}\|\xx - \uu^{(k)} \|_2^2 + \frac{\lambda}{\alpha^{(k)}}g(\xx)\},\\ & \doteq & G_{\alpha^{(k)}}(\xx^{(k)}), \end{array} \label{eq:IST-objective}$$ where $$\uu^{(k)} = \xx^{(k)} - \frac{1}{\alpha^{(k)}}\nabla f(\xx^{(k)}).$$ In , the Hessian $\nabla^2 f(\xx^{(k)})$ is approximated by a diagonal matrix $\alpha^{(k)} I$. If we replace $g(\xx)$ in by the $\ell_1$-norm $\|\xx\|_1$, which is a separable function, then $G_{\alpha^{(k)}}(\xx^{(k)})$ has a closed-form solution w.r.t. each component: $$\begin{array}{lll} x_i^{(k+1)} &= &\arg\min_{\xx_i}\left\{ \frac{(x_i - u_i^{(k)})^2}{2} + \frac{\lambda |x_i|}{\alpha^{(k)}} \right\}\\ &=& \mbox{soft}\left(u_i^{(k)}, \frac{\lambda}{\alpha^{(k)}} \right), \end{array} \label{eq:soft-thresholding}$$ where $$\begin{array}{rcl} \mbox{soft}(u, a) &\doteq& \mbox{sgn}(u)\max\{|u|-a, 0 \} \\ &= & \left\{\begin{matrix}\mbox{sgn}(u)(|u| - a) &\mbox{if } |u|>a \\ 0 & \mbox{otherwise}\end{matrix}\right. \end{array}$$ is the *soft-thresholding* or *shrinkage* function [@DonohoD1995]. There are two free parameters in , namely, the regularizing coefficient $\lambda$ and the coefficient $\alpha^{(k)}$ that approximates the hessian matrix $\nabla^2 f$. Different strategies for choosing these parameters have been proposed. Since $\alpha I$ mimics the Hessian $\nabla^2 f$, we require that $\alpha^{(k)}(\xx^{(k)} - \xx^{(k-1)})\approx \nabla f(\xx^{(k)}) - \nabla f(\xx^{(k-1)})$ in the least-squares sense. Hence, $$\begin{array}{rcl} \alpha^{(k+1)} &=& \arg\min_{\alpha} \|\alpha (\xx^{(k)} - \xx^{(k-1)}) \\ & & - \left(\nabla f(\xx^{(k)}) - \nabla f(\xx^{(k-1)})\right) \|_2^2\\ & =& \frac{\left(\xx^{(k)} - \xx^{(k-1)}\right)^T \left(\nabla f(\xx^{(k)}) - \nabla f(\xx^{(k-1)}\right)}{\left(\xx^{(k)} - \xx^{(k-1)}\right)^T \left(\xx^{(k)} - \xx^{(k-1)}\right)}. \end{array} \label{eq:soft-thresholding-update-alpha}$$ This is known as the *Barzilai-Borwein* equation [@BarzilaiJ1988; @SerafiniT2004; @WrightS2008]. For choosing $\lambda$, instead of using a fixed value, several papers have proposed a *continuation* strategy [@HaleE2007; @FigueiredoM2007], in which is solved for a decreasing sequence of $\lambda$. As mentioned in Section \[sec:homotopy\], recovers the optimal $\ell_1$-min solution when $\lambda \rightarrow 0$. However, it has been observed that in practice the performance degrades by directly solving for small values of $\lambda$, which has been dubbed as a “cold” starting point. Instead, *continuation* employs a warm-starting strategy by first solving for a larger value of $\lambda$, then decreasing $\lambda$ in steps towards its desired value. The IST algorithm is summarized in Algorithm \[alg:thresholding\].[^9] The interested reader is referred to [@MalekiA2009] for guidelines on choosing good soft-thresholding parameters to improve accuracy. [**Input:**]{} A full rank matrix $A\in{{\mathbb{R}}}^{d\times n}$, $d<n$, a vector $\bb\in{{\mathbb{R}}}^d$, Lagrangian $\lambda_0$, initial values for $\xx^{(0)}$ and $\alpha^{0}$, $k\leftarrow 0$. Generate a reducing sequence $\lambda_0>\lambda_1>\cdots>\lambda_N$. $\lambda \leftarrow \lambda_i$ $k \leftarrow k+1$. $\xx^{(k)} \leftarrow G(\xx^{(k-1)})$. Update $\alpha^{(k)}$ using . [**Output:**]{} $\xx^* \leftarrow \xx^{(k)}$. Proximal Gradient Methods {#sec:Nesterov} ------------------------- Proximal Gradient (PG) algorithms represent another class of algorithms that solve convex optimization problems of the form . Assume that our cost function $F(\cdot)$ can be decomposed as the sum of two functions $f$ and $g$, where $f$ is a smooth convex function with Lipschitz continuous gradient, and $g$ is a continuous convex function. The principle behind these algorithms is to iteratively form quadratic approximations $Q(\xx,\yy)$ to $F(\xx)$ around a carefully chosen point $\yy$, and to minimize $Q(\xx,\yy)$ rather than the original cost function $F$. For our problem, we define $g(\xx) = \|\xx\|_1$ and $f(\xx) = \frac{1}{2}\|A\xx-\bb\|_2^2$. We note that $\nabla f(\xx) = A^T(A\xx - \bb)$ is Lipschitz continuous with Lipschitz constant $L_f \doteq \|A\|^2$. Define $Q(\xx,\yy)$ as: $$Q(\xx,\yy)\doteq f(\yy) + \langle \nabla f(\yy), \xx-\yy \rangle + \frac{L_f}{2}\,\|\xx-\yy\|^2 + \lambda\,g(\xx).$$ It can be shown that $F(\xx) \leq Q(\xx,\yy)$ for all $\yy$, and $$\arg\min_\xx\: Q(\xx,\yy) = \arg\min_\xx\left\{ \lambda g(\xx) + \frac{L_f}{2} \left \| \xx - \uu \right\|^2 \right\}, \label{eq:argmin-Q}$$ where $\uu = \yy - \frac{1}{L_f}\nabla f(\yy)$ by the same trick used in . For the $\ell_1$-min problem, has a closed-form solution given by the soft-thresholding function: $$\arg\min_\xx\: Q(\xx,\yy) = \mathrm{soft}\left(\uu,\frac{\lambda}{L_f}\right).$$ However, unlike the iterative thresholding algorithm described earlier, we use a smoothed computation of the sequence $\yy_k$. It has been shown that choosing $$\yy^{(k)} = \xx^{(k)} + \frac{t_{k-1}-1}{t_k}\left( \xx^{(k)} - \xx^{(k-1)}\right),$$ where $\{t_k\}$ is a positive real sequence satisfying $t_k^2 - t_k \leq t_{k-1}^2$, achieves an accelerated non-asymptotic convergence rate of $O(k^{-2})$ [@NesterovY1983-SMD; @NesterovY2007; @BeckA2009]. To further accelerate the convergence of the algorithm, one can also make use of the continuation technique described in Section \[sec:thresholding\]. Finally, for large problems, it is often computationally expensive to directly compute $L_f = \|A\|^2$.[^10] A *backtracking* line-search strategy [@BeckA2009] can be used to generate a scalar sequence $\{L_k\}$ that approximates $L_f$. We define $$Q_L(\xx, \yy) \doteq f(\yy) + (\xx - \yy)^T\nabla f(\yy) + \frac{L}{2}\| \xx - \yy\|^2 + \lambda\,g(\xx).$$ Suppose that $\eta > 1$ is a pre-defined constant. Then, given $\yy^{(k)}$ at the $k$th iteration, we set $L_k = \eta^j\,L_{k-1}$, where $j$ is the smallest nonnegative integer such that the following inequality holds: $$F(G_{L_k}(\yy^{(k)})) \le Q_{L_k}(G_{L_k}(\yy^{(k)}), \yy^{(k)}), \label{eq:backtracking}$$ where $G_L(\yy) \doteq \arg\min_\xx \: Q_L(\xx,\yy) = \mathrm{soft}\left(\uu,\frac{\lambda}{L}\right)$ for $\uu \doteq \yy - \frac{1}{L}\nabla f(\yy)$. The algorithm, dubbed FISTA in [@BeckA2009], is summarized as Algorithm \[alg:FISTA\].[^11] The convergence behavior of FISTA is given by $$F(\xx^{(k)}) - F(\xx^*) \le \frac{2L_f\|\xx^{(0)} - \xx^* \|^2}{(k+1)^2}, \quad \forall \, k.$$ The interested reader may refer to [@NesterovY2007; @BeckA2009; @BeckerS2009] for a proof of the above result. $\bb \in {{\mathbb{R}}}^m$, $A \in {{\mathbb{R}}}^{m \times n}$. Set $\xx^{(0)} \leftarrow 0$, $\xx^{(1)} \leftarrow 0$, $t_0 \leftarrow 1$, $t_1 \leftarrow 1$, $k \leftarrow 1$. Initialize $L_0$, $\lambda_1$, $\beta \in (0,1)$, $\bar{\lambda} > 0$. $\yy^{(k)} \leftarrow \xx^{(k)} + \frac{t_{k-1}-1}{t_k}\,\left( \xx^{(k)} - \xx^{(k-1)}\right)$. Update $L_k$ using with $\yy^{(k)}$. $\uu^{(k)} \leftarrow \yy^{(k)} - \frac{1}{L_k}\,\, A^T(A\yy^{(k)} - \bb)$. $\xx^{(k+1)} \leftarrow \mathrm{soft}\left(\uu^{(k)}, \frac{\lambda_k}{L_k}\right)$. $t_{k+1} \leftarrow \frac{1+\sqrt{4t_k^2+1}}{2}$. $\lambda_{k+1} \leftarrow \max(\beta\,\lambda_k, \bar{\lambda})$. $k \leftarrow k+1$. $\xx^* \leftarrow \xx^{(k)}$. \[alg:fista\] Augmented Lagrange Multiplier Methods {#sec:ADM} ------------------------------------- Lagrange multiplier methods are a popular class of algorithms in convex programming. The basic idea is to eliminate equality constraints and instead add a penalty term to the cost function that assigns a very high cost to infeasible points. Augmented Lagrange Multiplier (ALM) methods differ from other penalty-based approaches by simultaneously estimating the optimal solution and Lagrange multipliers in an iterative fashion. We consider the general $\ell_1$-min problem with the optimal solution $\xx^*$. The corresponding augmented Lagrange function is given by $$L_\mu (\xx,\yy) = \|\xx\|_1 + \langle \yy, \bb-A\xx \rangle + \frac{\mu}{2} \|\bb - A\xx\|_2^2,$$ where $\mu > 0$ is a constant that determines the penalty for infeasibility, and $\yy$ is a vector of Lagrange multipliers. Let $\yy^*$ be a Lagrange multiplier vector satisfying the second-order sufficiency conditions for optimality (see section 3.2 in [@Bertsekas2003] for more details). Then, for sufficiently large $\mu$, it can be shown that $$\xx^* = \arg \min_\xx \: L_\mu (\xx,\yy^*).$$ The main problem with this solution is that $\yy^*$ is unknown in general. Furthermore, the choice of $\mu$ is not straightforward from the above formulation. It is clear that to compute $\xx^*$ by minimizing $L_\mu(\xx,\yy)$, we must choose $\yy$ close to $\yy^*$ and set $\mu$ to be a very large positive constant. The following iterative procedure has been proposed in [@Bertsekas2003] to simultaneously compute $\yy^*$ and $\xx^*$: $$\left \{ \begin{array}{lll} \xx_{k+1} & = & \arg\min_{\xx} \, L_{\mu_k} (\xx,\yy_k)\\ \yy_{k+1} & = & \yy_k + \mu_k (\bb - A\xx_{k+1}) \\ \end{array} \right . , \label{eq:alm}$$ where $\left\{\mu_{k}\right\}$ is a monotonically increasing positive sequence. We note that the first step in the above procedure is itself an unconstrained convex optimization problem. Thus, the above iterative procedure is computationally efficient only if it is easier to minimize the augmented Lagrangian function compared to solving the the original constrained optimization problem directly. We focus our attention on solving the first step of for the $\ell_1$-min problem. Although it is not possible to obtain a closed-form solution, the cost function has the same form as described in . Furthermore, the quadratic penalty term is smooth and has a Lipschitz continuous gradient. Therefore, we can solve it efficiently using proximal gradient methods (e.g., FISTA) described in Section \[sec:Nesterov\]. The entire algorithm is summarized as Algorithm \[alg:alm\], where $\tau$ represents the largest eigenvalue of the matrix $A^TA$, and $\rho > 1$ is a constant. Although the steps in Algorithm \[alg:alm\] closely resemble those of Algorithm \[alg:fista\], we will later see that ALM in general is more accurate and converges faster. $\bb \in {{\mathbb{R}}}^m$, $A \in {{\mathbb{R}}}^{m \times n}$. $t_1 \leftarrow 1$, $\zz_1 \leftarrow \xx_k$, $\uu_1 \leftarrow \xx_k$ $\uu_{l+1} \leftarrow \mathrm{soft}\left(\zz_l - \frac{1}{\tau}A^T\left(A\zz_l - \bb - \frac{1}{\mu_k}\yy_k\right), \frac{1}{\mu_k\tau}\right)$ $t_{l+1} \leftarrow \frac{1}{2}\left( 1 + \sqrt{1+4t_l^2}\right)$ $\zz_{l+1} \leftarrow \uu_{l+1} + \frac{t_l - 1}{t_{l+1}}(\uu_{l+1} - \uu_l)$ $ \xx_{k+1} \leftarrow \uu_{l+1}$ $\yy_{k+1} \leftarrow \yy_k + \mu_k (\bb - A\xx_{k+1})$ $\mu_{k+1} \leftarrow \rho\cdot\mu_k$ $\xx^* \leftarrow \xx_k$. \[alg:alm\] We would like to point out that a similar algorithm called Alternating Directions Method (ADM) [@YangJ2009] essentially shares the same idea as above.[^12] The major difference is that ADM would approximate the solution to the first step of the ALM iteration in . This is done by computing only one iteration of the FISTA algorithm. Although this approximation yields a gain in computational speed, we have observed from experiments that the convergence behavior of ADM may vary depending on the distribution of the matrix $A$. Algorithm \[alg:alm\] summarizes the ALM method applied to the primal problem , which is referred to as Primal ALM (PALM) in this paper. Interestingly, the principles of ALM can also be applied to its dual problem [@YangJ2009]: $$\max_{\yy} \bb^T\yy \quad \mbox{subj. to}\quad A^T\yy \in {\boldsymbol}{B}_1^{\infty},$$ where ${\boldsymbol}{B}_1^{\infty} = \{\xx\in {{\mathbb{R}}}^n: \|\xx\|_{\infty}\leq 1\}$. Under certain circumstances, this may result in a more computationally efficient algorithm. In the rest of the section, we will briefly explain the Dual ALM (DALM) algorithm. Simple computation shows that DALM solves the following problem: $$\begin{array}{rl} \min_{\yy, \xx, \zz} &-\bb^T \yy - \xx^T(\zz-A^T\yy) + \frac{\beta}{2}\|\zz-A^T\yy\|_2^2 \\ \subj& \zz \in {\boldsymbol}{B}_1^{\infty}. \end{array}$$ Here, $\xx$ as the primal variable becomes the associated Lagrange multiplier in the dual space [@YangJ2009]. Since it is difficult to solve the above problem simultaneously w.r.t. $\yy$, $\xx$ and $\zz$, we adopt a strategy by alternately updating the primal variable $\xx$ and the dual variables $\yy$ and $\zz$. On one hand, for $\xx=\xx_k$ and $\yy = \yy_k$, the minimizer $\zz_{k+1}$ with respect to $\zz$ is given by $$\zz_{k+1} = \mathcal{P}_{{\boldsymbol}{B}_1^{\infty}}(A^T\yy_k + \xx_k / \beta),$$ where $\mathcal{P}_{{\boldsymbol}{B}_1^{\infty}}$ represents the projection operator onto ${\boldsymbol}{B}_1^{\infty}$. On the other hand, given $\xx = \xx_k$ and $\zz = \zz_{k+1}$, the minimization with respect to $\yy$ is a least squares problem, whose solution is given by the solution to the following equation: $$\beta AA^T\yy = \beta A \zz_{k+1} - (A\xx_k - \bb). \label{eqn:dual_alm}$$ Suppose that $AA^T$ is invertible. Then, we directly use its inverse to solve . However, for large scale problems, this matrix inversion can be computationally expensive. Therefore, in such cases, we can approximate the solution with one step of the Conjugate Gradient algorithm in the $\yy$ direction at each iteration, as proposed in [@YangJ2009]. The basic iteration of the DALM algorithm is given by[^13] $$\left\{ \begin{array}{l} \zz_{k+1} = \mathcal{P}_{{\boldsymbol}{A}_1^{\infty}}(A^T\yy_k + \xx_k / \beta), \\ \yy_{k+1} = (AA^T)^{-1}(A \zz_{k+1} - (A\xx_k - \bb) /\beta), \\ \xx_{k+1} = \xx_{k} - \beta(\zz_{k+1} - A^T\yy_{k+1}), \end{array}\right.$$ As we will see in Section \[sec:experiment\], when the matrix $AA^T$ is easily invertible, it is possible to solve exactly. Since all the subproblems now are solved exactly, the convergence of the dual algorithm is guaranteed. Furthermore, since its major computation lies in solving for the dual variable $\yy$, when the number of dual variables are much smaller than the number of primal variables (i.e., when $d \ll n$), the dual algorithm could be more efficient than the primal algorithm. Simulation: Random Sparse Signals {#sec:simulation} ================================= In this section, we present two sets of experiment to benchmark the performance of the five fast $\ell_1$-min algorithms on random sparse signals, namely, L1LS/TNIPM, Homotopy, SpaRSA/IST, FISTA, and DALM, together with the basic BP algorithm (i.e., PDIPA). All experiments are performed in MATLAB on a Dell PowerEdge 1900 workstation with dual quad-core 2.66GHz Xeon processors and 8GB of memory. Before we present the results, we need to caution that the performance of these algorithms is affected by multiple factors, including the algorithm implementation, the chosen ad-hoc parameters, and even the MATLAB environment itself. One factor that we should pay special attention to is the stopping criteria used in benchmarking these algorithms. As we first mentioned in Section \[sec:BP\], choosing a good stopping criterion is important to properly exit an iteration when the estimate becomes close to a local or global optimum. On one hand, in general, straightforward rules do exist, such as the relative change of the objective function: $$\frac{\|F(\xx^{(k+1)})-F(\xx^{(k)}) \|}{\|F(\xx^{(k)}) \|}, \label{eq:relative-function-stopping}$$ or the relative change of the estimate: $$\frac{\|\xx^{(k+1)}-\xx^{(k)} \|}{\|\xx^{(k)} \|}. \label{eq:relative-estimate-stopping}$$ However, their efficacy depends on a proper step size of the update rule: If the step size is poorly chosen, the algorithm may terminate prematurely when the estimate is still far away from the optimum. On the other hand, certain special criteria are more effective to some algorithms than the others. For example, for PDIPA, it is natural to use the (relative) duality gap between the primal and dual solutions; for Homotopy, it is easy to measure the change of the sparse support as the stopping criterion. In this section, since the experiment is performed on synthetic data, we use a simple threshold that compares the $\ell_2$-norm difference between the $\ell_1$-min estimate $\xx^*$ and the ground truth $\xx_0$ as the stopping criterion. $\rho$-$\delta$ Plot in the Noise-Free Case {#sec:experiment-noise-free} ------------------------------------------- The first experiment is designed to measure the accuracy of the various algorithms in recovering sparse signals in the noise-free case $(P_1)$. We evaluate each algorithm using a $\rho$-$\delta$ plot, where the sparsity rate $\rho=k/n \in (0,1]$ and the sampling rate $\delta=d/n\in (0,1]$. At each $\delta$, the percentages of successes that an $\ell_1$-min algorithm finds the ground-truth solution $\xx_0$ (with a very small tolerance threshold) are measured over different $\rho$’s. Then a fixed success rate, say of 95%, over all $\delta$’s can be interpolated as a curve in the $\rho$-$\delta$ plot. In general, the higher the success rates, the better an algorithm can recover dense signals in the $(P_1)$ problem. Figure \[fig:success-rate-curve\] shows the 95% success-rate curves for the six algorithms. In the simulation, the ambient dimension $d=1000$ is fixed. For a fixed pair $(\rho,\delta)$, the support of the ground-truth signal $\xx_0$ is chosen uniformly at random, and the values of the non-zero entries are drawn from the standard normal distribution. In addition, the vector is normalized to have unit $\ell_2$-norm. The measurement matrix $A$ is generated as a random Gaussian matrix, each of whose entries is i.i.d. randomly generated from the standard normal distribution and then normalized to have unit column norm. We choose to compute the average success rates over 100 trials on the vertices of a grid of $(\rho,\delta)$ pairs for each of the $\ell_1$-min algorithms, and the coordinates of the 95% rate are interpolated from the vertex values. ![The $\rho$-$\delta$ plot (in color) shows the 95% success-rate curves for the six fast $\ell_1$-min algorithms.[]{data-label="fig:success-rate-curve"}](rho-delta-plot.png){width="2.4in"} The observations of the experiment are summarized below: 1. Without concerns about speed and data noise, the success rates of the interior-point method PDIPA is the highest of all the algorithms compared in Figure \[fig:success-rate-curve\], especially when the signal becomes dense. 2. The accuracy for the remaining five algorithm is separated in three groups. In particular, the success rates of L1LS and Homotopy are similar w.r.t. different sparsity rates and sampling rates, and they are also very close to PDIPA. In the low sampling-rate regime, Homotopy is slightly better than L1LS. 3. The success rates of FISTA and DALM are comparable over all sampling rates. In the noise-free case, they are not as accurate as the other exact $\ell_1$-min solutions. However, their performance shows a significant improvement over the IST algorithm, namely, SpaRSA. Performance with Moderate Data Noise {#sec:experiment-noisy} ------------------------------------ We are more interested in comparing the $\ell_1$-min algorithms when the measurement contains moderate amounts of data noise. In the second simulation, we rank the six algorithms under two scenarios: Firstly, we measure the performance in the low-sparsity regime, where the ambient dimension $n=2000$ and the sparsity ratio $\rho\doteq k/n=0.1$ are fixed, and the dimension $d$ of the Gaussian random projection varies from 800 to 1900. Secondly, we measure the performance when $\xx$ becomes dense w.r.t. a fixed sampling rate, where $n=2000$ and $d=1500$ are fixed, and the sparsity ratio $\rho=k/n$ varies from 0.1 to 0.26. The maximum number of iterations for all algorithms is limited to 5000. The results are shown in Figure \[fig:vary-dimension-comparison\] and \[fig:vary-sparsity-comparison\], respectively. In both experiments, we corrupt the measurement vector $\bb$ with $\ee$, an additive white noise term whose entries are i.i.d. according to a Gaussian distribution with zero mean and variance $0.01$. Firstly, when a low sparsity ratio of $\rho=0.1$ is fixed in Figure \[fig:vary-dimension-comparison\], $\ell_1$-min becomes better conditioned as the projection dimension increase. However, the computation cost also increases with the projection dimension. We then compare the performance of the six algorithm in Figure \[fig:vary-dimension-comparison\]: 1. The computational complexity of PDIPA grows much faster than the other fast algorithms. More importantly, in contrast to its performance in the noise-free case, the estimation error also grows exponentially, in which case the algorithm (i.e., the update rule ) fails to converge to an estimate that is close to the ground truth. 2. As the projection dimension increases, the five fast $\ell_1$-min algorithms all converge to good approximate solutions. The estimation error of Homotopy is slightly higher than the rest four algorithms. 3. In terms of speed, L1LS and Homotopy take much longer time to converge than SpaRSA, FISTA, and DALM. 4. The average run time of DALM is the shortest over all projection dimensions, which makes it the best algorithm in this comparison. Secondly, when the projection dimension $d=1500$ is fixed in Figure \[fig:vary-sparsity-comparison\], we compare both the average run time and the average estimation error when the sparsity varies: 1. PDIPA significantly underperforms the rest five fast algorithms in terms of both accuracy and speed, consistent with the result in the previous simulation with noisy data. 2. The average run time of Homotopy grows almost linearly with the sparsity ratio, while the other algorithms are relatively unaffected. Thus, Homotopy is more suitable for scenarios where the unknown signal is expected to have a very small sparsity ratio. 3. The computational cost of the rest four algorithms, namely, L1LS, SpaRSA, FISTA, and DALM, remains largely unchanged when the random projection dimension $d$ is fixed. 4. DALM is the fastest algorithm in the low-sparsity regime, but its run time approaches that of the other first-order methods in the high-sparsity regime. Overall, DALM is the best algorithm in this scenario. To summarize, in the presence of low levels of Gaussian noise, PDIPA performs quite badly in comparison with the other $\ell_1$-min algorithms that employ more sophisticated techniques to handle noisy data. Perhaps a more surprising observation is that when the coefficients of $A$ are randomly drawn from a Gaussian distribution, under a broad range of simulation conditions, FISTA is *not* significantly better than the original IST algorithm.[^14] However, in the next section, we will see that when the dictionary $A$ is replaced by real-world data matrices that contain training images in face recognition, the performance of the six $\ell_1$-min algorithms would be dramatically different from synthetic data. Face Recognition Experiment I: Corruption and Disguise Compensation {#sec:experiment} =================================================================== In this section, we benchmark the performance of the six algorithms in robust face recognition. The experiment is set up to estimate sparse representation of real face images based on the *cross-and-bouquet* (CAB) model introduced in [@WrightJ2008-IT]. More specifically, It has been known in face recognition that a well-aligned frontal face image $\bb$ under different lighting and expression lies close to a special low-dimensional linear subspace spanned by the training samples from the same subject, called a face subspace [@BelhumeurP1997-PAMI; @BasriR2003-PAMI]: $$A_i = [\vv_{i,1}, \vv_{i,2}, \cdots, \vv_{i,n_i}] \in {{\mathbb{R}}}^{d\times n_i},$$ where $\vv_{i,j}$ represents the $j$-th training image from the $i$-th subject stacked in the vector form. Given $C$ subjects and a new test image $\bb$ (also in the vector form), we seek the sparsest linear representation of the sample with respect to all training examples: $$\bb = [A_1, A_2, \cdots, A_C] [\xx_1; \xx_2; \cdots; \xx_C] = A\xx, \label{eq:intro-classification}$$ where $A\in{{\mathbb{R}}}^{d\times n}$ collects all the training images. Clearly, if $\bb$ is a valid test image, it must be associated with one of the $C$ subjects. Therefore, the corresponding representation in has a sparse representation $\xx = [\cdots; \zero; \xx_i; \zero; \cdots ]$: on average only a fraction of $\frac{1}{C}$ coefficients are nonzero, and the dominant nonzero coefficients in sparse representation $\xx$ reveal the true subject class. In addition, we consider the situation where the query image $\bb$ may be severely occluded or corrupted. The problem is modeled by a corrupted set of linear equations $\bb = A\xx + \ee$, where $\ee\in{{\mathbb{R}}}^{d}$ is an unknown vector whose nonzero entries correspond to the corrupted pixels. In [@WrightJ2009-PAMI], the authors proposed to estimate $\ww\doteq[\xx; \ee]$ jointly as the sparsest solution to the extended equation: $$\min\|\ww \|_1 \mbox { subject to } \bb = [A,\ I]\ww. \label{eq:CAB}$$ The new dictionary $[A,\ I]$ was dubbed a cross-and-bouquet model in the following sense. The columns of A are highly correlated, as the convex hull spanned by all face images of all subjects occupies an extremely tiny portion of the ambient space. These vectors are tightly bundled together as a “bouquet,” whereas the vectors associated with the identity matrix and its negative $\pm I$ form a standard “cross” in ${{\mathbb{R}}}^{d}$, as shown in Figure \[fig:cross-and-bouquet\]. ![The cross-and-bouquet model for face recognition. The raw images of human faces expressed as columns of $A$ are clustered with very small variance. (Courtesy of John Wright [@WrightJ2008-IT])[]{data-label="fig:cross-and-bouquet"}](cross-and-bouquet.png){width="3in"} In this section, the performance of the six $\ell_1$-min algorithms using the CAB model is benchmarked on the CMU Multi-PIE face database [@GrossR2006]. A subset of 249 subjects from the data set (Session 1) is used for this experiment. Each subject is captured under 20 different illumination conditions with a frontal pose. The images are then manually aligned and cropped, and down-sampled to $40\times 30$ pixels. Out of the 20 images for each subject, images $\{0, 1, 7, 13, 14, 16, 18 \}$ with extreme illumination conditions are chosen as the training images, and the remaining 13 images are used as test images. Finally, a certain number of image pixels are randomly corrupted by a uniform distribution between $[0,255]$, with the corruption percentage from 0% to 90%, as four examples shown in Figure \[fig:occlusion\]. In Table \[tab:multi-pie-accuracy\], the recognition rates between 50% and 70% pixel corruption are highlighted for more detailed comparison. ![An aligned face image of Subject 1 in Multi-PIE, Session 1, under the ambient lighting condition (No. 0) is shown on the left. On the right, 20%, 40%, 60%, and 80% of image pixels are randomly selected and corrupted with uniformly distributed values in \[0, 255\], respectively.[]{data-label="fig:occlusion"}](Multi-PIE-occlusion.png){width="3.2in"} We measure the performance of the algorithms in terms of the final recognition rate and the speed of execution. In choosing a proper stopping criterion, the stopping threshold is individually tuned for each algorithm to achieve the highest recognition rate. The results are shown in Figure \[fig:multi-pie\]. In addition, Figure \[fig:recovered-faces\] shows the reconstructed face images after corruption compensation, namely, $\hat{\bb} = \bb - \ee$, given by the Homotopy solver. ![ Average recognition accuracy (in percentage) on the Multi-PIE database. [**Right:**]{} Average run time (in second) on the Multi-PIE database.[]{data-label="fig:multi-pie"}](multi_pie_accuracy.png "fig:"){width="2.4in"} ![ Average recognition accuracy (in percentage) on the Multi-PIE database. [**Right:**]{} Average run time (in second) on the Multi-PIE database.[]{data-label="fig:multi-pie"}](multi_pie_runtime.png "fig:"){width="2.4in"} Corr. PDIPA L1LS Homotopy SpaRSA FISTA DALM ------- -------------- ------ -------------- -------- ------- ------ 50% [**99.8**]{} 99.5 [**99.8**]{} 97.6 96.2 99.5 60% 98.6 96.6 [**98.7**]{} 90.5 86.8 96.2 70% 84.1 76.3 [**84.6**]{} 63.3 58.7 78.8 : Average recognition rates (in percentage) between 50% and 70% random pixel corruption on the multi-PIE database. The highest rates are in boldface.[]{data-label="tab:multi-pie-accuracy"} In terms of accuracy, Homotopy achieves the best overall performance. For instance, with 60% of the pixels randomly corrupted, its average recognition rate based on the CAB model is about 99%. The performance of PDIPA is very close to Homotopy, achieving the second best overall accuracy. On the other hand, FISTA obtains the lowest recognition rates, followed by SpaRSA. In terms of speed, Homotopy is also one of the fastest algorithm. In particular, when the pixel corruption percentage is small, the sparsity of the coefficient vector $\ww = [\xx; \ee]$ is very small. As Homotopy iteratively adds or removes nonzero coefficients one at a time, the algorithm can quickly terminate after just a few iterations. On the other hand, as $\ww$ becomes dense when the pixel corruption percentage increases, the complexity of Homotopy increases superlinearly and becomes inefficient. It is also important to note that the speed of the two accelerated iterative-thresholding algorithms, namely, FISTA and DALM, does not increase significantly as the sparsity of $\ww$ increases. It is more interesting to separately compare PDIPA, L1LS, and Homotopy, which provably solve the $(P_1)$ problem, and SpaRSA, FISTA, and DALM, which essentially solve a relaxed version of $(P_1)$. Overall, PDIPA, L1LS and Homotopy perform similarly in terms of the accuracy, consistent with our previous observations in the simulation. However, L1LS is significantly more expensive from a computation point of view to achieve the same recognition rates as Homotopy. Among the three iterative soft-thresholding methods, the experiments suggest that DALM is the most accurate. Face Recognition Experiment II: Face Alignment {#sec:alignment} ============================================== In this section, we benchmark the performance of the $\ell_1$-min algorithms in robust face alignment within the sparse representation framework. While the previous section has demonstrated that solving the CAB model achieves state-of-the-art recognition accuracy on public datasets even when the test face images are severely corrupted or occluded, its success still relies on the assumption that the test images are well aligned with the training images. However, this assumption is not always satisfied in practice. As illustrated in Figure \[fig:promo\] top, a test face image obtained from an off-the-shelf face detector is likely to have some registration error against the training images. The registration error may contribute to erroneous linear representation of the query image, even if sufficient illuminations are present in the training set. Fortunately, this alignment issue can be naturally addressed within the sparse representation framework. It has been shown in [@WagnerA2009-CVPR] that the face alignment problem can be solved by a series of linear problems that iteratively minimize the sparsity of the registration error, which leads to an effective algorithm that works for face images under a large range of variations in scale change, 2-D translation and rotation, and different 3-D poses. Furthermore, it inherits the robustness property from the sparse representation framework, and hence is also able to deal with faces that are partially occluded. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![[**Effect of registration**]{}. The task is to identify the girl among 20 subjects, by computing the sparse representation of her input face with respect to the entire training set. The absolute sum of the coefficients associated with each subject is plotted on the right. We also show the faces reconstructed with each subject’s training images weighted by the associated sparse coefficients. The red line (cross) corresponds to her true identity, subject 12. [**Top:**]{} The face region is extracted by Viola and Jones’ face detector (the black box). [**Bottom:**]{} Informative representation obtained by using a well-aligned face region (the white box). (Courtesy of [@WagnerA2009-CVPR]).[]{data-label="fig:promo"}](detector.png "fig:"){height="0.9in"} ![[**Effect of registration**]{}. The task is to identify the girl among 20 subjects, by computing the sparse representation of her input face with respect to the entire training set. The absolute sum of the coefficients associated with each subject is plotted on the right. We also show the faces reconstructed with each subject’s training images weighted by the associated sparse coefficients. The red line (cross) corresponds to her true identity, subject 12. [**Top:**]{} The face region is extracted by Viola and Jones’ face detector (the black box). [**Bottom:**]{} Informative representation obtained by using a well-aligned face region (the white box). (Courtesy of [@WagnerA2009-CVPR]).[]{data-label="fig:promo"}](sci_with_axis_face_case1.png "fig:"){height="0.9in" width="1.7in"} ![[**Effect of registration**]{}. The task is to identify the girl among 20 subjects, by computing the sparse representation of her input face with respect to the entire training set. The absolute sum of the coefficients associated with each subject is plotted on the right. We also show the faces reconstructed with each subject’s training images weighted by the associated sparse coefficients. The red line (cross) corresponds to her true identity, subject 12. [**Top:**]{} The face region is extracted by Viola and Jones’ face detector (the black box). [**Bottom:**]{} Informative representation obtained by using a well-aligned face region (the white box). (Courtesy of [@WagnerA2009-CVPR]).[]{data-label="fig:promo"}](alignment.png "fig:"){height="0.9in"} ![[**Effect of registration**]{}. The task is to identify the girl among 20 subjects, by computing the sparse representation of her input face with respect to the entire training set. The absolute sum of the coefficients associated with each subject is plotted on the right. We also show the faces reconstructed with each subject’s training images weighted by the associated sparse coefficients. The red line (cross) corresponds to her true identity, subject 12. [**Top:**]{} The face region is extracted by Viola and Jones’ face detector (the black box). [**Bottom:**]{} Informative representation obtained by using a well-aligned face region (the white box). (Courtesy of [@WagnerA2009-CVPR]).[]{data-label="fig:promo"}](sci_with_axis_face_case3.png "fig:"){height="0.9in" width="1.7in"} ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Recall that a well-aligned test image $\bb_0 \in {{\mathbb{R}}}^{d}$ can be represented as a sparse linear combination $A\xx_0$ of all of the images in the database, plus a sparse error $\ee_0$ due to occlusion: $\bb_0 = A\xx_0 + \ee_0$. Now suppose that $\bb_0$ is subject to some misalignment, so instead of observing $\bb_0$, we observe the warped image $$\bb = \bb_0 \circ \tau^{-1}$$ for some transformation $\tau\in T$, where $T$ is a finite-dimensional group of transformations acting on the image domain. As seen in Figure \[fig:promo\], directly seeking a sparse representation of $\bb$ against the (well-aligned) training images is no longer appropriate. However, if the true deformation $\tau^{-1}$ can be efficiently found, then we can apply its inverse $\tau$ to the test image, and it again becomes possible to find a good sparse representation of the resulting image: $$\bb \circ \tau = A \xx + \ee.$$ ------------------------------------ ---------------------------------------- ------------------------------------------ ----------------------------------------- {width="1.6in"} {width="1.6in"} {width="1.6in"} {width="1.6in"} {width="1.6in"} {width="1.6in"} {width="1.6in"} {width="1.6in"} {width="1.6in"} {width="1.6in"} {width="1.6in"} {width="1.6in"} (a) (b) (c) (d) ------------------------------------ ---------------------------------------- ------------------------------------------ ----------------------------------------- Naturally, we would like to use the sparsity as a cue for finding the correct deformation $\tau$. This can be formulated as the following optimization problem: $$\min_{\xx,\ee,\tau\in T}\|\xx\|_1 + \|\ee\|_1 \quad \textup{subj to} \quad \bb \circ \tau = A \xx + \ee. \label{eq:global-alignment}$$ However, this is a difficult nonconvex optimization problem. Furthermore, due to the concern of local minima, directly solving may simultaneously align the test image to different subjects in the database. Therefore, it is more appropriate to seek the best alignment with each subject $i$ in the database [@WagnerA2009-CVPR]: $$\hat{\tau_i} = \arg \min_{\xx,\ee,\tau_i\in T}\|\ee\|_1 \quad \textup{subj to} \quad \bb \circ \tau_i = A_i \xx + \ee. \label{eqn-align}$$ In , $\|\xx\|_1$ is not penalized, since $A_i\in {{\mathbb{R}}}^{d\times n_i}$ only contains the images of subject $i$ and $\xx$ is no longer expected to be sparse. While is still nonconvex, when a good initial estimation for the transformation is available, e.g., from the output of a face detector, one can refine this initialization to an estimate of the true transformation by repeatedly linearizing about the the current estimate of $\tau_i$. This idea leads to the final problem formulation as a convex optimization problem as follows: $$\min_{\xx,\ee,\Delta \tau_i\in T}\|\ee\|_1 \quad \textup{subj to} \quad \bb\circ \tau_i + J_i\Delta \tau_i = A_i \xx + \ee. \label{eqn-align-linear}$$ Here, $J_i = \frac{\partial}{\partial \tau_i}\bb \circ \tau_i \in {{\mathbb{R}}}^{d\times q_i}$ is the Jacobian of $\bb \circ \tau_i$ with respect to the transformation parameters $\tau_i$, and $\Delta \tau_i$ is the update direction w.r.t. $\tau_i$. During each iteration $k$, the current alignment parameters $\tau_i^{(k)}$ correct the observation as $b_i^{(k)} = b\circ \tau_i^{(k)}$. Denote $B_i^{(k)} = [A_i, -J_i^{(k)}]\in {{\mathbb{R}}}^{d\times (n_i+q_i)}$ and $\ww = [\xx^T, \Delta\tau_i^T]^T$, then the update $\Delta\tau_i$ can be computed by solving the following problem: $$\min_{\ww,\ee}\|\ee\|_1 \quad \textup{subj to} \quad \bb_i^{(k)} = B_i^{(k)}\ww+\ee. \label{eqn-align-simple}$$ Interested readers are referred to [@WagnerA2009-CVPR] for more details about the algorithm and comprehensive experimental studies. In this section, we benchmark the various $\ell_1$-min algorithms on the CMU Multi-PIE face database. In order to solve , we had to modify the code for each algorithm so that $\|\xx\|_1$ is no longer penalized and to take care of the new constraint. To further speed up the algorithms, we take advantage of the fact that the matrix $B$ is usually a tall matrix with much fewer columns than rows. For our experiments, the training set contains $n_i=7$ images per subject, and we choose the transformation group $T$ to be the set of similarity transforms (therefore $q_i=4$). So, the number of columns in $B$ is just $n_i+q_i = 11$, while the number of rows is equal to the number of pixels in each image. Firstly, we modify the PDIPA to obtain a faster algorithm for this problem. We recall that PDIPA is computationally expensive mainly because at each iteration, we need to solve a linear system of equations the size of the dictionary $[B, I]$ for . For instance, if $d=40\times 30 = 1200$, it results in inverting a matrix of size larger than $1200\times 1200$. However, if we know that $B$ has very few columns, it is not efficient to consider the dictionary $[B, I]$ as a whole. Instead, we can explicitly write down the closed-form solution to the large linear system with $B$ and $I$ considered separately. This leads to a much smaller problem that only requires inverting a $(n_i+q_i) \times (n_i+q_i) = 11 \times 11$ matrix. Hence, the complexity of PDIPA is significantly reduced. We note that similar ideas can be also applied to speed up the other algorithms. In particular, when matrix $B$ is typically an overdetermined matrix, the ALM algorithm that operates in the primal space, namely, PALM, is more efficient than DALM in the dual space. In the Appendix, modifications for three $\ell_1$-min algorithms are briefly explained, which solve based on GP, Homotopy, and IST approaches, respectively. The performance of the six $\ell_1$-min algorithms is again evaluated using the face images from the Multi-PIE database. In this experiment, the 249 subjects from the Session 1 are used. Out of 20 illuminations, $\{0,1,7,13,14,16,18\}$ are chosen as the training images and the illumination $10$ is used as the query image. All images are cropped and down-sampled to $40\times 30$ pixels. Furthermore, the test images are manually perturbed up to $8$ pixels along the $x$ and $y$-axes in the canonical frame (with size $40\times 30$) or up to $45^{\circ}$ in-plane degree, respectively. Moreover, to test the robustness of the $\ell_1$-min algorithms to occlusion, we replace a random located block of size $10\%$ of the face image with the image of a baboon. See Figure \[fig:alignment-result\](a) for an example. The accuracy of each algorithm is measured by both the average pixel error and the rate of success. We consider an alignment successful if the difference between the final alignment is within 3 pixels of the ground truth in the original image frame ($640\times 480$ image size). As shown in Figure \[fig:alignment-result\](b) and (c), in terms of alignment accuracy, PDIPA, PALM and L1LS achieve the best performance, with Homotopy and FISTA slightly worse. Meanwhile, the performance of SpaRSA degrades much faster than the others as the perturbation level increases. On the other hand, Figure \[fig:alignment-result\](d) shows that L1LS, SpaRSA, and FISTA are more computational expensive. Overall, PALM is the fastest algorithm, which takes $25\%$ to $50\%$ less time compared to PDIPA and Homotopy. In conclusion, Both PALM and PDIPA performs very well for the alignment problem. Taking into account the implementation complexity and the fact that PALM is slightly faster than PDIPA, we conclude that PALM is the best choice for this problem. Conclusion and Discussion {#sec:conclusion} ========================= We have provided a comprehensive review of five fast $\ell_1$-min methods, i.e., *gradient projection*, *homotopy*, *soft shrinkage-thresholding*, *proximal gradient*, and *augmented Lagrange multiplier*. Through extensive experiments, we have shown that under a wide range of data conditions, there is no clear winner that always achieves the best performance in terms of both speed and accuracy. For perfect, noise-free data, the primal-dual interior point method (PDIPA) is more accurate than the other algorithms, albeit at a much lower speed. For noisy data, first-order $\ell_1$-min techniques (i.e., SpaRSA, FISTA, and ALM) are more efficient in recovering sparse signals in both the low-sparsity and high-sparsity regimes. We have also considered special cases of the $\ell_1$-min problem where it is applied to robust face recognition. In this application, Homotopy and ALM achieve the highest recognition rates, and their computational cost is also the lowest compared to the other $\ell_1$-min algorithms. All experiments described in this paper were carried out in MATLAB. The benchmarking in terms of computational speed may vary significantly when the algorithms are implemented in other programming languages such as C/C++. The accuracy of different algorithms may also be affected by the precision of different floating-point/fixed-point arithmetic logic. To aid in peer evalution, all the experimental scripts and data have been made available on our website. Appendix ======== In this section, we briefly discuss how to more efficiently solve the objective function by taking advantage of the intrinsic structure of its dictionary $[B, I]$. For clarity, we will rewrite using the following notation: $$\min_{\ww,\ee}\|\ee\|_1 \quad \textup{subj to} \quad \bb = B\ww+\ee. \label{eqn-align-appendix}$$ Solving Face Alignment via Gradient Projection Methods ------------------------------------------------------ The objective function can be optimized relatively straightforward following the gradient projection approach. In particular, using *truncated Newton interior-point method*, the objective function is rewritten as: $$\begin{aligned} \min & \frac{1}{2}\|\bb-B\ww -\ee\|^2 + \lambda \sum_{i=1}^{d} u_i\\ \mbox{subj. to} & -u_i\le e_i \le u_i, \quad i=1,\cdots, d.\end{aligned}$$ Then a logarithmic barrier to bound the domain of $-u_i\le e_i \le u_i$ can be constructed as: $$\Phi(\ee, \uu) = -\sum \log(u_i + e_i) - \sum \log(u_i - e_i).$$ Over the domain of $(\ww, \ee, \uu)$, the central path is calculated by minimizing the following convex function $$F_t(\ww, \ee, \uu) = t(\frac{1}{2} \|\bb -B\ww - \ee \|^2 + \lambda \sum_{i=1}^d u_i) + \Phi(\ee, \uu),$$ where $t$ is a nonnegative parameter. Then, the update direction $(\Delta \ww, \Delta \ee, \Delta \uu)$ is solved using Newton’s method: $$\nabla^2F_t(\ww, \ee, \uu) \cdot \left[\begin{matrix} \Delta \ww\\ \Delta \ee\\ \Delta \uu \end{matrix}\right] = - \nabla F_t(\ww, \ee, \uu).$$ Solving Face Alignment via Homotopy Methods ------------------------------------------- In the Homotopy formulation, define a composite objective function: $$\begin{aligned} F(\ww, \ee) &=& \frac{1}{2}\|\bb - B\ww - \ee\|^2 + \lambda \|\ee\|_1\\ &=& f(\ww, \ee) + \lambda g(\ee). \label{eq:composite}\end{aligned}$$ Setting the subgradient $\partial F = 0 \Leftrightarrow -\partial f = \lambda \partial g$, we obtain the following equalities: $$\lambda \frac{\partial}{\partial \ww} g(\ee) = 0 = B^T (\bb - B\ww - \ee) \label{eq:1}$$ and $$\cc(\ee)\doteq \lambda \frac{\partial}{\partial \ee} g(\ee) = (\bb - B\ww - \ee). \label{eq:2}$$ In initialization, set $\ee^{(0)} = \zero$ and $\ww = B^{\dagger} \bb$. Since we know $\partial \|\ee\|_1 \in [-1, 1]$, we initialize the maximal $\lambda$ using : $$\lambda_0 = \max(\mbox{abs}(\cc)),$$ and the support ${\mathcal{I}}\doteq \{i : | c_i |=\lambda \}$. At each iteration $k$, first assuming $\ww$ is fixed, the update direction for $\ee$ is $$\dd^{(k)}({\mathcal{I}}) = -\lambda\cdot\mbox{sgn}(\cc^{(k)}({\mathcal{I}})),$$ and the direction for the coefficients that are not in the support ${\mathcal{I}}$ is manually set to zero. Hence, $$\ee^{(k+1)} = \ee^{(k)} + \gamma \dd^{(k)}. \label{eq:update}$$ In , a proper step size value $\gamma$ remains to be determined. In Homotopy methods, $\gamma$ is set as the minimal step size that leads to either a coefficient of $\cc$ that is in the support ${\mathcal{I}}$ crosses zero, and hence should be removed from ${\mathcal{I}}$; or the magnitude of a coefficient of $\cc$ outside the support ${\mathcal{I}}$ reaches or exceed $\lambda$, and hence should be added to ${\mathcal{I}}$. The first situation is easy to evaluate: $$\gamma^- = \min_{i\in{\mathcal{I}}}\{-e_i/d_i \}.$$ However, choosing $\gamma^+$ for the second situation presents a problem, as in any update in $\ee$ does not change the values of $\cc$ outside the support. In other words, due to the fact that the dictionary for $\ee$ where its coefficients are sparse is indeed an *identity* matrix, the original strategy in Homotopy could not add any new indices into the support of $\ee$. An alternative solution that we use in our implementation is to determine $\gamma$ and $\ww$ separately using the first constraint , and then iterate. Solving Face Alignment via Iterative Soft-Thresholding ------------------------------------------------------ The composite objective function $F(\ww, \ee) = f(\ww, \ee) + \lambda g(\ee)$ in can be also solved by *iterative soft-thresholding* (IST) methods. To simplify the notation, denote $\yy = [\ww; \ee]\in{{\mathbb{R}}}^{n+d}$. Then the objective function becomes: $$F(\yy) = \frac{1}{2} \|\bb - [B, I] \yy \|^2 + \lambda \|\ee \|_1.$$ Then the update rule to minimize $F(\yy)$ is computed by substitutting a linearized approximation of $f$ as shown below: $$\begin{aligned} {rcl} \yy^{(k+1)} & = &\arg\min_\yy \,\{(\yy-\yy^{(k)})^T\nabla f(\yy^{(k)}) +Ê \nonumber \\ & & \frac{1}{2}\|\yy - \yy^{(k)} \|^2\cdot \nabla^2f(\yy^{(k)}) + \lambda g(\ee)\}. \label{eq:IST-objective-appx}\end{aligned}$$ We further introduce an intermediate variable $$\uu^{(k)} = \yy^{(k)} - \frac{1}{\alpha^{(k)}}\nabla f(\yy^{(k)}),$$ where the hessian $\nabla^2f(\yy^{(k)})$ is approximated by the diagonal matrix $\alpha^{(k)} I$. Then based on the IST approach, has a closed-form solution as follows: For $i\le n$, $$y_i^{(k+1)} = w_i^{(k+1)}= \arg\min_{y_i}\left\{\frac{(y_i - u_i^{(k)})^2}{2} \right \} = u_i^{(k)};$$ and for $i>n$, $$\begin{array}{rcl} y_i^{(k+1)} & = & e_{i-n}^{(k+1)} \\ &=& \arg\min_{y_i}\left\{\frac{(y_i - u_i^{(k)})^2}{2} + \frac{\lambda | y_i |}{\alpha^{(k)}}\right\}\\ & =& \mbox{soft}\left(u_i^{(k)}, \frac{\lambda}{\alpha^{(k)}}\right). \end{array}$$ Acknowledgements {#acknowledgements .unnumbered} ================ The authors would like to thank Dr. John Wright and Dr. Zhouchen Lin at Microsoft Research Asia for his valuable comments. Yang also appreciates the hospitality of the Visual Computing Group at Microsoft Research Asia during his visit there in 2009. [^1]: A. Yang and S. Sastry are with the Department of Electrical Engineering and Computer Sciences, University of California, Berkeley, USA. A. Ganesh, Z. Zhou, and Y. Ma are with the Coordinated Science Laboratory, University of Illinois, Urbana, USA. Y. Ma is also with the Visual Computing Group, Microsoft Research Asia, Beijing, China. Corresponding author: Allen Yang, Cory Hall, University of California, Berkeley, CA 94720. Email: yang@eecs.berkeley.edu. Tel: 1-510-643-5798. Fax: 1-510-643-2356. [^2]: This work was partially supported by NSF IIS 08-49292, NSF ECCS 07-01676, ONR N00014-09-1-0230, ARO MURI W911NF-06-1-0076, and ARL MAST-CTA W911NF-08-2-0004. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Laboratory or the U.S. Government. The U.S. Government is authorized to reproduce and distribute for Government purposes notwithstanding any copyright notation hereon. An extended abstract of the work was previously published in the Proceedings of the International Conference on Image Processing in 2010 [@YangA2010-ICIP]. [^3]: By sparse, we mean that most entries in the vector are zero. [^4]: This constraint can be easily removed by considering the linear system $\bb = [A, -A][\xx_+; \xx_-]$, where $[\xx_+; \xx_-]$ is also nonnegative. [^5]: In general, any smooth function $\Psi$ that satisfies $\Psi(0^+)=-\infty$ is a valid barrier function [@HertogD1992]. [^6]: There are multiple versions of the primal-dual interior-point solver implemented in MATLAB. Notable examples include SparseLab at <http://sparselab.stanford.edu/>, the CVX environment at <http://cvxr.com/cvx/>, and the $\ell_1$magic package at <http://www.acm.caltech.edu/l1magic/>. [^7]: A MATLAB implementation of GPSR is available at <http://www.lx.it.pt/~mtf/GPSR>. A MATLAB Toolbox for TNIPM called L1LS is available at <http://www.stanford.edu/~boyd/l1_ls/>. [^8]: A MATLAB implementation [@AsifM2008] can be found at <http://users.ece.gatech.edu/~sasif/homotopy/>. [^9]: A MATLAB implementation called *Sparse Reconstruction by Separable Approximation* (SpaRSA) [@WrightS2008] is available at <http://www.lx.it.pt/~mtf/SpaRSA/>. [^10]: This problem occurs in the IST algorithm as well. [^11]: An implementation of FISTA is provided on the website of this paper. Another Matlab toolbox called NESTA [@BeckerS2009] is available at: <http://www.acm.caltech.edu/~nesta/>. [^12]: A MATLAB toolbox of the ADM algorithm called YALL1 is provided at: <http://yall1.blogs.rice.edu/>. [^13]: The PALM and DALM algorithms in MATLAB can be downloaded from the website of this paper. [^14]: We have observed that the performance of FISTA may vary widely depending on its chosen parameters. If the parameters are tuned at different noise level, its speed of convergence can be much improved compared to IST. However, for consistency, all algorithms in this simulation are assigned a fixed set of parameters.

--- abstract: 'Kirkwood-Buff (KB) integrals are notoriously difficult to converge from a canonical simulation because they require estimating the grand-canonical radial distribution. The same essential difficulty is encountered when attempting to estimate the direct correlation function of Ornstein-Zernike theory by inverting the pair correlation functions. We present a new theory that applies to the entire, finite, simulation volume, so that no cutoff issues arise at all. The theory gives the direct correlation function for closed systems, while smoothness of the direct correlation function in reciprocal space allows calculating canonical KB integrals via a well-posed extrapolation to the origin. The present analysis method represents an improvement over previous work because it makes use of the entire simulation volume and its convergence can be accelerated using known properties of the direct correlation function. Using known interaction energy functions can make this extrapolation near perfect accuracy in the low-density case. Because finite size effects are stronger in the canonical than the grand-canonical ensemble, we state ensemble correction formulas for the chemical potential and the KB coefficients. The new theory is illustrated with both analytical and simulation results on the 1D Ising model and a supercritical Lennard-Jones fluid. For the latter, the finite-size corrections are shown to be small.' author: - 'David M. Rogers' title: ' Extension of Kirkwood-Buff Theory to the Canonical Ensemble' --- Introduction ============= Kirkwood and Buff’s theory of solutions exploits a Gibbs relation to determine derivatives of activities from fluctuations in particle number at constant volume, $$\left( {\ensuremath{\frac{\partial{\beta \mu_\alpha}}{\partial{n_\gamma}}}}\right)_{n_\eta,V,T} = \left[{\ensuremath{\left\langle{n_{\alpha'} n_{\gamma'} | \mu,V,T}\right\rangle}} - V^2 \rho_{\alpha'} \rho_{\gamma'} \right]_{\alpha\gamma}^{-1},$$ with the notation $\rho_\alpha = {\ensuremath{\left\langle{n_\alpha | \mu, V, T}\right\rangle}}/V$, the average number of molecules of type $\alpha$ in the grand-canonical ensemble (where $\mu, V, T$ are given), and using $\mu \in \mathbb R^\nu$ to denote a vector of chemical potential values and $n \in \mathbb N^\nu$, a vector of molecule numbers for all $\nu$ solution components. The inverse indicated above is the Moore-Penrose pseudoinverse of the $\nu\times\nu$ matrix of number fluctuations. KB theory is exact even when applied to periodic simulation volumes with finite unit cells, but requires an [*open*]{} system (i.e. data from the $\mu,V,T$ ensemble). In this work, we consider only such finite unit cells and use the additional assumption that no fixed external potential is present that breaks the translational symmetry. We introduce a theory that applies equally to open and closed systems and then investigate its zero-frequency limit (which is only rigorous for the open system). Also, we maintain constant temperature throughout and hence omit it in further notation. KB theory is commonly cast in terms of the grand-canonical radial distribution function, $$g_{\alpha\gamma}(r | \mu, L) \equiv \frac{ {\ensuremath{\left\langle{\hat \rho_\alpha(r) \hat \rho_\gamma(0) | \mu, L}\right\rangle}} - \delta_{\alpha\gamma} \delta(r) \rho_\alpha }{ \rho_\alpha \rho_\gamma },\label{e:g}$$ whose integral reports on number fluctuations in the grand-canonical ensemble. Here, $r \in \mathbb R^3 / L$ is a coordinate vector in the periodic unit cell with lattice vectors given as rows of a 3$\times$3 matrix, $L$, and $$\hat \rho_\alpha(r) \equiv \sum_{i=1}^{n_\alpha} \delta(r - r_{\alpha,i})$$ is a sum of delta-functions located at the centers of mass for all particles of type $\alpha$ (so that $\int_V \hat \rho_\alpha(r) dr = n_\alpha$ and $\rho_\alpha = {\ensuremath{\left\langle{\hat \rho_\alpha(0) | \mu, L}\right\rangle}}$). Directly integrating Eq. \[e:g\] gives Kirkwood and Buff’s Eq. 7,[@kb] confirming that it applies to [*finite*]{} volumes with the entire volume as the domain of integration. A recurring problem is that KB integrals estimated from closed systems (i.e. those under NVT or NPT conditions) contain substantial finite-size effects.[@nmatu96; @pgang13] Proposed corrections fall into two categories: those aimed at computing number density fluctuations directly in small sub-volumes,[@mrove90; @sschn11; @rcort16] and those addressing the cutoff error in truncating the KB integral at a finite maximum radius.[@nmatu96; @pkrug13; @pgang13] Both categories recognize that [*two*]{} independent corrections are necessary. First, the sub-volume or truncation radius is typically extrapolated to infinite size. Second, the radial distribution (or equivalently the number correlations) themselves must be corrected for the effect of using a closed system. These corrections were defined as implicit and explicit by Salacuse et. al. [@jsala96; @jsala96b], who derived and compared calculations of the closed system correction. However, the finite truncation radius remained as a source of ambiguity, which more recent works have addressed with renewed interest.[@pkrug13] In this work, we re-introduce the structure factor as the preferred method for computing both the direct correlation function and the KB integral as its zero-frequency limit. This corrects a minor defect in Ref.  by using the unique vectors in reciprocal space that correspond to a finite simulation volume, and yields exact direct correlation functions corresponding to the canonical ensemble simulated. For the problem of estimating the thermodynamic limit, our method improves on earlier works as it remains invariant to shifts of $g(r|n, L)$ by a constant, and makes use of the entire available simulation volume. In the process, we find two major conclusions. First, extrapolation of the direct correlation function for closed systems to zero frequency gives good estimates of the second derivatives of the Helmholtz free energy. Second, both closed (canonical) and open (grand-canonical) KB integrals can be estimated under near-critical conditions from box lengths on the order of 5 to 10 times the correlation length. Theory ======= Definitions ------------ We introduce the structure factors using Fourier transforms of the density operators, $$\begin{aligned} \hat S_\alpha(m) &\equiv \mathcal F[\hat \rho_\alpha](m) \equiv \int_V \hat \rho_\alpha(r) e^{-2\pi i m\cdot r} \; dr . \intertext{Explicit formulas for these transforms are,} \hat S_\alpha(m) &= \sum_{j=1}^{n_\alpha} e^{-2\pi i m\cdot r_{\alpha,j}}, \, \hat \rho_\alpha(r) = \frac{1}{V} \sum_m \hat S_\alpha(m) e^{2\pi i m\cdot r} ,\end{aligned}$$ where the sum over $m$ runs over the infinite reciprocal lattice with vectors $m = L^{-1} u$ ($u\in\mathbb Z^3$), and the volume of the periodic unit cell is $V = |L|$. These sums can be computed efficiently using the fast Fourier transform in molecular simulations.[@uessm95] The results in this work were computed with our own implementation of the method, which is available with documentation in Ref. . For any fixed-volume ensemble with partition function $Z(\Gamma,L)$ (e.g. $\Gamma = n,T$ for canonical or $\Gamma = \mu,T$ for grand-canonical), and probability distribution $P(\tau | \Gamma,L)$ (where $\tau$ represents the combined positions and momenta of all molecules present), define an extended ensemble that characterizes its response to a set of externally applied fields, $\phi = \{\phi_\alpha\}$, using the partition function, $$\begin{aligned} \Theta(\phi, \Gamma, L) &= Z(\Gamma, L) \int e^{{\ensuremath{ \left\langle \hat \rho, \phi \right\rangle }}} \; P(\tau | \Gamma, L) d\tau \\ {\ensuremath{ \left\langle X,Y \right\rangle }} &\equiv \sum_\alpha \int_V X_\alpha(r)^\dagger Y_\alpha(r) dr \notag \\ &= \frac{1}{V} \sum_\alpha \sum_m \mathcal F[X_\alpha](m)^\dagger \mathcal F[Y_\alpha](m) \label{e:inner}.\end{aligned}$$ The last two lines define a convenient notation for inner products between two functions of $r$ in the unit cell and state the appropriate Poisson summation formula.[@lax] We also define the function, $$\begin{aligned} \tilde \phi_\alpha(m) &\equiv \mathcal F[\phi_\alpha](m) / V \\ \intertext{to absorb the normalization constant in Eq.~\ref{e:inner}. This way,} {\ensuremath{ \left\langle \hat \rho, \phi \right\rangle }} &= \sum_\alpha \sum_m \hat S_\alpha^\dagger(m) \tilde \phi_\alpha(m), \\ \text{and } \tilde \phi_\alpha(0) / \beta &= \mu_\alpha,\end{aligned}$$ the chemical potential of species $\alpha$. The first and second derivatives of $\ln \Theta$ generate the densities and correlations, $$\begin{aligned} {\ensuremath{\frac{\partial{\ln \Theta}}{\partial{\tilde \phi_\alpha(m)^\dagger}}}} &= {\ensuremath{\left\langle{ \hat S_\alpha(m)}\right\rangle}} \equiv S_\alpha(m) \label{e:S} \\ {\ensuremath{\frac{\partial{^2 \ln \Theta}}{\partial{\tilde \phi_\alpha(m)^\dagger \tilde \phi_\gamma(m')}}}} &= {\ensuremath{\frac{\partial{S_\alpha(m)}}{\partial{\tilde \phi_\gamma(m')}}}} = V \delta_{m,m'} Q_{\alpha\gamma}(m) \label{e:corr} \\ Q_{\alpha\gamma}(m) &\equiv \frac{1}{V} {\ensuremath{\left\langle{ \Delta \hat S_\alpha(m) \Delta \hat S_\gamma(m)^\dagger }\right\rangle}} \\ &= \mathcal F[{\ensuremath{\left\langle{\hat\rho_\alpha(r) \hat\rho_\gamma(0)}\right\rangle}}](m) - \delta_{m,0}V\rho_\alpha\rho_\gamma, \label{e:Qg}\end{aligned}$$ where $\Delta \hat S_\alpha(m) \equiv \hat S_\alpha(m) - S_\alpha(m)$. All quantities in Eqns. \[e:S\]-\[e:Qg\] are conditional on the macrostate $(\phi, \Gamma, L)$ and evaluated at points where $\phi$ is spatially uniform. The vanishing of $m \ne m'$ cross-derivatives in Eq. \[e:corr\] can be proven from translational symmetry in this case. The matrix $Q/\rho$ is the structure factor of scattering theory.[@jsala96] It is connected to the Fourier transform of the radial distribution function via Eqns. \[e:Qg\] and \[e:g\]. It is also related to the integral theory of solutions via the definition,[@pdt6] $$Q^{-1}_{\alpha\gamma}(r) = \delta(r) {\ensuremath{\frac{\partial{\beta \mu^\text{id}_\alpha}}{\partial{\rho_\gamma}}}} - c_{\alpha\gamma}(r) ,\label{e:OZ}$$ of the Ornstein-Zernike direct correlation function, $c_{\alpha\gamma}(r)$. Here, $\mu_\alpha^\text{id}$ is the chemical potential of an ideal gas, whose derivative is $\partial \beta\mu_\alpha^\text{id}/\partial\rho_\alpha = 1/\rho_\alpha$. Canonical Kirkwood-Buff Coefficients ------------------------------------- It is well-known that the matrix inverse of Eq. \[e:corr\] gives the shift in the potential required to maintain a small change in the density profile appropriate for an ensemble with fixed $(S, \Gamma, L)$. Eqns. \[e:S\]-\[e:Qg\] apply exactly to both canonical ($\Gamma = n,T$) and grand-canonical ($\Gamma = \mu,T$) ensembles. From this connection, the most relevant quantities to canonical KB theory are the fit coefficients, $c_0$ and $c_2$, for $$\lim_{m\to 0} Q_{\alpha\gamma}^{-1}(m | n,L) = {\ensuremath{\frac{\partial{\beta \mu^\text{id}_\alpha}}{\partial{\rho_\gamma}}}} + c_{0,\alpha\gamma}(0) + c_{2,\alpha\gamma} (2\pi m)^2 + O(m^4) .\label{e:fit}$$ There is no linear term since $Q$ must be an even function of $m$ by symmetry. For a sufficiently large number of molecules of type $\alpha$, the first coefficient provides the density dependence of their excess chemical potential, $$\lim_{L /L_c \to\infty} c_{0,\alpha\gamma}(n, V) = \beta V \left( \mu^\text{ex}_\alpha(n,V) - \mu^\text{ex}_\alpha(n - \delta_\gamma,V) \right) ,\label{e:dmuex}$$ where we define $$\beta\mu^\text{ex}_\alpha(n,V) \equiv -\ln {\ensuremath{\left\langle{ e^{-\beta \Delta U_\alpha} | n, V}\right\rangle}} ,$$ to be the test particle insertion free energy in the canonical ensemble.[@pdt2] The coefficient $c_{0,\alpha\gamma}$ also converges to $\partial \beta\mu_\alpha^\text{ex} / \partial\rho_\gamma$ in the thermodynamic limit. Convergence to the particle insertion free energy is faster though, because the Helmholtz free energy can be rigorously stated as the sum of $n$ successive test particle insertions, $$\beta A(n,V,T) = \beta A^\text{id}(n,V,T) + \sum_{j=0}^{n-1} \beta\mu^\text{ex}(j,V,T) ,$$ while the ideal term is indistinguishable from $n \ln \rho$ after $n \ge 100$. As further justification of the limit, we invoke the well-known result that density fluctuations must become uncorrelated as the wavelength, $1/m$ grows much larger than the correlation length.[@jlebo67] Thus, under these conditions $Q^{-1}(m)$ is continuous as $m$ approaches zero and its extrapolation becomes easier as the cell size increases and the range of interactions becomes shorter. Away from $m=0$, the matrix $Q$ reports on structural features that cannot be calculated from bulk thermodynamic data. The second coefficient, $c_2$, in Eq. \[e:fit\] provides the correlation length for the $\alpha\gamma$ radial distribution,[@davis9] $$L_c^{(\alpha\gamma)} = \sqrt{ \left | Q_{\alpha\gamma}(0) c_{2,\alpha\gamma} \right |} \label{e:clen} .$$ The coefficient $c_2$ is generally negative for oscillatory decay and positive for monotonic decay of the radial distribution, $g(r)$. The identification of Eq. \[e:dmuex\] as the proper limiting form for the canonical ensemble allows us to carry through the rest of the Kirkwood-Buff analysis unchanged. All the other derived quantities of Kirkwood-Buff theory can be stated by noting that the matrix $Q_{\alpha\gamma}(0|\mu, L)$ is identical to $B_{\alpha\gamma}$ defined in Eq. 9 of Ref. . The supplementary information provides explicit expressions. In the thermodynamic limit, the density-derivatives of the chemical potential are the same for the canonical and grand-canonical ensembles. However, at finite volume, the grand-canonical ensemble provides observables closer to the thermodynamic limit than the canonical. From this identification, we conclude that the most effective method for estimating KB integrals from canonical simulations is to use density fluctuations over the entire simulation cell to estimate $Q(m)$ and then to extrapolate the value of $Q(0)$. Our numerical results below indeed show this works, but caution that there are two sources of error. The first is due to the difficulty of extrapolating to $m=0$. The second is due to finite size corrections, which become increasingly important for successive density derivatives. This work shows that the first type of error can be entirely eliminated within the canonical ensemble when the simulation box size is larger than about ten correlation lengths (which can be easily estimated using Eq. \[e:clen\]). Finite Size Corrections ------------------------ Finite size corrections (which extrapolate a result at finite volume to the thermodynamic limit) are more severe for the canonical than the grand-canonical ensemble. This section derives relations connecting averages in the canonical and grand-canonical ensembles. It is usually assumed that the grand-canonical value is close enough to the thermodynamic limit that this ensemble correction is a good estimate of the finite size correction. The correction to the distribution function was given by Lebowitz [@jlebo67], $$\begin{split} Q_{\alpha\gamma}&(m | \mu, L) - Q_{\alpha\gamma}(m | n, L) \\ &= \frac{1}{2V} \sum_{\eta\zeta} Q_{\eta\zeta}(0 | \mu, L) {\ensuremath{\frac{\partial{^2}}{\partial{\rho_{\eta} \partial \rho_{\zeta}}}}} Q_{\alpha\gamma}(m | \mu, L) , \end{split} \label{e:finite}$$ which is correct to relative order $1/V$ unless $m=0$. At the point $m=0$, $Q_{\alpha\gamma}(0 | n, L) = 0$, and the entire contribution to $Q_{\alpha\gamma}(0 | \mu, L)$ comes from the second term. Eq. \[e:finite\] was shown to match the thermodynamic limit in Ref. . Using the same method, we derive in the appendix the correction to the excess chemical potential valid to first order in $1/V$, $$\begin{split} \beta \mu_\alpha^\text{ex}&(\mu,V) - \beta \mu_\alpha^\text{ex}(n,V) \\ &= \frac{1}{2V}\left( \frac{2}{\rho_\alpha} - Q^{-1}_{\alpha\alpha}(0) + \sum_{\gamma} {\ensuremath{\frac{\partial{Q^{-1}_{\alpha\gamma}}}{\partial{\beta\mu_\gamma}}}}(0) \right). \end{split} \label{e:dmu}$$ For a 1-component system, $Q(0) = \rho^2\kappa / \beta$ is proportional to the isothermal compressibility, $\kappa$, and the correction (Eq. \[e:dmu\]) specializes to the derivative of the excess compressibility, $$\beta \Delta \mu^\text{ex}(\mu,V) = \frac{-1}{2V\kappa }{\ensuremath{\frac{\partial{( \kappa - \beta/\rho )}}{\partial{\rho}}}} .\label{e:dmu2}$$ This result agrees with Ref. , where it was first derived and tested against the thermodynamic limit for hard-sphere fluids. Lattice Gas Model ================== To illustrate the convergence properties of the structure factor approach, we apply it to a simple one-dimensional, one-component periodic lattice gas on $L$ sites with nearest-neighbor interaction energy $\beta J$This system is isomorphic to the 1D Ising model.[@sfrie17] The operators, $\hat \rho$, are defined as indicator functions on the $L$ sites, and all Fourier transforms become discrete Fourier transforms over $L$. Otherwise, the theory above goes through as expected except for a change in the ideal chemical potential, [@jrein00] $${\ensuremath{\frac{\partial{\beta\mu^\text{id}}}{\partial{\rho}}}} = \frac{1}{\rho(1-\rho)} .\label{e:c1}$$ ![Discrete radial distribution functions. The marked points show canonical, closed simulation data, while the solid lines show corresponding grand-canonical radial distributions. The upper panel (a) shows the antiferromagnetic model ($x=-2$), while the lower panel (b) shows the ferromagnetic $x=2$. For both plots, the grand-canonical distributions overlap so as to be indistinguishable, showing strong insensitivity to the cell size, $L$.[]{data-label="f:gr"}](decay.eps){width="3in"} This system, along with several generalizations, were shown to have a strictly nearest-neighbor direct correlation function in a series of original studies,[@cborz87; @cteje87] yet it still poses a nontrivial estimation problem since the correlation length goes to infinity at strong coupling. It is ideal for illustrating estimation of the direct correlation function because the densities and pair distributions can be computed analytically in the grand-canonical ensemble using the transfer matrix method.[@davis] We also calculated the canonical ensemble densities and pair distributions exactly with the use of the absolutely convergent cluster summation technique of Ref. . The most important properties of this model depend on the parameter $y \equiv \pm\exp(-1/L_c)$, where $L_c$ is the correlation length over which the radial distribution functions decay exponentially. The complete expression for $y$ in terms of $\beta J$ and $\phi (= \beta\mu)$ is given in the supplementary information. It can vary from $-1$ in an antiferromagnetic system when $\beta J \to -\infty$ to $+1$ in a ferromagnetic system when $\beta J \to +\infty$. ![Inverse correlation function from Eq. \[e:Qg\] without any ensemble corrections. The marked points show canonical simulation data, while the smooth lines show the low-density limiting form, $-2\beta J \cos(2\pi m) +\text{const}$. As in Fig. \[f:gr\], (a) and (b) are antiferromagnetic and ferromagnetic, respectively.[]{data-label="f:sigma"}](sigma.eps){width="3in"} Fig. \[f:gr\] compares the radial distribution function computed in the canonical ensemble with the corresponding grand-canonical distributions (at the chemical potential $\phi = -\beta J$, for which ${\ensuremath{\left\langle{n | \mu, L}\right\rangle}} = L/2$). The convergence is slow with increasing $L$, even though the canonical correlation functions appear to flatten out to a constant at large distance. These shifts in $g(r | n, L)$ by a constant do not affect our analysis, since they merely change the structure factor at $m=0$ – which our analysis disregards. The interesting ensemble effect shown in this plot is that the $r=1$ point appears to have a different vertical shift than the large-distance asymptote. For the antiferromagnetic case (a), the short-range behavior is relatively insensitive to the ensemble correction. In contrast (and disregarding the smallest simulation size), ensemble dependence shows up most strongly at short range in the ferromagnetic case (b). Fig. \[f:sigma\] compares the inverse correlation function, $Q^{-1}(m)$ for the three systems above in the canonical and grand-canonical ensembles. The function $Q^{-1}$ can be calculated exactly and simplifies in the limit as $L/L_c \to \infty$, $$Q^{-1}(m|\mu) = \frac{1}{\rho(1-\rho)} \left( \frac{1-y}{1+y} - \frac{2 y}{1-y^2} \left( \cos(2\pi m)-1 \right) \right) \label{e:c2} .$$ Fig. \[f:extrap\] compares two estimates for the KB coefficients corresponding to extrapolating to $m=0$ following Eq. \[e:fit\] (panels a,b) and correcting for ensemble dependence, fitting every two successive lengths to an empirical expression, $$Q^{-1}_\text{extrap}(1/L) \sim Q_\infty^{-1}(0) + c/L ,\label{e:empire}$$ for panels (c,d). The latter two plots used Eq. \[e:empire\] rather than Eq. \[e:finite\] because it gave much faster convergence. Several coupling values, $\beta J$, and several simulation sizes are shown, and fall on a single curve when scaled by their correlation lengths. The lines in (a,b) plot Eq. \[e:dmuex\] for second derivatives (times $L$) of the Helmholtz free energy for the lattice gas as $n$ is varied. Its agreement with the extrapolated values shows that the canonical ensemble $Q^{-1}(0 | n,V)$ converges to \[e:dmuex\] first, and then converges to $\partial \beta\mu^\text{ex}/\partial\rho$ as $V$ increases. {width="90.00000%"} The error visible at $\beta J = -1$ (panel a) gives another important message for estimating the KB coefficients. Rather than fit to Eq. \[e:fit\], we should have subtracted off the known contribution to the direct correlation function, $c_\text{LR} = -2\beta J \cos(2\pi m)$, in Eq. \[e:OZ\] and fit the remainder. Since the correlation length is very short for this weakly coupled limit, the resulting estimate would give perfect agreement with the line for $\beta J = -1$ (not shown). The lower panels (c,d) show extrapolation to $1/L = 0$ using the assumption[@pkrug13] that ${\ensuremath{\frac{\partial{^2}}{\partial{\rho^2}}}} Q(m | \mu, L)$ is independent of $L$ in Eq. \[e:finite\] and that $Q(1/L) \sim Q(0)$. Attempts to combine both types of extrapolation, in either order, resulted in worse convergence (not shown). Comparing the ferromagnetic and anti-ferromagnetic cases shows that the oscillating radial distribution function of the anti-ferromagnetic system leads to better overall convergence. This is associated with better short-range behavior of $g(r)$ (Fig. \[f:gr\]a). The strong ensemble-dependence of the ferromagnetic $g(r)$ carries over to Fig. \[f:extrap\]a, where it appears as worse agreement at small inverse-distance, $m$. It is because of the slow convergence of canonical and grand-canonical chemical potentials that Fig. \[f:extrap\]b shows $Q(0 | n,L)$ is 10% lower than $Q(0 | \mu, L)$, even at very large lengths. Luckily, this issue appears to be limited to the 1D case. Lennard Jones Model ==================== This section presents results of applying the formalism to a supercritical Lennard-Jones (LJ) gas. To mirror the conditions of the lattice model, we fix the density to a near-critical value, $\rho^* \sigma^3 \approx 0.36$, and set $\beta^{-1} = k_B T = 1.5 \epsilon$. This is above the critical temperature, $T^* = 1.36 \epsilon/k_B$. In the lattice model analogy, the coupling constant, $|\beta J|$, would be $\epsilon / k_B (T - T^*)$ if the temperature varied while $\epsilon$ remained constant. Fluctuations in this system are interesting because they report on cluster formation processes.[@hmart96] After summarizing target ‘thermodynamic limit’ results computed in the grand-canonical ensemble and using analytical formulas, we present the scaling behavior in the canonical ensemble. The LJ gas has become an indispensable simulation model, for which analytical approximations many properties are available. The most important for the present case are the equation of state, which gives $Q(0)$ via the compressibility, and the radial distribution function, which gives $Q(m)$ via Eq. \[e:g\]. We compare our simulation results to the MBWR equation of state,[@jnico79; @jjohn93] which was parameterized from molecular dynamics and Monte Carlo simulation data, and includes the effects from tail corrections. We compare the radial distribution function with a numerical solution of the Percus-Yevick (PY) closure of the Ornstein-Zernike equation, computed following the method of Fries and Patey.[@pfrie86] Our grand-canonical Monte Carlo (GC) simulations were performed with Towhee,[@mmart13] at an excess chemical potential of $\mu^\text{ex} = -1.868 \epsilon$ in a cubic box of length $10.357\sigma$. Pair interactions were truncated at 5$\sigma$ after which long-range corrections were added for the energy and pressure. Simulations were run long enough to collect 10,000 configurations for analysis. The number of trial moves between configurations was 40,000, since the potential energy showed an exponential autocorrelation function with a decay time of 13,200 trial moves. Moves were selected randomly with 25% probability for particle insertion/deletion and 75% probability for translation. Because our GC simulation gave a slightly different density ($0.356 \sigma^{-3}$), the GC properties summarized in Table \[t:GC\] were obtained at $\rho = 0.36\sigma^{-3}$ by differentiating the partition function estimate, $$\begin{aligned} \beta P V &= \Theta(\beta\mu) \approx \Theta(\beta\mu_0) e^{C(\beta\mu - \beta\mu_0)} \\ C(\lambda | \mu, L) &\equiv \ln \sum_{n=0}^\infty f_n(\mu,L) e^{\lambda n} \label{e:C} ,\end{aligned}$$ where $f_n$ is the frequency of observing $n$ particles in the GC simulation. Eq. \[e:C\] is a maximum likelihood estimate appropriate for a large number of samples, $S$, when the probability for observing a histogram, $f$, is given by Sanov’s theorem. We computed the estimation error using the saddle-point approximation (taking the second derivative of the relative entropy) to yield the covariance, $${\ensuremath{\left\langle{\delta C(\lambda)\delta C(\lambda')}\right\rangle}} S = e^{C(\lambda + \lambda') - C(\lambda) - C(\lambda')} - 1 .$$ L/$\sigma$ $\mu^\text{ex} / \epsilon$ $\beta P/\rho$ $Q^{-1}(0)/\sigma^3$ $L_c / \sigma$ ${\ensuremath{\frac{\partial{Q^{-1}(0)}}{\partial{(\beta\mu)}}}}/\sigma^3$ ------------ ---------------------------- ---------------- ---------------------- ---------------- ---------------------------------------------------------------------------- 6.525 -1.887(3) 0.4597(11) 1.89$^*$ 0.70$^*$ 8.434 -1.886(3) 0.4670(7) 1.81 0.97 10.36 -1.884(3) 0.4697(5) 1.53 1.16 12.65 -1.882(3) 0.4720(4) 1.32 1.36 16.44 -1.882(3) 0.4730(3) 1.24 1.49 19.68 0.47367 1.27 1.50 25.30 0.47312 1.25 1.55 GC -1.878(3) 0.47591 1.084(9) 1.50 4.0(3) PY 1.28 1.77 4.0724 MBWR -1.854 0.48218 1.1330 3.3978 : Comparison of excess chemical potential, pressure, and its derivatives from canonical simulations (top 7 rows), and grand-canonical and equation of state calculations (bottom rows). Numbers in parentheses display the uncertainty in the last digit. $^*$Only the first two points were used in the extrapolation to $m=0$ for the smallest simulation size.[]{data-label="t:GC"} The MBWR equation of state is fairly accurate near the state point studied here, but predicts $Q^{-1}(0)$ that is too low by $0.06 \sigma^{3}$ and $\partial Q^{-1}(0)/\partial (\beta\mu)$ that is too low by $0.7 \sigma^{-3}$. Although its parameterization[@jjohn93] did not include the particular state point studied here, it can be noted that the MBWR slightly over-estimates $\beta P / \rho$ in this region. That quantity reaches a minimum near the present simulation conditions, and it is the shape of the minimum which the derivatives in Table \[t:GC\] report on. We estimated the finite-size correction for $Q(m)$ in Eq. \[e:finite\], from GC, PY and MBWR methods and all three agree on $\frac{Q(0)}{2} {\ensuremath{\frac{\partial{^2 Q(0)}}{\partial{\rho^2}}}} = -20$, although the error in the GC estimate is 50%. Canonical simulations were performed with Towhee for $L=6.525\sigma$ to $L=16.44\sigma$. The smallest two simulations used pair interaction cutoffs of $3.25\sigma$ and 4.2$\sigma$, respectively. All larger cell sizes used a cutoff of 5$\sigma$. All results include long-range corrections for the energy and pressure. The largest two simulations were run with LAMMPS using a timestep of $0.003$ LJ units and collecting 10$^4$ samples – each separated in time by 10$^4$ dynamics steps. The $Q(0)$ and $L_c$ values reported in Table \[t:GC\] were fit to the smallest 3 $m$-points following Eq. \[e:fit\]. For the $L=6.525\sigma$ simulation, only the smallest 2 points were used in the fit. ![Estimates of $Q(0)$ using the radial distribution function integral (Eq. \[e:Gint\]) from different simulation sizes. The value at the origin is the ideal gas contribution, $\rho = 0.36\sigma^{-3}$, while the final estimate must be positive. Dotted lines show the partial integral without the correction of Ref. . The inverse of the final values (labeled on the plot) should be compared to Tbl. \[t:GC\]. The PY number is smaller than 1.28 due to the integral cutoff.[]{data-label="f:Gint"}](Gint.pdf){width="45.00000%"} Figure \[f:Gint\] shows a running integral of the radial distribution function from several of the simulations. The solid lines use the correction proposed by Krüger et. al.,[@pkrug13] $$Q_\text{est} = \rho + \rho^2 \int_0^{L/2} 4\pi r^2 (g(r)-1) \left(1 - (\tfrac{r}{L/2})^3 \right) ,\label{e:Gint}$$ where $L$ is the length of one side of the simulation cell. The dotted lines were not corrected (setting the $1 - 2r/L$ term to $1$ in Eq. \[e:Gint\]). The figure also shows that the PY closure yields a reasonable estimate of the RDF up to $4\sigma$, while the remainder is not possible to judge based on our simulation sizes. The long correlation length of this near-critical point system makes for a challenging test case. ![Estimates of $Q(m)$ from the Lennard-Jones system for the same simulations shown in Fig. \[f:Gint\]. The $L = 25.30\sigma$ line shows artificial damping at large $m$ caused by the smoothing spline FFT method on our fixed 128$^3$ grid. Otherwise, all the lines in (a) overlap one another. In (b), only the smallest three $k$-points from each simulation are shown, and the horizontal axis is scaled to show $(2\pi m)^2$, anticipating the role of the correlation length (Eq. \[e:clen\]). The dotted line plots Eq. \[e:uessm\] plus a constant using $\eta = 1$.[]{data-label="f:LJ"}](zero.eps){width="45.00000%"} Fig. \[f:LJ\]a plots the structure factor, $Q$, using spectroscopy units, $k = 2\pi m$. It should be compared with Figs. 3 and 4 of Ref. , where the relationship is $\rho^2 \tilde h(k) = Q(k/2\pi) - \rho$. That reference shows the shape of $Q(m)$ near $m=0$ was very flat, and easy to extrapolate. The present case is much more difficult because it is near the critical point and a very large peak has appeared near $m=0$. Fig. \[f:LJ\]b shows the inverse of Fig. \[f:LJ\]a near the origin. It is the analogue of Fig. \[f:sigma\] for the lattice gas. For clarity, only the first three points with smallest $m$ from each simulation are shown. These particular $m$-points correspond to averages over plane wave shaped density fluctuations oriented along faces, edges, and corners of the cubic unit cell. All five of the canonical simulation points shown at $m=0$ are extrapolations from these points. Their numerical values are in Table \[t:GC\]. Fig. \[f:LJ\]b also shows two other important points. First, $Q(m)$ computed from the largest simulation size is artificially noisy and scaled downward at high wavenumbers. The scale artifact is due to the inaccuracy of B-spline smoothing past $K_\text{max}/2$, where $K_\text{max} = 128/L$ is set by our numerical grid, which contained 128$^3$ points for every $L$. A similar artifact also appears for the $L=12.65$ simulation just past the right boundary of Fig. \[f:LJ\]b. The noise is caused by the discrete nature of the 128$^3$ grid points, which appear at irregularly spaced distances from the origin. It could be eliminated by resamping the data to uniform intervals in $|m|$. Second, the long-range portion of the LJ potential can be Fourier-transformed according to Ref.  to give, $$c_\text{LR}(x = m/\eta) = \frac{4\epsilon}{3} \left[ (1-2x^2) e^{-x^2} + 2 x^3 \sqrt{\pi} {\operatorname{erfc}}(x) \right] ,\label{e:uessm}$$ where $\eta$ is an inverse distance corresponding to a division between short-range and long-range forces. Recognizing that this term makes an additive contribution to Eq. \[e:OZ\] can accelerate the convergence of the fit in Eq. \[e:fit\]. Indeed, we find that subtracting this term from $Q^{-1}$ makes the first few points much more linear. The result can be visualized from Fig. \[f:LJ\]b as the distance between the red and black lines. However, to keep our discussion simple, we have not included it in our estimates. Similar to the antiferromagnetic lattice gas in Fig. \[f:extrap\]a, the estimates of $Q(0)$ are larger than the grand-canonical value, and decrease with increasing cell size. The former figure showed that the extrapolated points, $Q(0)$, quickly converged to second differences of the Helmholtz free energy (Eq. \[e:dmuex\]). We cannot test this numerically here, since the MD estimates of $\beta\mu^\text{ex}$ at finite volumes have uncertainties on the order of the volume. Instead, we have tested the finite-size corrections to $\beta\mu^\text{ex}$ in Eq. \[e:dmu\]. Using the value of $Q^{-1}$ and its derivative from Table \[t:GC\] gives $\Delta \beta \mu^\text{ex} = 4.18\sigma^3 / V$. This is a good fit to the size-dependence of $\mu^\text{ex}$ seen in simulations. In the limit as $m \to 0$, the finite-size corrections to $Q$ and to $\beta\mu^\text{ex}$ are related by a density derivative. Computing the finite-size correction to $Q(0)$ from Eq. \[e:finite\] gives $\lim_{m\to 0} \Delta Q(m) = -20 / V$, which agrees with the density derivative of Eq. \[e:dmu\]. The simulation points do not allow a single estimation of the slope near $m=0$ for the largest simulations, but the slope of the $c_0$ values vs. $1/V$ appears to be +200 rather than -20 for simulations at intermediate cell-sizes. ![Contour plot of the $1/V$ coefficient in the finite-size correction to the free energy, $\partial(\kappa - \beta/\rho)/\partial\rho / 2\kappa$ (Eq. \[e:dmu2\]) as calculated using the MBWR equation of state. Contours are shown every 2$\sigma^3$, while variations above $\pm 10 \sigma^3$ were colored at the max/min values. The axes are LJ temperature and density. Values in the liquid-vapor coexistence region (bottom, under the saturation curve) have been set to zero for clarity.[]{data-label="f:finite"}](FEcorr.pdf){width="45.00000%"} Because finite size corrections are relatively more important for canonical than grand-canonical simulations, we have plotted the finite size correction to the free energy predicted by the MBWR equation of state in Fig. \[f:finite\]. The result reassures us that the chemical potential calculated by test particle insertion has negligible volume dependence in both gas and liquid states.Near the solidification line, test particle insertion should be expected to slightly overestimate the excess chemical potential, while we get an underestimation for high-temperature liquids near the critical point. Conclusions ============ Kirkwood-Buff integrals are notoriously difficult to calculate from canonical, closed simulations. We have shown that this problem can be helpfully mapped onto the problem of estimating the direct correlation function of Ornstein-Zernike theory in a canonical ensemble. All earlier work used truncated real-space estimates, for which the process of extrapolating to infinite truncation radius, $R$, is ill-posed.[@nmatu96; @pgang13] Our exact theory for the canonical ensemble uses the entire simulation volume at once, is manifestly invariant to shifts of the radial distribution function by a constant, and replaces the problem of extrapolating to infinite $R$ with the well-posed problem of extrapolating to zero $m$ within a finite simulation volume. The density-dependence of the radial distribution function becomes important when estimating the thermodynamic limit.[@jlebo67; @jsala96] These higher-order derivatives are very difficult to estimate accurately from simulation data, and it is simpler to test for finite size effects by scaling up the system.[@pkrug13] This work showed that the principle difficulty in reaching the thermodynamic limit is the size-dependence of the canonical ensemble itself. We then provided a simple, general formula for estimating finite-size corrections of the excess chemical potential from KB coefficients. These corrections are small for the LJ fluid. Our data in Fig. \[f:LJ\] show excellent agreement for the entire correlation function at low wave-vectors in reciprocal space. Higher wave-vectors require smaller grid-spacing to avoid numerical artifacts, though. Fits using Eq. \[e:fit\] provide an estimation of both the KB coefficient and the correlation length which are more accurate than radial-distribution function integrals and fit well into the context of integral theories of solution. We have shown that the theory is robust by treating difficult cases with long-range correlations. These can be identified in simulation work from the shape of $Q(m)$ near $m=0$. We have shown that the zero-frequency limit of the canonical KB theory gives second derivatives of the Helmholtz free energy for simulation sizes larger than about 5 times the correlation length. Further, this theory has several nice properties. In the low density limit, the direct correlation function can be predicted from the long-range form of the interaction energy function. This leads to even better extrapolation to zero frequency. Further information on the applicability of this process is given in the supplementary material. Alternatively, the direct correlation function can be computed from simulation data and used to test common assumptions for OZ closure relations.[@pdt6] A companion paper uses this process to compute the spatial dielectric response of water.[@droge18a] This use of $Q$ to estimate pair interaction energies was first proposed by Madden and Rice [@wmadd80]. Supplementary Material {#supplementary-material .unnumbered} ======================= The supplementary material includes derivations and extended discussion of several formulas from the main text. It provides a definition of the finite-volume radial distribution, free from cutoffs, and consistent with Eq. \[e:g\]. It also provides explicit expressions for the other derived quantities of Kirkwood-Buff theory in terms of the matrix $Q_{\alpha\gamma}(0|\mu)$. It derives the ensemble correction to the free energy (Eq. \[e:dmu\]). Then it gives computable expressions for the correlation function of the finite-size Ising model and an approximate expression for the correlation length of the Lennard-Jones fluid. Both correlation lengths are plotted as a function of state point. Those results provide additional intuitive insight on the long-range correction proposed above. Acknowledgments {#acknowledgments .unnumbered} ================ This work was supported by the USF Research Foundation and NSF MRI CHE-1531590. [10]{} John G. Kirkwood and Frank P. Buff. The statistical mechanical theory of solutions. i. , 19(6):774–777, 1951. Nobuyuki Matubayasi and Ronald M. Levy. Thermodynamics of the hydration shell. 2. excess volume and compressibility of a hydrophobic solute. , 100(7):2681–2688, 1996. Pritam Ganguly and Nico F. A. [van der Vegt]{}. Convergence of sampling [Kirkwood-Buff]{} integrals of aqueous solutions with molecular dynamics simulations. , 9(3):1347–1355, 2013. M. Rovere, D. W. Heermann, and K Binder. The gas-liquid transition of the two-dimensional lennard-jones fluid. , 2:7009, 1990. Sondre K. Schnell, Xin Liu, Jean-Marc Simon, André Bardow, Dick Bedeaux, Thijs J. H. Vlugt, and Signe Kjelstrup. Calculating thermodynamic properties from fluctuations at small scales. , 115(37):10911–10918, 2011. R. Cortes-Huerto, K. Kremer, and R. Potestio. Communication: [Kirkwood-Buff]{} integrals in the thermodynamic limit from small sized molecular dynamics simulations. , 145:141103, 2016. Peter Krüger, Sondre K. Schnell, Dick Bedeaux, Signe Kjelstrup, Thijs J. H. Vlugt, and Jean-Marc Simon. integrals for finite volumes. , 4(2):235–238, 2013. J. J. Salacuse, A. R. Denton, and P. A. Egelstaff. Finite-size effects in molecular dynamics simulations: Static structure factor and compressibility. i. theoretical method. , 53:2382–2390, 1996. J. J. Salacuse, A. R. Denton, P. A. Egelstaff, M. Tau, and L. Reatto. Finite-size effects in molecular dynamics simulations: Static structure factor and compressibility. ii. application to a model krypton fluid. , 53(3):2390–2401, 1996. U. Essmann, L. Perera, M. L. Berkowitz, T. Darden, H. Lee, and L. G. Pedersen. A smooth particle mesh [Ewald]{} method. , 103:8577–8592, 1995. David M. Rogers. . GitHub, 2017. https://github.com/frobnitzem/EwaldCorrel. Peter D. Lax. , chapter 30, page 348. Wiley-Interscience, 2002. T. L. Beck, M. E. Paulaitis, and L. R. Pratt. , chapter 6, pages 123–141. Cambridge, New York, 2006. T. L. Beck, M. E. Paulaitis, and L. R. Pratt. , chapter 2, pages 23–31. Cambridge, New York, 2006. J. L. Lebowitz, J. K. Percus, and L. Verlet. Ensemble dependence of fluctuations with application to machine computations. , 153(1):250–254, Jan 1967. H. Ted Davis. , chapter 9, pages 442–448. VCH Publishers, New York, 1996. J. I. Siepmann, I. R. McDonald, and D. Frenkel. Finite-size corrections to the chemical potential. , 4(3):679–691, 1992. Sacha Friedli and Yvan Velenik. , chapter 3, pages 97–183. Cambridge University Press, Cambridge, UK, 2017. J. Reinhard, W. Dieterich, P. Maass, and H. L. Frisch. Density correlations in lattice gases in contact with a confining wall. , 61(1):422–428, 2000. C. Borzi, G. Ord, and J. K. Percus. The direct correlation function of a one-dimensional [Ising]{} model. , 46(1–2):51–66, 1987. Carlos F Tejero. One-dimensional inhomogeneous [Ising]{} model: A new approach. , 48(3):531–538, 1987. H. Ted Davis. , chapter 3, pages 128–130. VCH Publishers, New York, 1996. Juraj Vavro. Exact solution for the lattice gas model in one dimension. , 63:057104, 2001. Hernan L. Martinez, R. Ravi, and Susan C. Tucker. Characterization of solvent clusters in a supercritical [Lennard-Jones]{} fluid. , 104:1067, 1996. J. J. Nicolas, K. E. Gubbins, W. B. Streett, and D. J. Tildesley. Equation of state for the [Lennard-Jones]{} fluid. , 37(5):1429–1454, 1979. J. Karl Johnson, John A. Zollweg, and Keith E. Gubbin s. The [Lennard-Jones]{} equation of state revisited. , 78(3):591–618, 1993. P. H. Fries and G. N. Patey. The solution of the [Percus-Yevick]{} approximation for fluids with angle-dependent pair interactions. , 85(12):7307–7311, 1986. Marcus G. Martin. : a tool for [Monte Carlo]{} molecular simulation. , 39(14–15):1212–1222, 2013. John D. Weeks, David Chandler, and Hans C. Andersen. Role of repulsive forces in determining the equilibrium structure of simple liquids. , 54:5237, 1971. David M. Rogers. Fluctuation theory of ionic solvation potentials. 2018. submitted. William G. Madden and Stuart A. Rice. The mean spherical approximation and effective pair potentials in liquids. , 72:4208, 1980. Ensemble Dependence ==================== Explicit expressions for the radial distribution functions help to show how ensemble dependence arises in the KB integral. Eq. 1 of the main text can be written as an average over canonical simulations, $$\begin{aligned} {\ensuremath{\left\langle{\hat \rho_\alpha(r) \hat \rho_\gamma(0) | \mu, L}\right\rangle}} &= \sum_{n} P(n | \mu, L) {\ensuremath{\left\langle{\hat \rho_\alpha(r) \hat \rho_\gamma(0) | n, L}\right\rangle}} \notag \\ {\ensuremath{\left\langle{\hat \rho_\alpha(r) \hat \rho_\gamma(0) | n, L}\right\rangle}} &= \frac{n_\gamma}{V} \sum_{j=1}^{n_\alpha} P(r_{\alpha,j} = r | r_{\gamma,1} = 0, n, L) \label{e:n2}\end{aligned}$$ The conditional probability in the last line gives the probability that the $j^\text{th}$ molecule of type $\alpha$ has its center of mass at $r$ given that the first molecule of type $\gamma$ is centered at the origin. When $\alpha = \gamma$, this has the effect of decomposing the sum over $j$ into $j=1$, where we know $r_j$ is already at the origin, and $j \ne 1$, where $j$ is distinct from the molecule at the origin. This gives a delta function at the origin plus $n_\alpha - 1$ times the pair distribution function for fixed $n$. Explicitly, $$P(r = r_{\alpha,j} | r_{\gamma,1} = 0, n, L) = \begin{cases} \delta(r), & \alpha = \gamma, j = 1 \\ g_{\alpha\gamma}(r | n, L)/V, & \text{o.w.} \end{cases} \label{e:Pr2}$$ and $${\ensuremath{\left\langle{\hat \rho_\alpha(r) \hat \rho_\gamma(0) | n, L}\right\rangle}} = \frac{n_\gamma}{V} \left( \frac{n_\alpha - \delta_{\alpha\gamma}}{V} g_{\alpha\gamma}(r | n, L) + \delta_{\alpha\gamma}\delta(r) \right) \label{e:gV}$$ The identification of $g(r | n, L)$ in Eq. \[e:gV\] agrees with its method of computation from canonical ensemble simulation data by estimating the conditional probability, $P(r = r_{\alpha,j} | r_{\gamma,1} = 0, n, L)$. Of course, when the two particles are independent, the conditional probability is $1/V$, and $g(r | n, L)$ is 1. However, even for real molecules with excluded volume, $g(r | n, L)$ always integrates to the volume, $V$. This makes the integral of $g(r | n, L)$ uninformative for number fluctuations – since there are none in a canonical simulation. Explicitly, if we insert Eqns. \[e:gV\] and \[e:n2\] into Eq. 1 and integrate, then the $g(r | n, L)$ term drops out and the entire Kirkwood-Buff integral is due to the ensemble correction. For completeness, we re-state the derived quantities of Kirkwood-Buff theory in terms of the $\nu\times\nu$ matrix ${\bf Q} = [Q_{\alpha\gamma}(0|\mu,L)]$ and using the vector of $\nu$ densities, $\rho$, below. $$\begin{aligned} \kappa/\beta &= 1 / \rho^T {\bf Q}^{-1} \rho \\ \bar V &= \frac{\kappa}{\beta} {\bf Q}^{-1} \rho \\ \beta V \left[ {\ensuremath{\frac{\partial{\mu_\alpha}}{\partial{n_\gamma}}}} \right]_{n,P,T} &= {\bf Q}^{-1} - \frac{\beta}{\kappa} \bar V \bar V^T \\ \left[ {\ensuremath{\frac{\partial{\Pi}}{\partial{\rho_\alpha}}}} \right]_{\mu_1,\rho_{\gamma \ne 1},T} &= {\bf Q_2}^{-1} \rho_2\end{aligned}$$ These formulas express the isothermal compressibility, vector of partial molar volumes, matrix of constant-pressure chemical potential derivatives, and vector of $\nu-1$ osmotic pressure derivatives (respectively). The derivative of the osmotic pressure requires a special notation, ${\bf Q_2}$, for the $\nu-1\times\nu-1$ sub-matrix of ${\bf Q}$ discarding the first row and column, and a similar notation, $\rho_2$ for the $\nu-1$ component vector of solute densities. Note that the entries are discarded from $\bf Q$ before inverting. The equation for $\partial \mu/\partial \rho$ at constant temperature and pressure requires using the relation $dn_1 \bar V_1 + dn_2 \bar V_2 = 0$ for a constant pressure variation of $n_1$ at constant (arbitrary) volume, which does not generalize beyond 2-components. Ensemble Correction for the Excess Chemical Potential ====================================================== The exponent of the excess chemical potential can be written in the grand-canonical ensemble as, $$e^{-{\beta\mu^\text{ex}_{\alpha}}} = e^{\beta(\mu^\text{id}_\alpha - \mu_\alpha)} \equiv Z ,$$ which we call $Z$ for compactness. The ensemble correction for $Z$ requires alternating derivatives,\[15\]$$Z(\mu) = Z(n) + \frac{V}{2 V} \sum_\gamma {\ensuremath{\frac{\partial{}}{\partial{(\beta\mu_\gamma)}}}} {\ensuremath{\frac{\partial{}}{\partial{n_\gamma}}}} Z(\mu) , \label{e:corr}$$ evaluating the first at constant chemical potentials ($\mu_\alpha$) and the second at constant densities ($\mu_\alpha^\text{id}$). The volume is constant in both, and can be used to transform $n_\gamma / V = \rho_\gamma$. The first derivative is, $${\ensuremath{\frac{\partial{}}{\partial{\rho_\gamma}}}} Z(\mu) = Z(\mu) \left( {\ensuremath{\frac{\partial{{\beta\mu^\text{id}_{\alpha}}}}{\partial{\rho_\gamma}}}} - Q^{-1}_{\alpha\gamma}(0) \right) .$$ The next step requires the other derivative, $${\ensuremath{\frac{\partial{}}{\partial{(\beta\mu_\gamma)}}}} Z(\mu) = Z(\mu) \left( {\ensuremath{\frac{\partial{{\beta\mu^\text{id}_{\alpha}}}}{\partial{\rho_\alpha}}}} Q_{\alpha\gamma}(0) - \delta_{\alpha\gamma} \right) ,$$ which uses the fact that $\mu^\text{id}_\alpha$ does not depend on any densities other than $\rho_\alpha$. Using these in Eq. \[e:corr\] yields, (to order $1/V$) $$\frac{Z(\mu)}{Z(n)} = 1 + \frac{1}{2V} \left( - 2 {\ensuremath{\frac{\partial{{\beta\mu^\text{id}_{\alpha}}}}{\partial{\rho_\alpha}}}} + Q_{\alpha\alpha}(0) \left[ \left({\ensuremath{\frac{\partial{{\beta\mu^\text{id}_{\alpha}}}}{\partial{\rho_\alpha}}}} \right)^2 + {\ensuremath{\frac{\partial{^2{\beta\mu^\text{id}_{\alpha}}}}{\partial{\rho_\alpha^2}}}} \right] + Q^{-1}_{\alpha\alpha}(0) - \sum_\gamma {\ensuremath{\frac{\partial{Q^{-1}_{\alpha\gamma}}}{\partial{(\beta\mu_\gamma)}}}} \right) .\label{e:corr2}$$ Since $\partial{\beta\mu^\text{id}_{\alpha}} / \partial \rho_\alpha = 1/\rho_\alpha$, the $Q_{\alpha\alpha}(0)$ term cancels. The final ensemble correction given in the main text is the first term in the expansion of $-\ln Z(\mu)/Z(n)$. Since $Q^{-1}$ is naturally a function of $\rho$, while its derivative in Eq. \[e:corr2\] is taken at constant $\rho$, it is useful to change the independent variables from $\mu$ to $\rho$, $${\ensuremath{\frac{\partial{Q^{-1}_{\alpha\gamma}}}{\partial{(\beta\mu_\gamma)}}}} = \sum_\delta Q_{\gamma\delta}(0) {\ensuremath{\frac{\partial{Q^{-1}_{\alpha\gamma}}}{\partial{\rho_\delta}}}} .$$ Another alternate formula is, $${\ensuremath{\frac{\partial{Q^{-1}_{\alpha\gamma}}}{\partial{(\beta\mu_\gamma)}}}} = -\sum_{ij} Q^{-1}_{\alpha i} {\ensuremath{\frac{\partial{Q_{ij}}}{\partial{(\beta\mu_\gamma)}}}} Q^{-1}_{j\gamma} .$$ Ising Model Distribution Function ================================== The lattice gas in the main text has potential energy function, $$\beta H(\hat \rho | L) = -x \sum_{j=1}^L \hat\rho_{j-1} \hat\rho_j ,$$ where we have defined $\beta J \equiv x$ and used periodic boundary conditions, $\hat\rho_L = \hat\rho_0$. To compute powers of the transfer matrix, $$T \equiv \begin{bmatrix} e^{\phi + x} & e^{\phi/2} \\ e^{\phi/2} & 1 \end{bmatrix} ,$$ we write it as $T = \alpha I + r_x \sigma_x + r_z \sigma_z$, where $\sigma_x$ and $\sigma_z$ are $2\times 2$ Pauli spin matrices and $$\alpha \equiv (e^{\phi + x} + 1)/2, \quad r \equiv [ e^{\phi/2}, 0, (e^{\phi + x}-1)/2 ]^T .$$ We also define the scale parameter, $$y \equiv \frac{\alpha - |r|}{\alpha + |r|}.$$ The partition function can then be written as, $$\begin{aligned} Z(\phi, L) &= {\operatorname{Tr}\left[ T^L \right] } = (\alpha + |r|)^L (1 + y^L) \notag \\ \frac{T^m}{(\alpha + |r|)^m/2} &= (1 + y^m) I + (1 - y^m) (r_x \sigma_x + r_z \sigma_z)/|r| , \\ \intertext{and it is easy to find expectation values of $\hat \sigma_z(j)$,} {\ensuremath{\left\langle{\hat \sigma_z(0)}\right\rangle}} &= {\operatorname{Tr}\left[ T^L \sigma_z \right] }/Z = \left( \frac{1 - y^L}{1 + y^L} \right) \frac{r_z}{|r|} \\ {\ensuremath{\left\langle{\hat \sigma_z(0) \hat \sigma_z(j)}\right\rangle}} &= {\operatorname{Tr}\left[T^{L-j}\sigma_z T^j\sigma_z \right] }/Z \notag \\ &= \frac{r_z^2}{|r|^2} + \frac{r_x^2}{|r|^2} \frac{ y^j + y^{L-j}}{1 + y^L} \label{e:gj} .\end{aligned}$$ These can be transformed into average densities and radial distribution functions by substituting, $\hat \sigma_z(j) = 2 \hat \rho(j) - I$. ![Scale of the cosine term in Eq. 18 compared with the correlation length, $L_c$, and the interaction energy, $y/\rho(1-\rho) \sim x$ as $x\to 0$. Points mark the scale observed in closed simulations from $L=10$ to $L=512$. The largest $L$ values approach the line most closely.[]{data-label="f:gc"}](gc.eps){width="45.00000%"} Fig. \[f:gc\] shows the prefactor of the cosine-term in $Q^{-1}(m)$, $2y/(1-y^2)$. Since the prefactor dictates the correlation length estimated by Eq. 16 of the main text, it is important to compare to the analytical $L_c$ in this model. In the high temperature limit, $|x| \to 0$, and the prefactor simplifies to $2x \rho (1-\rho)$. At large coupling, the prefactor approaches the analytical correlation length, $L_c$. Both limits are shown in the figure. The finite-size effect on $L_c$ was calculated by estimating the prefactor for canonical, closed systems. This estimation is very nearly exact, since $Q^{-1}$ for closed systems maintains its cosine shape, even at high coupling. Those estimates are shown by crosses in the figure. All these estimates show a larger effective correlation at finite sizes. Our simulation lengths were spaced logarithmically in $L$, so the figure suggests that the prefactor scales linearly with $L/L_c$. Correlation Length of the Lennard-Jones System =============================================== The main text suggests estimating $Q(m)$ near $m=0$ by subtracting a long-range estimate for the direct correlation function, $$-\tilde c_{LR}(m) = \int e^{-2\pi i m\cdot r} u_{LR}(r) dr .\label{e:mod}$$ In a companion paper,\[30\]we show that this estimate is near quantitative accuracy for ionic systems away from any phase transitions. To apply this ansatz here, we modify the WCA approximation by using a fictitious reference system. WCA showed that the hard-sphere reference system gets the high-wavenumber behavior of $\tilde c(m)$ right. Since $\tilde c_{LR}(m)$ is small at high wavenumbers, we simply add it to the hard-sphere solution in the mean spherical approximation (MSA), $$c_A(r) = c_d(r; \rho\sigma^3) - \beta \epsilon u_{LR}(r) .\label{e:modA}$$ In variance with the main text, we let $u_\text{LR}$ be the long-range part of the WCA potential, $$u_{LR}(r) = \begin{cases} -1, & r < 2^{1/6} \\ 4(r^{-12} - r^{-6}), & r \ge 2^{1/6} \end{cases} ,$$ and $r$ is in units of $\sigma$. The function $c_d$ is the cubic polynomial on $[0,d)$ given by the MSA solution of the hard sphere system for a diameter chosen by a consistency condition on $Q(0)$. We then multiply the solution by $e^{-\beta\epsilon u_0}$ as described in Ref. \[29\]to smooth the first peak around $r = d$. The solution has better agreement with $\tilde c(m)$ at small wavenumber, but still predicts a first peak that is lower than simulation (and lower than the PY solution used in the main text). However, it is a good approximation for predicting the correlation length. It can be found from Eq. 20 in the main text, using $$\begin{aligned} 1 - \rho c_0 &\equiv \rho Q^{-1}(0) = \frac{\beta}{\rho\kappa} = 1 - 4\pi\rho \int_0^\infty r^2 c(r) dr \label{e:compr} \\ &= \frac{(2\eta + 1)^2}{(1 - \eta)^4} - \tfrac{32}{9} \pi\rho\beta\epsilon \sqrt{2} \label{e:cxn} \\ \rho c_2 &= \frac{4\pi\rho}{6} \int_0^\infty r^4 c(r) dr = \tfrac{48}{35} \pi\rho\beta\epsilon 2^{5/6} - \frac{d^2}{(1-\eta)^4}\Big( \\ &\quad (2\eta + 1)^2 (\tfrac{\eta}{20} - \tfrac{3}{16}) + 3(\eta + \tfrac{1}{8})/2 \Big) .\end{aligned}$$ The second part in Eq. \[e:cxn\] is the zeroth-order perturbation, $\int u_{LR}(r) dr$, which matches the corresponding term in the WCA perturbation energy.\[29\] ![Correlation length of the LJ system predicted from LR perturbation, Eqns. \[e:compr\]-\[e:cxn\]. The correlation length increases sharply around phase transition points. The model predicts a critical temperature that is too high and a liquid density line that is too dilute.[]{data-label="f:LJcorr"}](LJcorr.eps){width="50.00000%"} Fig. \[f:LJcorr\] plots Eq. \[e:compr\]-\[e:cxn\] for the correlation length along a few isotherms. The numerical value at $\beta^{-1} = 1.5\epsilon$ and $\rho = 0.36\sigma^{-3}$ is $L_c = 2 \sigma$, which is higher than simulation, but smaller than the PY approximation. The correlation length shows the basic features of the phase diagram, which is remarkable because the hard-sphere system does not have a liquid-vapor transition. Fig. \[f:LJcorr\] is based on an approximation to the WCA model, so its phase behavior is only qualitative. The qualitative agreement, however, allows us to conclude that Eq. \[e:mod\] is a good approximation to the small-wavenumber behavior of $\tilde c(m)$ and its inclusion will aid extraction of $\tilde c(m)$ from simulation. It also shows that $L_c$ is only few molecular diameters away from phase transitions. This short-ranged behavior translates to a small second derivative of $Q$, which makes estimation of $Q(0)$ possible from the method in the main text without corrections under ordinary circumstances.

--- abstract: 'The purpose of this paper is to expound and clarify the mathematics and explanations commonly employed in certain notable areas of astronomy and astrophysics. The first section concentrates upon the mathematics employed to represent and understand stellar structure and evolution. The second section analyses two different explanations for the structure of spiral galaxies.' author: - Gordon McCabe title: Mathematics and explanation in astronomy and astrophysics --- Stellar structure and evolution =============================== A star forms when a body of gas contracts under its own gravitational attraction, and the pressure and temperature created at the centre of the agglomerated mass is sufficient to ignite the nuclear fusion of atomic nuclei. The radiation released as a by-product of the nuclear fusion eventually reaches the surface of the star and becomes starlight. Newtonian astrophysics represents the gaseous content of a star as a fluid which occupies a compact open subset $\Omega \subset \mathbb{R}^3$ of three-dimensional Euclidean space. The fluid is considered to possess a mass density scalar field $\rho$ and a pressure scalar field $p$. To study stellar structure in Newtonian astrophysics, it is common to assume that the star is spherically symmetric and static (time-independent). The requirement that the star be static is equivalent to the requirement that the star is in hydrostatic and thermal equilibrium. For a static spherically symmetric star, the fluid occupies a solid ball of radius $r$, the mass density and pressure are time independent, and the velocity vector of the fluid vanishes. Using spherical polar coordinates, $(r,\theta,\phi)$, the pressure and the density are independent of the angular coordinates $(\theta,\phi)$, but vary as a function of the radius $r$. The fluid is assumed to be self-gravitating and to satisfy the Poisson equation of Newtonian gravitation, $$\nabla^2 \Phi = - 4 \pi G \rho \;,$$ with respect to a gravitational potential scalar field $\Phi$. $\nabla^2$ is the Laplacian, also denoted as $\Delta$ in many texts. There are four differential equations which govern the structure of a static, spherically symmetric star. The four equations taken together constitute a coupled (‘simultaneous’) set of ordinary first-order, *non-linear* differential equations. We now proceed to introduce and discuss these equations. Assuming spherical symmetry and time-independence, the mass density is a function of radius alone $\rho(r)$, and from the definition of mass density and Euclidean geometry, it follows that $m(r)$, the mass enclosed within the surface of radius $r$, is given by the expression: $$m(r) = \int_0^r 4 \pi (r')^2 \rho(r') dr' \;.$$ (The prime here simply indicates the use of a dummy variable for the integration.) This gives us our first differential equation for stellar structure, sometimes called the mass continuity equation: $$\frac{dm(r)}{dr} = 4 \pi r^2 \rho(r) \;.$$ The Newtonian gravitational potential for a self-gravitating, spherically symmetric body satisfying the Poisson equation, is $\Phi = -Gm(r)r^{-1}$. Given a gravitational potential $\Phi$, the gravitational force exerted per unit mass is specified by the vector field $ - \text{grad} \; \Phi$, and the gravitational force per unit volume is specified by $F = - \rho \; \text{grad} \; \Phi$. Most texts denote the gradient operator as $\nabla$, but this can be confused with the covariant derivative operator in differential geometry. Whilst the gradient of a scalar field is a contravariant vector field, the covariant derivative of a scalar field is a *covariant* vector field. Hence, to avoid confusion, we shall write ‘$\text{grad}$’ rather than $\nabla$. Given a choice of coordinates $(x_1,...,x_n)$ on a manifold, and given a metric tensor field $g_{ij} dx^i \otimes dx^j$ which specifies the geometry of space, the gradient of an arbitrary scalar field $f$ can be expressed as $$\text{grad} \; f = g^{ij} \frac{\partial f}{\partial x_i}\frac{\partial}{\partial x_j} \;,$$ where $g^{ij}$ is the inverse of the matrix $g_{ij}$. In the spherical polar coordinates we have chosen, $(x_1,x_2,x_3) = (r,\theta,\phi)$, the flat Euclidean metric takes the form $$ds^2 = dr \otimes dr + r^2 d\theta \otimes d \theta + r^2\sin^2\theta d\phi \otimes d\phi \;,$$ hence the gradient of $f$ takes the form $$\text{grad} \; f = \frac{\partial f}{\partial r}\frac{\partial}{\partial r} + r^2 \frac{\partial f}{\partial \theta}\frac{\partial}{\partial \theta} + r^2 \sin^2 \theta \frac{\partial f}{\partial \phi}\frac{\partial}{\partial \phi} \;.$$ Assuming spherical symmetry of the gravitational potential $\Phi$ entails that $\partial \Phi/\partial \theta$ and $\partial \Phi/\partial \phi$ vanish, hence $$\eqalign{\text{grad} \; \Phi &= \frac{d \Phi}{dr} \frac{\partial}{\partial r} \cr &= \frac{d (-Gm(r)r^{-1})}{dr} \frac{\partial}{\partial r} \cr &= Gm(r)r^{-2} \frac{\partial}{\partial r} \;.}$$ The requirement of hydrostatic equilibrium corresponds to the requirement that the inward gravitational force $ - \rho \; \text{grad} \; \Phi$ is balanced at each point by the outward pressure gradient force $ - \text{grad} \; p$:[^1] $$\text{grad} \; p + \rho \: \text{grad} \; \Phi = 0 \;.$$ Spherical symmetry entails that $\partial p/\partial \theta$ and $\partial p/\partial \phi$ vanish, hence $$\text{grad} \; p = \frac{dp}{dr}\frac{\partial}{\partial r} = -Gm(r)\rho(r) r^{-2}\frac{\partial}{\partial r} \;.$$ This gives us our second differential equation for stellar structure: $$\frac{dP(r)}{dr} = - \frac{Gm(r)\rho(r)}{r^2} \;.$$ Let $l(r)$ denote the amount of energy passing through the surface of radius $r$ per unit time. Let $\epsilon(r)$ denote the energy production coefficient, the amount of energy released at radius $r$ in the star per unit time and per unit mass. It follows that $4 \pi r^2 \rho(r) \epsilon (r)$ is the energy released per unit time at radius $r$. A star in thermal equilibrium is an open system for which there are large energy flows out of the system, and in which there are significant temperature and pressure gradients, but in which, nevertheless, the temperature and pressure profile of the star remain constant. For a star in thermal equilibrium, none of the energy released is used to heat up the star or change its volume. Hence, for such a star, the rate of energy flow at radius $r$ is given by the expression: $$l(r) = \int_0^r 4 \pi (r')^2 \rho(r') \epsilon(r') dr' \;.$$ This gives us our third differential equation of stellar structure, $$\frac{dl(r)}{dr} = 4 \pi r^2 \rho(r) \epsilon(r) \;.$$ For a star of radius $R$ the *luminosity* is defined to be $L = l(R)$, the amount of energy passing through the outer surface of the star per unit time. When a star resides in a state of thermal equilibrium, the total amount of energy produced by the star per unit time equals the amount of energy radiated from the outer surface per unit time. Hence, the luminosity $L$ of such a star is given by the equation: $$L = \int_0^R 4 \pi r^2 \rho(r) \epsilon(r) dr \;.$$ The energy per unit time $l(r)$ passing through the sphere of radius $r$ can be expressed as $$l(r) = 4 \pi r^2 F(r) \;,$$ where $F$ is the energy flux, the energy flow per unit time per unit area. When there is a temperature gradient in a body, energy flow will occur by conduction and radiation at the very least. Assuming, for simplicity, the absence of convection, the energy flux in a star can be broken down into a conductive energy flux $F_{cond}(r)$ and a radiative flux $F_{rad}(r)$. In both cases, the energy flow is proportional to the temperature gradient as follows (Tayler 1994, p63): $$F_{cond}(r) = - \frac{4acT^3(r)}{3\kappa_{cond}(r)\rho(r)}\frac{dT(r)}{dr} \;,$$ and $$F_{rad}(r) = - \frac{4acT^3(r)}{3\kappa_{rad}(r)\rho(r)}\frac{dT(r)}{dr} \;.$$ $\kappa_{cond}$, the coefficient of conductive opacity, measures the resistance to the flow of heat by conduction, and $\kappa_{rad}$, the coefficient of radiative opacity, measures the resistance to the flow of heat by radiation. $a$ is the radiation density constant, $a= 7.57 \times 10^{-15} \; \text{erg cm}^{-3} \text{K}^{-4}$. Assuming that energy flow in a star is due to conduction and radiation,[^2] one obtains (Tayler 1994, p64): $$l(r) = 4\pi r^2(F_{cond}(r) + F_{rad}(r)) = - \frac{16\pi acr^2T^3(r)}{3\kappa(r) \rho(r)}\frac{dT(r)}{dr} \;,$$ where $$\frac{1}{\kappa(r)} = \frac{1}{\kappa_{cond}(r)} + \frac{1}{\kappa_{rad}(r)} \;.$$ Re-arranging, one obtains the following differential equation for the temperature gradient in a star, which is the fourth equation governing stellar structure: $$\frac{dT(r)}{dr} = - \frac{3\kappa(r) \rho(r)}{4 ac T^3(r)} \frac{l(r)}{4 \pi r^2} \;.$$ The chemical composition of a spherically symmetric and static star can be specified by a set of functions $X_i(r)$, which specify the fraction of unit mass consisting of nuclei of type $i$, for $i = 1,...,I$. As such, $\Sigma_i X_i(r) = 1$.[^3] Before the equations of stellar structure can be solved, one must specify an equation of state for the pressure, $$P = P(\rho, T, X_1,...,X_I) \;,$$ an expression for the energy production coefficient $$\epsilon = \epsilon(\rho, T, X_1,...,X_I) \;,$$ and an expression for the opacity $$\kappa = \kappa(\rho, T, X_1,...,X_I) \;.$$ These expressions are sometimes called the constitutive equations. If these three functions are specified, and four boundary conditions are specified, then one can solve the four differential equations of stellar structure. The four boundary conditions can be chosen to be: $$m =0 \;, \quad l= 0 \quad \text{at} \;\; r =0 \;,$$ and $$\rho = 0 \;, \quad T = 0 \quad \text{at} \;\; r = R \;.$$ Note here that the star is idealised as an *open* solid ball of radius $R$, hence the boundary of the star at radius $R$ is not itself considered to be part of the star. Choosing $\rho = 0$ (or $P=0$) and $T = 0$ at a finite radius $r = R$ is merely a simple idealisation. In reality, the density, pressure and temperature of a star gradually descend to the non-zero value of the interstellar medium in the neighbourhood of the star. Solving the four differential equations for stellar structure then gives the mass, pressure, energy flow and temperature as a function of radius. However, in practical terms, one might wish to specify the total mass $M$ and chemical composition of a star, and then obtain, amongst other things, the radius of such a star in hydrostatic and thermal equilibrium. It is therefore conventional to re-cast the differential equations with mass $m$, rather than radius $r$, as the independent variable. Doing so, one obtains (Tayler 1994, p70-71): $$\frac{dr(m)}{dm} = \frac{1}{4 \pi r^2(m) \rho(m)} \;,$$ $$\frac{dP(m)}{dm} = - \frac{Gm}{4 \pi r^4(m)} \;,$$ $$\frac{dl(m)}{dm} = \epsilon(m) \;,$$ $$\frac{dT(m)}{dm} = - \frac{3\kappa(m) l(m)}{64 \pi^2 ac r^4(m) T^3(m)} \;.$$ The chemical composition of the star is then expressed in terms of functions $\{X_i(m): i=1,...,I \}$. One can then choose boundary conditions for the re-formulated equations of stellar structure, such as: $$r =0 \;, \quad l= 0 \quad \text{at} \;\; m =0 \;,$$ and $$\rho = 0 \;, \quad T = 0 \quad \text{at} \;\; m = M \;.$$ Again, one could choose $P =0$ at $m=M$ rather than $\rho = 0$, and, again, this is merely the simplest idealisation on offer. Now consider the relationship between theory and observation. The observational state of a star at a moment in time can be specified by just two parameters: its luminosity and ‘surface’ temperature. Let us take these in turn: 1. [If one calculates the distance to an observable star, say by parallax, then one can infer the luminosity of the star from its apparent luminosity.]{} 2. [Whilst the temperature of the surface of a star was idealised in the boundary conditions above to be zero, a more sophisticated model divides a star into its interior and atmosphere. The lowest level of the atmosphere is called the photosphere, and the ‘surface’ temperature of a star is deemed to be the temperature of the photosphere. The temperature of a star’s surface, in this sense, is intimately related to the spectral type of the star, defined by the absorption lines in the spectrum of light emitted by the photosphere. Given that the temperature of a star’s photosphere largely determines the type of its spectrum, the observational state of a star can be specified by its luminosity and spectral type. There are ten spectral types which, in order of decreasing temperature, are referred to as O, B, A, F, G, K, M, R, N, and S.]{} If one knows both the luminosity and temperature of a star, then one can calculate the radius of the star. The luminosity of a star is a function of the radius $R$ of the star and the effective[^4] surface temperature $T_e$ of the star, according to the following expression, $$L = 4 \pi R^2 \sigma T_e^4 \;,$$ where $\sigma$ is the Stefan-Boltzmann constant, related to the radiation density constant $a$ by $\sigma = ac/4$. Given $L$ and $T_e$, one can obviously use this expression to calculate $R$. The observational state of a star can be represented by a point on the Hertzsprung-Russell diagram, a two-dimensional rectangle, coordinatized by luminosity on the vertical axis, and either spectral type or temperature on the horizontal axis. The observational history of a star traces out a path on the Hertzsprung-Russell diagram. The luminosity and temperature of a star in hydrostatic and thermal equilibrium are the observational properties which can be explained by solving the differential equations of stellar structure. This explanation falls under the aegis of the deductive-nomological (D-N) account of scientific explanation. In such explanations, one explains certain phenomenal facts by logically deriving them from the conjunction of general laws and particular specified circumstances. In this case, the mass and chemical composition of a star are the specified circumstances, and given a specification of the constitutive equations, one can explain the luminosity and temperature of a star from the conjunction of the equations of stellar structure and the given mass and chemical composition. The initial mass and chemical composition of different stars vary, and these are the characteristics which determine the history and lifetime of a star on the Hertzsprung-Russell diagram. Stars which are burning hydrogen into helium, in a state of hydrostatic and thermal equilibrium, occupy a roughly diagonal channel, running from the top left to the bottom right of the Hertzsprung-Russell diagram, called the *main sequence*. The initial mass, and to a lesser extent, the initial chemical composition of a star, determine its initial location in the main sequence. Stars with greater initial mass are more luminous, and occupy positions further up the main sequence slope. The larger the mass and luminosity of a star, the shorter the length of time it will spend on the main sequence until its hydrogen fuel is expended. The Vogt-Russell ‘theorem’, conjectured independently by Heinrich Vogt (1926) and Henry Norris Russell (Russell *et al* 1927, p910), states that with the total mass and chemical composition specified, the equations of stellar structure admit a unique solution. If true, the Vogt-Russell theorem would entail that each mass and chemical composition corresponds to a unique *equilibrium* configuration. This, of course, is consistent with the fact that stars of the same mass but different chemical profile, can possess different equilibrium states. The chemical composition of a star will change during its lifetime as nuclear fusion converts lighter elements into heavier elements, and a star will tend to pass through a sequence of different equilibrium configurations. However, whilst the Vogt-Russell theorem is essentially true, it is not strictly true, and it is certainly no theorem. A system of four coupled, *non-linear*, ordinary differential equations, with boundary conditions specified at two different points, does not necessarily admit a unique solution. Whilst there is usually a unique equilibrium configuration for a given mass and chemical composition, for some combinations of mass and chemical composition there are multiple solutions to the equations of stellar structure. The notion that there was actually a ‘theorem’ was promulgated in textbooks, and never given a rigorous proof. As a conjecture, it is now known to be false. In fact, both the existence and uniqueness parts of the conjecture fail. For example, there is no solution composed of helium with a mass less than about 0.3 solar masses, and Cox and Salpeter (1964) demonstrated that there are two different solutions for stars of the same mass, burning purely helium, close to this minimum mass. This is referred to as the *double-valued* helium main sequence, in the sense that, for some range of masses, two values of radius $R$ are possible for each value of mass. Kähler (1978) tracked the post main sequence evolution of the helium core of a star of two solar masses, and found different equilibrium configurations of different radius, but with the same mass. Gabriel and Noels-Grötsch (1968) demonstrated that the minimum mass of a solution composed entirely of carbon is about 0.9 solar masses, and that, once again, there are two possible solutions for each mass value greater than the minimum mass. In one branch of such double-valued main sequences, the radius and luminosity increase with increasing mass, whilst in the other branch, there are smaller radii, and the luminosity decreases with increasing mass (Hansen 1978, p23). The latter, anomalous branches, correspond physically to the presence of electron degeneracy. It is therefore not strictly true to say that the mass and chemical composition of a star in equilibrium uniquely determine its structure. However, as Hansen *et al* point out, “the idea of uniqueness is still useful in that among a set of models all having the same mass and run of composition, usually only one seems to correspond to a real star or to have come from some realistic line of stellar evolution. The others are unstable in some fundamental way (as far as we know)," (2004, p331). Given a fixed mass and chemical composition, each solution to the differential equations of stellar structure corresponds to a pair of values $(P_c,T_c)=(P(0),T(0))$ for the central pressure and temperature of the star, or, equivalently, to a pair of values $(R,L)=(r(M),L(M))$ for the surface radius and luminosity. The significance of this is that the radius $R$ and luminosity $L$ of a star are both inferrable from observation. If the mass and chemical composition determined a unique solution, then the mass and chemical composition of a star in equilibrium would uniquely determine its radius and luminosity $(R,L)$. If, on the contrary, there is no unique solution for a particular combination of mass and chemical composition, then there are multiple corresponding pairs $\{(R,L)_i: i =1,...,n\}$. If one drops the requirements of hydrostatic and thermal equilibrium, then one can obtain the equations of stellar structure and evolution which govern the history of a Newtonian star. Assuming spherical symmetry, the equations are (Kippenhahn and Weigert 1990, p64): $$\frac{\partial r(m,t)}{\partial m} = \frac{1}{4 \pi r^2(m,t) \rho(m,t)} \;,$$ $$\frac{\partial P(m,t)}{\partial m} = - \frac{Gm}{4 \pi r^4(m,t)} - \frac{1}{4 \pi r^2(m,t)}\frac{\partial^2 r(m,t)}{\partial t^2} \;,$$ $$\frac{\partial l(m,t)}{\partial m} = \epsilon(m) - c_P \frac{\partial T(m,t)}{\partial t} + \frac{\delta(m,t)}{\rho(m,t)} \frac{\partial P(m,t)}{\partial t}\;,$$ $$\frac{\partial T(m,t)}{\partial m} = - \frac{3\kappa(m,t) l(m,t)}{64 \pi^2 ac r^4(m,t) T^3(m,t)} \;.$$ Notice that these are *partial* differential equations, whilst the equations of stellar structure are *ordinary* differential equations. $c_P$ is the specific heat at constant pressure and $\delta \equiv -(\partial \ln \rho/\partial \ln T)_P$, (see Kippenhahn and Weigert, p19). The fourth equation here continues to assume that energy transport is due to radiation and conduction alone, but this equation can be generalized. Time-dependence requires one to introduce additional equations for the time-dependence of the mass fractions $X_i$: $$\frac{\partial X_i(m,t)}{\partial t} = \frac{m_i}{\rho(m,t)}\left (\Sigma_j r_{ji} - \Sigma_k r_{ik} \right ) \;, \quad \quad i = 1,...,I \;.$$ $m_i$ is the mass of the nuclei of type $i$, and $r_{ij}$ is the rate at which nuclei of type $i$ are transformed into nuclei of type $j$ per unit volume. Solving these equations of stellar structure and evolution requires the specification of initial conditions as well as boundary conditions. The histories of different star types can intersect when plotted as paths on the Hertzsprung-Russell diagram. For example, after a star of around one solar mass has expended all the hydrogen in its core, it leaves the main sequence, increasing in radius and decreasing in temperature, to the effect of a net increase in luminosity. This evolutionary path takes the star up the Hertzsprung-Russell diagram to become a red giant. It then subsequently increases in temperature, crossing the main sequence itself from right to left, before it sheds its outer layers in the form of a so-called planetary nebula. Where this path crosses the main sequence, the mass of the star is considerably below the mass of stars at that point on the main sequence. If two different types of star can occupy the same point on the Hertzsprung-Russell diagram, it entails that knowledge of the luminosity and temperature of a star alone is not sufficient to determine the unique evolutionary history of a star. The Hertzsprung-Russell diagram does not provide a state space for stars in the sense of a state space (‘phase space’) in Hamiltonian mechanics. Nuclear fusion inside stars is responsible for creating almost all of the atoms in our universe which are heavier than hydrogen or helium. These heavier elements are generically called ‘metals’ by astrophysicists. During their lifetime, and particularly at the end of their lifetime, stars will eject a proportion of their mass into the interstellar medium, and this mass will contain, in some proportion, the metallic elements created by fusion. The mass returned into the interstellar medium is re-cycled in the formation of subsequent generations of stars. The stars in our galaxy are classified, in a coarse-grained fashion, as either Population $\rm{I}$ or Population $\rm{II}$. The Population $\rm{II}$ stars were the first generation of stars formed in our galaxy, and as such, are metal-poor stars. Population $\rm{I}$ stars were formed from an interstellar medium which already contained the metallic elements created by the first generation of stars, hence the Population $\rm{I}$ stars have a greater proportion of metallic elements in their chemical constitution. Galaxies ======== A galaxy is a collection of $\sim 10^{11}$ stars, which is gravitationally bound together, and which is $\sim 10^5$ light years in diameter. Stars within a galaxy are separated by distances of $\sim 10$ light years, (Rindler 2001, p350). There are two main types of galaxy: spiral and elliptical. Most of the stars in a spiral galaxy are concentrated in a flattened disk, which rotates about its center. The rotation, however, is not rigid, and the angular velocity of the rotation decreases as a function of radius. The average rotation period in a spiral galaxy is about $100$ million years, (ibid.). The disk of a spiral galaxy is surrounded by a more diffuse ‘halo’ of stars, distributed in a spherically symmetric fashion. The halo is populated only by Population $\rm{II}$ stars. Whilst successive generations of stars form throughout most of the disk in a spiral galaxy, there is typically a bulge at the center of the disk, and this bulge also tends to contain Population $\rm{II}$ stars. At the very center of the disk there is typically a supermassive black hole. A spiral galaxy possesses an interstellar medium which consists of gas and dust.[^5] The stars in the disk of a spiral galaxy are not actually distributed in spiral patterns. Rather, the spirals trace the brightest regions in the disk of the galaxy. The brightest regions are those which contain both the O- and B-type stars and the H$\rm{II}$ regions of ionized hydrogen, created by the ultraviolet radiation from those high-temperature O and B-type stars. Because the brightest (high mass) stars are also the stars with the shortest lifetimes, the bright high-mass stars can only be found in regions of recent star formation, and these regions happen to be the spiral arms. In an elliptical galaxy, the stars are uniformly distributed within a $3$-dimensional ellipsoid, and there is no overall rotation. In fact, the velocities of the stars in an elliptical galaxy are randomly distributed in all directions. Much like the halo of a spiral galaxy, there appears to be only one generation of star formation in an elliptical galaxy. Our own galaxy, the Milky Way, is a so-called barred-spiral, in which the spiral arms emanate from the ends of an apparently rigid bar of luminous matter, (Nicolson 1999, p206). There appear to be two main spiral arms in the Milky Way: the Norma arm and the Sagittarius arm. In addition, there are a number of smaller arms, and the Sun can be found within the Orion arm or ‘spur’, which lies outside the Sagittarius arm, and inside the Perseus arm (Nicolson p202). The Sun resides inside a ‘bubble’ three-hundred light years in diameter, in which the interstellar medium has been cleared by a supernova explosion (Smolin 1997, p124). From this perspective, of the $10^{11}$ stars in our galaxy, only about $7000$ can be seen with the naked eye, (Rindler p350). Moreover, only $1\%$ of the light which falls upon the Earth comes from beyond our galaxy, (Disney 2000, p4). The mutual gravitation of the stars in a galaxy define an escape velocity, and, at any one time, most of the stars in a galaxy will be moving with a velocity less than the escape velocity. In this sense alone, a galaxy cannot be treated as a collection of isolated, independent entities. Rather, a galaxy must be represented as a system, in the sense that it constitutes a collection of mutually interacting parts. In a spiral galaxy in particular, one can treat the stars as the members of a population, and the interstellar medium as the environment with which that population interacts, via feedback loops. In this sense, it has been suggested that a spiral galaxy can be treated as an ecological system, (Smolin 1997, Chapter 9). Perhaps, however, it would be more accurate to say that biological ecosystems and spiral galaxies are both instances of the same type of formal system. In other words, there is a type of formal system which contains a population, an environment, and a set of feedback loops between the population and the environment, and spiral galaxies provide instances of this system-type just as much as biological ecosystems do. Smolin asserts that “There are processes by which the matter of the interstellar medium is converted into stars and there are processes by which matter is returned from the stars to the interstellar medium. To understand what a galaxy is, and especially to understand it as a system, is then primarily to understand the processes that govern the flow of matter and energy between the stars and the interstellar medium," (1997 p118). He points out (p124) that the interstellar medium is a system far from thermodynamic equilibrium, consisting of a number of different components of different densities, temperatures, and compositions. The different components include normal atomic gas; cold, dense Giant Molecular Clouds (GMCs); and regions of hot, dilute plasma, otherwise known as H$\rm{II}$ regions, denoting ionized hydrogen. The relative amount of matter in these different components remains approximately constant over time. Smolin asserts that for a system to be in such a stable, but far-from-equilibrium state, there must be processes which cycle the material among the different components, and the rates of these processes must be controlled by feedback mechanisms. The arms of a spiral galaxy trail behind with respect to the direction of galactic rotation, hence the obvious explanation for their shape is simply the fact that the angular rotation velocity decreases with radial distance from the centre of a galaxy. However, at the very least, this *differential rotation* cannot be the sole explanation because spiral galaxies have existed for at least $10$ billion years, and given a typical rotation period of $100$ million years, this means that a typical spiral galaxy has completed at least $100$ rotations, and “would long ago have wrapped its arms to extinction if they had been created early in its history and had not been sustained by some ongoing process," (Nicolson 1999, p208). Observations also suggest that the spiral arms do not rotate with the galaxy, but at a slightly slower rate. Note carefully, however, that an explanation which uses differential rotation in part is not excluded; it may still be possible to explain the spiral arms by invoking differential rotation and an ongoing process which continually creates the arms. What cannot be done is to explain spiral arms simply in terms of differential rotation acting upon the initial conditions under which galaxies are created. The burden of explanation has been shifted from the initial conditions under which a spiral galaxy is created to the ongoing processes which continually operate within such a galaxy. There are two processes which have been suggested to explain the ongoing formation of spiral arms: density-waves and self-propagating star formation. As Nicolson (p208) points out, both processes may participate in spiral arm formation, and may be of differing importance in different types of spiral galaxy. There are so-called ‘grand-design’ spirals, which possess thin, long and well-defined spiral arms (see Figure \[M51\]); there are flocculent spirals, which possess many fluffy, poorly-defined spiral arms (see Figure \[NGC4414\]); and there is a continuum of intermediate cases, of which the Milky Way is one such. Karttunen *et al* point out that “in multiarmed galaxies the spiral arms may be short-lived, constantly forming and disappearing, but extensive, regular, two-armed patterns have to be more long-lived," (2003 p359). As Nicolson suggests, density-wave mechanisms are more appropriate for grand-design spirals, and self-propagating star formation is more appropriate for flocculent spirals. It was Bertil Lindblad who suggested in the 1920s that spiral arms could be produced by density waves propagating in a galaxy. The peaks and troughs of a density wave are independent of any particular elements from the medium through which the density wave is propagating. Thus, the density-wave explanation accepts that stars pass in and out of the spiral arms. However, it also accepts that spiral arms are genuine indicators of higher-than-average stellar density. The postulated density wave travels faster than the speed of sound in the interstellar medium, and the shock wave purportedly triggers compression of the interstellar medium, and the observed star formation in the spiral arms. The spiral arms therefore trace the brightest regions in the disk of a galaxy because the brightest regions are those which contain the shortlived O- and B-type stars. As Tayler states, “the most massive stars have such a short lifetime that, by the time they have ceased to be luminous, the spiral pattern has hardly changed its position. They should therefore only be found in the spiral regions…Stars of lower mass, such as the Sun, have a main sequence lifetime which is equal to many rotation periods of the pattern. Such stars should therefore be observed throughout the disk with no significant correlation with the present position of the spiral arms," (1993, p145). C.C.Lin and Frank Shu developed the density-wave idea further in 1964 by postulating that the stars in a spiral galaxy travel in slightly elliptical orbits which precess with the passage of time.[^6] This postulate was complemented by the suggestion of J. Kalnajs in 1973 that if the orientation of the major axis of these ellipses varies by a small angular increment at increasing distance from the galactic centre, then the ellipses fail to be uniformly separated, and where they come close together, they produce the appearance of grand-design spiral arms. Kalnajs’s explanation of spiral arms can be seen as a *structural explanation*, in the sense that the spiral arms are represented as structural elements in a geometrical model, and these structural elements are linked to other structural elements in the model, namely the precessing elliptical orbits, each aligned at a small angle to the one inside it. Nicolson suggests that density-wave spiral arms “are probably sustained either by the asymmetric gravitational field associated with a central bar structure (typical of grand-design spirals) or by gravitational disturbances caused by neighboring galaxies, or by a combination of both," (1999, p208). In the 1970s Mueller and Arnett suggested an alternative explanation for spiral arms, which rejects the idea that all spiral arms are density waves, and postulates instead that at least some are self-propagating waves of star formation. Unlike density waves, these waves purportedly do not require one to postulate external mechanisms, such as the asymmetric gravitational field of the central bar, or the tidal forces caused by other galaxies. Let us, then, examine star formation processes in a little more detail. As stated in Section 1, the mass and chemical composition of stars vary, and these are the characteristics which determine the history and lifetime of a star. New stars are continually born in a spiral galaxy, but they are not born from other stars, and consequently the mass of a new star is not in any sense inherited. The first generation of stars must have formed from clouds of almost pure hydrogen and helium. Subsequent generations of stars are observed to form from the cold, dense Giant Molecular Clouds of the interstellar medium. Given an environment which already contains surrounding stars, GMCs are only able to remain cool because they contain dust, which acts as a shield to starlight, and because they contain organic molecules, which are able to radiate excess heat (Smolin p110). Organic molecules contain carbon, a by-product of the nuclear fusion processes in previous generations of stars. The star-formation processes within GMCs are thought to be triggered by the shock-waves from the supernovae explosions of nearby high-mass stars, formed within neighbouring GMCs $10$ million years previously, or less. The shock-waves purportedly compress the interstellar medium, and instigate star formation. Because one spell of star-formation is triggered by the death of high-mass stars from a previous spell of star-formation, this process is referred to as self-propagating star-formation (Smolin p128). Indeed, the inner lanes of spiral arms are often observed to be dark and dusty, while the outer sides contain the star-forming regions. This suggests the star formation is directed against the direction of rotation, which would explain why spiral arms rotate slightly slower than a host galaxy. Nicolson (p208) suggests that after a burst of such self-propagating star formation, differential rotation stretches the region into a spiral arm. He suggests that in a flocculent spiral, random bursts of star formation produce numerous spiral arms by this mechanism, each of which fades away after star formation has ceased. Subsequent star formation elsewhere in the galaxy then creates more spiral arms. This differs somewhat from Smolin’s notion of self-propagating star formation, in which he states that “the waves of star formation neither die out, nor grow uncontrollably, but propagate at exactly the right rate to persist in the galaxy indefinitely," (1997 p135). The debate between the two different theories which attempt to explain spiral structure is particularly interesting when one appreciates that they take opposite stances on the basic cause and effect relationships which operate in a spiral galaxy. The density wave theory represents a spiral to be a spiral of density waves, and explains star formation as the consequence of the density waves. In contrast, the theory of self-propagating star formation represents spiral structure to be the consequence of waves of star formation, (Tayler 1993, p145). It would not be correct to say that the chemical composition of a new star is independent of the chemical composition of previous generations of stars. The chemical composition of the interstellar medium in a spiral galaxy is changing with the passage of time. The material expelled from one generation of stars provides a high-metallicity contribution to the medium from which the next generation of stars is composed. High metallicity purportedly inhibits the formation of higher-mass stars, hence the relative birthrate of lower-mass stars in a spiral galaxy increases with the passage of time. Accordingly, the birthrate of stars in the solar neighbourhood is claimed to be inversely proportional to mass $m$ according to the Salpeter form, $m^{-7/3}$, (Tayler 1993, p149). It is also suggested that the formation rate of high-mass stars is less than the formation rate of low-mass stars because the energy from a newly-created high-mass star heats up the medium from which it was born to such an extent that the star formation process halts, (Smolin p127). In general statistical terms, the number density $N_m$ of type-$m$ objects in such a population is the product $N_m = B_m \cdot l_m$ of the type-$m$ birthrate $B_m$ with the type-$m$ life-time $l_m$. If we let $N_m$ denote the number density of mass-$m$ stars, then it follows that because small mass stars have a greater lifetime, even if they have the same birth-rate as high-mass stars, they will come to dominate the population. The eventual domination of lower-mass stars is therefore the consequence of both statistics and the physics of star formation processes. In a spiral galaxy, however, where the birthrates are far from being spatially homogeneous, and the majority of star formation occurs in the spiral arms, one cannot infer birthrates from the known lifetimes of stars and the observed number densities in regions which are remote from the star formation in spiral arms. The population of stars in a spiral galaxy is therefore a type of population in which: (i) there are variable characteristics distributed within the population, and a small subset of these characteristics define different types within the population; (ii) each member of the population has a finite lifetime, determined by the values of the type-defining characteristics; and (iii) new members of the population are not reproduced from existing members of the population, i.e., new members of the population are born without having any parent(s) in the existing population. Suppose in addition that (iv) population members of each type are created at approximately the same rate, or short-lifetime members are created at a lower rate than long-lifetime members. Such a population evolves to be dominated by the long-lifetime objects. Such a population is neither evolving randomly, nor is it evolving by natural selection. [99]{} Cox, J.P., Salpeter, E.E. (1964). Equilibrium Models for Helium-Burning Stars. III. Semi-Degenerate Stars of Small Mass, *Astrophysical Journal*, vol. 140, p.485. Disney, M.J. (2000). *The case against cosmology*, arXiv:astro-ph/0009020 v1 1 Sep 2000. Gabriel, M., Noels-Grötsch, A. (1968). Stabilité séculaire des étoiles de carbone pur, *Annales d’Astrophysique*, vol. 31, p.167. Hansen, C.J. (1978). Secular stability: applications to stellar structure and evolution. *Annual review of astronomy and astrophysics*, volume 16, pp15-32. Hansen, C.J., Kawaler, S.D., Trimble, V. (2004). *Stellar Interiors: Physical Principles, Structure and Evolution*, 2nd Edition. Berlin-Heidelberg-New York: Springer Verlag. Kähler, H. (1978). The Vogt-Russell theorem, and new results on an old problem, pp303-311 in *The HR diagram - The 100th anniversary of Henry Norris Russell*, A.G. Davis Philip and D.S.Hayes (eds.). Dordrecht: Reidel. Karttunnen, H., Kröger, P., Oja, H., Poutanen, M., Donner, K.J. (2003). *Fundamental astronomy*, Fourth Edition. Berlin-Heidelberg-New York: Springer Verlag. Kippenhahn, R., Weigert, A. (1990). *Stellar Structure and Evolution*. Berlin-Heidelberg-New York: Springer Verlag. Nicolson, I. (1999). *Unfolding our universe*. Cambridge: Cambridge University Press. Pasha, I.I. (2004). Density-wave spiral theories in the 1940s, I. arXiv:astro-ph/0406142 Rindler, W. (2001). *Relativity: Special, General, and Cosmological*. Oxford: Oxford University Press. Russell, H.N., Dugan, R.S., Stewart, J.Q. (1927). *Astronomy* Vol. 2. Boston: Ginn and Co. Smolin, L. (1997a). *The Life of the Cosmos*. London: Weidenfeld and Nicolson. Tayler, R.J. (1993). *Galaxies: structure and evolution*. Cambridge: Cambridge University Press. Tayler, R.J. (1994). *The Stars: their structure and evolution*. Cambridge: Cambridge University Press. Vogt, H. (1926). Die Beziehung zwischen den Massen und den absoluten Leuchtkräften der Sterne, *Astronomische Nachrichten*, volume 226, p.301. [^1]: The pressure at each radius is a sum of the gas pressure and the radiation pressure. [^2]: Conduction is only significant in ‘compact’ stars, i.e., white dwarf and neutron stars. [^3]: It is a common approximation to use only three fractions $(X,Y,Z)$, which are such that $X+Y+Z =1$, and which specify, respectively, the proportion of hydrogen, the proportion of helium, and the proportion of all other elements in the constitution of a star. [^4]: Strictly, the effective surface temperature $T_e$ is defined as the temperature of a black body with the same radius and luminosity as the star. [^5]: In relativistic cosmology, ‘dust’ means a pressure-less fluid. In astronomy and astrophysics, however, dust means tiny grains of solid matter (Nicolson 1999, p157). [^6]: See Pasha (2004) for a detailed history of these ideas.

--- abstract: 'We show that, assuming Vojta’s height conjecture, the height of a rational point on an algebraically hyperbolic variety can be bounded “uniformly” in families. This generalizes a result of Su-Ion Ih for curves of genus at least two to higher-dimensional varieties. As an application, we show that, assuming Vojta’s height conjecture, the height of a rational point on a curve of general type is uniformly bounded. Finally, we prove a similar result for smooth hyperbolic surfaces with $c_1^2 > c_2$.' author: - Kenneth Ascher AND Ariyan Javanpeykar bibliography: - 'heights\_stacks.bib' title: Bounding heights uniformly in families of hyperbolic varieties --- Introduction ============ The celebrated work of Caporaso, Harris, and Mazur [@chm], sparked an interest in discovering implications of Lang’s conjecture for rational points on varieties of general type. In fact, they show that Lang’s conjecture implies a uniform bound, based solely on $k$ and the genus, of the number of $k$-points on a curve of general type defined over a number field $k$ (cf. [@av; @has]). As Vojta’s height conjecture (Conjecture \[conj:Vojta\]) implies the conjecture of Lang, the aforementioned results show that Vojta’s height conjecture also implies a uniform version of Lang’s conjecture. In particular, it seems reasonable to suspect that Vojta’s height conjecture also has consequences for “uniform” height bounds. However, one cannot expect uniform height bounds in the naive sense. Indeed, for all $P \in\mathbb P^2(\mathbb Q)$ and all $d\geq 4$, there is a smooth curve $X$ of degree $d$ in $\mathbb P^2_{\mathbb{Q}}$ with $P \in X(\mathbb{Q})$. Thus, for all $d\geq 4$, there is no real number $c>0$ depending only on $d$ such that for all smooth degree $d$ hypersurfaces $X \subset \mathbb P^2_{\mathbb{Q}}$ and all $P \in X(\mathbb Q)$ the inequality $h(P) \leq c$ holds. In particular, there is no real number $c>0$ such that for all smooth quartic hypersurfaces $X \subset \mathbb P^2_{\mathbb{Q}}$ and all $P \in X(\mathbb Q)$ the inequality $h(P) \leq c$ holds. Thus, it is at first sight not clear what is meant by “uniform” height bounds. However, Su-Ion Ih has shown [@ih] that Vojta’s height conjecture implies that the height of a rational point on a smooth proper curve of general type is bounded uniformly in families with the bound depending linearly on the height of the curve. Ih later showed in [@ih2] that the same is true for integral points on elliptic curves. The goal of this paper is to generalize Ih’s results in [@ih] by investigating consequences of Vojta’s height conjecture for families of (algebraically) hyperbolic varieties of general type. In this paper, a proper scheme $X$ over a field $k$ is called (algebraically) hyperbolic if all integral subvarieties of $X $ are of general type; see Definition \[defn:hyperbolic\]. In the statement of our main result we consider morphisms of algebraic stacks $f\colon X\to Y$ which are representable by schemes, i.e., for all schemes $S$ and all morphisms $S\to Y$, the algebraic stack $X\times_Y S$ is (representable by) a scheme. Furthermore, a substack of an algebraic stack is constructible if it is a finite union of locally closed substacks. Moreover, we will use the relative discriminant $d_k(\mathcal T_P)$ of a point on an algebraic stack over a number field $k$; we refer the reader to Section \[section:discr\] for a precise definition of the relative discriminant $d_k({\mathcal{T}}_P)$. Also, to state our theorem, we will use heights on stacks as discussed in Section \[section:heights\_on\_stacks\]. \[thm:stackthm\] Let $k$ be a number field and let $f \colon X \to Y$ be a proper surjective morphism of proper Deligne-Mumford stacks over $k$ which is representable by schemes. Let $h$ be a height function on $X$ and let $h_Y$ be a height function on $Y$ associated to an ample divisor with $h_Y \geq 1$. Assume Vojta’s height conjecture (Conjecture \[conj:Vojta\]). Let $U\subset Y$ be a constructible substack such that, for all $t\in U$, the variety $X_t$ is smooth and hyperbolic. Then there is a real number $c > 0$ depending only on $k$, $Y$, $X$, and $f$ such that, for all $P$ in $X(k)$ with $f(P)$ in $U$, the following inequality holds $$h(P) \leq c \cdot \big(h_Y(f(P)) + d_k({\mathcal{T}}_P)\big).$$ Note that Ih proves Theorem \[thm:stackthm\] under the additional assumptions that the fibres are one-dimensional, and $Y$ is a scheme; see [@ih Theorem 1.0.1]. If one assumes that $Y$ is a scheme, then the discriminant term $d_k({\mathcal{T}}_P)$ can be omitted (as it equals zero). Ih’s theorem for families of curves is slightly more general than Theorem \[thm:stackthm\], as he treats points of bounded degree, and not merely rational points. To keep the proofs slightly more transparent, we have restricted our attention to rational points. However, the transition from rational points to points of bounded degree can be made easily. Furthermore, the generalization of Ih’s theoem to stacks is unavoidable if one desires applications to all curves simultaneously; see Theorem \[thm1\] below, and the discussion following it. One cannot expect a stronger uniformity type statement for heights on (not necessarily hyperbolic) varieties of general type. Indeed, if $k$ is a number field and $f \colon X\to Y$ is a smooth proper morphism of $k$-schemes whose geometric fibres are varieties of general type and $t$ is a point in $Y$ such that $X_t$ contains a copy of $\mathbb P^1_{k(t)}$, then there is no real number $c>0$ such that for all $P \in X_t$, the inequality $h(P) \leq c \cdot h_Y(f(P))$ holds. Our proof of Theorem \[thm:stackthm\] uses the recent [@avap], which shows that Vojta’s conjecture actually implies a version of the conjecture for stacks. Moreover, to prove Theorem \[thm:stackthm\] we follow the strategy of Ih. Indeed, we combine an induction argument with an application of Vojta’s conjecture to a desingularization of $X$ (Proposition \[prop:mainprop\]). This line of reasoning was also used in Ih’s work [@ih; @ih2]. We argue that it is more natural to work in the stacks setting, as this allows us to apply our results to moduli stacks of hyperbolic varieties, thus obtaining more concrete results. In fact, as a first corollary of Theorem \[thm:stackthm\] we obtain the following uniformity statement for curves. \[thm1\] Assume Conjecture \[conj:Vojta\]. Let $g\geq 2$ be an integer and let $k$ be a number field. There is a real number $c$ depending only on $g$ and $k$ satisfying the following. For all smooth projective curves $X$ of genus $g$ over $k$, and all $P$ in $X(k)$, the following inequality holds $$h(P) \leq c(g, k) \cdot \big(h(X) +d_k(\mathcal T_{X})\big).$$ The discriminant term $d_k(\mathcal{T}_{X})$ can not be omitted in Theorem \[thm1\] (and neither in Theorem \[thm:stackthm\]). To explain this, for an integer $n\geq 1$, define $d_n := n^5+1$ and define the genus $2$ curve $C_n$ by $d_n y^2 =x^5+1$. Note that the height of $C_n$ is equal to the height of $C_1$, as $C_{n,{\overline{\mathbb Q}}}\cong C_{1,{\overline{\mathbb Q}}}$ and the height is a “geometric” invariant. Let $P_n := (1,n) \in \mathbb{Q}^2$ and note that $P_n$ defines a $\mathbb{Q}$-rational point of $C_n$. Since $h(P_n)$ tends to infinity as $n$ gets larger, we can not omit the discriminant term in Theorem \[thm1\]. It is not clear how to deduce Theorem \[thm1\] from Ih’s results, as Ih’s results only apply to families of curves parametrized by schemes. Finally, we also obtain a uniformity statement for certain hyperbolic surfaces. \[thm2\] Assume Conjecture \[conj:Vojta\]. Fix an even integer $a$ and a number field $k$. There is a real number $c$ depending only on $a$ and $k$ satisfying the following. For all smooth hyperbolic surfaces $S$ over $k$ with $c_1^2(S) = a> c_2(S)$ and all $P$ in $S(k)$, the following inequality holds $$h(P) \leq c \cdot \big( h(S) + d_k(\mathcal T_{S}) \big).$$ We refer the reader to Section \[section:apps\] for precise definitions of the height functions appearing in Theorems \[thm1\] and \[thm2\]. We prove Theorems \[thm1\] and \[thm2\] by applying Theorem \[thm:stackthm\] to the universal family of the moduli space of curves and the moduli space of surfaces of general type, respectively. The technical difficulty in applying Theorem \[thm:stackthm\] is to prove the constructibility of the locus of points corresponding to hyperbolic varieties. In the setting of curves (Theorem \[thm1\]) this is simple, whereas the case of surfaces (Theorem \[thm2\]) requires deep results of Bogomolov and Miyaoka [@bogomolov; @miyaoka]. Theorem \[thm:stackthm\] applies to any family of varieties of general type for which the locus of hyperbolic varieties is constructible on the base. However, as we show in Section \[section:apps\], verifying the constructibility of the latter locus is not straightforward. We note that a conjecture of Lang (see [@langsurvey]) asserts that our notion of hyperbolicity for $X$ is equivalent to being *Brody hyperbolic*, i.e., that there are no non-constant holomorphic maps $f: \mathbb{C} \to X(\mathbb{C})$. In particular, as the property of being Brody hyperbolic is open in the analytic topology [@brody], Lang’s conjecture implies that the property of being hyperbolic is open in the analytic topology. In particular, assuming Lang’s conjecture, if the locus of smooth projective hyperbolic surfaces is constructible in the moduli stack of smooth canonically polarized surfaces, then [@SGA1 Exposé XII, Corollaire 2.3] implies that it is (Zariski) open. Acknowledgements {#acknowledgements .unnumbered} ---------------- We would like to thank Dan Abramovich, Dori Bejleri, Marco Maculan, and Siddharth Mathur for useful comments and suggestions. We are most grateful to the referee for many comments and remarks which helped improve this paper. K.A. was supported in part by funds from NSF grant DMS-1162367 and an NSF Postdoctoral Fellowship. A.J. gratefully acknowledges support from SFB/Transregio 45. Hyperbolicity ============= In this section the base field $k$ is a field of arbitrary characteristic. Let $X$ be a proper Deligne-Mumford stack of dimension $n$ over $k$. A divisor $D$ on $X$ is *big* if $h^0(X, {\mathcal{O}}_X(mD)) > c \cdot m^n$ for some $c>0$ and $m \gg1$. Recall that a projective geometrically irreducible variety $X$ over $k$ is of general type if for a desingularization $\widetilde{X} \to X_{\textrm{red}}$ of the reduced scheme $X_{\textrm{red}}$, the sheaf $\omega_{\widetilde{X}}$ is big. Note that, if $X$ is of general type and $\widetilde{X}\to X_{\textrm{red}}$ is any desingularization, then $\omega_{\widetilde{X}}$ is big. \[defn:hyperbolic\] A projective scheme $X$ over $k$ is *hyperbolic (over $k$)* if for all its closed subschemes $Z$, any irreducible component of $Z_{\overline{k}}$ is of general type. Note that, if $X$ is a hyperbolic projective scheme over $k$, then $X$ and all of its closed subvarieties are of general type. Moreover, if $L/k$ is a field extension, then $X$ is hyperbolic over $k$ if and only if $X_L$ is hyperbolic over $L$. For example, a smooth proper geometrically connected curve $X$ over $k$ is hyperbolic if and only if the genus of $X$ is at least two. If $X$ is a smooth projective scheme over $\mathbb{C}$ such that the associated complex manifold $X^{\textrm{an}}$ admits an immersive period map (i.e., there exists a polarized variation of ${\mathbb{Z}}$-Hodge structures over $X^{\textrm{an}}$ whose differential is injective at all points), then $X$ is hyperbolic. This follows from the proof of [@JL2 Lemma 6.3] which uses Zuo’s theorem [@Zuo] (cf. [@Brunebarbe]). Finally, let $X$ be a smooth projective scheme over $\mathbb{C}$ and suppose that there exists a smooth proper morphism $Y\to X$ whose fibres have ample canonical bundle such that, for all $a$ in $X(\mathbb C)$, the set of $b$ in $X(\mathbb C)$ with $X_a \cong X_b$ is finite. Then $X$ is hyperbolic. This is a consequence of Viehweg’s conjecture for “compact” base varieties [@patakfalvi]. Kodaira’s criterion for bigness ------------------------------- We assume in this section that $k$ is of characteristic zero. Recall that for a big divisor $D$ on a projective variety, there exists a positive integer $n$ such that $nD \sim_{\mathbb{Q}}A +E$, where $A$ is ample and $E$ is effective [@km Lemma 2.60]. We state a generalization of this statement (see Lemma \[lem:kod\_crit\]) which is presumably known; we include a proof for lack of reference. \[lem:bigness\] Let $\pi \colon X\to Y$ be a quasi-finite morphism of proper Deligne-Mumford stacks over $k$. Let $D$ be a divisor on $Y$. The divisor $D$ is big on $Y$ if and only if $\pi^\ast D$ is big on $X$. This follows from the definition of bigness, and the fact that $\pi_\ast \pi^\ast D $ is linearly equivalent to $ m D$, where $m\geq 1$ is some integer. If $D$ is a divisor on a finite type separated Deligne-Mumford stack $\mathcal X$ over $k$ with coarse space $\mathcal X\to \mathcal X^c$, then $D$ is *ample* (resp. *effective*) on $\mathcal X$ if there exists a positive integer $n$ such that $nD$ is the pull-back of an ample (resp. *effective*) divisor on $\mathcal X^c$. Note that, if $\mathcal X$ has an ample divisor, then $\mathcal X^c$ is a quasi-projective scheme over $k$. \[lem:kod\_crit\] Let ${\mathcal{X}}$ be a proper Deligne-Mumford stack over $k$ with projective coarse moduli space ${\mathcal{X}}^c$. If $D$ is a big divisor on ${\mathcal{X}}$, then there exists a positive integer $n$ such that $nD~\sim_{\mathbb{Q}}~{\mathcal{A}}+ {\mathcal{E}}$, where ${\mathcal{A}}$ is ample and ${\mathcal{E}}$ is effective. Let $\pi \colon {\mathcal{X}}\to {\mathcal{X}}^{c}$ denote the morphism from ${\mathcal{X}}$ to its coarse moduli space ${\mathcal{X}}^c$. It follows from [@Olsson Proposition 6.1] that there exists a positive integer $m$ such that $mD$ is $\mathbb Q$-linearly equivalent to the pullback of a divisor $D_0$ on ${\mathcal{X}}^c$. As $mD$ is a big divisor on ${\mathcal{X}}$, the divisor $D_0$ is big on ${\mathcal{X}}^c$ (Lemma \[lem:bigness\]). By Kodaira’s criterion for bigness, there exists a positive integer $m_2$ such that $m_2 D_0 $ is $\mathbb{Q}$-linearly equivalent to $ A + E$, where $A$ is an ample divisor on ${\mathcal{X}}^c$ and $E$ is an effective divisor on ${\mathcal{X}}^c$. Write $n = m\cdot m_2$. We now see that $nD = m \cdot m_2\cdot D \sim_{\mathbb Q} \pi^\ast m_2 D_0 \sim_{\mathbb Q} \pi^\ast (A+E)$. Since $\mathcal A\colonequals \pi^\ast A$ is ample, and $\mathcal E \colonequals \pi^\ast E$ is effective, this concludes the proof of the lemma. Vojta’s conjecture for varieties and stacks =========================================== In this section, we let $k$ be a number field. We begin by recalling Vojta’s conjecture for heights of points on schemes, using [@avap] and [@Vojta]. Our statement of the conjecture is perhaps not the most standard, but is more natural for our setting as we will need the extension of the conjecture to algebraic stacks. Discriminants of fields ----------------------- Before defining the conjecture, we recall discriminants of fields following Section 2 of [@avap]. Given a finite extension $E / k$, define the *relative logarithmic discriminant* to be: $$\begin{aligned} d_k(E) &= \dfrac{1}{[E:k]} \log|{\mathrm{Disc}}({\mathcal{O}}_E)| - \log|{\mathrm{Disc}}({\mathcal{O}}_k)| = \dfrac{1}{[E:k]} \deg(\Omega_{{\mathcal{O}}_E / {\mathcal{O}}_k}),\end{aligned}$$ where the second equality follows from the equality of ideals $({\mathrm{Disc}}({\mathcal{O}}_k)) = N_{k/{\mathbb{Q}}} \det \Omega_{{\mathcal{O}}_k / {\mathbb{Z}}}.$ Heights ------- In this paper we will use *logarithmic* (Weil) heights. For more details, we refer the reader to [@bg; @hs]. \[def:ht\] Let $d $ be the degree of $k$ over ${\mathbb{Q}}$ and let $M_k$ be a complete set of normalized absolute values on $k$. The (logarithmic) height of a point $P = \left[x_0 : \dots : x_n\right]~\in~{\mathbb{P}}^n(k)$ is defined to be: $$h_k(P) = \frac{1}{d} \sum_{v \in M_k} \log(\max_{0 \leq i \leq n} \{ \|x_i\|_v \}).$$ If $X$ is a projective variety with a projective embedding $\phi: X \hookrightarrow {\mathbb{P}}^n$, we can define a height function $h_\phi \colon X \to \mathbb{R}$ given by $$h_\phi(P) = h(\phi(P)).$$ More generally, given a very ample divisor $D$ on $X$, we define $h_D(P) = h(\phi_D(P))$, where $\phi_D$ is the natural embedding of $X$ in ${\mathbb{P}}^n$ given by $D$. (We stress that $h_D$ is well-defined, up to a bounded function.) \[properties1\] The following statements hold. 1. If $f \colon X \to Y$ is a morphism, then $h_{X,f^*D} = h_{Y,D} + O(1)$. 2. If $D$ and $E$ are both divisors, then $h_{D+E} = h_D + h_E + O(1)$. 3. If $D$ is effective, $h_D \geq O(1)$ for all points not in the base locus of $D$. See [@hs Theorems B.3.2.b, B.3.2.c, and B.3.2.e]. Vojta’s conjecture ------------------ We now state Vojta’s conjecture for schemes. We stress that this conjecture (Conjecture \[conj:Vojta\]) implies a version for stacks; see Proposition \[prop:vstacks\]. [@Vojta Conjecture 2.3]\[conj:Vojta\] Let $X$ be a smooth projective scheme over $k$. Let $H$ be a big line bundle on $X$, let $r$ be a positive integer, and fix $\delta > 0$. Then there exists a proper Zariski closed subset $Z \subset X$ such that, for all closed points $x \in X$ with $x\not\in Z$ and $[k(x):k]\leq r$, $$h_{K_X}(x) - \delta h_H(x) \leq d_k(k(x)) + O(1).$$ Note that the discriminant term $d_k(k(x))$ equals zero when $x$ is a $k$-rational point of $X$. Vojta’s conjecture for stacks ----------------------------- Before stating the version of Vojta’s conjecture for Deligne-Mumford stacks, we introduce some preliminaries, following Section 3 of [@avap]. If $S$ is a finite set of finite places of $k$, we let ${\mathcal{O}}_{k,S}$ be the ring of $S$-integers in $k$. ### The stacky discriminant Let ${\mathcal{X}}\to \operatorname{Spec}({\mathcal{O}}_{k,S})$ be a finite type separated Deligne-Mumford stack with generic fibre $X\to \operatorname{Spec}k$. Given a point $x \in {\mathcal{X}}(\overline{k}) = X(\overline{k})$, we define ${\mathcal{T}}_x \to {\mathcal{X}}$ to be the normalization of the closure of $x$ in $\mathcal X$. Note that ${\mathcal{T}}_x$ is a normal proper Deligne-Mumford stack over ${\mathcal{O}}_{k,S}$ whose coarse moduli scheme is $\operatorname{Spec}({\mathcal{O}}_{k(x), S_{k(x)}})$. Here $S_{k(x)}$ is the set of finite places of $k(x)$ lying over $S$. ### Relative discriminants for stacks {#section:discr} Let $E$ be a finite field extension of $k$, and let $\mathcal T$ be a normal separated Deligne-Mumford stack over $\mathcal O_E$ whose coarse moduli scheme is $\operatorname{Spec}{\mathcal{O}}_E$. We define the relative discriminant of $\mathcal T$ over ${\mathcal{O}}_k$ as follows: $$\begin{aligned} d_k({\mathcal{T}}) &= \dfrac{1}{\deg({\mathcal{T}}/ {\mathcal{O}}_k)}\deg(\Omega_{{\mathcal{T}}/ \operatorname{Spec}({\mathcal{O}}_k)}). \end{aligned}$$ Note that $d_k(\mathcal T)$ is a well-defined real number, and that $\exp(d_k(\mathcal{T}))$ is a rational number. ### Heights on stacks {#section:heights_on_stacks} Let $X$ be a finite type Deligne-Mumford stack over $k$ with finite inertia whose coarse space $X^c$ is a quasi-projective scheme over $k$. Fix a finite set of finite places $S$ of $k$ and a finite type separated Deligne-Mumford stack $\mathcal X\to \operatorname{Spec}({\mathcal{O}}_{k,S})$ such that ${\mathcal{X}}_k \cong X$. Let $H$ be a divisor on $X$. Let $n\geq 1$ be an integer such that $nH$ is the pull-back of a divisor $H^c$ on $X^c$. Fix a height function $h_{H^c}$ for $H^c$ on $X^c$. We define the height function $h_H$ on $X(k)$ with respect to $H$ to be $$h_H(x) \colonequals \frac{1}{n} h_{H^c}(\pi(x)).$$ Note that $h_H$ is a well-defined function on $X(\overline{k})$ which is independent of the choice of $n$ and $H^c$. We now give another way to compute the height function, under suitable assumptions on $X$. By [@kv Theorem 2.1], a finite type separated Deligne-Mumford stack over $k$ which is a quotient stack and has a quasi-projective coarse moduli space admits a finite flat surjective morphism $f \colon Y \to {\mathcal{X}}$, where $Y$ is a quasi-projective scheme. Fix a height function $h_{f^\ast H}$ on $Y$. We define the height $h_H(x)$ of $x\in {\mathcal{X}}(\overline{k})$ as follows. If $x \in {\mathcal{X}}(\overline{k})$, then we choose $y \in Y(\overline{k})$ to be a point over $x$, and we define $$h_H(x) \colonequals h_{f^*(H)}(y).$$ It follows from the projection formula (which holds for Deligne-Mumford stacks, in particular see the introduction of [@Vistoli]) that $h_H$ is a well-defined function on $\mathcal X(\overline{k})$. Moreover, if $H$ is ample, for all $d\geq 1$ and $C\in \mathbb R$, the set of isomorphism classes of $\overline{k}$-points $x$ of $\mathcal X$ such that $h_H(x) \leq C$ and $[k(x):k] \leq d$ is finite. The analogous finiteness statement for $k$-isomorphism classes can fail. However, the set of $k$-isomorphism classes of $k$-points $x$ of $\mathcal X$ such that $h_H(x) + d_k(\mathcal T_x) \leq C$ and $[k(x):k] \leq d$ is finite. In particular, as $h_H(x) + d_k(\mathcal T_x)$ has the Northcott property, the expression $h_H(x) + d_k(\mathcal T_x)$ can be considered as “the” height of $x$ [@avap]. \[prop:vstacks\] Assume Conjecture \[conj:Vojta\] holds and fix $\delta > 0$. Let $S$ be a finite set of finite places of $k$. Let ${\mathcal{X}}$ be a smooth proper Deligne-Mumford stack over ${\mathcal{O}}_{k,S}$ whose generic fibre $X = {\mathcal{X}}_k$ is geometrically irreducible over $k$ and has a projective coarse space. Let $H$ be a big line bundle on $X$. Then, there is a proper Zariski closed substack $Z \subset X$ such that, for all $x\in (X\setminus Z)(k)$ the following inequality holds $$h_{K_X }(x) - \delta h_H(x) \leq d_k({\mathcal{T}}_x) + O(1).$$ This is [@avap Proposition 3.2]. Applying the stacky Vojta conjecture ==================================== We prove a generalization of [@ih2 Proposition 2.5.1] to morphisms of proper Deligne-Mumford stacks, under suitable assumptions. We stress that our reasoning follows Ih’s arguments in *loc. cit.* in several parts of the proof. Let $k$ be a number field, and let $f:\mathcal X\to \mathcal Y$ be a proper morphism of proper integral Deligne-Mumford stacks over $B = \operatorname{Spec}{\mathcal{O}}_{k,S}$, where $\mathcal{X}$ is smooth with a projective coarse moduli space. Let $h$ be a height function on ${\mathcal{X}}$ and let $h_{\mathcal{Y}}$ be a height function on ${\mathcal{Y}}$ associated to an ample divisor such that $h_{\mathcal{Y}}\geq 1.$ Let $\eta$ be the generic point of $\mathcal{Y}$, let ${\mathcal{X}}_{\eta}$ be the generic fibre of $f \colon {\mathcal{X}}\to {\mathcal{Y}}$, and let ${\mathcal{X}}_k$ be the generic fibre of ${\mathcal{X}}\to B$. Note that $\mathcal X_k$ is a smooth proper Deligne-Mumford stack over $k$ with a projective coarse space. \[prop:mainprop\] Assume Conjecture \[conj:Vojta\]. Suppose that the morphism $f$ is representable by schemes, and that ${\mathcal{X}}_\eta$ is smooth and of general type. Then there exists a real number $c(k,S, {\mathcal{Y}},f)$ and a proper Zariski closed substack ${\mathcal{Z}}\subset {\mathcal{X}}$ such that, for all $P$ in ${\mathcal{X}}(B)\setminus {\mathcal{Z}}$, the following inequality holds: $$h(P) \leq c(k, {\mathcal{Y}},f) \cdot \big(d_k(\mathcal T_P)+ h_{\mathcal{Y}}(f(P))\big).$$ Let $\Delta$ be an ample divisor on ${\mathcal{X}}$ such that the associated height $h_\Delta$ on ${\mathcal{X}}$ satisfies $h_\Delta \geq 1$. Note that the push-forward of $\Delta$ to the coarse space is ample. Recall that $\mathcal{X}_k$ denotes the generic fibre of $\mathcal{X}\to B$. Moreover, Vojta’s conjecture (Conjecture \[conj:Vojta\]) implies Vojta’s conjecture for stacks (Proposition \[prop:vstacks\]). Therefore, by Vojta’s conjecture for stacks (Proposition \[prop:vstacks\]) applied to ${\mathcal{X}}_k$, there exists a proper Zariski closed substack $Z\subset {\mathcal{X}}_k$ such that, for all $P \in {\mathcal{X}}_k(k)\setminus Z$, the following inequality $$h_{K_{{\mathcal{X}}_k}}(P) - \dfrac{1}{2} \epsilon h_{\Delta}(P) \leq d_k({\mathcal{T}}_P) + O(1)$$ holds, where we compute all invariants with respect to the model $\mathcal X$ for ${\mathcal{X}}_k$ over $B$. In particular, there exists a proper closed substack ${\mathcal{Z}}$ of ${\mathcal{X}}$ (namely, the closure of $Z$ in $\mathcal X$) such that, for all $P$ in ${\mathcal{X}}(B) $ not in ${\mathcal{Z}}$, the following inequality holds $$\begin{aligned} \label{eqn:vojta} h_{K_{\mathcal{X}}}(P) - \dfrac{1}{2} \epsilon h_{\Delta}(P) \leq d_k({\mathcal{T}}_P) + O(1). \end{aligned}$$ Since $f$ is representable, ${\mathcal{X}}_\eta$ is a scheme. Moreover, since ${\mathcal{X}}_\eta$ is smooth and of general type, by the Kodaira criterion for bigness (Lemma \[lem:kod\_crit\]), there exists an ample divisor $A$ on ${\mathcal{X}}_\eta$, an effective divisor $E$ on ${\mathcal{X}}_\eta$, and a positive integer $n$ such that $$n(K_{{\mathcal{X}}_\eta}) \sim_{\mathbb{Q}}A + E.$$ For a small enough $\epsilon \in {\mathbb{Q}}_{>0}$, we can rewrite $$\begin{aligned} (K_{\mathcal{X}}- \epsilon\Delta)|_\eta &= K_{{\mathcal{X}}_\eta} - \epsilon\Delta|_\eta \sim_{\mathbb Q} \bigg(\dfrac{1}{n}A + \dfrac{1}{n}E\bigg) - \epsilon\Delta|_\eta \\ &= \bigg(\dfrac{1}{n}A - \epsilon\Delta|_\eta\bigg) + \dfrac{1}{n}E. \end{aligned}$$ Thus, there exists an effective divisor $E'$ on ${\mathcal{X}}_\eta$ and a positive integer $m$ such that $$m\bigg(\bigg(\dfrac{1}{n}A -\epsilon\Delta|_\eta\bigg) + \dfrac{1}{n}E\bigg) \sim_{\mathbb Q} E'.$$ Taking Zariski closures of these divisors in ${\mathcal{X}}$, it follows that there exists a vertical ${\mathbb{Q}}$-divisor ${\mathcal{F}}$ on ${\mathcal{X}}$ and an effective divisor $\mathcal E$ on ${\mathcal{X}}$ such that $$K_{\mathcal{X}}-\epsilon \Delta + {\mathcal{F}}\sim_{\mathbb Q} \dfrac{1}{m}{\mathcal{E}}.$$ Since $\mathcal F$ is a vertical divisor on ${\mathcal{X}}$, there is an effective divisor $\mathcal G$ on ${\mathcal{Y}}$ such that $\mathcal F \leq f^*\mathcal G$. Therefore, by Proposition \[properties1\], the inequality $h_{\mathcal F} \leq h_{f^*\mathcal{G}} +O(1)$ holds, outside of $\mathrm{Supp} \ (f^*\mathcal{G})$, and $h_{f^*\mathcal{G}} = (h_{\mathcal G} \circ f) + O(1)$. In particular, since $h_{\mathcal{Y}}$ is a height associated to an ample divisor, we see that $h_{\mathcal{G}} \leq O(h_{\mathcal{Y}})$ by [@lang Proposition 5.4]. Therefore, for all points $t$ in ${\mathcal{Y}}(k)$ and all $P \in {\mathcal{X}}_t(B) \setminus \textrm{Supp}(f^*G)$, the inequality $$\begin{aligned} h_{\mathcal {F}}(P) & \leq & h_{f^*\mathcal {G}}(P) + O(1)= h_{\mathcal G} ( f(P)) + O(1)\leq O\big(h_{\mathcal{Y}}(f(P))\big) + O(1) \end{aligned}$$ holds, outside of $\mathrm{Supp} \ (f^*\mathcal{G})$. In particular, replacing $\mathcal Z$ by the union of $\mathcal Z$ with $\mathrm{Supp}(f^\ast \mathcal G)$, it follows that $$\begin{aligned} \label{eqn:vertical}h_{\mathcal {F}} \leq O(h_{\mathcal{Y}}\circ f) + O(1) \end{aligned}$$ outside $\mathcal Z$. Since $K_{\mathcal X} - \epsilon \Delta +F$ is effective, it follows that, replacing $\mathcal Z$ by a larger proper closed substack of $\mathcal X$ if necessary, the inequality $$\begin{aligned} \label{eqn:effectivity} h_{K_{\mathcal{X}}- \epsilon \Delta + F} \geq O(1) \end{aligned}$$ holds outside $\mathcal Z$ by Proposition \[properties1\] (3). Let $d_k(\mathcal T)$ be the function that assigns to a point $P$ in $\mathcal X(\overline{k})$ the real number $d_k(\mathcal T_P)$. In particular, we obtain that $$\begin{aligned} O(1) &\leq h_{K_{\mathcal{X}}-\epsilon \Delta + F}\leq (h_{K_{{\mathcal{X}}}} - \frac{1}{2} \epsilon h_\Delta) - \frac{1}{2} \epsilon h_\Delta + h_F + O(1)\\ & \leq (h_{K_{{\mathcal{X}}}} - \frac{1}{2} \epsilon h_\Delta) - \frac{1}{2} \epsilon h_\Delta + O(h_{\mathcal{Y}}\circ f) + O(1) \leq d_{k}(\mathcal T) - \frac{1}{2} \epsilon h_\Delta + O(h_{\mathcal{Y}}\circ f) + O(1), \end{aligned}$$ where the inequalities follow from Equation (\[eqn:effectivity\]), Proposition \[properties1\].(2), Equation (\[eqn:vertical\]), and Vojta’s conjecture (\[eqn:vojta\]) respectively. We conclude that, for all $t$ in $\mathcal Y(B)$ and all $P$ in ${\mathcal{X}}_t(B) \setminus {\mathcal{Z}}$ the inequality $$\frac{1}{2} \epsilon h_\Delta(P) \leq d_k(\mathcal T_P)+ O(h_{\mathcal{Y}}(t)) + O(1)$$ holds. Therefore, there is a real number $c>0$ such that, for all $t$ in $\mathcal Y(t)$ and all $P$ in $\mathcal X_t$ not in $\mathcal Z$, the inequality $$h_\Delta(P) \leq c\cdot \bigg(d_k(\mathcal T_P)+ O\big(h_{\mathcal{Y}}(t)\big)\bigg) + O(1)$$ holds. In particular, replacing $c$ by a larger real number if necessary, we conclude that $$h_\Delta(P) \leq c\cdot \bigg(d_k(\mathcal T_P)+ h_{\mathcal{Y}}(t)\bigg) + O(1).$$ As $\Delta$ is ample and $h_\Delta \geq 1$, we conclude that, using [@lang Proposition 5.4] and replacing $c$ by a larger real number if necessary, for all $t$ in $\mathcal Y(t)$ and all $P$ in $\mathcal X_t$ not in $\mathcal Z$, the inequality $$h(P) \leq O\big(h_\Delta(P)\big) \leq c \cdot \big(d_k(\mathcal T_P) + h_{{\mathcal{Y}}}(f(P))\big) + O(1)$$ holds. In particular, replacing $c$ by a larger real number $c(k,\mathcal Y, f)$ if necessary, we conclude that the following inequality $$h(P) \leq c(k,{\mathcal{Y}},f) \cdot (d_k(\mathcal T_P)+ h_{\mathcal{Y}}(f(P)))$$ holds. Uniformity results ================== Let $k$ be a number field. In this section we prove Theorem \[thm:stackthm\]. \[lem:desing\] Let $f \colon X \to Y$ be a proper surjective morphism of proper Deligne-Mumford stacks over $k$ which is representable by schemes. Let $h$ be a height function on $X$ and let $h_Y$ be a height function on $Y$ associated to an ample divisor with $h_Y \geq 1$. Assume Conjecture \[conj:Vojta\]. Suppose that the generic fibre $X_\eta$ of $f\colon X\to Y$ is smooth and of general type. There exists a proper Zariski closed substack $Z\subset X$ and a real number $c$ depending only on $k$, $X$, $Y$, and $f$, such that, for all $P$ in $X(k)\setminus Z$, the following inequality holds $$h(P) \leq c \cdot \big(h_Y(f(P)) + d_k({\mathcal{T}}_P)\big).$$ We may and do assume that $X$ and $Y$ are geometrically integral over $k$. Let $\mu \colon {\widetilde}{X}\to X$ be a desingularization of $X$; see [@temkin Theorem 5.3.2]. Note that ${\widetilde}{f} \colon {\widetilde}{X} \to Y$ is a proper surjective morphism of proper Deligne-Mumford stacks whose generic fibre is of general type. Define $X_{exc}\subset X$ to be the exceptional locus of $\mu \colon {\widetilde}{X}\to X$, so that $\mu$ induces an isomorphism of stacks from ${\widetilde}{X}\setminus \mu^{-1}(X_{exc})$ to $X\setminus X_{exc}$. Note that $X_{exc}$ is a proper closed substack of $X$, as $X$ is reduced. Let ${\widetilde}{h}$ be the height function on ${\widetilde}{X}$ associated to $h$, so that, for all ${\widetilde}{P}$ in ${\widetilde}{X}$, we have ${\widetilde}{h}({\widetilde}{P}) = h(P)$. As we are assuming Conjecture \[conj:Vojta\], it follows from Proposition \[prop:mainprop\] that there exists a proper Zariski closed substack ${\widetilde}{Z}\subset {\widetilde}{X}$ such that, for all ${\widetilde}{P}$ in ${\widetilde}{X}(k)\setminus {\widetilde}{Z}$, the following inequality $${\widetilde}{h}(P) \leq c \cdot \big( h_Y({\widetilde}{f}(P)) + d_k(\mathcal T_P)\big)$$ holds, where $c$ is a real number depending only on $k$, $Y$, $X$, and $f$. (Here we use that ${\widetilde}{X}\to X$ only depends on $X$.) Define $Z$ to be the closed substack $\mu({\widetilde}{Z})\cup X_{exc}$ in $X$. Note that $\mu$ induces an isomorphism from ${\widetilde}{X}\setminus \mu^{-1}(Z) $ to $X\setminus Z$. Therefore, we conclude that, for all $P$ in $X(k)\setminus Z$, the inequality $$h(P) = {\widetilde}{h}({\widetilde}{P}) \leq c \cdot \big(h_Y(f(P)) + d_k(\mathcal T_P)\big)$$ holds, where ${\widetilde}{P}$ is the unique point in ${\widetilde}{X}$ mapping to $X$. Since $U$ is constructible, we have that $U = \cup_{i=1}^n U_i$ is a finite union of locally closed substacks $U_i \subset Y$. Let $Y_i$ be the closure of $U_i$ in $Y$, let $X_i = X\times_Y Y_i$, and let $f_i: X_i \to Y_i$ be the associated morphism. Note that $U_i$ is open in $Y_i$. In particular, to prove the theorem, replacing $X$ by $X_i$, $Y$ by $Y_i$, $U$ by $U_i$, and $f:X\to Y$ by $f_i: X_i\to Y_i$ if necessary, we may and do assume that $U$ is open in $Y$. We now argue by induction on $\dim X$. If $\dim X = 0$, then the statement is clear. As we are assuming Conjecture \[conj:Vojta\], it follows from Lemma \[lem:desing\] that there exists a proper Zariski closed substack $Z \subset X$ and a real number $c_0>0$ depending only on $k$, $X$, $Y$, and $f$ such that, for all $P$ in $X(k)\setminus Z$, the inequality $$\begin{aligned} \label{ineq1} h(P) & \leq & c_0\cdot \big( h_Y(f(P)) + d_k(\mathcal T_P)\big) \end{aligned}$$ holds. Let $X_1,\ldots, X_s\subset Z$ be the irreducible components of $Z$. For $i\in \{1,\ldots,s\}$, let $Y_i = f(X_i)$ be the image of $Y_i$ in $Y$. Note that $f_i \colonequals f|_{X_i} \colon X_i\to Y_i$ is a proper morphism of proper integral Deligne-Mumford stacks which is representable by schemes. Moreover, for $t$ in the open subscheme $Y_i\cap U$ of $Y_i$, the proper variety $X_{i,t}$ is hyperbolic, as $X_{i,t}$ is a closed subvariety of the hyperbolic variety $X_t$. Let $h_i$ be the restriction of $h$ to $X_i$, and let $h_{Y_i}$ be the restriction of $h_Y$ to $Y_i$. Since $X_i$ is a proper Zariski closed substack of $X$, it follows that $\dim X_i < \dim X$. Therefore, by the induction hypothesis, we conclude that there is a real number $c_i>0$ depending only on $k$, $X_i$, $Y_i$, and $f_i$ such that, for all $P$ in $X_i(k)$, the following inequality $$\begin{aligned} \label{ineq2} h(P) = h_i(P) & \leq & c_i \cdot \big(h_{Y_i}(f_i(P)) + d_k(\mathcal T_{P})\big) = c_i \cdot \big(h_Y(f(P)) + d_k(\mathcal T_P)\big).\end{aligned}$$ holds. Let $c' \colonequals \max(c_1,\ldots,c_s)$. By (\[ineq2\]), we conclude that, for all $P$ in $Z(k)$, the inequality $$\begin{aligned} \label{ineq3} h(P) &\leq & c'\cdot \big( h_Y(f(P)) + d_k(\mathcal T_P)\big)\end{aligned}$$ holds. Combining (\[ineq1\]) and (\[ineq3\]), we conclude the proof of the theorem with $c\colonequals \max(c_0,c')$. \[lem:disc\_for\_rep\] Let $f \colon {\mathcal{X}}\to {\mathcal{Y}}$ be a proper surjective morphism of proper Deligne-Mumford stacks over $\mathcal{O}_K$ which is representable by schemes. If $P\in {\mathcal{X}}(k)$, then $d_k(\mathcal{T}_P) = d_k(\mathcal{T}_{f(P))})$. Since the normalization morphism of an integral algebraic stack is representable and ${\mathcal{X}}\to{\mathcal{Y}}$ is representable, we see that the morphism $\mathcal{T}_P\to\mathcal{T}_{f(P)}$ is representable. Therefore, we see that $\mathcal{T}_P\to \mathcal{T}_{f(P)}$ is proper surjective and representable by schemes. Moreover, it is birational, and it has a section generically. This implies that the morphism $\mathcal{T}_P\to \mathcal{T}_{f(P)}$ is a proper birational quasi-finite representable morphism. Since $\mathcal{T}_{f(P)}$ is normal, the lemma follows from Zariski’s Main Theorem for stacks. \[cor:thm\] Let $f \colon {\mathcal{X}}\to {\mathcal{Y}}$ be a proper surjective morphism of proper Deligne-Mumford stacks over $\mathbb{Z}$ which is representable by schemes. Let $X:={\mathcal{X}}_{\mathbb{Q}}$ and $Y:={\mathcal{Y}}_{\mathbb{Q}}$. Let $h$ be a height function on $X$ and let $h_Y$ be a height function on $Y$ associated to an ample divisor with $h_Y \geq 1$. Assume Vojta’s height conjecture (Conjecture \[conj:Vojta\]). Let $U\subset Y$ be a constructible substack such that, for all $t\in U$, the variety $X_t$ is smooth and hyperbolic. Then there is a real number $c > 0$ depending only on $k$, $Y$, $X$, and $f$ such that, for all $P$ in $X(k)$ with $f(P)$ in $U$, the following inequality holds $$h(P) \leq c \cdot \big(h_Y(f(P)) + d_k({\mathcal{T}}_{f(P)})\big).$$ Combine Theorem \[thm:stackthm\] and Lemma \[lem:disc\_for\_rep\]. Applications {#section:apps} ============ In this section we apply our main result (Theorem \[thm:stackthm\]) to some explicit families of hyperbolic varieties, and prove Theorems \[thm1\] and \[thm2\]. Application to curves {#section:apps_curves} --------------------- For $g\geq 2$ an integer, let $\mathcal M_g$ be the stack over $\mathbb Z$ of smooth proper genus $g$ curves. Let $\overline{\mathcal M}_g$ be the stack of stable genus $g$ curves. Note that $\mathcal M_g$ and $\overline{\mathcal M}_g$ are smooth finite type separated Deligne-Mumford stacks. Moreover, $\mathcal M_g\to \overline{\mathcal M}_g$ is an open immersion, and $\overline{\mathcal M}_g$ is proper over $\mathbb Z$ with a projective coarse space [@Kollar Theorem 5.1]. These properties of $\mathcal M_g$ and $\overline{\mathcal M}_g$ are proven in [@DeligneMumford]. We fix an ample divisor $H$ on $\overline{\mathcal M}_g$. If $X$ is a smooth projective curve of genus at least two over a number field $k$, we let $h \colon X(\overline{k})\to \mathbb R$ be the height with respect to the canonical embedding $X\to \mathbb P^{5g-6}_k$. Moreover, we define the height of $X$ to be the height of the corresponding $k$-rational point of $\overline{\mathcal M}_g$ with respect to the fixed ample divisor $H$ on $\overline{\mathcal M}_g$ (following Section \[section:heights\_on\_stacks\]). If $X$ is a smooth projective curve of genus at least two over $k$ and $P$ is a $k$-rational point of $X$, then the pair $(X,P)$ defines a point on the universal curve $\mathcal{M}_{g,1}$ over $\mathcal{M}_g$. We let $d_k(\mathcal T_{(X,P)})$ denote the discriminant of this point, as defined in Section \[section:discr\]. Since $U:=\mathcal M_g$ is open in $Y:=\overline{\mathcal M_g}$, we can apply Corolary \[cor:thm\] to the universal family of stable genus $g$ curves $f \colon X \to Y$. \[rem:faltings\_height\] In Theorem \[thm1\], one can also use the (stable) Faltings height $h_{\mathrm{Fal}}(X)$ of $X$ (instead of the height $h$ introduced above). Indeed, it follows from [@Faltings; @Pazuki] that the Faltings height $h_{\mathrm{Fal}}(X)$ is bounded by $ h(X) +c$, where $c$ is a real number depending only on the genus of $X$. Hyperbolic surfaces ------------------- Recall that, if $S$ is a smooth projective surface, then $c_1^2(S) = K_S^2$ and $c_2(S) = e(S)$ is the topological Euler characteristic. Moreover, by Noether’s lemma, they are related by the following equality: $$\chi(S,\mathcal O_S) = \dfrac{c_1(S)^2 + c_2(S)^2}{12}.$$ In particular, the information of $K_S^2$ and $\chi(S)$ is equivalent to the data of $c_1(S)$ and $c_2(S)$. Finally, we note that $c_2(S) \geq 1$ for any surface of general type $S$ [@beauville X.1 and X.4]. A smooth proper morphism $f:X\to Y$ of schemes is a canonically polarized smooth surface over $Y$ if, for all $y$ in $Y$, the scheme $X_y$ is a connected two-dimensional scheme and $\omega_{X_y/k(y)}$ is ample. If $a$ and $b$ are integers, we let $\mathcal M_{a,b}$ over $\mathbb Z$ be the stack of smooth canonically polarized surfaces $S$ with $c_1(S)^2 = a $ and $c_2(S) = b$. Note that $\mathcal M_{a,b}$ is a finite type algebraic stack over $\mathbb Z$ with finite diagonal (cf. [@mm; @tankeev]). \[lem:smoothness\_argument\] If $S $ is a smooth hyperbolic surface over a field $k$, then $S$ is canonically polarized. If $S$ is a (smooth) minimal surface of general type, then the canonical model $S^{c}$ is obtained by contracting all rational curves with self intersection $-2$ [@Liu Chapter 9]. Consequently, the singularities on a singular surface in $\mathcal M_{a,b}(k)$ are rational double points arising from the contraction of these $-2$ curves. As having a $-2$ rational curve would contradict $S$ being hyperbolic, we see that $S^{c}$ must be smooth, and thus equal to $S$. As the canonical bundle on $S^c$ is ample, we conclude that $S$ is canonically polarized. Let $\mathcal M_{a,b}^{h}\subset \mathcal M_{a,b}$ be the substack of hyperbolic surfaces, i.e., for a scheme $S$, the objects $f:X\to S$ of the full subcategory $\mathcal M_{a,b}^{h}(S)$ of $\mathcal M_{a,b}(S)$ satisfy the property that, for all $s$ in $S$, the surface $X_s$ is hyperbolic (Definition \[defn:hyperbolic\]). We do not know of any result on the algebraicity of $\mathcal M_{a,b}^h$ (nor the algebraicity of $\mathcal M_{a,b}^h\times_{\mathbb{Z}}\operatorname{Spec}{\mathbb{C}}$). However, if $S$ is a minimal projective surface of general type over $\mathbb C$ and $c_1^2(S) > c_2(S)$, then Bogomolov proved [@bogomolov] that $S$ contains only a finite number of curves of bounded genus, and thus $S$ contains only finitely many rational and elliptic curves. In [@miyaoka Theorem 1.1] Miyaoka proved a more effective version of Bogomolov’s result, showing that in fact the canonical degree of such curves is bounded in terms of $c_1^2$ and $c_2$. Using these results we are able to prove the following. \[lem:constructibility\] If $a>b$, then $\mathcal M_{a,b}^h\times_{{\mathbb{Z}}}\operatorname{Spec}\mathbb{C}$ is a constructible substack of ${\mathcal{M}}_{a,b} \times_{{\mathbb{Z}}} \operatorname{Spec}\mathbb{C}$. Let $a$ and $b$ be integers such that $a>b$. Let $N$ be an integer such that, for all $S$ in $\mathcal M_{a,b}(\mathbb C)$, the ample line bundle $\omega_{S/\mathbb C}^{\otimes N}$ is very ample. In particular, $S$ is embedded in $\mathbb P^n \cong \mathbb P(\mathrm{H}^0(S,\omega_{S/\mathbb C}^{\otimes N}))$. Let $\mathrm{Hilb}_{a,b}$ be the Hilbert scheme of $N$-canonically embedded smooth surfaces, and note that $\mathcal M_{a,b} = [\mathrm{Hilb}_{a,b}/\mathrm{PGL}_{n+1}]$. Let $\mathrm{H}_d$ be the Hilbert scheme of (possibly singular) curves of canonical degree $d$ in $\mathbb P^n$. Let $\mathrm{H}_d^{\textrm{int}}$ be the subfunctor of geometrically integral curves. Since the universal family over $\mathrm{H}_d$ is flat and proper, the subfunctor $\mathrm{H}_d^{\textrm{int}}$ is an open subscheme of $\mathrm{H}_d$; see [@GW Appendix E.1.(12)]. Let ${\mathcal{W}}_{a,b,d} \subset \mathrm{H}_d^{\textrm{int}}\times \mathrm{Hilb}_{a,b}$ be the incidence correspondence subscheme parametrizing parametrizing pairs $(C,S)$ where the curve $C$ is inside the surface $S$. (Note that ${\mathcal{W}}_{a,b,d}$ is a closed subscheme of $\mathrm{H}_d^{\textrm{int}}\times \mathrm{Hilb}_{a,b}$.) By Miyaoka’s theorem [@miyaoka Theorem 1.1], there exist integers $d_1,\ldots,d_m$ which depend only on $a$ and $b$ with the following property. A surface $S \in \mathcal M_{a,b}(\mathbb{C})$ is hyperbolic if and only if, for all $i=1,\ldots,m$, it does not contain an integral curve of degree $d_i$. Note that, by Chevalley’s theorem, for all $d\in {\mathbb{Z}}$, the image of the composed morphism $${\mathcal{W}}_{a,b,d} \subset \mathrm{H}_d \times\mathrm{Hilb}_{a,b}\to \mathrm{Hilb}_{a,b}\to \mathcal M_{a,b}$$ is constructible. Let $\mathcal M_{a,b,d_i}$ be the stack-theoretic image of ${\mathcal{W}}_{a,b,d_i}$ in $\mathcal M_{a,b}$. Since a finite union of constructible substacks is constructible, the union $\bigcup_{i=1}^m \mathcal M_{a,b,d_i} $ is a constructible substack of $\mathcal M_{a,b}$. Finally, by construction, a surface $S$ in $\mathcal M_{a,b}(\mathbb C)$ is hyperbolic if and only if it is not (isomorphic to an object) in the constructible substack $\bigcup_{i=1}^m \mathcal M_{a,b,d_i} $. As the complement of a constructible substack is constructible, we conclude that $\mathcal M_{a,b}\times_{\mathbb{Z}}\operatorname{Spec}\mathbb{C}$ is a constructible substack of $\mathcal M_{a,b}\times_{\mathbb{Z}}\mathbb{C}$. Let $\mathcal{U}_{a,b}\to \mathcal{M}_{a,b}$ denote the universal family. We let $\overline{\mathcal M_{a,b,\mathbb Q}}$ be a compactification of $\mathcal M_{a,b,\mathbb Q}$ with a projective coarse moduli space; see [@hacking Section 2.5] (or [@Kollar Corollary 5.6]) for an explicit construction of such a compactification. (As the stack of smooth canonically polarized surfaces is open in the stack of canonical models, it suffices to compactify the latter, as is achieved in *loc. cit.* for all $a$ and $b$.) We now choose $\overline{\mathcal M}_{a,b}$ to be a compactification of $\mathcal M_{a,b}$ over ${\mathbb{Z}}$ whose generic fibre $\overline{\mathcal M}_{a,b}\times_{\mathbb{Z}}\operatorname{Spec}{\mathbb{Q}}$ is isomorphic to $\overline{\mathcal M_{a,b,{\mathbb{Q}}}}$, and we also choose a representable proper morphism $\overline{\mathcal{U}}_{a,b}\to \overline{\mathcal{M}}_{a,b}$ extending the universal family over $\mathcal{M}_{a,b}$ compatibly with the universal family over $\overline{\mathcal{M}_{a,b,\mathbb{Q}}}$. If $S$ is a smooth projective canonically polarized hyperbolic surface over a number field $k$, we let $h:S(\overline{k})\to \mathbb R$ be the height with respect to the very ample divisor $\omega_{S/k}^{\otimes 34}$ (see [@tankeev]). Moreover, we define the height of $S$ in $\mathcal M_{a,b,{\mathbb{Q}}}(k)$ to be the height of the corresponding $k$-rational point of $\overline{\mathcal M}_{a,b}$ with respect to some fixed ample divisor $H$ on $\overline{\mathcal M}_{a,b,{\mathbb{Q}}}$ (following Section \[section:heights\_on\_stacks\]). If $S$ is a smooth projective surface and $P$ is a $k$-rational point of $S$, then the pair $(S,P)$ defines a point on the universal surface $\mathcal{U}_{a,b}$ over $\mathcal{M}_{a,b}$. We let $d_k(\mathcal T_{(S,P)})$ denote the discriminant of this point, as defined in Section \[section:discr\]. By Lemma \[lem:constructibility\] and standard descent arguments (cf. [@stacks-project Tag 02YJ]), we conclude that ${\mathcal{M}}^h_{a,b} \times_{{\mathbb{Z}}} \operatorname{Spec}\mathbb{Q}$ is a constructible substack of ${\mathcal{M}}_{a,b}\times_{{\mathbb{Z}}}\operatorname{Spec}\mathbb{Q}$. Also, a smooth hyperbolic surface is canonically polarized by Lemma \[lem:smoothness\_argument\]. Therefore, the result follows from an application of Corollary \[cor:thm\] to the universal family over ${\mathcal{Y}}:= \overline{\mathcal{M}_{a,b}}$, with $Y:= \overline{\mathcal M}_{a,b,{\mathbb{Q}}}$, and the constructible substack $U:= \mathcal M_{a,b,\mathbb{Q}}^h$ in $Y$. There are many examples of surfaces of general type with $c_1^2 > c_2$. Some of the simplest examples are surfaces $S$ with ample canonical bundle such that there exist a smooth proper curve $C$ and a smooth proper morphism $S\to C$ (see for instance [@Kodaira]).

--- abstract: 'The Probabilistic Object Detection Challenge evaluates object detection methods using a new evaluation measure, Probability-based Detection Quality (PDQ), on a new synthetic image dataset. We present our submission to the challenge, a fine-tuned version of Mask-RCNN with some additional post-processing. Our method, submitted under username pammirato, is currently second on the leaderboard with a score of 21.432, while also achieving the highest spatial quality and average overall quality of detections. We hope this method can provide some insight into how detectors designed for mean average precision (mAP) evaluation behave under PDQ, as well as a strong baseline for future work.' author: - | Phil Ammirato\ UNC-Chapel Hill\ ammirato@cs.unc.edu\ - | Alexander C. Berg\ UNC-Chapel Hill\ aberg@cs.unc.edu bibliography: - 'final.bib' title: 'A Mask-RCNN Baseline for Probabilistic Object Detection' --- Introduction ============ [^1] Object detection is a both very popular task in the computer vision research community and a useful tool for many real world applications. Many such applications are in robotics. A robot aided with information of the objects around it will be better equipped to have successful interactions with its environment. Traditionally, object detection methods are designed and evaluated for the Mean Average Precision (mAP) metric. This metric is very useful for evaluating detection methods for a variety of tasks, but may not be ideal for robotics contexts. mAP ranks the bounding boxes output by a detection system based on each box’s score. The raw values of the scores are hard to interpret, other than a higher score is means the system is more confident. The ranking system generally encourages the detectors to output many detections for each image, as usually this will only increase its mAP score, as shown in [@pdq]. This large output may not be ideal for a robotics context, as processing resources may be limited, and there is no reliable/consistent measure for confidence. Recently, a new evaluation technique for object detectors was proposed, the Probability-based Detection Quality (PDQ) [@pdq]. This measure treats detection scores as real probabilities, not just as means for a ranking mechanism. It also allows for some uncertainty to be modeled in the objects location, represented as a *probabilistic bounding box* (pbox). A covariance matrix is given for each corner of a pbox, describing the uncertainty of the box’s shape. This gives a much more felixble representation that a traditional bounding box, which is very rigid in what is an object and what is not. The goal of this work is to establish a strong baseline for probabilistic object detection using an existing well known object detector, Mask-RCNN [@he2017maskrcnn]. We first fine-tune a Mask-RCNN model trained for MSCOCO [@lin2014microsoftCOCO] mAP object detection for the detection on the PODC data. We then develop some post-processing routine on the detectors output to be better suited to PDQ evaluation as opposed to mAP. We tested our work as a part of the Probablistic Object Detection Challenge (PODC) at CVPR 2019 and achevied 2nd place. Mask-RCNN ========= Mask-RCNN [@he2017maskrcnn] is a very popular deep-learning method for object detection and instance segmentation that achieved state-of-the art results on the MSCOCO[@lin2014microsoftCOCO] dataset when published. While a few detectors have since passed Mask-RCNN in mAP performance, they have done so by only a few points and are usually based on the Mask-RCNN architecture. Mask-RCNN is a two-stage recognition pipeline. In its first stage, features are extracted from the image using a backbone convolutional neural network (CNN), and class agnostic region proposals are predicted. These proposals are then refined and classified in the second stage, to become either labeled bounding boxes for object detection or segmentation masks for the instance segmentation task. We use an implementation of Mask-RCNN based in Pytorch [@pytorch2017], [@maskrcnn_benchmark_code]. The authors of [@maskrcnn_benchmark_code] provide models trained for object detection and instance segmentation on MSCOCO, using backbone networks that are pre-trained on the ImageNet [@russakovsky2015imagenet] classification dataset. Their is also code for training models from scratch or fine-tuning. See their public github repository for more details. We base our system for the PODC challenge on the models provided by this code base. Training for PODC data ====================== We started building our system using a model from [@maskrcnn_benchmark_code] that used a ResNet-101 [@resnet] backbone with a Feature Pyramid Network [@feature_pyramid], and was trained for both object detection and instance segmentation on the 80 object classes of MSCOCO. This model achieves 42.2 mAP on the 80 MSCOCO classes. Our first step was to remove all mask prediction heads and the object detection classifier heads for the 50 classes in MSCOCO that are not included in PODC. This left us with an object detector for 30 classes designed and trained to maximize mAP on MSCOCO, we will refer to this as Mask-RCNN-30. The lack of training data for this task presents a unique domain transfer challenge. Usually systems aim to transfer from synthetic images to real world images, but in this task the test data is synthetic while there is some real training data available. The PODC validation and testing data is generated from a high-fidelity simulation, and so the images are likely from a different domain than the real-world images collected in MSCOCO. To see how well our initial model can generalize to this new domain, we test test Mask-RCNN-30 on one scene of the PODC validation data, which will refer to as PODC-val0. As shown in Table \[tab:training\_mAP\], the model trained only on MSCOCO does not perform well on PODC-val0 when compared to its performance on MSCOCO. Since there is no training data associated with the PODC, we looked for another source of synthetic data to help Mask-RCNN-30 generalize better. We first looked at the AI2-Thor [@ai2thor] environment, another synthetic renderer built for simulating robotic motion around indoor home-style scenes. Unfortunately, the object class set of AI2-Thor does not fully cover the set of objects in PODC, only about 9 of PODC classes are labeled. In fact there are object classes, such as oven, that are present in the images of AI2-Thor but are not labeled. These false negatives in the ground truth data would likely cause confusion during training. The SunCG [@suncg] dataset, however, has a much higher overlap of labeled classes with PODC with 25 of the 30 classes covered. House3D [@house3d] uses the 3D scene models and annotations to provide environments similar to AI2-Thor. We use tools from the publicly available House3d code to generate a training set of 173,250 images with bounding box annotations, which we refer to as House3D-train. While the SunCG data is synthetic, it is much less realistic than the PODC validation and testing data. We fine-tune Mask-RCNN-30 on both the original MSCOCO images it was trained on and the new House3D-train set, in a weak attempt at domain randomization with respect to image source. The hope is that the model will learn to be robust to different image sources, real or synthetic. We can see in Table \[tab:training\_mAP\] that this does help performance on PODC-val0 with respect to mAP. We add one more data augmentation during training, namely adjusting the image brightness and contrast via Pytorch transforms. There are two intended effects of this augmentation: to further improve the generalization of the model to different image source domains, and to help with performance on the ‘night’ scenes present in PODC. We can see another small improvement in mAP in Table \[tab:training\_mAP\]. In Table \[tab:training\_mAP\], we can see as we improve mAP, PDQ scores actually decrease. This is mostly due to an increase in false positives, and we will give some heursitcs for remedying this in the next section. Model mAP PDQ ------------------------- ------- ------ Mask-RCNN-30 12.09 7.2 MSCOCO + SunCG 13.53 6.16 MSCOCO + SunCG + jitter 14.08 5.73 : mAP and PDQ on PODC-val0 for different training data. mAP improves as we add more variation to the training data. PDQ gets worse mainly due to an increase of false positives. []{data-label="tab:training_mAP"} {width="80.00000%"} {width="80.00000%"} Post Processing for PODC evaluation =================================== Even though we have managed to improve our detectors mAP score on PODC-val0, the PDQ score is still very low and has been getting worse. We identify various causes of our low PDQ score, and propose simple post-processing techniques to improve them. These improvements are not meant to solve the problem of adapting detectors for probabilistic object detection. They hopefully provided some insights into the problem, and will serve as a useful baseline for future techniques. PDQ Calculation --------------- We first give a short outline of how PDQ [@pdq] is calculated, so the motivations for our solutions will be clear. There are two important two sub-measures of PDQ we must first compute, spatial quality ($Q_{S}$) and label quality ($Q_{L}$). Let $T_{j}^{i}$ be the probability distribution across all possible object classes assigned to the $j^{th}$ in the $i^{th}$ image. Then the label quality of the detection, $D_{j}^{i}$ with respect to the $k^{th}$ ground truth object, $G_{k}^{i}$ is: $$Q_{L} (G_{k}^{i}, D_{j}^{i})= T_{j}^{j}(c_{k}^{i})$$ where $c_{k}^{i})$ is the object class of $G_{k}^{i}$. This is a number on $[0,1]$, that is equal to the detector’s score for the ground truth object, regardless of any other detections or objects. The spatial quality, $Q_{S}$, is calculated based on the ground truth’s objects segmentation mask and the probabilistic bounding box outputted by the detector. Essentially, assigning higher probabilities to pixels that belong to the object as foreground improve the score, and including any background pixels outside of the ground truth bounding box hurt the score. The pairwise PDQ for one detection and ground truth is then calculated using the geometric mean of these two measures. $$pPDQ(G_{k}^{i}, D_{j}^{i}) = \sqrt{Q_{S} * Q_{L}}$$ A matching routine then finds the best matching of detections and ground truth objects. The average score after the matching routine for all true positive detections is $aPDQ$. the final PDQ score is calculated. The score is scaled by three quantities: the number of false positives, $N_{fp}$, the number of false negatives, $N_{fn}$ and the number of true postives, $N_{tp}$. $$PDQ = \frac{1}{N_{fp} + N_{fn} +N_{tp}} * N_{tp}*aPDQ$$ False Negatives --------------- Perhaps the largest source of error when evaluating raw Mask-RCNN output with PDQ is the large number of false negatives. The simple solution is to remove all detections with score less than .5. We experimented with changing this score threshold, but .5 consistently gave the best performance as show in Table \[tab:training\_thresh\]. Score Threshold PDQ TP FP FN ----------------- ----------- ------ ------- ------- 0.0 5.73 7262 41771 7978 0.3 14.78 5331 6412 9909 0.4 15.58 4994 4155 10246 **0.5** **15.79** 4694 2875 10546 0.6 15.63 4296 1988 10944 0.7 15.21 3915 1210 11325 : How PDQ score changes when changing the score threshold for discarding detections. This includes the heuristic of setting all detection scores to 1.0 and a set covariance of 7.5. TP=True Positives, FP=False Positives, FN=False Negatives. []{data-label="tab:training_thresh"} Label Quality ------------- The label quality is only counted for true positive detections. The only penalty for assigning high scores to false positive detections comes from the spatial quality measure. We found that given the large weight of label quality in the final PDQ score, and the other influences on spatial quality besides detection score, reassigning a value of 1.0 to all detection scores resulted in the highest PDQ. This seems to not be in the spirit of PDQ as now the scores have even less meaning than the original detector output, while PDQ seems to encourage a meaningful uncertainty measure. We hope that with this baseline established, future work will have to find some better representation for detection scores to improve. Confusing Objects {#sec:confuse} ----------------- While removing all low scoring detections was a simple and effective solution for reducing false positives, it also removed a lot of false positives. We added another simple heuristic to add back in certain low scoring detections that had boxes with high intersection over union (IOU) with other low scoring detections. Essentially here we are trying to add in detections that are likely objects of interest, as the detector has at least two outputs for the object, but have ambiguous class to the detector. As can be seen in the equation 3, while $N_{fn}$ &gt;0, adding a true positive detection has improve the score more than it will be reduced from adding a false positive. So adding keeping two detections, even though one is almost guaranteed to be a false positive, is worth it as long as one is a true positive. This method increases the performance only a small amount, and is likely not worth the expensive computation. Probabilistic Bounding Box -------------------------- We found that adding some fixed value in the covariance matrix to our bounding boxes was slightly better than nothing, but scaling the covariance according to the box size was even better. After trying various values, adding a covariance between 20-30% of the bounding box size was best. For example, scaling at 30%, a bounding box with width of 100 and height of 50 would be given a covariance matrix of: $$\begin{bmatrix} 30.0 & 0.0 \\ 0.0 & 15.0 \\ \end{bmatrix}$$ for both the top left and bottom right corners. See Table \[tab:training\_cov\] for an ablation study. Covariance PDQ Avg Spatial ------------ ----------- ------------- -- -- 0.0 6.51 17.50 5.0 15.45 44.26 7.5 15.79 45.04 10.0 15.97 45.26 15.0 16.12 45.17 20.0 16.13 44.75 25.0 16.08 44.19 10% 16.06 45.89 **20%** **16.36** **46.05** : How PDQ score changes when changing how to set the covariance of every bounding box. Compares both setting the covariance to a set number (i.e. 5.0) and as a function of the box dimensions (i.e. 20% box). Includes heuristics of removing detections with scores below 0.5 and then setting all detection scores to 1.0. *Avg Spatial* is the average spatial quality as defined by the evaluation code provided by the PODC. PDQ goes up slightly when covariance goes from 10.0 to 15.0 even though spatial quality decreases, because the number of True Positives also changes slightly. []{data-label="tab:training_cov"} Reducing Box Size ----------------- The spatial quality measure can be reduced greatly by assigning high probabilities to pixels that are truly in the background. Since all of our detections have a score of 1.0, this can really hurt our overall PDQ. We reduce the bounding box size of all detections by 10% in width and height respectively. This gives us a smaller box with high confidence centered on each object, and our large covariance values allow us to capture the rest of the foreground pixels without such a high probability assigned to the background. Box Reduction Factor PDQ Avg Spatial ---------------------- ------- ------------- -- -- 0.0 16.36 46.05 **0.1** 18.11 55.00 0.2 14.72 37.76 : How PDQ score changes when reducing the size of every bounding box. Includes heuristics of removing detections with scores below 0.5 and then setting all detection scores to 1.0, as well as setting the covariance as 20% of the reduced box size. *Avg Spatial* is the average spatial quality as defined by the evaluation code provided by the PODC. []{data-label="tab:training_boxsize"} Final Model Details =================== For our final model, we fine-tune Mask-RCNN-30 on the MSCOCO + SunCG + jitter training set for 10,000 iterations at a batch size of 8 and learning rate or .0005. We then add the PODC validation data to the training set and fine-tune for another 5,000 iterations. We post process the detections by thresholding at a score of 0.5, then setting all scores to 1.0. We then add back in boxes as in section \[sec:confuse\]. Finally we reduce the bounding boxes by 10% and add in a covariance based on 30% of the box width and height. Conclusion and Future Work ========================== With a final score of 21.432, we see lots of room for improvement in future work, though the lack of training data is a limiting factor. Most of the improvements in this work are based on simple heuristics taking advantage of the structure of the PDQ calculation. We hope future work can use this to improve, and provide more meaningful uncertainty predictions. [^1]: Supported by NSF NRI grant 1526367

--- abstract: 'We show how a cluster-tilted algebra of finite representation type is related to the corresponding tilted algebra, in the case of algebras defined over an algebraically closed field.' address: - | Institutt for matematiske fag\ Norges teknisk-naturvitenskapelige universitet\ N-7491 Trondheim\ Norway - | Institutt for matematiske fag\ Norges teknisk-naturvitenskapelige universitet\ N-7491 Trondheim\ Norway author: - Aslak Bakke Buan - Idun Reiten title: 'From tilted to cluster-tilted algebras of Dynkin type' --- \[section\] \[lem\][Proposition]{} \[lem\][Corollary]{} \[lem\][Theorem]{} \[lem\][Remark]{} \[lem\][Definition]{} Introduction ============ The cluster categories ${\operatorname{\mathcal C}\nolimits}_H$ of finite dimensional hereditary algebras $H$ were introduced and studied in [@bmrrt] in order to give a categorical model for some of the central ingredients of the theory of cluster algebras introduced by Fomin and Zelevinsky [@fz]. In particular tilting theory for cluster categories was investigated. Associated with a tilting module $T$ over a hereditary algebras is, in addition to the tilted algebra ${\Lambda}= {\operatorname{End}\nolimits}_{H}(T)^{{\operatorname{op}\nolimits}}$, also the cluster-tilted algebra ${\Gamma}= {\operatorname{End}\nolimits}_{{\operatorname{\mathcal C}\nolimits}_H}(T)^{{\operatorname{op}\nolimits}}$, investigated in [@bmr1; @bmr2; @bmr3]. It is of interest to compare the algebras ${\Lambda}$ and ${\Gamma}$ associated with the same tilting module. In general we do not know how to construct ${\Gamma}$ from ${\Lambda}$. Here we deal with the case where $H$ is of Dynkin type over an algebraically closed field $K$. In this case we show how the cluster-tilted algebra ${\Gamma}$ is determined by the quiver and relations of the tilted algebra ${\Lambda}$. Background ========== In this section we recall some notions and basic results. Let $K$ be an algebraically closed field, and $Q$ a finite quiver with no oriented cycles. Then the path algebra $H = KQ$ is a hereditary finite-dimensional algebra. Let ${\operatorname{mod}\nolimits}H$ be the category of (left) finitely generated modules over $H$ and let $\underline{r}$ denote the Jacobson radical of an algebra. Tilting modules and tilted algebras ----------------------------------- A module $T$ in ${\operatorname{mod}\nolimits}H$ is called a [*tilting module*]{} if ${\operatorname{Ext}\nolimits}^1_H(T,T) = 0$ and if $T$ is maximal with respect to this property, i.e. if ${\operatorname{Ext}\nolimits}^1_H(T \amalg X,T \amalg X) = 0$, then $X$ is a direct summand in a direct sum of copies of $T$. The endomorphism-algebras of tilting modules are called [*tilted algebras*]{}. See [@h] for basic properties of tilted algebras and further references. Mutation of quivers ------------------- In connection with their definition of cluster algebras, Fomin and Zelevinsky [@fz] defined a [*mutation*]{}-operation on skew-symmetrizable integer matrices. Mutation of quivers can be seen as a special case of this, and can be defined for any finite quiver with no loops and no oriented cycles of length two. A quiver is called [*double path avoiding*]{} if it does not contain multiple arrows, and if any quiver obtained by repeated mutating and/or factoring (i.e. removing one or more vertices) also contains no multiple arrows. See Section 1 of [@bmr3] for more details on this. A fact that we will freely use here is that the quiver of a cluster-tilted algebra of finite representation type is double path avoiding. Cluster categories and cluster-tilted algebras ---------------------------------------------- Cluster categories were introduced in [@bmrrt] in order to give a categorical approach to cluster algebras as defined by Fomin and Zelevinsky [@fz]. In [@ccs1] a category was introduced for Dynkin quivers of type $A_n$, which was shown to be equivalent to the cluster category. The idea was to model the combinatorics of clusters on the set of tilting objects in a cluster category. The first link from cluster algebras to tilting theory was given by Marsh, Reineke and Zelevinsky [@mrz]. Given a hereditary algebra $H$, the cluster category is defined as follows. Let ${\operatorname{\mathcal D}\nolimits}= D^b({\operatorname{mod}\nolimits}H)$ be the bounded derived category of $H$. This category has a triangulated structure. Let $[1]$ denote the suspension functor. It also has AR-triangles, by [@h]. Let $\tau$ denote the corresponding auto-equivalence on ${\operatorname{\mathcal D}\nolimits}$ with quasi-inverse $\tau^{-1}$. Consider the auto-equivalence $F = \tau^{-1}[1]$. The cluster category ${\operatorname{\mathcal C}\nolimits}$ is the orbit category ${\operatorname{\mathcal C}\nolimits}/F$. That is, the objects of ${\operatorname{\mathcal C}\nolimits}$ are the objects of ${\operatorname{\mathcal D}\nolimits}$, while ${\operatorname{Hom}\nolimits}_{{\operatorname{\mathcal C}\nolimits}}(X,Y) = \amalg_i {\operatorname{Hom}\nolimits}_{{\operatorname{\mathcal D}\nolimits}}(X, F^i Y)$. The indecomposable objects of ${\operatorname{\mathcal D}\nolimits}$ are stalk-complexes, and we identify ${\operatorname{mod}\nolimits}H$ with the full subcategory of ${\operatorname{\mathcal D}\nolimits}$ where the objects are stalk-complexes in degree 0. It was shown by Keller [@k] that the cluster category is also triangulated and that the canonical functor ${\operatorname{\mathcal D}\nolimits}\to {\operatorname{\mathcal C}\nolimits}$ is a triangle functor. It was shown in [@bmrrt] that ${\operatorname{\mathcal C}\nolimits}$ is a Krull-Schmidt category with AR-triangles, and that the canonical functor ${\operatorname{\mathcal D}\nolimits}\to {\operatorname{\mathcal C}\nolimits}$ preserves AR-triangles. An object $T$ of ${\operatorname{\mathcal C}\nolimits}$ is called a (cluster-)tilting object if ${\operatorname{Ext}\nolimits}^1_{{\operatorname{\mathcal C}\nolimits}}(T,T) = 0$ and $T$ is maximal with respect to this property, i.e. if ${\operatorname{Ext}\nolimits}^1_{{\operatorname{\mathcal C}\nolimits}}(T \amalg X,T \amalg X) = 0$, then $X$ is a direct summand in a direct sum of copies of $T$. The endomorphism-algebra ${\operatorname{End}\nolimits}_{{\operatorname{\mathcal C}\nolimits}}(T)^{{\operatorname{op}\nolimits}}$ of a tilting object $T$ is called a [*cluster-tilted algebra*]{}. A cluster-tilted algebra coming from $H$ is of finite representation type if and only if $H$ is of finite representation type [@bmr1]. We also need the following results from [@bmrrt], [@bmr1] and [@bmr2] about tilting objects and cluster-tilted algebras. \[gb\] Let ${\operatorname{\mathcal C}\nolimits}= {\operatorname{\mathcal C}\nolimits}_H$ for a hereditary algebra $H$, and let $T$ be a tilting object in ${\operatorname{\mathcal C}\nolimits}$. - [Let $Q$ be the quiver of a cluster-tilted algebra ${\Gamma}= {\operatorname{End}\nolimits}_{{\operatorname{\mathcal C}\nolimits}}(T)^{{\operatorname{op}\nolimits}}$. Let $Q'$ be the quiver obtained by mutating $Q$ at the vertex $k$. Furthermore, let $T_k$ be the indecomposable direct summand of $T$ corresponding to the vertex $k$ and assume $T = \bar{T} \amalg T_k$. Then there is a tilting object $T' = \bar{T} \amalg T_k^{\ast}$, such that $Q'$ is the quiver of ${\operatorname{End}\nolimits}_{{\operatorname{\mathcal C}\nolimits}}(T')^{{\operatorname{op}\nolimits}}$. ]{} - [Assume $H$ is of finite type, and let $T_a$ and $T_b$ be indecomposable direct summands in $T$. Then ${\operatorname{dim}\nolimits}{\operatorname{Hom}\nolimits}_{{\operatorname{\mathcal C}\nolimits}}(T_a,T_b) \leq 1$. If there is a non-zero map $T_a \to T_b$ in ${\operatorname{\mathcal C}\nolimits}$, which lifts to a map $T_a \to FT_b$ in ${\operatorname{\mathcal D}\nolimits}$ with $T_a, T_b$ in ${\operatorname{mod}\nolimits}H$, then any non-zero map $T_b \to T_a$ in ${\operatorname{\mathcal C}\nolimits}$ lifts to a map $T_b \to T_a$ in ${\operatorname{\mathcal D}\nolimits}$, with $T_a, T_b$ in ${\operatorname{mod}\nolimits}H$. ]{} Cluster-tilted algebras of finite representation type ----------------------------------------------------- The proof of the main result of this paper strongly depends on results in [@bmr3]. In that paper, the main aim was to show that the cluster-tilted algebras of finite type are determined by their quivers and to describe the relations. Some of the results and notions needed to prove this are also useful here, so we recall them. An oriented cycle in a quiver is called [*full*]{} if there are no repeated vertices and if the subquiver generated by the cycle contains no further arrows. If there is an arrow $i \to j$ in a quiver $Q$, then a path from $j$ to $i$ is called [*a shortest path*]{} if the induced subquiver is a full cycle. Let ${\Gamma}= KQ/I$ be a cluster-tilted algebra. The elements in $I$ are called [*relations*]{} if they are linear combinations $k_1 \omega_1 + \cdots + k_m \omega_m$ of paths $\omega_i$ in $Q$, all starting in the same vertex and ending in the same vertex, and with each $k_i$ non-zero in $K$. If $m=1$, we call the relation a [*zero-relation*]{}. If $m=2$, we call it a [*commutativity-relation*]{} (and say that the paths $\omega_1$ and $\omega_2$ commute). A relation $\rho$ is called [*minimal*]{} if whenever $\rho = \sum_i \beta_i \circ \rho_i \circ \gamma_i$, where $\rho_i$ is a relation for every $i$, then there is an index $j$ such that both $\beta_j$ and $\gamma_j$ are scalars. For ease of notation we use the same symbol $\omega$ to denote a path, the corresponding element in the path algebra, and the corresponding element in $KQ/I$. The following was the main result of [@bmr3]. A consequence is that a cluster-tilted algebra of finite representation type is determined by its quiver, up to isomorphism. \[gc\] Let ${\Gamma}= KQ/I$ be a cluster-tilted algebra of finite representation type, and let $i,j$ be vertices in $Q$. - [The ideal $I$ is generated by minimal zero-relations and minimal commutativity relations.]{} - Assume there is an arrow $i \to j$. Then there are at most two shortest paths from $j$ to $i$. - [If there is exactly one, then this is a minimal zero-relation.]{} - [If there are two, $\omega$ and $\mu$, then $\omega$ and $\mu$ are not zero in ${\Gamma}$, and there is a minimal relation $\omega + \lambda \mu$ for some $\lambda \neq 0$ in $K$. ]{} - [Up to multiplication by non-zero elements of $K$, there are no other minimal zero-relations or commutativity relations than the ones coming from (b).]{} We also need some results which were used to prove the above theorem in [@bmr3]. \[help\] Let ${\Gamma}= KQ/I$ be a cluster-tilted algebra of finite representation type, and let $i,j$ be vertices in $Q$. A relation $\rho$ is minimal if and only if $\rho = \rho' +\rho''$, where $\rho'$ is a minimal commutativity relation or a minimal zero-relation, and $\rho''$ is any relation. [Given an arrow $i \to j$, and two distinct shortest paths $\omega$ and $\mu$ from $j$ to $i$, $\omega$ and $\mu$ are both non-zero, and there is a minimal commutativity relation $\omega + \lambda \mu$ for some $\lambda$ in $K$, where $\lambda$ can be assumed to be $-1$.]{} [For any minimal commutativity relation involving paths $\omega$ and $\mu$, the subquiver generated by these paths has the form $$\xy \xymatrix{ & \cdot \ar[r] & \cdot \ar@{.}[r] & \cdot \ar[r] & \cdot \ar[dr] & \\ j \ar[dr] \ar[ur] & & & & & i \ar[lllll] \\ & \cdot \ar[r] & \cdot \ar@{.}[r] & \cdot \ar[r] & \cdot \ar[ur] & } \endxy$$ ]{} [In a full cycle in $Q$ of length $n$, the composition of $n-2$ arrows is always non-zero.]{} [If the quiver $Q$ is an oriented cycle of length $n$, then $I = \underline{r}^{n-1}$.]{} Main result =========== Let $K$ denote an algebraically closed field, and let $H$ be the path algebra over $K$ of some Dynkin quiver. Let $T$ be a basic tilting module in ${\operatorname{mod}\nolimits}H$, and let ${\Lambda}= {\operatorname{End}\nolimits}_H(T)^{{\operatorname{op}\nolimits}}$ and ${\Gamma}= {\operatorname{End}\nolimits}_{{\operatorname{\mathcal C}\nolimits}_H}(T)^{{\operatorname{op}\nolimits}}$ be the corresponding tilted and cluster-tilted algebra, respectively. Assume ${\Lambda}= KQ'/I'$ and ${\Gamma}= KQ/I$. We will compare the quivers $Q$ and $Q'$. Let ${\operatorname{add}\nolimits}T$ denote the full subcategory of ${\operatorname{mod}\nolimits}H$ with objects the direct summands of directs sums of copies of $T$. We let $({\operatorname{add}\nolimits}T)_{{\operatorname{\mathcal C}\nolimits}}$ denote the full subcategory of ${\operatorname{\mathcal C}\nolimits}= {\operatorname{\mathcal C}\nolimits}_H$ with objects the direct summands of directs sums of copies of $T$. The vertices $Q'_0$ of $Q'$ correspond to the indecomposable objects in ${\operatorname{add}\nolimits}T$, while the vertices $Q_0$ of $Q$ correspond to the indecomposable objects in $({\operatorname{add}\nolimits}T)_{{\operatorname{\mathcal C}\nolimits}}$. It is clear that if we write $T$ as a direct sum of indecomposables $T = T_1 \amalg \dots \amalg T_n$ in ${\operatorname{mod}\nolimits}H$, this is a also a decomposition into indecomposables in ${\operatorname{\mathcal C}\nolimits}$. We therefore identify the sets $Q_0$ and $Q'_0$. Let $T_a$ and $T_b$ be indecomposable objects of ${\operatorname{add}\nolimits}T$. Using the definition of the cluster category ${\operatorname{\mathcal C}\nolimits}= {\operatorname{\mathcal C}\nolimits}_H$ and the fact that ${\operatorname{dim}\nolimits}{\operatorname{Hom}\nolimits}_{{\operatorname{\mathcal C}\nolimits}}(T_a,T_b) \leq 1$ by Theorem \[gb\] (b), we have that any non-zero map $T_a \to T_b$ in $({\operatorname{add}\nolimits}T)_{{\operatorname{\mathcal C}\nolimits}}$ is either the image of a map $T_a \to T_b$ in ${\operatorname{\mathcal D}\nolimits}$, or the image of a map $T_a \to FT_b$ in ${\operatorname{\mathcal D}\nolimits}$. Maps of the first kind are called [*$m$-maps*]{}, while maps of the second kind are called [*$f$-maps.*]{} Up to scalars, irreducible maps in $({\operatorname{add}\nolimits}T)_{{\operatorname{\mathcal C}\nolimits}}$ correspond to arrows in the quiver $Q$, while irreducible maps in ${\operatorname{add}\nolimits}T$ correspond to arrows in the quiver $Q'$. The arrows in $Q$ corresponding to irreducible $m$-maps, we call [*$m$-arrows*]{}, the other arrows are called [*$f$-arrows*]{}. For the set of arrows $Q_1$ in $Q$ we thus have a partition $Q_1 = Q_1^m \cup Q_1^f$, where $Q_1^m$ are the $m$-arrows and $Q_1^f$ are the $f$-arrows. A path in the quiver $Q$ is called an [*$f$-path*]{} if it contains at least one $f$-arrow. The other paths are called [*$m$-paths*]{}. \[preserved\] - [An irreducible map $T_a \to T_b$ in ${\operatorname{add}\nolimits}T$ is also irreducible as a map in $({\operatorname{add}\nolimits}T)_{{\operatorname{\mathcal C}\nolimits}}$.]{} - [Every irreducible map $T_a \to T_b$ in $({\operatorname{add}\nolimits}T)_{{\operatorname{\mathcal C}\nolimits}}$ which is the image of a map in ${\operatorname{add}\nolimits}T$ is also irreducible in ${\operatorname{add}\nolimits}T$.]{} (a): Trivial, since an $m$-map in ${\operatorname{\mathcal C}\nolimits}_H$ can not factor through an $f$-map. (b): Also trivial. This means that the set of arrows $Q'_1$ in $Q'$ corresponds to the set of $m$-arrows $Q_1^m$ in $Q$. We want to show that the $f$-arrows are determined by the minimal relations for ${\Lambda}$. We first compare the minimal relations in $I' \subset KQ'$ with the minimal relations in $I \subset KQ$. A minimal relation in $I' \subset KQ'$ is also a minimal relation in $I \subset KQ$. Given two indecomposable direct summands $T_a$ and $T_b$ in a tilting module $T$, we have ${\operatorname{dim}\nolimits}_k {\operatorname{Hom}\nolimits}_{{\operatorname{\mathcal C}\nolimits}}(T_a, T_b) \leq 1$ by Theorem \[gb\] (b). A consequence is that there cannot be both non-zero $m$-maps and non-zero $f$-maps from $T_a$ to $T_b$. Assume that $\rho'$ is a minimal relation in $I' \subset KQ'$. It is clear that $\rho'$ is also a relation in $I \subset KQ$. Since $I$ is generated by minimal commutativity-relations and minimal zero-relations, by Theorem \[gc\] (a), we can write $\rho' = \sum_{i=1}^m \beta_i \rho_i \gamma_i$ in $KQ$, with $\rho_i$ minimal commutativity-relations for $i= 1, \dots, r$ and $\rho_i$ minimal zero-relations for $i= r+1, \dots, m$. Write $\rho_i$ as a sum of scalar multiples of paths $\rho_i = \omega_i^{(0)} + \omega_i^{(1)}$, for $i= 1, \dots, r$. It follows from Theorem \[gb\] (b) that for $i= 1, \dots, r$ we have that $\omega_i^{(0)}$ is an $f$-path if and only if $ \omega_i^{(1)}$ is an $f$-path. Hence, the terms containing $f$-paths in the sum $\sum_{i=1}^m \beta_i \rho_i \gamma_i$ cancel, and we are left with an expression of $\rho'$ as a linear combination of relations in $I' \subset KQ'$. By the minimality assumption on $\rho'$ in $I'$, this means that for some $i$ we have that $\beta_i$ and $\gamma_i$ are scalars. By Proposition \[help\] (a), we have that $\rho'$ is also minimal in $I \subset KQ$. Using Theorem \[gc\] (c) we obtain as a consequence of the above that for any minimal relation $\rho$ in $I' \subset KQ'$, there is an $f$-arrow $\alpha_{\rho}$ in $Q$. It remains to show that actually all $f$-arrows arise this way. Recall that Theorem \[gc\] (a) says that the the set of minimal zero-relations and minimal commutativity relations in $I \subset KQ$ generate $I$. \[main\] Let $\rho_1, \dots , \rho_d$ be the minimal relations in $I'$ which are either zero-relations or commutativity relations. Then the quiver $Q$ is given by $Q_0 = Q'_0$ and $Q_1 = Q_1' \cup \{\alpha_{\rho_i} \}$. Let us fix some notation used in this proof. Recall that if there is an arrow $i \to j$, then a path from $j$ to $i$ is called shortest if the induced oriented cycle is full. A [*walk*]{} in a quiver is a sequence of arrows $\alpha_1, \dots, \alpha_t$ such that $\alpha_j$ connects the vertices $i_j$ and $i_{j+1}$ for $j= 1, \dots, t-1$. If all arrows involved are $m$-arrows it is called an $m$-walk. A path in a quiver is hence a directed walk, and an $m$-path is a directed $m$-walk. It remains only to show that for every $f$-arrow, say $\alpha \colon j \to i$, there is a minimal relation involving paths from $i$ to $j$ in $I' \subset KQ'$. We first show that there is an $m$-path from $i$ to $j$, which is a shortest path. We show this by induction. Note that a walk of length two between $j$ and $i$ is necessarily an oriented path from $i$ to $j$, since the quiver $Q$ is double path avoiding. We have the following observations. \[four\] Every full oriented cycle of length $n \geq 4$ contains exactly one $f$-arrow. Since $Q'$ does not contain oriented cycles, it is clear that there is always at least one $f$-arrow on every cycle. It is also clear that every path containing two (or more) $f$-arrows must be zero, since such a path corresponds to a map $T_a \to F^2T_b$ in ${\operatorname{\mathcal D}\nolimits}$. On the other hand the composition of $n-2$ or less arrows on a cycle of length $n$ is non-zero by Proposition \[help\] (d). From this it follows that any full oriented cycle of length $\geq 5$ contains exactly one $f$-arrow. It also follows that two consecutive $f$-arrows in a full oriented cycle of length 4 is impossible. To exclude a full oriented cycle $c_1 \circ c_2 \circ c_3 \circ c_4$ where $c_1$ and $c_3$ are $f$-arrows and $c_2$ and $c_4$ are $m$-arrows, note that in this case $c_1 \circ c_2$ is non-zero and $c_3 \circ c_4$ is non-zero, so we have a contradiction to Theorem \[gb\] (b). It will actually follow from the remaining part of the proof, that also full oriented cycles of length 3 have this property. \[commdia\] Assume $Q$ has a full subquiver of the form $$\xy \xymatrix{ & k \ar[dr] & \\ j \ar[dr] \ar[ur] & & i \ar[ll]_{\alpha} \\ & \cdot \ar[ur] & } \endxy$$\ where $\alpha$ is an $f$-arrow. Then the other arrows in this full subquiver are all $m$-arrows. By Theorem \[gb\] (b) and Proposition \[help\] (c), it follows that either both paths from $j$ to $i$ are $f$-paths, or both are $m$-paths. Assume both are $f$-paths. Mutating at $k$ one obtains a full subquiver which is a 4-cycle. $$\xy \xymatrix{ & k^{\ast} \ar[dl] & \\ j \ar[dr] & & i \ar[ul] \\ & \cdot \ar[ur] & } \endxy$$\ Assume ${\Gamma}= {\operatorname{End}\nolimits}_{{\operatorname{\mathcal C}\nolimits}}(T)^{{\operatorname{op}\nolimits}}$, for a tilting object $T$ in ${\operatorname{\mathcal C}\nolimits}_H$, and let $T_i, T_j$ be indecomposable direct summands in $T$ corresponding to the vertices $i$ and $j$, respectively. Since ${\operatorname{dim}\nolimits}{\operatorname{Hom}\nolimits}_{{\operatorname{\mathcal C}\nolimits}}(T_i, T_j) \leq 1$, by Theorem \[gb\] (b), and using Proposition \[help\] (c), it is clear that the arrow $\alpha$ and the composition $i \to k^{\ast} \to j$ represent the same map (up to scalars) in $({\operatorname{add}\nolimits}T)_{{\operatorname{\mathcal C}\nolimits}}$. Thus the composition $i \to k^{\ast} \to j$ contains an $f$-arrow, and we have a contradiction to Lemma \[four\]. Thus all arrows except $\alpha$ are $m$-arrows. \[cycles\] Assume there is an arrow $x \to y$ in $Q$, such that there is a walk $\delta$ between $y$ to $x$. Then there is a shortest path from $y$ to $x$ only passing through vertices on $\delta$. This follows from the fact that the quivers of cluster-tilted algebras of finite representation type are double-path avoiding, and the fact [@bmr3] that any full subquiver which is a non-oriented cycle, is not double path avoiding. The quiver $Q'$ is connected, since $H$ is connected. There is therefore always an $m$-walk between vertices of $Q$. It is a direct consequence of Lemma \[cycles\] that all $f$-arrows lie on oriented cycles. Assume now that there is an $f$-arrow $\alpha \colon i \to j$ such that there is no shortest $m$-path from $j$ to $i$. For each such $f$-arrow there is at least one $m$-walk between $i$ and $j$. Let $m(\alpha)$ be the length of the $m$-walk between $i$ and $j$ of minimal length. Amongst all $f$-arrows with the property that there is no $m$-path from $j$ to $i$, choose an $\alpha$ with smallest possible $m(\alpha)$. By Lemma \[cycles\] there is a shortest path from $j$ to $i$. By assumption this is not an $m$-path. Hence, by Lemma \[four\], it is of length two. Assume the path is composed by the arrows $\beta \colon j \to k$ and $\gamma \colon k \to i$. At least one of the arrows $\beta$ and $\gamma$ is an $f$-arrow. Consider also an $m$-walk $\delta$ between $i$ and $j$ of minimal length. Assume the walk is of length $t$, and consists of arrows $\delta_1, \dots, \delta_t$, where $\delta_1$ is an arrow either starting or ending in $j$ and $\delta_t$ is an arrow either starting or ending in $i$. $$\xy \xymatrix{ & \cdot \ar@{-}[r]^{\delta_2} & \cdot \ar@{--}[r] & \cdot \ar@{--}[r] & \cdot \ar@{-}[r] & \cdot \ar@{-}[dr]^{\delta_t} & \\ j \ar[drrr]_{\beta} \ar@{-}[ur]^{\delta_1} & & & & & & i \ar[llllll]^{\alpha} \\ & & & k \ar[urrr]_{\gamma} & & & } \endxy$$\ First assume that $\gamma$ is an $f$-arrow, and that $\beta$ is an $m$-arrow. Consider the case where $k$ is a vertex on the path $\delta$. Then by the assumption that $\delta$ is a walk of minimal length, we must have that $\beta = \delta_1 \colon j \to k$. The induced $m$-walk $\delta'$ from $k$ to $i$ given by the arrows $\delta_2, \dots, \delta_t$ obviously has length smaller than the length of $\delta$, so since $\gamma$ is an $f$-arrow, and using the mimimality assumption, there exists a directed path of $m$-arrows from $i$ to $k$. It is clear that $j$ is not a vertex on this path, so we have the following commutative diagram. $$\xy \xymatrix{ & \cdot \ar[r] & \cdot \ar@{--}[r] & \cdot \ar@{--}[r] & \cdot \ar[r] & \cdot \ar[dr] & \\ i \ar[drrr]_{\alpha} \ar[ur] & & & & & & k \ar[llllll]^{\gamma} \\ & & & j \ar[urrr]_{\beta} & & & } \endxy$$\ By the minimality assumptions on the walk $\delta$ it is clear that the two paths from $i$ to $k$ are such that the subquiver generated by these paths is as in Proposition \[help\] (c), and thus are non-zero and commute. By Theorem \[gb\] (b) this is a contradiction, since one path is an $m$-path and the other is an $f$-path. Now consider the case where $k$ is not on the walk $\delta$. Then, by Lemma \[cycles\], there is an alternative shortest path from $j$ to $i$ only passing through vertices on $\delta$. By assumption on $\alpha$ this path contains at least one $f$-arrow. Thus, it is of length two. We get the following quiver $$\xy \xymatrix{ & \cdot \ar[dr] & \\ i \ar[ur] \ar[dr] & & k \ar[ll] \\ & j \ar[ur] & } \endxy$$\ with two shortest paths from $i$ to $k$ containing $f$-arrows, and hence we have a contradiction to Lemma \[commdia\]. This finishes the case where $\gamma$ is an $f$-arrow, while $\beta$ is an $m$-arrow. The case where $\beta$ is an $f$-arrow, while $\gamma$ is an $m$-arrow can be excluded by similar arguments. Now assume that both $\beta$ and $\gamma$ are $f$-arrows. Consider a shortest path $\omega$ from $j$ to $i$ induced by the minimal $m$-walk, i.e. the path only passes through vertices on the chosen minimal $m$-walk. This path is necessarily of length two, by Lemma \[four\]. There are two cases. We first consider the case that $\omega$ is different from $\beta \circ \gamma$. In this case, the subquiver generated by these paths is as in Lemma \[commdia\], and hence we have a contradiction to this lemma. The second case is when these paths are equal. Then consider the $f$-arrow $\beta \colon j \to k$. The minimal $m$-walk between $i$ and $j$ induces a minimal $m$-walk $\delta''$ between $j$ and $k$. By minimality, it is clear that $i$ is not a vertex on $\delta''$. Thus the shortest path $\phi$ from $k$ to $j$ induced by $\delta''$ is disjoint from the path $\gamma \circ \alpha \colon k \to i \to j$. Since $\phi$ clearly contains an $f$-arrow, it must be of length two. So, we have two different paths of length two from $k$ to $j$, and since $Q$ is double path avoiding it is clear that the subquiver generated by these paths is $$\xy \xymatrix{ & \cdot \ar[dr] & \\ k \ar[ur] \ar[dr] & & j \ar[ll] \\ & i \ar[ur] & } \endxy$$ We thus have a contradiction by Lemma \[commdia\]. Let us now complete the proof of the theorem. We have established that for every $f$-arrow $i \to j$, there is at least one shortest path from $j$ to $i$ which is an $m$-path. By Theorem \[gc\] there are either exactly one or exactly two shortest paths from $j$ to $i$. From the same theorem it follows that in case there is exactly one, this is a minimal zero-relation. Furthermore, in case there is a second shortest path from $j$ to $i$, this path must represent a non-zero map, and hence is an $m$-path. Thus there is a minimal commutativity relation involving these two paths, by Propostion \[help\] (b). Hence, in both cases there is a minimal relation in $I \subset KQ$ involving $m$-paths from $j$ to $i$. It is clear that this is also a minimal relation in $I' \subset KQ'$. The following is obtained as a consequence of the last part of the proof of our main theorem, using that there are no oriented cycles in the quiver of a tilted algebra. Every full cycle in the quiver $Q$ of a cluster-tilted algebra of finite representation type contains exactly one $f$-map. As mentioned in the introduction, it is shown in [@bmr3] that a cluster-tilted algebra of finite representation type is determined by its quiver. This has the following consequence. Let $H$ be the path algebra of some Dynkin quiver over an algebraically closed field $K$. Given a tilted algebra ${\Lambda}= {\operatorname{End}\nolimits}_{H}(T)^{{\operatorname{op}\nolimits}} = KQ'/I'$ of type $H$, the corresponding cluster-tilted algebra ${\Gamma}= {\operatorname{End}\nolimits}_{{\operatorname{\mathcal C}\nolimits}_H}(T)^{{\operatorname{op}\nolimits}}$ is uniquely defined by $Q'$ and $I'$. Note that it is well known that a tilted algebra ${\Lambda}= {\operatorname{End}\nolimits}_{H}(T)^{{\operatorname{op}\nolimits}}$ can be of finite type, even though $H$ is not of finite type. It would be interesting to know in which generality the main theorem and the above corollary hold. We do not know any example where it does not hold. An example ========== The path algebra of the quiver $$\xy \xymatrix{ & \cdot \ar[dr]^{\beta} \ar[dl]_{\alpha} \ar@{.}@/^1.1pc/[drr] \ar@{.}[dd] & & \\ \cdot \ar[dr]_{\gamma} & & \cdot \ar[dl]^{\delta} \ar[r]^{\epsilon} & \cdot \\ & \cdot & & } \endxy$$\ with relations $\alpha \gamma = \beta \delta$ and $\beta \epsilon = 0$ as indicated, is a tilted algebra of type $D_5$. The corresponding cluster-tilted algebra is the path algebra of the quiver $$\xy \xymatrix{ & \cdot \ar[dr]^{\beta} \ar[dl]_{\alpha} & & \\ \cdot \ar[dr]_{\gamma} & & \cdot \ar[dl]^{\delta} \ar[r]^{\epsilon} & \cdot \ar@/_1.1pc/[ull]_{\pi} \\ & \cdot \ar[uu]^{\mu}& & } \endxy$$\ with relations given by the commutativity relation $\alpha \gamma = \beta \delta$ and all other compositions of two arrows equal to zero. [99]{} Buan A., Marsh R., Reiten I. *Cluster-tilted algebras*, preprint (2004) math.RT/0402075, to appear in Trans. Amer. Math. Soc. Buan A., Marsh R., Reiten I. *Cluster mutation via quiver representations*, preprint (2004) math.RT/0412077 Buan A., Marsh R., Reiten I. *Cluster-tilted algebras of finite representation type*, preprint (2005) math.RT/0509198 Buan A., Marsh R., Reineke M., Reiten I., Todorov G. *Tilting theory and cluster combinatorics*, preprint math.RT/0402054, to appear in Adv. Math. (2004) Caldero P., Chapoton F., Schiffler R. *Quivers with relations arising from clusters ($A_n$ case)*, Trans. Amer. Math. Soc., to appear Fomin S., Zelevinsky A. *Cluster Algebras I: Foundations*, J. Amer. Math. Soc. 15, no. 2, 497–529 (2002) Happel D. *Triangulated categories in the representation theory of finite-dimensional algebras*, London Mathematical Society Lecture Note Series, 119. Cambridge University Press, Cambridge (1988) Keller B. *On triangulated orbit categories*, preprint (2005), math.RT/0503240 Marsh R., Reineke M., Zelevinsky A. *Generalized associahedra via quiver representations* Trans. Amer. Math. Soc. 355, no. 1, 4171–4186 (2003)

--- abstract: 'This technical note describes a monotone and continuous fixpoint operator to compute the answer sets of programs with aggregates. The fixpoint operator relies on the notion of [*aggregate solution*]{}. Under certain conditions, this operator behaves identically to the three-valued immediate consequence operator $\Phi^{aggr}_P$ for aggregate programs, independently proposed in [@Pelov04; @PelovDB04]. This operator allows us to closely tie the computational complexity of the answer set checking and answer sets existence problems to the cost of checking a solution of the aggregates in the program. Finally, we relate the semantics described by the operator to other proposals for logic programming with aggregates.' author: - | TRAN CAO SON and ENRICO PONTELLI\ Department of Computer Science\ New Mexico State University\ {tson,epontell}@cs.nmsu.edu title: A Constructive Semantic Characterization of Aggregates in Answer Set Programming --- Aggregates, answer set programming, semantics Introduction ============ Several semantic characterizations of answer sets of logic programs with aggregates have been proposed over the years (e.g., [@KempS91; @MumickPR90; @a-prolog; @faberLP04; @PelovDB04]). Most of these proposals have their roots in the answer set semantics of normal logic programs without aggregates [@GL88]. Nevertheless, it is known that a straightforward generalization of the definition of answer sets to programs with aggregates may yield [*non-minimal*]{} and/or unintuitive answer sets. Consider the following example. \[exp2\] Let $P$ be the program $$\begin{array}{lcl} p(1) \leftarrow & \hspace{1cm} & p(2) \leftarrow \hspace{4cm} p(3) \leftarrow \\ p(5) \leftarrow q && q \leftarrow \textnormal{\sc Sum}(\{X \mid p(X)\}) > 10 \end{array}$$ The aggregate $ \textnormal{\sc Sum}(\{X \mid p(X)\}) > 10$ is satisfied by any interpretation $M$ of $P$ where the sum of $X$ such that $p(X)$ is true in $M$ is greater than 10. A straightforward extension of the original definition of answer sets [@GL88] defines $M$ to be an answer set of $P$ if and only if $M$ is the minimal model of the reduct $P^M$, where $P^M$ is the program obtained by [*(i)*]{} removing from $P$ all the rules containing in their body at least an aggregate or a negation-as-failure literal which is false in $M$; and [*(ii)*]{} removing all the aggregates and negation-as-failure literals from the remaining rules. Effectively, this definition treats aggregates in the same fashion as negation-as-failure literals. It is easy to see that for $A = \{p(1),p(2),p(3)\}$ and $B = \{p(1),p(2),p(3),p(5),q\}$, $$\begin{array}{lll} P^A = \left\{ \begin{array}{l} p(1) \leftarrow \\ p(2) \leftarrow \\ p(3) \leftarrow \\ p(5) \leftarrow q \\ \end{array} \right \} & \hspace*{0.5cm} & P^B = \left \{ \begin{array}{l} p(1) \leftarrow \\ p(2) \leftarrow \\ p(3) \leftarrow \\ p(5) \leftarrow q \\ q \leftarrow \\ \end{array} \right \} \end{array}$$ and $A$ and $B$ are minimal model of $P^A$ and $P^B$ respectively. Thus, both $A$ and $B$ are answer sets of $P$. As we can see, treating aggregates like negation-as-failure literals yields non-minimal answer sets. Accepting $B$ as an answer set seems counter-intuitive, since $p(5)$ “supports” itself through the aggregate. Different approaches have been proposed to deal with this problem. Early works concentrate on finding syntactic (e.g., stratification [@MumickPR90; @KempS91]) and semantic (e.g., monotonic aggregates [@RossS97; @KempS91]) restrictions on aggregates which guarantee minimality, and often uniqueness, of answer sets. In this technical note, we present a fixpoint operator that allows us to compute answer sets of normal logic programs with [*arbitrary aggregates*]{}. It is a straightforward extension of the Gelfond-Lifschitz definition, making use of the *same* notion of reduct as in [@GL88], and relying on a continuous fixpoint operator for computing selected minimal models of the reduct (corresponding to our notion of answer sets). This fixpoint operator is a natural extension of the traditional immediate consequence operator $T_P$ to programs with aggregates. It takes into consideration the provisional answer set while trying to verify that it is an answer set. This fixpoint operator makes use of the notion of *aggregate solutions*, and it captures the *unfolding semantics* for normal logic programs with aggregates, originally proposed in [@ElkabaniPS04] and completely developed in [@SonPE05]. This semantics builds on the principle of unfolding of intensional set constructions, as developed in [@DovierPR01]. This operator corresponds to the $\Phi_P^{aggr}$ operator proposed in [@PelovDB04; @Pelov04], when ultimate approximating aggregates are employed and 2-valued stable models are considered. In particular, the two operators are identical when they are applied to the construction of a correct answer set $M$. The proposed fixpoint operator allows us also to easily demonstrate the existence of a large class of logic programs with aggregates (which includes recursively defined aggregates and non-monotone aggregates) for which the problems of answer set checking and of determining the existence of an answer set is in [**P**]{} and [**NP**]{} respectively. Finally, we relate our work to recently proposed semantics for programs with aggregates [@faberLP04; @PelovDB04; @SonPE05]. Preliminary Definitions {#sec2} ======================= Language Syntax --------------- Let us consider a signature $\Sigma_L = \langle {\cal F}_L\cup {\cal F}_{Agg}, {\cal V}\cup {\cal V}_l, \Pi_L \rangle$, where ${\cal F}_L$ is a collection of constants, ${\cal F}_{Agg}$ is a collection of unary function symbols, ${\cal V} \cup {\cal V}_l$ is a denumerable collection of variables (such that ${\cal V} \cap {\cal V}_l = \emptyset$), and $\Pi_L$ is a collection of predicate symbols. In the rest of this paper, we will always assume that the set $\mathbb{Z}$ of the integers is a subset of ${\cal F}_L$—i.e., there are distinct constants representing the integer numbers. We will refer to $\Sigma_L$ as the *ASP signature*. We will also refer to $\Sigma_P = \langle {\cal F}_P,{\cal V}\cup {\cal V}_l, \Pi_P\rangle$ as the *program signature*, where ${\cal F}_P \subseteq {\cal F}_L$, $\Pi_P \subseteq \Pi_L$, and ${\cal F}_P$ is finite. We will denote with ${\cal H}_P$ the $\Sigma_P$-Herbrand universe, containing the ground terms built using symbols of ${\cal F}_P$, and with ${\cal B}_P$ the corresponding $\Sigma_P$-Herbrand base. An ASP-atom is an atom of the form $p(t_1,\dots,t_n)$, where $t_i \in {\cal F}_P \cup {\cal V}$ and $p\in \Pi_P$; an ASP-literal is either an ASP-atom or the negation as failure ($not\:A$) of an ASP-atom. We will use the traditional notation $\{t_1,\dots,t_k\}$ to denote an extensional set of terms, and the notation $\{\!\!\{t_1,\dots,t_k\}\!\!\}$ to denote an extensional multiset (or bag) of terms. \[intensionalset\] An *intensional set* is a set of the form $ \{X \:\mid\: p(X_1,\dots,X_k)\} $ where $X\in {\cal V}_l$, $X_i$’s are variables or constants (in ${\cal F}_P$), $\{X_1,\dots,X_k\} \cap {\cal V}_l = \{X\}$, and $p$ is a $k$-ary predicate in $\Pi_P$. Similarly, an *intensional multiset* is a multiset of the form $$\{\!\!\{ X \:\mid\: \exists Z_1,\dots,Z_r {\textit{.}}\: p(Y_1,\dots,Y_m)\}\!\!\}$$ where $\{X,Z_1,\dots,Z_r\}\subseteq {\cal V}_l$, $Y_i$ are variables or constants (of ${\cal F}_P$), $\{Y_1,\ldots,Y_m\} \cap {\cal V}_l= \{X,Z_1,\dots,Z_r\}$, and $X \notin \{Z_1,\dots,Z_r\}$. We call $X$ the *grouped variable*, $Z_1,\dots,Z_r$ the *local variables*, and $p$ the *grouped predicate* of the intensional set/multiset. Intuitively, in an intensional multiset, we collect the values of $X$ for which $p(Y_1,\ldots,Y_m)$ is true, under the assumptions that the variables $Z_1,\dots,Z_r$ are locally, existentially quantified. Multiple occurrences of the same value of $X$ can appear. For example, if $p(X,Z)$ is true for $X=1, Z=2$ and $X=1, Z=3$, then the multiset $\{\!\{X\mid\exists Z. p(X,Z)\}\!\}$ will correspond to $\{\!\!\{1,1\}\!\!\}$. Definition \[intensionalset\] can be easily extended to allow more complex types of sets, e.g., sets with a tuple as the grouped variable and sets with conjunctions of atoms as property of the intensional construction. Observe that the variables from ${\cal V}_l$ are used exclusively as grouped or local variables in defining intensional sets/multisets, and they cannot occur anywhere else. We write $\bar{X}$ to denote $X_1,\ldots,X_n$. - An *aggregate term* is of the form $aggr(s)$, where $s$ is an intensional set/multiset, and $aggr \in {\cal F}_{Agg}$ (called the *aggregate function*). - An *aggregate atom* has the form $aggr(s) \:\: \texttt{op} \:\: Result$, where [op]{} is a relational operator in the set $\{=, \neq, <, >, \leq, \geq\}$ and $Result \in {\cal V} \cup (\mathbb{Z} \cap {\cal F}_P)$—i.e., it is either a variable or a numeric constant. In our examples, we will focus on the traditional aggregate functions, e.g., [Count, Sum, Min]{}. For an aggregate atom $\ell$ of the form $aggr(s) \:\: \texttt{op} \:\: Result$, we refer to the grouped variable and predicate of $s$ as the grouped variable and predicate of $\ell$. The set of ASP-atoms constructed from the grouped predicate of $\ell$ and the terms in ${\cal H}_P$ is denoted by ${\cal H}(\ell)$. \[[${ASP}^A$]{} Rule/Program\] - An [${ASP}^A$]{} rule is of the form $$\label{agg-rule} A \leftarrow C_1,\ldots,C_m,A_1,\ldots,A_n,\naf B_1, \dots, \naf B_k$$ where $A$, $A_1$, $\dots$, $A_n$, $B_1$, $\dots$, $B_k$ are ASP-atoms, while $C_1, \dots, C_m$ are aggregate atoms ($m\geq 0$, $n\geq 0$, $k\geq 0$). - An [${ASP}^A$]{} program is a finite collection of [${ASP}^A$]{} rules. For an [${ASP}^A$]{} rule $r$ of the form (\[agg-rule\]), $head(r)$, $agg(r)$, $pos(r)$, and $neg(r)$ denote respectively $A$, $\{C_1, \ldots, C_m\}$, $\{A_1, \ldots, A_n\}$, and $\{B_1, \ldots, B_k\}$. Furthermore, $body(r)$ denotes the right-hand side of the rule $r$. Observe that grouped and local variables in an aggregate atom $\ell$ have a scope limited to $\ell$. As such, given an [${ASP}^A$]{} rule, it is always possible to rename such variables occurring in the aggregate atoms $C_1,\ldots,C_m$ apart, so that they are pairwise different. Observe also that the grouped and local variables represent the only occurrences of variables from ${\cal V}_l$, thus they will not occur in $A$, $A_1$, $\dots$, $A_n$, $B_1$, $\ldots$, $B_k$. For this reason, without loss of generality, whenever we refer to an [${ASP}^A$]{} rule $r$, we will assume that the grouped and local variables of its aggregate atoms are pairwise different and do not appear in the rest of the rule. Given a term, literal, aggregate atom, rule $\alpha$, let us denote with $fvars(\alpha)$ the set of variables from ${\cal V}$ present in $\alpha$. The entity $\alpha$ is ground if $fvars(\alpha) = \emptyset$. A ground substitution $\sigma$ is a set $\{X_1/c_1,\ldots,X_n/c_n\}$ where $X_i$’s are distinct variables from ${\cal V}$ and $c_i$’s are constants in ${\cal F}_P$. For an ASP-atom $p$ (an aggregate atom $\ell$), $p \sigma$ ($\ell \sigma$) denotes the ASP-atom (the aggregate atom) which is obtained from $p$ ($\ell$) by simultaneously replacing every occurrence of $X_i$ with $c_i$. Let $r$ be a rule of the form (\[agg-rule\]) and $\{X_1,\ldots,X_t\}$ be the set of free variables occurring in $A$, $C_1,\ldots,C_m$, $A_1,\ldots,A_n$, and $B_1,\ldots,B_k$—i.e., $fvars(r) = \{X_1,\dots,X_t\}$. Let $\sigma$ be a ground substitution $\{X_1/c_1,\ldots,X_t/c_t\}$. The ground instance of $r$ w.r.t. $\sigma$, denoted by $r\sigma$, is the ground rule obtained from $r$ by simultaneously replacing every occurrence of $X_i$ with $c_i$. By $ground(r)$ we denote the set of all ground instances of the rule $r$. For a program $P$, the set of all ground instances of the rules in $P$, denoted by $ground(P)$, is called the ground instance of $P$, i.e., $ground(P) = \bigcup_{r \in P} ground(r)$. Aggregate Solutions ------------------- In this subsection we provide the basic definitions of satisfaction and solution of an aggregate atom. The domain of our interpretations is the set ${\cal D} = {\cal H}_P \cup 2^{{\cal H}_P} \cup {\cal M}({\cal H}_P)$, where $2^{{\cal H}_P}$ is the set of (finite) subsets of ${\cal H}_P$ and ${\cal M}({\cal H}_P)$ is the set of finite multisets of elements from ${\cal H}_P$. An interpretation $I$ is a pair $\langle {\cal D}, (\cdot)^I\rangle$, where $(\cdot)^I$ is a function that maps ground terms to elements of $\cal D$ and ground atoms to truth values. Given a constant $c$, its interpretation $c^I$ is equal to $c$. Given a ground intensional set $s$ of the form $\{X \mid p(\bar{X})\}$, its interpretation $s^I$ is the set $\{a_1,\dots,a_n\} \subseteq {\cal H}_P$, where $(p(\bar{X}))\{X / a_i\}^I$ is equal to true for $1\leq i\leq n$, and no other value for $X$ has such property. Given a ground intensional multiset $s$ of the form $\{\!\!\{X\:\mid\: \exists \bar{Z}{\textit{.}}p(\bar{X},\bar{Z})\}\!\!\}$, its interpretation $s^I$ is the multiset $\{\!\!\{a_1,\dots,a_k\}\!\!\} \in {\cal M}({\cal H}_P)$ where, for each $1\leq i\leq k$, there is a ground substitution $\eta_i$ for $\bar{Z}$ such that $p(\bar{X},\bar{Z})\gamma_i^I$ is true for $\gamma_i = \eta_i \cup \{X/a_i\}$, and no other elements satisfy this property. Given the aggregate term $aggr(s)$, its interpretation is $aggr^I(s^I)$, where $$aggr^I: 2^{ {\cal H}_P }\cup {\cal M}({ {\cal H}_P }) \rightarrow \mathbb{Z}.$$ Given a ground [${ASP}^A$]{} atom $p(t_1,\dots,t_n)$, its interpretation is $p^I(t_1^I,\dots,t_n^I)$, where $p^I: {\cal D}^n \rightarrow \{\texttt{true},\texttt{false}\}$. Given a ground aggregate atom $\ell$ of the form $aggr(s) \; {\tt op} \; Result$, its interpretation $\ell^I$ is true if $op^I(aggr(s)^I,Result^I)$ is true, where ${\tt op}^I: \mathbb{Z} \times \mathbb{Z} \rightarrow \{\texttt{true},\texttt{false}\}$. We will assume that the traditional aggregate functions are interpreted in the usual way. E.g., [Sum]{}$^I$ is the function that maps a set/multiset of numbers to its sum, and [Count]{}$^I$ is the function that maps a set/multiset of constants to its cardinality. Similarly, we assume that the traditional relational operators (e.g., $\leq$, $\neq$) are interpreted according to their traditional meaning. Given a literal $not\:p$, its interpretation $(not\:p)^I$ is true (false) iff $p^I$ is false (true). Given an atom, literal, or aggregate atom $\ell$, we will denote with $I\models \ell$ the fact that $\ell^I$ is true. $I$ [*satisfies the body of a ground rule*]{} $r$ (denoted by $I \models body(r)$), if [ ]{} $pos(r) \subseteq I$; $neg(r) \cap I = \emptyset$; $I \models c$ for every $c \in agg(r)$. $I$ [*satisfies a ground rule*]{} $r$ if $I \models head(r)$ or $I \not\models body(r)$. Having specified when an interpretation satisfies an aggregate atom or a [${ASP}^A$]{} rule, we can define the notion of model of a program. Let $P$ be an [${ASP}^A$]{} program. An interpretation $M$ is a [*model*]{} of $P$ if $M$ satisfies every rule in $ground(P)$. In our view of interpretations, we assume that the interpretation of the aggregate functions and relational operators is fixed. In this perspective, we will still be able to keep the traditional view of interpretations as subsets of ${\cal B}_P$. $M$ is a *minimal model* of $P$ if $M$ is a model of $P$ and there is no proper subset of $M$ which is also a model of $P$. We will now define a notion called [*aggregate solution*]{}. Observe that the satisfaction of an ASP-atom $a$ is *monotonic*, in the sense that if $I \models a$ and $I \subseteq I'$ then we have that $I' \models a$. On the other hand, the satisfaction of an aggregate atom is possibly non-monotonic, i.e., $I \models \ell$ and $I \subseteq I'$ do not necessarily imply $I' \models \ell$. For example, $\{p(1)\} \models \textnormal{\sc Sum}(\{X \mid p(X)\}) \ne 0$ but $\{p(1),p(-1)\} \not\models \textnormal{\sc Sum}(\{X \mid p(X)\}) \ne 0$. The notion of aggregate solution allows us to define an operator where the monotonicity of satisfaction of aggregate atoms is used in verifying an answer set. Let $\ell$ be a ground aggregate atom. An [*aggregate solution*]{} of $\ell$ is a pair $\langle S_1, S_2 \rangle$ of disjoint subsets of ${\cal H}(\ell)$ such that, for every interpretation $I$, if $S_1 \subseteq I$ and $S_2 \cap I = \emptyset$ then $I \models \ell$. ${\cal SOLN}(\ell)$ is the set of all the solutions of $\ell$. It is obvious that if $I \models \ell$ then $\langle I \cap {\cal H}(\ell), {\cal H}(\ell) \setminus I \rangle$ is a solution of $\ell$. Let $S = \langle S_1, S_2 \rangle$ be an aggregate solution of an aggregate atom; we denote with $S{\textit{.}}p$ and $S{\textit{.}}n$ the two components $S_1$ and $S_2$ of the solution. \[ex-sol\] Consider the aggregate atom ${\textnormal{\sc Sum}(\{X \mid p(X)\}) > 10}$ from the program in Example \[exp2\]. This atom has a unique solution: $\langle \{p(1), p(2), p(3), p(5)\}, \emptyset\rangle$. On the other hand, the aggregate atom ${\textnormal{\sc Sum}(\{X \mid p(X)\}) > 6}$ has the following solutions: $$\begin{array}{lrclr} \langle \{p(3), p(5)\}, &\emptyset\rangle & \hspace{.8cm} & \langle \{p(3), p(5)\}, &\{p(1),p(2)\}\rangle \\ \langle \{p(3), p(5)\}, &\{p(1)\}\rangle & & \langle \{p(3), p(5)\}, &\{p(2)\}\rangle \\ \langle \{p(2), p(5)\}, &\emptyset\rangle & & \langle \{p(2), p(5)\}, &\{p(1),p(3)\}\rangle \\ \langle \{p(2), p(5)\}, &\{p(1)\}\rangle & & \langle \{p(2), p(5)\}, &\{p(3)\}\rangle \\ \langle \{p(1), p(2), p(5)\}, &\emptyset\rangle & & \langle \{p(1), p(2), p(5)\}, &\{p(3)\}\rangle \\ \langle \{p(1), p(3), p(5)\}, &\emptyset\rangle & & \langle \{p(1), p(3), p(5)\}, &\{p(2)\}\rangle \\ \langle \{p(1), p(2), p(3), p(5)\}, &\emptyset\rangle & & \langle \{p(2), p(3), p(5)\}, &\emptyset\rangle \\ \langle \{p(2), p(3), p(5)\}, &\{p(1)\}\rangle \\ \end{array}$$ A Fixpoint Operator based on Aggregate Solutions {#semantics} ================================================ In this section, we construct the semantics for [${ASP}^A$]{} programs, through the use of a monotone and continuous fixpoint operator. For the sake of simplicity, we will assume that programs, ASP-atoms, and aggregate atoms referred to in this section are ground[^1]. As we will show in Section \[pelov\], this fixpoint operator behaves as the 3-valued immediate consequence operator of [@PelovDB04] under certain conditions (e.g., use of ultimate approximating aggregates). Let $P$ be an [${ASP}^A$]{} program and let $M$ be an interpretation. The reduct of $P$ with respect to $M$, denoted by $^M\!P$, is defined as $$^M\!P=\left\{ head(r) \leftarrow pos(r),agg(r) \mid r \in ground(P), \;\; M \cap neg(r) = \emptyset \right\}$$ Observe that, for a program $P$ without aggregates, the process of checking whether $M$ is an answer set [@GL88] requires first computing the Gelfond-Lifschitz reduct of $P$ w.r.t. $M$ ($P^M$), and then verifying that $M$ is the least model of $P^M$. This second step is performed by using the van Emden-Kowalski operator $T_{P^M}$ to regenerate $M$, by computing the least fixpoint of $T_{P^M}$. I.e., we compute the sequence $M_0, M_1, M_2, \dots$ where $M_0 = \emptyset$ and $M_{i+1} = T_{P_M}(M_i)$. In every step of regenerating $M$, an atom $a$ is added to $M_{i+1}$ iff there is a rule in $P^M$ whose head is $a$ and whose body is contained in $M_i$. This process is monotonic, in the sense that, if $a$ is added to $M_i$, then $a$ will belong to $M_j$ for all $j \ge i$. Our intention is to define a $T_P$-like operator for programs with aggregates. Specifically, we would like to verify that $M$ is an answer set of $P$ by generating a monotone sequence of interpretations $M_0 \subseteq M_1 \subseteq \ldots \subseteq M_n \subseteq \ldots = M$. To do so, we need to specify when a rule of $^M\!P$ can be used, i.e., when an ASP/aggregate atom is considered satisfied by $M_i$. We also need to ensure that, at each step $i+1$, $M_{i+1}$ will [*still*]{} satisfy all ASP-atoms and the aggregate atoms that are satisfied by $M_{i}$. This observation leads us to define the notion of *conditional satisfaction* of an atom (ASP-atom or aggregate atom) over a pair of sets of atoms $(I,M)$—where $I$ is an interpretation generated at some step of the verification process, and $M$ is the answer set that needs to be verified. \[cond-sat\] Let $\ell$ be an ASP-atom or an aggregate atom, and $I$, $M$ be two interpretations[^2]. We define the *conditional satisfaction* of $\ell$ w.r.t. $I$ and $M$, denoted by $(I,M) \models \ell$, as: [$\bullet$]{}[ ]{} if $\ell$ is ASP-atom, then $(I,M) \models \ell \:\:\Leftrightarrow\:\:I\models \ell$ if $\ell$ is an aggregate atom, then $$(I,M) \models l \:\:\Leftrightarrow\:\: \langle I\cap M\cap {\cal H}(\ell),\: {\cal H}(\ell)\setminus M\rangle \textit{ is a solution of $\ell$}$$ The first bullet says that an ASP-atom is satisfied by a pair $(I,M)$ if it is satisfied by $I$. The second bullet states that $I$ contains enough information of $M$ to guarantee that any successive expansion of $I$ towards $M$ will satisfy the aggregate. Conditional satisfaction is naturally extended to conjunctions of atoms. The following lemma trivially holds. \[l0\] Let $\ell$ be an ASP-atom or an aggregate atom and $I, J, M$ be interpretations such that $I \subseteq J$. Then, $(I,M) \models \ell$ implies $(J,M) \models \ell$. We are now ready to define the consequence operator for [${ASP}^A$]{}programs. Let $P$ be an [${ASP}^A$]{} program and $M$ be an interpretation. We define the consequence operator on $P$ and $M$, called $K_M^P$, as $$K_M^P(I) = \{\:head(r) \:\:\mid\:\: r\in {^M\!P}\:\wedge\: (I,M) \models body(r)\:\}$$ for every interpretation $I$ of $P$. By definition, we have that $K_M^P(I) = T_P(I)$ for definite programs without aggregate atoms. Thus, $K_M^P$ can be viewed as an extension of $T_P$ to the class of programs with aggregates. The following lemma is a consequence of Lemma \[l0\]. \[l4\] Let $P$ be a program and $M$ be an interpretation. Then, $K_M^P$ is monotone and continuous over the lattice $\langle 2^{{\cal B}_P}, \subseteq \rangle$. The above lemma allows us to conclude that the least fixpoint of $K_M^P$, denoted by $lfp(K_M^P)$, exists and it is equal to $K_M^P \uparrow \omega$. Here, $K^P_M \uparrow n$ denotes $$\underbrace{K^P_M(K^P_M(\dots(K^P_M}_{n-times \; K^P_M}(\emptyset)\dots)))$$ and $K^P_M \uparrow \omega $ denotes $\lim_{n \rightarrow \infty} K^P_M \uparrow n$. We are now ready to define the concept of *answer set* of an [${ASP}^A$]{} program. Let $P$ be an [${ASP}^A$]{} program and let $M$ be an interpretation. $M$ is a *fixpoint answer set* of $P$ iff $M=lfp(K_M^P)$. Whenever it is clear from the context, we will simply talk about *answer sets* of $P$ instead of fixpoint answer sets. \[exp4\] Let us continue with the program $P$ from Example \[exp2\]. Since $P$ does not contain negation-as-failure literals, $^M\!P = P$ for any interpretation $M$ of $P$. Any answer set of $P$ must contain $p(1)$, $p(2)$, and $p(3)$. We will now show that $A = \{p(1), p(2), p(3)\}$ is the unique fixpoint answer set of $P$. It is easy to see that $$\begin{array}{lcl} K_A^P \uparrow 0 & = & \emptyset\\ K_A^P \uparrow 1 & = & K_A^P (K_A^P \uparrow 0) = \{p(1), p(2), p(3)\} \\ K_A^P \uparrow 2 & = & \{p(1), p(2), p(3)\} = K_A^P \uparrow 1 \end{array}$$ Thus, $A$ is indeed a fixpoint answer set of $P$. Let us consider $B = \{p(1), p(2), p(3),p(5), q\}$. We have that $^B\!P = P$ and it is easy to verify that $lfp(K_B^P) = \{p(1), p(2), p(3)\}$. Therefore, $B$ is not a fixpoint answer set of $P$. It is easy to check that no proper superset of $A$ is a fixpoint answer set of $P$, i.e., $A$ is the unique answer set of $P$. In the next example, we show how this definition works when the programs contain negation-as-failure literals. Let $P$ be the program[^3]: $$\begin{array}{lll} p(a) & \leftarrow & \textnormal{\sc Count}(\{X \mid p(X)\}) > 0 \\ p(b) & \leftarrow & \naf q \\ q & \leftarrow & \naf p(b) \\ \end{array}$$ We will show now that the program has two answer sets $A = \{q\}$ and $B = \{p(b),p(a)\}$. We have that - $^A\!P$ consists of the first rule and the fact $q$. The verification that $A$ is an answer set of $P$ is shown next. $$\begin{array}{lcl} K_A^P \uparrow 0 & = & \emptyset\\ K_A^P \uparrow 1 & = & K_A^P (K_A^P \uparrow 0) = \{q\}\\ K_A^P \uparrow 2 & = & \{q\} = K_A^P\uparrow 1 \end{array}$$ $p(a)$ cannot belong to $K_A^P \uparrow 1$ since $\langle \emptyset, \emptyset \rangle$ is not a solution of the aggregate atom $\textnormal{\sc Count}(\{X \mid p(X)\}) > 0$. - $^B\!P$ consists of the first rule and the fact $p(b)$. $$\begin{array}{lcl} K_B^P \uparrow 0 & = & \emptyset\\ K_B^P \uparrow 1 & = & K_B^P (K_B^P \uparrow 0) = \{p(b)\}\\ K_B^P \uparrow 2 & = & \{p(b),p(a)\} \\ K_B^P \uparrow 3 & = & \{p(b),p(a)\} = K_B^P \uparrow 2 \end{array}$$ $p(a)$ belongs to $K_B^P \uparrow 2$ since $\langle \{p(b)\}, \emptyset \rangle$ is a solution of the aggregate atom $\textnormal{\sc Count}(\{X \mid p(X)\}) > 0$. It is easy to see that $P$ does not have any other answer sets. Related Work and Discussion {#discuss} =========================== In this section, we will relate our proposal to the unfolding semantics presented in [@SonPE05] and to two other recently proposed semantics for programs with aggregates[^4]—i.e., the ultimate stable model semantics [@PelovDB03; @PelovDB04; @Pelov04] and the minimal answer set semantics [@faberLP04]. We will also investigate some of the computational complexity issues related to determining the fixpoint answer sets of [${ASP}^A$]{} programs. Equivalence of Fixpoint Semantics and Unfolding Semantics --------------------------------------------------------- We will show that the notion of fixpoint answer set corresponds to the [*unfolding semantics*]{} presented in [@SonPE05]. To make this note self-contained, let us recall the basic definition of the unfolding semantics. For a ground aggregate atom $c$ and an interpretation $M$, let $${\cal S}(c, M) = \left\{S_c \:\mid\: S_c \in {\cal SOLN}(c), \:S_c{\textit{.}}p \subseteq M,\: S_c {\textit{.}}n \cap M = \emptyset \right\}$$ Intuitively, ${\cal S}(c, M)$ is the set of solutions of $c$ which are satisfied by $M$. For a solution $S_c \in {\cal S}(c,M)$, the unfolding of $c$ in $M$ w.r.t. $S_c$ is the conjunction $\bigwedge_{a \in S_c{\textit{.}}p} a$. We say that $c'$ is an unfolding of $c$ with respect to $M$ if $c'$ is an unfolding of $c$ in $M$ with respect to some $S_c \in {\cal S}(c, M)$. When ${\cal S}(c, M) = \emptyset$, we say that $false$ is the unfolding of $c$ in $M$. The unfolding of a rule $r \in ground(P)$ with respect to $M$ is the set of rules $unfolding(r,M)$ defined as follows: 1. If $neg(r) \cap M \ne \emptyset$ or there is some $c \in agg(r)$ such that $false$ is the unfolding of $c$ in $M$ then $unfolding(r,M) = \emptyset$; 2. If $neg(r) \cap M = \emptyset$ and $false$ is not the unfolding of $c$ for every $c \in agg(r)$, then $r' \in unfolding(r,M)$ where 1. $head(r') = head(r)$ 2. $neg(r') = neg(r)$ 3. there is a sequence of aggregate solutions $\langle S_c \rangle_{c \in agg(r)}$ for the aggregates in $agg(r)$, such that $S_c \in {\cal S}(c,M)$ for every $c \in agg(r)$ and $pos(r') = pos(r) \cup \bigcup_{c \in agg(r)} S_c{\textit{.}}p$. For a program $P$, $unfolding(P,M)$ denotes the set of unfolding rules of $ground(P)$ w.r.t. $M$. $M$ is an [${ASP}^A$]{}-answer set of $P$ iff $M$ is an answer set of $unfolding(P, M)$. This notion of unfolding derives from the work on unfolding of intensional sets [@DovierPR01], and has been independently described in [@PelovDB03]. \[l1\] Let $c$ be an aggregate atom, let $M$ be an interpretation, and let $S_c$ be a solution of $c$ such that $S_c \in {\cal S}(c,M)$. Then, $\langle S_c{\textit{.}}p, {\cal H}(c)\setminus M\rangle$ is a solution of $c$. Let us consider an interpretation $I$ such that $S_c{\textit{.}}p \subseteq I$ and $I \cap ({\cal H}(c)\setminus M) = \emptyset$. Because $S_c{\textit{.}}n \subseteq {\cal H}(c) \setminus M$, $I \cap S_c{\textit{.}}n = \emptyset$. Since $S_c$ is a solution, $I\models c$. Since this holds for every interpretation $I$ satisfying $S_c{\textit{.}}p\subseteq I$ and $I \cap ({\cal H}(c)\setminus M) = \emptyset$, we have that $\langle S_c{\textit{.}}p, {\cal H}(c)\setminus M\rangle$ is a solution of $c$. \[l2\] Let $R=unfolding(P,M)$. Then $T_R\uparrow i = K_M^P\uparrow i$ for $i\geq 0$. Let us prove the result by induction on $i$. [*Base:*]{} for $i=0$, we have that $T_R \uparrow 0 = \emptyset = K_M^P\uparrow 0$, and the result is obviously true. Let us consider the case $i=1$. Let $p \in T_R\uparrow 1 = \{ \ell \:\mid\: (\ell\leftarrow) \in R\}$. If $p\leftarrow$ is a fact in $P$, then it is also a fact in $^M\!P$. This means that $p\leftarrow$ is an element of $^M\!P$, and thus $p$ is in $K_M^P\uparrow 1$. Otherwise, there is a rule $r$ in $P$, such that [-]{}[ ]{} $head(r) = p$; $pos(r) = \emptyset$; $neg(r) \cap M = \emptyset$; and for each $\ell \in agg(r)$ we have that there exists a solution of $\ell$ of the form $\langle \emptyset,J\rangle$ such that $M \cap J =\emptyset$. The rule $p\leftarrow agg(r)$ is a rule in $^M\!P$. From Lemma \[l1\] we can conclude that $(\emptyset,M) \models agg(r)$, thus ensuring that $p \in K^P_M\uparrow 1$. Let $p \in K^P_M\uparrow 1$. Thus, there exists a rule $r'\in {^M\!P}$ such that $(\emptyset,M)\models body(r)$ and $head(r')=p$. This means that there is a rule $r \in P$ such that [-]{}[ ]{} $head(r) = head(r') = p$; $M \cap neg(r) = \emptyset$; $pos(r) = \emptyset$; and $agg(r) = agg(r')$. Since $(\emptyset,M) \models agg(r)$, we have that, for each $c\in agg(r)$, $\langle \emptyset, {\cal H}(c)\setminus M\rangle$ is a solution of $c$. This means that the rule $p \leftarrow $ is in $unfolding(P,M)$. This also means that $p\in T_R\uparrow 1$. *Step:* Let us assume that the result holds for $i \leq k$ and consider the iteration $k+1$. [$\bullet$]{}[ ]{} Let $p \in T_R\uparrow (k+1)$ and $p \not\in T_R\uparrow k$. Thus, there is a rule $r'$ in $R$ such that [-]{}[ ]{} $head(r')=p$; and $pos(r') \subseteq T_R\uparrow k$. This implies that there is a rule $r\in P$ such that [-]{}[ ]{} $head(r) = p$; $pos(r) \subseteq T_R\uparrow k$; $M \cap neg(r) =\emptyset$; and for each $c \in agg(r)$, there is a solution $S_c$ s.t. $S_c{\textit{.}}p \subseteq T_R\uparrow k$ and $M \cap S_c{\textit{.}}n = \emptyset$. This also means that $p \leftarrow pos(r),agg(r)$ is a rule in $^M\!P$. We already know that $pos(r) \subseteq K^P_M\uparrow k$. Now we wish to show that $(K^P_M\uparrow k, M) \models agg(r)$. Lemma \[l1\] shows that, for each $c \in agg(r)$, $\langle S_c{\textit{.}}p, {\cal H}(c)\setminus M\rangle$ is a solution of $c$. This allows us to conclude that $p \in K^P_M\uparrow (k+1)$. Let $p \in K^P_M\uparrow (k+1)$ and $p \not\in K^P_M\uparrow k$. This means that there is a rule $r'$ in $^M\!P$ such that [-]{}[ ]{} $head(r') = p$; $pos(r') \subseteq K^P_M\uparrow k$; and $( K^P_M\uparrow k, M) \models body(r')$ This also means that there is a rule $r$ in $P$ such that [-]{}[ ]{} $head(r) = head(r') = p$; $agg(r) = agg(r')$; $pos(r) = pos(r')$; $neg(r) \cap M = \emptyset$; and for each $c \in agg(r)$, $S_c = \langle K^P_M\uparrow k\cap M\cap {\cal H}(c), {\cal H}(c) \setminus M \rangle$ is a solution of $c$. This means that there is a rule $r''$ in $unfolding(P,M)$ such that: [-]{}[ ]{} $head(r'') = p$ $pos(r'') = pos(r) \cup \bigcup_{c\in agg{r}} S_c{\textit{.}}p$ Since each $S_c{\textit{.}}p \subseteq K^P_M\uparrow k=T_R\uparrow k$ for each $c \in agg(r)$ and $pos(r) \subseteq K^P_M\uparrow k=T_R\uparrow k$, we have that $p \in T_R\uparrow (k+1)$. \[fixpointasp\] Let $P$ be a program with aggregates. $M$ is an answer set of $unfolding(P,M)$ iff $M$ is a fixpoint answer set of $P$. Let $R = unfolding(P,M)$. We have that $M$ is an answer set of $P$ iff $M = T_R\uparrow \omega$ iff $M = K^P_M \uparrow \omega$ (Lemma \[l2\]). The results from [@SonPE05] and Theorem \[fixpointasp\] provide us a direct connection between fixpoint answer sets and other semantics for logic programs with aggregates. Faber et al.’s Minimal Model Semantics -------------------------------------- The notion of answer set proposed in [@faberLP04] is based on a new notion of reduct, defined as follows. Given a program $P$ and a set of ASP-atoms $M$, the [*reduct of P with respect to M*]{}, denoted by $\Gamma(M,P)$, is obtained by removing from $ground(P)$ those rules whose body cannot be satisfied by $M$. In other words, $\Gamma(M,P) = \{r \mid r \in ground(P), M \models body(r)\}$. \[FLP-answer set, [@faberLP04]\] \[d-faberLP04\] For a program $P$, $M$ is an [*FLP-answer set*]{} of $P$ if it is a minimal model of $\Gamma(M,P)$. The following theorem derives directly from Theorem \[fixpointasp\] and [@SonPE05]. \[th2\] Let $P$ be a program with aggregates. If $M$ is a fixpoint answer set, then $M$ is an FLP-answer set of $P$. Observe that there are cases where FLP-answer sets are not fixpoint answer sets. Consider the program $P$ where $$\begin{array}{lll} p(1)& \leftarrow & \textnormal{\sc Sum}(\{X \mid p(X)\}) \ge 0 \\ p(-1)& \leftarrow & p(1) \\ p(1)& \leftarrow & p(-1) \\ \end{array}$$ It can be checked that $M = \{p(1), p(-1)\}$ is an FLP-answer set of $P$. It is possible to show that $\textnormal{\sc Sum}(\{X \mid p(X)\}) \geq 0$ has the following solutions: $\langle \emptyset, \{p(1),p(-1)\} \rangle$, $\langle \{p(1)\}, \{p(-1)\} \rangle$, $\langle \{p(1)\}, \emptyset \rangle$, and $\langle \{p(1),p(-1)\}, \emptyset \rangle$. We have that $K^P_M(\emptyset) = \emptyset$ since $\langle \emptyset, \emptyset \rangle$ is not a solution of $\textnormal{\sc Sum}(\{X \mid p(X)\}) \ge 0$. This implies that $lfp(K^P_M) = \emptyset$. Thus, $M$ is not a fixpoint answer set of $P$. It can be easily verified that $P$ does not have any fixpoint answer set. If we replace in $P$ the rule $p(1) \leftarrow \textnormal{\sc Sum}(\{X \mid p(X)\}) \ge 0$ with the intuitively equivalent [Smodels]{} weight constraint rule $$p(1) \leftarrow 0 [p(1) = 1, p(-1)=-1]$$ we obtain a program that does not have answer sets in [Smodels]{}. The above example shows that our characterization differs from [@faberLP04]. Our definition is closer to [Smodels]{}’ understanding of aggregates. Approximation Semantics for Logic Programs with Aggregates {#pelov} ---------------------------------------------------------- The work of Pelov et al. [@PelovDB03; @Pelov04; @PelovDB04] contains an elegant generalization of several semantics of logic programs to logic programs with aggregates. The key idea in this work is the use of approximation theory in defining several semantics for logic programs with aggregates (e.g., two-valued semantics, ultimate three-valued stable semantics, three-valued stable model semantics). In particular, in [@PelovDB04], the authors describe a fixpoint operator, called $\Phi^{appr}_P$, operating on 3-valued interpretations and parameterized by the choice of approximating aggregates. It is possible to show the following results: - Whenever the approximating aggregate used in $\Phi^{appr}_P$ is the *ultimate approximating aggregate* [@PelovDB04], then the fixpoint semantics defined by the operator $K^P_M$ coincides with the two-valued stable model semantics defined by the operator $\Phi^{appr}_P$. - It is possible to prove a stronger result, showing that, if $I \subseteq M$ then $K_M^P(I) = \Phi^{aggr,1}_P(I,M)$, where $\Phi^{aggr,1}_P(I,M)$ denotes the first component of $\Phi^{aggr}_P(I,M)$. In other words, when ultimate approximating aggregates are employed and $M$ is an answer set, then the fixpoint operator of Pelov et al. and $K_M^P$ behave identically. We will prove next the first of these two results. The proof of the second result (kindly contributed by one of the anonymous reviewers) can be found in Appendix A. We will make use of the translation of logic programs with aggregates to normal logic programs, denoted by $tr$, described in [@PelovDB03]. The translation in [@PelovDB03] and the unfolding described in the previous subsection are similar[^5]. For the sake of completeness, we will review the translation of [@PelovDB03], presented using the notation of our paper. Given a ground logic program with aggregates $P$, $tr(P)$ denotes the ground normal logic program obtained after the translation. The process begins with the translation of each aggregate atom $\ell$ of the form $aggr(s) \:\: \texttt{op} \:\: Result$ into a disjunction $tr(\ell) = \bigvee F^{{\cal H}(\ell)}_{(s_1,s_2)}$, where $s_1 \subseteq s_2 \subseteq {\cal H}(\ell)$, and each $F^{{\cal H}(\ell)}_{(s_1,s_2)}$ is a conjunction of the form $$\bigwedge_{l \in s_1} l \wedge \bigwedge_{l \in {\cal H}(\ell) \setminus s_2} \naf l$$ The construction of $tr(\ell)$ considers only the pairs $(s_1,s_2)$ that satisfy the following condition: each interpretation $I$ such that $s_1 \subseteq I$ and ${\cal H}(\ell) \setminus s_2 \cap I = \emptyset$ must satisfy $\ell$. The translation $tr(P)$ is then created by replacing rules with disjunction in the body by a set of standard rules in a straightforward way. For example, the rule $$a \leftarrow (b \vee c), d$$ is replaced by the two rules $$\begin{array}{lcl} a \leftarrow b, d & \hspace{1cm} & a \leftarrow c, d\\ \end{array}$$ From the definitions of $tr(\ell)$ and of aggregate solutions, we have the following simple lemma: \[tr1\] For every aggregate atom $\ell$ of the form $aggr(s) \:\: \texttt{op} \:\: Result$, $S$ is a solution of $\ell$ if and only if $F^{{\cal H}(\ell)}_{(S{\textit{.}}p,{\cal H}(\ell) \setminus S{\textit{.}}n)}$ is a disjunct in $tr(\ell)$. We next show that fixed point answer sets of $P$ are answer sets of $tr(P)$. For a program $P$, $M$ is a fixpoint answer set of $P$ iff $M$ is an answer set of $tr(P)$. Let $M$ be an interpretation of $P$ and $R = unfolding(P,M)$. We have that $R$ is a positive program. Furthermore, let $Q$ denote the result of the Gelfond-Lifschitz reduction of $tr(P)$ with respect to $M$, i.e., $Q = (tr(P))^M$. We will prove by induction on $k$ that if $M$ is an answer set of $Q$ then $T_Q \uparrow k = T_R \uparrow k$ for every $k \ge 0$. The equation holds trivially for $k=0$. Let us consider now the case for $k$, assuming that $T_Q \uparrow l = T_R \uparrow l$ for $0 \le l < k$. 1. Consider $p \in T_Q \uparrow k$. This means that there exists some rule $r' \in Q$ such that $head(r') = p$ and $body(r') \subseteq T_Q \uparrow (k-1)$. $r' \in Q$ if and only if there exists some $r \in P$ such that $r' \in tr(r)$. Together with Lemma \[tr1\], we can conclude that there exists a sequence of aggregate solutions $\langle S_c \rangle_{c \in agg(r)}$ for the aggregate atoms in $body(r)$ such that $pos(r') = pos(r) \cup \bigcup_{c \in agg(r)} S_c{\textit{.}}p$, and $(neg(r) \cup \bigcup_{c \in agg(r)} S_c{\textit{.}}n) \cap M = \emptyset$. This implies that $r' \in R$. Together with the inductive hypothesis, we can conclude that $p \in T_R \uparrow k$. 2. Consider $p \in T_R \uparrow k$. This implies that there exists some rule $r' \in R$ such that $head(r') = p$ and $body(r') \subseteq T_R \uparrow (k-1)$. From the definition of $R$, we conclude that there exists some rule $r \in ground(P)$ and a sequence of aggregate solutions $\langle S_c \rangle_{c \in agg(r)}$ for the aggregate atoms in $body(r)$ such that $pos(r') = pos(r) \cup \bigcup_{c \in agg(r)} S_c{\textit{.}}p$, and $(neg(r) \cup \bigcup_{c \in agg(r)} S_c{\textit{.}}n) \cap M = \emptyset$. Using Lemma \[tr1\], we can conclude that $r' \in Q$. Together with the inductive hypothesis, we can conclude that $p \in T_Q \uparrow k$. Similar arguments can be used to show that if $M$ is an answer set of $R$, $T_Q \uparrow k = T_R \uparrow k$ for every $k \ge 0$, which means that $M$ is an answer set of $Q$. In [@PelovDB03], it is shown that answer sets of $tr(P)$ coincide with the *two-valued partial stable models* of $P$ (defined by the operator $\Phi^{aggr}_P$). This, together with the above lemma and Theorem \[fixpointasp\], allows us to conclude the following theorem. For a program with aggregates $P$, $M$ is an fixpoint answer set of $P$ if and only if it is a fixpoint of the operator $\Phi^{aggr}_P$ of [@PelovDB04]. Complexity Considerations ------------------------- We will now discuss the complexity of computing fixpoint answer sets. In what follows, we will assume that the program $P$ is given and it is a ground program whose language is finite. By the *size* of a program, we mean the number of rules and atoms present in it, as in [@faberLP04]. Observe that, in order to support the computation of the iterations of the $K^P_M$ operator, we need the ability to determine whether a given $\langle I,J \rangle$ is a solution of an aggregate atom. For this reason, we classify programs with aggregates by the computational complexity of its aggregates. We define a notion, called $C$-[*decidability*]{}, where $C$ denotes a complexity class in the complexity hierarchy, as follows. Given an aggregate atom $\ell$ and an interpretation $M$, we say that $\ell$ is $C$-[*decidable*]{} if its truth value with respect to $M$ can be decided by an oracle of the complexity $C$. A program $P$ is called [*$C$-decidable*]{} if the aggregate atoms occurring in $P$ are $C$-decidable. It is easy to see that aggregate atoms built using the standard aggregate functions ([Sum, Min, Max, Count, Avg]{}) and relations ($=,\ne,\geq,>,\leq,<$) are polynomially decidable. The solution checking problem is defined as follows. \[(SCP) Solution Checking Problem\] [***Given***]{} an aggregate atom $\ell$, its language extension ${\cal H}(\ell)$, and a pair of disjoint sets $I, J \subseteq {\cal H}(\ell)$, [***Determine***]{} whether $\langle I,J \rangle$ is a solution of $\ell$. We have the following lemma. \[l5\] The SCP is in [**co[-]{}NP**]{}$^C$ for $C$-decidable aggregate atoms. We will show that the complexity of the inverse problem of the SCP is in [**NP**]{}$^C$, i.e., determining whether $\langle I, J \rangle$ is not a solution of $\ell$ is in [**NP**]{}$^C$. By definition, $\langle I, J \rangle$ is not a solution of $\ell$ if there exists an interpretation $M$ such that $I \subseteq M$, $J \cap M = \emptyset$, and $M \not\models \ell$. To answer this question, we can guess an interpretation $M$ and check whether $\ell$ is false in $M$. If it is, we conclude that $\langle I, J \rangle$ is not a solution of $\ell$. Because $\ell$ is $C$-decidable and there are at most $2^{|{\cal H}(\ell) \setminus (I \cup J)|}$ interpretations that can be used in checking whether $\langle I, J \rangle$ is not a solution of $\ell$, we conclude that the complexity of the inverse problem is in [**NP**]{}$^C$. We will now address the problem of answer set checking and determining the existence of answer set. \[(ACP) Answer Set Checking Problem\] [***Given***]{} an interpretation $M$ of $P$, [***Determine***]{} whether $M$ is an answer set of $P$. \[(AEP) Answer Set Existence Problem\] [***Given***]{} a program $P$, [***Determine***]{} whether $P$ has a fixpoint answer set. The following theorem follows from Lemma \[l5\]. The ACP of $C$-decidable programs is in [**co[-]{}NP**]{}$^C$. The main tasks in checking whether $M$ is an answer set of $P$ are [*(i)*]{} computing $^M\!P$; and [*(ii)*]{} computing $lfp(K_M^P)$. Obviously, $^M\!P$ can be constructed in time linear in the size of $P$, since the reduction relies on the satisfiability test of a negation-as-failure literal $\ell$ w.r.t. $M$. Computing $lfp(K_M^P)$ requires at most $na$ iterations, i.e., $lfp(K_M^P) = K_M^P \uparrow na$, where $na$ is the number of atoms of $P$, each step is in [**co[-]{}NP**]{}$^C$, due to the requirement of solution checking. This theorem allows us to conclude the following result. The AEP for $C$-decidable program is in [**NP**]{}$^{\textnormal{\bf co{-}NP}^C}$. So far, we discussed the worst case analysis of answer set checking and determining the existing of an answer set based on a general assumption about the complexity of computing the aggregate functions and checking the truth value of aggregate atoms. Next we analyze the complexity of these problems w.r.t. the class of programs whose aggregate atoms are built using standard aggregate functions and operators. ### Complexity of Solution Checking for Standard Aggregates We will now focus on the class of programs defined in Section \[sec2\] with standard aggregate functions ([Sum, Min, Max, Count, Avg]{}) and relations ($=$, $\geq$, $>$, $\leq$, $<$, $\ne$). It is easy to see that all aggregate atoms involving these functions and relations are [**P**]{}-decidable. Therefore, by Lemma \[l5\], the SCP for standard aggregates will be at most [**co[-]{}NP**]{}. We will now show that it is [**co[-]{}NP**]{}-complete. \[t10\] The SCP for standard aggregates is [**co[-]{}NP**]{}-complete. Membership follows from Lemma \[l5\]. To prove hardness, we will translate a well-known [**NP**]{}-complete problem, namely the subset sum problem [@cormen], to the complement of the solution checking problem. An instance $Q$ of the subset sum problem is given by a set of non-negative integers $S$ and an integer $t$, and the question is to determine whether there exists any non-empty subset $A$ of $S$ such that $\sum_{x \in A} x = t$. Let ${\cal H}(\ell) = \{p(x) \mid x \in S\}$ for some unary predicate $p$. We define an instance of the solution checking problem, $s(Q)$, by setting $I=\emptyset$, $J=\emptyset$, and $\ell = \textnormal{\sc Sum}(\{X \mid p(X)\}) \ne t$. It is easy to see that $s(Q)$ is equivalent to $Q$ as follows: if $\langle I,J \rangle$ is a solution of $\ell$ then $Q$ does not have an answer; if $\langle I,J \rangle$ is not a solution to $\ell$ then $Q$ has an answer. This proves the desired result. The above theorem shows that, in general, the inclusion of standard aggregates implies that the answer set checking problem and the problem of determining the existing of an answer set are in [**co[-]{}NP**]{} and [**NP**]{}$^{\textnormal{\bf co{-}NP}}$ respectively. Fortunately, there is a large class of programs with standard aggregates for which the complexity of these two problems are in [**P**]{} and [**NP**]{} respectively, as shown next. \[l10\] Let $\ell$ be an aggregate of the form $\textnormal{\sc Sum}(\{X \mid p(X)\}) = v$, where $v$ is a constant in $\mathbf{R}$. Let $I,J \subseteq {\cal H}(\ell)$ such that $I \cap J = \emptyset$. Then, determining whether $\langle I, J \rangle$ is a solution of $\ell$ can be done in time polynomial in the size of ${\cal H}(\ell)$. Let us denote with $\pi$ the function that projects an element $p$ of ${\cal H}(\ell)$ to the value that $p$ assigns to the collected variable. This value will be denoted by $\pi(p)$. We prove the lemma by providing a polynomial algorithm for determining whether $\langle I, J \rangle$ is a solution of $\ell$. iiiiīiiiīiiiiīiiiīiiiīiiiīiiiīiiiīiii1: ($v$, $\langle I, J\rangle$, ${\cal H}(\ell)$)\ 2: compute $s = \Sigma_{p \in I} \pi(p)$\ 3: $s \ne v$ [**then return false**]{}\ 4: ${\cal H}(\ell) \setminus (I \cup J) = \emptyset$ [**then return true**]{};\ 5: ($p \in {\cal H}(\ell) \setminus (I \cup J)$)\ 6: $\pi(p) \ne 0$ [**then return false**]{}\ 7:\ 8:\ It is easy to see that the above algorithm returns true (resp. false) if and only if $\langle I, J \rangle$ is (resp. is not) a solution of $\ell$. Furthermore, the time complexity of the above algorithm is polynomial in the size of ${\cal H}(\ell)$. This proves the lemma. The above lemma shows that the solution checking problem can be solved in polynomial time for a special type of standard aggregate atoms. Indeed, this can be proven for all standard aggregates but those of the form $\textnormal{\sc Sum} \ne v$ and $\textnormal{\sc Avg} \ne v$. Let $\ell$ be the aggregate $agg(s) \;\mathbf{ op }\; v$ where $agg \not\in \{\textnormal{\sc Sum, Avg}\}$ or $agg \in \{\textnormal{\sc Sum, Avg}\}$ and $\mathbf{op}$ is not ‘$\ne$’. Let $I, J \subseteq {\cal H}(\ell)$, $I \cap J = \emptyset$, and $v \in \mathbf{R}$. Then, checking if $\langle I, J \rangle$ is a solution of $\ell$ can be done in time polynomial in the size of ${\cal H}(\ell)$. The proof can be done similarly to the proof of Lemma \[l10\]: for each type of atom, we develop an algorithm, which returns true (resp. false) if $\langle I,J\rangle$ is (resp. is not) a solution of $\ell$. For brevity, we only discuss the steps which need to be done. It should be noted that each of these steps can be done in polynomial time in the size of ${\cal H}(\ell)$, which implies the conclusion of the lemma. [Sum:]{} Let $s = \sum_{p\in I} \pi(p)$. All cases can be handled in time $O(|{\cal H}(\ell)|)$. Let us consider the various cases for [**op**]{}. [$\bullet$]{}[ ]{} The case [**op**]{} is ’=’ has been discussed in Lemma \[l10\]. For ${\bf op} \in\{\geq, >\}$, let $H_1 = \{p \mid p \in {\cal H}(\ell) \setminus (I \cup J),\; \pi(p)<0\}$. We have that $\langle I,J \rangle$ is a solution of $\ell$ if and only if $s \; \mathbf{op} \; v$ and $\sum_{p \in H_1} \pi(p) + s \; \mathbf{op} \; v$. For $\mathbf{op} \in \{\leq,<\}$, let $H_1 = \{p \mid p \in {\cal H}(\ell) \setminus (I \cup J),\; \pi(p)>0\}$. We have that $\langle I,J \rangle$ is a solution of $\ell$ if and only if $s \; \mathbf{op} \; v$ and $\sum_{p \in H_1} \pi(p) + s \; \mathbf{op} \; v$. [Count:]{} Let $c = |I|$ and $H_1 = {\cal H}(\ell) \setminus (I \cup J)$. All cases can be handled in time $O(|{\cal H}(\ell)|)$. [$\bullet$]{}[ ]{} If $\mathbf{op} \in \{>,\ge\}$, then $\langle I, J\rangle$ is a solution of $\ell$ if and only if $c \; \mathbf{op} \; v$. If $\mathbf{op} \in \{=,<,\le\}$, then $\langle I, J\rangle$ is a solution of $\ell$ if and only if $c \; \mathbf{op} \; v$ and $c + |H_1| \; \mathbf{op} \; v$. If $\mathbf{op}$ is $\ne$, then $\langle I, J\rangle$ is a solution of $\ell$ if and only if either [*(i)*]{} $|I| > v$; or [*(ii)*]{} $|I| < v$ and $|H_1| < v-|I|$. [Min:]{} Let $c = \min \{\pi(p) \mid p \in I\}$ and $c_1 = \min \{\pi(p) \mid p \in {\cal H}(\ell) \setminus (I \cup J)\}$. All cases can be handled in time $O(|{\cal H}(\ell)|)$. [$\bullet$]{}[ ]{} If $\mathbf{op}$ is $=$ then we have that $\langle I,J \rangle$ is a solution of $\ell$ if and only if $c = v $ and $c_1 \ge v$. If $\mathbf{op} \in \{\leq,<\}$ then $\langle I,J \rangle$ is a solution of $\ell$ if and only if $c \; \mathbf{op} \; v $. If $\mathbf{op} \in \{\geq,>\}$ then $\langle I,J \rangle$ is a solution of $\ell$ if and only if $c \; \mathbf{op} \; v$ and $c_1 \; \mathbf{op} \; v$. If $\mathbf{op}$ is $\ne$ then $\langle I,J \rangle$ is a solution of $\ell$ if and only if either [*(i)*]{} $c < v$; or [*(ii)*]{} $c > v$ and for every $p \in H_1$, $\pi(p) \ne v$. [Max:]{} Let $c = \max \{\pi(p) \mid p \in I\}$ and $c_1 = \max \{\pi(p) \mid p \in {\cal H}(\ell) \setminus (I \cup J)\}$. All cases can be handled in time $O(|{\cal H}(\ell)|)$. [$\bullet$]{}[ ]{} If $\mathbf{op}$ is $=$ then $\langle I,J \rangle$ is a solution of $\ell$ if and only if $c = v $ and $c_1 \le v$. If $\mathbf{op} \in \{\geq,>\}$ then $\langle I,J \rangle$ is a solution of $\ell$ if and only if $c \; \mathbf{op} \; v $. If $\mathbf{op} \in \{\leq,<\}$ then $\langle I,J \rangle$ is a solution of $\ell$ if and only if $c \; \mathbf{op} \; v$ and $c_1 \; \mathbf{op} \; v$. If $\mathbf{op}$ is $\ne$ then $\langle I,J \rangle$ is a solution of $\ell$ if and only if either [*(i)*]{} $c > v$; or [*(ii)*]{} $c < v$ and for every $p \in H_1$, $\pi(p) \ne v$. [Avg:]{} Let $a = \frac{\sum_{p\in I} \{\pi(p)\}}{|I|}$ and $H_1 = {\cal H}(\ell)\setminus (I\cup J)$. [$\bullet$]{}[ ]{} If $\mathbf{op}$ is $=$ then $\langle I,J\rangle$ is a solution of $\ell$ if and only if $a = v$ and for every $p \in H_1$, $\pi(p) = v$. This can be done in time $O(|{\cal H}(\ell)|)$. If $\mathbf{op} \in \{\geq,>\}$ then let $e_1,\dots, e_r$ be an enumeration of $H_1$ such that $\pi(e_i) \leq \pi(e_{i+1})$ for $1 \leq i\leq r-1$. $\langle I,J\rangle$ is a solution of $\ell$ if and only if $a \;\mathbf{op}\; v$ and for each $ 0 \leq h \leq r$, $$\sum_{p\in I}\pi(p) + \sum_{i=1}^h \pi(e_i) \;\mathbf{op}\; v\cdot |I|+ v \cdot h.$$ This can be accomplished in time $O(|{\cal H}(\ell)|^2)$. If $\mathbf{op} \in \{\leq, <\}$ then let $e_1,\dots, e_r$ be an enumeration of $H_1$ such that $\pi(e_i) \geq \pi(e_{i+1})$ for $1 \leq i\leq r-1$. $\langle I,J\rangle$ is a solution of $\ell$ if and only if $a \;\mathbf{op}\; v$ and for each $ 0 \leq h \leq r$, $$\sum_{p\in I}\pi(p) + \sum_{i=1}^h \pi(e_i) \; \mathbf{op} \; v\cdot |I|+ v \cdot h.$$ This can be accomplished in time $O(|{\cal H}(\ell)|^2)$. The above lemma shows that there is a large class of programs with aggregates for which the problem of checking an answer set and the problem of determining the existence of an answer set belongs to the class [**P**]{} and [**NP**]{} respectively. Observe that similar results can be extrapolated from the discussion in Pelov’s doctoral dissertation [@Pelov04]. Conclusions and Future Work =========================== In this technical note, we defined $K^P_M$, a fixpoint operator for verifying answer sets of programs with aggregates. We showed that the semantics for programs with aggregates described by this operator provides a new characterization of the semantics of [@SonPE05] for logic programs with aggregates. This operator converges to the same semantics as in [@Pelov04] when ultimate approximating aggregates are used. We also related this semantics to recently proposed semantics for aggregate programs. We discussed the complexity of the answer set checking problem and the problem of determining the existence of an answer set. We showed that, for the class of programs with standard aggregates without the relation $\ne$ for [Sum]{} and [Avg]{}, the complexity of these two problems remains unchanged comparing to that of normal logic programs. In the future, we would like to use this idea in an efficient implementation of answer set solvers with aggregates. ### Acknowledgments {#acknowledgments .unnumbered} The authors wish to thank the anonymous reviewers for their insightful comments and for pointing out relationships with existing literature, and Dr. Hing Leung for his suggestions. The research has been partially supported by NSF grants HRD-0420407, CNS-0454066, and CNS-0220590. 2001\. . MIT Press, Cambridge, MA. , [Pontelli, E.]{}, [and]{} [Rossi, G.]{} 2001. Constructive negation and constraint logic programming with sets.  [*19,*]{} 3, 209–256. , [Pontelli, E.]{}, [and]{} [Rossi, G.]{} 2003. Intensional Sets in CLP. In [*International Conference on Logic Programming*]{}, Springer, 284–299. , [Pontelli, E.]{}, [and]{} [Son, T. C.]{} 2004. Smodels with CLP and its Applications: a Simple and Effective Approach to Aggregates in ASP. In [*International Conference on Logic Programming*]{}, Springer, 73–89. , [Leone, N.]{}, [and]{} [Pfeifer, G.]{} 2004. Recursive Aggregates in Disjunctive Logic Programs: Semantics and Complexity. In [*JELIA*]{}, Springer, 200–212. 2002\. . In [*Computational Logic: Logic Programming and Beyond*]{}, Springer Verlag, 413–451. 1988\. The Stable Model Semantics for Logic Programming. In [*International Conf. and Symp. on Logic Programming*]{}, MIT Press, 1070–1080. 1991\. Semantics of Logic Programs with Aggregates. In [*ISLP*]{}, MIT Press, 387–401. 1987\. *Foundations of Logic Programming*. Springer Verlag. , [Pirahesh, H.]{}, [and]{} [Ramakrishnan, R.]{} 1990. The Magic of Duplicates and Aggregates. In [*Int. Conf. on Very Large Data Bases*]{}, Morgan Kaufmann, 264–277. 2004\. . Ph.D. thesis, Katholieke Universiteit Leuven. , [Denecker, M.]{}, [and]{} [Bruynooghe, M.]{} 2003. . In [*[ASP: Advances in Theory and Implementation]{}*]{}, CEUR Workshop Proceedings. [29–42]{}. , [Denecker, M.]{}, [and]{} [Bruynooghe, M.]{} 2004. Partial Stable Models for Logic Programs with Aggregates. In [*LPNMR*]{}, Springer, 207–219. 1997\. Monotonic Aggregation in Deductive Database.  [*54,*]{} 1, 79–97. , [Pontelli, E.]{}, [and]{} [Elkabani, I.]{} 2005. . Tech. Rep. [CS-2005-006]{}, New Mexico State University. [www.cs.nmsu.edu/CSWS/php/techReports.php?rpt\_year=2005](www.cs.nmsu.edu/CSWS/php/techReports.php?rpt_year=2005). , [Arni, N.]{}, [and]{} [Ong, K.]{} 1993. Negation and Aggregates in Recursive Rules: the LDL++ Approach. In [*DOOD*]{}. 204–221. Appendix A — Correspondence between $K_M^P$ and $\Phi^{aggr}_P$ {#appendix-a-correspondence-between-k_mp-and-phiaggr_p .unnumbered} =============================================================== We assume that the readers are familiar with the notations and definitions introduced in [@PelovDB04]. The three-valued immediate consequence operator $\Phi_P^{aggr}$ of a program $P$ in [@PelovDB04], maps 3-valued interpretations to 3-valued interpretations. But 3-valued interpretations can be split up in pairs $(I, J)$ of two valued interpretations such that $I \subseteq J$. Hence, an operator $\Phi_P^{aggr}$ can be viewed as an operator from pairs $(I,J)$ to pairs $\Phi_P^{aggr}(I,J) = (I',J')$ of 2-valued interpretations. It follows that $\Phi_P^{aggr}$ determines two component operators $\Phi_P^{aggr,1} (I, J) = I'$ and $\Phi_P^{aggr,2} (I, J) = J'$. The correspondence between $K_M^P$ and $\Phi^{aggr}_P$ is shown in the following claim. [**Claim.**]{} For every $I \subseteq M$, $K_M^P(I) = \Phi^{aggr,1}_P(I,M)$. First, let us identify the aggregate atoms $agg(s) \;\; \mathbf{op} \;\; v$ in this paper with aggregate atoms $R(s, v)$ of [@PelovDB04]. E.g., [Max]{}$(s) = v$ corresponds to [Max]{}$(s, v)$; [Max]{}$(s) \le v$ corresponds to [Max]{}$_\le (s, v)$. Now we compare the definition of $K_M^P$ and $\Phi^{aggr,1}_P$ in the case that $I \subseteq M$. For simplicity let us assume that atom $a$ is defined by only one ground rule, say $r$. $a \in K_M^P(I)$ iff $pos(r)$ is true in $I$, $neg(r)$ is false in $M$, and for each $\ell \in aggr(r)$, $l$ has a solution $(I \cap M \cap {\cal H}(\ell), {\cal H}(\ell) \setminus M)$. $a \in \Phi^{aggr,1}_P(I,M)$ iff $pos(r)$ is true in $I$, $neg(r)$ is false in $M$, and for each $\ell \in aggr(r)$, $l$ evaluates to true, i.e., if $U^1_R(s^{(I,M)})) = t$. Here, $U^1_R$ is the first component of the three-valued aggregate, and $s^{(I,M)}$ is the evaluation of the set expression under the 3-valued interpretation $(I,M)$. All that remains to be done is to show that $(I \cap M \cap {\cal H}(\ell), {\cal H}(\ell) \setminus M)$ is a solution for $l$ iff $U^1_R(s^{(I,M)}) = t$. Recall that we are considering the case where $I \subseteq M$, therefore the first expression simplifies to $(I \cap {\cal H}(\ell), {\cal H}(\ell) \setminus M)$. Let us focus on set aggregates but the argument for multisets is the same. Let us consider an aggregate atom $$\ell = agg(s)\;\; \mathbf{op} \;\;v$$ where $$s = \{X \mid p(d_1,\ldots,d_{i-1},X, d_{i+1},\ldots,d_n)\}$$ and $X$ is the only variable and $d_1,\ldots,d_n$ are members of the Herbrand universe. For any $I \subseteq M$,\ $(I \cap {\cal H}(\ell), {\cal H}(\ell) \setminus M)$ is a solution for $\ell$\ iff for each $J$ such that $I \cap {\cal H}(\ell) \subseteq J$ and $J \cap ({\cal H}(\ell) \setminus M) = \emptyset$, $J \models \ell$\ iff for each $J$ such that $I \subseteq J \subseteq M$, $J \models \ell$. The latter equivalence is perhaps not entirely trivial but it follows easily from the fact that $J \models \ell \Leftrightarrow J'\models \ell$ whenever $J \cap {\cal H}(\ell) = J' \cap {\cal H}(\ell)$. In [@PelovDB04], the value $s^{(I,M)}$ is a three-valued (multi-)set, which can be written as a pair of two valued sets $(S_1, S_2)$ where $$S_1 = \{d \mid I \models p(d_1,\ldots,d_{i-1},d,d_{i+1},\ldots,d_n)\}$$ and $$S_2 = \{d \mid M \models p(d_1,\ldots,d_{i-1},d,d_{i+1},\ldots,d_n)\}.$$ By definition of $U^1_R$, $U^1_R(s^{(I,M)})) = t$ iff for each set $S$ such that $S_1 \subseteq S \subseteq S_2$, $R(S, v)$ is true. It is straightforward to see that the conditions in this paragraph and the previous one are equivalent. [^1]: A program $P$ with variables can be viewed as a shorthand for $ground(P)$. [^2]: Recall that an interpretation is a set of atoms in ${\cal B}_P$. [^3]: We would like to thank Vladimir Lifschitz for providing us this example. [^4]: A detailed comparison between the semantics in [@SonPE05] and earlier proposals for programs with aggregates can be found in the same report. [^5]: It should be noted that our translation builds on our previous work on semantics of logic programming with sets and aggregates [@DovierPR01; @DovierPR03; @ElkabaniPS04] and was independently developed w.r.t. the work in [@PelovDB03].

--- abstract: | We investigate the axisymmetric structure of collimating, relativistic, strongly magnetized (force-free) jets. In particular, we include the differential rotation of the foot points of the field lines in our treatment. The magnetic flux distribution is determined by the solution of the Grad-Shafranov equation and the regularity condition along the light surface. With differential rotation, i.e. the variation of the iso-rotation parameter $\omf $, the shape of the light surface is not known a priori and must be calculated in an iterative way. For the first time, we have calculated the force-free magnetic structure of truly two-dimensional, relativistic jets, anchored in a differentially rotating disk. Such an approach allows for a direct connection between parameters of the central source (mass, rotation) and the extension of the radio jet. In particular, this can provide a direct scaling of the location of the asymptotic jet light cylinder in terms of the central mass and the accretion disk magnetic flux distribution. We demonstrate that differentially rotating jets must be collimated to a smaller radius in terms of the light cylinder if compared to jets with rigid rotation. Also, the opening angle is smaller. Further we present an analytical estimate for the jet opening angle along the asymptotic branches of the light surface. In general, differential rotation of the iso-rotation parameter leads to an increase of the jet opening angle. Our results are applicable for highly magnetized, highly collimated, relativistic jets from active galactic nuclei and Galactic superluminal jet sources. Comparison to the M87 jet shows agreement in the collimation distance. We derive a light cylinder radius of the M87 jet of 50 Schwarzschild radii. author: - Christian Fendt - Elisabetta Memola date: 'Received ??; accepted ??' subtitle: The axisymmetric field structure of relativistic jets and the example of the M87 jet title: 'Collimating, relativistic, magnetic jets from rotating disks' --- \ psid[\_[D]{}]{} Formation of magnetic jets ========================== Observations of astrophysical jet sources have now established the idea that jet formation is always connected to the presence of an accretion disk and strong magnetic fields. This holds for various scales of energy output, jet velocity and nature of the jet emitting objects. Examples are jets from active galactic nuclei (AGN), Galactic superluminal jet sources, the example of a mildly relativistic jet from a neutron star (SS 433) and the numerous class of protostellar jets (see Zensus et al. 1995; Mirabel & Rodriguez 1995; Mundt et al. 1990; Ray et al. 1996). Magnetic jets are believed to originate very close to the central object in the interaction region with the accretion disk. Beside observational arguments also theoretical considerations have shown that magnetic fields play an important role in jet formation and propagation (Blandford & Payne 1982; Pudritz & Norman 1983; Shibata & Uchida 1985; Sakurai 1985; Camenzind 1987; Lovelace et al. 1991). If the central object is a black hole as it is the case for AGN and Galactic superluminal jet sources, the surrounding accretion disk is the only possible location for a field generation (by dynamo action or/and advection of flux). In the case of stellar objects (protostars, white dwarfs or neutron stars), the central star also carries a relatively strong magnetic field and it is not yet clear, whether the jet magnetosphere originates in the disk or in the star. However, a strong interaction between stellar field and accretion flow is evident. The jet formation process itself is not yet fully understood theoretically. In particular, for the mass transfer from the disk into the jet and the process of magnetic field generation a complete physical model is missing. However, over the last decades the basic ideas of Blandford & Payne (1982) have been extended by various authors. The general picture is the following. Matter is lifted from the disk into the magnetosphere and becomes magnetically accelerated (Ferreira 1997). Toroidal magnetic fields, generated by inertial back-reaction of the plasma on the poloidal field, may collimate the disk magnetosphere into a highly collimated jet flow (Camenzind 1987; Chiueh et al. 1991; Lovelace et al. 1991). In general, due to the complexity of the MHD equations, stationary relativistic models of magnetic jets has to rely on simplifying assumptions such as self-similarity (Contopoulos 1994, 1995), some other prescription of the field structure (Li 1993, Beskin 1997) or the restriction to asymptotic regimes (Chiueh et al. 1991; Appl & Camenzind 1993; Nitta 1994, 1995). For highly magnetized jets the force-free limit applies. This allows for a truly two-dimensional calculation of the magnetic field structure (Fendt et al. 1995; Fendt 1997a). Relativistic magnetohydrodynamics implies that poloidal electric fields, which are not present in Newtonian MHD, cannot be neglected anymore. From the observations we know that extragalactic jets as well as Galactic superluminal jets and protostellar jets are collimated almost to a cylindrical shape (Zensus et al. 1995; Ray et al. 1996; Mirabel & Rodriguez 1995). Theoretically, it has been shown that current carrying relativistic jets must collimate to a cylinder (Chiueh et al. 1991). For the asymptotic limit of a cylindrically collimated magnetic relativistic jet, Appl & Camenzind (1993) found a non-linear analytical solution for the trans-field force-balance in the case of a constant iso-rotation parameter. These results were further generalized for jets with differential rotation (Fendt 1997b). Such an asymptotic field distribution is especially interesting for jets emerging from (differentially rotating) accretion disks. In previous papers, we applied the asymptotic jet model of Appl & Camenzind (1993) as a boundary condition for the calculation of [*global, two-dimensional,*]{} force-free jet magnetospheres for rapidly rotating stars (Fendt et al. 1995) or rapidly rotating black holes (Fendt 1997a). In this paper, we continue our work on 2D force-free jet magnetospheres applying an asymptotic jet with [*differential rotation*]{} of the iso-rotation parameter $\omf$ as boundary condition for the global jet structure. Such an approach allows for a connection between the differentially rotating jet source – the accretion disk – and the asymptotic collimated jet. Since jet motion seems intrinsically connected to the accretion disk, differential rotation of the field lines should be a natural ingredient for any magnetic jet structure. As a principal problem for differentially rotating relativistic jet magnetospheres, position and shape of the singular light surface are not known [*a priori*]{}, but have to be calculated in an iterative way together with the magnetic flux distribution. In Sect.2 we recall some basic equations of the theory of relativistic magnetospheres and discuss several difficulties with the solution of the Grad–Shafranov (hereafter GS) equation. After some comments on the numerical approach in Sect.3, we discuss our results in Sect.4. A summary is given in Sect.5. Structure of magnetic jets ========================== Throughout the paper we apply the following basic assumptions: [*axisymmetry*]{}, [*stationarity*]{} and [*ideal MHD*]{}. We use cylindrical coordinates $(R,\phi,Z)$ or $(x,\phi,z)$ if normalized. The term [*’asymptotic’*]{} always denotes the limit of $Z >> R$ unless explicitly stated otherwise. We consider jets with a [*finite*]{} radius, $R < \infty$ for $Z\rightarrow \infty$. The force-free, cross–field force–balance ----------------------------------------- With the assumption of axisymmetry, a magnetic flux function can be defined $$\Psi = \frac {1}{2 \pi} \int {\vec {B}}_{\rm P} \cdot d{\vec{A}} ,\quad\quad R{\vec{B}}_{\rm P} = \nabla \Psi \wedge {\vec{e}}_{\phi },$$ measuring the magnetic flux through a surface element with radius $R$ and, in a similar way, the poloidal current through the same area $$I = \int {\vec{j}}_{\rm P} \cdot d{\vec{A}} = -\frac{c}{2}\,RB_{\phi},$$ which, in the force-free case, flows parallel to the flux surfaces, $I = I(\Psi)$. The structure of the magnetic flux surfaces is determined by the toroidal component of Ampére’s law, $\nabla \times \bp = 4\pi j_{\phi} /c $, where the toroidal electric current density has to be calculated from the equation of motion projected perpendicular to the flux surfaces (Camenzind 1987; Fendt et al. 1995). For strong magnetic fields, inertial forces of the matter can be neglected. This is the [*force-free*]{} limit and the equation of motion reduces to $0=c \rho_c\vec{E} + \vec{j}\times \vec{B}$ with the charge density $\rho_c$. Combining Ampére’s law and the force-free equation of motion the cross-field force-balance can be written as the modified relativistic GS equation, $$\begin{aligned} R\nabla\cdot\left({\frac{1-(R\omf(\Psi)/c)^2}{R^2}}\nabla\Psi\right) = & - & \frac{4}{c^2}\frac{1}{R}\frac{1}{2}\left(I^2(\Psi)\right)'\\ & - & R\,|\nabla\Psi|^2 \frac{1}{2}\left(\omf^2(\Psi)\right)', \nonumber \end{aligned}$$ where the primes denote the derivative $\frac{d}{d\Psi}$ (see Camenzind 1987; Okamoto 1992). $\omf$ is conserved along the flux surfaces, $\omf = \omf(\Psi)$. We will call it the [*iso-rotation*]{} parameter, defined by Ferraro’s law of iso-rotation. It can be understood as the angular velocity of the plasma, reduced by the slide along the field lines. Sometimes, it is called the angular velocity of the field lines. Both, the current distribution $I(\Psi)$ and the rotation law $\omf(\Psi)$ determine the source term for the GS equation and govern the structure of the magnetosphere. We apply the following normalization, $$\begin{aligned} R, Z & \Leftrightarrow & x\,\ro , z\,\ro,\nonumber \\ \omf & \Leftrightarrow & \omf\,(c/\ro)\,, \nonumber \\ \Psi & \Leftrightarrow & \Psi\,{\Psi }_{\rm max}\,,\nonumber\\ I & \Leftrightarrow & I\,I_{\rm max}\,. \nonumber\end{aligned}$$ As the length scale for the GS equation (3) the asymptotic radius $\ro$ of the light surface is selected (see below). In order to allow for a direct comparison to rigidly rotating magnetospheres, the normalization was chosen such that $\omf = 1$ at $ x = 1 $. With the chosen normalization, Eq.(3) can be written dimensionless, $$\begin{aligned} x\nabla\cdot\left({\frac{1-x^2\omf^2(\Psi)}{x^2}}\nabla\Psi\right) = & - & \frac{1}{x}\frac{g}{2}\left(I^2(\Psi)\right)' \nonumber \\ & - & x|\nabla\Psi|^2\frac{1}{2}\left(\omf^2(\Psi)\right)'.\end{aligned}$$ $g$ is a coupling constant describing the strength of the current term in the GS equation, $$g = \frac {4 I_{\rm max}^2 \ro^2}{c^2 {\Psi }_{\rm max}^2} = 4\, \left(\frac{I_{\rm max}}{10^{18} {\rm A}}\right)^{\!\!2} \left(\frac{\ro}{10^{16} {\rm cm}}\right)^{\!\!2} \left(\frac{{\Psi }_{\rm max}}{10^{33}\,{\rm Gcm}^2}\right)^{\!\!-2}$$ where the parameters are chosen for extragalactic jets. In this paper, $g$ is in accordance with the definition in Fendt et al. (1995) and differs from the definition in Appl & Camenzind (1993) by a factor of two, $g_{\rm Fendt} = 2\,g_{\rm AC}$[^1]. Interestingly, a coupling constant, defined in a similar way also for the differential rotation term, would be equal to unity. The GS equation is numerically solved applying the method of finite elements (see Appendix). Along the light surface, where $D\equiv 1-x^2\omf^2(\Psi)=0$, the GS equation reduces to the regularity condition, $$\nabla\Psi\cdot\nabla D = -g\frac{1}{2}\left(I(\Psi)^2\right)' - \frac{1}{2} |\nabla\Psi|^2\left(\ln\left(\omf(\Psi)^2\right)\right)',$$ which is equivalent to a Neumann boundary condition. However, for differentially rotating magnetospheres with $\omf = \omf(\Psi)$ the shape of this surface is not known a priori and has to be calculated in an iterative way together with the two-dimensional solution of the GS equation. For constant $\omf$ the light surface is of cylindrical shape. As we have shown in a previous publication (Fendt et al. 1995), our finite element code satisfies the regularity condition [*automatically*]{}, since the surface integral along the light surface vanishes. Discussion of the force-free assumption --------------------------------------- It is clear that relativistic jets must be highly magnetized. Only a high plasma magnetization gives jet velocities close to the speed of light (Fendt & Camenzind 1996). Therefore, for the calculation of [*field structure*]{} the force-free limit seems to be reasonable. However, one may question the assumption of a force-free [*asymptotic*]{} jet. In a fully self-consistent picture of magnetic jet formation, the asymptotic jet is located beyond the collimating, non force-free wind region and beyond the fast magnetosonic surface. The asymptotic jet parameters are determined by the wind motion. Thus, poloidal current and iso-rotation parameter of the field are [*not*]{} functions free of choice. The force-free region of a jet is located close to its origin, where the speed is low. Beyond the Alfvén surface plasma kinetic energy dominates the magnetic energy, which is just the contrary to force-freeness. For small plasma density, the Alfvén surface of the wind flow approaches the light surface. In this case the fast magnetosonic surface moves to infinity for a conical flow. Okamoto (1999) argues that a force-free field distribution extending to infinity in both $x$ and $z$ direction will asymptotically be of conical shape, i.e. un-collimated. However, his approach differs from ours in the sense that he [*assumes*]{} that all field lines will cross the light cylinder. Such an assumption [*per se*]{} prohibits any collimation. On the other hand, perfect jet collimation is an observational fact. Astrophysical jets (of very different energy scales) appear collimated to cylinders of finite radius. In general, the non force-free relativistic GS equation shows three inertial contributions, $$\begin{aligned} 0 & = & - \tilde{\kappa} \left(1 - M^2 - x^2\omf^2\right) + \left(1 - x^2\omf^2\right) \frac{\nabla_{\!\!\perp}\bp^2}{8\pi} + \frac{\nabla_{\!\!\perp}B_{\phi}^2}{8\pi} \nonumber \\ & + & \nabla_{\!\!\perp}P + \left(\frac{B_{\phi}^2}{4\pi} -\rho u_{\phi}^2\right) \frac{\nabla_{\!\!\perp}x}{x} - \frac{\bp^2\omf}{4\pi}\nabla_{\!\!\perp}(x^2\omf),\nonumber \end{aligned}$$ where $\nabla_{\!\!\perp}$ indicates the gradient perpendicular to $\Psi$, $\tilde{\kappa} \equiv \kappa\bp^2/4\pi = {\bf n}\cdot ({\bf B_p}\cdot\nabla){\bf B_p}/4\pi $ the poloidal field curvature, $\rho$ the mass density, $u_{\phi}$ the toroidal velocity, $P$ the gas pressure and $M$ the Alfvén Mach number (Chiueh et al. 1991). One can show that in the asymptotic, cylindrical jet the contribution of inertial terms in the force-balance across the field is weak. The contribution of gas pressure is usually negligible in astrophysical jets. Also, the centrifugal force does not play a role for radii larger than the Alfvén radius, since outside the Alfvén surface (where $M^2=1-x^2\omf^2$) the plasma moves with constant angular momentum. The curvature term vanishes in cylindrical geometry. Therefore, since for cylindrical jets the contribution from inertial terms is weak, the configuration is comparable to the force-free case. The source term of the GS equation may be reduced to a form similar to the common force-free limit. We suggest the phrase “quasi force-free” for such a configuration because the GS equation looks force-free even if the physical system is not magnetically dominated. In the force-free limit of a highly magnetized plasma the previous arguments also apply. However, in difference to the asymptotic regime considered above, the low plasma density implies that inertial terms are weak over the [*whole*]{} two-dimensional jet region. The centrifugal term $\rho u_{\phi}$ is weak even if the Alfvén surface now comes close to the light surface. Numerical calculations of the plasma motion along the field have shown that, for a high magnetization, the Alfvén Mach number $M$ grows almost linearly with radius but remains relatively low (Fendt & Camenzind 1996). Thus, the inertial curvature term should not play a dominant role. Contopoulos & Lovelace (1994) find from self-similar solutions that centrifugal forces are dominated by magnetic forces leading to a re-collimation of the outflow. In summary, our discussion of the inertial terms in the force-balance equation has shown that these terms are generally weak in the case of a high magnetization. We therefore think that for the calculation of the magnetic field structure in relativistic jets the force-free assumption is acceptable. The main motivation of the force-free assumption is clearly the reason of simplification. There is yet no other way to calculate a truly two-dimensional field distribution for relativistic jets. Naturally, with a force-free solution, nothing can be said about the flow acceleration itself. Location of the asymptotic light cylinder ----------------------------------------- The radius of the asymptotic light cylinder $\ro $ is the natural length-scale for the GS solution. A scaling of the GS solution in terms of the properties of the central object (e.g. its mass) relies on the proper connection between the asymptotic jet and the accretion disk. This is feasible only if differential rotation $\omf(\Psi)$ is included in the treatment (see Sect.3). In the following we consider the location of the light surface and its relation to the relativistic character of the magnetosphere from a general point of view. The light cylinder of a flux surface $\Psi$ is defined as a cylinder with radius $R =\rl(\Psi) \equiv c/\omf(\Psi)$. At this position the GS equation becomes singular. However, this light cylinder is only important for the field line [*if*]{} the field line actually intersects it as for $\Psi_{\rm out}$ in Fig.1. Only then, relativistic effects become dominant. For example, the poloidal electric field scales with the radius in units of the light cylinder radius, $E_{\rm P} = (R/\rl) B_{\rm P}$. On the other hand, in the case of $\Psi_{\rm in}$ in Fig.1, the asymptotic radius of the flux surface is smaller than its light cylinder radius $\rl(\Psi_{\rm in})$ (located between $\Psi_{\rm in}$ and $\Psi_0$), therefore relativistic effects are small. For jet solutions with rigid rotation $\omf$ all flux surfaces have the same light cylinder radius. Thus, the singular light surface of the magnetosphere is a cylinder. For jet solutions with differential rotation $\omf$ the flux surfaces have different light cylinder radii. The singular surface of the magnetosphere is not a cylinder anymore. It is now interesting to note that the case of differential rotation $\omf(\Psi)$ allows for a hypothetical field distribution where (i) the light radius of most of the flux surfaces is located within the jet radius, but where also (ii) the asymptotic radius of the flux surfaces is always smaller than their light radius. Such a field distribution would not have a singular light surface and could be considered as “non relativistic”, even if the hypothetical light radii of many field lines are inside the jet radius. Such a situation is not possible if the magnetosphere is constrained by a constant rotation $\omf$. This underlines the importance of the treatment of differential rotation for jets from accretion disks. A relativistic description for the jet magnetosphere is always required if the jet contains a flux surface for which the light radius is smaller than the asymptotic radius. (10,0) The regularity condition and the jet opening angle -------------------------------------------------- (0,0) (0,50) The regularity condition (5) is the natural boundary condition along the light surface. Although it is impossible to solve equation (5) explicitly, a general relation concerning the jet opening angle can be derived. First, we rewrite Eq.(5) as $$B_z = \frac{1}{4} g (I^2)' - \frac{1}{4} \bp^2\left(\frac{1}{\omf^2}\right)',$$ where $\omf(\Psi) = 1 / \xl(\Psi) \equiv \ro/\rl(\Psi)$ has been applied. From Eq.(6) it follows for the radial field component $B_x^2 = -g (I^2(\Psi))'/(1/\omf^2(\Psi))'$, if $\Psi$ intersects the light surface with vanishing $B_z$. On the other hand, considering a field line perpendicular to the light surface, $\nabla\Psi\perp\nabla D$, this provides a condition for the axial field component, $$B_z = \frac{g}{2} (I^2)' = \frac{\bp^2}{2}\left(\frac{1}{\omf^2}\right)'.$$ Interestingly, this is [*either*]{} only a function of the current distribution $I(\Psi)$ [*or*]{} depends only from the specification of the rotation law $\omf(\Psi)$. Further, in this case it is always $B_z > 0$, since $(1/\omf^2)' = (\xl^2)' > 0$. In particular, for the asymptotic ($z\rightarrow\infty$) part of the magnetosphere, this implies that [*only collimating field lines can cross the light surface*]{}. Now we consider the asymptotic branches of the light surface. For the asymptotic branch in $z$-direction it holds $(\nabla D)_x >> (\nabla D)_z \simeq 0$. Further, it is $B_x (\ln \omf^2)' = 0$, implying either a collimated field structure, $B_x \equiv 0$ or rigid rotation, $(\omf(\Psi))' \equiv 0$. From this we conclude that in the asymptotic regime of a cylindrical light surface, also the flux surfaces along this light cylinder must be of cylindrical shape. Collimation must occur in the non-asymptotic region of the jet. If we now assume that there exists an asymptotic part of the light surface in $x$-direction (where $x >> z$) and that $(\nabla D)_z >> (\nabla D)_x \simeq 0$, we derive an equation for the half jet [*opening angle*]{}, $$\alpha(\Psi) = \tan^{-1}\left(\sqrt{1 + \frac{1}{4}g\frac{(I^2(\Psi))'(\omf^2(\Psi))'}{\omf^4(\Psi)}}\right),$$ for the flux surfaces in this region. As a general example we apply the analytical solution obtained for the asymptotic jet (Fendt 1997b), $$\begin{aligned} \Psi(x) & \equiv & \frac{1}{b} \ln\left(1+\left(\frac{x}{a}\right)^2\right), \quad b \equiv \ln\left(1+\left(\frac{x_{\rm jet}}{a}\right)^2\right), \nonumber \\ \omf^2(\Psi) & = & \frac{g\,b^2}{4}\left(\frac{I^2(\Psi)}{(1-e^{-b\Psi})^2} - \frac{1}{(1-e^{-b})^2}\right) + \omf^2(1)\end{aligned}$$ for Eq.(9). Here, $b$ is a measure for the ratio of jet radius to jet core radius $a$. Finally, we obtain the half opening angle for the outermost flux surface $\Psi =1$, $$\alpha = \tan^{-1}\left(\sqrt{ 1 + \Reo \left(4 \Reo\left(\frac{e^b-1}{be^b}\right)^2 +\frac{g/2}{e^b-1}\right) }\right),$$ where $\Reo$ is defined as $(\omf(1))'/\omf(1)$. Note, that Eq.(10) is only valid in the limit of $(\nabla D)_x \simeq 0$. Fig.2 shows the variation of the opening angle $\alpha$ with the parameters $a$ and $b$ for two choices of the strength of the poloidal electric current. In general, jets with a strong differential rotation $\omf(\Psi)$ (i.e. large $\Reo$) have a larger opening angle. Also, jets with a large ratio of jet to core radius have a smaller opening angle. Therefore, jets originating in a small part of the accretion disk, equivalent to small value of $\Reo$, will be collimated to a smaller opening angle. It is interesting to note that, in the case of a rigid rotation, the limiting half opening angle is $45\degr$, [*independently*]{} of $g$ and $b$. The jet-disk connection, providing the true length scale of the GS solution =========================================================================== The natural length scale of the relativistic GS equation (3) is the asymptotic light cylinder $\ro $ (see Sect.2.3). Its size is related to the iso-rotation parameter $\omf(\Psi)$, which itself is connected to the angular rotation of the foot points of the field lines. Concerning the GS equation, the size of $\ro $ follows purely from electro-magnetic quantities, if the coupling constant $g$ is chosen. The GS solution can be scaled to any central object from stars to galactic nuclei as long as the interrelation of the parameters ${\Psi }_{\rm max}$, $I_{\rm max}$ and $\ro$ provides the same $g$. So far, no connection has been made to the type of central object. Here, we treat the question where the asymptotic light cylinder is located in physical units. In the case of rigid rotation, the light cylinder radius is usually estimated by choosing a distinct radial distance from the central object and defining $\omf$ under the assumption that the jet magnetosphere is anchored in that point. If the central object is a black hole, the marginally stable orbit implies an upper limit for $\omf$. For jets in AGN this estimate leads to the common conclusion that the light cylinder radius is at about $10\rs$ and the typical jet radius at about $100\,\rl$. Clearly, such arguments relies on the [*assumption*]{} that the field line emerging at this very special radius defining $\omf$ also extends to the light cylinder radius $\rl$ (see Sect.2.3). This picture changes, if differential rotation $\omf(\Psi)$ is considered. Different flux surfaces anchor at different foot point radii and have different light radii (Sect.2.3). Assuming a Keplerian rotation, the light surface radius $\rl(\Psi)$ is located at $$\rl(\Psi) = 4\times 10^{15}\cm \left(\frac{\rf(\Psi)}{R_{\rm S}}\right)^{\!\!3/2} \!\!\left(\frac{M}{10^{10}\,\msun} \right)\,,$$ where $R_{\rm S}$ is the Schwarzschild radius of a point mass and $\rf(\Psi)$ the foot point of the flux surface $\Psi$ on a Keplerian disk. A more general equation is $$\frac{\rl(\Psi)}{R_{\rm S}} = \sqrt{2}\left(\frac{\rf(\Psi)}{R_{\rm S}}\right)^{3/2}.$$ $\rf(\Psi)$ is determined from the magnetic flux distribution along the disk and is defined by a certain disk model. Fig.3 shows the location of the light radius $\rl$ for a field line anchored at a foot point radius $\rf$ in a Keplerian disk around a central object of mass $M$ (see Eq.11). Note that the unit of the field line footpoint radius in Eq.(12) and Fig.3 is the Schwarzschild radius. Therefore, Fig.3 is appropriate only for relativistic jets. The footpoint radii for protostellar jets are several [*stellar*]{} radii, corresponding to about $10^6$ Schwarzschild radii (which would be located far above the box in Fig.3). So far, nothing can be said about the location of the asymptotic radius of the field lines in general. The essential question is where the asymptotic radius of a flux surface is located in respect to its light cylinder. This question can only be answered by a detailed model considering the [**two-dimensional**]{} field distribution [*including differential rotation*]{} $\omf(\Psi)$. Only in such a model, the flux distribution of the asymptotic jet can be connected to the flux distribution of the “star”-disk system. Certainly, both boundary conditions – asymptotic jet and accretion disk – rely on model assumptions. However, in a self-consistent model these boundary conditions have to follow certain constraints (see Sect.4.1, 4.3). (5,0) The two-dimensional jet solution ================================ Disk and jet boundary condition ------------------------------- Three important boundary conditions determine the two-dimensional flux distribution. The first boundary condition is in the asymptotic region. Here we assume a cylindrically collimated jet. We apply the magnetic flux distribution derived by Fendt (1997b), where the rigidly rotating jet model of Appl & Camenzind (1993) is extended for differential rotation $\omf(\Psi)$. In particular, our asymptotic jet shows the typical jet core-envelope structure of magnetic flux and electric current, i.e. a configuration where most of the magnetic flux and poloidal electric current is concentrated within a “core” radius. The asymptotic model provides not only the asymptotic magnetic flux boundary condition but also the $\omf(\Psi)$ and $I(\Psi)$ distribution for the whole two-dimensional jet magnetosphere. In the model of Fendt (1997b) these functions follow from the solution of the one-dimensional (asymptotic) GS equation across the cylindrical jet and the prescription of $I(x) = (x/a)^2/(1+(x/a)^2)$ together with $\omf^2(x) = e^{h - h x}$, where $a$ is the core radius of the electric current distribution and $h$ governs the steepness of the $\omf$-profile[^2]. The second boundary condition is the magnetic flux distribution along the disk. This distribution is in general [*not known*]{} as a solution of the full MHD disk equations. Typical models rely on various simplifying assumptions, as e.g. stationarity, the distribution of magnetic resistivity or the disk turbulence governing a dynamo process. We apply an analytic flux distribution similar to the model of Khanna & Camenzind (1992), who calculated the stationary accretion disk magnetic field structure around a super massive black hole. The typical behavior of the magnetic flux distribution is (i) a slight increase of magnetic flux along the innermost disk, (ii) a small or vanishing flux at the inner disk radius, (iii) a strong increase of magnetic flux at intermediate radius (the core radius) and (iv) a saturating behavior for large radii. Using the normalization introduced above, we choose the following disk boundary magnetic flux distribution, $$\Psi_{\rm disk}(x) = \frac{1}{\tilde{b}} \ln\left(1 + \left(\frac{x-x_{\rm in}}{\tilde{a}}\right)^2\right)\,$$ with $\tilde{b} = \ln(1 + (x_{\rm disk}-x_{\rm in})^2/\tilde{a}^2))$ (see Fig.4). The parameters are: the core radius $\tilde{a}$, the disk outer radius, $x_{\rm disk}$ and the disk inner radius, $x_{\rm in}$. For simplicity, we choose $x_{\rm in} \simeq 0$ without loss of generality. Such a choice will definitely not influence the global jet solution which is normalized to the asymptotic light cylinder radius. The third boundary condition is the jet boundary $x_{\rm jet}(z)$. Along this boundary the flux distribution is $\Psi = 1$ by definition. However, the shape of this boundary is not known [*a priori*]{}. It must be determined by the regularity of the solution across the light surface (see also Fendt et al. 1995). A slightly different shape may give the same solution. However, the main features of the solution as opening angle or locus of the collimation are fixed by the internal equilibrium. Therefore, the regularity condition governs the shape of the jet boundary. For a given $I(\Psi)$, $\omf(\Psi)$, disk and jet boundary condition, the jet radius $x_{\rm jet}(z)$ is uniquely determined. The inner spherical grid boundary with radius $x_{\star}$ close to the origin, indicates the regime of the central source, possibly enclosing a collapsed object. Neutron stars or magnetic white dwarfs may carry their own magnetic field, a black hole can only be thread by the disk magnetic field. In any case, the magnetic flux distribution is a combination of “central” magnetic flux and disk magnetic flux $\Psi = \Psi_{\star} + \Psi_{\rm disk}$. For simplicity we assume that the magnetic flux increases monotonously from the axis to the disk edge and $ \Psi(x_{\star}) = \Psi_{\star}(x_{\star}) $ and $\Psi_{\rm disk}(x_{\star}) = 0$. (0,0) (-10,0) (80,0) The two-dimensional collimating magnetic field structure -------------------------------------------------------- Results of numerical solutions of the GS equation are presented in Fig.5. Shown is the two-dimensional structure of the magnetic flux surfaces as projection of the helical field lines onto the meridional plane. In general, for a choice of the “free functions” $I(\Psi)$ and $\omf(\Psi)$, here taken from the asymptotic cylindrical jet solution, the field structure is determined by the boundary conditions and the regularity condition along the light surface. We calculated two solutions with a different choice for the steepness parameter in the iso-rotation $\omf$. The first solution is for $h=0.2$ (Fig.5, [*left*]{}). This is more close to the case of rigid rotation. Indeed the solution look rather similar to the solutions presented in Fendt et al. (1995). The second solution is for $h=0.5$ (Fig.5, [*right*]{}). The steeper profile for the rotation law implies a smaller asymptotic jet radius (Fendt 1997b). This can be seen in comparison with the rigid rotation solutions (Fendt et al. 1995). However, a larger poloidal electric current can balance the effect of differential rotation. Therefore, the $h=0.2$ solution (with $g=2.5$) collimates to a smaller asymptotic jet radius than the $h=0.5$ solution (with $g=2.0$). A $h=0.2$ solution with $g=2.0$ would have a jet radius of $2.4$. The second solution with the steeper profile of the rotation law $\omf(\Psi)$ would better fit to a Keplerian disk rotation. A perfect match would require an even steeper $\omf(\Psi)$-profile (see below). The mean half opening angle of the jet magnetospheres is about $60\degr$. As discussed above, the shape of the outermost flux surface ($\Psi = 1$) is [*not*]{} prescribed but is a result of our calculation eventually determined by the regularity condition. After crossing the light surface the jets collimate to their asymptotic radius within a distance from the source of about $1 - 2\,\ro$ along the jet axis. The opening angle of the second solution is smaller, however, the jet collimation is achieved only at a larger distance from the central source. In our examples, the “jet expansion rate”, which we define as the ratio of the asymptotic jet radius to the foot point jet radius (the “disk radius”), is about 10. The true scaling of the jet magnetosphere in terms of the size of the central object can be determined by connecting the jet iso-rotation parameter $\omf(\Psi)$ to the disk rotation (see next Sect.). We note that, although in our computations the jet boundary $x_{\rm jet}(z)$ is determined by the force-balance within the jet, and therefore subject to the regularity condition, with our results we do not prove the magnetohydrodynamic [*self-collimation process*]{} of the jet flow. Clearly, the calculated jet magnetosphere is self-collimated in the sense that its structure has been determined only by the internal properties. However, the actual collimation process of the jet flow from an un-collimated conical outflow into a cylinder could only be investigated by time-dependent simulations taking into account the interaction with the ambient medium. On the other hand, we can assume that our jet solution is embedded in an ambient medium. If we further assume an equilibrium between the internal pressure (magnetically dominated) and external (gas) pressure along the jet boundary, we may derive the gas pressure distribution in the ambient medium, since we know the magnetic pressure distribution along the collimating jet radius. In this case, the jet solution may be considered as collimated by ambient pressure. To our understanding one may claim a self-collimation only, if the jet flow collimates [*independently*]{} from external forces. Since in our treatment we do not consider the interrelation with the medium outside the jet, we cannot decide whether the flow is self-collimated or pressure collimated. The field structure is governed by the choice of the functions $I(\Psi)$ and $\omf(\Psi)$, here taken from an asymptotic jet solution. In combination with the disk magnetic flux distribution (13) we can determine two parameters interesting for the jet-disk interaction. These are (i) the magnetic angular momentum loss per unit time and unit radius from disk into the jet and (ii) the toroidal magnetic field distribution along the disk. With $I(\Psi)$ as the angular momentum flux per unit time per unit flux tube, the (normalized) angular momentum flux per unit time per unit radius is $d\dot{J}/dx = - x B_{\rm z} I(x)$ along the disk. Fig.6 shows the behavior of both quantities for our jet model with the steeper profile of the rotation law, $h=0.5$. As we see, most of the magnetic angular momentum is lost in the outer parts of the disk. This may have interesting applications for accretion disk models taking into account a magnetized wind as a boundary condition. The total magnetic angular momentum loss is determined by the normalization, $\dot{J} = - (I_{\rm max} \Psi_{\rm max}/c ) \int I(\Psi) d\Psi $ or $\dot{J} = - (\sqrt{g}\Psi_{\rm max}/2\ro ) \int I(\Psi) d\Psi $. The magnetic toroidal field distribution along the disk has a maximum at about half the disk radius. Clearly, these parameters are biased by the magnetic flux disk boundary condition (13) of our model. However, we believe that the main features are rather general and valid for any poloidal current and magnetic flux distribution with the typical core-envelope structure. (5,50) (5,0) Scaling relations of disk and jet --------------------------------- As discussed above, the two-dimensional magnetic field distribution connecting the asymptotic jet region with the lower disk boundary allows for a direct scaling of the jet in terms of the size of the central object. This is simply based on the assumption that the foot points of the field lines are rotating with Keplerian speed, $\omf = \omk$ and to the fact that in ideal MHD the iso-rotation parameter $\omf$ is conserved along the field lines. It is therefore possible to construct a self-consistent model of the whole “star”-disk-jet system with only a small set of free parameters. In the following we will motivate such a model. The first example demonstrates how the connection between the asymptotic jet and the disk, applied for our very special model assumption, provides a specific estimate for the asymptotic light cylinder $\ro$. Normalizing the Keplerian velocity $\omk$ in the same way as $\omf$ (Sect.2.1), we obtain the expression $$\ro = \frac{GM}{c^2 \omk^2}\frac{1}{x^3} = \frac{GM/c^2}{\omf^2(\Psi=1) x_{\rm disk}^3} = \frac{0.5 R_{\rm S}}{\omf^2(\Psi=1) x_{\rm disk}^3}.$$ Iso-rotation parameter $\omf(\Psi)$ and disk radius $x_{\rm disk}$ are fixed by our model. Therefore, the asymptotic light cylinder is proportional to the mass of the central object. For $\omf^2(1)=0.54$ (which refers to the $h=0.5$ model) and $x_{\rm disk}=0.2$ the asymptotic light cylinder is $\ro = 116 R_{\rm S}$, which is about 2 times larger compared to the jet solution with a rigid rotation $\omf \equiv 1$ and will increase for larger values of $h$. With the choice of $g$, the value of $\ro$ constraints the maximum poloidal magnetic flux and electric current. Here, no assumption is made about the flux distribution along the disk. In the second example we determine the disk magnetic flux distribution $\Psi(x)$ combining the asymptotic jet rotation law $\omf(\Psi)$ with a Keplerian disk rotation $\omk(x)$. From Eq.(14) follows that $\omf(\Psi)/\omf(1) = \omk(x)/\omf(\Psi=1)= (x/x_{\rm disk})^{-3/2}$. In combination with the numerically derived $\omf(\Psi)$ this gives the $\Psi(x)$ along the disk (Fig.7). The figure shows that the disk flux distribution derived from the asymptotic jet is distributed only over the outer part of the disk. This can be interpreted in two ways. First it may imply a relatively large inner disk radius and, hence, an asymptotic jet radius small in terms of radii of the central object. Secondly, it just underlines the fact that in our model the distribution of the asymptotic jet iso-rotation parameter is too flat in order to be truly connected to a disk magnetic flux with an extended radial distribution. For a model taking into account the disk Keplerian rotation in a fully self-consistent way, the magnetic flux distribution which has to be used as disk boundary condition for the GS solution is the one derived in Fig.7. On the other hand, the assumption of a Keplerian disk rotation together with a certain disk magnetic flux distribution provides an expression for the iso-rotation parameter $\omf(\Psi) = \omk(x(\Psi)) = (GM/\ro c^2) \left( {\tilde{a}}^{2}\left(e^{\tilde{b}\Psi}-1\right)\right)^{-3/2}$. Here, the disk magnetic flux distribution (13) has been used. Eventually, one finds $$\frac{\omk(\Psi)}{\omf(1)} = \left(\frac{x_{\rm disk}}{\tilde{a}}\right)^{3/2} \left(\left(1 + \left(\frac{x_{\rm disk}}{\tilde{a}}\right)^2\right)^{\Psi} -1\right)^{-3/4}.$$ This function is definitely steeper compared to the $\omf(\Psi)$-distributions which have been derived in Fendt (1997b) and are used in the present paper. Here, we see the limitation of our ansatz. A steeper profile for rotation law is not yet possible to treat with our code due to the lack of numerical resolution. The non-linear character of the GS equation becomes more problematic due to the gradients in the $\omf$-source term. In summary, only a model including differential rotation $\omf(\Psi)$ may provide a connection between the asymptotic jet, the disk magnetic flux distribution and also the size of the central object. With our model we have presented a reasonable first solution for a self-consistent treatment. (0,0) Application to the M87 jet -------------------------- The jet of M87 shows superluminal motion clearly indicating a highly relativistic jet velocity (Biretta et al. 1999). Recent radio observations have been able to resolve the innermost region of the M87 jet formation region with $0.33 \times 0.12$ mas beam resolution (Junor et al. 1999), corresponding to $2.5 - 7.0 \times10^{16}\cm$. Assuming a central supermassive black hole of $3\times10^9\msun$ (Ford et al. 1994), this is equivalent to about $30\,\rs$! The derived jet full opening angle is $60\degr$ up to a distance of 0.04 pc from the source with a “strong collimation” occurring afterwards (Junor et al. 1999). We now apply our two-dimensional jet model to these observations and compare the geometrical scales. Such a comparison is not possible for e.g. self-similar models. From the observed radio profile resolving the inner M87 jet (see Fig.1 in Junor et al. 1999), we deduce a jet radius of about 120 Schwarzschild radii. With this, the first important conclusion is that the ratio of jet radius to light cylinder radius must be definitely less than the value of 100 which is usually assumed in the literature. A number value of 3-10 seems to be much more likely. Numerical models of two-dimensional general relativistic magnetic jets fitting in this picture were calculated by Fendt (1997a). These solutions, however, do not take into account the differential rotation $\omf(\Psi)$. Junor et al. (1999) claim that the M87 jet radius in the region “where the jet is first formed cannot be larger than” their resolution of $30\,\rs$. Our conclusion is that the expansion rate is limited in both directions. The new radio observations give a minimum value of 3. Theoretical arguments limit the expansion rate to the value of about 20, since the jet mass flow must originate outside the marginally stable orbit which is located at $3-6 \rs$. Clearly, if the jet radius is really as small as observed in M87, general relativistic effects may vary the field structure in the jet formation region. From our model solutions, we derive a light cylinder radius of the M87 jet of about $50\,\rs$. The value derived from Eq.(14) differs from that by a factor of two, but is biased by the unknown size of the disk radius $x_{\rm disk}$. This parameter, however, does not affect the global solution. Considering the standard relativistic MHD theory, nothing special is happening at the light cylinder. For a highly magnetized plasma wind the light surface corresponds to the usual Alfvén surface which itself does not affect the flow of matter. Hence, the light cylinder is un-observable. Also the opening angle in our numerical solution is larger than the observed value by a factor of two. This cannot be due to projection effects since any inclination between jet axis and the line of sight will increase the observed opening angle. We hypothesize that a numerical model with a steeper profile for the iso-rotation parameter will give a smaller jet opening angle comparable to the observed data. This is not surprising, since the jet footpoint anchored in a Keplerian disk rotates faster than in our model. Nevertheless, comparing the collimation distance observed in the M87 jet and assuming a similar ratio of jet radius to light cylinder radius as in our model with $h=0.5$, we find good agreement. The collimation distance is $2\ro$. In summary, we conclude that the example of the M87 jet gives clear indication that the light cylinder of AGN jets might not be as large as previously thought. Although our model does not fit the observed geometrical properties of the inner M87 jet perfectly, we find in general a close compatibility. Conclusions =========== We have investigated the two-dimensional magnetic field distribution in collimating, relativistic jets. The structure of the axisymmetric magnetic flux surfaces is calculated by solving the relativistic force-free Grad-Shafranov equation numerically. In relativistic MHD, [*electric fields*]{} become important in difference to Newtonian MHD. The simplifying assumption of the force-free limit has been applied as relativistic jets must be highly magnetized. The central point of our paper is the consideration of differential rotation of the foot points of the field lines, i.e. a variation of the iso-rotation parameter $\omf(\Psi)$. The underlying model is that of a magnetic jet anchored in an accretion disk. Two main problems had to be solved in order to calculate a two-dimensional field distribution: a) to determine the [*a priori*]{} unknown location of the light surface, b) the proper treatment of the regularity condition along that light surface. The light surface is the force-free equivalent of the Alfvén surface and provides a singularity in the Grad-Shafranov equation. We summarize our results as follows. \(1) We find numerical solutions for the two-dimensional magnetic flux distribution connecting the asymptotic cylindrical jet with a differentially rotating disk. In our example solutions the asymptotic jet radius is about 2.5 times the asymptotic light cylinder radii. This is the first truly two-dimensional relativistic solution for a jet magnetosphere including differential rotation of the iso-rotation parameter $\omf(\Psi)$. The physical solution, being characterized by a smooth transition across the light surface, is unique for a certain parameter choice for the rotation law $\omf$. \(2) The half opening angle of the numerical jet solution is about 60 degrees. Cylindrical collimation is achieved already after a distance of 1-2 asymptotic light cylinder radii along the jet axis. Differential rotation decreases the jet opening angle, but increases the distance from the jet origin where collimation is achieved. The “jet expansion rate”, defined as the ratio of the asymptotic jet radius to the jet radius at the jet origin, is about 10. \(3) From the analytical treatment of the regularity condition along the asymptotic branches of the light surface we derive a general estimate for the jet opening angle. We find that the jet half opening angle is larger than $45\degr$ and increases for a steeper profile of the differential rotation $\omf$. \(4) Our two-dimensional ansatz, in combination with the treatment of differential rotation, allows for a connection of the asymptotic jet solution with the accretion disk. Certain disk properties can be deduced from the asymptotic jet parameters. Examples are the disk toroidal magnetic field distribution, with a maximum at half of the disk radius and the angular momentum flux per unit time and unit radius. This is interesting as a boundary condition for accretion disk models. We find that most of the angular momentum is lost in the outer part of the disk. \(5) Application of our model to the M87 jet gives good agreement qualitatively. From our numerical solution we derive an asymptotic light cylinder of the M87 jet of about 50 Schwarzschild radii. Collimation of the jet would be achieved after a distance of two asymptotic light cylinder radii from the source. This value is comparable with the observations, however, the opening angle in our model is larger by a factor of two. In summary, we have presented the first global two-dimensional solutions for a relativistic jet magnetosphere taking into account differential rotation of the jet footpoints. From our [*jet*]{} model we may determine certain physical quantities in the [*disk*]{} that are not possible to observe, as e.g. the angular momentum flux distribution at the disk-jet interface. Comparison with the M87 jet shows that our model seems to be consistent with the observations, therefore allowing for a derivation of the collimation distance, the light cylinder radius and the jet expansion rate for that example. Clearly, such features as the time-dependent ejection of knots and the interaction process between disk, jet and central source cannot be answered by our approach. Time-dependent relativistic MHD simulations of the whole “star”-disk-jet system would be necessary, however, such codes are not yet fully developed. Part of this work was initiated when C.F. was holding a guest stipend of the Sonderforschungsbereich (SFB) 328 of the University of Heidelberg. E.M. acknowledges a grant (FE490/1-1) from the Deutsche Forschungsgemeinschaft (DFG). numerical methods ================= For the solution of the two-dimensional GS equation we apply the method of finite elements as developed by Camenzind (1987) and Fendt et al. (1995). Differential rotation $\omf(\Psi)$ implies two major complications for the numerical computation. The first one is the fact that position and shape of the light surface $D=0$ is not known [*a priori*]{}. Along the light surface the boundary condition is the regularity condition, which, however, itself depends on the two-dimensional solution $\Psi(x,z)$. The second problem is the GS source term for the differential rotation, containing the gradient of the magnetic flux, $|\nabla\Psi|^2$. Compared to the case of rigid rotation, this introduces another (and stronger) non-linearity in the GS equation. Therefore, a fragile numerical convergence process can be expected. An additional complication is that our grid of finite elements of second order may be inadequate for a calculation of monotonous gradients between the elements if the numerical resolution is too low. However, for appropriate numerical parameters as grid size, element size and iteration step size, we were finally able to overcome these difficulties. Determination of the light surface ---------------------------------- Here we discuss the iteration procedure we use to determine the location of the light surface. Because the rotation law $\omf(\Psi)$ is prescribed, the radius where the light surface, $D=0$, intersects the jet boundary, $\Psi = 1$, is known, $$\xl(\Psi=1) = 1/\omf(\Psi=1).$$ However, the corresponding position in $z$-direction is not known. Some estimates can be made about shape and inclination of the light surface in the limit of large radii (see Sect.2.4), but a general solution is not yet known. We start the iteration procedure calculating the inner solution (defined as the field distribution inside the light surface) with an outer grid boundary at $x=1$ (for comparison see Fig.1). This choice is equivalent to the light cylinder in the case of rigid rotation. For differential rotation the radius $x=1$ is defined as asymptotic light cylinder (for large $z$). For low $z$-values the boundary $x=1$ is located inside the light surface $\xl(\Psi) = 1/\omf(\Psi)$. Along this outer boundary (of the [*inner*]{} solution), we apply a homogeneous Neumann boundary condition. Usually, this implies that the field lines will cross that boundary perpendicularly. However, in our case the homogeneous Neumann boundary condition transforms into the regularity condition [*if*]{} the boundary becomes equivalent to the singular light surface. As shown in Fendt et al. (1995), this transformation applies “automatically” in our finite element code. This is due to the facts that (i) finite element code solves the [*integrated*]{} GS equation and (ii) the boundary integral, which is proportional to $D=1-x^2\omf^2$, vanishes along the light surface. With the GS solution of the first iteration step we estimate the deviation of the chosen outer boundary from the true light surface by calculating $D=1-x^2\omf^2(\Psi)$. For the lowest $z$-value prescribed, we know that $D=1-x^2\omf^2(\Psi=1)$. Then, the outer grid boundary $(x,z)$ is slowly moved to a larger radius with $\Delta x \sim D(x,z)^2$. As a consequence of the different numerical grid, the field distribution will change. The value of $D$ will, however, decrease. This procedure is repeated until $D$ is below a certain limit, $D\simeq0$. Having obtained the solution inside the light surface, that field distribution is taken as inner boundary condition for the outer solution. Appl S., Camenzind M., 1993, A&A, 274, 699 Beskin V.S., 1997, Phys. Uspekhi, 40, 659 Biretta J.A., Sparkes W.B., Maccetto F., 1999, ApJ, 520, 621 Blandford R.D., Payne D.G., 1982, MNRAS, 199, 883 Camenzind M., 1987, A&A, 184, 341 Camenzind M., 1990, RvMA, 3, 234 Chiueh T., Li Z.-Y., Begelman M.C., 1991, ApJ, 377, 462 Contopoulos J., Lovelace R.V.E., 1994, ApJ, 429, 139 Contopoulos J., 1994, ApJ, 432, 508 Contopoulos J., 1995, ApJ, 446, 67 Fendt C., Camenzind M., Appl S., 1995, A&A, 300, 791 Fendt C., Camenzind M., 1996, A&A, 313, 591 Fendt C., 1997a, A&A, 319, 1025 Fendt C., 1997b, A&A, 323, 999 Ferreira J., 1997, A&A, 319, 340 Ford H.C., et al., 1994, ApJ, 435, L27 Junor W., Biretta J.A., Livio M., 1999, Nature, 401, 891 Khanna R., Camenzind M., 1992, A&A, 263, 401 Li Z.-Y., 1993, ApJ, 415, 118 Lovelace R.V.E., Berk H.L., Contopoulos J., 1991, ApJ, 379, 696 Mirabel I.F., Rodriguez L.F., 1995, Superluminal motions in our Galaxy. In: Böhringer H., Morfill G.E., Trümper J.E. (eds) 17th Texas Symposium on Relativistic Astrophysics and Cosmology. Annals of the New York Academy of Sciences, New York, p.21 Mundt R., Bührke T., Solf J., Ray T.P., Raga A.C., 1990, A&A, 232, 37 Nitta S.-Y., 1994, PASJ, 46, 217 Nitta S.-Y., 1995, MNRAS, 276, 825 Okamoto I., 1992, MNRAS, 254, 192 Okamoto I., 1999, MNRAS, 307, 253 Ray T.P., Mundt R., Dyson J.E., Falle S.A.E.G., Raga A.C., 1996, ApJ, 468, L103 Pudritz R.E., Norman C.A., 1983, ApJ, 274, 677 Shibata K., Uchida Y., 1985, PASJ, 37, 31 Sakurai T., 1985, A&A, 152, 121 Zensus J.A., Cohen M.H., Unwin S.C., 1995, ApJ, 443, 35 [^1]: Due to the fact that the jet radius (where $\Psi=1$) is not known before the asymptotic GS equation has been solved (Fendt 1997b), the normalization with $g$ leads to a current distribution $I(\Psi)$ which is [*not*]{} normalized to unity. This difference in normalization is “hidden” in the coupling constant $g$, which could, in principal, be re-scaled appropriately. [^2]: For figures of these functions and the related $\Psi(x)$, $\omf(\Psi)$ and $I(\Psi)$ distribution, we refer to Fendt (1997b).

--- abstract: 'A tight far-infrared - radio correlation similar to that in star-forming late-type galaxies is also indicated in star-forming blue early-type galaxies, in which the nuclear optical-line emissions are primarily due to star-forming activities without a significant contribution from active galactic nuclei. The average value of far-infrared to 1.4 GHz radio flux-ratio commonly represented as the $''q''$ parameter is estimated to be $2.35\pm0.05$ with a scatter of 0.16 dex. The average star formation rate estimated using 1.4 GHz radio continuum is $\sim6$ M$_{\odot}$ yr$^{-1}$ in good agreement with those estimated using far-infrared and H$\alpha$ luminosities. The radio emission is detected mainly from central region which could be associated with the star-forming activities, most likely triggered by recent tidal interactions. The average thermal contribution to total radio flux is estimated to be $\sim12$ per cent. The average value of the magnetic field strengths in the star-forming early-type galaxies is estimated to be 12$^{+11}_{-4}$ $\mu$G. These magnetic fields are very likely generated via fast amplification in small-scale turbulent dynamos acting in the star-bursting regions.' author: - | A. Omar,$^{1}$[^1] A. Paswan$^{1,2}$\ $^{1}$Aryabhatta Research Institute of observational sciences, Manora Peak, Nainital, India\ $^{2}$Pt. Ravishankar Shukla University, Raipur, India date: 'Accepted XXX. Received YYY; in original form ZZZ' title: 'Far-infrared - radio correlation and magnetic field strength in star-forming early-type galaxies' --- \[firstpage\] Galaxies: magnetic field – Galaxies: radio continuum – Galaxies: early-type – Galaxies: star formation Introduction ============ A tight correlation between radio-continuum (RC) emission and infrared (IR) emission from the nuclei of late-type star-forming (SF) galaxies is known since the early studies of van der Kruit (1971, 1973). Later, a far-infrared - radio (FIR-RC) correlation was confirmed in the disks of different types of galaxies such as optical and infrared bright galaxies (Condon, Anderson & Helou 1991; Yun, Reddy & Condon 2001; hereafter Y01), dwarf irregular galaxies (e.g., Roychowdhury & Chengalur 2012), Seyferts and low-ionization nuclear emission regions - LINERs (e.g., Morić et al. 2010), and Wolf-Rayet galaxies with nascent star-formation (e.g., Jaiswal & Omar 2016). A large number of studies on the FIR-RC correlations based on observations and theoretical modeling have been carried out. Almost all studies on SF galaxies indicate a tight FIR-RC correlation. Most of the FIR-RC correlations are constructed using total radio flux at cm-wavelengths, where galaxies usually have significant thermal radio emission (free-free) mixed with the dominating non-thermal (synchrotron) radio emission. Tight FIR-RC correlations have also been seen at meter-wavelengths (Basu et al. 2012) where the thermal radio emission is negligible, and also in faint dwarf irregular galaxies where the thermal radio emission can be almost equal to the non-thermal radio component (Roychowdhury & Chengalur 2012). The FIR-RC correlation and its tightness are one of the unsolved mysteries in astronomy. The tight FIR-RC correlation is most often linked to the star-forming activities in galaxies. Harwit & Pacini (1975) proposed that both RC and FIR emission may be linked to massive stars in SF galaxies. According to a widely discussed scenario (Völk 1989; Xu 1990), ultra-violet photons from the massive ($M\ge$8 M$_\odot$) supernovae progenitor stars are considered as the main source for heating of dust in the inter-stellar medium (ISM), which in turn emits in IR wavebands. The same massive star population, which is also short-lived ($\le$10 Myr), undergo type-II supernova explosions in which the cosmic-ray (CR) particles are accelerated to the relativistic energies. These high-energy CR particles loose energy mainly via the synchrotron process in presence of galactic magnetic fields. This scenario requires a fine balance to be maintained between the complex processes of generating CR particles, escaping of CR electrons, synchrotron cooling time, magnetic field strengths and heating of dust so that a universal FIR-RC correlation can be obtained. Such a requirement appears almost like a ’conspiracy’ and explanations for such a balance in galaxies with widely varying luminosities over several orders of magnitude are challenging. Nevertheless, several explanations highlighting the possibilities of inter-connections between star-formation, dust heating, cosmic-rays, and magnetic field to understand this apparent conspiracy have been given (Helou & Bicay 1993; Lisenfeld, Völk & Xu 1996; Nicklas & Beck 1997; Thompson et al. 2006; Lacki, Thompson & Quataert 2010). The FIR-RC correlation is also explained via amplification of random turbulent magnetic fields through supernovae explosions in the ISM (Schleicher & Beck 2013). Some alternate explanations such as the CR heating of molecular clouds and regulation of FIR and radio emission by hydrostatic pressure have also been discussed (Suchkov, Allen & Heckman 1993; Murgia et al. 2005). The powerful active galactic nuclei (AGN) also display a FIR-RC correlation, however, the relations in AGNs are different and not very tight (Bally & Thronson 1989; Condon et al. 1991; Wrobel & Heeschen 1991; Colina & Pérez-Olea 1995; Miller & Owen 2001). The tightness of the FIR-RC correlation in the SF galaxies has often been used to distinguish the AGN hosting SF galaxies from the SF-dominant galaxies (e.g., Y01; Omar & Dwarakanath 2005). Studies on the FIR-RC correlation in SF early-type galaxies (ETG; lenticulars and ellipticals) are rare although a presence of ISM and star-formation in a large number of ETGs is known since the detection of dust in ETGs via the far-infrared ($25\mu$m$-100\mu$m) emission from the Infrared Astronomical Satellite ($IRAS$; Neugebauer et al. 1984; Jura 1986). The prime reason for this missing study in ETGs is the lack of strong star formation and a high possibility for the presence of AGN in these galaxies. Therefore, even if low-level SF activities were previously detected in a few ETGs, it was difficult to separate emissions from the ISM with that related to an AGN (Bally & Thronson 1989; Walsh et al. 1989; Wrobel & Heeschen 1991). More recently, there are some studies on the FIR-RC correlation in nearby ETGs (Combes et al. 2007; Crocker et al. 2011; Nyland et al. 2017). These studies reported larger scatter in the correlation in ETGs as compared to that in the late-type galaxies. Nyland et al. (2017) reported FIR-excess in many ETGs, selected from a sample of nearby ETGs in ATLAS$^{\rm{3D}}$ project (Cappellari et al. 2011). These ETGs have low star formation rates (SFR) with a median value at nearly $\sim0.15$ M$_{\odot}$ yr$^{-1}$ (Davis et al. 2014). A study of the FIR-RC correlation in five ETGs with high SFR ($1 - 5$ M$_{\odot}$ yr$^{-1}$) by Lucero & Young (2007) indicated a normal relation similar to that known in the late-type galaxies. Recently, Paswan & Omar (2016) found three ETGs with radio power $L_{\rm 1.4GHz}>10^{21}$ W Hz$^{-1}$ having a normal FIR-RC relation. Since a cool ’cirrus’ component related to low-mass/evolved stars is known to contribute to the total IR emission in galaxies (Helou 1986), the FIR emission from galaxies with very low SFR is not expected to be tightly linked to the diffuse radio emission, related to massive star formation. The cirrus component can also introduce a non-linearity in the FIR-RC correlation towards low FIR luminosity (Y01). As there is already a hint that the FIR-RC correlation in ETGs with high SFR may be similar to that in the late-type galaxies, such studies need to be expanded to a larger sample of galaxies. The synchrotron radio detection from the SF galaxies can also be used to infer galactic magnetic fields (Beck & Krause 2005). All star-forming late-type galaxies without any exception are detected with synchrotron radio emission (Condon 1992). This detection establishes magnetic field in almost all late-type SF galaxies. The strengths of the magnetic fields derived using the synchrotron radio emission range from a few $\mu$G in galaxies with low SFR up to 100 $\mu$G in extreme star-burst galaxies (see Beck et al. 1996; Beck 2016). The typical strengths are around 10 $\mu$G in normal SF spiral galaxies (e.g., Fitt & Alexander 1993). Due to lack of strong star formation, studies of magnetic fields based on synchrotron radio detection are rare in ETGs. Nyland et al. (2017) made an attempt to constrain magnetic fields in ETGs, however, their sample has low SFR and could also be significantly contaminated by weak AGNs. The best examples close to the early-type morphologies are two early-type disk galaxies NGC 4736 (SAab) (Chyży & Buta 2008) and NGC 4594 (SAa) (Krause, Wielebinski & Dumke 2006), in which both regular and random magnetic fields have been detected. In this paper, we present a study based on the FIR and RC emissions from a sample of ETGs with high SFR. The galaxies are selected from the Sloan Digital Sky Survey (SDSS) based galaxy-zoo project in which a unique population of nearby ($z\sim0.02 - 0.05$) blue star-forming ETGs without AGN was found (Schawinski et al. 2009; hereafter S09). The average SFR is inferred as $\sim5$ M$_{\odot}$ yr$^{-1}$ from an analysis of the H$\alpha$ emission from these ETGs. The nature of the optical emission-lines from the central region was inferred as due to star formation, based on the SDSS optical spectroscopy data. The majority of these galaxies are bulge-dominated ETGs with Sérsic index in the range 2 to 8 (George 2017). The low redshift of these ETGs made it possible to search for RC emission in NRAO (National Radio Astronomy Observatory) VLA (Very Large Array) Sky Survey (NVSS; Condon et al. 1998) and Faint Images of the Radio Sky at Twenty-centimeters (FIRST; Becker, White & Helfand 1995), and the FIR emission in the $IRAS$ data-base. The estimates for the RC-based SFR and magnetic field are also presented in this paper. The value of the Hubble constant is taken at 68 km s$^{-1}$ Mpc$^{-1}$ (Planck Collaboration 2016). Physical properties of the sample ================================= A population of blue ETGs with more than 200 galaxies was detected in the VI data release (Adelman-McCarthy et al. 2008) of the SDSS in the redshift range of $0.02 < z < 0.05$ (S09). The SDSS provides broad-band five-color photometry and optical spectroscopy within a $3''$ region in the visible wavelength range for about one-third of the sky using a 2.5-meter optical telescope (Gunn et al. 2006). S09 presented a comprehensive analysis in terms of color, classifications based on the optical-line emissions, star-formation rates based on the H$\alpha$ emission, masses, and environments of these galaxies. A summary of the properties of these galaxies based on the analysis presented in S09 is given here. The $u - r$ optical colors of the blue ETGs are in the range of $1.3 - 2.5$ in comparison to the redder colors ($2.5 - 3.0$) of quiescent ETGs. The nature of the emission from the central ($\sim3''$) region of these galaxies was classified based on two well-established SF-AGN diagnostics schemes (Baldwin, Phillips & Terlevich 1981; Kauffmann et al. 2003) using the optical-line ratios of \[NII\] $\lambda$6583/H$\alpha$ $\lambda$6563 and \[OIII\] $\lambda$5007/H$\beta$ $\lambda$4861. Using these schemes, 55 blue ETGs were identified as star-forming without a significant contribution from AGN. A larger number of 73 blue ETGs were found to have an AGN or a mix of AGN and SF. The identified AGNs were either Seyferts or Low Ionization Nuclear Emission Regions (LINER). The remaining 76 galaxies were termed as weak emission-line (WEL) galaxies in which it was not possible to discern AGN and SF-related emissions due to poor signal-to-noise ratio (SNR) in the optical spectrum. These 55 blue star-forming galaxies without a significant contribution from AGN constitute our sample for the present study. S09 also found that the SF ETGs have central stellar velocity dispersions ($\sigma$) between 40 and 150 km s$^{-1}$ and such blue ETGs are nearly absent at $\sigma$ above 200 km s$^{-1}$. The blue ETGs in this sample are of intermediate stellar masses (a few times $10^{10}$ M$_{\odot}$) without any massive red-sequence early-type system. These blue ETGs are rare and make up nearly 6 per cent of the low-redshift ETG population. These blue ETGs are residing in low galaxy-density environments. The difference in the local density was also visible as a function of color in the sense that the bluer galaxies are found in lower density environments compared to their red counterparts at the same velocity dispersion (S09). These blue ETGs are not residing in the central regions of cluster of galaxies. These properties make it a unique sample of star-forming ETGs, which was not available before. The SFRs were estimated in the range of $0.5 - 21$ M$_{\odot}$ yr$^{-1}$ from the H$\alpha$ line-flux within the central $3''$ region extrapolated to the optical extents of the galaxies (S09). The bluer early-type galaxies in this sample tend to show higher SFR on average (Paswan & Omar 2016). They also showed a hint that the SFR is linked to the central stellar velocity dispersion in the sense that the SFR decreases in galaxies with $\sigma$ beyond $\sim100$ km s$^{-1}$. This trend in SFR with $\sigma$ is consistent with some semi-analytical models in which less-massive galaxies with lower $\sigma$ within 80 - 240 km s$^{-1}$ are predicted to experience star formation (e.g., Schawinski, Khochfar & Kaviraj 2006). Therefore, it appears that the star-forming ETGs in the S09 sample are having unique environmental and physical properties which are allowing these galaxies to harbor intense star-forming activities. The morphological designation of these galaxies as presented in S09 was somewhat uncertain as it was based on visual classifications. Most of the galaxies were classified as E/S0. Recently, George (2017) presented structural analysis of the SF galaxies in the S09 sample through Sérsic profile fitting to the optical surface brightness. This analysis provides a quantitative parameter Sérsic index, which can be better than qualitative visual classification of galaxy morphology. In our radio-detected sample, thirty-two galaxies showed Sérsic index in the range $2 - 8$ indicating bulge dominance in these systems, and 6 galaxies have the index $<2$ indicating a possibility for the presence of a disc component. The analysis presented by George (2017) also showed that after subtracting a best-fit Sérsic profile, the residual images of several galaxies show significant deviations from a smooth light distribution. These deviations are identified as signatures of tidal debris in form of tidal tails, shells and asymmetric excess-light. The residual images also show circum-nuclear ring-like structure. These galaxies also follow the tight Kormendy scaling-relation (see Kormendy 1977) between the effective radius and the mean surface brightness at the effective radius. The Kormendy relation is followed by early-type galaxies. This analysis confirms these galaxies as early-types without an ambiguity. [llcccccc]{} S09 & Galaxy & Redshift & $S^{\rm FIRST}_{1.4 \mathrm{GHz}}$ & $S^{\rm NVSS}_{1.4 \mathrm{GHz}}$ & $S_{60\mu \mathrm{m}}$ & $S_{100\mu \mathrm{m}}$ & Size$^{a}$\ id & & $z$ & \[mJy\] & \[mJy\] & \[Jy\] & \[Jy\]& \[arcsec\]\ 8 &J010358+151450 & 0.04176 & x & $2.3\pm0.4$ & $0.20\pm0.07$ & $<1.3$ & 32\ 7 &J014143+134032 & 0.04539 & x & $3.1\pm0.4$ & $0.74\pm0.06$ & $0.94\pm0.18$ & 25\ 68 &J030126-000425 & 0.04285 & $1.02\pm0.12$ & $<1.5$ & $<0.1$ & $<0.8$ & 28\ 108 &J074723+222041 & 0.04549 & $<0.5$ & $<1.5$ & $<0.2$ & $<0.9$ & 22\ 84 &J075420+255133 & 0.04167 & $<0.5$ & $<1.5$ & $0.20\pm0.05$ & $<1.2$ & 25\ 139 &J075608+172250 & 0.02899 & $<0.5$ & $<1.5$ & $0.17\pm0.04$ & $<0.4$ & 37\ 137 &J075636+184417 & 0.03988 & $<0.5$ & $<1.5$ & $<0.24$ & $<0.7$ & 33\ 129 &J075912+533326 & 0.03479 & $2.10\pm0.17$ & $4.8\pm0.5$ & $0.76\pm0.05$ & $0.90\pm0.13$ & 30\ 130 &J081020+561226 & 0.04623 & $<0.5$ & $<1.5$ & $<0.2$ & $<0.8$ & 23\ 172 &J081756+470719 & 0.03901 & $3.30\pm0.15$ & $2.6\pm0.5$ & $0.27\pm0.07$ & $1.46\pm0.25$ & 34\ 207 &J082909+524906 & 0.04842 & $<0.5$ & $<1.5$ & $<0.1$ & $<0.5$ & 22\ 213 &J084346+313452 & 0.04756 & $0.73\pm0.13$ & $<1.5$ & $0.25\pm0.04$ & $0.49\pm0.17$ & 16\ 86 &J085311+370806 & 0.04980 & $2.40\pm0.13$ & $4.2\pm0.4$ & $0.28\pm0.04$ & $1.30\pm0.20$ & 23\ 175 &J092429+534137 & 0.04590 & $<0.5$ & $<1.5$ & $0.10\pm0.02$ & $0.40\pm0.11$ & 20\ 41 &J101628+033502 & 0.04848 & $<0.5$ & $<1.5$ & $0.13\pm0.04$ & $<0.4$ & 33\ 148 &J102034+291410 & 0.04846 & $<0.5$ & $<1.5$ & $0.11\pm0.04$ & $<0.5$ & 20\ 160 &J102524+272506 & 0.04973 & $<0.5$ & $<1.5$ & $<0.1$ & $<0.5$ & 26\ 215 &J102654-003229 & 0.03463 & $2.80\pm0.14$ & $2.7\pm0.4$ & $0.65\pm0.07$ & $2.32\pm0.41$ & 29\ 66 &J105437+553946 & 0.04787 & $0.50\pm0.12$ & $<1.5$ & $0.21\pm0.03$ & $0.43\pm0.12$ & 24\ 2 &J112327-004248 & 0.04084 & $0.66\pm0.13$ & $<1.5$ & $0.23\pm0.07$ & $<1.1$ & 19\ 149 &J113122+324222 & 0.03368 & $4.20\pm0.15$ & $4.9\pm0.4$ & x & x & 40\ 192 &J115205+455706 & 0.04316 & $<0.5$ & $<1.5$ & $0.13\pm0.04$ & $<0.5$ & 19\ 180 &J120617+633819 & 0.03974 & $5.40\pm0.17$ & $5.0\pm0.4$ & $1.29\pm0.04$ & $2.13\pm0.17$ & 27\ 23 &J120647+011709 & 0.04124 & $<0.5$ & $<1.5$ & $<0.12$ & $0.50\pm0.10$ & 28\ 5 &J120823+000637 & 0.04081 & $<0.5$ & $<1.5$ & $0.13\pm0.04$ & $0.37\pm0.13$ & 34\ 92 &J122023+085137 & 0.04894 & $<0.5$ & $<1.5$ & $<0.21$ & $<0.9$ & 20\ 77 &J122037+562846 & 0.04381 & $<0.5$ & $<1.5$ & $0.11\pm0.04$ & $0.77\pm0.10$ & 21\ 14 &J123502+662233 & 0.04684 & x & $1.9\pm0.4$ & $0.49\pm0.04$ & $0.76\pm0.12$ & 29\ 177 &J130141+044049 & 0.03836 & $<0.5$ & $<1.5$ & $0.31\pm0.07$ & $<0.7$ & 26\ 147 &J132620+314159 & 0.04999 & $<0.5$ & $<1.5$ & $0.29\pm0.04$ & $1.18\pm0.15$ & 28\ 121 &J134747+111627 & 0.03942 & $2.80\pm0.13$ & $3.2\pm0.5$ & $0.33\pm0.05$ & $0.98\pm0.20$ & 30\ 50 &J135707+051506 & 0.03967 & $2.60\pm0.13$ & $2.6\pm0.3$ & $0.46\pm0.04$ & $0.84\pm0.12$ & 37\ 101 &J140248+523000 & 0.04361 & $<0.5$ & $<1.5$ & $<0.1$ & $0.37\pm0.10$ & 24\ 58 &J140656-013541 & 0.02916 & $4.40\pm0.15$ & $6.2\pm0.5$ & $0.82\pm0.06$ & $1.45\pm0.24$ & 40\ 102 &J140747+523809 & 0.04381 & $<0.5$ & $<1.5$ & $0.18\pm0.03$ & $0.51\pm0.10$ & 25\ 190 &J141433+404522 & 0.04185 & $1.70\pm0.14$ & $2.5\pm0.4$ & $0.37\pm0.04$ & $0.58\pm0.09$ & 26\ 202 &J141732+362019 & 0.04712 & $<0.5$ & $<1.5$ & $0.20\pm0.04$ & $<0.3$ & 21\ 182 &J143222+565108 & 0.04302 & $0.65\pm0.12$ & $3.0\pm0.5$ & $0.36\pm0.03$ & $0.80\pm0.14$ & 30\ 206 &J143733+080443 & 0.04987 & $<0.5$ & $<1.5$ & $0.12\pm0.04$ & $<0.45$ & 34\ 52 &J145115+620014 & 0.04306 & $0.79\pm0.12$ & $2.7\pm0.4$ & $0.36\pm0.02$ & $1.00\pm0.05$ & 26\ 195 &J145323+390413 & 0.03153 & $<0.5$ & $<1.5$ & $0.14\pm0.03$ & $0.40\pm0.08$ & 32\ 30 &J151719+031918 & 0.03749 & $<0.5$ & $<1.5$ & $<0.1$ & $<0.4$ & 25\ 4 &J152347-003823 & 0.03747 & $0.80\pm0.13$ & $<1.5$ & $0.18\pm0.06$ & $<1.2$ & 25\ 151 &J154451+175122 & 0.04521 & $<0.5$ & $<1.5$ & $0.15\pm0.03$ & $<0.5$ & 26\ 103 &J155000+415811 & 0.03391 & $1.25\pm0.14$ & $2.8\pm0.4$ & $0.22\pm0.03$ & $0.65\pm0.09$ & 28\ 119 &J155335+321820 & 0.04985 & $<0.5$ & $<1.5$ & $0.13\pm0.04$ & $<0.2$ & 23\ 157 &J160439+164443 & 0.04599 & $<0.5$ & $<1.5$ & $<0.2$ & $<0.6$ & 25\ 146 &J160754+200303 & 0.03165 & $0.94\pm0.14$ & $1.4\pm0.4$ & $0.31\pm0.04$ & $0.62\pm0.13$ & 29\ 209 &J161818+340640 & 0.04733 & $<0.5$ & $<1.5$ & $<0.1$ & $0.31\pm0.09$ & 30\ 124 &J164430+195626 & 0.02300 & $3.90\pm0.14$ & $4.80\pm0.4$ & $0.69\pm0.05$ & $0.93\pm0.17$ & 35\ 105 &J165116+280652 & 0.04724 & $1.90\pm0.12$ & $2.20\pm0.4$ & $0.19\pm0.03$ & $<2.3$ & 21\ 56 &J172324+274846 & 0.04845 & x & $3.50\pm0.5$ & $0.25\pm0.05$ & $0.87\pm0.23$ & 25\ 61 &J221516-091547 & 0.03843 & $2.90\pm0.13$ & $3.00\pm0.4$ & $0.51\pm0.05$ & $<0.5$ & 41\ \ \ \ \[table:tab1\] Radio and far-infrared properties of the sample =============================================== The radio emission at 1.4 GHz was searched in 55 star-forming ETGs using the NVSS and FIRST images. These images were made using the aperture-synthesis interferometric technique (Ryle & Hewish 1960). The NVSS is a low angular resolution ($45''$ beam) survey with typical rms of $\sim0.45$ mJy beam$^{-1}$ in the images while the FIRST is a relatively higher angular resolution ($5''$ beam) survey with a typical rms of $\sim0.15$ mJy beam$^{-1}$. The radio images were visually examined for possible detections at the optical locations of the galaxies. Four locations have no observations in the FIRST survey and two source locations appeared confused with a strong background radio source in the vicinity. This reduces our sample to 53 galaxies. The rms (noise) variations among the images or within the images were normally within $\pm0.02$ mJy beam$^{-1}$ for the FIRST images, and $\pm0.05$ mJy beam$^{-1}$ for the NVSS images. We detected radio emission in 22 galaxies in the FIRST and 21 galaxies in the NVSS with a total of 17 common detections in FIRST and NVSS. The radio detections are listed in Table \[table:tab1\]. The radio images of two galaxies are presented in Fig. \[figure:fig1\] and the remaining images are provided in the Appendix. The upper limits to the radio flux density for the undetected galaxies are taken as 0.5 mJy and 1.5 mJy in the FIRST and NVSS images respectively. The SNRs in these detections are above 5, except in two cases where the SNR is nearly 3. Although the radio detections reported here do not have very high SNR, the detections are made in two independent surveys and therefore are highly reliable. Among the common detections made in the FIRST and NVSS images, it was found that the radio flux densities of the sources estimated from the two surveys differ slightly. In the interferometric aperture synthesis imaging technique, the sensitivity to detect extended radio emission depends on the source size and the density of measurements at short interferometric baselines. Due to low-resolution ($45''$) and high density of measurements taken at short baselines, the NVSS images are not expected to loose any significant flux provided the radio extents of the sources are similar to the optical extents (up to $40''$) of the galaxies in our sample. Becker et al. (1995) showed that the FIRST images can reliably detect sources up to $\sim30''$ and $\sim9''$ extent for total flux greater than nearly 30 mJy and 3 mJy respectively. The radio flux densities of the sources detected here are less than 6 mJy. Therefore, it is possible that some sources (S09 ID: 129, 86, 58, 182, 52, 103), for which the flux densities estimated from the FIRST images are substantially lower compared to that estimated using the NVSS images, are extended beyond $10''$ extent. The maximum difference is $\sim2.7$ mJy among these sources. We also noticed that some sources (S09 ID: 172, 215, 180) have flux density estimated from the FIRST image slightly higher than that estimated using the NVSS image. The differences in all the cases are less than $\sim0.7$ mJy. When comparing the point-like radio sources in the FIRST and the NVSS images, White et al. (1997) found that the difference in flux density estimated from the two surveys is a function of flux density. According to their analysis, the sources with flux densities above $\sim10$ mJy show little scatter, however, weaker sources ($<5$ mJy) can show difference up to $\pm1$ mJy. This difference arises mainly due to noise, flux estimation errors and deconvolution errors. We do not find any source for which the flux density estimated from the FIRST image was substantially ($>1$ mJy) higher compared to that estimated using the NVSS image. Therefore, the minor differences ($<0.7$ mJy) in the flux densities estimated from the two surveys can be due to estimation errors and not due to any intrinsic variation in the source flux. The far-infrared emissions at 60 $\mu$m and 100 $\mu$m were searched using the online $IRAS$ Scan Processing and Integration (<span style="font-variant:small-caps;">scanpi</span>) tool, which performs 1-dimensional averaging of all the $IRAS$ scans taken for a location in the sky (Helou et al. 1988). The $IRAS$ revised bright galaxy sample is constructed using an earlier version of this tool (Sanders et al. 2003). The <span style="font-variant:small-caps;">scanpi</span> processing tool is known to provide a factor of 2 to 5 gain in the sensitivity over that in the $IRAS$ point source catalogue. The [<span style="font-variant:small-caps;">scanpi</span>]{} has been reliably used in some studies to get FIR fluxes for weak extended sources (e.g., Lisenfeld et al. 2007). The effective resolution of the $IRAS$ data products is $\sim2'$. We extracted the peak flux densities from the noise-weighted mean of all the $IRAS$ scans using the <span style="font-variant:small-caps;">scanpi</span> tool. Typical rms were nearly 0.04 Jy and 0.12 Jy at 60 $\mu$m and 100 $\mu$m respectively. The detections using the <span style="font-variant:small-caps;">scanpi</span> are made here above 0.1 Jy and 0.3 Jy at 60$\mu$m and 100$\mu$m respectively. In comparison, the $IRAS$ faint source catalogue has sources above 0.2 Jy and 1.0 Jy at 60$\mu$m and 100$\mu$m respectively. The detections with 60 $\mu$m and 100 $\mu$m flux densities are listed in Table \[table:tab1\]. The upper limits to the flux for the undetected galaxies are taken as 2.5 times the local rms in the <span style="font-variant:small-caps;">scanpi</span> co-added scan. Analysis and Results ==================== Far-infrared radio correlation ------------------------------ The flux densities at the FIR and radio bands were converted to luminosities using the standard relations A1 and A2 given in the Appendix A. Out of 53 galaxies listed in Table \[table:tab1\], a total of 32 galaxies with Sérsic index $\ge2$ and 6 galaxies with Sérsic index $<2$ have radio detections, for which the RC and FIR luminosities are listed in Table \[table:tab2\]. Since the galaxies with Sérsic index $<2$ are likely to have a disk, we do not consider such galaxies as early-type. The SF ETGs with Sérsic index $\ge2$ have $\langle \log$ ($L_{\rm 1.4 GHz}$/W Hz$^{-1}\rm )\rangle$ $\sim22.02\pm0.08$. The FIR luminosity $\langle \log$ ($L_{\rm FIR}$/$L_{\odot}\rm ) \rangle$ for the SF ETGs with Sérsic index $\ge2$ is $\sim10.33\pm0.07$ and $\sim10.22\pm0.07$ for the radio-detected and the FIR detected sample respectively. The upper end ($L_{\rm 1.4 GHz}\sim10^{23}$ W Hz$^{-1}$) of the radio luminosity range of the galaxies in this sample is consistent with that seen in the star-forming late-type galaxies in groups of galaxies (e.g. Omar & Dwarakanath 2005) and also in nearby clusters of galaxies (e.g., Reddy & Yun 2004). The ratio of total FIR to radio flux density (or luminosity), commonly represented by a logarithmic $'q'$ parameter (see equations A3 and A4 in the Appendix A) is plotted in Fig. \[figure:fig3\] against the FIR luminosity and the radio luminosity. The reference average value for the $'q'$ parameter is taken as 2.34 from Y01, along with two values at 1.86 and 2.82, indicating a deviation (excess or deficiency) of about three times in radio or FIR luminosity. Our choice of three-times deviation is somewhat stricter than the five-times deviation criterion used in Y01 to identify normal SF late-type galaxies without an AGN component. We used only the NVSS flux density for estimating the $'q'$ parameter as both $IRAS$ and NVSS surveys are sensitive to detect emissions to the full optical extent of the galaxies studied here. Fig. 2 indicates that all the SF ETGs are within the limits of three-times deviation in FIR or RC luminosity. The AGNs are normally identified as radio-excess galaxies with $'q'$ values significantly smaller than the average (Y01, Omar & Dwarakanath 2005). We do not find any radio-excess galaxies in our sample of SF ETGs. The FIR-RC correlations are known to be slightly non-linear (e.g., Nikalas & Beck 1997; Y01, Morić et al. 2010; Schleicher & Beck 2013) at IR luminosities below $\sim10^{9}$ L$_{\odot}$ and radio luminosities above $\sim10^{23}$ W Hz$^{-1}$. This non-linearity can be checked from Fig. 2. For a non-linear relation, $'q'$ will not be a constant and depend upon radio or FIR luminosity. For our sample with limited luminosity range, $'q'$ is consistent with being constant. These plots also indicate that the galaxies studied here are not falling within that range of luminosities, where non-linearities in the FIR-RC correlations are prominent and were reported previously. The scatter (rms of residuals) in the $'q'$ parameter for the SF ETGs is estimated to be 0.16 dex, which is similar to that known for the late-type galaxies. The average value of the $'q'$ parameter for the SF ETGs is estimated here at 2.35 $\pm$ 0.05, which is very close to the known average value of 2.34 $\pm$ 0.01 for a large sample of SF late-type galaxies (Y01). This is a strong indication that a tight FIR-RC correlation also exists in the SF early-type galaxies. Nature of radio emission ------------------------ The FIRST images show centrally concentrated radio emission in all the galaxies in our sample. The angular resolution of $5''$ in the FIRST images corresponds to nearly 2 kpc and 5 kpc of physical scales at the redshift of 0.02 and 0.05 respectively. Therefore, the FIRST images are sampling a large circum-nuclear region in these galaxies. It is also interesting to note that the residual optical images after a Sérsic profile fit to the light distribution in these galaxies show a circum-nuclear ring-like morphology of typical extent less than $10''$ (George 2017). This indicates that the star formation in these galaxies is most likely taking place in the circum-nuclear rings. Such circum-nuclear SF rings are common in early-type galaxies (Pogge & Eskridge 1993; Comerón et al. 2014; Kostiuk & Silchenko 2015). Intense far-ultraviolet emission indicating ongoing star-formation has been detected from the rings seen in the Hubble space telescope images of some S0 galaxies (Salim et al. 2012). These SF rings around galaxies are believed to be formed by resonances (Schommer & Sullivan 1976) due to non-axisymmetric gravitational potential and accretion of gas (Buta & Combes 1996). Therefore, we believe that the radio emission detected in the FIRST images is most likely associated with the star-forming circum-nuclear regions in these galaxies. Since the resolution ($45''$) in the NVSS images is much lower compared to that in the FIRST images ($5''$) and also compared to the optical sizes ($<41''$) of the galaxies, it is difficult to compare morphology of the NVSS radio emission with that of the FIRST radio emission. The powerful AGNs usually have their 1.4 GHz radio luminosities above $\sim10^{23}$ W Hz$^{-1}$. Since the galaxies studied here have radio luminosities in the range $10^{21.5} - 10^{22.4}$ W Hz$^{-1}$, large contaminations due to AGN are not expected. The optical line analyses as presented in S09 carefully isolated AGN hosting SF galaxies from the SF-dominant galaxies. The analyses of population fraction and star-formation history of various types (SF, AGN+SF, WEL) of ETGs indicate that the peak of AGN activity starts roughly 0.5 Gyr after the star-burst phase (Schawinski et al. 2007). Paswan & Omar (2016) analysed colors of the galaxies in the S09 sample studied here, and predicted similar time-scales for AGN trigger, which within $<1$ Gyr was also predicted to suppress star-formation. The star-bursts in interacting disk galaxies do not last more than nearly 50 Myr (Bernlöhr 1993; Mihos & Hernquist 1994). Therefore, intense SF and peak AGN activities in ETGs appear well separated in time. Despite these isolations, a contamination from a weak AGN in SF galaxies can not be ruled out in absence of higher resolution radio images. Nevertheless, some more arguments in favor of the radio emission being mainly related to the star formation in these galaxies can be given. A radio image stacking analysis of the WEL sample of ETGs drawn from the same S09 sample studied here, showed a residual average 1.4 GHz radio luminosity as $10^{20.3}$ W Hz$^{-1}$ in absence of intense star-formation (Paswan & Omar 2006). Therefore, AGNs are weak in these galaxies in general. The similarity in the average value and the scatter of the $'q'$ parameter in these ETGs with those for the late-type galaxies is also a strong indicator that the majority of radio and FIR emission is likely of a non-AGN origin. Morić et al. (2010) found in a sample of SF and AGN galaxies, that the $'q'$ value decreases with increasing radio luminosity. Their analysis showed a weak negative trend between the $'q'$ parameter and the 1.4 GHz radio luminosity in the range $10^{22}$ and $10^{25}$ W Hz$^{-1}$. The lower $'q'$ values at higher radio luminosities are possibly due to increased contribution of AGNs to the radio luminosity. The trend becomes prominent at radio luminosities above $10^{23}$ W Hz$^{-1}$. Ivison et al. (2010) also reported that the $'q'$ parameter starts showing significantly increased contribution in the radio continuum at $L_{\rm 1.4 GHz} \ge 10^{22.7}$ W Hz$^{-1}$. Moreover, Colina et al. (1995) showed that in radio-quite AGNs, pure starburst model can successfully reproduce the observed FIR-RC correlation in majority of galaxies. Since the radio luminosities of the galaxies in our sample are below $10^{22.4}$ W Hz$^{-1}$, we expect that our sample is not contaminated significantly by an AGN-related radio emission. The nearby SF galaxies, in which high-resolution radio images are available, typically $<10$ per cent of the total radio emission comes from a compact un-resolved core indicative of an AGN (Hill et al. 2001). Therefore, we assume a minor contribution at $10$ per cent level to the total radio emission from the AGN-related emission. We have subtracted this component while estimating SFR and magnetic field. Star formation rates -------------------- The star formation rates can be estimated from the radio and FIR luminosities using the standard conversion relations (see equations A5, A6 and A9 in the Appendix). For several cases, a radio detection corresponding to the FIR or the H$\alpha$ detection is missing. We find that the radio non-detections are consistent with the sensitivity limits of the two radio surveys, given a low level of star forming activities in these galaxies indicated by a weaker FIR or H$\alpha$ emission. The average SFR for the SF ETGs with Sérsic index $\ge2$ estimated from radio, FIR and H$\alpha$ are $\sim5.9$, $\sim6.5$, and $\sim5.3$ M$_{\odot}$ yr$^{-1}$ respectively. The average SFRs for the common sources detected in both radio and H$\alpha$ are nearly identical at the values of 5.9 and 6.1 M$_{\odot}$ yr$^{-1}$ respectively. Therefore, the average SFRs estimated using three indicators (radio, FIR, and H$\alpha$) are in general consistent with each other. It is to be noted that the radio-SFRs are estimated after subtracting 10 per cent flux (i.e., assumed AGN component) from the total radio flux. In majority of cases, differences in the SFRs estimated using luminosities in different wave-bands are well within a factor of two. This difference can arise due to several uncertainties inherent in the calibration factor to convert luminosity into SFR. The SFR estimates are sensitive to variation in the initial mass function (IMF) and the star formation history (Calzetti 2012). The FIR emission usually traces star formation up to $\sim100$ Myr, while the H$\alpha$ emission traces more recent ($\sim10$ Myr) star formation. In case of constant star formation over the past 100 Myr period, different SFR tracers are likely to provide similar estimates for the SFR in a galaxy. If IMFs are bottom-heavy (La Barbera, Ferreras & Vazdekis 2015), less number of the massive stars responsible for the synchrotron radio emission will be formed. Such IMFs can increase FIR to radio luminosity ratio. The FIR emission is also sensitive to additional parameters such as fraction of the ionizing photons absorbed by hydrogen and dust absorption efficiency of the non-ionizing photons from the young stars (e.g., Inoue, Hirashita & Kamaya 2000). Moreover, total FIR flux from galaxies without an AGN originates in two components, viz., cold cirrus heated by low-mass/evolved stars and warm dust heated by the young massive stars (Helou 1986). The cirrus component is not linked to recent massive star formation whereas the dominating synchrotron radio emission traces massive star formation activities over a few 100 Myr. A presence of weak AGN can also modify luminosities to some extent in different wave-bands. Overall, comparisons between SFRs estimated from different tracers are complicated due to the reasons discussed above and it is often worth to estimate SFRs in a galaxy from all the wavebands. We noticed that SFRs estimates for only three galaxies (S09 IDs - 61,77,175) deviate significantly (more than a factor of 3) in different wavebands. These deviations may be due to combinations of several possibilities discussed above. Thermal-radio fraction ---------------------- The radio emissions from the SF galaxies contain a small thermal component due to the free-free emission from the non-relativistic electrons in the star-forming regions. The thermal radio emission can be estimated from the H$\alpha$ flux density (or luminosity) as both the emissions are linked to the ionized region powered by the massive stars. The H$\alpha$ luminosities were estimated using the H$\alpha$ based SFR provided in S09. Since the H$\alpha$ measurements as provided in S09 were extrapolated values for the full extents of galaxies from the measurements made via $3''$ SDSS fibre values, such values may have large uncertainty. Therefore, we also estimated thermal-radio fraction using FIR emission as the FIR emission also traces star formation and the FIR measurements from the $IRAS$ are un-resolved for the sizes of the galaxies in this sample. The average of the H$\alpha$-based and the FIR-based estimates for the thermal-radio fraction is provided in Table 2. This analysis is reproduced in the Appendix (see equations A5, A6, A7, A8). The average value of the thermal-radio fraction at 1.4 GHz for the SF ETGs with Sérsic index $\ge2$ is estimated to be $\sim12$ per cent. It may be noted that both FIR and H$\alpha$ emission provided almost same mean value for the thermal-radio fraction in the range of $3 - 33$ % for these galaxies. The average value of the thermal-radio fraction in SF ETGs in our sample is similar to that obtained for the late-type star-forming galaxies (e.g., Condon 1992; Condon, Cotton & Broderick 2002). Magnetic field in star-forming circum-nuclear region ---------------------------------------------------- The magnetic fields in late-type galaxies usually have a regular field component and a turbulent field component (see e.g., Beck 2016). The regular and turbulent field components can be separated with the help of sensitive radio polarization measurements which are not available for our sample. Here, we estimate total magnetic field in the star-forming ETGs based on the derived synchrotron radio flux from the ISM. The synchrotron flux is taken as 0.78 times the observed radio flux in the FIRST images to remove the assumed 10 per cent contribution due to a possibility of weak AGN as discussed in Sect. 4.2 and 12 per cent average contribution from the thermal-radio component as estimated in Sect. 4.4. The magnetic fields can be estimated via an assumption made from one of the three plausible scenarios - (i) An equipartition between the total energy densities of CR particles and that of the magnetic field, (ii) A pressure equilibrium between the energy densities of CR particles and that of the magnetic field, (iii) Minimizing total energy density. Beck & Krause (2005) discussed these scenarios in detail and provide formulae to estimate magnetic field. The methods to estimate magnetic field in the ISM are described in the Appendix (see equations A10, A11). The radio synchrotron surface brightness (mJy arcsec$^{-2}$) and total path length in the emitting medium are needed to estimate magnetic field using equation A11. These values are not available directly in absence of a detailed high-resolution map of the radio emission. As star-formations in the ETGs are most likely taking place in a ring-like structure, we assume a flattened geometry for estimating field strengths. The source size fitting through deconvolution is not expected to be fruitful due to poor SNR in the radio images. The FIRST images are known to recover radio flux up to $\sim9''$ extent for a flux density $\sim3$ mJy (Becker et al. 1995). We used a radio source size of nearly $10''\times10''$ for all the galaxies in our sample. This size is also consistent with the typical extent of the circum-nuclear rings detected in the optical surface brightness analysis presented by George (2017). The thickness of the synchrotron emitting medium at 1.4 GHz was taken as 1 to 2 kpc in the estimates of field strengths made for the late-type galaxies (Fitt & Alexander 1993). On the other hand, the turbulent field scale-heights and the CR diffusion scale-lengths in the diffuse ISM of galaxies are estimated up to several kpc from the disk (see Beck et al. 1996). We used here the path length as 4 kpc to estimate total magnetic field in absence of high (sub-kpc) angular resolution radio images. Our choice for the path length is somewhere between that used by Fitt & Alexander (1993) and estimates for the diffuse ISM given in Beck et al. (1996). The non-thermal radio spectral indices for the ETGs studied here are not known. We assumed spectral index as $\sim0.75$ following the observed values of spectral index in the late-type star-forming galaxies (Condon 1992). The fields were estimated using the formulation provided in Eq. A11. The estimates of the magnetic fields can be highly uncertain due to poor constraints on the geometry of the synchrotron emitting region. Therefore, we also estimated lower and upper bounds to the estimated field strengths assuming a range of parameters. We took the bounds for the source size as $7''$ and $13''$, and those for the path-length as 2 kpc and 6 kpc. We obtain an error due to these uncertainties (source size, path-length and spectral index) as $+11/-4~\mu$G for the average estimated field at $\sim12~\mu$G for the SF ETGs. It is to be noted here that some other fixed parameters (e.g., electron to proton ratio at 100, and spectral index as $\sim0.75$) used in the calculation are also uncertain, and their effects are not included here. Discussions =========== The S09 sample of ETGs with high SFR provided an opportunity to study FIR and RC relation in ETGs. We found that the average and scatter of the $'q'$ parameter of the SF ETGs are similar to the SF late type galaxies. This is a strong indication that a linear relation between radio continuum and far-infrared luminosity is obeyed by star-forming ETGs also. Previously, Nyland et al. (2017) reported high FIR to radio ratio (high $'q'$ value) in about 20 per cent galaxies in the ATLAS$^{\mathrm{3D}}$ sample of ETGs, and discussed several radio-deficiency scenarios in ETGs including a possibility of lower magnetic fields. We do not find any large deviation in the $'q'$ parameter for the SF ETGs in the S09 sample. A close inspection of the properties of the galaxies in the S09 and ATLAS$^{\mathrm{3D}}$ samples reveals two major differences. The ATLAS$^{\mathrm{3D}}$ galaxies have low median SFR ($\sim0.15$ M$_{\odot}$ yr$^{-1}$; Davis et al. 2014) and high median stellar velocity dispersion (130 km s$^{-1}$; Cappellari et al. 2013) compared to the S09 sample ($<$SFR$> = 5 - 6$ M$_{\odot}$ yr$^{-1}$ and median stellar velocity dispersion $80$ km s$^{-1}$). The radio luminosities of the ATLAS$^{\mathrm{3D}}$ galaxies are in fact closer to those estimated for the WEL galaxies in the S09 sample (Paswan & Omar 2016). Our estimates for the magnetic field do not indicate a possibility of weak field in the star-forming ETGs at high SFRs. It is also worth to make a comparison of our study with the studies on the FIR-RC correlation on ’blue-cloud’ galaxies at intermediate and high redshifts (Sargent et al. 2010; Jarvis et al. 2010; Ivison et al. 2010; Bourne et al. 2011; Basu et al. 2015). The blue-cloud galaxies also show tight FIR-RC correlation, constructed using stacking of radio and IR images. The blue-cloud galaxies are also star-forming galaxies, selected based on their blue colors and contain similar galaxies in terms of stellar mass and SFR, as studied in this paper. However, the optical morphologies of the blue-cloud galaxies are presently unknown. Therefore, the sample of the blue ETGs studied here is unique in the sense that it is morphologically confirmed sample of blue star-forming early-type galaxies selected based on colors and optical spectroscopy. It is worth to point out some possibilities which may affect a FIR-RC relation in ETGs. The IMF variations in ETGs with high stellar velocity dispersion may affect the FIR-RC correlation as the IMFs are known to be strongly linked to stellar velocity dispersion in galaxies (La Barbera et al. 2015). The IMFs may become bottom-heavy at higher stellar velocity dispersions. A bottom-heavy IMF produces relatively more number of low-mass stars as compared to those expected in the Kroupa IMF (Kroupa 2002) for a fixed total mass of all the stars. A bottom-heavy IMF can therefore substantially alter supernovae rates and hence the radio synchrotron emission. The FIR-RC correlation in massive ETGs may be affected by a bottom-heavy IMF. A significant bottom-heavy IMF is not expected in the S09 sample of galaxies due to their low stellar velocity dispersion. Therefore, the S09 sample of ETGs has some unique properties, which make these galaxies suitable for having a normal and tight FIR-RC correlation similar to the late-type galaxies. It appears that the radio luminosities ($10^{21.5} - 10^{22.4}$ W Hz$^{-1}$) of the ETGs in the S09 sample are in a range, above which the the AGN contribution to the radio flux becomes more significant and below which the cirrus component contribution to the IR flux becomes more significant. The FIR-RC relation is also linked to field amplification in turbulence driven by supernova explosions of the massive stars (Tabatabaei et al. 2013; Schleicher & Beck 2013; Schober, Schleicher & Klessen 2016). Although the estimates of magnetic field in galaxies can be uncertain due to various unknown parameters and assumptions made, the field strengths estimated in the SF ETGs are not too different from those measured in SF late-type galaxies using nearly identical assumptions and choice of parameters. The current theories explain magnetic fields in spiral galaxies via a 3-stage process of seeding, amplifying and ordering. The seed fields can be generated in the young Universe by various possibilities such as purely primordial (Durrer & Neronov 2013), cosmological structure formation (Lazar et al. 2009), plasma fluctuations in protogalaxies (Schlickeiser & Felten 2013), Biermann mechanism in primordial supernovae remnants (Hanayama et al. 2005), or jets from the first black holes (Rees 2005). The cosmological seed fields can be amplified in galaxies to the equipartition field strengths at nearly $10~\mu$G level in $10 - 100$ Myr of time, in highly efficient small-scale dynamos driven by the turbulence generated in the supernova explosions (Balsara et al. 2004; Arshakian et al. 2009; Schober et al. 2012; Beck et al. 2012; Schleicher & Beck 2013). The turbulent kinetic energy is converted into the magnetic energy in small-scale dynamos. This amplified field is random. The field can become regular and ordered by the large-scale dynamos acting in the differential rotation of the ISM in spiral galaxies over a few giga-years of time (Beck et al. 2016). Regular fields are not expected here due to lack of ordered motions in elliptical galaxies. However, it may be noted that a possibility for field regularization over a few giga-year of time-scales in some dynamical components such as the rings in ETGs has been discussed (Moss et al. 2016). Large-scale magnetic fields can also be generated via cosmic-ray driven dynamos as discussed in Hanasz et al. (2009). All these scenarios will be worth to examine with new observations. Our estimates for the magnetic fields ($\sim12~\mu$G) in the circum-nuclear SF regions are therefore in good agreement with those predicted from small-scale dynamos acting in turbulence, driven by the type-II supernovae in the star-bursting regions. The star formation in the blue ETGs is very likely triggered by recent tidal interactions, which could be inferred from the optical images (George 2017). Since the random magnetic fields decay as a power-law in absence of continuous injection of turbulent energy (e.g., Beck et al. 2012), we do not expect quiescent low-mass ETGs to have strong magnetic fields after a few hundred Myr of cessation of a previous episode of star-burst. Strong central starbursts in interacting disk galaxies are predicted to last for durations shorter than $50$ Myr (Bernlöhr 1993; Mihos & Hernquist 1994). This in turn implies that the observed magnetic fields in the star-forming ETGs are very likely amplified during the present starburst. If turbulent field amplifications were taking significantly longer as compared to the star-formation period, this would have possibly resulted into a large scatter or deviation in FIR-RC relations (e.g., the $'q'$ parameter) in ETGs which was not seen in the present study. The fast field-amplification in turbulent dynamos is also consistent with the observed correlation between the turbulent field strength and the SFR on sub-galactic scales in some other galaxies (e.g., Tabatabaei et al. 2013; Chyży 2008). Our inference is in good agreement with the theories of random field amplification in turbulent dynamos. Summary and concluding remarks ============================== We presented here the FIR-RC correlation in terms of the ratio of FIR to RC luminosity ($'q'$ parameter) in a sample of star-forming ETGs in the redshift range of $0.02 - 0.05$, selected from the SDSS. The early-type morphology was identified with the Sérsic index values equal to or above 2. The star-forming nature was identified using the optical-line diagnostics schemes. These galaxies are characterized by blue colors ($u-r < 2.5$), intermediate stellar masses (a few times $10^{10}$ M$_{\odot}$) and central stellar velocity dispersion between 40 and 150 km s$^{-1}$. This sample is a unique morphologically confirmed sample of star-forming early-type galaxies, which is not contaminated from powerful central AGN sources. The main conclusions from this study are as follows:\ (1) The average value of the $'q'$ parameter for the star-forming ETGs is estimated to be $2.35\pm0.05$ with a scatter of 0.16 dex. These values are similar to those seen in the late-type star-forming galaxies.\ (2) The 1.4 GHz radio emission in the VLA FIRST images is detected mainly from the circum-nuclear star-bursting region.\ (3) The thermal radio continuum fraction in the star-forming ETGs is estimated at around 12 per cent, similar to that in the late-type galaxies.\ (4) The average star formation rate in the star-forming ETGs in our sample estimated using the radio continuum is $\sim6$ M$_{\odot}$ yr$^{-1}$. This value is consistent with those estimated using FIR and H$\alpha$ emission.\ (5) The average value of the magnetic field strengths in star-forming early-type galaxies is estimated to be 12$^{+11}_{-4}$ $\mu$G in the circum-nuclear region for a range of synchrotron path length as $2 - 6$ kpc, a range of source angular size as $7'' - 13''$, and for an assumed radio spectral index as $\sim0.75$.\ (6) These magnetic fields are very likely generated via fast amplification in small-scale turbulent dynamos driven by supernovae explosions in the star-bursting regions.\ It will be worth to take up follow-up studies on this sample of ETGs. The high resolution ($<1$ kpc-scale) multi-frequency radio imaging will be very valuable as it will provide estimates for the spectral index and resolve emissions on sub-galactic scales. The magnetic fields can then be better constrained in compact star-bursting regions. This unique sample of ETGs with high SFR will be an excellent target for radio polarization measurements to understand further evolution of magnetic field, particularly to assess the existence of large-scale magnetic field structure and their possible origin. [lccccccccccc]{} Gal. & Sérsic$^{a}$ & log & log & log & log & SFR$_{\rm rad}$ & SFR$_{\rm FIR}$ & SFR$_{\rm H\alpha}^{b}$ & $q$ & $R$$_{\rm th}$ & $B$\ ID & index & ($\frac{L_{radio}^{\rm FIRST}}{\rm W~Hz^{-1}}$) & ($\frac{L_{radio}^{\rm NVSS}}{\rm W~Hz^{-1}}$) & ($\frac{L_{{60\mu}{\rm m}}}{\rm L_\odot}$) & ($\frac{L_{\mathrm{FIR}}}{\rm L_\odot}$) &M$_{\odot}$ yr$^{-1}$ &M$_{\odot}$ yr$^{-1}$ & M$_{\odot}$ yr$^{-1}$ & & % & $\mu$G\ \ 146 & 8.0 & $21.34\pm0.07$ & $21.52\pm0.13$ & $9.79\pm0.06$ & $10.04\pm0.05$ & $1.8\pm0.5$ & $3.3\pm0.4$ & 4.8 & $2.53\pm0.14$ & $19.4$ & 9.9\ 84 & 8.0 & $<21.31$ & $<21.78$ & $9.84\pm0.12$ & $<10.36$ & $<3.6$ & $<6.9$ & 1.7 & — & — & —\ 14 & 8.0 & x & $21.99\pm0.09$ & $10.33\pm0.04$ & $10.54\pm0.04$ & $5.2\pm1.1$ & $10.5\pm1.0$ & 6.1 & $2.55\pm0.10$ & 13.4 & —\ 50 & 7.3 & $21.98\pm0.02$ & $21.98\pm0.05$ & $10.16\pm0.04$ & $10.39\pm0.04$ & $5.0\pm0.5$ & $7.4\pm0.7$ & 6.6 & $2.41\pm0.07$ & 11.6 & 13.2\ 61 & 6.3 & $22.00\pm0.02$ & $22.02\pm0.06$ & $10.18\pm0.04$ & $<10.32$ & $5.6\pm0.7$ & $<6.3$ & 21.0 & $<2.30$ & 20.5 & 13.6\ 195 & 5.4 & $<21.07$ & $<21.54$ & $9.44\pm0.10$ & $9.77\pm0.07$ & $<2.0$ & $1.8\pm0.3$ & 1.2 & $>2.23$ & — & —\ 206 & 5.3 & $<21.46$ & $<21.94$ & $9.78\pm0.17$ &$<10.17$ & $<5.1$ & $<4.5$ & 3.6 & — & — & —\ 8 & 4.9 & x & $21.97\pm0.08$ & $9.84\pm0.18$ & $<10.39$ & $5.0\pm0.9$ & $<7.4$ & 7.0 & $<2.42$ & 11.8 & —\ 149 & 4.8 & $22.05\pm0.02$ & $22.11\pm0.04$ & x & x & $6.8\pm1.2$ & — & 6.3 & — & 7.8 & 15.2\ 190 & 4.5 & $21.84\pm0.04$ & $22.01\pm0.07$ & $10.11\pm0.05$ & $10.32\pm0.04$ & $5.4\pm0.9$ & $6.3\pm0.6$ & 5.2 & $2.31\pm0.08$ & 8.9 & 11.7\ 124 & 4.3 & $21.68\pm0.02$ & $21.77\pm0.04$ & $9.86\pm0.03$ & $10.05\pm0.04$ & $3.1\pm0.3$ & $3.4\pm0.3$ & 5.2 & $2.27\pm0.06$ & 11.5 & 14.8\ 129 & 4.2 & $21.77\pm0.04$ & $22.13\pm0.05$ & $10.26\pm0.03$ & $10.43\pm0.03$ & $7.2\pm0.8$ & $8.2\pm0.6$ & 13.0 & $2.30\pm0.06$ & 12.2 & 12.4\ 4 & 4.2 & $21.42\pm0.08$ & $<21.69$ & $9.70\pm0.17$ & $<10.26$ & $>1.4^{c}$ & $<5.5$ & 2.5 & — & — & 9.4\ 215 & 4.1 & $21.90\pm0.02$ & $21.88\pm0.07$ & $10.19\pm0.05$ & $10.57\pm0.05$ & $4.0\pm0.7$ & $11.3\pm1.3$ & 7.6 & $2.69\pm0.09$ & 19.6 & 13.5\ 5 & 3.7 & $<21.29$ & $<21.77$ & $9.64\pm0.16$ & $9.96\pm0.12$ & $<3.5$ & $2.8\pm1.5$ & 3.5 & $>2.2$ & — & —\ 121 & 3.7 & $22.01\pm0.02$ & $22.07\pm0.07$ & $10.01\pm0.07$ & $10.34\pm0.06$ & $6.2\pm1.0$ & $6.6\pm0.9$ & 5.5 & $2.28\pm0.09$ & 8.2 & 13.5\ 172 & 3.6 & $22.07\pm0.02$ & $21.97\pm0.09$ & $9.91\pm0.13$ & $10.41\pm0.07$ & $5.0\pm1.0$ & $7.8\pm1.3$ & 8.1 & $2.44\pm0.11$ & 13.5 & 14.2\ 2 & 3.2 & $21.41\pm0.09$ & $<21.77$ & $9.88\pm0.15$ & $<10.34$ & $>1.3^{c}$ & $<6.6$ & 4.5 & — & — & 8.9\ 192 & 2.9 & $<21.34$ & $<21.82$ & $9.68\pm0.16$ & $<10.08$ & $<3.9$ & $<3.6$ & 1.7 & — & — & —\ 68 & 2.8 & $21.64\pm0.05$ & $<21.81$ & $<9.6$ & $<10.18$ & $>2.3^{c}$ & $<4.6$ & 3.2 & — & — & 10.1\ 177 & 2.8 & $<21.24$ & $<21.71$ & $9.96\pm0.11$ & $<10.23$ & $<3.0$ & $<5.1$ & 1.5 & — & — & —\ 102 & 2.8 & $<21.35$ & $<21.83$ & $9.84\pm0.08$ & $10.16\pm0.06$ & $<4.0$ & $4.4\pm0.6$ & 4.2 & $>2.33$ & — & —\ 56 & 2.7 & x & $22.28\pm0.07$ & $10.06\pm0.09$ & $10.44\pm0.08$ & $10.1\pm1.6$ & $8.3\pm1.6$ & 3.2 & $2.16\pm0.11$ & 4.8 & —\ 151 & 2.7 & $<21.38$ & $<21.86$ & $9.79\pm0.09$ & $<10.15$ & $<4.3$ & $<4.3$ & 3.7 & — & — & —\ 86 & 2.6 & $22.14\pm0.02$ & $22.39\pm0.04$ & $10.14\pm0.07$ & $10.59\pm0.05$ & $13.0\pm1.2$ & $11.8\pm1.4$ & 10.0 & $2.20\pm0.06$ & 7.0 & 12.9\ 182 & 2.5 & $21.45\pm0.09$ & $22.11\pm0.08$ & $10.12\pm0.04$ & $10.39\pm0.04$ & $6.8\pm1.2$ & $7.4\pm0.7$ & 6.1 & $2.28\pm0.09$ & 8.3 & 8.9\ 105 & 2.4 & $22.00\pm0.03$ & $22.06\pm0.09$ & $9.93\pm0.07$ & $<10.68$ & $6.1\pm1.2$ & $<14.6$ & 4.0 & $<2.62$ & 5.9 & 12.1\ 66 & 2.3 & $21.43\pm0.15$ & $<21.91$ & $9.98\pm0.06$ & $10.23\pm0.07$ & $>1.4^{c}$ & $5.1\pm0.9$ & 6.5 & $>2.32$ & — & 8.2\ 77 & 2.3 & $<21.35$ & $<21.83$ & $9.63\pm0.19$ & $10.20\pm0.06$ & $<4.0$ & $4.8\pm0.7$ & 0.5 & $>2.37$ & — & —\ 147 & 2.2 & $<21.47$ & $<21.94$ & $10.16\pm0.06$ & $10.57\pm0.04$ & $<5.1$ & $11.3\pm1.1$ & 6.7 & $>2.63$ & — & —\ 175 & 2.1 & $<21.39$ & $<21.87$ & $9.62\pm0.09$ & $10.03\pm0.09$ & $<4.4$ & $3.2\pm0.7$ & 0.6 & $>2.16$ & — & —\ 103 & 2.0 & $21.53\pm0.05$ & $21.88\pm0.07$ & $9.70\pm0.06$ & $10.04\pm0.04$ & $4.0\pm0.7$ & $3.3\pm0.3$ & 3.6 & $2.16\pm0.08$ & 7.3 & 10.7\ \ 180 & 1.8 & $22.27\pm0.02$ & $22.30\pm0.04$ & $10.61\pm0.02$ & $10.82\pm0.02$ & $10.6\pm1.0$ & $20.0\pm0.9$ & 18.0 & $2.52\pm0.05$ & 16.5 & 16.3\ 58 & 1.6 & $21.94\pm0.02$ & $22.09\pm0.04$ & $10.14\pm0.03$ & $10.37\pm0.04$ & $6.6\pm0.6$ & $7.1\pm0.7$ & 5.7 & $2.28\pm0.06$ & 8.3 & 15.4\ 52 & 1.6 & $21.53\pm0.07$ & $22.07\pm0.07$ & $10.12\pm0.02$ & $10.44\pm0.02$ & $6.2\pm1.0$ & $8.3\pm0.4$ & 3.9 & $2.37\pm0.07$ & 8.5 & 9.4\ 213 & 1.3 & $21.59\pm0.08$ & $<21.90$ & $10.05\pm0.07$ & $10.30\pm0.08$ & $>2.1^{c}$ & $6.0\pm1.1$ & 7.3 & $>2.40$ & — & 9.2\ 148 & 1.3 & $<21.44$ & $<21.92$ & $9.71\pm0.19$ &$<10.15$ & $<4.9$ & $<4.3$ & 1.0 & — & — & —\ 7 & 1.0 & x & $22.17\pm0.06$ & $10.48\pm0.04$ & $10.66\pm0.04$ & $7.8\pm1.1$ & $13.9\pm1.3$ & 12.0 & $2.49\pm0.07$ & 13.8 & —\ \ \ \ \[table:tab2\] Acknowledgements {#acknowledgements .unnumbered} ================ We thank the referee, Aritra Basu for making insightful comments which greatly improved the paper. We acknowledge some preparatory work carried out by P. Aromal and Suvendu Rakshit. The NVSS and FIRST sky survey were conducted by the National Radio Astronomical Observatory (NRAO) using the Very Large Array (VLA). The NRAO is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. The NASA/ IPAC Infrared Science Archive, which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. This research has made use of the SAO/NASA Astrophysics Data System (ADS) operated by the Smithsonian Astrophysical Observatory (SAO) under a NASA grant. [99]{} Adelman-McCarthy J.K. et al., 2008, ApJs, 175, 297 Arshakian T.G., Beck R., Krause M., Sokoloff D., 2009, A&A, 494, 21 Baldwin J.A., Phillips M.M., Terlevich R., 1981, PASP, 93, 5 Bally J., Thronson H.A., 1989, AJ, 97, 69 Balsara D.S., Kim J., Mac Low M., Mathews G.J., 2004, ApJ, 617, 339 Basu A., Roy S., Mitra, D., 2012, ApJ, 756, 141 Basu A., Wadadekar Y., Beelen A., Singh V., Archana K.N., Sirothia, S., Ishwara-Chandra C.H., 2015, ApJ, 803, 51 Beck A.M., Lesch H., Dolag K., Kotarba H., Geng A., Stasyszyn F.A., 2012, MNRAS, 422, 2152 Beck R., Brandenburg A., Moss D., Shukurov A., Sokoloff D., 1996, ARA&A, 34, 155 Beck R., 2016, Astron. Astrophys. Rev., 24, 4 Beck R., Krause, M., 2005, Astronomische Nachrichten, 326, 414 Becker R.H., White R.L., Helfand D.J., 1995, ApJ, 450, 559 Bernlöhr K., 1993, A&A, 268, 25 Bourne N., Dunne L., Ivison R.J., Maddox S.J., Dickinson M., Frayer D.T., 2011, MNRAS, 410, 1155 Buta R., Combes F., 1996, Fundamentals of Cosmic Physics, 17, 95 Calzetti D., 2013, in Falcon-Barroso J., Knapen J.H., eds, Secular Evolution of Galaxies, Cambridge Univ. Press, Cambridge, p. 419 Chyży K.T., 2008, A&A, 482, 755 Chyży K.T., Buta R.J., 2008, ApJ, 677, L17 Cappellari M. et al., 2011, MNRAS, 413, 813 Cappellari M. et al., 2013, MNRAS, 432, 1709 Colina L., Pérez-Olea, D.E., 1995, MNRAS, 277, 845 Combes F., Young L.M., Bureau M., 2007, MNRAS, 377, 1795 Comerón S. et al., 2014, A&A, 562, A121 Crocker A.F., Bureau M., Young L.M., Combes F., 2011, MNRAS, 410, 1197 Condon J.J., 1992, ARA&A, 30, 575 Condon J.J., Anderson M.L., Helou G., 1991, ApJ, 376, 95 Condon J.J., Cotton W.D., Greisen E.W., Yin Q.F., Perley R.A., Taylor G.B., Broderick J.J., 1998, AJ, 115, 1693 Condon J.J., Cotton W.D., Broderick J.J., 2002, ApJ, 124, 675 Davis T.A. et al., 2014, MNRAS, 444, 3427 Dopita M.A., Pereira M., Kewley L.J., Capaccioli M., 2002, ApJs, 143, 47 Durrer R., Neronov A., 2013, Astron. Astroph. Rev, 21, 62 Fitt A.J., Alexander P., 1993, MNRAS, 261, 445 George K., 2017, A&A, 598, A45 Govoni F., Feretti L., 2004, Intl. J. Mod. Phys. D., 13, 1549 Gunn, J.E. et al., 2006, AJ, 131, 2332 Hanasz M., Woltanski D., Kowalik K., 2009, ApJ, 706, L155 Hanayama H., Takahashi K., Kotake K., Oguri M., Ichiki K., Ohno H., 2005, ApJ, 633, 941 Harwit M., Pacini F., 1975, ApJ, 200, L127 Helou G., 1986, ApJ, 311, 33 Helou G., Bicay M.D., 1993, ApJ, 415, 93 Helou G., Kahn I.R., Malek L., Boehmer L., 1988, ApJS, 68, 151 Hill T.L., Heisler C.A., Norris R.P., Reynolds J.E., Hunstead R.W., 2001, AJ, 121, 128 Inoue A.K., Hirashita H., Kamaya H., 2000, PASJ, 52, 539 Ivison R.J. et al., 2010, A&A, 518, L31 Jaiswal S., Omar A., 2016, MNRAS, 462, 92 Jarvis M.J. et al., 2010, MNRAS, 409, 92 Jura M., 1986, ApJ, 306, 483 Kauffmann G. et al., 2003, MNRAS, 346, 1055 Kennicutt, R.C., 1998, ApJ, 498, 541 Kormendy J., 1977, ApJ, 218, 333 Kostiuk I.P., Silchenko O.K., 2015, Astrophys. Bull., 70, 280 Krause M., Wielebinski R., Dumke M., 2006, Astronomische Nachrichten, 327, 499 Kroupa P., 2002, Science, 295, 82 La Barbera F., Ferreras I., Vazdekis A., 2015, MNRAS, 449, L137 Lacki B.C., Thompson T.A., Quataert E., 2010, ApJ, 717, 1 Lazar M., Schlickeiser R., Wielebinski R., Poedts S., 2009, ApJ, 693, 1133 Lisenfeld U., Völk H.J., Xu C., 1996, A&A, 306, 677 Lisenfeld U. et al., 2007, A&A, 462, 507 Longair M.S., 1994, In High Energy Astrophysics, 2nd edition, Cambridge Univ. Press, vol. 2 Lucero D.M., Young, L.M., 2007, AJ, 134, 2148 Mihos J.C., Hernquist L., 1994, ApJ, 431, L9 Miller N.A., Owen F.N., 2001, AJ, 121 1903 Morić I., Smolčić V., Kimball A., Riechers D.A., Ivezić, Ž., 2010, ApJ, 724, 779 Moss D., Mikhailov E., Silchenko O., Sokoloff D., Horellou C., Beck R., 2016, A&A, 592, A44 Murgia M., Helfer T.T., Ekers R., Blitz L., Moscadelli L., Wong, T., Paladino R., 2005, A&A, 437, 389 Neugebauer G. et al., 1984, ApJ, 278, L1 Niklas S., Beck R., 1997, A&A, 320, 54 Nyland K. et al., 2017, MNRAS, 464, 1029 Omar A., Dwarakanth K.S., 2005, JAA, 26, 89 Paswan A., Omar A., 2016, MNRAS, 459, 233 Planck Collaboration, 2016, A&A, 594, A13 Pogge R.W., Eskridge P.B., 1993, AJ, 106, 1405 Reddy N.A., Yun, M.S., 2004, ApJ, 600, 695 Rees M., 2005, Lecture Notes in Physics, 664, 1 Roychowdhury S., Chengalur J.N., 2012, MNRAS, 423, L127 Ryle M., Hewish A., 1960, MNRAS, 120, 220 Salim S., Fang J.J., Rich R.M., Faber S.M., Thilker D.A., 2012, ApJ, 755,105 Sargent M.T. et al., 2010, ApJs,186, 341 Schawinski K., Khochfar S., Kaviraj S., 2006, Nature, 442, 888 Schawinski K., Thomas D., Sarzi M., Maraston C., Kaviraj S., Joo S., Yi S.K., Silk J., 2007, MNRAS, 382, 1415 Schawinski K. et al., 2009, MNRAS, 396, 818 Schleicher D.R.G., Beck R., 2013, A&A, 556, A142 Schlickeiser R., Felten T., 2013, ApJ, 778, 39 Schober J., Schleicher D.R.G., Federrath C., Klessen R.S., Banerjee R., 2012, Phys. Rev. E, 85, 26303 Schober J., Schleicher D.R.G., Klessen R.S., 2016, ApJ, 827, 109 Schommer R.A., Sullivan W.T., 1976, Astrophys. Lett., 17, 191 Suchkov A., Allen R.J., Heckman T.M., 1993, ApJ, 413, 542 Tabatabaei F.S. et al., 2013, A&A, 552, A19 Thompson T.A., Quataert E., Waxman E., Murray N., Martin C.L., 2006, ApJ, 645, 186 van der Kruit P.C., 1971, A&A, 15, 110 van der Kruit P.C., 1973, A&A, 29, 263 Völk H.J., 1989, A&A, 218, 67 Walsh D.E.P., Knapp G.R., Wrobel J.M., Kim, D.-W., 1989, ApJ, 337, 209 White R.L., Becker R.H., Helfand D.J., Gregg M.D., 1997, ApJ, 475, 479 Wrobel J.M., Heeschen D.S., 1991, AJ, 101, 148 Xu C., 1990, ApJ, 365, 47 Yun M.S., Reddy N.A., Condon J.J., 2001, ApJ, 554, 803 Estimation of luminosities, thermal radio flux, star formation rates, and magnetic field strengths ================================================================================================== A brief overview of the methods used to estimate different parameters presented in this paper is provided here. The radio and far-infrared flux densities were converted to luminosities using the relations: $${\rm log}~\frac{L_{\rm 1.4GHz}}{\rm W~Hz^{-1}} = 20.08 + 2~{\rm log}~\frac{D}{\rm Mpc} + {\rm log}~ \frac{S_{\rm 1.4GHz}}{\rm Jy} \label{equ:RADLUM}$$ $${\rm log}~\frac{L_{\rm FIR}}{\rm L_{\odot}} = 5.602 + 2~{\rm log}~\frac{D}{\rm Mpc} + {\rm log}~\frac{2.58~S_{{60\mu}{\rm m}} + S_{{100\mu}{\rm m}}}{\rm Jy} \label{equ:FIRLUM}$$ where flux and luminosity are denoted by $S$ and $L$ with a subscript indicating radio or infrared wave-band. $D$ represents the distance to the galaxy. Since the redshifts ($z$) of the galaxies studied here are low ($1 + z \approx$ 1), no $k$-correction was performed. Following Helou et al. (1998), the $'q'$ parameter was defined as: $$q = {\rm log} \left (\frac{2.58~S_{{60\mu}{\rm m}} + S_{{100\mu}{\rm m}}}{2.98~\rm Jy} \right) - {\rm log} \left(\frac{S_{\rm 1.4GHz}}{\rm Jy} \right) \label{equ:qpar}$$ or equivalently, $$q \approx 14.0 + {\rm log}~\frac{L_{\rm FIR}}{\rm L_{\odot}} - {\rm log}~\frac{L_{\rm 1.4GHz}}{\rm W~Hz^{-1}}$$ Kennicutt (1998) SFR relations were used to derive ${\rm H\!\alpha}$ flux from the ${\rm H\!\alpha}$ SFR provided in S09. The FIR luminosity was converted to SFR and also used to estimate equivalent ${\rm H\!\alpha}$ flux using the Kennicutt (1998) SFR relations assuming that both FIR and ${\rm H\!\alpha}$ trace same SFR: $$\frac{\rm SFR}{\rm ~M_\odot~yr^{-1}} = {{L_{\rm FIR} (\rm L_{\odot})} \over {3.3 \times 10^9} }$$ $$\frac{\rm SFR}{\rm ~M_\odot~yr^{-1}} = 9\times 10^8 \left(\frac{D}{\rm Mpc} \right)^2 \left(\frac{S_{\rm H\!\alpha}}{\rm erg~s^{-1}~cm^{2}}\right)$$ Subsequently, two values for the thermal-radio component for each galaxy were estimated, one value using the derived $\rm {H\!\alpha}$ flux from the FIR luminosity and other using the directly observed $\rm {H\!\alpha}$ flux from S09 and the relation given by Dopita et al. (2002): $$\frac{S_{\rm 1.4GHz}^{\rm thermal}}{\rm Jy} = 1.21 \times 10^{9} ~\left(\frac{S_{\rm H\!\alpha}}{\rm erg~s^{-1} cm^{2}}\right)$$ where $S_{\rm H\!\alpha}$ denotes $\rm {H\!\alpha}$ flux and SFR is the star formation rate. The SFR estimates are generally valid for a constant SFR in the past 100 Myr and are sensitive to variations in the initial mass function and SF history. The thermal-radio fraction (in per cent) was estimated as: $$R_{\rm th} = 100~\frac{S_{\rm 1.4GHz}^{\rm thermal}}{S_{\rm 1.4GHz}^{\rm NVSS}}$$ The SFR using the total (thermal plus non-thermal) 1.4 GHz radio flux density was estimated using the relation given in Yun et al. (2001): $$\frac{\rm SFR}{\rm ~M_\odot~yr^{-1}} \approx 5.9 (\pm1.8) \times 10^{-22} \frac{L_{\rm 1.4GHz}}{\rm W~Hz^{-1}}$$ The magnetic field ($B$) at low redshifts ($1+z\approx1$) can be estimated using the following relation obtained after minimizing the total energy density (CR and magnetic field) in the ISM: $$B (\mu {\rm G}) = A_{\alpha,\nu_{min},\nu_{max}}~(1 + K)^{\beta}~\nu^{\alpha\beta}~I^{\beta}~d^{-\beta}$$ where $\alpha$ is radio spectral index, $\nu$ is radio frequency in MHz, $I$ is radio surface flux density in mJy arcsec$^{-2}$ measured at frequency $\nu$ and $d$ is the path length of the synchrotron emitting region in kpc. In classical derivations (see Longair 1994) the index $\beta$ is constant at a value of 2/7. In the modified estimates of Beck & Krause (2005), the value of $\beta$ is 1/(3+$\alpha$). $K$ is the ratio of the total energy of CR nuclei to that of the synchrotron emitting electrons and positrons in the classical formulae while $K$ is taken as the ratio of number densities of CR protons and electrons per particle energy interval within the energy range traced by the observed synchrotron emission. The $K$ is normally taken as a constant value at 100, consistent with the values measured in the local Galactic CR data near the sun at particle energies of a few GeV and predictions from Fermi shock acceleration and hadronic interaction models (see Beck & Krause 2005 for details). The factor $A$($\alpha,\nu_{min},\nu_{max}$) can be estimated from Govoni & Feretti (2004) and Beck & Krause (2005) for the classical and modified cases respectively. The spectral indices $\alpha$ are not known for the galaxies studied here. We assumed $\alpha$ as $\sim0.75$ consistent with the average values of radio spectral index observed in the star-forming galaxies (Condon 1992). For $K$ = 100 and $\alpha\sim0.75$, we found that the derived field strengths using formulae with classical index 2/7 and modified index 1/(3+$\alpha$) give nearly identical results. At $\alpha\sim0.75$, the uncertainties in the field estimates due to unknown values of path length and radio source size in our sample are much higher than the uncertainty arising from the use of the classical formulae or the modified formulae. Therefore, the field strengths can be estimated using the classical formulae or the modified formulae for $\alpha \approx 0.75$, without introducing a significant error. In the absence of spectral-index measurements, the simplified classical relation to estimate galactic magnetic field in terms of the synchrotron surface flux density measured at 1.4 GHz, and assuming $\alpha\sim0.75$, can be written as: $$B (\mu {\rm G}) \approx 60 \left(\frac{I_{\rm {1.4 GHz}}}{\rm {mJy~arcsec^{-2}}}\right)^{2/7} \left(\frac{d}{\mathrm{kpc}}\right)^{-2/7}$$ It is to be noted here that at significantly flatter or steeper spectral index than 0.75, the above equation requires significant corrections as pointed out in Beck & Krause (2005). Radio images ============ The remaining radio-contour images overlaid upon the grey-scale SDSS g-band images are presented here. The images are made following the same contour scheme as in the Fig. 1. \[lastpage\] [^1]: E-mail: aomar@aries.res.in

LPENSL-Th 05/99\ solv-int/9903016\ **Canonicity of : $r$-matrix approach. I.** E K Sklyanin [On leave from: Steklov Mathematical Institute at St. Petersburg, Fontanka 27, St. Petersburg 191011, Russia. E-mail: [sklyanineuclid.pdmi.ras.ru]{}]{} 0.4cm Laboratoire de Physique [UMR 5672 du CNRS et de l’ENS Lyon]{} , Groupe de Physique Théorique, ENS Lyon,\ 46 allée d’Italie, 69364 Lyon 07, France For the Hamiltonian integrable systems governed by $SL(2)$-invariant $r$-matrix (such as Heisenberg magnet, Toda lattice, nonlinear Schrödinger equation) a general procedure for constructing  is proposed. The corresponding BT is shown to preserve the Poisson bracket. The proof is given by a direct calculation using the $r$-matrix expression for the Poisson bracket. 25 March, 1999 In 1976 Flaschka and McLaughlin [@FM76] has demonstrated that the standard  (BT) for the KdV equation is canonical with respect to the associated symplectic structure. Subsequently, the canonicity of BT[s]{} has been proved for some more integrable models, see e. g. Appendix to [@MS91]. The aim of the present paper is to apply the $r$-matrix formalism [@FT87] to the problem of proving the canonicity of BT. Our proof is quite general and requires only that the Poisson brackets of the corresponding Lax operator could be written down in the $r$-matrix form. In our treatment of  we adopt the approach of [@KS5] where a program has been formulated of reexamining BT[s]{} from the Hamiltonian point of view. Consider an integrable Hamiltonian system possessing a Lax matrix $L(u)$ depending on the dynamical variables and a complex parameter $u$ (spectral parameter). The spectral invariants of $L(u)$ are supposed to generate the commuting Hamiltonians of the system. Defining a  as a canonical transformation preserving the Hamiltonians of the system (see [@KS5] for a more detailed list of properties of BT) we conclude that it has to preserve as well the spectral invariants of $L(u)$. As a consequence, the original matrix $L(u)$ and the transformed one $\L(u)$ must be related by a similarity transformation Ł(u)=M(u)L(u)M(u)\^[-1]{}. \[def-Lambda\] (see [@MS91] for a detailed account of the theory of BT as gauge transformations). An important practical question arising in the theory of BT is how to find, given a Lax matrix $L(u)$, such a matrix $M(u)$ which would generate a BT. To check that a matrix $M(u)$ is admissible one needs, first, to verify that the system of equations resulting from is self-consistent, and, second, to proof that the resulting transformation of dynamical variables is canonical, that is preserves the Poisson bracket. Below we solve the both problems for the class of integrable systems governed by the $SL(2)$-invariant $r$-matrix. Suppose that $L(u)$ is a matrix of order $2\times2$ and the Poisson brackets between the entries of $L(u)$ can be expressed in the $r$-matrix form [@FT87] {L(u),L(v)}=\[r\_[12]{},L(u)L(v)\] \[pb-LL\] where $\one L=L\otimes\id$, $\two L=\id\otimes L$, and $ r_{12}=\kappa(u-v)^{-1}\P_{12}$ is the standard $SL(2)$-invariant solution to the classical Yang-Baxter equation [@FT87], $\kappa$ being a constant and $\P_{12}$ being the permutation operator in $\C^2\otimes\C^2$. The class of integrable models thus defined includes such well-known models as the nonlinear Schrödinger equation, Heisenberg magnetic chain, Toda lattice [@FT87]. When choosing an ansatz for the matrix $M(u)$ we shall follow [@QNCL84; @NP89] where it was observed that, in the cases of Heisenberg magnetic chain and of the lattice Landau-Lifshitz equation, the matrix $M(u)$ happens to have the same form, as a function of $u$, as the corresponding elementary Lax matrix $L(u)$ for the chain consisting only of one atom. As shown below, such choice of $M(u)$ is valid for any integrable model governed by the $r$-matrix specified above. Our ansatz for $M(u)$ mimics the form of elementary Lax matrix for the isotropic Heisenberg magnetic chain [@FT87; @QNCL84]: M(u)=(u-ł)+\[eq:ansatz-M\] where =0, =-\^2 \[eqs-S\] $\l$ and $\mu$ being free parameters. It is convenient to perform a reparametrization $\l_1=\l+\mu$, $\l_2=\l-\mu$, so that $\mu=(\l_1-\l_2)/2$. The constraints on the matrix $\Sb$ leave two more undetermined parameters. Denoting them $p$ and $q$ and choosing a particular parametrization of $\Sb$ we fix the following ansatz for $M(u)$ M(u)=( [cc]{} u-ł\_1+pq & p\ -pq\^2+2q & u-ł\_2-pq ). \[def-M\] We shall consider $\l_1$, $\l_2$ as the free parameters of BT. The parameters $p$, $q$ are then to be determined from the equations . Let us introduce the eigenbasis of $M(u)$ =( [c]{} 1\ -q ), =( [c]{} p\ 2-pq ) and the dual eigenbasis =(2-pq,-p), =(q,1), as well as the corresponding spectral projectors P\_[ij]{}= , i,j{1,2} satisfying P\_[ij]{}P\_[kl]{}=P\_[il]{}\_[jk]{}. In terms of the projectors $P_{ij}$ the matrix $M(u)$ and its inverse are written down, respectively, as M(u)=(u-ł\_1)P\_[11]{}+(u-ł\_2)P\_[22]{} and M(u)\^[-1]{}=(u-ł\_1)\^[-1]{}P\_[11]{}+(u-ł\_2)\^[-1]{}P\_[22]{}. \[M\^[-1]{}\] Note that M(u)=(u-ł\_1)(u-ł\_2). Let us derive now from the equations determining the parameters $p$, $q$. Suppose that $L(u)$ is polynomial in $u$ (this covers all lattice models, the continuous models can then be obtained as appropriate limits). Requiring that the transformation preserves the polynomiality of $L(u)$ we conclude that the apparent poles of the right-hand-side of at $u=\l_{1,2}$ due to should vanish, 0=\_[u=ł\_i]{}Ł(u)=(ł\_i-ł\_[1-i]{})P\_[1-i,1-i]{}L(ł\_i)P\_[ii]{}, which gives us two equations for determining $p$ and $q$ as functions of the dynamical variables of the system: P\_[12]{}L(ł\_1)=0, P\_[21]{}L(ł\_2)=0. \[eq-pq\] With the parameters $p$ and $q$ determined by the equations , the matrix $\L(u)$ is defined by as a function of the dynamical variables and free parameters $\l_{1,2}$. The next step is to show that $\L(u)$ satisfies the same Poisson bracket relations as $L(u)$: {Ł(u),Ł(v)}=\[r\_[12]{},Ł(u)Ł(v)\]. \[pb-LmbLmb\] The calculation of the Poisson brackets for $\L(u)$, though cumbersome, is quite straightforward, since all the necessary ingredients are already prepared. Substituting into $\{\one\L(u),\two\L(v)\}$ and differentiating the products of matrices we obtain a rather long expression. To write it down in a more compact form let us introduce the following notation &=& M(u){L(u),M(v)}M(u)\^[-1]{}M(v)\^[-1]{},\ &=& M(v){M(u),L(v)}M(u)\^[-1]{}M(v)\^[-1]{},\ &=& M(u)M(v){M(u),M(v)}M(u)\^[-1]{}M(v)\^[-1]{}, \[def&lt;&gt;\] r\_[12]{}=M(u)M(v)r\_[12]{}M(u)\^[-1]{}M(v)\^[-1]{}. \[def-rtilde\] Using , one can write down the left-hand-side of as {Ł(u),Ł(v)}&=& Ł(u)Ł(v) +Ł(v) -Ł(u)Ł(v)\ &&+Ł(u) +\[r\_[12]{},Ł(u)Ł(v)\] -Ł(u)\ &&-Ł(v)Ł(u) -Ł(v) +Ł(u)Ł(v). \[9terms\] Our aim is to show that the resulting expression is equal to $[r_{12},\one\L(u)\two\L(v)]$. Note, first, that using the identity ¶\_[12]{}=P\_[11]{}P\_[11]{}+P\_[12]{}P\_[21]{}+P\_[21]{}P\_[12]{}+ P\_[22]{}P\_[22]{} one can show that r\_[12]{}=r\_[12]{}+2( -). \[r-tilde\] It remains then to calculate the Poisson brackets between $L(u)$ and $M(u)$. To do it, we recollect the equations defining implicitely $p$ and $q$. The calculate the Poisson brackets for $p$ and $q$ we shall use the trick employed in a similar situation in [@Skl38]. For any function $f$ on the phase space we have 0&=&{f,P\_[12]{}L(ł\_1)}\ &=& {f,p}L(ł\_1) +{f,q}L(ł\_1) +P\_[12]{}{f,L(ł\_1)},\ 0&=&{f,P\_[21]{}L(ł\_2)}\ &=& {f,p}L(ł\_2) +{f,q}L(ł\_2) +P\_[21]{}{f,L(ł\_2)}. \[pb-fpq\] The Poisson brackets $\{f,p\}$ and $\{f,q\}$ are then determined by solving the linear system . Using the equalities =0, =(P\_[11]{}-P\_[22]{})+P\_[12]{}, =P\_[11]{}-P\_[22]{}, =(P\_[11]{}-P\_[22]{})-P\_[21]{}, =2p, =p\^2P\_[12]{}+P\_[21]{} and introducing the notation w\_i(ł)=P\_[ii]{}L(ł), w(ł)=w\_1(ł)-w\_2(ł) one obtains then {f,M(v)}=-P\_[21]{}{f,L(ł\_1)}P\_[12]{} -P\_[12]{}{f,L(ł\_2)}P\_[21]{} (note that the last formula does not depend on parametrization of ). Now, using it is easy to calculate the brackets {L(u),M(v)}&=& - (w\_1(ł\_1)P\_[12]{}L(u)P\_[21]{} -w\_2(ł\_1)L(u)P\_[12]{}P\_[21]{})\ &&- (w\_2(ł\_2)P\_[21]{}L(u)P\_[12]{} -w\_1(ł\_2)L(u)P\_[21]{}P\_[12]{}), {M(u),L(v)}&=& (w\_1(ł\_1)P\_[21]{}P\_[12]{}L(v) -w\_2(ł\_1)P\_[21]{}L(v)P\_[12]{})\ &+& (w\_2(ł\_2)P\_[12]{}P\_[21]{}L(v) -w\_1(ł\_2)P\_[12]{}L(v)P\_[21]{}), {M(u),M(v)}= -[w(ł\_1)w(ł\_2)]{} (P\_[12]{}P\_[21]{}-P\_[21]{}P\_[12]{}), Recalling the notation one arrives to the expressions &=& -( P\_[12]{}Ł(u)P\_[21]{} -Ł(u)P\_[12]{}P\_[21]{})\ &&-( P\_[21]{}Ł(u)P\_[12]{} -Ł(u)P\_[21]{}P\_[12]{}), \[LM\] &=& ( P\_[21]{}P\_[12]{}Ł(v) -P\_[21]{}Ł(v)P\_[12]{})\ &&+( P\_[12]{}P\_[21]{}Ł(v) -P\_[12]{}Ł(v)P\_[21]{}), \[ML\] &=& -[w(ł\_1)w(ł\_2)]{}\ &&( -). \[MM\] Finally, substituting , , together with into , we obtain, after massive cancellations, the equality which finishes the proof of the canonicity of . Discussion {#discussion .unnumbered} ========== In this paper, we have studied only the case of the XXX type $r$-matrix. There is little doubt that our results can be generalized to the cases of XXZ and XYZ type $r$-matrices. Taking the limit of linear Poisson brackets {(u),(v)}=\[r\_[12]{},(u)+(v)\] one obtains, as a corollary, the canonicity of BT for the Gaudin model [@Skl38]. The proof of canonicity of BT given in the present paper, being quite simple and straightforward, does not explain, however, the reasons why we should choose for $M$ the same ansatz as for $L$. An answer to this question is given in the second part of our paper which is being prepared for publication. Acknowledgements {#acknowledgements .unnumbered} ================ I am grateful to V. B. Kuznetsov and F. Nijhoff for their interest in the work and useful discussions. This work was started during my stay at the Department of Mathematics, the University of Leeds. and benefited from the financial support of EPSRC. [12]{} H. Flaschka and D.W. McLaughlin, [*Some comments on , canonical transformations, and the Inverse Scattering Method,*]{} In: Bäcklund Transformations, ed. by R.M. Miura, Lect. Notes in Math.  [**515**]{} (1976) 253–295 V.B. Matveev and M.A. Salle, Darboux Transformations and Solitons, Berlin: Springer (1991). L.D. Faddeev and L.A. Takhtajan, Hamiltonian Methods in the Theory of Solitons, Berlin: Springer (1987). V.B. Kuznetsov and E.K. Sklyanin, [*On for many-body systems,*]{} J. Phys. A: Math. Gen. [**31**]{} (1998) 2241–2251. G.R.W. Quispel, F.W. Nijhoff, H.W. Capel and J. van der Linden, [*Linear integral equations and nonlinear difference-difference equations,*]{} Physica [**125A**]{} (1984) 344–380. F.W. Nijhoff and V. Papageorgiou, [*Lattice equations associated with the Landau-Lifshitz equations,*]{} Phys. Lett. A [**141**]{} (1989) 269–274. E.K. Sklyanin, [*Separation of variables. New trends.*]{} Progr. Theor. Phys. Suppl. [**118**]{} (1995) 35–60. E.K. Sklyanin. [*Canonicity of : $r$-matrix approach. II,*]{} [solv-int/9903017]{}; to be published in: [*Trudy MIAN, v. 226 (1999), Moskow, Nauka.*]{}

--- abstract: 'This paper investigates optimal consumption, investment, and healthcare spending under Epstein-Zin preferences. Given consumption and healthcare spending plans, Epstein-Zin utilities are defined over an agent’s random lifetime, partially controllable by the agent as healthcare reduces Gompertz’ natural growth rate of mortality. In a Black-Scholes market, the stochastic optimization problem is solved through the associated Hamilton-Jacobi-Bellman (HJB) equation. Compared with classical Epstein-Zin utility maximization, the additional controlled mortality process complicates the uniqueness of Epstein-Zin utilities and verification arguments. A combination of probabilistic arguments and analysis of the HJB equation are required to resolve the challenges. In contrast to prior work under time-separable utilities, Epstein-Zin preferences largely facilitate calibration. In four different countries we examined, the model-generated mortality closely approximates actual mortality data; moreover, the calibrated efficacy of healthcare is in broad agreement with empirical studies on healthcare across countries.' author: - 'Joshua Aurand[^1]' - 'Yu-Jui Huang[^2]' bibliography: - 'refsMathSci.bib' title: | Mortality and Healthcare:\ an Analysis under Epstein-Zin Preferences --- **MSC (2010):** 91G10, 93E20. **JEL:** G11, I12 **Keywords:** Consumption-investment problem, Healthcare, Mortality, Gompertz’ law, Epstein-Zin utilities, Random horizons, Backward stochastic differential equations. Introduction {#sec:intro} ============ Mortality, the probability that someone alive today dies next year, exhibits an approximate exponential growth with age, an observation made by Gompertz [@Gompertz25] in the early 19th century. Despite the steady decline of mortality at all age groups [*across*]{} different generations, the exponential growth of mortality [*within*]{} each generation has remained remarkably stable, which is called the Gompertz law. Figure \[fig:cal\_US\] displays this clearly: in the US, mortality of the cohort born in 1900 (blue dots) and that of the cohort born in 1940 (red dots) grew exponentially at a similar rate; the latter is essentially shifted down from the former.[^3] ![Mortality rates (vertical axis, in logarithmic scale) at adults’ ages for the cohorts born in 1900 and 1940 in the US. The dots are actual mortality data (Source: Berkeley Human Mortality Database), and the lines are model-implied mortality curves.[]{data-label="fig:cal_US"}]({Model_Calibration/Updates/US_Mortality2.pdf}){width=".75\linewidth"} At the intuitive level, the steady decline of mortality across generations can be ascribed to the continuous improvement of healthcare and accumulation of wealth. Understanding the precise relations among healthcare, wealth, and mortality demands a general model in which wealth evolution, healthcare choices, and the resulting mortality are all [*endogenous*]{}. Standard models of optimal consumption and investment do not seem to serve the purpose: the majority, e.g. [@Yaari1965], [@Richard1975], [@rosen1988value], and [@ShepardZeckhauser1984], consider no more than exogenous mortality, leaving no room for healthcare.[^4] Recently, Guasoni and Huang [@Huang19] directly modeled the effect of healthcare on mortality: healthcare reduces Gompertz’ natural growth rate of mortality, through an [*efficacy function*]{} that characterizes the effect of healthcare spending in a society. Healthcare, as a result, indirectly increases utility from consumption accumulated over a longer lifetime. Under the constant relative risk aversion (CRRA) utility function $U(x) = \frac{x^{1-\gamma}}{1-\gamma}$, $0<\gamma<1$, an optimal strategy of consumption, investment, and healthcare spending is derived in [@Huang19], where the constraint $0<\gamma<1$ is justified by interpreting $1/\gamma$ as an agent’s [*elasticity of intertemporal substitution*]{} (EIS). Specifically, to model mortality endogenously, we need to be cautious of potential preference for death over life. To avoid this, [@Huang19] assumes that an agent can leave a fraction $\zeta\in (0,1]$, not necessarily all, of his wealth at death to beneficiaries, reflecting the effect of inheritance and estate taxes. It is shown in [@Huang19] that the optimization problem is ill-posed for $\gamma>1$. Indeed, with $\gamma>1$, or EIS less than one, the income effect of future loss of wealth at death is so substantial that the agent reduces current consumption to zero, leading to the ill-posedness; see below [@Huang19 Proposition 3.2] for details. Despite the progress in [@Huang19], the artificial relation that EIS is the reciprocal of [*relative risk aversion*]{}, forced by CRRA utility functions, significantly restricts its applications. Although a preliminary calibration was carried out in [@Huang19 Section 5], it was not based on the full-fledged model in [@Huang19], but a simplified version without any risky asset. Indeed, once a risky asset is considered, it is unclear whether $\gamma$ should be calibrated to relative risk aversion or EIS. More crucially, empirical studies largely reject relative risk aversion and EIS being reciprocals to each other: it is widely accepted that EIS is larger than one (see e.g. [@Bansal04], [@Bansal07], [@Bhamra10], and [@Benzoni11]), while numerous estimates of relative risk aversion are also larger than one (see e.g. [@Vissing03], [@Bansal04], and [@Hansen07]). In this paper, we investigate optimal consumption, investment, and healthcare spending under preferences of Epstein-Zin type, which disentangle relative risk aversion (denoted by $0<\gamma\neq 1$) and EIS (denoted by $\psi>0$). In particular, we assume throughout the paper $$\label{specification} \psi >1\quad \hbox{and}\quad \gamma>{1}/{\psi},$$ which implies a preference for early resolution of uncertainty (as explained in [@Skiadas98]), and conforms to empirical estimations mentioned above. Our Epstein-Zin utility process has several distinctive features. First, it is defined on a random horizon $\tau$, the death time of an agent. Prior studies on Epstein-Zin utilities focus on a fixed-time horizon; see e.g. [@DuffieLions], [@Schroder96], [@Kraft13], [@Seifried16], [@Kraft17], and [@Xing]. To the best our knowledge, random-horizon Epstein-Zin utilities are developed for the first time in Aurand and Huang [@AH19], where the horizon is assumed to be a stopping time adapted to the market filtration. Our studies complement [@AH19], by allowing for a stopping time (i.e. the death time) that need not depend on the financial market. Second, the random horizon $\tau$ is [*controllable*]{}: one slows the growth of mortality via healthcare spending, which in turn changes the distribution of $\tau$. Note that a controllable random horizon is rarely discussed in stochastic control, even under time-separable utilities. Third, to formulate our Epstein-Zin utilities, we need not only a given consumption stream $c$ (as in the literature), but also a specified healthcare spending process $h$. Given the pair $(c,h)$, the Epstein-Zin utility is defined as the right-continuous process $\widetilde V^{c,h}$ that satisfies a random-horizon dynamics (i.e. below), with a jump at time $\tau$. Thanks to techniques of filtration expansion, we decompose $\widetilde V^{c,h}$ as a function of $\tau$ and a process $V^{c,h}$ that solves an infinite-horizon backward stochastic differential equation (BSDE) under [*solely*]{} the market filtration; see Proposition \[prop:decompose wtV\]. That is, the randomness from death and from the market can be dealt with separately. By deriving a comparison result for this infinite-horizon BSDE (Proposition \[prop:comparison\]), we are able to uniquely determine the Epstein-Zin utility $\widetilde V^{c,h}$ for any $k$-admissible strategy $(c,h)$ (Definition \[k-admissible\]); see Theorem \[thm:Existence and Uniqueness\]. In a Black-Scholes financial market, we maximize the time-0 Epstein-Zin utility $\widetilde V^{c,h}_0$ over [*permissible*]{} strategies $(c,\pi,h)$ of consumption, investment, and healthcare spending (Definition \[cP\]). First, we derive the associated Hamilton-Jacobi-Bellman (HJB) equation, from which a candidate optimal strategy $(c^*,\pi^*,h^*)$ can be deduced. Taking advantage of a scaling property of the HJB equation, we reduce it to a nonlinear ordinary differential equation (ODE), for which a unique classical solution exists on strength of the Perron method construction in [@Huang19]. This, together with a general verification theorem (Theorem \[Verification\]), yields the optimality of $(c^*,\pi^*,h^*)$; see Theorem \[Thm:AH\], the major result of this paper. Compared with classical Epstein-Zin utility maximization, the additional controlled mortality process $M^h$ in our case adds nontrivial complexity. In deriving the comparison result Proposition \[prop:comparison\], standard Gronwall’s inequality cannot be applied due to the inclusion of $M^h$. As shown in Appendix \[subsec:proof of comparison\], a transformation of processes, as well as the use of both forward and backward Gronwall’s inequalities, are required to circumvent this issue. On the other hand, in carrying out verification arguments, we need to contain the growth of $M^h$ properly to ensure that the Epstein-Zin utility is well-defined. This is done through a combination of probabilistic arguments and analysis of the aforementioned nonlinear ODE; see Appendix \[subsec:proof of Thm:AH\] for details. Our model is calibrated to mortality data in the US, the UK, the Netherlands, and Bulgaria. There are three intriguing findings. First, our model-implied mortality closely approximates the actual mortality data. Under the simplifying assumptions that the cohort born in 1900 had no healthcare and the cohort born in 1940 had full access to healthcare, we use our model to generate an endogenous mortality curve for the 1940 cohort. Figure \[fig:cal\_US\] shows that the model-implied mortality (the red line) essentially reproduces actual data (the red dots). Our model performs well for not only the US data, but those from other countries as well; see Figure \[fig:calibration\]. Second, as shown in Figure \[fig:CalibratedEfficiency\], the calibrated efficacy functions of healthcare (one for each country) indicate a ranking in terms of the effectiveness of healthcare spending: across realistic levels of spending, healthcare is more effective in the Netherlands than in the UK, in the UK than in the US, and in the US than in Bulgaria. This ranking, somewhat surprisingly, is in broad agreement with empirical studies on healthcare across countries; see Section \[subsec:g\]. Third, healthcare spendings in these four countries all increase steadily with age, relative to both total wealth and total spending, but differ markedly in magnitude; see Figure \[fig: healthcarespending\]. This, together with the ranking of efficacy in Figure \[fig:CalibratedEfficiency\], reveals that higher efficacy of healthcare induces lower healthcare spending. ![Calibrated efficacy of healthcare $g(h)$, measured by the reduction in the growth of mortality, given proportions of wealth $h$ spent on healthcare in different countries.[]{data-label="fig:CalibratedEfficiency"}]({Model_Calibration/Updates/Calibrated_Efficiency_FL.pdf}){width=".8\linewidth"} The rest of the paper is organized as follows. Section \[sec:EZ\] establishes Epstein-Zin utilities over one’s random lifetime, with healthcare spending incorporated. Section \[sec:PF\] introduces the problem of optimal consumption, investment, and healthcare spending under Epstein-Zin preferences, and derives the related HJB equation and a general verification theorem. Section \[sec:Results\] presents the main results in order of complexity, characterizing optimal consumption, investment, and healthcare spending in three different settings of aging and access to healthcare. Section \[sec:calibration\] calibrates our model to mortality data in four countries, and discusses important implications. Most proofs are collected in Appendices. Epstein-Zin Preferences with Healthcare Spending {#sec:EZ} ================================================ Let $(\Omega,{\mathcal{F}}, {\mathbb{P}})$ be a probability space equipped with a filtration $\mathbb F= ({\mathcal{F}}_t)_{t\ge 0}$ that satisfies the usual conditions. Consider another probability space $(\Omega',{\mathcal{F}}', {\mathbb{P}}')$ supporting a random variable $Z$ that has an exponential law $$\label{exp law} {\mathbb{P}}'(Z>z)= e^{-z},\quad z\ge0.$$ We denote by $(\bar\Omega,\bar {\mathcal{F}}, \bar {\mathbb{P}})$ the product probability space $(\Omega\times\Omega', {\mathcal{F}}\times{\mathcal{F}}', {\mathbb{P}}\times{\mathbb{P}}')$. The expectations taken under ${\mathbb{P}}$, ${\mathbb{P}}'$, and $\bar {\mathbb{P}}$ will be denoted by ${\mathbb{E}}$, ${\mathbb{E}}'$, and $\bar{{\mathbb{E}}}$, respectively. Consider an agent who obtains utility from consumption, partially determines his lifespan through healthcare spending, and has bequest motives to leave his wealth at death to beneficiaries. Specifically, we assume that the mortality rate process $M$ of the agent evolves as $$\label{Mortality} dM_{t} = (\beta-g(h_{t}))M_{t}^{}dt,\quad M_{0} = m>0,$$ where $h=(h_t)_{t\ge 0}$, a nonnegative $\mathbb F$-progressively measurable process, represents the proportion of wealth spent on healthcare at each time $t$, while $g:{\mathbb{R}}_+\to{\mathbb{R}}_+$ is the [*efficacy function*]{} that prescribes how much the natural growth rate of mortality $\beta>0$ is reduced by healthcare spending $h_t$. For any $\bar\omega = (\omega,\omega')\in\bar\Omega$, the random lifetime of the agent is formulated as $$\label{tau} \tau(\bar\omega) := \inf\left\{t\ge 0 : \int_{0}^{t}M_{s}^{h}(\omega)ds \ge {Z}(\omega') \right\}.$$ The information available to the agent is then defined as $\mathbb G = ({\mathcal{G}}_t)_{t\ge 0}$ with $$\label{cG} {\mathcal{G}}_{t} := {\mathcal{F}}_{t}\vee{\mathcal{H}}_t,\quad \hbox{where}\quad {\mathcal{H}}_t := \sigma\left({\mathbbm{1}}_{\{\tau\le u\}}, u\in[0,t]\right).$$ That is, at any time $t$, the agent knows the information contained in ${\mathcal{F}}_t$ and whether he is still alive (i.e. whether $\tau>t$ holds); he has no further information of $\tau$, as the random variable $Z$ is inaccessible to him. Finally, we assume that the agent can leave a fraction $\zeta\in (0,1]$, not necessarily all, of his wealth at death to beneficiaries, reflecting the effect of inheritance and estate taxes. The controlled mortality , introduced by Guasoni and Huang [@Huang19], is based on the assumption that healthcare expenses affect mortality growth relative to wealth rather than in absolute terms. While this is a modeling simplification, there are empirical and theoretical justifications for it; see [@Huang19 p.319] for details. Now, let us define a non-standard Epstein-Zin utility process that incorporates healthcare spending. First, recall the Epstein-Zin aggregator $f:{\mathbb{R}}_+\times{\mathbb{R}}\to {\mathbb{R}}$ given by $$\label{EZ_aggregator} \begin{split} f(c,v)&:=\delta\frac{(1-\gamma)v}{1-\frac{1}{\psi}}\left(\bigg(\frac{c}{((1-\gamma)v)^{\frac{1}{1-\gamma}}}\bigg)^{1-\frac{1}{\psi}}-1\right)\\ &=\delta\frac{c^{1-\frac{1}{\psi}}}{1-\frac{1}{\psi}}\big((1-\gamma)v\big)^{1-\frac{1}{\theta}}-\delta\theta v,\qquad\text{with}\quad \theta := \frac{1-\gamma}{1-\frac{1}{\psi}}, \end{split}$$ where $\gamma$ and $\psi$ represent the agent’s [relative risk aversion]{} and EIS, respectively, as stated in Section \[sec:intro\]. Given a consumption stream $c=(c_t)_{t\ge 0}$, assumed to be nonnegative $\mathbb F$-progressively measurable, and a healthcare spending process $h=(h_t)_{t\ge 0}$ introduced below , we define the [*Epstein-Zin utility on the random horizon $\tau$*]{} to be a $\mathbb G$-adapted semimartingale $(\widetilde{V}_{t}^{c,h})_{t\ge 0}$ satisfying $$\label{Vtilde_Utility} \widetilde{V}_{t}^{c,h} = \bar{{\mathbb{E}}}_t\left[\int_{t\wedge\tau}^{T\wedge\tau}f(c_s,\widetilde{V}_{s}^{c,h})ds + \zeta^{1-\gamma}\widetilde{V}_{\tau-}^{c,h}{\mathbbm{1}}_{\{\tau\le T\}} + \widetilde{V}_{T}^{c,h}{\mathbbm{1}}_{\{\tau>T\}} \right],\text{ for all }0\le t\le T<\infty,$$ where we use the notation $\bar {\mathbb{E}}_{t}\left[\cdot\right]=\bar{\mathbb{E}}\left[\cdot|{\mathcal{G}}_{t}\right]$. In , we assert that the loss of wealth at death results in a decreased bequest utility, by a factor of $\zeta^{1-\gamma}$. This assertion will be made clear and justified in Section \[sec:Results\], where a financial model is in place; see Remark \[rem:confirms\] particularly. To solve for the process $\widetilde V^{c,h}$, let us first introduce a general formulation of infinite-horizon BSDEs that will be used throughout the paper. \[def: BSDE\] Let $V$ be an $\mathbb F$-progressively measurable process satisfying ${\mathbb{E}}[\sup_{s\in[0,t]}|V_{s}|]<\infty$ for all $t\ge 0$. For any $G:\Omega\times{\mathbb{R}}_{+}\times{\mathbb{R}}\to {\mathbb{R}}$ such that $\big(G(\cdot,t,V_t(\cdot))\big)_{t\ge 0}$ is $\mathbb F$-progressively measurable, we say $V$ is a solution to the infinite-horizon BSDE $$\begin{aligned} \label{BSDE} d V_{t}= -G(\omega,t,V_t) dt + d\mathscr{M}_{t},$$ if for any $T>0$ there exists an $\mathbb{F}$-martingale $(\mathscr{M}_t)_{t\in[0,T]}$ such that holds for $0\le t\le T$. Without a terminal condition, can have infinitely many solutions. Indeed, as long as $G$ admits proper monotonicity, there are solutions to that satisfy “$\lim_{t\to\infty}V_{t} = \xi$ for ${\mathcal{F}}$-measurable random variable $\xi$” or “$\lim_{t\to\infty}{\mathbb{E}}\left[e^{\rho t}V_{t}\right]\rightarrow 0$ for $\rho>0$”; see [@Briand03] and [@Fan15]. We will address this non-uniqueness issue by enforcing appropriate “terminal behavior”; see Remark \[rem:address\]. The next result shows that the $\mathbb G$-adapted $\widetilde V$ in can be expressed as a function of $\tau$ and an $\mathbb F$-adapted process $V$ that satisfies an infinite-horizon BSDE. \[prop:decompose wtV\] Let $c, h$ be nonegative $\mathbb F$-progressively measurable and $\widetilde V$ be a $\mathbb G$-adapted semimartingale, with $\bar {\mathbb{E}}[\sup_{s\in[0,t]}|\widetilde V_{s}|]<\infty$ for all $t\ge 0$, that satisfies . Then, $$\label{wtV decompose} \widetilde V_t = V_t {\mathbbm{1}}_{\{t<\tau\}} + \zeta ^{1-\gamma}V_{\tau-}{\mathbbm{1}}_{\{t\ge\tau\}}\qquad \forall t\ge 0,$$ where $V$ is an $\mathbb F$-adapted semimartingale, with ${\mathbb{E}}[\sup_{s\in[0,t]}|V_{s}|]<\infty$ for all $t\ge 0$, that satisfies the infinite-horizon BSDE $$\label{BSDE_V} dV_{t}^{} = -F(c_{t},M_{t}^{h},V_{t}^{})ds + d\mathscr{M}_{t},$$ with $F:{\mathbb{R}}_+\times{\mathbb{R}}_+\times{\mathbb{R}}\to{\mathbb{R}}$ defined by $$\label{F func} F(c,m,v):= f(c,v) - (1-\zeta^{1-\gamma})m v.$$ See Section \[subsec:proof of prop:decompose wtV\]. In view of Proposition \[prop:decompose wtV\], to uniquely determine the Epstein-Zin utility process $\widetilde V$, we need to find a suitable class of stochastic processes among which there exists a unique solution to . To this end, we start with imposing appropriate integrability and transversality conditions. \[cE\] For any $k\in{\mathbb{R}}$, define $\Lambda:= \delta\theta + (1-\theta)k$. Then, for any nonnegative $\mathbb F$-progressively measurable $h$, we denote by ${\mathcal{E}}^h_k$ the set of all $\mathbb F$-adapted semimartingales $Y$ that satisfy the following integrability and transversality conditions: $$\begin{aligned} \label{TransversalityCondition} {\mathbb{E}}\bigg[\sup\limits_{s\in[0,t]}\left|Y_{s}\right|\bigg]<\infty\ \ \forall t>0\quad\hbox{and}\quad\lim\limits_{t\rightarrow\infty}e^{-\Lambda t}{\mathbb{E}}\left[e^{-\gamma(\psi-1)\frac{1-\zeta^{1-\gamma}}{1-\gamma}\int_{0}^{t}M_{s}^{h}ds}|Y_{t}|\right] = 0. \end{aligned}$$ Condition is similar to [@Melnyk17 (2.3)], but the controlled mortality $M^h$ in our case complicates the transversality condition: unlike [@Melnyk17 (2.3)], the exponential term no longer contains a constant rate, but a stochastic one involving $M^h$. This adds nontrivial complexity to deriving a comparison result (Proposition \[prop:comparison\]) and the use of verification arguments (Theorem \[Thm:AH\]). \[rem:negative Lambda\] The constant $\Lambda:= \delta\theta + (1-\theta)k$ in can be negative, even when $k>0$ (as will be assumed in Section \[sec:Results\]). In such a case, stipulates that $M^h$ must increase fast enough to neutralize the growth of $e^{-\Lambda t}$, such that the transversality condition can be satisfied. We now introduce the appropriate collection of strategies $(c,h)$ we will focus on. \[k-admissible\] Let $c, h$ be nonnegative $\mathbb F$-progressively measurable. For any $k\in{\mathbb{R}}$, we say $(c,h)$ is $k$-admissible if there exists $V\in {\mathcal{E}}^h_k$ satisfying and $$\label{comparison_condition} V_{s}\le \delta^{\theta}\left(k+(\psi-1)\frac{1-\zeta^{1-\gamma}}{1-\gamma} M^h_{s}\right)^{-\theta}\frac{c_{s}^{1-\gamma}}{1-\gamma},\quad \forall s\ge 0.$$ \[rem:address\] Condition is the key to a comparison result for , as shown in Proposition \[prop:comparison\] below. In a sense, - is the enforced “*terminal behavior*", under which a solution to can be uniquely identified. Technically, is similar to typical conditions imposed for infinite-horizon BSDEs, such as [@Briand03 (H1’)] and the one in [@Fan15 Theorem 5.1]: all of them require the solution to be bounded from above by a tractable process. Moreover, for classical Epstein-Zin utilities (without healthcare), a similar condition was imposed in [@Melnyk17 (2.5)]. In fact, Definition \[k-admissible\] is in line with [@Melnyk17 Definition 2.1], but adapted to include the controlled mortality $M^h$. A comparison result for BSDE can now be established. \[prop:comparison\] Let $k\in{\mathbb{R}}$ and $c, h$ be nonnegative $\mathbb F$-progressively measurable processes. Suppose that $V^1\in{\mathcal{E}}_{k}^{h}$ is a solution to and $V^2\in{\mathcal{E}}_{k}^{h}$ is a solution to . If $V^{1}$ satisfies and $F(c_{t},M_{t},V_{t}^2) \le G(t,V_{t}^2)\ d{\mathbb{P}}\times dt$-a.e., then $V_{t}^{1}\le V_{t}^{2}$ for $t\ge 0$ ${\mathbb{P}}$-a.s. See Section \[subsec:proof of comparison\]. The next result is a direct consequence of Propositions \[prop:decompose wtV\] and \[prop:comparison\]. \[thm:Existence and Uniqueness\] Fix $k\in{\mathbb{R}}$. For any $k$-adimissible $(c,h)$, there exists a unique solution $V^{c,h}\in{\mathcal{E}}^h_k$ to that satisfies . Hence, the Epstein-Zin utility $\widetilde{V}^{c,h}$ can be uniquely determined via . Problem Formulation {#sec:PF} =================== Let $B=(B_t)_{t\ge 0}$ be an $\mathbb F$-adapted standard Brownian motion. Consider a financial market with a riskfree rate $r>0$ and a risky asset $S_{t}$ given by $$\begin{aligned} \label{S} dS_{t} & = (\mu+r)S_{t}dt + \sigma S_{t}dB_{t},\end{aligned}$$ where $\mu\in{\mathbb{R}}$ and $\sigma>0$ are given constants. Given initial wealth $x>0$, at each time $t\ge0$, an agent consumes a lump-sum $c_{t}$ of his wealth, invests a fraction $\pi_t$ of his wealth on the risky asset, and spends another fraction $h_{t}$ on healthcare. The resulting dynamics of the wealth process $X$ is $$\begin{aligned} \label{Wealth} dX_{t} &= X_{t}\left(r+\mu\pi_{t}-h_{t}\right)dt -c_{t} dt + X_{t}\sigma\pi_{t}dB_{t},\quad X_0 =x.\end{aligned}$$ \[H\_k\] For all $k\in{\mathbb{R}}$, let ${\mathcal{H}}_k$ be the set of strategies $(c,\pi,h)$ such that $(c,h)$ is $k$-admissible (Definition \[k-admissible\]), $\pi$ is $\mathbb F$-progressively measurable, and a unique solution $X^{c,\pi,h}$ to exists. The agent aims to maximize his lifetime Epstein-Zin utility $\widetilde V_0^{c,h}$ by choosing $(c,\pi,h)$ in a suitable collection of strategies ${\mathcal{P}}$, i.e. $$\begin{aligned} \label{problem} \sup\limits_{(c,\pi,h)\in{\mathcal{P}}}\widetilde{V}_{0}^{c,h} &=\sup\limits_{(c,\pi,h)\in{\mathcal{P}}}V_{0}^{c,h},\end{aligned}$$ where the equality follows from . In this section, we only require ${\mathcal{P}}$ to satisfy $$\label{cP condition} {\mathcal{P}}\subseteq {\mathcal{H}}_k\quad \hbox{for some}\ k\in{\mathbb{R}}.$$ A more precise definition of ${\mathcal{P}}$, depending on the specification of $\beta$, $\gamma$, and $\zeta$, will be introduced in Definition \[cP\] below. The focus of this section is to derive the Hamilton-Jacobi-Bellman (HJB) equation for and establish a versatile verification theorem under the general condition . A General Verification Theorem {#DPP} ------------------------------ Under the current Markovian setting (i.e. and ), we take $$\label{problem'} v(x,m) : = \sup\limits_{(c,\pi,h)\in{\mathcal{P}}}V_{0}^{c,h}.$$ Namely, the optimal value should be a function of the current wealth and mortality. The relation , derived from , suggests the following dynamic programming principle (DPP): With the shorthand notation ${\mathfrak{p}}= (c,\pi,h)$ and ${\mathfrak{p}}_s = (c_s,\pi_s,h_s)$ for $s\ge0$, for any $T>0$, $$\begin{aligned} \label{ValueFn} &v(x,m) =\nonumber\\ &\sup\limits_{{\mathfrak{p}}\in{\mathcal{P}}}{\mathbb{E}}\left[\int_{0}^{T}e^{-\int_{0}^{s}M_{r}^{h}dr}\left(f(c_{s},v(X_{s}^{{\mathfrak{p}}},M_{s}^{h})) + \zeta^{1-\gamma}M_{s}^{h}v(X_{s}^{{\mathfrak{p}}},M_{s}^{h})\right)ds + e^{-\int_{0}^{T}M_{s}ds}v(X_{T}^{{\mathfrak{p}}},M_{T}^{h}) \right].\end{aligned}$$ By applying Itô’s formula to $e^{-\int_{0}^{t}M_{s}^{h}ds}v(X_{t}^{{\mathfrak{p}}},M_{t}^{h})$, assuming enough regularity of $v$, we get $$\begin{aligned} &e^{-\int_{0}^{T}M_{s}^{h}ds}v(X_{T}^{{\mathfrak{p}}},M_{T}^{h}) - v(x,m)\\ &= \int_{0}^{T}\left(L^{{\mathfrak{p}}_s}[v](X_{t}^{{\mathfrak{p}}},M_{t}^{h})dt - M_{t}^{h}v(X_{t}^{{\mathfrak{p}}},M_{t}^{h})\right)dt + \int_{0}^{T}e^{-\int_{0}^{t}M_{s}^{h}ds}\sigma\pi X_{t}^{{\mathfrak{p}}}v_{x}(X_{t}^{{\mathfrak{p}}},M_{t}^{h})dB_{t},\end{aligned}$$ where the operator $L^{a,b,d}[\cdot]$ is defined by $$\label{L} L^{a,b,d}[\kappa](x,m):= \left((r+\mu b-d)x-a\right) \kappa_{x}(x,m)+(\beta-g(d))m \kappa_{m}(x,m) + \frac{1}{2}\sigma^{2}b^{2}x^{2}\kappa_{xx}(x,m),$$ for any $\kappa\in C^{2,1}({\mathbb{R}}_+\times{\mathbb{R}}_+)$. We can then rewrite as $$0 = \sup\limits_{{\mathfrak{p}}\in{\mathcal{P}}}{\mathbb{E}}\left[\int_{0}^{T}e^{-\int_{0}^{s}M_{t}^{h}dt}\left(f(c_{s},v(X_{s}^{{\mathfrak{p}}},M_{s}^{h}))ds + (\zeta^{1-\gamma}-1)M_{s}^{h} v(X_{s}^{{\mathfrak{p}}},M_{s}^{h})+L^{{\mathfrak{p}}_s}[v](X_{s}^{{\mathfrak{p}}},M_{s}^{h})\right)ds\right].$$ The HJB equation associated with $v(x,m)$ is then $$\begin{aligned} \label{HJB} 0 &= \sup\limits_{c\in{\mathbb{R}}_{+}}\left\{f(c,w(x,m)) - cw_{x}(x,m)\right\} +\sup\limits_{h\in{\mathbb{R}}_{+}}\left\{-g(h)mw_{m}(x,m) -hxw_{x}(x,m)\right\}\nonumber\\ &\hspace{0.2in}+ \sup\limits_{\pi\in{\mathbb{R}}}\left\{\mu\pi xw_{x}(x,m) + \frac{1}{2}\sigma^{2}\pi^{2}x^{2}w_{xx}(x,m)\right\}\\ &\hspace{0.2in}+ rxw_{x}(x,m) +\beta mw_{m}(x,m) + (\zeta^{1-\gamma}-1)m w(x,m),\quad \forall (x,m)\in{\mathbb{R}}_{+}^2.\nonumber\end{aligned}$$ Equivalently, this can be written in the more compact form $$\label{HJB'} \sup\limits_{c, h\in{\mathbb{R}}_+, \pi\in{\mathbb{R}}}\left\{L^{c,\pi,h}[w](x,m) + f(c,w(x,m))\right\} + (\zeta^{1-\gamma}-1)m w(x,m)= 0,\quad \forall (x,m)\in{\mathbb{R}}_{+}^2.$$ We establish a general verification theorem for $v(x,m)$ in . \[Verification\] Let $w\in C^{2,1}({\mathbb{R}}_{+}\times{\mathbb{R}}_+)$ be a solution to , and ${\mathcal{P}}$ satisfy . Suppose that for any $(c,\pi,h)\in{\mathcal{P}}$, the process $w(X_{t}^{c,\pi,h},M_{t}^{h})$, $t\ge 0$, belongs to ${\mathcal{E}}^h_k$ (where $k\in{\mathbb{R}}$ is specified by ) and $$\label{veri condition} {\mathbb{E}}\bigg[\sup_{s\in[0,t]} \pi_s X^{c,\pi,h}_s w_x(X_{s}^{c,\pi,h},M_{s}^{h})\bigg]<\infty,\quad \forall t>0.$$ Then, the following holds. - $w(x,m)\ge v(x,m)$ on ${\mathbb{R}}_{+}\times {\mathbb{R}}_+$. - Suppose further that there exist Borel measurable functions $\bar c, \bar\pi, \bar h:{\mathbb{R}}^2_+\to{\mathbb{R}}$ such that $\bar c(x, m)$, $\bar\pi(x,m)$, and $\bar h(x, m)$ are maximizers of $$\begin{aligned} &\sup\limits_{c\in{\mathbb{R}}_{+}}\left\{f(c,w(x,m)) - cw_{x}(x,m)\right\},\quad \sup\limits_{\pi\in{\mathbb{R}}}\left\{\mu\pi xw_{x}(x,m) + \frac{1}{2}\sigma^{2}\pi^{2}x^{2}w_{xx}(x,m)\right\},\label{sup1, 2}\\ &\hspace{1.2in} \sup\limits_{h\in{\mathbb{R}}_{+}}\left\{-g(h)mw_{m}(x,m) -hxw_{x}(x,m)\right\},\label{sup3}\end{aligned}$$ respectively, for all $(x,m)\in{\mathbb{R}}^2_+$. If $(c^*,\pi^*,h^*)$ defined by $$\label{optimal strategies} c^*_t:= \bar c(X_t,M_t),\quad \pi^*_t:= \bar \pi(X_t,M_t),\quad h^*_t:=\bar h(X_t,M_t),\qquad t\ge 0,$$ belongs to ${\mathcal{P}}$ and $W^*_t:=w(X_{t}^{c^*,\pi^*,h^*},M_{t}^{h^*})$ satisfies (with $V$, $c$, $h$ replaced by $W^*$, $c^*$, $h^*$), then $(c^*,\pi^*,h^*)$ optimizes and $w(x,m)=v(x,m)$ on ${\mathbb{R}}_{+}\times {\mathbb{R}}_+$. \(i) Fix $(x,m)\in{\mathbb{R}}^2_+$. Consider an arbitrary ${\mathfrak{p}}=(c,\pi,h)\in{\mathcal{P}}$. For any $T\ge0$ and $t\in[0,T]$, by applying Itô’s formula to $w(X^{{\mathfrak{p}}}_s,M^h_s)$, we get $$\begin{aligned} w(X_{T}^{{\mathfrak{p}}},M_{T}^{h}) = w(X_{t}^{{\mathfrak{p}}},M_{t}^{h})+\int_{t}^{T}L^{{\mathfrak{p}}_{s}}[w](X_{s}^{{\mathfrak{p}}},M_{s}^{h})ds +\int_{t}^{T}\sigma\pi_{s}X_{s}^{{\mathfrak{p}}}w_{x}(X_{s}^{{\mathfrak{p}}},M_{s}^{h})dB_{s}, \end{aligned}$$ where the operator $L^{a,b,d}[\cdot]$ is defined in . Thanks to , $u\mapsto \int_{t}^{u}\sigma\pi_{s}X_{s}^{{\mathfrak{p}}}w_{x}(X_{s}^{{\mathfrak{p}}},M_{s}^{h})dB_{s}$ is a true martingale. Hence, the above equality shows that $W_s:= w(X_{s}^{{\mathfrak{p}}},M_{s}^{h})$ is a solution to BSDE , with $G(\omega, s, v) := -L^{{\mathfrak{p}}_{s}(\omega)}[w](X_{s}^{{\mathfrak{p}}}(\omega),M_{s}^{h}(\omega))$. On the other hand, implies that $(c,h)$ is $k$-admissible, so that there exists a unique solution $V^{c,h}\in{\mathcal{E}}^h_k$ to that satisfies (Theorem \[thm:Existence and Uniqueness\]). Since $w$ is a solution to , and equivalently to , we have $$\label{verif inequality} F(c_s,M^h_s, W_s) = f(c_{s},W_{s}) + (\zeta^{1-\gamma}-1)M_{s}^{h}W_{s} \le -L^{{\mathfrak{p}}_{s}}[w](X_{s}^{{\mathfrak{p}}},M_{s}^{h}).$$ We then conclude from Proposition \[prop:comparison\] that $W_{t}\ge V_{t}^{c,h}$ for all $t\ge0$. In particular, $w(x,m)=W_0\ge V_0^{c,h}$. By the arbitrariness of $(c,\pi,h)\in{\mathcal{P}}$, $w(x,m)\ge \sup_{(c,\pi,h)\in{\mathcal{P}}}V_0^{c,h}=v(x,m)$, as desired. \(ii) Fix $(x,m)\in{\mathbb{R}}^2_+$. If $(c^*,\pi^*,h^*)\in{\mathcal{P}}$, we can repeat the arguments in part (a), obtaining with the inequality replaced by equality. This shows that $W^*_t=w(X^{c^*,\pi^*,h^*}_t,M^{h^*}_t)\in {\mathcal{E}}^{h^*}_k$ is a solution to . Also, implies that $(c^*,h^*)$ is $k$-admissible, so that there is a unique solution $V^{c^*,h^*}\in{\mathcal{E}}^{h^*}_k$ to satisfying (Theorem \[thm:Existence and Uniqueness\]). As $W^*$ also satisfies , we must have $W^*_t= V^{c^*,h^*}_t$ for all $t\ge 0$; particularly, $w(x,m)= W^*_0 = V^{c^*,h^*}_0$. With $w(x,m)\ge \sup_{(c,\pi,h)\in{\mathcal{P}}}V_0^{c,h}=v(x,m)$ in part (a), we conclude $w(x,m)=v(x,m)$ and $(c^*,\pi^*,h^*)\in{\mathcal{P}}$ is an optimal control. Reduction to an Ordinary Differential Equation ---------------------------------------------- If we assume heuristically that $w_{xx}<0$, $w_{m}<0$, $g$ is differentiable, and the inverse of $g'$ is well-defined, then the optimizers stated in Theorem \[Verification\] (ii) can be uniquely determined as $$\label{candidates} \begin{split} \bar c(x,m) &= \delta^{\psi}\frac{\left[(1-\gamma)w(x,m)\right]^{\psi(1-\frac{1}{\theta})}}{w_{x}(x,m)^{\psi}},\quad \bar\pi(x,m) = - \frac{\mu}{\sigma^{2}}\frac{w_{x}(x,m)}{xw_{xx}(x,m)},\\ &\hspace{0.5in}\bar h(x,m) = (g')^{-1}\left(-\frac{xw_{x}(x,m)}{mw_{m}(x,m)}\right). \end{split}$$ Plugging these into yields $$\begin{aligned} \label{HJB_S} 0 &= \frac{\delta^{\psi}}{\psi-1}\frac{\left[(1-\gamma)v(x,m)\right]^{\psi(1-\frac{1}{\theta})}}{v_{x}(x,m)^{\psi-1}} - \delta\theta v(x,m)- \frac{1}{2}\left(\frac{\mu}{\sigma}\right)^{2}\frac{v_{x}(x,m)^{2}}{v_{xx}(x,m)} + rxv_{x}(x,m) +\beta mv_{m}(x,m)\nonumber\\ &\hspace{0.2in} + (\zeta^{1-\gamma}-1)mv(x,m)- mv_{m}(x,m)\sup\limits_{h\in{\mathbb{R}}_{+}}\left\{g(h) +\frac{hxv_{x}(x,m)}{mv_{m}(x,m)}\right\}.\end{aligned}$$ Using the ansatz $w(x,m) = \delta^{\theta}\frac{x^{1-\gamma}}{1-\gamma}u(m)^{-\frac{\theta}{\psi}}$, the above equation reduces to $$\begin{aligned} \label{ODE} 0 &= u(m)^{2}-\tilde{c}_{0}(m) u(m)-\beta mu'(m) + mu'(m)\sup\limits_{h\in{\mathbb{R}}_{+}}\left\{g(h) -(\psi-1)\frac{u(m)}{ mu'(m)}h\right\},\quad m>0,\end{aligned}$$ where $$\begin{aligned} \label{c_0} \tilde{c}_{0}(m) &:= \psi\delta + (1-\psi)\bigg(\frac{(\zeta^{1-\gamma}-1)m}{1-\gamma}+r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2}\bigg). $$ Moreover, the maximizers in now become $$\label{bar's} \bar c(x,m) = x u(m),\quad \bar \pi \equiv \frac{\mu}{\gamma\sigma^{2}}, \quad \bar h(m) = (g')^{-1}\left((\psi-1)\frac{u(m)}{mu'(m)}\right).$$ These maximizers indeed characterize optimal consumption, investment, and healthcare spending, as will be shown in the next section. The Main Results {#sec:Results} ================ In this section, we present the main results progressively: Section \[beta=0, g=0\] deals with the simplest case with neither aging nor healthcare; Section \[beta&gt;0, g=0\] handles the scenario with aging but without healthcare, which serves as a baseline for the general model with both aging and healthcare in Section \[beta&gt;0, g&gt;0\]. Let us now formulate the set $\mathcal P$ of [*permissible*]{} strategies $(c,\pi,h)$ in the optimization problem . First, take $k\in{\mathbb{R}}$ in Definition \[cE\] to be $$\label{k^*} k^*:= \delta\psi + (1-\psi)\bigg(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2}\bigg),$$ so that $\Lambda\in{\mathbb{R}}$ in Definition \[cE\] becomes $$\begin{aligned} \label{Lambda^*} \Lambda^* := \delta\theta + (1-\theta)k^* =\delta\gamma\psi + (1-\gamma\psi)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2}\right).\end{aligned}$$ Let ${\mathcal{P}}_1$ the set of strategies $(c,\pi,h)$ such that $(c,\pi,h)\in {\mathcal{H}}_{k^*}$, $(X^{c,\pi,h})^{1-\gamma}$ satisfies (with $\Lambda\in{\mathbb{R}}$ therein taken to be $\Lambda^*$) as well as ${\mathbb{E}}\big[\sup_{s\in[0,t]} \pi_s(X^{c,\pi,h}_s)^{1-\gamma}\big]<\infty$ for $t\ge 0.$ Let ${\mathcal{P}}_2$ be defined as ${\mathcal{P}}_1$, except that the second part of is replaced by $$\label{Permissible} \lim_{t\rightarrow\infty}e^{-\Lambda^* t}{\mathbb{E}}\left[e^{-\eta\gamma(\psi-1)\frac{1-\zeta^{1-\gamma}}{1-\gamma}\int_{0}^{t}M_{s}^{h}ds}(X_{t}^{c,\pi,h})^{1-\gamma}\right] = 0,\quad \hbox{for some $\eta\in(1-\frac{1}{\gamma},1)$}.$$ \[cP\] The set of [*permissible*]{} strategies $(c,\pi,h)$, denoted by $\mathcal P$, is defined as follows. - For the case $\beta=0$ and $g\equiv 0$ (i.e. with neither aging nor healthcare), ${\mathcal{P}}:={\mathcal{P}}_1$; - For the case $\beta>0$ (i.e. with aging), $${\mathcal{P}}:= \begin{cases} {\mathcal{P}}_1,\quad &\hbox{if}\ \gamma\in\big(\frac{1}{\psi},1\big)\ \hbox{or}\ \zeta= 1,\\ {\mathcal{P}}_2,\quad &\hbox{if}\ \gamma> 1\ \hbox{and}\ \zeta\in(0,1), \end{cases}$$ When there is aging ($\beta>0$), for the case $\gamma> 1$ and $\zeta\in(0,1)$, we need $(X^{c,\pi,h})^{1-\gamma}$ to satisfy the slightly stronger condition (than the transversality condition in ), so that the general verification Theorem \[Verification\] can be applied; see Appendix \[subsec:proof of Thm:AH\] for details. Neither Aging nor Healthcare {#beta=0, g=0} ---------------------------- When the natural growth rate of mortality is zero ($\beta=0$) and healthcare is unavailable ($g\equiv 0$), the mortality process is simply constant, i.e. $M_{t}\equiv m$. Consequently, in the HJB equation , the derivatives in $m$ should all vanish; moreover, as $v(x,m)$ is nondecreasing in $x$ by definition, the second supremum in should also be zero. Corresponding to this largely simplified HJB equation, reduces to $$0 = u(m)^{2}-\tilde{c}_{0}(m) u(m),$$ which directly implies $u(m)=\tilde c_0(m)$. The problem can then be solved explicitly. \[prop:NoAging\] Assume $\beta=0$ and $g\equiv 0$. For any $m\ge 0$, if $\tilde{c}_{0}(m)$ defined in is strictly positive, then $$v(x,m) = \delta^{\theta}\frac{x^{1-\gamma}}{1-\gamma}\tilde{c}_{0}(m)^{-\frac{\theta}{\psi}}\quad \hbox{for $x> 0$.}$$ Furthermore, $c_{t}^{*}:= \tilde{c}_{0}(m)X_t$, $\pi^{*}_{t}:= \frac{\mu}{\gamma\sigma^{2}}$, and $h^*_t:= 0$, for $t\ge 0$, form an optimal control for . See Section \[subsec:proof of prop:NoAging\]. Proposition \[prop:NoAging\] shows that in the absence of aging and healthcare, the optimal investment rate is the classical Merton’s proportion, while the optimal consumption rate is the constant $\tilde{c}_{0}(m)$, dictated by the fixed mortality $m$. By , for the case $\zeta=1$, $\tilde c_0(m)\equiv \psi\delta + (1-\psi)\big(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2}\big) $ no longer depends on $m$. Indeed, when there is no loss of wealth (and thus utility) at death, dying sooner or later (i.e. how large $m$ is) does not make a difference to an agent who maximizes lifetime utility plus bequest utility at death. Note that for the specific case $\psi=1/\gamma$, Proposition \[prop:NoAging\] reduces to [@Huang19 Proposition 3.1 and Theorem 4.1] under time-separable utilities. As $\frac{\zeta^{1-\gamma}-1}{1-\gamma}<0$ for all $0<\gamma\neq 1$, we observe from that a larger mortality rate $m$ induces a larger consumption rate due to EIS $\psi>1$. This can be explained by the usual substitution effect in response to negative wealth shocks: a larger mortality rate encourages the agent to consume more (i.e. consumption substitutes for saving) before the loss of wealth at death. Aging without Healthcare {#beta>0, g=0} ------------------------ When the natural growth of mortality is positive ($\beta>0$) but healthcare is unavailable ($g\equiv 0$), the mortality grows exponentially, i.e. $M_{t}=me^{\beta t}$, consistently with the Gompertz law. Thanks to $g\equiv 0$ and $v(x,m)$ being nondecreasing in $x$ by definition, the second supremum in the HJB equation vanishes. Corresponding to this, reduces to $$\label{ODE'} 0 = u(m)^{2}-\tilde{c}_{0}(m) u(m)-\beta mu'(m),\quad m>0.$$ The next result shows that this type of differential equations can be solved explicitly. \[ANH\_ODE\] Fix $q>0$, and define the function $u_q:{\mathbb{R}}_+\to{\mathbb{R}}_+$ by $$\label{u_q} u_{q}(m) := \left(\frac{1}{q}\int_{0}^{\infty}e^{\frac{\psi-1}{q(1-\gamma)}(\zeta^{1-\gamma}-1)my}(y+1)^{-\left(1+\frac{k^*}{q}\right)}dy\right)^{-1}.$$ If $k^*$ in is strictly positive, then $u_q$ is the unique solution to the ordinary differential equation $$\label{ANHp_ODE} 0= u^{2}(m)-\tilde{c}_{0}(m)u(m)-q m u'(m),\quad\forall m >0,$$ such that $\lim_{q\rightarrow0}u_{q}(m) = \tilde{c}_{0}(m)$. Moreover, $u_q$ satisfies - $u_{q}(0) = \tilde{c}_{0}(0) = k^*>0$, $\lim_{m\rightarrow\infty}\left[u_{q}(m) - (\tilde{c}_{0}(m) + q)\right] = 0$, and $$\label{ANHbound} \tilde{c}_{0}(m)< u_{q}(m)<\tilde{c}_{0}(m) + q,\quad \forall m>0.$$ - $u_{q}'(0+) = \infty$, $u_{q}'(\infty) = (\psi-1)\frac{1-\zeta^{1-\gamma}}{{1-\gamma}}$. As the results follow from analogous arguments in [@Huang19], we only sketch the proof. First, similarly to (A.8) in [@Huang19], admits the general solution $$u(m) = q e^{\frac{\psi}{\theta q}(\zeta^{1-\gamma}-1)m}\left(C\beta m^{\frac{k}{\beta}}+ \int_{1}^{\infty}e^{\frac{\psi}{\theta q}(\zeta^{1-\gamma}-1)mv}v^{-(1+\frac{k}{q})}dv\right)^{-1},\quad \hbox{with}\ C\in{\mathbb{R}}.$$ To ensure $\lim_{q\rightarrow0}u(m) = \tilde{c}_{0}(m)$, we need $C=0$, which identifies the corresponding solution as $$\begin{aligned} u_{q}(m) &= q e^{\frac{\psi}{\theta q}(\zeta^{1-\gamma}-1)m}\left( \int_{1}^{\infty}e^{\frac{\psi}{\theta q}(\zeta^{1-\gamma}-1)mv}v^{-(1+\frac{k}{q})}dv\right)^{-1}. \end{aligned}$$ A straightforward change of variable then gives the formula . Now, replacing the positive constants $\frac{\delta + (\gamma-1)r}{\gamma}$, $\beta$, and $\frac{1-\zeta^{1-\gamma}}{\gamma}$ in [@Huang19 Lemma A.1] by $k^*$, $q$, and $-\frac{\psi-1}{1-\gamma}(\zeta^{1-\gamma}-1)$ in our setting, we immediately obtain the remaining assertions. Relying on Lemma \[ANH\_ODE\], the problem can be solved. \[prop:Aging\] Assume $\beta>0$ and $g\equiv 0$. If $k^*$ defined in is strictly positive, then $$v(x,m) = \delta^{\theta}\frac{x^{1-\gamma}}{1-\gamma}u_\beta(m)^{-\frac{\theta}{\psi}},\qquad (x,m)\in {\mathbb{R}}_+^2,$$ where $u_{\beta}:{\mathbb{R}}_+\to{\mathbb{R}}_+$ is defined as in , with $q=\beta$. Furthermore, $c^*_t:= u_\beta(me^{\beta t})X_t$, $\pi^*_t := \frac{\mu}{\gamma\sigma^{2}}$, and $h^*_t:=0$, for $t\ge 0$, form an optimal control for . See Section \[subsec:proof of prop:Aging\]. Proposition \[prop:Aging\], together with Lemma \[ANH\_ODE\], admits interesting implications. First, we note from and that $$\label{c_0=k^*+...} \tilde c_0(m)= k^* + (\psi-1)\frac{(1-\zeta^{1-\gamma})m}{1-\gamma}.$$ As $\psi>1$ and $\frac{1-\zeta^{1-\gamma}}{1-\gamma}>0$ for all $0<\gamma\neq 1$, the condition $k^*>0$ ensures $\tilde c_0(m)>0$ for all $m>0$. This, together with $u_\beta>\tilde c_0$ ( with $q=\beta$), shows that $k^*>0$ in Proposition \[prop:Aging\] is essentially a well-posedness condition, which guarantees that the optimal consumption rate $u_\beta(me^{\beta t})$ is strictly positive for all $t\ge 0$. Moreover, with $q=\beta$, Lemma \[ANH\_ODE\] (a) stipulates that aging enlarges consumption rate, but the increase does not exceed the growth of aging $\beta>0$; in addition, this upper bound is asymptotically reached as mortality increases indefinitely. Note that the increase in consumption results from the same substitution effect as discussed below Proposition \[prop:NoAging\]. Lemma \[ANH\_ODE\] (b) further describes how the optimal consumption rate increases with mortality: as the agent is young (i.e. $m$ is small), it increases steeply; as the agent is old (i.e. $m$ is large), it grows asymptotically linearly, with the same slope $(\psi-1)\frac{1-\zeta^{1-\gamma}}{{1-\gamma}}$ as in the case without aging. Note that for the specific case $\psi=1/\gamma$, Proposition \[prop:Aging\] reduces to [@Huang19 Proposition 3.2 and Theorem 4.1] under time-separable utilities. Aging and Healthcare {#beta>0, g>0} -------------------- For the general case where the natural growth of mortality is positive ($\beta>0$) and healthcare is available ($g\not\equiv 0$), we need to deal with the equation in its full complexity. To this end, we impose the following condition on $g$. \[assump:AH\] Let $g:{\mathbb{R}}_{+}\to{\mathbb{R}}_{+}$ be twice differentiable with $g(0)=0$, $g'(h)>0$ and $g''(h)<0$ for $h>0$, and satisfies the Inada condition $$g'(0+) = \infty\quad \text{and}\quad g'(\infty) = 0,$$ as well as $$\label{g<beta} g\left(I\left(\psi-1\right)\right)<\beta\quad \text{with}\quad I:= (g')^{-1}.$$ Condition was first introduced in [@Huang19]. Its purpose will be made clear after the optimal healthcare spending strategy $h^*$ is introduced in Theorem \[Thm:AH\]; see Remark \[rem:why g&lt;beta\]. \[ODE\_Solution\] Suppose Assumption \[assump:AH\] holds. If $k^*$ in is strictly positive, there exists a unique nonnegative, strictly increasing, strictly concave, classical solution $u^*:{\mathbb{R}}_+\to{\mathbb{R}}_+$ to . Furthermore, define $$\underline{\beta}:= \beta - \sup\limits_{h\ge0}\left\{g(h)-(\psi-1)h\right\}\in(0,\beta). $$ Then, $$\label{u^* bounds} u_{\underline{\beta}}(m)\le u^*(m)\le \min\{u_{\beta}(m),\tilde{c}_{0}(m)+\underline{\beta}\}\quad \forall m>0,$$ and $\lim_{m\rightarrow\infty}\left[u^*(m) - (\tilde{c}_{0}(m) + \underline\beta)\right] = 0$. Replacing the positive constants $\frac{1-\gamma}{\gamma}$, $\frac{\delta + (1-\gamma)r}{\gamma}$, and $\frac{1-\zeta^{1-\gamma}}{\gamma}$ in [@Huang19 Appendix A.3] (particularly Theorems 3.1 and 3.2 therein) by $\psi-1$, $k^*$, and $ -\frac{\psi-1}{1-\gamma}(\zeta^{1-\gamma}-1)$ in our setting, we immediately get the desired results. \[Thm:AH\] Suppose Assumption \[assump:AH\] holds. If $k^*$ in is strictly positive, then $$\label{v's form} v(x,m) = \delta^{\theta}\frac{x^{1-\gamma}}{1-\gamma}u^*(m)^{-\frac{\theta}{\psi}},\quad (x,m)\in{\mathbb{R}}^2_+,$$ where $u^*:{\mathbb{R}}_{+}\to{\mathbb{R}}_{+}$ is the unique nonnegative, strictly increasing, strictly concave, classical solution to . Furthermore, $(c^*,\pi^*,h^*)$ defined by $$c^{*}_{t} := u^*(M_{t})X_t,\quad \pi_{t}^{*}:= \frac{\mu}{\gamma\sigma^{2}},\quad h^{*}_{t}:= (g')^{-1}\left((\psi-1)\frac{u^{*}(M_{t})}{M_{t}(u^*)'(M_{t})}\right),\qquad t\ge 0$$ is an optimal control for . See Section \[subsec:proof of Thm:AH\]. Theorem \[Thm:AH\] identifies the marginal efficacy of optimal healthcare spending, $g'(h^*_t)$, to be inversely proportional to $\frac{m(u^*)'(m)}{u^*(m)}$, the elasticity of consumption with respect to mortality, where the constant of proportionality depends on EIS $\psi$. Note that a larger EIS implies less healthcare spending, as $(g')^{-1}$ is strictly decreasing. In a sense, healthcare spending is like saving: it crowds out current consumption, but potentially enlarges future consumption by extending one’s lifetime. Since a larger EIS means a stronger substitution effect (as discussed below Proposition \[prop:NoAging\]), one substitutes more consumption for saving-like healthcare spending with a larger $\psi$. Although the optimal consumption rate $u^*(m)$, the solution to , does not admit an explicit formula, it does have simple upper and lower bounds thanks to . The lower bound is $u_{\underline{\beta}}(m)$, the consumption rate in a model where healthcare is unavailable, but mortality grows at the lower rate $\underline\beta<\beta$, low enough that the agent would be willing to give up access to healthcare in exchange for such a slower natural growth of mortality. The upper bound is the minimum between the consumption rate in the case of aging without healthcare (i.e. $u_\beta(m)$ in Section \[beta&gt;0, g=0\]), and the consumption rate in the case of neither aging nor healthcare (i.e. $\tilde c_0(m)$ in Section \[beta=0, g=0\]) plus the adjusted growth rate $\underline\beta$. These bounds for $u^*$ are crucial for verification arguments in the proof of Theorem \[Thm:AH\], as well as the calibration in Section \[sec:calibration\]. Note that for the specific case $\psi=1/\gamma$, Theorem \[Thm:AH\] reduces to [@Huang19 Theorems 3.4 and 4.1] under time-separable utilities. \[rem:why g&lt;beta\] As the same argument in [@Huang19 Lemma A.2] implies $\frac{u^*(m)}{ m(u^{*}(m))'}\ge1$ for $m>0$, $$\label{g(h^*)<beta} g(h^*_t) = g\left( I\left((\psi-1)\frac{u^{*}(M_{t})}{M_{t}(u^*)'(M_{t})}\right) \right)\le g(I(\psi-1)) <\beta,$$ where the last inequality is due to . In other words, stipulates that optimizing healthcare spending can only reduce, but not reverse, the growth of mortality. \[rem:confirms\] Since the transferred wealth at death is $\zeta X^{c^*,\pi^*,h^*}_{\tau-}$, indicates that the resulting bequest utility is $$\delta^{\theta}\frac{(\zeta X^{c^*,\pi^*,h^*}_{\tau-})^{1-\gamma}}{1-\gamma}u^*(M^{h^*}_{\tau-})^{-\frac{\theta}{\psi}} = \zeta^{1-\gamma} v(X^{c^*,\pi^*,h^*}_{\tau-}, M^{h^*}_{\tau-}),$$ i.e. the loss of wealth at death reduces utility by a factor of $\zeta^{1-\gamma}$, confirming the setup in . Calibration and Implications {#sec:calibration} ============================ In this section, we calibrate the model in Section \[beta&gt;0, g&gt;0\] to mortality data in various countries. We focus on examining (i) whether our model properly explains the change of mortality over time, and (ii) the shape of the efficacy function $g$ and its implications to a country’s healthcare quality. We take as given the following values: $r=1\%$, $\delta=3\%$, $\psi=1.5$, $\gamma=2$, $\zeta=50\%$, $\mu = 5.2\%$, and $\sigma = 15.4 \%$. A safe rate of $r=1\%$ approximates the long-term average real rate on Treasury bills reported in [@beeler2012long], and the time preference $\delta=3\%$ is also consistent with estimates therein; $\psi=1.5$ corresponds to the estimate obtained in [@Bansal04]; $\gamma=2$ follows the specification in [@Kraft17] and [@Xing]; market parameters $\mu = 5.2\%$ and $\sigma = 15.4 \%$ are taken from the long-term estimates in [@Imperial18]; $\zeta=50\%$ is a rough estimate of inheritance and estate taxes in developed countries. Note that these values ensure $k^*>0$ in . On the other hand, the efficacy function $g:{\mathbb{R}}_+\to{\mathbb{R}}_+$ is taken to be $$\label{g'} g(z) = a\frac{z^{q}}{q},\quad \hbox{with}\ a>0\ \hbox{and}\ q\in(0,1).$$ The equation then becomes $$\label{ODE_Example} u^{2}(m) - \tilde{c}_{0}(m)u(m) - \beta mu'(m) + \frac{1-q}{q}a^{\frac{1}{1-q}}\left((\psi-1)u(m)\right)^{\frac{-q}{1-q}}(mu'(m))^{\frac{1}{1-q}}=0,$$ and the optimal healthcare spending process $h^*$ is now given by $$h_{t}^* = \left(a^{-1}(\psi-1)\frac{u^*(M_{t})}{M_{t}(u^*)'(M_{t})}\right)^{\frac{-1}{1-q}},$$ where $u^*$ is the unique solution to . The endogenous mortality is then $$\label{Mortality_Example} dM_{t} = M_{t}\left(\beta - \frac{1}{q}a^{\frac{1}{1-q}}\left((\psi-1)\frac{u^*(M_{t})}{M_{t}(u^*)'(M_{t})}\right)^{\frac{-q}{1-q}}\right)dt,\quad M_{0} = m_{0}>0.$$ We calibrate $\beta>0$, $a>0$, $q\in(0,1)$, and $m_0>0$ to mortality data in the US, the UK, the Netherlands, and Bulgaria. For each country, the natural growth rate of mortality $\beta>0$ is estimated from mortality data for the cohort born in 1900, assuming no healthcare available. Given this estimated $\beta>0$, healthcare parameters $a>0$ and $q\in (0,1)$ in , as well as the initial mortality $m_0>0$, are calibrated by matching the endogenous mortality curve with mortality data for the cohort born in 1940, through minimizing the mean squared error (MSE). Calibration results are listed in Table \[table1\]. [l|c|c|c|c|c|c]{} Country & $\beta$ (%) & $m_{0}\times10^{4}$ & $a$ & $q$ & Model MSE $\times10^{6}$ & MSE $\times10^{6}$\ United States (US) & 7.24069 & 1.34995 & 0.19 & 0.61 & 0.0436896 & 0.128984\ United Kingdom (UK) & 7.79605 & 0.843827 & 0.19 & 0.60 & 0.0249924 & 0.12755\ Netherlands (NL) & 8.65832 & 0.477551 & 0.16 & 0.53 & 0.0478583 & 0.207779\ Bulgaria (BG) & 8.86593 & 0.892038 & 0.14 & 0.56 & 0.923716 & 2.85819\ Mortality rates impacted during WWII were excluded when calculating $\beta$. Incomplete data for the 1900 cohort. $\beta$ estimated from age range 47-77. \[table1\] Our calibration can potentially be improved in a few directions. First, we take advantage of the upper and lower bounds in to approximate the solution $u^*$ to , instead of solving directly. There are nontrivial challenges in solving numerically: the natural initial condition $u(0) = 0$ leads to multiple solutions; to resolve this, our investigation suggests enforcing Neumann boundary conditions $u'(0) = \infty$ and $u'(\infty) = 0$, and solving via a sequential approximation technique. This is computationally taxing, even for a fixed pair of parameters $(a,q)$. As the calibration needs to explore numerous possibilities of $(a,q)$, we do not proceed with this approach. Despite this, an efficient numerical scheme for solving , if developed, is likely to improve the accuracy of our calibration. Second, we use a grid size of $0.01$ for the search of optimal $(a,q)$. A finer grid might allow for better distinction between countries (such as the US and UK). Mortality --------- Figure \[fig:cal\_US\] presents the model performance under two simplifying assumptions. First, we assume that the cohort born in 1900 essentially had no access to healthcare, so that its mortality grew exponentially with the Gompertz law. The blue line is then obtained simply by linearly regressing actual mortality data (the blue dots). Second, we assume that the cohort born in 1940 had full access to healthcare. We then compare the model-implied mortality (the red curve) with actual mortality date (the red dots). Although these assumptions are crude approximations, they are in place due to several realistic considerations, as explained in [@Huang19 Section 5.2]. It is confirmed in Figure \[fig:cal\_US\] that our model has the ability to reproduce declines in mortality that are very close to the ones observed historically. Moreover, when compared with [@Huang19 Figure 5.2], Figure \[fig:cal\_US\] clearly provides a better fit. This improvement can be attributed to the use of Epstein-Zin utilities (so that $\gamma$ and $\psi$ can [*both*]{} take empirically relevant values), the inclusion of risky assets, modifications of calibration methods, or a combination of all three. The success of our model is not limited to the US data. Figure \[fig:calibration\] evaluates the model performance using mortality data in the UK, the Netherlands, and Bulgaria. In all cases, the model-implied mortality (the red line) desirably approximates actual mortality data (the red dots). Another way to evaluate the performance of our model is to compare it with linear regression. Indeed, without any idea of healthcare, one can model the actual mortality of the 1940 cohort by a simple linear regression (as we did for the 1900 cohort). As shown in the last two columns of Table \[table1\], our model outperforms linear regression significantly, across all countries considered. Specifically, the sixth column of Table \[table1\] reports the MSEs under our model, much smaller than those under linear regression in the seventh column. [.48]{} ![Mortality rates (vertical axis, in logarithmic scale) at adults’ ages for the cohorts born in 1900 and 1940 in three countries. The dots are actual mortality data (Source: Berkeley Human Mortality Database), and the lines are model-implied mortality curves.[]{data-label="fig:calibration"}]({Model_Calibration/Updates/UK_Mortality.pdf} "fig:"){width=".9\linewidth"} [.48]{} ![Mortality rates (vertical axis, in logarithmic scale) at adults’ ages for the cohorts born in 1900 and 1940 in three countries. The dots are actual mortality data (Source: Berkeley Human Mortality Database), and the lines are model-implied mortality curves.[]{data-label="fig:calibration"}]({Model_Calibration/Updates/Netherlands_Mortality.pdf} "fig:"){width=".9\linewidth"} \ [.48]{} ![Mortality rates (vertical axis, in logarithmic scale) at adults’ ages for the cohorts born in 1900 and 1940 in three countries. The dots are actual mortality data (Source: Berkeley Human Mortality Database), and the lines are model-implied mortality curves.[]{data-label="fig:calibration"}]({Model_Calibration/Updates/Bulgaria_Mortality.pdf} "fig:"){width=".9\linewidth"} Healthcare Spending {#subsec:h^*} ------------------- Figure \[fig: healthcarespending\] demonstrates the model-implied optimal healthcare spending in the four countries. The left panel shows that the proportion of wealth spent on healthcare is negligible at age 40, but increases quickly to around 0.5-1% at age 80. This is in line with the finding in [@Huang19 Section 5.1] under time-separable utilities, and broadly consistent with the US data reported in [@HartmanCatlinLassmanEtAl2008]. The right panel further shows that as one gets older, healthcare spending increases much faster than consumption and investment combined: it accounts for less than 5% of total spending at age 40, but increases continuously to 13-30% at age 80. [.5]{} ![Optimal healthcare spending in the US, UK, Netherlands (NL), and Bulgaria (BG). Left panel: Healthcare-wealth ratio (vertical, log-scale) at adult ages (horizontal). Right panel: Healthcare as a fraction of total spending in consumption, investment, and healthcare (vertical) at adult ages (horizontal).[]{data-label="fig: healthcarespending"}]({Model_Calibration/Updates/Healthcare_Wealth_80L.pdf} "fig:"){width=".85\linewidth"} \[fig:Healthcare-Wealth\] [.48]{} ![Optimal healthcare spending in the US, UK, Netherlands (NL), and Bulgaria (BG). Left panel: Healthcare-wealth ratio (vertical, log-scale) at adult ages (horizontal). Right panel: Healthcare as a fraction of total spending in consumption, investment, and healthcare (vertical) at adult ages (horizontal).[]{data-label="fig: healthcarespending"}]({Model_Calibration/Updates/Healthcare_Spending_80L.pdf} "fig:"){width=".85\linewidth"} \[fig:Healthcare-Spending\] For the US, UK, and Netherlands, healthcare-spending ratios reported above are in broad agreement with actual healthcare expenditure as a percentage of GDP, as displayed in Figure \[fig:LifeExpectancy\]. Bulgaria is distinctively different: model-implied healthcare-spending ratios largely outsize its healthcare expenditure as a percentage of GDP at 8.4%. This might indicate that Bulgaria’s healthcare expenditure is less than optimal, while a detailed empirical investigation is certainly needed here. ![Life expectancy v.s. healthcare spending as a percentage of GDP (2017) for countries in OECD and European Union (Source: OECD Health Statistics Database and [@OECD-EU]).[]{data-label="fig:LifeExpectancy"}]({Model_Calibration/Updates/LifeExpectancy2017.pdf}){width=".7\linewidth"} \[fig:kyg\] The Efficacy Function $g$ {#subsec:g} ------------------------- Figure \[fig:CalibratedEfficiency\] presents the calibrated efficacy function $g(h) = a \frac{h^q}{q}$ for each of the four countries. Intriguingly, it indicates a ranking among countries in term of the effectiveness of healthcare spending: across realistic levels of spending (0-30% of wealth), healthcare is more effective (in reducing mortality growth) in the Netherlands than in the UK, in the UK than in the US, and in the US than in Bulgaria. This leads to two important observations. First, this ranking of efficacy, together with healthcare spending illustrated in Figure \[fig: healthcarespending\], reveals that lower efficacy of healthcare is compensated by larger healthcare spending, relative to both total wealth and total spending. In other words, in the face of enhanced efficacy, our model simply stipulates less healthcare spending, instead of more to take advantage of the reduced marginal cost to curtail mortality growth. Second, this model-implied ranking of efficacy is in broad agreement with empirical studies on healthcare across countries. A common measure of the effectiveness of healthcare spending in a country is life expectancy versus healthcare spending as a percentage of GDP. Figure \[fig:LifeExpectancy\] presents such data for numerous countries, and the black line represents the average effectiveness of healthcare. The Netherlands is further away above the average than the UK, while the US and Bulgaria are two outliers below average. This is generally consistent with the ranking in Figure \[fig:CalibratedEfficiency\]. Certainly, there are more comprehensive, multifaceted measures of healthcare. Tandon et al. [@Evans00], recently rated by [@SAF18] as the most reproducible and transparent ranking of healthcare systems, evaluated healthcare across countries based on quality of care, access to care, efficiency, equity, and general healthiness of citizens. Among the 191 countries evaluated, the Netherlands, the UK, the US, and Bulgaria ranked number 17, 18, 37, and 102, respectively. This is again in line with the ranking suggested by Figure \[fig:CalibratedEfficiency\]. Proofs for Section \[sec:EZ\] ============================= Proof of Proposition \[prop:decompose wtV\] {#subsec:proof of prop:decompose wtV} ------------------------------------------- In view of and , for any $0\le t \le s$, it holds for $\bar{\mathbb{P}}$-a.e. $\bar\omega=(\omega,\omega')\in\bar\Omega$ that $$\label{tau>s} \bar {\mathbb{P}}(\tau>\ell\mid {\mathcal{F}}_s\vee{\mathcal{H}}_t)(\bar\omega) = e^{-\int_{t}^{\ell}M_{u}^{h}(\omega)du} {\mathbbm{1}}_{\{\tau>t\}}(\bar\omega),\quad \forall t\le\ell\le s.$$ Also, since $\widetilde V$ is a $\mathbb G$-adapted semimartingale, it follows from that there exists an $\mathbb F$-adapted semimartingale $V$ such that $$\label{wtV=V} \widetilde V_t = V_t\quad \hbox{$\bar{\mathbb{P}}$-a.s. on $\{t<\tau\}$},\qquad \forall t\ge 0.$$ Indeed, for any fixed $\omega\in\Omega$, consider $A_t(\omega):=\{\omega'\in\Omega': t<\tau(\omega,\omega')\}$ for all $t\ge 0$. As $\widetilde V$ is $\mathbb G$-adapted, implies $\widetilde V_t(\omega,\omega')$ is constant ${\mathbb{P}}'$-a.s. on $A_t(\omega)$. By defining $V_t(\omega) = \widetilde V_t(\omega,A_t(\omega))$ for all $t\ge 0$, $V$ is an $\mathbb F$-adapted semimartingale satisfying . Also note that ${\mathbb{E}}[\sup_{s\in[0,t]}|V_{s}|]<\infty$, as $\bar {\mathbb{E}}[\sup_{s\in[0,t]}|\widetilde V_{s}|]<\infty$, for all $t\ge 0$. Now, observe that $$\begin{aligned} \bar{{\mathbb{E}}}\left[\int_{t\wedge\tau}^{T\wedge\tau}f(c_s,\widetilde{V}_{s}^{c,h})ds\ \middle|\ {\mathcal{G}}_t\right]&=\bar {\mathbb{E}}\left[\int_{t}^{T}{\mathbbm{1}}_{\{s<\tau\}}f(c_s,\widetilde{V}_{s}^{c,h})ds\ \middle|\ {\mathcal{F}}_t\vee{\mathcal{H}}_t\right]\nonumber\\ &= \int_{t}^{T}\bar {\mathbb{E}}\left[{\mathbbm{1}}_{\{s<\tau\}}f(c_s,\widetilde{V}_{s}^{c,h})\ \middle|\ {\mathcal{F}}_t\vee{\mathcal{H}}_t\right]ds\nonumber\\ &= \int_{t}^{T}\bar {\mathbb{E}}\left[\bar {\mathbb{E}}\big[{\mathbbm{1}}_{\{s<\tau\}}f(c_s,{V}_{s}^{c,h})\mid {\mathcal{F}}_s\vee{\mathcal{H}}_t\big]\ \middle|\ {\mathcal{F}}_t\vee{\mathcal{H}}_t\right]ds\nonumber\\ &= \int_{t}^{T}\bar {\mathbb{E}}\left[f(c_s,{V}_{s}^{c,h})\ \bar {\mathbb{E}}\big[{\mathbbm{1}}_{\{s<\tau\}}\mid {\mathcal{F}}_s\vee{\mathcal{H}}_t\big]\ \middle|\ {\mathcal{F}}_t\vee{\mathcal{H}}_t\right]ds\nonumber\\ &= \int_{t}^{T}\bar {\mathbb{E}}\left[f(c_s,{V}_{s}^{c,h}) {\mathbbm{1}}_{\{t<\tau\}}e^{-\int_{t}^{s}M_{u}^{h}du}\ \middle|\ {\mathcal{F}}_t\vee{\mathcal{H}}_t\right]ds\nonumber\\ &= \bar {\mathbb{E}}\left[\int_{t}^{T}{\mathbbm{1}}_{\{t<\tau\}}e^{-\int_{t}^{s}M_{u}^{h}du}f(c_s,{V}_{s}^{c,h})ds\ \middle|\ {\mathcal{G}}_t\right],\label{11}\end{aligned}$$ where the second and the last equalities follow from Fubini’s theorem for conditional expectations (see e.g. [@Schilling-book-2017 Theorem 27.17]), the third equality is due to the tower property of conditional expectations and , the fourth equality results from $c_s\in {\mathcal{F}}_s$ and $V^{c,h}_s\in {\mathcal{F}}_s$, and the fifth equality is a consequence of . On the other hand, for $\bar{\mathbb{P}}$-a.e. fixed $\bar\omega=(\omega,\omega')\in\bar\Omega$, consider the cumulative distribution function of $\tau$ given the information ${\mathcal{F}}_T\vee{\mathcal{H}}_t$, i.e. $$F(s):= \bar{\mathbb{P}}(\tau\le s\mid {\mathcal{F}}_T\vee{\mathcal{H}}_t)(\bar\omega),\quad s\ge0.$$ Thanks to , $F(s) = 1- e^{-\int_{t}^{s}M_{u}^{h}(\omega)du} {\mathbbm{1}}_{\{\tau>t\}}(\bar\omega)$ for $t\le s\le T$. The density function of $\tau$ given the information ${\mathcal{F}}_T\vee{\mathcal{H}}_t$ is then $$\label{tau density} \eta(s) = F'(s) = M_{s}^{h}(\omega)e^{-\int_{t}^{s}M_{u}^{h}(\omega)du} {\mathbbm{1}}_{\{\tau>t\}}(\bar\omega),\quad \hbox{for}\ t\le s\le T.$$ It follows that $$\begin{aligned} \bar{\mathbb{E}}\left[\widetilde{V}_{\tau-}^{c,h}{\mathbbm{1}}_{\{\tau\le T\}}\ \middle|\ {\mathcal{G}}_t\right] &= \bar{\mathbb{E}}\left[{V}_{\tau-}^{c,h}{\mathbbm{1}}_{\{\tau\le T\}}\ \middle|\ {\mathcal{G}}_t\right] {\mathbbm{1}}_{\{\tau\le t\}} +\bar{\mathbb{E}}\left[{V}_{\tau-}^{c,h}{\mathbbm{1}}_{\{\tau\le T\}}\ \middle|\ {\mathcal{G}}_t\right] {\mathbbm{1}}_{\{\tau>t\}}\nonumber\\ &={V}_{\tau-}^{c,h}{\mathbbm{1}}_{\{\tau\le t\}}+ \bar{\mathbb{E}}\left[ \bar{\mathbb{E}}\big[{V}_{\tau-}^{c,h}{\mathbbm{1}}_{\{t<\tau\le T\}}\mid {\mathcal{F}}_T\vee{\mathcal{H}}_t\big]\ \middle|\ {\mathcal{F}}_t\vee{\mathcal{H}}_t\right]\nonumber\\ &={V}_{\tau-}^{c,h}{\mathbbm{1}}_{\{\tau\le t\}}+ \bar{\mathbb{E}}\left[ \int_{t}^{T}{\mathbbm{1}}_{\{t<\tau\}}M_{s}^{h}e^{-\int_{t}^{s}M_{u}^{h}du}{V}_{s}^{c,h}ds\ \middle|\ {\mathcal{G}}_t \right],\label{22}\end{aligned}$$ where the first line results from $\widetilde V_{\tau-}=V_{\tau-}$ (thanks to ), the second line follows from the tower property of conditional expectations, and the third line is due to the density formula in . Note that since $V$ is right-continuous, it has at most countably many jumps on $[t,T]$, so that we may use $V_s$ (instead of $V_{s-}$) in the last term of . Finally, $$\begin{aligned} \bar{\mathbb{E}}\left[\widetilde{V}_{T}^{c,h}{\mathbbm{1}}_{\{\tau> T\}}\ \middle|\ {\mathcal{G}}_t \right] &= \bar{\mathbb{E}}\left[ \bar{\mathbb{E}}\big[{V}_{T}^{c,h}{\mathbbm{1}}_{\{\tau>T\}}\mid {\mathcal{F}}_T\vee{\mathcal{H}}_t\big]\ \middle|\ {\mathcal{F}}_t\vee{\mathcal{H}}_t\right]\nonumber\\ &=\bar{\mathbb{E}}\left[ {V}_{T}^{c,h} \bar{\mathbb{E}}\big[{\mathbbm{1}}_{\{\tau>T\}}\mid {\mathcal{F}}_T\vee{\mathcal{H}}_t\big]\ \middle|\ {\mathcal{F}}_t\vee{\mathcal{H}}_t\right]\nonumber\\ &= \bar {\mathbb{E}}\left[{\mathbbm{1}}_{\{t<\tau\}}e^{-\int_{t}^{T}M_{u}^{h}du}{V}_{T}^{c,h}\ \middle|\ {\mathcal{G}}_t \right],\label{33}\end{aligned}$$ where the first equality follows from the tower property of conditional expectations and , the second equality is due to ${V}_{T}^{c,h}\in {\mathcal{F}}_T$, and the third equality is a consequence of . Now, combining , , and , we obtain from and $\widetilde V_{\tau-}=V_{\tau-}$ that $$\begin{aligned} \widetilde{V}_{t}^{c,h} & ={\mathbb{E}}_{t}\bigg[\int_{t}^{T}e^{-\int_{t}^{s}M_{r}^{h}dr}\left(f(c_{s},{V}_{s}^{c,h}) + \zeta^{1-\gamma}M_{s}^{h}{V}_{s}^{c,h}\right)ds + e^{-\int_{t}^{T}M_{s}^{h}ds}{V}_{T}^{c,h} \bigg] {\mathbbm{1}}_{\{t<\tau\}}\nonumber\\ &\hspace{2.8in}+ \zeta^{1-\gamma}{V}_{\tau-}^{c,h}{\mathbbm{1}}_{\{t\ge\tau\}},\ \ \text{ for all }0\le t\le T<\infty,\label{wtV'}\end{aligned}$$ where we use the notation ${\mathbb{E}}_{t}\left[\cdot\right]={\mathbb{E}}\left[\cdot|{\mathcal{F}}_{t}\right]$. This, together with , particularly implies $$\label{with ind} V_t(\omega) {\mathbbm{1}}_{\{t<\tau\}(\omega,\omega')} = \widetilde V_t(\omega,\omega') {\mathbbm{1}}_{\{t<\tau\}(\omega,\omega')} = E_{t,T}(\omega){\mathbbm{1}}_{\{t<\tau\}(\omega,\omega')},$$ where $$E_{t,T}(\omega) := {\mathbb{E}}_{t}\bigg[\int_{t}^{T}e^{-\int_{t}^{s}M_{r}^{h}dr}\left(f(c_{s},{V}_{s}^{c,h}) + \zeta^{1-\gamma}M_{s}^{h}{V}_{s}^{c,h}\right)ds + e^{-\int_{t}^{T}M_{s}^{h}ds}{V}_{T}^{c,h}\bigg](\omega).$$ For any $\omega\in\Omega$, since there exists $\omega'\in\Omega'$ such that ${\mathbbm{1}}_{\{t<\tau\}(\omega,\omega')}=1$ (in view of and ), we conclude from that $V_t(\omega) = E_{t,T}(\omega)$. We can then simplify as $$\begin{aligned} \label{wtV'''} \widetilde{V}_{t} & =V_t {\mathbbm{1}}_{\{t<\tau\}} + \zeta^{1-\gamma}{V}_{\tau-}{\mathbbm{1}}_{\{t\ge\tau\}},$$ where $V$ satisfies $$\label{V eqn} V_t = {\mathbb{E}}_{t}\bigg[\int_{t}^{T}e^{-\int_{t}^{s}M_{r}^{h}dr}\left(f(c_{s},{V}_{s}) + \zeta^{1-\gamma}M_{s}^{h}{V}_{s}\right)ds + e^{-\int_{t}^{T}M_{s}^{h}ds}{V}_{T}\bigg],\ \ \forall 0\le t\le T<\infty$$ Now, note that the above equation directly implies $$V'_t := e^{-\int_{0}^{t}M_{r}^{h}dr}V_t = \mathscr M'_t - \int_0^t e^{-\int_{0}^{s}M_{r}^{h}dr} \left(f(c_{s},{V}_{s}) + \zeta^{1-\gamma}M_{s}^{h}{V}_{s}\right)ds,$$ where $$\mathscr M'_t := {\mathbb{E}}_{t}\bigg[\int_{0}^{T}e^{-\int_{0}^{s}M_{r}^{h}dr}\left(f(c_{s},{V}_{s}^{}) + \zeta^{1-\gamma}M_{s}^{h}{V}_{s}^{}\right)ds + e^{-\int_{0}^{T}M_{s}^{h}ds}{V}_{T}^{}\bigg]$$ is an $\mathbb F$-martingale on $[0,T]$, thanks to . Applying generalized Itô’s formula for semimartingales (see [@Jacod03 Theorem I.4.57]) to $V_t = e^{\int_{0}^{t}M_{r}^{h}dr} V'_t$ gives $$dV_t = -F(c_t,M^h_t,V_t) + e^{\int_{0}^{t}M_{r}^{h}dr} d\mathscr M'_t.$$ Since $0\le M^h_t \le me^{\beta t}$ by definition (see ), $\mathscr M_t:= \int_0^t e^{\int_{0}^{s}M_{r}^{h}dr} d\mathscr M'_s$ is again an $\mathbb F$-martingale. Hence, $V$ is a solution to BSDE . This, together with , yields the desired result. Derivation of Proposition \[prop:comparison\] {#subsec:proof of comparison} --------------------------------------------- Recall the generator $F$ defined in . \[Lemma1\] Let $c, h, V$ and $W$ be $\mathbb F$-progressively measurable processes with $W_{s}\le V_{s}$ for all $s\ge 0$. If there exists $k\in{\mathbb{R}}$ such that $V$ satisfies , then $$\label{monotonicity_V} F(c_{s},M^h_{s},V_{s}) - F(c_{s},M^h_{s},W_{s})\le -\Gamma(\Lambda, M^h_{s})(V_{s}-W_{s}),$$ where $\Lambda:=\delta\theta +(1-\theta)k$ (as in Definition \[cE\]) and the function $\Gamma$ is defined by $$\label{Gamma} \Gamma(\lambda, m):= \lambda+\frac{\gamma(\psi-1)}{1-\gamma}(1-\zeta^{1-\gamma})m.$$ As in the proof of [@Melnyk17 Lemma B.1], holds by the mean value theorem provided that $F_{v}(c_{s},M^h_{s},u)\le -\Gamma(\Lambda, M^h_{s})$ for all $u\in[W_{s},V_{s}]$. To this end, note that $$F_{v}(c_{s},M^h_{s},u) = -\bigg(\delta\theta + (1-\zeta^{1-\gamma})M^h_{s} + \delta(1-\theta)\bigg(\frac{c_{s}^{1-\gamma}}{(1-\gamma)u}\bigg)^{{1}/{\theta}}\bigg).$$ Thanks to , a direct calculation shows $F_{vv}(c_{s},M^h_{s},u)>0$, i.e. $F_{v}(c_{s},M^h_{s},u)$ is increasing in $u$. This, together with $V$ satisfying , implies that for all $u\in[W_{s},V_{s}]$, $F_{v}(c_{s},M^h_{s},u)\le F_{v}(c_{s},M^h_{s},\hat u)$, where $\hat u := \delta^{\theta}\big(k-\frac{\psi-1}{1-\gamma}(\zeta^{1-\gamma}-1)M^h_{s}\big)^{-\theta}\frac{c_{s}^{1-\gamma}}{1-\gamma}$. By direct calculation, $$\begin{aligned} F_{v}(c_{s},M^h_{s},\hat u) &= -\bigg(\delta\theta + (1-\zeta^{1-\gamma})M^h_{s} + (1-\theta)\left(k-\frac{\psi-1}{1-\gamma}(\zeta^{1-\gamma}-1)M^h_{s}\right)\bigg)\\ &= -\left(\Lambda+\frac{\gamma(\psi-1)}{1-\gamma}(1-\zeta^{1-\gamma})M^h_{s}\right)=-\Gamma(\Lambda, M^h_{s}), \end{aligned}$$ where the second equality follows from the definition of $\Lambda$ and $\theta=\frac{1-\gamma}{1-1/\psi}$. To prove Proposition \[prop:comparison\], we intend to follow the idea in the proof of [@Melnyk17 Theorem 2.2]. The involvement of the controlled mortality $M^h$ in , as well as the possibility that $\Lambda$ therein can be negative (Remark \[rem:negative Lambda\]), result in additional technicalities. The proof below combines arguments in [@Melnyk17 Theorem 2.2] and [@Fan15 Theorem 2.1], adapted to weaker regularity of processes. Recall the function $\Gamma$ in . Fix $0\le t_0<T$, define $$\label{Delta} \Delta_{t}:=e^{-\int_{t_0}^{t}\Gamma(0,M^h_{s})ds}\left(V_{t}^{1}-V_{t}^{2}\right),\quad t\in[t_{0},T],$$ and consider the stopping time $\theta := \inf\left\{s\ge t_0: V_{s}^{1}\le V_{s}^{2}\right\}$. Applying generalized Itô’s formula (see [@Jacod03 Theorem I.4.57]) to $e^{-\int_{0}^{t}\Gamma(0,M^h_{s})ds}V_{t}^{i}$, $i=1,2$, yields $$\begin{aligned} d\left(e^{-\int_{0}^{t}\Gamma(0,M^h_{s})ds}V_{t}^{1}\right) &= -e^{-\int_{0}^{t}\Gamma(0,M^h_{s})ds}\left[\Gamma(0,M^h_{s})V_{t}^{1} +F(c_{t},M^h_{t},V_{t}^{1})\right]dt + e^{-\int_{0}^{t}\Gamma(0,M^h_{s})ds}d\mathscr{M}_{t}^{1},\\ d\left(e^{-\int_{0}^{t}\Gamma(0,M^h_{s})ds}V_{t}^{2}\right) &= -e^{-\int_{0}^{t}\Gamma(0,M^h_{s})ds}\left[\Gamma(0,M^h_{s})V_{t}^{2} +G(t,V_{t}^{2})\right]dt + e^{-\int_{0}^{t}\Gamma(0,M^h_{s})ds}d\mathscr{M}_{t}^{2}, \end{aligned}$$ where $\mathscr M^1$, $\mathscr M^2$ are some $\mathbb F$-martingales on $[0,T]$. As $0\le \Gamma(0,M^h_{t})\le \frac{\gamma(\psi-1)}{1-\gamma}(1-\zeta^{1-\gamma})me^{\beta t}$ by the definition of $M^h$ in , the processes $\int_{t_{0}}^{\cdot}e^{-\int_{0}^{t}\Gamma(0,M^h_{s})ds}d\mathscr{M}_{t}^{i}$ is a true martingales for $i=1,2$. It follows that $$\begin{aligned} \Delta_{t}= {\mathbb{E}}_{t}\left[\int_{t}^{T}{\mathbbm{1}}_{\{s<\theta\}}\left[\left(F(c_{s},M^h_{s},V_{s}^{1})-G(s,V_{s}^{2})\right) + \Gamma(0,M^h_{s})\left(V_{s}^{1}-V_{s}^{2}\right)\right]e^{-\int_{t_0}^{s}\Gamma(0,M^h_{r})dr}ds + \Delta_{T\wedge\theta}\right]. \end{aligned}$$ Observe that $$\begin{aligned} {\mathbbm{1}}_{\{s<\theta\}}\left(F(c_{s},M^h_{s},V_{s}^{1})-G(s,V_{s}^{2})\right)&= {\mathbbm{1}}_{\{s<\theta\}}\left(F(c_{s},M^h_{s},V_{s}^{1})-F(c_{s},M^h_{s},V_{s}^{2})\right)\\ & \hspace{0.5in}+{\mathbbm{1}}_{\{s<\theta\}}\left(F(c_{s},M^h_{s},V_{s}^{2}) - G(s,V_{s}^{2})\right)\\ &\le {\mathbbm{1}}_{\{s<\theta\}}\left(F(c_{s},M^h_{s},V_{s}^{1})-F(c_{s},M^h_{s},V_{s}^{2})\right)\\ &\le {\mathbbm{1}}_{\{s<\theta\}}\left( -\Gamma(\Lambda,M^h_{s})\left(V_{s}^{1}-V_{s}^{2}\right)\right), \end{aligned}$$ where the first inequality follows from $F(c_{s},M^h_{s},V_{s}^2)\le G(s,V_{s}^2)$, and the second inequality is due to Lemma \[Lemma1\]. Note that Lemma \[Lemma1\] is applicable here as $V_{s}^{1}>V_{s}^{2}$ for $s\in[t,\theta)$. Thanks to the above inequality, $$\begin{aligned} \label{B1} \Delta_{t}&\le {\mathbb{E}}_{t}\left[\int_{t}^{T}{\mathbbm{1}}_{\{s<\theta\}}\left[-\Gamma(\Lambda,M^h_{s}) + \Gamma(0,M^h_{s})\right]\left(V_{s}^{1}-V_{s}^{2}\right)e^{-\int_{t_{0}}^{s}\Gamma(0,M^h_{r})dr}ds + \Delta_{T\wedge\theta}\right]\nonumber\\ & = {\mathbb{E}}_{t}\left[-\int_{t}^{T} {\mathbbm{1}}_{\{s<\theta\}}\Lambda\Delta_{s}ds + \Delta_{T\wedge\theta}\right], \end{aligned}$$ where the second line follows from $\Gamma(\Lambda,M^h_{s}) = \Lambda + \Gamma(0,M^h_{s})$ and . Multiplying both sides by ${\mathbbm{1}}_{\{t<\theta\}}$ yields $$\Delta_{t}{\mathbbm{1}}_{\{t<\theta\}}\le {\mathbb{E}}_{t}\left[-\int_{t}^{T}\Lambda\Delta_{s}{\mathbbm{1}}_{\{s<\theta\}}ds + \Delta_{T\wedge \theta}{\mathbbm{1}}_{\{t<\theta\}}\right]\le {\mathbb{E}}_{t}\left[-\int_{t}^{T}\Lambda\Delta_{s}{\mathbbm{1}}_{\{s<\theta\}}ds + \Delta_{T}{\mathbbm{1}}_{\{T<\theta\}}\right],$$ where the second inequality follows from the right continuity of $V^1$ and $V^2$: indeed, the right continuity implies $V_{\theta}^{1}\le V_{\theta}^{2}$, so that $\Delta_{T\wedge\theta} = \Delta_{\theta}{\mathbbm{1}}_{\{\theta\le T\}} + \Delta_{T} {\mathbbm{1}}_{\{T<\theta\}} \le \Delta_{T} {\mathbbm{1}}_{\{T<\theta\}}$. By defining $\Delta_{t}^{+} = \Delta_{t}{\mathbbm{1}}_{\{t<\theta\}}$, we write the previous inequality as $$\label{before GW} \Delta_{t}^{+}\le {\mathbb{E}}_{t}\left[-\int_{t}^{T}\Lambda\Delta_{s}^{+}ds + \Delta_{T}^{+}\right].$$ Now, consider $\Theta_{t}:= {\mathbb{E}}\left[\Delta^{+}_{t}\right]\ge 0$, which is well-defined as $\Gamma(0,M_{s})\ge0$ and ${\mathbb{E}}\big[\sup_{t\in[0,T]}|V_{t}^{i}|\big]<\infty$, thanks to $V^i\in{\mathcal{E}}^h_k$ (Definition \[cE\]), for $i=1,2$. We then obtain from that $$\label{GW} \Theta_{t} \le -\int_{t}^{T}\Lambda\Theta_{s}ds + \Theta_{T}.$$ If $\Lambda>0$, by writing $\Theta_{T}\ge \Theta_{t} + \int_{t}^{T}\Lambda\Theta_{s}ds$, we apply standard Gronwall’s inequality to get $\Theta_{T}\ge \Theta_{t}e^{\int_{t}^{T}\Lambda ds}$, or equivalently $$\label{Theta Bound} \Theta_{t}\le \Theta_{T}e^{-\int_{t}^{T}\Lambda ds},\quad t\in[t_0,T].$$ If $\Lambda<0$, applying backward Gronwall’s inequality (see [@Fan18 Proposition 2]) to also gives . By , , and , we obtain $$\Theta_{t_{0}}\le \Theta_{T}e^{-\int_{t_{0}}^{T}\Lambda ds}\le {\mathbb{E}}\left[e^{-\int_{t_{0}}^{T}\Gamma(\Lambda,M_{s})ds}\left(|V_{T}^{1}|+|V_{T}^{2}|\right)\right].$$ Since $T>0$ is arbitrary, the transversality condition in $\eqref{TransversalityCondition}$ for $V_{t}^{1}$ and $V_{t}^{2}$ immediately implies $$0\le \Theta_{t_{0}} \le \lim\limits_{T\rightarrow\infty}{\mathbb{E}}\left[e^{-\int_{t_{0}}^{T}\Gamma(\Lambda,M_{s})ds}\left(|V_{T}^{1}|+|V_{T}^{2}|\right)\right]= 0.$$ That is, $\Theta_{t_0}={\mathbb{E}}\left[\left(V_{t_0}^{1}-V_{t_0}^{2}\right)1_{\{t_0<\theta\}}\right] = 0$. This entails $\theta=t_0$, and thus $V_{t_{0}}^{1}\le V_{t_{0}}^{2}$. Since $t_0\ge 0$ is arbitrary, we conclude that $V_{t}^{1}\le V_{t}^{2}$ for all $t\ge 0$. Proofs for Section \[sec:Results\] ================================== Proof of Proposition \[prop:NoAging\] {#subsec:proof of prop:NoAging} ------------------------------------- For any fixed $m>0$ such that $\tilde c_0(m)>0$, define $w(x):=\delta^{\theta}\frac{x^{1-\gamma}}{1-\gamma}\tilde{c}_{0}(m)^{-\frac{\theta}{\psi}}$ for all $x>0$. In order to apply Theorem \[Verification\] to our setting, we need to verify all its conditions. First, it can be checked directly that $w$, as a one-variable function, solves in a trivial way, with derivatives with respect to $m$ in all being zero. For any $(c,\pi,h)\in{\mathcal{P}}={\mathcal{P}}_1$, since $(X^{c,\pi,h})^{1-\gamma}$ satisfies (with $\Lambda\in{\mathbb{R}}$ therein taken to be $\Lambda^*$), so does $ w(X^{c,\pi,h}_t)$; namely, $w(X^{c,\pi,h}_t)\in{\mathcal{E}}^h_{k^*}$. Also, by the definitions of ${\mathcal{P}}$ and $w$, ${\mathcal{P}}={\mathcal{P}}_1\subseteq {\mathcal{H}}_{k^*}$ and is satisfied. As $\tilde c_0(m)>0$, $w_x>0$ and $w_{xx}<0$ by definition. It follows that $\bar c(x,m):=x\tilde{c}_{0}(m)$ and $\bar \pi(x,m) :=\frac{\mu}{\gamma\sigma^2}$ are the unique maximizers of the two supremums in , respectively. Note that the supremum in is zero, as $g\equiv 0$ and $w_x>0$. Hence, $\bar h(x,m) := 0$ trivially maximizes . Now, the only condition in Theorem \[Verification\] that remains to be checked is “$(c^*,\pi^*,h^*)$ defined in belongs to ${\mathcal{P}}$ and $W^*_t:= w(X^{c^*,\pi^*,h^*}_t)$ satisfies ”. To this end, observe that a unique solution $X^*=X^{c^*,\pi^*,h^*}$ to exists, which is a geometric Brownian motion satisfying the dynamics $$\label{X^* GBM} dX_{t}^{*} = X_{t}^{*}\left(r+\frac{1}{\gamma}\left(\frac{\mu}{\sigma}\right)^{2} - \tilde{c}_{0}(m)\right)dt + X_{t}^{*}\frac{\mu}{\gamma\sigma}dB_{t},$$ This implies that $$\begin{aligned} \label{X^*} (X^{*}_t)^{1-\gamma}= x^{1-\gamma}\text{exp}\left((1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} - \tilde{c}_{0}(m) - \frac{(1-\gamma)}{2\gamma^{2}}\left(\frac{\mu}{\sigma}\right)^{2}\right)t + \frac{(1-\gamma)\mu}{\gamma\sigma}B_{t}\right), \end{aligned}$$ which is again a geometric Brownian motion that satisfies the dynamics $$\frac{dY_t}{Y_t} = (1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} - \tilde{c}_{0}(m)\right)dt + \frac{(1-\gamma)\mu}{\gamma\sigma} dB_t,\quad Y_0=x^{1-\gamma}.$$ Consequently, $$\label{first part} e^{-\Lambda^* t} {\mathbb{E}}\left[e^{-\gamma(\psi-1)\frac{1-\zeta^{1-\gamma}}{1-\gamma}mt}(X^*_t)^{1-\gamma}\right]=x^{1-\gamma}e^{(C-\Lambda^*)t},$$ where $$C:= (1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} - \tilde{c}_{0}(m) \right)-\gamma(\psi-1)\frac{1-\zeta^{1-\gamma}}{1-\gamma}m. $$ Remarkably, by the definitions of $\tilde c_0(m)$ and $\Lambda^*$ in and , a direct calculation shows that $C-\Lambda^* = -\tilde c_0(m) <0$, where the inequality follows from $\tilde c_0(m)>0$. It follows from that $$\label{de} \lim_{t\to\infty} e^{-\Lambda^* t} {\mathbb{E}}\left[e^{-\gamma(\psi-1)\frac{1-\zeta^{1-\gamma}}{1-\gamma}mt}(X^*_t)^{1-\gamma}\right]=0.$$ On the other hand, we can rewrite as $$\begin{aligned} \label{X^* with Z} (X^{*}_t)^{1-\gamma}= x^{1-\gamma}\text{exp}\left((1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} - \tilde{c}_{0}(m) \right)t \right)\cdot Z_t, \end{aligned}$$ where $Z$ is a geometric Brownian motion with the dynamics ${dZ_t} =Z_t \frac{(1-\gamma)\mu}{\gamma\sigma}dB_{t}$, $Z_0=1$. As $Z$ is a martingale, we can apply the Burkhölder-Davis-Gundy inequality to get $$\label{BDG} {\mathbb{E}}\bigg[\sup\limits_{s\in[0,t]}(X_{s}^{*})^{1-\gamma}\bigg]\le K x^{1-\gamma} e^{\left(|1-\gamma|\left|r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} - \tilde{c}_{0}(m) \right|\right) t} \frac{|1-\gamma|\mu}{\gamma\sigma} {\mathbb{E}}\left[\bigg(\int_0^t Z_s^2 ds\bigg)^{1/2}\right],$$ for some constant $K>0$. By Jensen’s inequality and Fubini’s theorem, $${\mathbb{E}}\bigg[\bigg(\int_0^t Z_s^2 ds\bigg)^{1/2}\bigg] \le \bigg(\int_0^t {\mathbb{E}}[Z_s^2] ds\bigg)^{1/2} = \bigg(\int_0^t e^{\frac{(1-\gamma)^2\mu^2}{\gamma^2\sigma^2}s} ds\bigg)^{1/2} = \frac{\gamma\sigma}{|1-\gamma|\mu} \bigg(e^{\frac{(1-\gamma)^2\mu^2}{\gamma^2\sigma^2}t}-1\bigg)^{1/2}.$$ We then conclude from the above two inequalities that $$\label{do} {\mathbb{E}}\bigg[\sup\limits_{s\in[0,t]}(X_{s}^{*})^{1-\gamma}\bigg]<\infty,\quad \forall t\ge0.$$ By and , $(X^{*})^{1-\gamma}$ satisfies (with $\Lambda^*$ in place of $\Lambda$), and so does the process $W^*_t:= w(X^*_t)=\delta^{\theta}\tilde{c}_{0}(m)^{-\frac{\theta}{\psi}}\frac{(X^*_t)^{1-\gamma}}{1-\gamma}$, i.e. $W^*\in {\mathcal{E}}^{h^*}_{k^*}$. Now, by applying Itô’s formula to $W^*_t$ and using the fact $$\label{dodo} {\mathbb{E}}\bigg[\sup_{s\in[0,t]}\pi^*_s(X_{s}^{*})^{1-\gamma}\bigg]<\infty\quad \hbox{for all}\ t\ge0,$$ which is a direct consequence of and $\pi^*_t\equiv \frac{\mu}{\gamma\sigma^2}$ being a constant process, we can argue as in the proof of Theorem \[Verification\] that $W^*_t$ is a solution to . Moreover, by direct calculation, $$W^*_t = \delta^\theta \tilde c_0(m)^{-\theta+(1-\gamma)}\frac{(X^*_t)^{1-\gamma}}{1-\gamma} = \delta^\theta \tilde c_0(m)^{-\theta}\frac{(c^*_t)^{1-\gamma}}{1-\gamma}$$ As $\tilde c_0(m)=k^*+(\psi-1)\frac{1-\zeta^{1-\gamma}}{1-\gamma}m$ (see ), this shows that is satisfied with $k=k^*$. Hence, $(c^*,h^*)$ is $k^*$-admissible, so that we can conclude $(c^*,\pi^*,h^*)\in{\mathcal{P}}$. Theorem \[Verification\] is then applicable to our setting, asserting that $w(x,m)=v(x,m)$ and $(c^*,\pi^*,h^*)$ optimizes . Proof of Theorem \[Thm:AH\] {#subsec:proof of Thm:AH} --------------------------- Define $w(x,m):=\delta^{\theta}\frac{x^{1-\gamma}}{1-\gamma}u^*(m)^{-\frac{\theta}{\psi}}$ for all $(x,m)\in{\mathbb{R}}^2_+$. In order to apply Theorem \[Verification\] to our setting, we need to verify all its conditions. First, $w\in C^{2,1}({\mathbb{R}}_+\times{\mathbb{R}}_+)$ by definition and it can be checked, as in -, that $w$ solves . In view of the definitions of ${\mathcal{P}}$ and $w$, ${\mathcal{P}}\subseteq {\mathcal{H}}_{k^*}$ and is satisfied for any $(c,\pi,h)\in{\mathcal{P}}$. Moreover, as $w_x>0$, $w_{xx}<0$, and $g$ satisfies Assumption \[assump:AH\], the functions $\bar c$, $\bar \pi$, and $\bar h$ in are unique maximizers of the supermums in and . In the following, we focus on showing (i) for any $(c,\pi,h)\in{\mathcal{P}}$, $w(X^{c,\pi,h}_t, M^h_t)\in{\mathcal{E}}^h_{k^*}$; (ii) $(c^*, \pi^*, h^*)$, defined using $\bar c$, $\bar\pi$, and $\bar h$ as in , belongs to ${\mathcal{P}}$ and $W^*_t:= w(X^{c^*,\pi^*,h^*}_t,M^{h^*}_t)$ satisfies . [**(i)**]{} Take any $\mathfrak p = (c,\pi,h)\in{\mathcal{P}}$, and set $W_{t} := w(X_{t}^{{\mathfrak{p}}},M_{t}^{h})$ for $t\ge 0$. We will prove that $W\in{\mathcal{E}}^h_{k^*}$, i.e. $W$ satisfies with $\Lambda$ therein taken as $\Lambda^*$ in . - [**Case (i)-1:**]{} $\gamma\in(\frac{1}{\psi},1)$. In view of , , and , we have $u^*(m)\ge \tilde c_0(m)\ge \tilde c_0(0)=k^*>0$. As $\theta>0$ when $\gamma\in(\frac{1}{\psi},1)$, this implies $$0<W_t=\delta^{\theta}\frac{(X^{{\mathfrak{p}}}_t)^{1-\gamma}}{1-\gamma}u^*(M^h_t)^{-\frac{\theta}{\psi}}\le \delta^{\theta}\frac{(X^{{\mathfrak{p}}}_t)^{1-\gamma}}{1-\gamma}(k^*)^{-\frac{\theta}{\psi}}\quad \forall t\ge 0,$$ Since $(X^{{\mathfrak{p}}})^{1-\gamma}$ satisfies (thanks to ${\mathfrak{p}}\in{\mathcal{P}}={\mathcal{P}}_1$), the above inequality implies that $W$ also satisfies . - [**Case (i)-2:**]{} $\gamma>1$ and $\zeta<1$. As ${\mathfrak{p}}\in{\mathcal{P}}={\mathcal{P}}_2$, there exists $\eta\in(1-\frac{1}{\gamma},1)$ such that holds. Consider the constants $$\label{alpha's} \alpha := -\eta\frac{\gamma(\psi-1)}{1-\gamma}(\zeta^{1-\gamma}-1)>0,\quad \alpha' := -(1-\eta)\frac{\gamma(\psi-1)}{1-\gamma}(\zeta^{1-\gamma}-1)>0,$$ as well as the process $$\label{F} F_t := \left(u_{\beta}(M_{t}^{h})\right)^{-\frac{\theta}{\psi}}\text{exp}\left(-\alpha'\int_{0}^{t}M_{s}^{h}ds\right),\quad t\ge 0.$$ First, we claim that the process $F$ is bounded from above; more specifically, $$\label{F bounded} \sup_{t\ge 0}F_t \le u_{\beta}\left(-\frac{\theta}{\alpha'\psi}\beta\right)^{-{\theta}/{\psi}}<\infty. $$ Observe that $$\begin{aligned} \frac{dF_t}{dt} &= -\left(\alpha' M_{t}^{h} + \frac{\theta}{\psi}u_{\beta}(M_{t}^{h})^{-1}u_{\beta}'(M_{t}^{h})\frac{dM_{t}^{h}}{dt}\right)F_t\nonumber\\ &= -\left(\alpha' M_{t}^{h} + \frac{\theta}{\psi\beta}(\beta-g(h_{t}))\big[u_{\beta}(M_{t}^{h}) - \tilde{c}_{0}(M_{t}^{h})\big]\right)F_t,\label{dF} \end{aligned}$$ where the second equality follows as $u_\beta$ solves with $q=\beta$. For each $\omega\in\Omega$, consider the collection of time points $$S(\omega) := \left\{t\ge 0: M_{t}^{h}(\omega)= \frac{-\theta}{\alpha'\psi\beta}(\beta-g(h_{t}))\big(u_{\beta}(M_{t}^{h}) - \tilde{c}_{0}(M_{t}^{h})\big)(\omega)\right\}.$$ We deduce from that $$\label{local max} \hbox{local maximizers of $t\mapsto F_t(\omega)$ must occur at time points in $S(\omega)$}.$$ Also, by $g\ge 0$ and , $$\label{L<} L_t (\omega):= \frac{-\theta}{\alpha'\psi\beta}(\beta-g(h_{t}))\big(u_{\beta}(M_{t}^{h}) - \tilde{c}_{0}(M_{t}^{h})\big)(\omega)\le -\frac{\theta}{\alpha'\psi}\beta,\quad \forall t\ge0.$$ This particularly implies that $$\label{M<} M^h_t(\omega) = L_t(\omega)\le -\frac{\theta}{\alpha'\psi}\beta,\quad \hbox{for each}\ t\in S(\omega).$$ Now, there are three distinct possibilities: 1) There exists $t^*\ge 0$ such that $M^h_t(\omega)< L_t(\omega)$ for all $t>t^*$. Then, $S(\omega)\subseteq [0,t^*]$ and implies $M^h_t(\omega)< -\frac{\theta}{\alpha'\psi}\beta$ for all $t> t^*$. It then follows from and that $$\label{<t^*} \sup_{t\le t^*} F_t(\omega)= \sup_{t\in S(\omega)} F_t(\omega) \le \sup_{t\in S(\omega)} u_{\beta}\big(M^h_t(\omega)\big)^{-\frac{\theta}{\psi}} \le u_{\beta}\left(-\frac{\theta}{\alpha'\psi}\beta\right)^{-{\theta}/{\psi}},$$ where the last inequality follows from . Moreover, $$\sup_{t> t^*} F_t(\omega)\le \sup_{t> t^*} u_{\beta}\big(M^h_t(\omega)\big)^{-\frac{\theta}{\psi}}\le u_{\beta}\left(-\frac{\theta}{\alpha'\psi}\beta\right)^{-{\theta}/{\psi}}.$$ Hence, holds. 2) There exists $t^*\ge 0$ such that $M^h_t(\omega)> L_t(\omega)$ for all $t>t^*$. By , $F_t(\omega)$ is strictly decreasing for $t>t^*$. Thus, $\sup_{t\ge 0} F_t(\omega) = \sup_{t\le t^*} F_t(\omega) = \sup_{t\in S(\omega)} F_t(\omega)$. Thanks to the estimate in , we conclude that holds. 3) Neither 1) nor 2) above holds. This implies $\sup\{t\ge 0 : t\in S(\omega)\}=\infty$. Hence, $\sup_{t\ge 0} F_t(\omega) = \sup_{t\in S(\omega)} F_t(\omega)$, so that holds thanks again to the estimate in . Now, since $u^*\le u_{\beta}$ (see ), $-\theta/\psi>0$, and $1-\gamma<0$, we have $$\begin{aligned} 0&\ge e^{\frac{\gamma(\psi-1)}{1-\gamma}(\zeta^{1-\gamma}-1)\int_{0}^{t}M^h_{s}ds}W_{t}\ge \delta^{\theta}\left(u_{\beta}(M_{t}^{h})\right)^{-{\theta}/{\psi}}e^{\frac{\gamma(\psi-1)}{1-\gamma}(\zeta^{1-\gamma}-1)\int_{0}^{t}M^h_{s}ds}\frac{(X_{t}^{{\mathfrak{p}}})^{1-\gamma}}{1-\gamma}\\ &=\delta^{\theta}F_t\ e^{-\alpha\int_{0}^{t}M^h_{s}ds}\frac{(X_{t}^{{\mathfrak{p}}})^{1-\gamma}}{1-\gamma}\ge \delta^{\theta}u_{\beta}\left(\frac{-\theta}{\alpha'\psi}\beta\right)^{-{\theta}/{\psi}}e^{-\alpha\int_{0}^{t}M^h_{s}ds}\frac{(X_{t}^{{\mathfrak{p}}})^{1-\gamma}}{1-\gamma}, \end{aligned}$$ where the equality follows from and , and the last inequality is due to . Recalling that ${\mathfrak{p}}\in{\mathcal{P}}={\mathcal{P}}_{2}$, we conclude from and the above inequality that $$\lim\limits_{t\rightarrow\infty}e^{-\Lambda^* t}{\mathbb{E}}\bigg[e^{\frac{\gamma(\psi-1)}{1-\gamma}(\zeta^{1-\gamma}-1)\int_{0}^{t}M_{s}^{h}ds}W_{t}\bigg] = 0.$$ On the other hand, since $M_{t}^{h}\le m e^{\beta t}$, $${\mathbb{E}}\bigg[\sup\limits_{s\in[0,t]}|W_{t}|\bigg]\le \frac{\delta^{\theta}}{|1-\gamma|}u_\beta(me^{\beta t})^{-\theta/\psi} {\mathbb{E}}\left[\sup\limits_{s\in[0,t]}(X_{s}^{{\mathfrak{p}}})^{1-\gamma}\right]<\infty,\quad \forall t\ge 0.$$ where the finiteness is a direct consequence of ${\mathfrak{p}}\in {\mathcal{P}}$. - [**Case (i)-3:**]{} $\gamma>1$ and $\zeta = 1$. In view of , $u_q \equiv k^*>0$ for any $q>0$. It then follows from that $u^*\equiv k^*>0$. The required properties then follow directly from ${\mathfrak{p}}\in{\mathcal{P}}={\mathcal{P}}_1$. [**(ii)**]{} Now, we show that $(c^*,\pi^*,h^*)\in{\mathcal{P}}$ and $W^*_t:= w(X^{c^*,\pi^*,h^*}_t,M^{h^*}_t)$ satisfies . Observe that a unique solution $M^*=M^{h^*}$ to exists. As $h^*$ by definition only depends on $u^*$, $g$, and the current mortality rate, $M^*$ is a deterministic process. Moreover, $t\mapsto M^*_t$ is strictly increasing thanks to . On the other hand, a unique solution $X^*=X^{c^*,\pi^*,h^*}$ to exists, which admits the explicit formula $$\begin{aligned} \label{X^*''} (X^{*}_t)^{1-\gamma}&=x^{1-\gamma}\exp\bigg(\int_{0}^{t}(1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} -(u^*(M^{h^*}_s)+h_{s}^*)- \frac{(1-\gamma)}{2\gamma^{2}}\left(\frac{\mu}{\sigma}\right)^{2}\right)ds\nonumber\\ &\hspace{4.25in}+ \frac{(1-\gamma)\mu}{\gamma\sigma}B_{t}\bigg). \end{aligned}$$ - [**Case (ii)-1:**]{} $\gamma\in(\frac{1}{\psi},1)$. As $M^*_t$ is strictly increasing, $u^*(M^*_t)\ge u^*(m)\ge \tilde c_0(m)$, where the second inequality follows from and . With this and $h^*_{t}\ge0$, we deduce from that holds with “$=$” therein replaced by “$\le$”. As $k^*>0$ entails $\tilde c_0(m)>0$ (see ), the same arguments in Proposition \[prop:NoAging\] can be applied to to show that $(X^*)^{1-\gamma}$ satisfies . With this, we can argue as in Case (i)-1 to show that $W^*_t:= w(X^*_t,M^*_t)$ belongs to ${\mathcal{E}}^{h^*}_{k^*}.$ - [**Case (ii)-2:**]{} $\gamma>1$ and $\zeta\neq 1$. As $u^*$ solves and $h^*$ is the maximizer of the supremum in , we have $$\begin{aligned} u^*(M_{t}^{*})-\tilde{c}_{0}(M_{t}^{*})-(\psi-1)h^*_t=\frac{M_{t}^{*}(u^*)'(M_{t}^{*})}{u^*(M_{t}^{*})}(\beta-g(h^*_t)) >0\quad \forall t>0, \end{aligned}$$ where the inequality follows from $g(h^*_t)<\beta$, thanks to . This gives $h^*_t< \frac{1}{\psi-1}(u^*(M_{t}^{*})-\tilde{c}_{0}(M_{t}^{*}))$, so that $$\label{B4} u^*(M_{t}^{*})+h^*_t <\frac{\psi}{\psi-1}u^*(M_{t}^{*}) - \frac{1}{\psi-1}\tilde{c}_{0}(M_{t}^{*})\le \frac{\psi}{\psi-1}u_\beta(M_{t}^{*}) - \frac{1}{\psi-1}\tilde{c}_{0}(M_{t}^{*}),$$ where the last inequality follows from $u^*(m)\le {u}_{\beta}(m)$ (see ). For any $\eta\in(1-\frac{1}{\gamma},1)$, consider $\alpha,\alpha'>0$ defined as in . Observe that ${u}_\beta(m)$ can be written as $${u}_\beta(m) = \beta\frac{e^{-m\frac{\psi}{\theta\beta}(1-\zeta^{1-\gamma})}\left(m\frac{\psi}{\theta\beta}(1-\zeta^{1-\gamma})\right)^{-\frac{k^*}{\beta}}}{\overline{\Gamma}\left(-\frac{k^*}{\beta},m\frac{\psi}{\theta\beta}(1-\zeta^{1-\gamma})\right)}$$ where $\overline{\Gamma}$ is the upper incomplete gamma function given by $\overline{\Gamma}(s,z):=\int_{z}^{\infty}t^{s-1}e^{-t}dt$. Similarly to the argument in [@Huang19 (A.6)-(A.7)], by using the fact $\lim\limits_{z\rightarrow\infty}\frac{\overline{\Gamma}(s,z)}{e^{-z}z^{s-1}} =1$, $$\label{lim>1} \lim\limits_{m\rightarrow\infty}\frac{\psi-1}{\psi}\frac{\left(\alpha+(\zeta^{1-\gamma}-1)\right) m}{(\gamma-1){u}_\beta(m)} = \frac{\psi-1}{\psi}\frac{\alpha+(\zeta^{1-\gamma}-1)}{(\psi-1)(\zeta^{1-\gamma}-1)} = \frac{\alpha+(\zeta^{1-\gamma}-1)}{\psi(\zeta^{1-\gamma}-1)}>1,$$ where the inequality follows from the definition of $\alpha$ and $\eta>1-\frac{1}{\gamma}$. This, together with $M^*$ being a strictly increasing deterministic process, implies the existence of $s^{*}>0$ such that $$\label{>for s>} (\alpha +(\zeta^{1-\gamma}-1))M_{s}^* > \frac{\psi(\gamma-1)}{\psi-1}{u}_\beta(M_{s}^*)\quad \hbox{for}\ s> s^{*}.$$ Consider the constant $$0\le K:= \max\limits_{t\in[0,s^*]}\left\{ \frac{\psi}{\psi-1}{u}_\beta(M^*_t) - \frac{\alpha +(\zeta^{1-\gamma}-1)}{\gamma-1}M^*_t\right\}<\infty.$$ In view of , , and the fact that $\tilde c_0(m)=k^*+(1-\psi)\frac{\zeta^{1-\gamma}-1}{1-\gamma}m$ (see ), $$\begin{aligned} &e^{-\alpha\int_{0}^{t}M^*_{s}ds}(X^{*}_t)^{1-\gamma}\\ &\le x^{1-\gamma}\text{exp}\left(\int_{0}^{t}(1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2}+\frac{k^*}{\psi-1} - \frac{\psi}{\psi-1}{u}_\beta(M_{s}^*)-\frac{\alpha +(\zeta^{1-\gamma}-1)}{1-\gamma} M_{s}^*\right)ds\right)\cdot Z_t\\ &\le x^{1-\gamma} e^{(1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} + \frac{k^*}{\psi-1} - K\right) s^*} e^{(1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} + \frac{k^*}{\psi-1} \right) (t-s^*)} Z_t, \end{aligned}$$ where $Z_t$ is the driftless geometric Brownian motion defined below , and the second inequality follows from . It follows that $$\begin{aligned} &e^{-\Lambda^* t}{\mathbb{E}}\left[e^{-\alpha\int_{0}^{t}M^*_{s}ds}(X_{t}^{*})^{1-\gamma}\right]\\ &\le x^{1-\gamma} e^{\left((1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} +\frac{k^*}{\psi-1} - K\right)-\Lambda^*\right) s^*} e^{\left((1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2}+ \frac{k^*}{\psi-1} \right)-\Lambda^*\right)(t-s^*)}\\ &= x^{1-\gamma} e^{\left((1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} + \frac{k^*}{\psi-1} - K\right)-\Lambda^*\right) s^*} e^{-(\gamma+\frac{\gamma-1}{\psi-1}) k^*(t-s^*)}\to\ 0\quad \hbox{as}\ t\to\infty, \end{aligned}$$ where the equality follows from a direct calculation using the definition of $\Lambda^*$ in , and the convergence is due to $k^*>0$. Namely, $X^*$ satisfies . On the other hand, by and $M^*_t\le me^{\beta t}$, we obtain from that $$\begin{aligned} (X^{*}_t)^{1-\gamma}\le x^{1-\gamma}\text{exp}\left(\int_0^t (1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} - \frac{\psi}{\psi-1}{u}_\beta(me^{\beta s}) \right) ds\right)\cdot Z_t, \end{aligned}$$ where $Z$ is again the driftless geometric Brownian motion defined below . By the Burkhölder-Davis-Gundy inequality, we obtain the estimate in with $-\tilde c_0(m)$ therein replaced by $\frac{\psi}{\psi-1}{u}_\beta(me^{\beta t})$. This then implies ${\mathbb{E}}\big[\sup_{s\in[0,t]}(X_{s}^{*})^{1-\gamma}\big]<\infty$, by the inequality preceding . Finally, under ${\mathbb{E}}\big[\sup_{s\in[0,t]}(X_{s}^{*})^{1-\gamma}\big]<\infty$ and , the same argument as in Case (i)-2 shows that $W^*_t:= w(X^*_t,M^*_t)$ belongs to ${\mathcal{E}}^{h^*}_{k^*}.$ - [**Case (ii)-3:**]{} $\gamma>1$ and $\zeta = 1$. By , ${u}_\beta(m) \equiv k^*>0$. As $M_{t}^*$ is strictly increasing, $\tilde c_0(M^*_t)\ge \tilde c_0(0)=k^*$. The estimate then becomes $u^*(M_{t}^{*})+h^*\le \frac{\psi}{\psi-1}k^* - \frac{1}{\psi-1}k^* = k^*$, so that we can deduce from that $$\begin{aligned} (X^{*}_t)^{1-\gamma}\le x^{1-\gamma}\text{exp}\left(\int_{0}^{t}(1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} -k^*- \frac{(1-\gamma)}{2\gamma^{2}}\left(\frac{\mu}{\sigma}\right)^{2}\right)ds + \frac{(1-\gamma)\mu}{\gamma\sigma}B_{t}\right). \end{aligned}$$ The arguments in Proposition \[prop:NoAging\] can then be applied to show that $(X^*)^{1-\gamma}$ satisfies . Then, we may argue as in Case (i)-3 to show that $W^*_t:= w(X^*_t,M^*_t)$ belongs to ${\mathcal{E}}^{h^*}_{k^*}.$ Finally, by applying Itô’s formula to $W^*_t$ and using the fact , which is a direct consequence of and $\pi^*_t\equiv \frac{\mu}{\gamma\sigma^2}$ being a constant process, we can argue as in the proof of Theorem \[Verification\] that $W^*_t$ is a solution to . Moreover, by direct calculation $$W^*_t = \delta^\theta u^*(M^*_t)^{-\theta+(1-\gamma)} \frac{(X^*_t)^{1-\gamma}}{1-\gamma} = \delta^\theta u^*(M^*_t)^{-\theta}\frac{(c^*_t)^{1-\gamma}}{1-\gamma}\le\delta^\theta \tilde c_0(M^*_t)^{-\theta}\frac{(c^*_t)^{1-\gamma}}{1-\gamma},$$ where the inequality follows from $u^*\ge \tilde c_0$ (see and ) and the fact that $\theta>0$ if $\gamma\in(\frac{1}{\psi},1)$ and $\theta<0$ if $\gamma>1$. This shows that $W^*$ satisfies with $k=k^*$. Hence, $(c^*,h^*)$ is $k^*$-admissible, and we can now conclude that $(c^*,\pi^*,h^*)$ belongs to ${\mathcal{P}}$. By Theorem \[Verification\], $v(x,m)=w(x,m)$ for all $(x,m)\in{\mathbb{R}}^2_+$ and $(c^*,\pi^*,h^*)$ optimizes . Proof of Proposition \[prop:Aging\] {#subsec:proof of prop:Aging} ----------------------------------- Define $w(x,m):=\delta^{\theta}\frac{x^{1-\gamma}}{1-\gamma}u_\beta(m)^{-\frac{\theta}{\psi}}$ for all $(x,m)\in{\mathbb{R}}^2_+$. In order to apply Theorem \[Verification\] to our setting, we need to verify all its conditions. First, $w\in C^{2,1}({\mathbb{R}}_+\times{\mathbb{R}}_+)$ by definition and it can be checked directly that $w$ solves , as $u_\beta$ is a solution to (by Lemma \[ANH\_ODE\]). In view of the definitions of ${\mathcal{P}}$ and $w$, ${\mathcal{P}}\subseteq {\mathcal{H}}_{k^*}$ and is satisfied for any $(c,\pi,h)\in{\mathcal{P}}$. By following part (i) of the proof of Theorem \[Thm:AH\], we immediately have $w(X^{c,\pi,h}_t, M^h_t)\in{\mathcal{E}}^h_{k^*}$ for any $(c,\pi,h)\in{\mathcal{P}}$ (The proof is actually simpler here, as $M^h_t=me^{\beta t}$ in the current setting). Moreover, as $w_x>0$, $w_{xx}<0$, $\bar c(x,m):=x u_\beta(m)$ and $\bar \pi(x,m) :=\frac{\mu}{\gamma\sigma^2}$ are the unique maximizers of the two supremums in , respectively. Note that the supremum in is zero, as $g\equiv 0$ and $w_x>0$. It follows that $\bar h(x,m) := 0$ trivially maximizes . It remains to show that $(c^*, \pi^*, h^*)$, defined using $\bar c$, $\bar\pi$, and $\bar h$ as in , belongs to ${\mathcal{P}}$ and $W^*_t:= w(X^{c^*,\pi^*,h^*}_t,M^{h^*}_t)$ satisfies . Observe that $M^{h^*}_t = me^{\beta t}$ as $h^*\equiv 0$, and a unique solution $X^*=X^{c^*,\pi^*,h^*}$ to exists, which satisfies the dynamics with $\tilde c_0(m)$ replaced by ${u}_\beta(me^{\beta t})$. This implies $$\label{X^*'} (X^{*}_t)^{1-\gamma}= x^{1-\gamma}\text{exp}\left(\int_{0}^{t}(1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} - {u}_\beta(me^{\beta s}) - \frac{(1-\gamma)}{2\gamma^{2}}\left(\frac{\mu}{\sigma}\right)^{2}\right)ds + \frac{(1-\gamma)\mu}{\gamma\sigma}B_{t}\right).$$ - [**Case 1:**]{} $\gamma\in(\frac{1}{\psi},1)$. As $1-\gamma>0$ and $u_\beta(m) \ge \tilde c_0(m)$ (see ), we deduce from that holds with “$=$” therein replaced by “$\le$”. As $k^*>0$ entails $\tilde c_0(m)>0$, the same arguments in Proposition \[prop:NoAging\] can be applied to show that $(X^*)^{1-\gamma}$ satisfies . With this, we can argue as in Case (i)-1 of the proof of Theorem \[Thm:AH\] to obtain $W^*_t:= w(X^*_t,M^*_t)\in {\mathcal{E}}^{h^*}_{k^*}.$ - [**Case 2:**]{} $\gamma>1$ and $\zeta\neq 1$. For any $\eta\in(1-\frac{1}{\gamma},1)$, consider the constant $\alpha>0$ defined in . Similarly to , using the fact that $\lim_{z\rightarrow\infty}\frac{\overline{\Gamma}(s,z)}{e^{-z}z^{s-1}} =1$ yields $$\label{lim>1'} \lim\limits_{m\rightarrow\infty}\frac{\alpha m}{(\gamma-1)\tilde{u}(m)} = \frac{\alpha}{(\psi-1)(\zeta^{1-\gamma}-1)} >1.$$ where the inequality follows from the definition of $\alpha$ and $\eta>1-\frac{1}{\gamma}$. This implies that there exists some $s^{*}>0$ such that $$\label{M dominates} \alpha me^{\beta s} \ge (\gamma-1)\tilde{u}(me^{\beta s})\quad \hbox{for all $s\ge s^{*}$}.$$ Consider the constant $$0\le K:= \max\limits_{t\in[0,s^*]}\left\{\tilde{u}(me^{\beta t}) - \frac{\alpha m e^{\beta t}}{\gamma-1}\right\}<\infty.$$ Now, as $M_t=me^{\beta t}$, we deduce from that $$\begin{aligned} &e^{-\alpha\int_{0}^{t}M_{s}ds}(X^{*})^{1-\gamma}\\ &=x^{1-\gamma}\exp\left(\int_{0}^{t}(1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} - {u}_\beta(me^{\beta s}) -\frac{\alpha me^{\beta s}}{(1-\gamma)} \right)ds \right)\cdot Z_t\\ &\le x^{1-\gamma} e^{(1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} - K\right)s^*} e^{(1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2}\right)(t-s^*)}Z_t, \end{aligned}$$ where $Z_t$ is the driftless geometric Brownian motion defined below , and the inequality follows from . It follows that $$\begin{aligned} e^{-\Lambda^* t}{\mathbb{E}}\left[e^{-\alpha\int_{0}^{t}M_{s}ds}(X_{t}^{*})^{1-\gamma}\right] &\le x^{1-\gamma} e^{\left((1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} - K\right)-\Lambda^*\right) s^*} e^{\left((1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2}\right)-\Lambda^*\right)(t-s^*)}\\ &= x^{1-\gamma} e^{\left((1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} - K\right)-\Lambda^*\right) s^*} e^{-\gamma k^*(t-s^*)}\to\ 0,\quad \hbox{as}\ t\to\infty, \end{aligned}$$ where the second line follows from a direct calculation using the definition of $\Lambda^*$ in , and the convergence is due to $k^*>0$. On the other hand, similarly to , we rewrite as $$\begin{aligned} (X^{*}_t)^{1-\gamma}= x^{1-\gamma}\text{exp}\left(\int_0^t (1-\gamma)\left(r+\frac{1}{2\gamma}\left(\frac{\mu}{\sigma}\right)^{2} - {u}_\beta(me^{\beta s}) \right) ds\right)\cdot Z_t, \end{aligned}$$ where $Z$ is again the driftless geometric Brownian motion defined below . By Burkhölder-Davis-Gundy’s inequality, we obtain the estimate in with $-\tilde c_0(m)$ therein replaced by ${u}_\beta(me^{\beta t})$. This then implies ${\mathbb{E}}\big[\sup_{s\in[0,t]}(X_{s}^{*})^{1-\gamma}\big]<\infty$, by the inequality preceding . Under ${\mathbb{E}}\big[\sup_{s\in[0,t]}(X_{s}^{*})^{1-\gamma}\big]<\infty$ and , the same argument as in Case (i)-2 of the proof of Theorem \[Thm:AH\] shows that $W^*_t:= w(X^*_t,M^*_t)$ belongs to ${\mathcal{E}}^{h^*}_{k^*}.$ - [**Case 3:**]{} $\gamma>1$ and $\zeta=1$. By , ${u}_\beta(m) \equiv k^*>0$. Then, in view of , we can apply the same arguments as in Proposition \[prop:NoAging\] to show that $(X^*)^{1-\gamma}$ satisfies . With this, we may argue as in Case (i)-3 in the proof of Theorem \[Thm:AH\] to obtain $W^*_t:= w(X^*_t,M^*_t)\in{\mathcal{E}}^{h^*}_{k^*}.$ Finally, by applying Itô’s formula to $W^*_t$ and using the fact , which is a direct consequence of and $\pi^*_t\equiv \frac{\mu}{\gamma\sigma^2}$ being a constant process, we can argue as in the proof of Theorem \[Verification\] that $W^*_t$ is a solution to . Moreover, by direct calculation $$W^*_t = \delta^\theta u_\beta(M^*_t)^{-\theta+(1-\gamma)} \frac{(X^*_t)^{1-\gamma}}{1-\gamma} = \delta^\theta u_\beta(M^*_t)^{-\theta}\frac{(c^*_t)^{1-\gamma}}{1-\gamma}\le\delta^\theta \tilde c_0(M^*_t)^{-\theta}\frac{(c^*_t)^{1-\gamma}}{1-\gamma},$$ where the inequality follows from $u_\beta \ge \tilde c_0$ (see ) and the fact that $\theta>0$ if $\gamma\in(\frac{1}{\psi},1)$ and $\theta<0$ if $\gamma>1$. This shows that $W^*$ satisfies with $k=k^*$. Hence, $(c^*,h^*)$ is $k^*$-admissible, and we can now conclude that $(c^*,\pi^*,h^*)$ belongs to ${\mathcal{P}}$. By Theorem \[Verification\], $v(x,m)=w(x,m)$ for all $(x,m)\in{\mathbb{R}}^2_+$ and $(c^*,\pi^*,h^*)$ optimizes . [^1]: University of Colorado, Department of Applied Mathematics, Boulder, CO 80309-0526, USA, email: `joshua.aurand@colorado.edu`. [^2]: University of Colorado, Department of Applied Mathematics, Boulder, CO 80309-0526, USA, email: `yujui.huang@colorado.edu`. Partially supported by National Science Foundation (DMS-1715439) and the University of Colorado (11003573). [^3]: Similar pattern is observed in various countries, as displayed in Figure \[fig:calibration\]. [^4]: As an exception, the literature on health capital, initiated by [@Grossman1972], considers endogenous healthcare. Despite its lasting development towards more realistic models (see e.g. [@EhrlichChuma1990], [@Ehrlich2000], [@Yogo2016], [@HugonnierPelgrinSt-Amour2012], and [@hall2007value]), the Gompertz law of mortality remains largely absent.

--- address: - 'Theory Division, CERN, Geneva 23, Switzerland' - 'Inst. of Theoretical Physics, Warsaw University, Hoża 69, 00-681 Warsaw, Poland' author: - 'PIOTR H. CHANKOWSKI [^1]' - STEFAN POKORSKI title: 'WHAT IS THE MASS OF THE LIGHTEST SUPERSYMMETRIC HIGGS BOSON? [^2]' --- \#1\#2\#3\#4[[\#1]{} [**\#2**]{}, \#3 (\#4)]{} CERN-TH/97-28\ hep–ph/9702431\ Introduction ============ Spontaneous breaking of the electroweak gauge symmetry $SU(2)\times U(1)$ is now confirmed experimentally with one per mille accuracy (see the Chapter by A. Blondel). However, the actual mechanism of this symmetry breaking still remains unknown and waits for experimental discovery. This is, by far, the most central question to particle physics and, in particular, to the experimental programs at LEP 2 and the LHC. It is very likely that the understanding of the mechanism of the electroweak symmetry breaking will not only provide us with the missing link in the Standard Model but, also, will be an important bridge to physics beyond it. The minimal model for spontaneous electroweak gauge symmetry breaking is the Higgs mechanism, whose minimal version (the minimal Standard Model) requires one scalar $SU(2)$ doublet (Higgs doublet). The scalar potential at some scale $\Lambda$ is: $$\begin{aligned} V(\Lambda) = m^2(\Lambda) |H(\Lambda)|^2 + {1\over2}\lambda(\Lambda)|H(\Lambda)|^4\end{aligned}$$ with the dependence on $\Lambda$ controlled by the renormalization group evolution (RGE). The mass of the physical scalar (Higgs boson) is $M^2_{\phi^0}= \lambda(M_Z)v^2(M_Z)$ where $v(M_Z)=\sqrt{4M^2_W/g^2_2(M_Z)}\approx246$ GeV. There exist the well known theoretical bounds on the Higgs boson mass (see [@QUIROS] for an extensive discussion) which follow from certain constraints on the behaviour of the self-coupling $\lambda(\Lambda)$. One can distinguish two types of bounds. The most general upper bound on $M_{\phi^0}$ follows from the requirement that the Standard Model is a unitary and weakly interacting theory at the energy scale ${\cal O}(M_Z)$. We get then $M_{\phi^0}\simlt{\cal O}(1$ TeV). Stronger bounds are $\Lambda$-dependent and are known under the names of the triviality (upper) bound and the vacuum stability (lower) bound. They follow respectively from the requirements that the theory remains perturbative ($\lambda(\Lambda)<16\pi^2$) and the vacuum remains stable ($\lambda(\Lambda)>0$) up to a certain scale $\Lambda$. Those bounds are particularly interesting in the presence of the heavy top quark, $m_t=(175\pm6)$ GeV. They are shown in Fig. 3 in ref. [@QUIROS] and lead to the striking conclusions: We see that the discovery of a light Higgs boson ($M_{\phi^0}\simlt80$ GeV) or a heavy one ($M_{\phi^0}\simgt500$ GeV) would be a direct information about the existence of new physics below the scale $\Lambda\sim {\cal O}(1$ TeV) (or at least of a strongly interacting Higgs sector). On the other hand, if the SM in its perturbative regime is to be valid up to very large scales $\Lambda$, of the order of the GUT scale $\Lambda\approx10^{16}$ GeV, one gets strong bounds 140 GeV $\simlt M_{\phi^0}\simlt180$ GeV. In this case we face the well known hierarchy problem in the SM: $M_{\phi^0}$, $v$ $\ll\Lambda$ and it is difficult to understand how the scalar potential remains stable under radiative corrections of the full theory. One way or another, the bounds on $M_{\phi^0}$ in the SM are strongly suggestive that the mechanism of spontaneous electroweak gauge symmetry breaking is directly related to the existence of a new scale (not much above the electroweak scale) in fundamental interactions. The central question can then be phrased as this: discover and investigate the next scale in fundamental interactions. Is it the scale of new strong interactions (strongly interacting Higgs sector or techicolour interactions or compositness scale) or the scale of soft supersymmetry breaking? We would like to stress the basic difference between these two lines of approach. In the first one, the new scale is also a cut-off scale for the perturbative validity of the electroweak theory. Supersymmetry offers a solution to the hierarchy problem while maintaining the perturbative nature of the theory up to the GUT or even Planck scale. This is a welcome feature if such facts as the gauge coupling unification are not to be considered as purely accidental. Another difference is in the expectations for the Higgs boson mass: in the strong interaction scenarios it is naturally heavy, with its mass close to the new scale $\Lambda$. In supersymmetric extensions of the SM the lightest Higgs scalar $h^0$ generically remains light, $M_{h^0}\sim {\cal O}(100$ GeV) and only logarithmically correlated with the scale of the soft supersymmetry breaking. It is the purpose of this Chapter to summarize the predictions for the Higgs sector in supersymmetric extensions of the SM. The Higgs sector in the Minimal Supersymmetric Standard Model (MSSM), considered as a low energy effective theory with all free parameters totally unconstrained has been discused in detail in [@HABER]. For the sake of easy reference and to establish our notation we summarize those results in Section 2. The main conclusion of that Section is the specification of the set of parameters which determine the Higgs boson masses and the existence of general upper bounds for the mass of the lightest supersymmetric Higgs boson $h^0$. Next, in the main part of this article, we collect the available results and arguments which constrain the general parameter space relevant for the Higgs mass in the MSSM and, therefore, give more specific predictions for $M_{h^0}$. One should stress that our interest in the MSSM is well motivated. It is structured in such a way that the success of the SM in describing the precision electroweak data is maintained. Moreover, its virtue is that it can be extrapolated up to the large energy scales (the scales where the soft supersymmetry breaking terms are generated) in unambigous and quantitative way. We shall mainly consider the supergravity scenario for supersymmetry breaking in which the MSSM is extrapolated up to the GUT scale. A brief discussion of the gauge mediated symmetry breaking models [@DIFISR] is, however, also included. Broadly speaking, the weak scale - large scale connection is the main source of constraints on the Higgs sector parameter space which we are going to present. Finally, in Section 5, we discuss the predictions for the Higgs sector in non-minimal versions of the supersymmetric extensions of the Standard Model. Higgs sector in MSSM - a brief summary ====================================== In the Minimal Supersymmetric Standard Model the Higgs sector is particularly simple and predictive. Supersymmetry and the minimal particle content imply that it consists of two Higgs doublets, each coupled to only one type ($H_1(H_2)$ couples to the down (up)) of fermions. The scalar Higgs potential reads: $$\begin{aligned} V = m^2_1 \overline H_1H_1 + m^2_2 \overline H_2H_2 + m_3^2 \left(\epsilon_{ab}H_1^aH_2^b + c.c\right)\nonumber\\ +{1\over8}(g_1^2+g_2^2)(\overline H_1H_1-\overline H_2H_2)^2 +{1\over2}g^2_2|\overline H_1H_2|^2 \label{eqn:treepot}\end{aligned}$$ where $\epsilon_{12}=-1$ and $m_1^2$, $m_2^2$ and $m_3^2$ are the soft supersymmetry breaking mass parameters. The crucial point about the potential (\[eqn:treepot\]) is that its quartic couplings are the electroweak gauge couplings (i.e. there is no $F$-term contribution to the scalar Higgs potential). The only free parameters are the three mass parameters. The tree level mass eigenstates of the Higgs bosons are: two $CP$-even ($h^0$, $H^0$), one $CP$-odd ($A^0$) and 2 charged ($H^\pm$) physical particles and three Goldstone bosons “eaten up” by the gauge bosons. An important parameter is $\tan\beta\equiv v_2/v_1$ where $v_i$ minimize the tree level potential (\[eqn:treepot\]) and are given by $v_1=v\cos\beta$, $v_2=v\sin\beta$ with $$\begin{aligned} v^2={8\over g^2_1+g_2^2}{m^2_1 - m^2_2\tan^2\beta\over\tan^2\beta-1} \label{eqn:vsquared}\end{aligned}$$ $$\begin{aligned} \sin2\beta={2m^2_3\over m^2_1+m^2_2} \label{eqn:sinbeta}\end{aligned}$$ Since $v$ is fixed by the $Z^0$ mass, all physical Higgs boson masses are expressed in terms of only two free parameters. They can be taken e.g. as $\tan\beta$ and the mass $M^2_{A^0}$ of the $CP-$odd Higgs scalar $A^0$ given by $M^2_{A^0}=m^2_1+m^2_2$. The $CP-$even Higgs boson masses then read: $$\begin{aligned} M^2_{h^0,H^0}={1\over2}\left(M^2_{A^0} + M^2_{Z^0} + \sqrt{\left(M^2_{A^0}-M^2_{Z^0}\right)^2- 4M^2_{A^0}M^2_{Z^0}\cos^22\beta} \right) \label{eqn:treemass}\end{aligned}$$ leading to the bound $M_{h^0}<M_{Z^0}$ and to the “natural” (i.e. independent of any other parameters) relation $M^2_{h^0}+M^2_{H^0}=M^2_{A^0}+M^2_{Z^0}$. The other relation is $M^2_{H^\pm}=M^2_{W^\pm} + M^2_{A^0}$ [@LISH]. The origin and the magnitude of radiative corrections to the Higgs boson masses can be easily understood. Let $M$ be the scale of the soft supersymmetry breaking sfermion masses. Neglecting terms suppressed by inverse powers of $M$, the dominant one-loop corrections to the effective potential $V_{eff}$, due to the top and stop loops, can be absorbed into renormalization of the parameters in the Higgs potential. One gets: $$\begin{aligned} V = \tilde m^2_1 \overline H_1H_1 + \tilde m^2_2 \overline H_2H_2 + \tilde m_3^2 \left(\epsilon_{ab}H_1^aH_2^b + c.c\right)\nonumber\\ + \lambda_1 |H_1|^4 +\lambda_2 |H_2|^4 +\lambda_3 |H_1|^2 |H_2|^2 +\lambda_4 |\overline H_1H_2|^2 \label{eqn:corrpot}\end{aligned}$$ The appearence of other couplings is protected by the symmetries of the model. It is clear on the dimensional grounds that $$\begin{aligned} \delta m^2_i = \tilde m^2_i - m^2_i \sim {\cal O}(M^2)\end{aligned}$$ They are logarithmically divergent but can be absorbed into the free parameters of the model. The corrections $\delta\lambda_i$ defined by $$\begin{aligned} \delta\lambda_1 = \lambda_1 - {1\over8}(g_1^2+g_2^2), ~~~~~~~~ \delta\lambda_2 = \lambda_2 - {1\over8}(g_1^2+g_2^2)\nonumber\\ \delta\lambda_3 = \lambda_3 + {1\over4}(g_1^2+g_2^2), ~~~~~~~~ \delta\lambda_4 = \lambda_4 - {1\over2}g_2^2\end{aligned}$$ are all ${\cal O}(\log M)$. Moreover, from the non-renormalization theorem, the corrections $\delta\lambda_i$ are calculable (finite) in terms of the remaining parameters of the model. From the top and stop loops with attached four Higgs boson legs one gets $$\begin{aligned} \delta\lambda_i\sim {12\over16\pi^2} h^4_t\log\left({M^2_{\tilde t}\over m_t^2}\right)\end{aligned}$$ where $h_t$ is the top quark Yukawa coupling (factor 12 comes from 4 top degrees of freedom multipied by the color factor of 3) and $M_{\tilde t}$ denotes the scale of the stop masses. Thus, the correction to the $h^0$ mass is [@OKYAYA] $$\begin{aligned} \delta M^2_{h^0}\sim {\cal O}\left({6g_2^2\over16\pi^2}{m_t^4\over M^2_W} \log\left({M^2_{\tilde t}\over m_t^2}\right)\right)\end{aligned}$$ In general, taking into account the full structure of the stop mass matrix, the lightest Higgs boson mass in the MSSM is parametrized by $$\begin{aligned} M_{h^0} = M_{h^0} \left(M_{A^0},\tan\beta,m_t,M_{\tilde t_1},M_{\tilde t_2},A_t,\mu,...\right)\end{aligned}$$ where $M_{\tilde t_i}$ are the physical stop masses, $A_t$ and $\mu$ determine their mixing angle (as well as some of their trilinear couplings to the Higgs bosons) and ellipsis stand for other parameters whose effects are not dominant (e.g. the sbottom sector parameters). In Fig. \[fig:mh\]a we show $M_{h^0}$ as a function of $M_{A^0}$ for two values of $\tan\beta$ and $M_{\tilde t_1}=M_{\tilde t_2}=1$ TeV, $\mu=0$ and two values of the $A_t$ parameter. We see that maximal $M_{h^0}$ is always obtained for $M_{A^0}\gg M_{Z^0}$ (in practice, the bound is saturated for $M_{A^0}\simgt250$ GeV). In this limit one gets from the effective potential approach (see ref. [@HABER] for details) particularly simple result for the one-loop corrected $M_{h^0}$ [@HE]: $$\begin{aligned} M^2_{h^0}=M^2_{Z^0}\cos^22\beta+{3\alpha\over4\pi s^2_W}{m^4_t\over M^2_W} \left[\log\left({M^2_{\tilde t_1}M^2_{\tilde t_2}\over m^4_t}\right) +\left({M^2_{\tilde t_1}-M^2_{\tilde t_2}\over2m^2_t}\sin^22\theta_{\tilde t} \right)^2\right.\nonumber\end{aligned}$$ $$\begin{aligned} \times\left. f(M^2_{\tilde t_1},M^2_{\tilde t_2}) + {M^2_{\tilde t_1}-M^2_{\tilde t_2}\over2m^2_t}\sin^22\theta_{\tilde t} \log\left({M^2_{\tilde t_1}\over M^2_{\tilde t_2}}\right)\right] \label{eqn:mhha}\end{aligned}$$ where $f(x,y)=2 - (x + y)/(x - y) \log(x/y)$. For large $M_{A^0}$ the separate dependence on the parameters $A_t$ and $\mu$ has disappeared and is replaced by the effective dependence on the left-right stop mixing angle $\theta_{\tilde t}$. One should also mention (see [@HABER]) that the two-loop corrections to $M_{h^0}$ are typically ${\cal O}(20$%) of the one-loop corrections and are negative. In Figs. \[fig:mh\] and \[fig:bsgamma\] they are taken into account in the approach proposed in [@CAESQUWA].The dependence on $\tan\beta$, illustrated in Fig. \[fig:mh\]b, is important for our further discussion. Experimental constraints on the parameters of the MSSM ====================================================== Having recalled the general parameter set relevant for the Higgs sector in supersymmetric models we proceed now to discuss constraints on those parameters which follow from various additional considerations. The parameters of the MSSM which enter into the prediction for $M_{h^0}$ are partially constrained by other experimental data. Although at present relatively weak, those are interesting constraints which correlate in the framework of the MSSM the Higgs boson mass(es) with other measurements without any additional assumptions. There are also stronger constraints which follow from embedding the MSSM into a SUSY GUT scenario, or more generally, from extrapolating the MSSM up to the GUT scale, supplemented with several plausible (simple) assumptions about physics at that scale. We begin our discussion with the former. A number of experimental data constrains the tree level parameters $M_{A^0}$ and $\tan\beta$. These are, first of all, $BR(b\rightarrow s\gamma)$ [@BUMIMUPO], $\Gamma(Z^0\rightarrow\overline bb)$ [@BUFI] and (for large $\tan\beta$) $BR(b\rightarrow c\tau\overline\nu_{\tau})$ [@KRPO]. Other processes (e.g. $B^0$-$\overline B^0$ mixing) give weaker constraints. The first two processes are sensitive to the $M_{H^\pm}$ and $\tan\beta$ via the charged Higgs boson - top quark loop contribution and the third one via the tree level $H^{\pm}$ exchange. The $btH^-$ coupling is given by $$\begin{aligned} {\cal L}_{btH^\pm} ={g_2\over\sqrt2M_W}H^+ \overline\psi_t \left(m_t\cot\beta P_L + m_b\tan\beta P_R\right)V_{tb}\psi_b + h.c. \label{eqn:coupl1}\end{aligned}$$ where $P_{L,R}\equiv(1\mp\gamma^5)/2$ (the coupling $bcH^\pm$ is similar with $\psi_t\rightarrow\psi_c$ and $V_{tb}\rightarrow V_{cb}$) and the coupling $\tau\nu_\tau H^\pm$ is given by: $$\begin{aligned} {\cal L}_{\tau\nu_\tau H^\pm} ={g_2\over\sqrt2}{m_\tau\over M_W}\tan\beta H^+ \overline\psi_{\nu_\tau} P_R\psi_\tau + h.c. \label{eqn:coupl}\end{aligned}$$ We see that $BR(b\rightarrow c\tau\overline\nu_{\tau})$ can be enhanced for large $\tan\beta$ and light $H^\pm$ and the measurement $BR(b\rightarrow c\tau\overline\nu_{\tau})=2.69\pm0.44$% gives the bound [@KRPO] $$\begin{aligned} \tan\beta\simlt0.52{M_{H^\pm}\over 1 ~{\rm GeV}}\end{aligned}$$ which is essentially independent of the other parameters of the MSSM and is shown in Fig. \[fig:bsgamma\]a by the dotted line. The first two processes get contributions also from diagrams with superpartners in the loop (the dominant one may come from the chargino-stop loop, due to the large Yukawa coupling for the higgsino component with the right-handed stop) and, in consequence, the exclusion regions in the $(M_{A^0}, \tan\beta)$ plane depend on supersymmetric parameters. The strongest dependence is on the $m_{C^\pm}$, $M_{\tilde t_i}$, $\theta_t$. In the limit of very heavy superpartners ($m_{C^\pm_i}$, $M_{\tilde t_i}\simgt$1 TeV) we get the exclusion limits shown in Fig. \[fig:bsgamma\]a by the solid curves. The dependence of these limits on $m_{C^\pm_1}$, $M_{\tilde t_1}$ (with all other superpartners at 1 TeV) is illustrated by the dashed (dot-dashed) curve which is obtained for $m_{C^\pm_1}=M_{\tilde t_1}=$500 (250) GeV. In each case a scan over the chargino composition and the left-right mixing angle in the stop sector have been performed. Negative contribution of a light charged Higgs boson - top quark loop to $R_b\equiv\Gamma(Z^0\rightarrow\overline bb)/\Gamma(Z^0\rightarrow hadrons)$ becomes too large for small values of $\tan\beta$ (see eq. (\[eqn:coupl1\])) [@BUFI] excluding, in the MSSM with heavy spectrum, the region of the ($\tan\beta$, $M_{A^0}$) plane to the left of the almost vertical solid line shown in Fig. \[fig:bsgamma\]a (we require that $R_b$ remains within $2\sigma$ of the presently measured value $R_b^{exp}=0.2179\pm0.0012$ [@LEPEWWG]). Positive contribution of the chargino - stop loop weakens significantly this bound (if both are light) as shown for $m_{C_1^\pm}=M_{\tilde t_1}=$250 GeV by the almost vertical dot-dashed line. Similar effects are observed for $BR(b\rightarrow s\gamma)$. In the limit of heavy superpartners one obtains strong bound [^3] on $M_{A^0}$: the $H^\pm$ - top loop adds positively to the SM contribution which, by itself, is in the upper edge of the experimentally allowed region. The bound is weakened in the presence of a light chargino - stop loop (which can interfere negatively with the $H^\pm$ -top loop) and even totally disappears for large values of $\tan\beta$. One should note, however, that for $\tan\beta\sim m_t/m_b$ the interference term is generically very large and consistency with the data requires a large amount of fine-tuning in the $(\theta_{\tilde t}, M_{\tilde t_1})$ parameter space [@CHPO] [^4]. Thus, one concludes that the large $\tan\beta$ scenario is unlikely to be consistent with a light $CP$-odd Higgs boson. For $M_{h^0}$ this implies the plateau region in Fig. \[fig:mh\]a. (The present experimental bound [@ALEPH] is $M_{A^0}\simgt55$ GeV for $\tan\beta\simgt50$.) Radiative corrections to $M_{h^0}$ are mainly dependent on the stop masses (for large $\tan\beta$ also on the sbottom mass) and on the parameters $A_t$ and $\mu$ (in the large $M_{A^0}$ limit they depend only on the combination $A_t+\mu\cot\beta$ which can be traded for $\theta_{\tilde t}$ as in eq. (\[eqn:mhha\])) which are constrained by the precision data. A light left-handed stop would introduce additional source of the custodial $SU_V(2)$ breaking. Since the SM fit to the LEP precision data is very good with the $SU_V(2)$ breaking given mainly by the $t$-$b$ mass spliting, additional sources of the custodial $SU_V(2)$ breaking would tend to destroy the quality of the SM fit. In Fig. \[fig:bsgamma\]b we show by the solid lines the limits in the ($M_{\tilde t_2}$, $\theta_{\tilde t}$) plane for $M_{\tilde t_1}=100$ GeV obtained from the requirement that $\Delta\chi^2<4$ compared with the minimum of a fit to the electroweak observables found for heavy $\tilde t_2$. Similar bounds exist for heavier $\tilde t_1$. In Fig. \[fig:bsgamma\]b, those limits are shown together with the contours of constant $M_{h^0}$. In summary, the present experimental data do not significantly improve the general upper bound on $M_{h^0}$. However they give constraints on SUSY parameters and there are interesting correlations between parameter regions allowed by other processes (e.g. a light $CP$-odd Higgs boson is consistent with $BR(b\rightarrow s\gamma)$ and $\Gamma(Z^0\rightarrow\overline bb)$ only if chargino and stop are also light) and the Higgs boson mass $M_{h^0}$. The weak scale - large scale connection ======================================= The parameter space of the low energy MSSM can be further reduced by introducing additional theoretical ideas. The first one involves the extrapolation of the MSSM up to very high energy scales (the “desert” hypothesis) and the observation that to a very good approximation the $SU(3)\times SU(2)\times U(1)$ gauge couplings converge to a common value. With the supersymmetry breaking scale of order 1 TeV or less the unification takes place at an energy scale of order $10^{16}$ GeV and depends weakly on the details of the GUT-scale theory. One can rephrase this result by saying that e.g. the Weinberg angle is correctly predicted in terms of the measured values of $\alpha$ and $\alpha_3$ from the hypothesis of the gauge coupling unification. This is one of the most compelling hints for the low energy supersymmetry and, this should be strongly stressed, for the fact that physics remains perturbative up to the GUT scale $\sim10^{16}$ GeV. Once assumed, the perturbative validity of the MSSM up to the scale $\sim10^{16}$ GeV has several interesting implications for the behaviour of the third generation Yukawa couplings and for the “interesting” (i.e. most plausible?) values of $\tan\beta$ and, in consequence, also for $M_{h^0}$. The first important notion is that of the quasi-infrared fixed point for the top quark (or top and bottom quarks for large $\tan\beta$) Yukawa coupling. We recall the fixed point structure of the top quark Yukawa coupling $Y_t$   ($Y_t\equiv h_t^2/4\pi$) in the MSSM. The renormalization group equations have the form: $$\begin{aligned} {dY_t\over dt}&=& Y_t (a_i^u\alpha_i - c_tY_t)\nonumber\\ {d\alpha_i\over dt}&=&-b_i\alpha_i^2 \label{eqn:yukrge}\end{aligned}$$ where $2\pi t=\log (M_{GUT}/Q)$,   $a_i^u=(13/15,3,16/3)$, $b_i=(11,1,-3)$, $c_t=6$ and $\alpha_i=g_i^2/4\pi$. Ignoring the smaller electroweak couplings, $Y_t$ is related to the QCD coupling $\alpha_3$: the fixed point solution for the ratio $Y_t/\alpha_3$ reads [@PERO]: $$Y_t^F(t) = {a^u_3 + b_3\over c_t}\alpha_3(t)$$ One can also solve eqs. (\[eqn:yukrge\]) explicitly [@IBLO]: $$Y_t(t) = {4\pi Y_t(0)E(t)\over4\pi+c_tY_t(0)F(t)}$$ with $$E(t)\approx\left({\alpha_3(0)\over\alpha_3(t)}\right)^{{a^u_3\over b_3}},~~~~~ F(t) = \int_0^t E(t')dt'$$ It may happen that $Y_t^F(t)$ is not reached because of too short a running but, nevertheless, $c_tY_t(0)F(t)\gg4\pi$ and $$Y_t(t)\approx Y_t^{QF}(t) = {4\pi E(t)\over c_tF(t)}$$ i.e.the low energy value of $Y_t$ no longer depends on the initial value $Y_t(0)$. This is called the quasi-infrared fixed point solution (Q-IR) [@HI] and we have: $$Y_t^{QF}(t)\approx Y_t^F(t){1\over1-\left({\alpha_3(t)\over\alpha_3(0)} \right)^{1+ {a^u_3\over b_3}}}$$ This situation indeed occurs in the MSSM (for $b_3$, $a^u_3$, $c_t$ of the MSSM) for $Y_t(0)\simgt\cal O$(0.1) (see Fig. \[fig:irfix\]a) i.e. for the initial values still in the perturbative regime! Thus, not only $Y_t^{QF}(M_Z)$ is the upper bound for $Y_t(M_Z)$ but it can be reached at the limit of perturbative physics [@ALPOWI]. The Q-IR prediction for the running top quark mass in the ${\overline MS}$ scheme, for $\alpha_3(M_Z)$= 0.11-0.13 and small and moderate values of $\tan\beta$ is approximately given by [@CAWA]: $$m_t^{QF}(m_t)\approx196 {\rm GeV}\left[1+2(\alpha_3(M_Z)-0.12)\right]\sin\beta \label{eqn:mtopir}$$ The physical top quark mass (pole mass) is obtained by including QCD corrections which contribute ${\cal O}$(10 GeV) to the final result. Eq. (\[eqn:mtopir\]) combined with $m_t=(175\pm6)$ GeV gives us a lower bound on $\tan\beta$: $\tan\beta\simgt1.4$ [^5] Similar results are also found in the large $\tan\beta$ region in which both the top and bottom Yukawa couplins are large[^6]. In the numerical calculation two-loop RG equations are used. It is clear from eq. (\[eqn:mtopir\]) that the experimental value of $m_t\approx 175$ GeV is very close to its perturbative upper bound in the MSSM. To know how close, we need to know $\tan\beta$ but one is tempted to speculate that $m_t\approx m^{QF}_t$ and then $\tan\beta$ must be either close to its lower bound or very large. An interesting observation is that the GUT assumption about unification of the bottom and tau Yukawa couplings gives independent support to the idea that $Y_t(M_Z)=Y^{QF}_t(M_Z)$ [@AR; @CAOLPOWA1]. Quantitatively, this conclusion depends on the values of $\alpha_3(M_Z)$ and $m_b(pole)$ and on the threshold corrections to the relation $Y_b=Y_{\tau}$. Generically, however, strong interaction renormalization effects for $Y_b$ are too strong and large top quark Yukawa coupling contribution to the running of $Y_b$ is needed to balance them (see Fig. \[fig:irfix\]b). For $Y_b=Y_{\tau}$ within 10%, $\alpha_3(M_Z)>0.11$ and $m_b(pole)<$5.2 GeV one gets $m_t\approx m_t^{QF}$ within 10%[^7]. Thus, the possibility of $m_t\approx m_t^{QF}$ is supported by several independent arguments (also models for dynamical determination of $Y_t$ give values close to the IR fixed point [@KUZWPA]) and, together with the measured value of $m_t=175\pm6$ GeV, it makes the region $1.5\simlt\tan\beta\simlt2$ particularly interesting. Another interesting region of $\tan\beta$ is $\tan\beta\approx m_t/m_b$. Large values of $\tan\beta$ have been discussed for some time as a solution to the $m_t/m_b$ hierarchy with full unification of the third generation Yukawa couplings $Y_t=Y_b=Y_\tau$ [@OLPO; @BA]. Such unification is predicted e.g. by simple versions of the SUSY $SO(10)$ GUT [@OLPO; @ANRADIHA] and remains to be an interesting possibility. In conclusion, although one cannot rule out intermediate values of $\tan\beta$, the arguments given above are compelling enough to consider the low and large $\tan\beta$ regions as “interesting” ones. In those regions the Higgs boson masses are bounded more strongly than in the general MSSM. For $m_t=175$ GeV, from Fig. \[fig:mh\]b we see that if $Y_t$ is $\simlt10$% away from its IR value (which corresponds to $\tan\beta\simlt1.85$) we have $M_{h^0}\simlt100$ GeV. For large $\tan\beta$ we have $M_{h^0}\approx M_{A^0}$ up to $\sim 120$ GeV and for heavier $M_{A^0}$ (for arguments in favour of such $M_{A^0}$, see the previous Section) the $M_{h^0}$ remains constant (independent of $M_{A^0}$) with the value fixed by radiative corrections. Finally, we are going to discuss several constraints on the range of the soft supersymmetry breaking masses $m^2_1$, $m^2_2$, $m^2_3$, the top squark masses and the mixing parameters $A_t$, $\mu$ (i.e. the remaining parameters relevant for $M_h$) which follow from the extrapolation of the MSSM to high energies. They are particularly interesting and easy to discuss under the assumption that the $h_t$ is not too far from its quasi-infrared fixed point limit. For the sake of definitness we shall mainly focus on the low and intermediate $\tan\beta$ region and present analytic results in the 1-loop approximation [@CAOLPOWA1; @CACHOLPOWA]. More complete 2-loop numerical calculations [@CAOLPOWA1; @CAWA] confirm very well these analytic considerations. In the 1-loop approximation and expanding in $y(t)\equiv Y_t(t)/Y^{QF}_t(t)$ the RG equations for the dimensionful parameters can be solved analytically [@CAOLPOWA1; @CACHOLPOWA]. Denoting by ${{\overline m}}^2(t)\equiv m^2_Q(t) + m^2_U(t) + m^2_{H_2}(t)$, $2\pi t\equiv\log{M\over Q}$ with $M=M_{GUT}$ or any intermediate scale, and with $m_K^2$, $K = H_i,Q,U,D,L,E$, standing for the soft supersymmetry breaking mass parameters of the Higgs, left-handed squark, right-handed up-type squark, right-handed down-type squark, left-handed slepton and right-handed slepton, respectively we get the the following results: $$\begin{aligned} {{\overline m}}^2(t) &=& (1-y){{\overline m}}^2(0) - y(1-y)A_t(0)\left(A_t(0)-2\hat\xi M_{1/2}\right) \nonumber\\ &+& (\overline\eta - y\hat\eta + y^2\hat\xi^2) M_{1/2}^2, \label{eqn:m2ov0}\end{aligned}$$ $$\begin{aligned} m_K^2(t) &=&m_K^2(0) - {c_K\over c_t}y{{\overline m}}^2(0) - {c_K\over c_t} y(1-y)A_t(0)\left(A_t(0)-2\hat\xi M_{1/2}\right) \nonumber\\ &+&\left[\eta_K-{c_K\over c_t}\left(y\hat\eta+y^2\hat\xi^2\right) \right]M_{1/2}^2 + D_K, \label{eqn:m2k0}\end{aligned}$$ $$\begin{aligned} D_K = -\kappa_K {{\overline m}}^2_Y(0) \left[1-\left({\alpha_1(0)\over\alpha_1(t)}\right)^{-13/33}\right], \label{eqn:dterm}\end{aligned}$$ $$\begin{aligned} {{\overline m}}^2_Y(t)&\equiv& -m^2_{H_1}(t) + m^2_{H_2}(t) \nonumber\\ &+&\sum_{gen}\left[m^2_E(t)- m^2_L(t) + m^2_Q(t) + m^2_D(t) -2 m^2_U(t)\right]. \label{eqn:m2yovt}\end{aligned}$$ Functions $\eta_K(t)$, $\hat\xi(t)$, $\hat\eta(t)$ ($\overline\eta\equiv\eta_Q+\eta_U+\eta_{H_2}$) are given in closed forms in terms of integrals over the gauge couplings and are defined in the Appendix of ref. [@CACHOLPOWA]. For $M=2\times10^{16}$ GeV and $\alpha_3(M_Z)=0.12$ they take values: $\hat\xi=2.23$, $\hat\eta=12.8$, $\eta_Q=7.04$, $\eta_U=6.62$, $\eta_{H_1}=\eta_{H_2}=0.513$ The coefficients $c_K$ and $\kappa_K$ read: $c_Q=1$, $c_U=2$, $c_{H_2}=3$, $c_L=c_E=c_D=c_{H_1}=0$; $\kappa_{H_1}=-\kappa_{H_2}=\kappa_L = -3/26$, $\kappa_E = 3/13$, $\kappa_Q = 1/26$, $\kappa_U = -2/13$, $\kappa_D = 1/13$. The evolution of the trilinear couplings $A_k$ are given by: $$\begin{aligned} A_i(t) = A_i(0) - {C_i\over c_t} y A_t(0) + \left({C_i\over c_t}y\hat\xi - \xi_i\right) M_{1/2}, \label{eqn:at0}\end{aligned}$$ Here $C_t=c_t=6$, $C_b=1$ and $C_\tau=0$. Factors $\xi_i(t)$ are defined in the Appendix of ref. [@CACHOLPOWA] and for $M=2\times10^{16}$ GeV and $\alpha_3(M_Z)=0.12$ $\xi_t=3.97$. Quantities at $t=0$ are the initial values of the parameters at the scale $M$, $M_{1/2}\equiv M_3(0)$ is the initial gaugino (gluino) mass (computing numerical values of $\hat\xi$, $\hat\eta$... we have assumed that $M_1/\alpha_1(0)=M_2/\alpha_2(0)=M_3/\alpha_3(0)$). There are several interesting observations about solutions (\[eqn:m2ov0\]-\[eqn:at0\]). Firstly, to a very good approximation squark mass parameters of the first two generations decouple from the running of the masses $m^2_{H_1}$ and $m^2_{H_2}$ (they enter only through small hypecharge $D$-term (\[eqn:dterm\])). As stressed in refs. [@DIGI; @CACHOLPOWA], this is very important for the “naturalness” problem. Moreover, we observe interesting “fixed point” behaviour for the parameter $A_t$ which, in the limit $y\rightarrow1$, becomes independent of its initial values and fixed in terms of the gaugino mass $M_{1/2}$ [@CAOLPOWA1] (unless $A_t(0)\gg M_{1/2}$ but large values of $A_t(0)$ are constrained by the requirement of the absence of the colour breaking minima). Having relations (\[eqn:m2ov0\]-\[eqn:at0\]) we can discuss the impact of the requirement of the proper electroweak symmetry breaking on the low energy parameter space, under various assumptions about the pattern of the soft supersymmetry breaking parameters at large scale. This requirement correlates the low energy soft supersymmetry breaking masses in the Higgs potential with the values of $M_{Z^0}$ and $\tan\beta$, as given in eqs. (\[eqn:vsquared\],\[eqn:sinbeta\]) and, in turn, with other parameters which enter into their RGE. Actually, under each specific assumption about the scale of supersymmetry breaking and the pattern of the soft terms, one can perform a global analysis which includes the electroweak breaking and the existing experimental constraints, and obtain the predictions for the Higgs sector. We present here the results of such an analysis for the following three scenarios: two supergravity scenarios with the GUT relation for the gaugino masses $M_1=M_2=M_3\equiv M_{1/2}$: one with universal soft scalar masses (and universal $A$-terms) and one with universal scalar masses in $SO(10)$ multiplets i.e. with universal sfermion masses (and universal $A$-terms) and two soft Higgs boson masses as independent parameters, and the gauge mediated supersymmetry breaking scenario. Part of the most important experimental constraints has already be discussed in Section 3. In the global analysis discussed now, we also include the following bounds: $m_{C^\pm}>85$ GeV, $\Gamma(Z^0\rightarrow N^0_1N^0_1)<4$ MeV, $BR(Z^0\rightarrow N^0_1N^0_2)<10^{-4}$. There has been often addressed the question of fine-tuning (large cancellations) in the Higgs potential in models with the soft terms generated at large scales [@BAGI; @DIGI; @ANCA]. Indeed, if supersymmetry is to be the solution to the hierarchy problem in the SM, it should not introduce another fine-tuning in the Higgs potential. The origin of the problem is easy to see. Combining (\[eqn:vsquared\]) and (\[eqn:m2ov0\]-\[eqn:at0\]) we can express $M_{Z^0}$ for a given $\tan\beta$ in terms of the initial values $m^2_K(0)$, $M_{1/2}$ and the $\mu$ parameter: $$\begin{aligned} M^2_{Z^0}&=&-2\mu^2(t) + a_{H_1}m^2_{H_1}(0) + a_{H_2}m^2_{H_2}(0) + a_{QU}\left(m^2_Q(0)+m^2_U(0)\right)\nonumber\\ &+& a_{AA}A_t^2(0) + a_{AM}A_t(0)M_{1/2} + a_MM^2_{1/2} \label{eqn:zmass}\end{aligned}$$ For $m_t=175$ GeV the generic values of the coefficients in eq. (\[eqn:zmass\]) in the supergravity scenario e.g. for $\tan\beta\approx1.65(2.2)$ corresponding to $y\approx0.95(0.85)$ are $a_{H_1}\approx1.2(0.5)$, $a_{H_2}\approx1.7(1.5)$, $a_{QU}\approx1.5(1.1)$, $a_{AA}\approx0.1(0.2)$, $a_{AM}\approx-0.3(-0.7)$, $a_M\approx15.0(10.8)$. Eq. (\[eqn:zmass\]) shows that for values of $\mu$, $M_{1/2}$ and/or $m^2_K(0)$ much larger than $M_{Z^0}$ one needs large cancellations. Asymptotically, we are back to the hierarchy problem in the SM. Although the idea of “naturalness” is only qualitative, one can at least correlate the magnitude of the necessary cancellations with the values of the parameters $\mu$, $m^2_K(0)$ and $M_{1/2}$ and, in consequence, with the low energy mass parameters. One notes, in particular, that the smalness of $a_{QU}$ (compared to $a_M$) puts weaker constraints on the “natural” values of $m^2_K(0)$ than large $a_M$ does on $M_{1/2}$. However, in the physical spectrum this effect is partially counterbalanced by the fact that the stop soft masses tend to be suppressed compared to $m_{Q,U}(0)$ by the running with large top quark Yukawa coupling. This effect is stronger for the right handed stop than for the left handed one and gives the hierarchy $M_{\tilde t_R}<M_{\tilde t_L}$. Important source of fine-tuning can also be the relation (\[eqn:sinbeta\]) which correlates the values of $\tan\beta$ and the $B_0$ parameter. “Naturalness” of a given parameter set can be quantified e.g. by calculating the derivatives of $M^2_{Z^0}$ with respect to the soft mass parameters [@BAGI; @DIGI]: $$\begin{aligned} \Delta_i\equiv\mid{a_i\over M_Z^2}{\partial M^2_Z\over\partial a_i}\mid \label{eqn:der}\end{aligned}$$ (other criteria have also been proposed [@ANCA] but will not be discussed here since, at any rate, the concept of naturalness is only qualitative.) In the global analysis, which combines the electroweak breaking with experimental constraints, it is interesting to check the “naturalness” of different parameter regions i.e. to check the values $(\Delta_i)$ for each parameter set. Before presenting the results for the three scenarios considered, it is worthwile to remember several general remarks. First, as already said, the “naturalness” criterion is only a qualitative one and it is unclear how big cancellations are “acceptable” ones. The hierarchy problem in the SM means fine-tuning of many orders of magnitude and from that perspective cancellations of one, two or even three orders of magnitude are still very small. Secondly, we do not know the theory in which soft supersymmetry breaking terms originate and it may well be that such a theory will give them correlated to each other, thus “explaining” the cancellations between them. Finally one can see from eqs. (\[eqn:vsquared\],\[eqn:sinbeta\]) and (\[eqn:zmass\]) that the necessary cancellations tend to increase in the small and large $\tan\beta$ regions advocated as the most interesting ones on the basis of earlier arguments. Indeed, the coefficients $a_i$ in eq. (\[eqn:zmass\]) have $1/(\tan^2\beta-1)$ singularity. Moreover, both for small and very large $\tan\beta$ eq. (\[eqn:sinbeta\]) gives very strong dependence on $B_0$ (i.e. large derivative with respect to $B_0$ in eq. (\[eqn:der\])). With those comments in mind we now present the results of our global analysis, for the three scenarios considered and for several values of $\tan\beta$. In each case the lightest Higgs boson mass is shown as a function of the heavier stop and the maximal and minimal values of $(\Delta_i)^{max}$ for the points on each plot are also noted. Widely discussed has been the so-called minimal supergravity model ([*Ansatz*]{}) with universal scalar and gaugino masses and universal trilinear soft terms. In this model all superpartner masses are given in terms of five parameters: $m_0^2$, $M_{1/2}$, $\mu$, $A_0$ and $B_0$. Two of them can be traded for $M_{Z^0}$ and $\tan\beta$. Thus, we get strongly correlated superpartner spectrum and correlated with the Higgs boson masses. It is now particularly simple to follow our global analysis and to determine the allowed range of the lightest Higgs boson mass as a function of the heavier stop mass. In Fig. \[fig:fintu1\] we show the results for $\tan\beta=1.65$ and 2.5 (corresponding to $y\approx0.95(0.80)$). The parameter space has been scanned up to $M_{\tilde t_2}=1$ TeV with $\mu_0$ and $B_0$ fixed by $M_{Z^0}$ and $\tan\beta$. We see that, in this model, requiring the proper breaking of the electroweak symmetry and with the imposed experimental constraints the lightest Higgs boson mass is bounded from below: $M_{h^0}\simgt75(85)$ GeV for $\tan\beta=1.65(2.5)$ (for $\tan\beta=10$ the lower bound is around 105 GeV). The model also gives lower bound on $M_{A^0}$ of about 500 GeV at $\tan\beta=1.65$ and decreasing to 300 GeV at $\tan\beta=10$. The heavier stop is bounded from below at $\sim450$ GeV. Of course, the crucial role in obtaining those bounds is played by the universality [*Ansatz*]{} combined with the existing experimental constraints. We note also an interesting dependence on the sign of the $\mu$ parameter (two clear branches in Fig. \[fig:fintu1\]a) which is a reflection of the acceptable region in the $(\theta_{\tilde t}, M_{\tilde t_2})$ plane shown in Fig. \[fig:bsgamma\]b. The mass $M_{h^0}$ is bounded from above at 95, 105 and 120 GeV for $\tan\beta=1.65$, 2.5 and 10, respectively. Thus the general bounds shown in Fig \[fig:mh\]b can be reached even in this constrained model. Turning now to the fine-tuning problem we observe first that the model does not admit at all solutions with all $\Delta_i<10$. This is mainly because of the imposed experimental limit $m_{C^\pm}>85$ GeV [@DIGI] which pushes $M_{1/2}$ into the region with $\Delta_{M_{1/2}}\simgt10$ for all $\tan\beta$ values [^8]. Moreover, close to the IR fixed point (for $\tan\beta\approx1.65$), there do not even exist solutions with all $\Delta_i<100$ [@CAOLPOWA1]. Actually, defining two different $\Delta$’s: $\Delta_{max}^\prime\equiv max\{\Delta_{M_{1/2}}, ~\Delta_{m_0}, ~ \Delta_{\mu_0}\}$ and $\Delta_{max}\equiv max\{\Delta_{max}^\prime, ~ \Delta_{B_0}, ~\Delta_{A_0}\}$, the points in Fig. \[fig:fintu1\]a(b) give $70(28)\simlt\Delta_{max}^\prime\simlt970(560)$,   $130(35)\simlt\Delta_{max}\simlt5400(750)$. As expected from the general arguments, cancellations become weaker with increasing $\tan\beta$. In Fig. \[fig:fintu2\]a (b) we show the results for $\tan\beta=2.5(10)$ with the cut $\Delta_{max}<100$. We note that in this case such a cut leaves a non-empty parameter region but gives stronger upper bounds on the Higgs boson mass for the same values of $M_{\tilde t_2}$. They result mainly from the bound on $A_0$ (i.e. on the left-right mixing) obtained due to increasing $\Delta_{M_{1/2}}$ with increasing $A_0$. Moreover, the cut $\Delta_{max}<100$ gives also an upper bound on $M_{\tilde t_2}$. A weaker cut, $\Delta_{max}^\prime<100$, does not change the results for $\tan\beta=10$ (as expected) but allows for broader range of $M_{\tilde t_2}$ for $\tan\beta=2.5$ (with the upper bound for $M_{h^0}$ as for $\Delta_{max}<100$). Finally we note one more interesting effect: a cut on $\Delta$’s gives almost flat (instead of logarithmic) dependence of $M_{h^0}$ on $M_{\tilde t_2}$. An increase in $M_{\tilde t_2}$ is balanced by a decrease in $A_0$ (i.e. in the left-right mixing) to keep $\Delta$’s below the imposed bound. In the next step one can study a less restrictive model, with the pattern of the soft terms consistent with the $SO(10)$ unification, i.e. with the universal sfermion masses and the two Higgs boson masses taken as independent parameters. It turns out that the predictions the $M_{h^0}$, $M_{A^0}$ and $M_{\tilde t_2}$ (as well as the degree of fine-tuning) are very similar to those in the universal case and need not be independently shown here. This similarity can be partly understood in terms of the important role played by the limit $m_{C^\pm}>85$ GeV and by the constraints from $b\rightarrow s\gamma$ and from precision data, which are not sensitive to the assumed non-universality in the Higgs boson mass parameters. Moreover, there is no real change in the values of $\Delta_i$’s since their expected decrease with the increasing number of free parameters is now reduced by the dissapearance of certain cancellations in $\Delta_i$’s which are present in the universal case. Finally, the considered values of $\tan\beta$ are enough above 1 to make the electroweak symmetry breaking easy even in the universal case and the considered breakdown of full universality does not significantly enlarge the parameter space consistent with the electroweak symmetry breaking. Thus, the results presented in Figs. \[fig:fintu1\] and \[fig:fintu2\] are representative also for the considered partial breakdown of universality. The difference between universal and non-universal boundary conditions becomes more visible in the large $\tan\beta$ region [@OLPO2; @HARASA]. In the latter case the parameters $\Delta$ are typically $\sim\tan\beta$ whereas in the former - ${\cal O}(1000)$, i.e. fine-tuning of order $10^{-3}$. From the point of view of the predictions for the MSSM spectrum, the large $\tan\beta$ region is characterized by the expectation of a relatively light $CP$-odd Higgs boson, $M_{A^0}\simlt {\cal O}(200$ GeV). For further details in the large $\tan\beta$ region, in particular for correlations with radiative corrections to the $b$-quark mass and with $b\rightarrow s\gamma$ decay, we refer the reader to [@OLPO2; @HARASA; @CAOLPOWA]. Finally, it is interesting to compare the supergravity scenario with models in which supersymmetry breaking is transmitted to the observable sector through ordinary $SU(3)\times SU(2)\times U(1)$ gauge interactions of the so-called messenger fields at scales $M\ll M_{GUT}$ [@DIFISR]. In general, these gauge-mediated models of SUSY breaking are characterized by two scales: the scale $M$, which is of the order of the average messenger mass and the scale $\sqrt{F}$ ($\sqrt{F}<M$) of supersymmetry breaking. Messenger fields are assumed to form complete [**5**]{}+$\overline{\rm\bf 5}$ (or [**10**]{}+$\overline{{\rm\bf 10}}$) $SU(5)$ representations. Their number $n$ is restricted to $n_{max}=4$ by the requirement of perturbativity of the gauge couplings up to the GUT scale. In those models the LSP is a very light gravitino ($m_{\tilde G}<1$ keV). For $\sqrt{F}<10^6$ GeV the decay length of the lightest neutralino into a photon and gravitino is such that this decay occurs within a typical detector. Hence, photons + missing energy become a signature of supersymmetry at LEP and Tevatron colliders. The absence of such events in the existing data strongly disfavours charginos and stops with masses below $\sim125$ GeV and $\sim140$ GeV respectively [@AMKAKRMAMR]. In terms of $M$ and $x\equiv F/M^2$ the soft supersymmetry breaking parameters of the MSSM at the scale $\sim M$ are given by: $$\begin{aligned} M_i = {\alpha_i(M)\over4\pi}M ~n ~x ~g(x) \equiv {\alpha_i(M)\over4\pi}M ~y\end{aligned}$$ $$\begin{aligned} m^2_{\tilde f} = 2M^2~n ~x^2 ~f(x) \sum_i \left({\alpha_i(M)\over4\pi}\right) C_i =2M^2 ~y^2 ~z \sum_i \left({\alpha_i(M)\over4\pi}\right) C_i \label{eqn:scalars}\end{aligned}$$ where $C_3=4/3$, $C_2=3/4$, $C_1=(3/5)Y^2$ ($Y$ being the hypercharge of the scalar ${\tilde f}$), the functions $g(x)$ and $f(x)$ ($g(0)=f(0)=1$, $g(1)\approx 1.4$, $f(1)\approx0.70$) can be found in ref. [@DIGIPO] and the factor $z\equiv f(x)/n g^2(x)$. Thus, for fixed messenger sector (i.e. fixed $n$) and fixed scale $M$ all soft supersymmetry breaking masses are predicted in terms of $y$ ($0<y<n_{max}~g(1)\approx5.6$) [^9]. In those models we also have $A_0\approx0$ as the $A_0$ parameter can be generated at two loop only [@DITHWE]. However, the values of the soft masses $m^2_{H_{1,2}}$ may differ significantly from their values given by eq. (\[eqn:scalars\]) since they can be modified by physics involved in generation of $B_0$ and $\mu_0$ parameters [@DVGIPO]. Therefore, in our scans we take $y$, $m_{H_1}$, $m_{H_2}$, $\mu_0$ and $B_0$ as free parameters (the last two are fixed by $M^2_{Z^0}$ and $\tan\beta$). To be general, the factor $z$ in eq. (\[eqn:scalars\]) is scanned between $z_{min}=f(1)/n_{max}g^2(1)\approx0.15$ and $z_{max}=1$. For definitness we will consider $M=10^5$ GeV only. Here we follow the same simple approach we used for the supergravity models. On the parameter space consistent with the electroweak symmetry breaking we impose the discussed earlier experimental constraints (now we require $m_{C^\pm}>120$ GeV, $M_{\tilde t_1}>140$ GeV). Very important rôle is played by $b\rightarrow s\gamma$. The requirement of good $b\rightarrow s\gamma$ rate reduces otherwise rather widely spread out $h^0$ and $A^0$ Higgs boson masses (for $\tan\beta=2.5$: $20<M_{h^0}<100$ GeV) to a narrow band ($80<M_{h^0}<100$ GeV, $M_{A^0}>200$ GeV). This effect can be easily understood (see Fig. \[fig:bsgamma\]) because in the model considered squarks and charginos are rather heavy [^10] so a light $A^0$ is not allowed by $b\rightarrow s\gamma$ and light $h^0$ is always associated with light $A^0$. Moreover, surviving small values of $M_{h^0}$ ($\sim80$ GeV for $\tan\beta=2.5$ are associated with lowest values of $M_{\tilde t_2}$ ($\simlt 500$ GeV) which are eliminated by imposing the $\Delta\chi^2<4$ cut. Finally, if we also require “naturalness” e.g. by demanding [^11] $\Delta_{max}<100$ ($\Delta_{max}=max\{\Delta_x, ~\Delta_{m_{H1}}, ~\Delta_{m_{H2}}, ~ \Delta_{\mu_0}, ~\Delta_{B_0}\}$), we constrain the heavier stop mass $M_{\tilde t_2}$ and $CP$-odd higgs boson mass $M_{A^0}$ from above to $\simlt700$ GeV. Final results are shown in Fig. \[fig:fintu3\] as a plot of $M_{h^0}$ versus the mass of the heavier stop $M_{\tilde t_2}$ predicted in models of gauge mediated supersymmetry breaking with $M=10^5$ GeV for $\tan\beta=2.5$ and 10. As in the case of supergravity models, the restriction of the chargino and stop masses eliminates solutions with $\Delta_{max}<10$. With all constraints imposed, $M_{h^0}$ turns out to be surprisingly tightly constrained. For $\tan\beta=2.5(10)$ values of the lightest scalar Higgs boson are bounded by 90(108) GeV $\simlt M_{h^0}\simlt97(115)$ GeV. Masses of the $CP$-odd Higgs boson are bounded to 280(200) GeV $\simlt M_{A^0}\simlt$700(850) GeV. These upper bounds should be compared to the ones obtained in [@RITOMO] in the restricted model of gauge mediated supersymmetry breaking (with $x=1$, $n=1$ and with $m_{H_{1,2}}$ as given by eq. (\[eqn:scalars\])) without imposing any additional constraints. It is interesting that in the much more general scenario described above, after imposing experimental and naturalness cuts, one gets upper bounds on $M_{h^0}$ not higher than those obtained in [@RITOMO]. Non-minimal SUSY models ======================= Non-minimal supersymmetric extensions of the Standard Model can go beyond the MSSM in several directions. Models with additional Higgs singlets [@FA; @DR], doublets and/or triplets have been considered in the literature [@ESQU1]. The gauge group of the MSSM can also be extended by e.g. additional $U(1)$ factors. Finally, models with $R-$parity spontaneously broken (Supersymmetric Singlet Majoron Model) have been proposed [@GIMAPIRI]. The important question is what can be said about the Higgs sector in such extensions. In the most popular extension, the so called Next-to-Minimal Supersymmetric Standard Model (NMSSM) [@FA; @DR], one introduces a gauge singlet superfield $N$ and replaces the term $\mu H_1 H_2$ in the MSSM superpotential by $$\begin{aligned} \lambda H_1 H_2 N + {\kappa\over3} N^3\end{aligned}$$ with the obvious motivation to avoid the $\mu$ problem [@HALY]. The model has been analysed in several papers. At the tree level, the upper bound for the mass of the lightest Higgs boson $$\begin{aligned} M^2_{h^0}\leq M_Z^2\left(\cos^22\beta + {2\lambda^2\over g^2} c^2_W \sin^22\beta\right)\end{aligned}$$ has been derived [@DR; @ESQU1]. It is important that this bound depends neither on the sfermion masses (which can be much larger than $M_Z$) nor on the vacuum expectation value of the singlet field $N$ (which is not constrained by $M_Z$). Therefore, the bound is controlled only by the dimensionless Yukawa coupling $\lambda$ which is constrained by the requirement of the perturbativity of the model, $\lambda\simlt4\pi$. This result, that the tree level bound does not depend on the parameters which can be of order of the supersymmetry breaking scale (and, hence, is always $\simlt {\cal O}(G_F^{-1/2})$), has been extended to models containing arbitrary numbers of Higgs singlets, doublets and triplets [@ESQU1; @ESQU2]. Finally it was also shown to be valid in the Supersymmetric Single Majoron Model [@ES]. Almost model-independent proofs of this remarkable fact has been given in [@KAKOWE; @ES]. Radiative corrections to the tree level bound [@ELRA] are under control as in the MSSM and has been shown to depend on soft SUSY breaking masses only logarithmically [@ELKIWH1; @ELKIWH2; @ES]. The tree level bounds can be further strenghtened by requiring that the additional Yukawa couplings remain perturbative up to a scale $\Lambda$ (GUT scale). For the NMSSM this has been done in the papers [@ESQU1; @BISA; @ELKIWH2]. Analysing coupled RGE for $\lambda$, $Y_t$ and $\alpha_s$ the upper limit on $\lambda$ has been found as a function of $Y_t$ for $\Lambda=10^{16}$ GeV. For large values of $m_t$, for which $Y_t$ is always close to its perturbativity limit, the coupling $\lambda$ is forced to be small and hence its effects become less important. Including all relevant one-loop radiative corrections it was found that, for $m_t$ in the present experimental range, the upper bound for the lightest Higgs boson in the NMSSM is only $5-15$ GeV larger than the corresponding upper bound in the MSSM and, for $M_{\tilde t_i}\leq1$ TeV it is smaller than 150 GeV. Effects of the additional assumptions like gauge coupling unification, universality of the soft SUSY breaking masses and requirement of the correct electroweak symmetry breaking have also been studied in this model [@KIWH]. Summary ======= In unconstrained minimal supersymmetric extension of the Standard Model there exists the well known upper bound on the mass of the lightest supersymmetric Higgs. The available parameter space of the model is now considerably reduced by the existing experimental data and can be further reduced by additional theoretical assumptions, mostly related to the extrapolation of the model to very large energy scales. Especially fruitful is the assumption about perturbative validity of the model up to the GUT scale and the requirement of the proper electroweak breaking combined with a simple [*Ansatz*]{} (such as universality or partial universality) for the pattern of the soft supersymmetry breaking terms. Such a reduction in the parameter space results in more definite expectations for $M_{h^0}$ than the general bounds. Several arguments point toward $M_{h^0}<$100 GeV. Both, the discovery or the absence of the Higgs boson in this mass range will have strong implications for the supersymmetric extension of the Standard Model. Acknowledgments {#acknowledgments .unnumbered} =============== Work supported by the Polish Commitee for Scientific Research under the grant 2 P03B 040 12 and by the European Union under contract CIPD-CT94-0016. P.H.Ch. would like to thank Dr. M. Misiak for cross checking his code for $b\rightarrow s\gamma$. References {#references .unnumbered} ========== [99]{} M. Quiros this volume. The figure we refer to can be also found (fig.1) in [*Higgs Physics at LEP 2*]{} M. Carena, P. Zerwas (convenors), CERN Yellow Report, G. Altarelli, F. Zwirner eds. (hep-ph/9602250). H.E. Haber this volume. M. Dine, W. Fischler, M. Srednicki [*Nucl. Phys.*]{} [**B189**]{} (1981) 575; S. Dimopoulos, S. Raby [*Nucl. Phys.*]{} [**B192**]{} (1981) 353; L. Alvarez-Gaumé, M. Claudson, M. Wise [*Nucl. Phys.*]{} [**B207**]{} (1982) 96; M. Dine, A. Nelson [*Phys. Rev.*]{} [**D48**]{} (1993) 1277; M. Dine, A. Nelson, Y. Shirman [*Phys. Rev.*]{} [**D51**]{} (1995) 1362; M. Dine, A. Nelson, Y. Nir, Y. Shirman [*Phys. Rev.*]{} [**D53**]{} (1996) 2658. S.P. Li, M. Sher [*Phys. Lett.*]{} [**B140**]{} (1984) 339; H.-P. Nilles, M. Nusbaumer [*Phys. Lett.*]{} [**B145**]{} (1984) 73; J. Gunion, H.E. Haber [*Nucl. Phys.*]{} [**B272**]{} (1985) 1. Y. Okada, M. Yamaguchi, T. Yanagida [*Prog. Theor. Phys.*]{} [**85**]{} (1991) 1; H.E. Haber, R. Hempfling [*Phys. Rev. Lett.*]{} [**66**]{} (1991) 1815; J. Ellis, G. Ridolfi, F. Zwirner [*Phys. Lett.*]{} [**B257**]{} (1991) 83, [**262**]{} (1991) 477. R. Hempfling Ph.D. Thesis, UCSC preprint SCIPP 92/28 (1992); J.L. Lopez, D.V. Nanopoulos [*Phys. Lett.*]{} [**266B**]{} (1991) 397. M. Carena, J.-R. Espinosa, M. Quiros, C.E.M. Wagner [*Phys. Lett.*]{} [**B355**]{} (1995) 209; M. Carena, M. Quiros, C.E.M. Wagner [*Nucl. Phys.*]{} [**B461**]{} (1996) 407. R. Barbieri, G.-F. Giudice [*Phys. Lett.*]{} [**309**]{} (1993) 86; A.J. Buras, M. Misiak, M. Münz, S. Pokorski [*Nucl. Phys.*]{} [**B424**]{} (1994) 374; J. Rosiek [*Phys. Lett.*]{} [**252B**]{} (1990) 135; A. Denner et al. [*Z. Phys.*]{} [**C51**]{} (1991) 695; M. Boulware, D. Finnell [*Phys. Rev.*]{} [**D44**]{} (1991) 2054. P. Krawczyk, S. Pokorski [*Phys. Rev. Lett.*]{} [**60**]{} (1988) 182; G. Isidori [*Phys. Lett.*]{} [**298B**]{} (1993) 409; Y. Grossman, H.E. Haber, Y. Nir [*Phys. Lett.*]{} [**357B**]{} (1995) 630. The Lep Electroweak Working Group, CERN preprint LEPEWWG/96-02. K.G. Chetyrkin, M. Misiak, M. Münz preprint ZU-TH 24/96 (hep-ph/9612313). P.H. Chankowski S. Pokorski [*Nucl. Phys.*]{} [**B475**]{} (1996) 3. The ALEPH collaboration, contribution to Int. Europhysics. Conf. on High Energy Physics, Brussels, Belgium July-August 1995, paper EPS0415. B. Pendleton, G.G. Ross [*Phys. Lett.*]{} [**B98**]{} (1981) 291; see also W. Zimmermann [*Comm. Math. Phys.*]{} [**97**]{} (1985) 211; J. Kubo, K. Sibold, W. Zimmermann [*Phys. Lett.*]{} [**B200**]{} (1989) 191. L. Ibañez, C. Lopez [*Nucl. Phys.*]{} [**B233**]{} (1984) 511; L. Ibañez, C. Lopez, C. Muñoz [*Nucl. Phys.*]{} [**B256**]{} (1985) 218; A. Bouquet, J. Kaplan, C.A. Savoy [*Nucl. Phys.*]{} [**B262**]{} (1985) 299. C.T. Hill [*Phys. Rev.*]{} [**D24**]{} (1981) 691; C.T. Hill C.N. Leung, S. Rao [*Nucl. Phys.*]{} [**B262**]{} (1985) 517. L. Alvarez-Gaumé, J. Polchinski, M.B. Wise [*Nucl. Phys.*]{} [**B221**]{} (1983) 495; J. Bagger, S. Dimopoulos, E. Masso [*Phys. Rev. Lett.*]{} [**55**]{} (1985) 920; M. Olechowski, S. Pokorski [*Phys. Lett.*]{} [**B257**]{} (1991) 388; C.D. Frogatt, I.D. Knowles, R.G. Moorhouse [*Phys. Lett.*]{} [**B249**]{} (1990) 273; M. Carena et al. [*Nucl. Phys.*]{} [**B369**]{} (1992) 33. M. Carena C.E.M. Wagner [*Nucl. Phys.*]{} [**B452**]{} (1995) 45. L.J. Hall, R. Ratazzi, U. Sarid [*Phys. Rev.*]{} [**D50**]{} (1994) 7048; R. Hempfling [*Phys. Rev.*]{} [**D49**]{} (1994) 6168. M. Carena, M. Olechowski, S. Pokorski, C.E.M. Wagner [*Nucl. Phys.*]{} [**B426**]{} (1994) 269; T. Blazek, S. Pokorski, S. Raby [*Phys. Rev.*]{} [**D52**]{} (1995) 4151. H. Arason et al. [*Phys. Rev. Lett.*]{} [**67**]{} (1991) 2933; S. Kelley, J.L. Lopez, D.V. Nanopoulos [*Phys. Lett.*]{} [**B278**]{} (1992) 140; V. Barger, M.S. Berger, P. Ohmann [*Phys. Rev.*]{} [**D47**]{} (1993) 1093; V. Barger, M.S. Berger, P. Ohmann, R.J.N. Phillips [*Phys. Lett.*]{} [**B314**]{} (1993) 35; P. Langacker, N. Polonsky [*Phys. Rev.*]{} [**D47**]{} (1993) 4028, [**D49**]{} (1994) 1454; M. Carena, S. Pokorski, C.E.M. Wagner [*Nucl. Phys.*]{} [**B406**]{} (1993) 59; W. Bardeen, M. Carena, S. Pokorski, C.E.M. Wagner [*Phys. Lett.*]{} [**B320**]{} (1994) 110; M. Carena, M. Olechowski, S. Pokorski, C.E.M. Wagner [*Nucl. Phys.*]{} [**B419**]{} (1994) 213; C. Kounnas, F. Zwirner, I. Pavel [*Phys. Lett.*]{} [**B335**]{} (1994) 403; P. Binetruy, E. Dudas [*Phys. Lett.*]{} [**B338**]{} (1994) 23. T. Banks [*Nucl. Phys.*]{} [**B303**]{} (1988) 172; M. Olechowski, S. Pokorski [*Phys. Lett.*]{} [**B214**]{} (1988) 393; G.-F. Giudice, G. Ridolfi [*Z. Phys.*]{} [**C41**]{} (1988) 447; B. Ananhtarayan, G. Lazarides, Q. Shafi [*Phys. Rev.*]{} [**D44**]{} (1991) 1613; S. Dimopoulos, L.J. Hall, S. Raby [*Phys. Rev. Lett.*]{} [**68**]{} (1992) 1984, [*Phys. Rev.*]{} [**D45**]{} (1992) 4192. S. Pokorski in Proc. XII Int. Workshop on Weak Interactions and Neutrinos, Ginosar, Israel [*Nucl. Phys. Proc. Suppl.*]{} [**B13**]{} (1990) 606; H.-P. Nilles in Proc. of the 1990 Theoretical Advanced Study Institute in Elementary Particle Physics, eds. M. Cvetic, P. Langacker, World Scientific, Singapore, p. 633; W. Majerotto, B. Mösslacher [*Z. Phys.*]{} [**C48**]{} (1990) 273; M. Drees, M.M. Nojiri [*Nucl. Phys.*]{} [**B369**]{} (1992) 54; B. Ananhtarayan, G. Lazarides, Q. Shafi [*Phys. Lett.*]{} [**B300**]{} (1993) 245. G.W. Anderson, S. Raby, S. Dimopoulos, L.J. Hall [*Phys. Rev.*]{} [**D47**]{} (1993) 3702; G.W. Anderson et al. [*Phys. Rev.*]{} [**D49**]{} (1994) 3660; L.J. Hall, S. Raby [*Phys. Rev.*]{} [**D51**]{} (1995) 6524. M. Carena et al. preprint CERN-TH/96-241 (hep-ph/9612261) [*Nucl. Phys.*]{} [**B**]{} in press. S. Dimopoulos, G.-F. Giudice [*Phys. Lett.*]{} [**B357**]{} (1995) 573. R. Barbieri, G.-F. Giudice [*Nucl. Phys.*]{} [**B306**]{} (1988) 63. G.W. Anderson, D.J. Castaño [*Phys. Lett.*]{} [**B347**]{} (1995) 300, [*Phys. Rev.*]{} [**D52**]{} (1995) 1693; A. Strumia preprint FT-UAM-96-40 (hep-ph/9609286); G.W. Anderson, D.J. Castaño, A. Riotto [*Phys. Rev.*]{} [**D55**]{} (1997) 2950. M. Olechowski, S. Pokorski [*Nucl. Phys.*]{} [**B404**]{} (1993) 590. B. de Carlos, J.A. Casas [*Phys. Lett.*]{} [**B309**]{} (1993) 320. M. Olechowski, S. Pokorski [*Phys. Lett.*]{} [**B344**]{} (1995) 201; F. Borzumati, M. Olechowski, S. Pokorski [*Phys. Lett.*]{} [**B349**]{} (1995) 311. S. Ambrosanio et al. [*Phys. Rev.*]{} [**D54**]{} (1996) 5395. S. Dimopoulos, G.-F. Giudice, A. Pomarol [*Phys. Lett.*]{} [**B389**]{} (1996) 37; S. Martin [*Phys. Rev.*]{} [**D55**]{} (1997) 3177. S. Dimopoulos, S. Thomas, J. Wells preprint SLAC-PUB-7237 (hep-ph/9609434). G. Dvali, G.-F. Giudice, A. Pomarol [*Nucl. Phys.*]{} [**B478**]{} (1996) 31. P. Ciafaloni, A. Strumia preprint FT-UAM 96/43 (hep-ph/9611204); G. Bhattacharyya, A. Romanino preprint IFUP-TH 67/96 (hep-ph/9611243). A. Riotto, O. Törnqvist, R.N. Mohapatra [*Phys. Lett.*]{} [**B388**]{} (1996) 599. P. Fayet [*Nucl. Phys.*]{} [**B90**]{} (1975) 104; H.-P. Nilles, M. Srednicki, D. Wyler [*Phys. Lett.*]{} [**B120**]{} (1983) 346; J.-P. Derendinger, C.A. Savoy [*Nucl. Phys.*]{} [**B237**]{} (1984) 307. M. Drees [*Int. J. Mod. Phys.*]{} [**A4**]{} (1989) 3635; J. Ellis, J. Gunion, H. Haber, L. Roszkowski, F. Zwirner [*Phys. Rev.*]{} [**D39**]{} (1989) 844. J.-R. Espinosa, M. Quiros [*Phys. Lett.*]{} [**B279**]{} (1992) 92. G.-F. Giudice, A. Masiero, M. Pietroni, A. Riotto [*Nucl. Phys.*]{} [**B396**]{} (1993) 243. L. Hall, J. Lykken, S. Weinberg [*Phys. Rev.*]{} [**D27**]{} (1983) 2359; J.E. Kim, H.-P. Nilles [*Phys. Lett.*]{} [**B138**]{} (1984) 150; K. Inoue, A. Kakuto, T. Takano [*Prog. Theor. Phys.*]{} [**75**]{} (1986) 664; A.A. Anselm, A.A. Johansen [*Phys. Lett.*]{} [**B200**]{} (1988) 331; G.-F. Giudice, A. Masiero [*Phys. Lett.*]{} [**B206**]{} (1988) 480. J.-R. Espinosa, M. Quiros [*Phys. Lett.*]{} [**B302**]{} (1993) 51. J.-R. Espinosa [*Phys. Lett.*]{} [**B353**]{} (1995) 243. G.L. Kane, C. Kolda, J. Wells [*Phys. Rev. Lett.*]{} [**70**]{} (1993) 2686. U. Ellwanger, M. Rausch de Traubenberg [*Z. Phys.*]{} [**C53**]{} (1992) 521; U. Ellwanger, M. Lindner [*Phys. Lett.*]{} [**B301**]{} (1993) 365; U. Ellwanger [**B303**]{} (1993) 271; W.T.A. ter Veldhuis preprint PURD-TH-92-11 (hep-ph/9211281). T. Elliott, S.F. King, P.L. White [*Phys. Lett.*]{} [**B305**]{} (1993) 71, [**B314**]{} (1993) 56, [**B351**]{} (1995) 213. T. Elliott, S.F. King, P.L. White [*Phys. Rev.*]{} [**D49**]{} (1994) 2435. P. Binetruy, C.A. Savoy [*Phys. Lett.*]{} [**B277**]{} (1992) 453. U. Ellwanger, M. Rausch de Traubenberg, C.A. Savoy [*Phys. Lett.*]{} [**B315**]{} (1993) 331, [*Z. Phys.*]{} [**C67**]{} (1995) 665; S.F. King, P.L. White [*Phys. Rev.*]{} [**D52**]{} (1995) 4183. [^1]: On leave of absence from the Institute of Theoretical Physics, Warsaw University, Hoża 69, 00-681 Warsaw, Poland. [^2]: To appear in “Perspectives on Higgs Physics II” ed. G.L. Kane, World Scientific, Singapore. [^3]: We require the branching ratio computed in the NLO approximation to fall into the interval $1\times10^{-4}<BR(b\rightarrow s\gamma)<4\times10^{-4}$. All uncertainties of the computation (for detailed discussion of the uncertainties see ref. [@CHMIMU]) are taken into account. [^4]: Those results are obtained under the assumption that the chargino -stop - bottom coupling is given by the Kobayashi-Maskawa angles. In principle, off-diagonal terms in the right-handed stop mass matrix are possible which would change those couplings. Such (1,3) and (2,3) terms are not constrained by any other data. However, other flavour off diagonal terms are strongly constrained so the presence of large (1,3) and (2,3) terms for right-handed squarks would mean strong flavour dependence in the squark mass matrices. [^5]: This bound can be slightly lowered to $\tan\beta\simgt1.1$, in gauge mediated SUSY breaking models, due to the presence of additional coloured matter fields at the intermediate scale $M\sim10^5-10^7$ GeV. [^6]: In this case the bottom pole mass may be significantly different from the running mass due to the supersymmetric loop corrections [@HARASA; @CAOLPOWA]. Similar corrections may be even important for the Kobayashi-Maskawa mixing angles [@BLPORA] [^7]: For large $\tan\beta$ this conclusion may be less strong. [^8]: Strictly speaking, this conclusion is valid as long as we work with the tree level Higgs potential renormalized at $M_{Z^0}$. It is well known [@OLPO1; @CACA; @ANCA; @DIGI] that inclusion of the full 1-loop corrections to the scalar potential diminishes somewhat the degree of fine-tuning and the bound $m_{C^\pm}>85$ GeV becomes marginally consistent with $\Delta_i<10$, for intermediate values of $\tan\beta$. We neglect this effect here, as it does not change our qualitative conclusions. [^9]: Nonzero lower bound on $y$ is set by Tevatron limit on gluino mass, $m_{\tilde g}\simgt150$ GeV, which is, however, subject to some restrictions [^10]: In addition, because $\mu$ values required by electroweak symmetry breaking are large, the lighter chargino turns out to have only small higgsino component and hence its $b \tilde t C^-$ coupling is weaker than that of the pure higgsino chargino which is responsible for the limit shown in Fig. \[fig:bsgamma\]. [^11]: Naturalness of the gauge mediated models has been analyzed by different methods in refs. [@CIST].

--- abstract: 'Rather than learning new control policies for each new task, it is possible, when tasks share some structure, to compose a “meta-policy” from previously learned policies. This paper reports results from experiments using Deep Reinforcement Learning on a continuous-state, discrete-action autonomous driving simulator. We explore how Deep Neural Networks can represent meta-policies that switch among a set of previously learned policies, specifically in settings where the dynamics of a new scenario are composed of a mixture of previously learned dynamics and where the state observation is possibly corrupted by sensing noise. We also report the results of experiments varying dynamics mixes, distractor policies, magnitudes/distributions of sensing noise, and obstacles. In a fully observed experiment, the meta-policy learning algorithm achieves 2.6x the reward achieved by the next best policy composition technique with 80% less exploration. In a partially observed experiment, the meta-policy learning algorithm converges after 50 iterations while a direct application of RL fails to converge even after 200 iterations.' author: - 'Richard Liaw$^{2}$, Sanjay Krishnan$^{2}$, Animesh Garg$^{2,3}$, Daniel Crankshaw$^{2}$, Joseph E. Gonzalez$^{2}$, Ken Goldberg$^{1,2}$ [^1]' title: | Composing Meta-Policies for Autonomous Driving\ Using Hierarchical Deep Reinforcement Learning --- =1 Acknowledgment {#acknowledgment .unnumbered} ============== Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Sponsors. This research was performed in part in the AUTOLAB at UC Berkeley in affiliation with the AMP Lab, BDD, BAIR, the CITRIS “People and Robots” (CPAR) Initiative: http://robotics.citris-uc.org Multilateral Manipulation by Human-Robot Collaborative Systems and by Google, Cisco, Siemens, Cloudminds. [^1]: $^{1}\,$IEOR, $^{2}\,$EECS, UC Berkeley, CA USA; $^{3}\,$CS, Stanford University, Stanford CA US; The AUTOLAB at UC Berkeley (automation.berkeley.edu); `{rliaw, sanjaykrishnan, animesh.garg, jegonzal, goldberg}@berkeley.edu, crankshaw@eecs.berkeley.edu`

--- author: - | [Noah Johnson]{}\ Oasis Labs\ University of California, Berkeley - | [Joseph P. Near]{}\ University of Vermont[^1] - | [Joseph M. Hellerstein]{}\ University of California, Berkeley - | [Dawn Song]{}\ Oasis Labs\ University of California, Berkeley bibliography: - 'refs.bib' - 'allegro-refs.bib' title: 'Chorus: Differential Privacy via Query Rewriting' --- Abstract {#abstract .unnumbered} ======== We present [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}, a system with a novel architecture for providing differential privacy for statistical SQL queries. The key to our approach is to embed a differential privacy mechanism into the query before execution so the query automatically enforces differential privacy on its output. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}is compatible with any SQL database that supports standard math functions, requires no user modifications to the database or queries, and simultaneously supports many differential privacy mechanisms. To the best of our knowledge, no existing system provides these capabilities. We demonstrate our approach using four general-purpose differential privacy mechanisms. In the first evaluation of its kind, we use [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}to evaluate these four mechanisms on real-world queries and data. The results demonstrate that our approach supports 93.9% of statistical queries in our corpus, integrates with a production DBMS *without any modifications*, and scales to hundreds of millions of records. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}is currently being deployed at [a large technology company]{} for its internal analytics tasks. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}represents a significant part of the company’s GDPR compliance efforts, and can provide both differential privacy and access control enforcement. In this capacity, [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}processes more than 10,000 queries per day. Introduction {#sec:intro} ============ Organizations are collecting more and more sensitive information about individuals. As this data is highly valuable for a broad range of business interests, organizations are motivated to provide analysts with flexible access to the data. At the same time, the public is increasingly concerned about privacy protection. There is a growing and urgent need for technology solutions that balance these interests by supporting general-purpose analytics while guaranteeing privacy protection. Differential privacy [@dworkdifferential2006; @dwork2014algorithmic] is widely recognized by experts as the most rigorous theoretical solution to this problem. Differential privacy provides a formal guarantee of privacy for individuals while allowing general statistical analysis of the data. In short, it states that the presence or absence of any single individual’s data should not have a large effect on the results of a query. This allows precise answers to questions about populations in the data while guaranteeing the results reveal little about any individual. Unlike alternative approaches such as anonymization and k-anonymity, differential privacy protects against a wide range of attacks, including attacks using auxiliary information [@sweeney1997weaving; @DBLP:journals/corr/abs-cs-0610105; @taxis; @de2013unique]. Current research on differential privacy focuses on development of new algorithms, called *mechanisms*, to achieve differential privacy for a particular class of queries. Researchers have developed dozens of mechanisms covering a broad range of use cases, from general-purpose statistical queries [@dwork2006calibrating; @nissim2007smooth; @mcsherry2007mechanism; @mcsherry2009privacy; @proserpio2014calibrating; @mohan2012gupt; @Blocki:2013:DPD:2422436.2422449] to special-purpose analytics tasks such as graph analysis [@hay2009accurate; @sala2011sharing; @karwa2011private; @kasiviswanathan2013analyzing; @Chen:2013:RMT:2463676.2465304], range queries [@hardt2012simple; @li2014data; @li2010optimizing; @li2015matrix; @zhang2014towards; @xiao2012dpcube; @cormode2012differentially; @qardaji2013differentially; @acs2012differentially; @xu2013differentially], and analysis of data streams [@dwork2010differential; @shi2011privacy]. Each mechanism works well for specific tasks and not as well, or not at all, on other tasks. Despite extensive academic research and an abundant supply of mechanisms, differential privacy has not been widely adopted in practice. Existing applications of differential privacy in practice are limited to specialized use cases such as web browsing statistics [@erlingsson2014rappor] and keyboard and emoji use [@apple]. There are two major challenges for practical adoption of differential privacy. The first is seamless integration into real-world data environments. These environments include highly customized data architectures and industrial-grade database engines carefully tuned for performance and reliability. Previous differential privacy systems require changes to the data pipeline [@allegro] or replacement of the database with a custom engine [@mohan2012gupt; @proserpio2014calibrating; @mcsherry2009privacy] and hence do not integrate easily into these environments. The second challenge is simultaneously supporting different mechanisms. Current evidence suggests that there is no single “best mechanism” that performs optimally for all queries. Rather, the best mechanism depends on both the query and the dataset. In fact, as demonstrated by Hay et al. [@DBLP:conf/sigmod/HayMMCZ16], the best mechanism can also vary with the size of the dataset even for a single query. A practical solution must provide flexibility for mechanism selection and easy integration of new mechanisms. Previous differential privacy systems [@allegro; @mohan2012gupt; @proserpio2014calibrating] implement only a single mechanism. #### The [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}system. This paper describes [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}, a system with a novel architecture for providing differential privacy for statistical SQL queries. The key insight of our approach is to embed the differential privacy mechanism into the SQL query before execution, so the query automatically enforces differential privacy on its own output. We define a SQL query with this property as an *intrinsically private query*. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}automatically converts untrusted input queries into intrinsically private queries. This approach enables a new architecture in which queries are rewritten, then submitted to an unmodified database management system (DBMS). This new architecture addresses the two major challenges outlined above. First, [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}is compatible with any SQL database that supports standard math functions ($\mathit{rand}$, $\mathit{ln}$, etc.) thus avoiding the need for a custom runtime or modifications to the database engine. By using a standard SQL database engine instead of a custom runtime, [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}can leverage the reliability, scalability and performance of modern databases, which are built on decades of research and engineering experience. Second, [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}enables the modular implementation and adoption of many different mechanisms, supporting a significantly higher percentage of queries than any single mechanism and providing increased flexibility for both general and specialized use cases. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}automatically selects a mechanism for each input query based on an extensible set of selection rules. To the best of our knowledge, no existing system provides these capabilities. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}also protects against an untrusted analyst: even if the submitted query is malicious, our transformation rules ensure that the executed query returns only differentially private results. The results can thus be returned directly to the analyst without post-processing. This enables easy integration into existing data environments via a single pre-processing step. We demonstrate the [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}approach with four existing general-purpose differential privacy mechanisms: Elastic Sensitivity [@allegro], Restricted Sensitivity [@Blocki:2013:DPD:2422436.2422449], Weighted PINQ [@dwork2006calibrating] and Sample & Aggregate [@mohan2012gupt]. These mechanisms support a range of SQL features and analytic tasks. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}contains query transformation rules for each mechanism which convert untrusted (non–intrinsically private) queries into intrinsically private queries. We also describe how additional mechanisms can be added to [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}. In an extended version of this paper [@chorusextended], we formalize the rewrite rules for transforming input queries into intrinsically private queries and prove them correct. #### Deployment. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}is currently being deployed at [a large technology company[^2]]{} for its internal analytics tasks. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}represents a significant part of the company’s General Data Protection Regulation (GDPR) [@gdpr] compliance efforts, and provides both differential privacy and access control enforcement. We have made [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}available as open source [[@chorusdownloadanon]]{} to enable additional deployments elsewhere. #### Evaluation. In the first evaluation of its kind, we use [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}to evaluate the utility and performance of the selected mechanisms on real data. Our dataset includes [18,774]{}real queries written by analysts at [the company in which [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}is deployed]{}. #### Contributions. In summary, we make the following contributions: 1. We present [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}, representing a novel architecture for enforcing differential privacy on SQL queries that simultaneously supports a wide variety of mechanisms and runs on any standard SQL database (§ \[sec:ipq\]). 2. We describe and formalize the novel use of rule-based query rewriting to automatically transform an untrusted SQL query into an intrinsically private query using four example general-purpose mechanisms. We describe how other mechanisms can be supported using the same approach (§ \[sec:intr-priv-quer-1\]). 3. We report on our experience deploying [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}to enforce differential privacy at [a large technology company]{}, where it processes more than 10,000 queries per day (§ \[sec:implementation\]). 4. We use [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}to conduct the first large-scale empirical evaluation of the utility and performance of multiple general-purpose differential privacy on real queries and data (§ \[sec:evaluation\]). Background {#sec:backgr-diff-priv} ========== [ &gt;p[3cm]{} c l l c &gt;p[3.5cm]{}]{} &**General-**&&&**Sensitivity**& ****\ **Mechanism**&**Purpose** & **Strengths** & **Supported Constructs** & **Measure** & **Algorithmic Requirements**\ Elastic Sensitivity [@allegro] & & General analytics & Counting queries w/ joins & Local & Laplace noise\ Restricted Sensitivity [@Blocki:2013:DPD:2422436.2422449] & & Graph analysis & Counting queries w/ joins & Global & Laplace noise\ WPINQ [@proserpio2014calibrating] & & Synthetic data gen. & Counting queries w/ joins & Global & Weighted dataset operations, Laplace noise\ Sample & Aggregate [@smith2011privacy] & & Statistical estimators & Single-table aggregations & Local & Subsampling, Widened Winsorized mean, Laplace noise\ Differential privacy provides a formal guarantee of *indistinguishability*. This guarantee is defined in terms of a *privacy budget* $\epsilon$—the smaller the budget, the stronger the guarantee. The formal definition of differential privacy is written in terms of the *distance* $d(x,y)$ between two databases, i.e. the number of entries on which they differ: $d(x,y) = |\{i : x_i \neq y_i\}|$. Two databases $x$ and $y$ are *neighbors* if $d(x,y) = 1$. A randomized mechanism $\mathcal{K} : D^n \rightarrow \mathbb{R}$ preserves $(\epsilon, \delta)$-differential privacy if for any pair of neighboring databases $x, y \in D^n$ and set $S$ of possible outputs: $$\mbox{Pr}[\mathcal{K}(x)\in S] \leq e^\epsilon \mbox{Pr}[\mathcal{K}(y) \in S] + \delta$$ Differential privacy can be enforced by adding noise to the non-private results of a query. The scale of this noise depends on the *sensitivity* of the query. The literature considers two different measures of sensitivity: *global* [@dwork2006calibrating] and *local* [@nissim2007smooth]. For more on differential privacy, see Dwork and Roth [@dwork2014algorithmic]. #### Statistical queries. Differential privacy aims to protect the privacy of individuals in the context of *statistical queries*. In SQL terms, these are queries using standard aggregation operators (`COUNT`, `AVG`, etc.) as well as histograms created via the `GROUP BY` operator in which aggregations are applied to records within each group. Differential privacy is not suitable for queries that return raw data (e.g. rows in the database) since such queries are inherently privacy-violating. We formalize the targeted class of queries in Section \[sec:preliminaries\]. In Section \[sec:histogram\_queries\] we discuss how our approach supports histogram queries, which require special care to avoid leaking information via the presence of absence of groups. #### Mechanism design. Research on differential privacy has produced a large and growing number of differential privacy mechanisms. Some mechanisms are designed to provide broad support for many types of queries [@dwork2006calibrating; @nissim2007smooth; @mcsherry2007mechanism; @mcsherry2009privacy; @proserpio2014calibrating; @mohan2012gupt; @Blocki:2013:DPD:2422436.2422449], while others are designed to produce maximal utility for a particular application [@hay2009accurate; @sala2011sharing; @karwa2011private; @kasiviswanathan2013analyzing; @Chen:2013:RMT:2463676.2465304; @hardt2012simple; @li2014data; @li2015matrix; @zhang2014towards; @xiao2012dpcube; @cormode2012differentially; @qardaji2013differentially; @acs2012differentially; @xu2013differentially; @erlingsson2014rappor; @DBLP:conf/sigmod/HayMMCZ16]. While mechanisms adopt unique strategies for enforcing differential privacy in their target domain, they generally share a common set of design choices and algorithmic components. For example, many mechanisms require addition of Laplace noise to the result of the query. Our approach is motivated by the observation that a wide range of distinct mechanisms can supported with a common set of algorithmic building blocks. In this paper we formalize several example building blocks via *transformation rules* that describe how each algorithmic component can be embedded within a SQL query. We demonstrate the flexibility of this design by showing that each mechanism can be implemented simply by composing these transformation rules according to the mechanism’s definition. #### General-purpose mechanisms. The four mechanisms in Table \[fig:mechanisms\] are general-purpose because they support a broad range of queries, including commonly used SQL constructs such as join. This paper focuses on these four mechanisms. Many more specialized mechanisms have substantially similar algorithmic requirements and can be supported as intrinsically private queries using variations of the transformation rules introduced in this paper. Section \[sec:disc-key-take\] discusses this subject in more detail. The [[**[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}**]{}]{}Architecture {#sec:ipq} ======================================================================================== This section presents the system architecture and advantages of [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}, and compares it against existing architectures for differentially private analytics. We first describe the design goals motivating the [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}architecture. Then, in Section \[sec:existing\_systems\], we describe the limitations of existing architectures preventing previous work from attaining these goals. Finally, Section \[sec:chorus\_arch\] describes the novel architecture of [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}and provides an overview of our approach. #### Design Goals. The design of [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}is motivated by the desire to enforce differential privacy for general-purpose analytics in a practical setting. To that end, [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}has the following design goals: 1. Usability for non-experts 2. Support for a wide variety of analytics queries 3. Easy integration with existing data environments As we will demonstrate in the next section, achieving these goals is challenging, and no existing system manages to achieve all three. To achieve the first goal, a system should work with standard query languages (e.g. SQL). To achieve the second goal, a system should be able to leverage the many differential privacy mechanisms listed in Table \[fig:mechanisms\], and should select a mechanism for a given query automatically. To achieve the third goal, a system should be deployable in the context of an existing database engine. Existing Architectures {#sec:existing_systems} ---------------------- {width=".95\linewidth"} -------------------------------------------- ---------------------------------------------- ----------------------------- **Pros:** **Pros:** **Pros:** - Broad mechanism support - DBMS-independent - DBMS-independent **Cons:** **Cons:** - Broad mechanism support - Higher software complexity - Supports only post-processing mechanisms - New system required for each mechanism -------------------------------------------- ---------------------------------------------- ----------------------------- Existing systems for enforcing differential privacy for data analytics tasks adopt one of two architecture types: they are either *deeply integrated* systems or *post processing* systems. These architectures are summarized in Figure \[fig:sample\_architectures\](a) and Figure \[fig:sample\_architectures\](b). PINQ [@mcsherry2009privacy], Weighted PINQ [@proserpio2014calibrating], GUPT [@mohan2012gupt], and Airavat [@roy2010airavat] follow the *deep integration* architecture: each one provides its own specialized DBMS, and cannot be used with a standard DBMS. Elastic sensitivity [@allegro] uses the *post processing* architecture, in which the original query is run on the database and noise is added to the final result. This approach supports mechanisms that do not modify the semantics of the original query (PINQ and Restricted sensitivity [@Blocki:2013:DPD:2422436.2422449] could also be implemented this way), and has the advantage that it is compatibile with existing DBMSs. However, the post processing architecture is *fundamentally incompatible* with mechanisms that change how the original query executes—including WPINQ and Sample & Aggregate, listed in Table \[fig:mechanisms\]. The *deeply integrated* and *post processing* architectures in Figure \[fig:sample\_architectures\](a) and (b) therefore both fail to address two major challenges in implementing a practical system for differentially private data analytics: - Custom DBMSs are unlikely to achieve parity with mature DBMSs for a wide range of features including rich SQL support, broad query optimization, high-performance transaction support, recoverability, scalability and distribution. - Neither architecture supports the simultaneous application of a large number of different mechanisms. The deeply integrated architecture requires building a new DBMS for each mechanism, while the post processing architecture is inherently incompatible with some mechanisms. The [[**[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}**]{}]{}Architecture {#sec:chorus_arch} ---------------------------------------------------------------------------------------- In [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}, we propose a novel alternative to the deeply integrated and post processing architectures used by existing systems. As shown in Figure \[fig:sample\_architectures\](c), [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}transforms the input query into an *intrinsically private query*, which is a standard SQL query whose results are *guaranteed to be differentially private*. An intrinsically private query provides this guarantee by embedding a differential privacy mechanism in the query itself. When executed on an unmodified SQL database, the embedded privacy mechanism ensures that the query’s results preserve differential privacy. The approach has three key advantages over previous work: - [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}is DBMS-independent (unlike the deeply integrated approach): it requires neither modifying the database nor switching to purpose-built database engines. Our approach can therefore leverage existing high-performance DBMSs to scale to big data. - [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}can implement a wide variety of privacy-preserving techniques. Unlike the post processing approach, [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}is compatible with all of the mechanisms listed in Table \[fig:mechanisms\], and many more. - [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{} eliminates the need for post-processing, allowing easy integration in existing data processing pipelines. Our approach enables a single data processing pipeline for all mechanisms. By adopting this novel architecture, [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}achieves all three design goals listed earlier. Since input queries are considered untrusted and the rewriting engine uses program analysis techniques, [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}preserves differential privacy even in the face of malicious analysts. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}’s architecture is specifically designed to be easily integrated into existing data environments. We report on the deployment of [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}[in a large technology company]{} in Section \[sec:implementation\]. #### Constructing intrinsically private queries. The primary challenge in implementing this architecture is transforming untrusted queries into intrinsically private queries. This process must be flexible enough to support a wide variety of privacy mechanisms and also general enough to support ad-hoc SQL queries. As described earlier, constructing intrinsically private queries automatically has additional advantages. This approach protects against malicious analysts by guaranteeing differentially private results by construction. It is also transparent to the analyst since it does not require input from the analyst to preserve privacy or select a privacy mechanism. As shown in Figure \[fig:sample\_architectures\](c), [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}constructs intrinsically private queries in two steps: 1. The *Mechanism Selection* component automatically selects an appropriate differential privacy mechanism to apply. 2. The *Query Rewriting* component embeds the selected mechanism in the input query, transforming it into an intrinsically private query. To select a mechanism, [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}leverages an extensible mechanism selection engine, discussed next. The query rewriting step then transforms the input query to produce an intrinsically private query embedding the selected mechanism. For this step, we employ a novel use of rule-based query rewriting, which has been studied extensively for query optimization but, to our knowledge, never applied to differential privacy. We introduce our solution by example in Section \[sec:intr-priv-quer-1\]and formalize it in Section \[sec:building-blocks\]. This paper focuses on differential privacy, but the same approach could be used to enforce other types of privacy guarantees or security policies [@stonebraker1974access]. #### Mechanism selection. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}implements an extensible *mechanism selection engine* that automatically selects a differential privacy mechanism for each input query. This engine can be extended based on available mechanisms, performance and utility goals, and to support custom mechanism selection techniques. For example, Hay et al. [@kotsogiannis2017pythia] demonstrate that a machine learning-based approach can leverage properties of the data to select a mechanism most likely to yield high utility. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}is designed to support any such approach. We describe an example of a syntax-based mechanism selection in Section \[sec:implementation\]. #### Privacy budget management. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}does not prescribe a specific privacy budget management strategy, as the best way to manage the available privacy budget in practice will depend on the deployment scenario and threat model. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}provides flexibility in how the budget is managed: the sole requirement is that rewriters are supplied with the $\epsilon$ value apportioned to each query.[^3] For the case of a single global budget, where [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}is deployed as the sole interface to the database, [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}can track the remaining budget according to the standard composition theorem for differential privacy [@dwork2006calibrating]. When a new query is submitted, [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}subtracts from the remaining budget the $\epsilon$ value allocated to that query, and refuses to process new queries when the budget is exhausted. In Section \[sec:disc-key-take\], we discuss more sophisticated methods that may yield better results for typical deployments. Query Rewriting {#sec:intr-priv-quer-1} =============== This section demonstrates our approach [by example]{}, using the four general-purpose differential privacy mechanisms listed in Table \[fig:mechanisms\]. For each mechanism, we briefly review the algorithm used. Then, we describe, using simple example queries, how an input SQL query can be systematically transformed into an intrinsically private query embedding that algorithm, and give an argument for the correctness of each transformation. Sensitivity-Based Mechanisms ---------------------------- We first consider two mechanisms that add noise to the final result of the query: Elastic Sensitivity [@allegro] and Restricted Sensitivity [@Blocki:2013:DPD:2422436.2422449]. Elastic Sensitivity is a bound on the local sensitivity of a query, while Restricted Sensitivity is based on global sensitivity. Both approaches add Laplace noise to the query’s result. For a query with sensitivity $s$ returning value $v$, the Laplace mechanism releases $v + \operatorname{Lap}(s / \epsilon)$, where $\epsilon$ is the privacy budget allocated to the query. Given a random variable $U$ sampled from the uniform distribution, we can compute the value of a random variable $X \sim \operatorname{Lap}(s / \epsilon)$: $$X= - \frac{s}{\epsilon}\,\mbox{sign}(U)\,\ln(1 - 2|U|)$$ In SQL, we can sample from the uniform distribution using `RANDOM()`. Consider the following query, which returns a (non–differentially private) count of trips in the database. This query can be transformed into an intrinsically private query as follows: SELECT COUNT(*) AS count FROM trips [[$\Downarrow$]{}]{} WITH orig AS (SELECT COUNT(*) AS count FROM trips), uniform AS (SELECT *, RANDOM()(*@-@*)0.5 AS u FROM orig) [[$\Downarrow$]{}]{} WITH orig AS (SELECT COUNT(*) AS count FROM trips), uniform AS (SELECT *, RANDOM()(*@-@*)0.5 AS u FROM orig) SELECT count(*@-@*)((*@ $s / \epsilon$ @*))*SIGN(u)*LN(1(*@-@*)2*ABS(u)) AS count FROM uniform The first step defines $U$ using `RANDOM()`, and the second uses $U$ to compute the corresponding Laplace noise. The correctness of this approach follows from the definition of the Laplace mechanism. The two transformation steps combined clearly generate Laplace noise with the correct scale, and add it to the sensitive query results. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}can calculate either Elastic Sensitivity with smoothing via smooth sensitivity [@nissim2007smooth][^4] or Restricted Sensitivity via a dataflow analysis of the query. Such an analysis is described in [@allegro]. We formalize the construction of intrinsically private queries via sensitivity-based approaches using the *Laplace Noise transformation*, defined in Section \[sec:building-blocks\]. Weighted PINQ {#sec:weighted-pinq} ------------- Weighted PINQ (WPINQ) enforces differential privacy for counting queries with equijoins. A key distinction of this mechanism is that it produces a differentially private *metric* (called a *weight*), rather than a count. These weights are suitable for use in a workflow that generates differentially private synthetic data, from which counts are easily derived. The workflow described in [@proserpio2014calibrating] uses weights as input to a Markov chain Monte Carlo (MCMC) process. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}’s implementation of WPINQ computes noisy weights for a given counting query according to the mechanism’s definition [@proserpio2014calibrating]. Since the weights are differentially private, they can be released to the analyst for use with any desired workflow. The WPINQ mechanism adds a weight to each row of the database, updates the weights as the query executes to ensure that the query has a sensitivity of 1, and uses the Laplace mechanism to add noise to the weighted query result. WPINQ has been implemented as a standalone data processing engine with a specialized query language, but has not been integrated into any SQL DBMS. Where a standard database is a collection of tuples in $D^n$, a weighted database (as defined in Proserpio et al. [@proserpio2014calibrating]) is a function from a tuple to its weight ($D \rightarrow \mathbb{R}$). In this setting, counting the number of tuples with a particular property is analogous to summing the weights of all such tuples. Counting queries can therefore be performed using *sum*. In fact, summing weights in a weighted dataset produces exactly the same result as the corresponding counting query on the original dataset, when the query does not contain joins. When the query does contain joins, WPINQ scales the weight of each row of the join’s output to maintain a sensitivity of 1. Proserpio et al. [@proserpio2014calibrating] define the weight of each row in a join as follows: $$A {\bowtie} B = \sum_k \frac{A_k \times B_k^T}{||A_k|| + ||B_k||} \label{eqn:wpinq_join}$$ Since the scaled weights ensure a sensitivity of 1, Laplace noise scaled to $1/\epsilon$ is sufficient to enforce differential privacy. WPINQ adds noise with this scale to the results of the weighted query. In SQL, we can accomplish the first task (adding weights) by adding a column to each relation. Consider transforming our previous example query: SELECT COUNT(*) FROM trips [[$\Downarrow$]{}]{} SELECT SUM(weight) FROM (SELECT *, 1 AS weight FROM trips) This transformation adds a weight of 1 to each row in the table, and changes the `COUNT` aggregation function into a `SUM` of the rows’ weights. The correctness of this transformation is easy to see: as required by WPINQ [@proserpio2014calibrating], the transformed query adds a weight to each row, and uses `SUM` in place of `COUNT`. We can accomplish the second task (scaling weights for joins) by first calculating the norms $||A_K||$ and $||B_k||$ for each key $k$, then the new weights for each row using $A_k \times B_k^T$. For a join between the `trips` and `drivers` tables, for example, we can compute the norms for each key: WITH tnorms AS (SELECT driver_id, SUM(weight) AS norm FROM trips GROUP BY driver_id), dnorms AS (SELECT id, SUM(weight) AS norm FROM drivers GROUP BY id) Then, we join the norms relations with the original results and scale the weight for each row: SELECT ..., (t.weight*d.weight)/(tn.norm+dn.norm) AS weight FROM trips t, drivers d, tnorm tn, dnorm dn WHERE t.driver_id = d.id AND t.driver_id = tn.driver_id AND d.id = dn.id The correctness of this transformation follows from equation . The relation `tnorms` corresponds to $||A_k||$, and `dnorms` to $||B_k||$. For each key, `t.weight` corresponds to $A_k$, and `d.weight` to $B_k$. Finally, we can accomplish the third task (adding Laplace noise scaled to $1/\epsilon$) as described earlier. We formalize the construction of intrinsically private queries via WPINQ using three transformations: the *Metadata Propagation* transformation to add weights to each row, the *Replace Aggregation Function* transformation to replace counts with sums of weights, and the *Laplace Noise* transformation to add noise to the results. All three are defined in Section \[sec:building-blocks\]. Sample & Aggregate ------------------ The Sample & Aggregate [@nissim2007smooth; @smith2011privacy] mechanism works for all statistical estimators, but does not support joins. Sample & Aggregate has been implemented in GUPT [@mohan2012gupt], a standalone data processing engine that operates on Python programs, but has never been integrated into a practical database. As defined by Smith [@smith2011privacy], the mechanism has three steps: 1. Split the database into disjoint *subsamples* 2. Run the query on each subsample independently 3. Aggregate the results using a differentially-private algorithm For differentially-private aggregation, Smith [@smith2011privacy] suggests *Widened Winsorized mean*. Intuitively, Widened Winsorized mean first calculates the *interquartile range*—the difference between the 25th and 75th percentile of the subsampled results. Next, the algorithm *widens* this range to include slightly more data points, then clamps the subsampled results to lie within the widened range. This step eliminates outliers, which is important for enforcing differential privacy. Finally, the algorithm takes the average of the clamped results, and adds Laplace noise scaled to the size of the range (i.e. the effective sensitivity) divided by $\epsilon$. In SQL, we can accomplish tasks 1 and 2 by adding a `GROUP BY` clause to the original query. Consider a query that computes the average of trip lengths: SELECT AVG(trip_distance) FROM trips [[$\Downarrow$]{}]{} SELECT AVG(trip_distance), ROW_NUM() MOD (*@ n @*) AS samp FROM trips GROUP BY samp This transformation generates $n$ subsamples and runs the original query on each one. The correctness of tasks 1 and 2 follows from the definition of subsampling: the `GROUP BY` ensures that the subsamples are disjoint, and that the query runs on each subsample independently. To accomplish task 3 (differentially private aggregation), we can use a straightforward implementation of Widened Winsorized mean in SQL, since the algorithm itself is the same for each original query. We formalize the construction of intrinsically private queries via Sample & Aggregate using the *Subsampling* transformation, defined in Section \[sec:building-blocks\]. Formalization & Correctness {#sec:building-blocks} =========================== This section formalizes the construction of intrinsically private queries as introduced in Section \[sec:intr-priv-quer-1\]. We begin by introducing notation (Section \[sec:preliminaries\]). We then define reusable transformation rules (Section \[sec:transformation-rules\]) that can be composed to construct mechanisms. Next, we formalize the four mechanisms described earlier in terms of these rules (Section \[sec:mech-defin\]). Finally, we prove a correctness property: our transformations do not modify the semantics of the input query (Section \[sec:corr-prop\]). By formalizing the transformation rules separately from the individual mechanisms, we allow the rules to be re-used in defining new mechanisms—taking advantage of the common algorithmic requirements demonstrated in Table \[fig:mechanisms\]. An additional benefit of this approach is the ability to prove correctness properties of the rules themselves, so that these properties extend to all mechanisms implemented using the rules. Preliminaries {#sec:preliminaries} ------------- #### Core relational algebra. We formalize our transformations as rewriting rules on a core relational algebra that represents general statistical queries. We define the core relational algebra in Figure \[fig:core-relational-algebra\]. This algebra includes the most commonly-used features of query languages like SQL: selection ($\sigma$), projection ($\Pi$), equijoins ($\bowtie$), and counting with and without aggregation. We also include several features specifically necessary for our implementations: constant values, random numbers, and the arithmetic functions *ln*, *abs*, and *sign*. We use standard notation for relational algebra with a few exceptions. We extend the projection operator $\Pi$ to attribute expressions, which allows projection to add new named columns to a relation. For example, if relation $r$ has schema $U$, then the expression $\Pi_{U \cup \mbox{weight}:1}r$ adds a new column named “weight” with the default value 1 for each row to the existing columns in $r$. In addition, we combine aggregation and grouping, writing $\underset{a_1..a_n}{{\mbox{\emph{Count}}}}$ to indicate a counting aggregation with grouping by columns $a_1..a_n$. We write $\underset{a_1..a_n}{{\mbox{\emph{Sum}}}_a}$ to indicate summation of column $a$, grouping by columns $a_1..a_n$. #### Notation for rewriting rules. Each transformation is defined as a set of inference rules that rewrites a relational algebra expression. A particular rule allows rewriting an expression as specified in its conclusion (below the line) if the conditions specified in its antecedent (above the line) are satisfied (either through syntactic properties of the query or by applying another inference rule). Our approach relies on the ability to analyze and rewrite SQL queries. This rewriting can be achieved by a rule-based query optimization engine [@calcite; @GPORCA]. Most of the conditions specified in our rewriting rules use standard notation. One exception is conditions of the form $Q : U$, which we use to denote that the query $Q$ results in a relation with schema $U$. We extend this notation to denote the schemas of database tables (e.g. $t : U$) and relational expressions (e.g. $r_1 : U$). Some of our rewriting rules have global parameters, which we denote by setting them above the arrow indicating the rewriting rule itself. For example, $r \overset{x}{\rightarrow} \Pi_x r$ allows rewriting $r$ to project only the column named $x$, where the value of $x$ is provided as a parameter. Most parameters are values, but parameters can also be functions mapping a relational algebra expression to a new expression. For example, $r \overset{f}{\rightarrow} f(r)$ indicates rewriting $r$ to the result of $f(r)$. $$\begin{array}{lcl} \multicolumn{3}{l}{\cellcolor{gray!15}\mbox{Attribute expressions}}\\ a & & \emph{attribute name}\\ e & ::= & a \;|\; a:v \\ v & ::= & a \;|\; n \in \mathbb{N} \;|\; v_1 + v_2 \;|\; v_1 * v_2 \;|\; v_1 / v_2\\ & | & {\mbox{\emph{rand}}}() \;|\; {\mbox{\emph{randInt}}}(n) \;|\; {\mbox{\emph{ln}}}(v) \;|\; {\mbox{\emph{abs}}}(v) \;|\; {\mbox{\emph{sign}}}(v) \\ \\[-2mm] \multicolumn{3}{l}{\cellcolor{gray!15}\mbox{Relational transformations}}\\ R & ::= & t \;|\; R_1 \underset{x = y}{\bowtie} R_2 \;|\; \Pi_{e_1, ..., e_n} R \;|\; \sigma_\varphi R \vspace*{1mm}\\ & | & {\mbox{\emph{Count}}}(R) \;|\; \underset{a_1..a_n}{{\mbox{\emph{Count}}}}(R) \\ \\[-2mm] \multicolumn{3}{l}{\cellcolor{gray!15}\mbox{Selection predicates}}\\ \varphi & ::= & e_1 \theta e_2 \;|\; e \theta v\\ \theta & ::= & < \;|\; \leq \;|\; = \;|\; \neq \;|\; \geq \;|\; >\\ \\[-2mm] \multicolumn{3}{l}{\cellcolor{gray!15}\mbox{Queries}}\\ Q & ::= & {\mbox{\emph{Count}}}(R) \;|\; \underset{a_1..a_n}{{\mbox{\emph{Count}}}}(R) \;|\; {\mbox{\emph{Sum}}}_a (R) \;|\; \underset{a_1..a_n}{{\mbox{\emph{Sum}}}_a} (R) \end{array}$$ Transformation Rules {#sec:transformation-rules} -------------------- #### Laplace Noise. All four mechanisms in Table \[fig:mechanisms\] require generating noise according to the Laplace distribution. We accomplish this task using the *Laplace Noise* transformation, defined in Figure \[fig:laplace-noise\]. This transformation has one parameter: $\gamma$, which defines the scale of the noise to be added to the query’s result. For a query with sensitivity 1 and privacy budget $\epsilon$, for example, $\gamma = 1/\epsilon$ suffices to enforce differential privacy. The Laplace Noise transformation defines a single inference rule. This rule allows rewriting a top-level query with schema $U$ according to the $\operatorname{Lap}$ function. $\operatorname{Lap}$ wraps the query in two projection operations; the first (defined in $\operatorname{Unif}$) samples the uniform distribution for each value in the result, and the second (defined in $\operatorname{Lap}$) uses this value to add noise drawn from the Laplace distribution. $$\begin{array}{c} \fbox{$Q {\overset{\gamma}{\rightarrow}}Q$} \quad \quad \inference {Q:U}{Q {\overset{\gamma}{\rightarrow}}\operatorname{Lap}(Q)} \\[2mm] \begin{aligned} \mbox{where}&\\ \operatorname{Unif}(Q) &= \Pi_{U \cup \{{\mbox{u}}_x : {\mbox{\emph{rand}}}() - 0.5 | x \in U\}}(Q))\\ \operatorname{Lap}(Q) &= \Pi_{\{x : x + - \gamma\, {\mbox{\emph{sign}}}({\mbox{u}}_x) {\mbox{\emph{ln}}}(1 - 2\, {\mbox{\emph{abs}}}({\mbox{u}}_x)) | x \in U \}}(\operatorname{Unif}(Q))\\ \end{aligned} \end{array}$$ #### Metadata Propagation. Many mechanisms require tracking metadata about each row as the query executes. To accomplish this, we define the *Metadata Propagation* transformation in Figure \[fig:metadata\]. The Metadata Propagation transformation adds a column to each referenced table and initializes its value for each row in that table, then uses the specified composition functions to compose the values of the metadata column for each resulting row of a join or an aggregation. The Metadata Propagation transformation has three parameters: ${i}$, a function defining the initial value of the metadata attribute for each row in a database table; ${j}$, a function specifying how to update the metadata column for a join of two relations; and ${c}$, a function specifying how to update the metadata column for subqueries. The inference rule for a table $t$ uses projection to add a new attribute to $t$’s schema to hold the metadata, and initializes that attribute to the value of ${i}()$. The rules for projection and selection simply propagate the new attribute. The rule for joins applies the ${j}$ function to perform a localized update of the metadata column. The rules for counting subqueries invoke the update function ${c}$ to determine the new value for the metadata attribute. Finally, the rules for counting queries eliminate the metadata attribute to preserve the top-level schema of the query. $$\begin{array}{c} \fbox{$R {\overset{{i},{j},{c}}{\rightarrow}}R$} \quad \inference{r_1 {\overset{{i},{j},{c}}{\rightarrow}}r_1' \quad r_2 {\overset{{i},{j},{c}}{\rightarrow}}r_2' \quad r_1':U_1 \quad r_2':U_2} {r_1 \underset{A = B}{\bowtie} r_2 {\overset{{i},{j},{c}}{\rightarrow}}{j}(r_1' \underset{A = B}{\bowtie} r_2')} \\[{8mm}] \inference {t:U \quad {\mbox{m}}\not \in U}{t {\overset{{i},{j},{c}}{\rightarrow}}\Pi_{U \cup {\mbox{m}}: {i}()} t} \quad \inference{r {\overset{{i},{j},{c}}{\rightarrow}}r'} {\Pi_U(r) {\overset{{i},{j},{c}}{\rightarrow}}\Pi_{U \cup \{{\mbox{m}}\}}(r')} \\[{8mm}] \inference{r {\overset{{i},{j},{c}}{\rightarrow}}r'} {\sigma_\phi(r) {\overset{{i},{j},{c}}{\rightarrow}}\sigma_\phi(r')} \quad \inference{r {\overset{{i},{j},{c}}{\rightarrow}}r' \quad {\mbox{\emph{Count}}}(r') : U} {{\mbox{\emph{Count}}}(r) {\overset{{i},{j},{c}}{\rightarrow}}\Pi_{U \cup {\mbox{m}}: {c}(r') } {\mbox{\emph{Count}}}(r')} \\[{8mm}] \inference{r {\overset{{i},{j},{c}}{\rightarrow}}r' \quad \underset{a_1..a_n}{{\mbox{\emph{Count}}}}(r') : U} {\underset{a_1..a_n}{{\mbox{\emph{Count}}}}(r) {\overset{{i},{j},{c}}{\rightarrow}}\Pi_{U \cup {\mbox{m}}: {c}(r') } \underset{a_1..a_n}{{\mbox{\emph{Count}}}}(r')} \\[{8mm}] \fbox{$Q {\overset{{i},{j},{c}}{\rightarrow}}Q$} \hspace*{1mm} \inference{r {\overset{{i},{j},{c}}{\rightarrow}}r'} {{\mbox{\emph{Count}}}(r) {\overset{{i},{j},{c}}{\rightarrow}}{\mbox{\emph{Count}}}(r')} \hspace*{1mm} \inference{r {\overset{{i},{j},{c}}{\rightarrow}}r'} {\underset{a_1..a_n}{{\mbox{\emph{Count}}}}(r) {\overset{{i},{j},{c}}{\rightarrow}}\underset{a_1..a_n}{{\mbox{\emph{Count}}}}(r')} \end{array}$$ #### Replacing Aggregation Functions. The *Replace Aggregation Function* transformation, defined in Figure \[fig:aggfn\], allows replacing one aggregation function ($\Gamma$) with another ($\Gamma'$). To produce syntactically valid output, $\Gamma$ and $\Gamma'$ must be drawn from the set of available aggregation functions. The antecedent is empty in the rewriting rules for this transformation because the rules operate only on the outermost operation of the query. $$\begin{array}{c} \fbox{$Q {\overset{\Gamma, \Gamma'}{\rightarrow}}Q$} \quad \inference{} {\Gamma(r) {\overset{\Gamma, \Gamma'}{\rightarrow}}\Gamma'(r)} \quad \inference{} {\underset{a_1..a_n}{\Gamma}(r) {\overset{\Gamma, \Gamma'}{\rightarrow}}\underset{a_1..a_n}{\Gamma'}(r)} \end{array}$$ #### Subsampling. The *Subsampling* transformation, defined in Figure \[fig:sampling\], splits the database into disjoint subsamples, runs the original query on each subsample, and aggregates the results according to a provided function. The Subsampling transformation is defined in terms of the Metadata Propagation transformation, and can be used to implement partition-based differential privacy mechanisms like Sample & Aggregate. The parameters for the Subsampling transformation are ${\mathcal{A}}$, a function that aggregates the subsampled results, and $n$, the number of disjoint subsamples to use during subsampling. Both inference rules defined by the transformation rewrite the queried relation using the Metadata Propagation transformation ($r {\overset{{i},{j},{c}}{\rightarrow}}r'$). The parameters for Metadata Propagation initialize the metadata attribute with a random *subsample number* between zero and $n$, and propagate the subsample number over counting subqueries. The update functions for joins and counting subqueries is undefined, because subsampling is incompatible with queries containing these features.[^5] In order to satisfy the semantics preservation property, the aggregation function ${\mathcal{A}}$ must transform the query results on each subsample into a final result with the same shape as the original query. Many aggregation functions satisfy this property (e.g. the mean of all subsamples), but not all of them provide differential privacy. $$\begin{array}{c} \fbox{$Q {\overset{\mathcal{A}}{\rightarrow}}Q$} \quad \quad \inference{r {\overset{{i},{j},{c}}{\rightarrow}}r'} {{\mbox{\emph{Count}}}(r) {\overset{\mathcal{A}}{\rightarrow}}{\mathcal{A}}( \underset{m}{{\mbox{\emph{Count}}}}(r'))} \\[{8mm}] \inference{r {\overset{{i},{j},{c}}{\rightarrow}}r'} {\underset{G_1..G_n}{{\mbox{\emph{Count}}}}(r) {\overset{\mathcal{A}}{\rightarrow}}{\mathcal{A}}( \underset{G_1..G_n, m}{{\mbox{\emph{Count}}}}(r') )} \\\\ \begin{aligned} \mbox{where } {i}&= {\mbox{\emph{randInt}}}(n), {j}= \bot, {c}(r) = \bot, \mbox{and } {\mathcal{A}}\mbox{ aggregates over } {\mbox{m}}\end{aligned} \end{array}$$ Mechanism Definitions {#sec:mech-defin} --------------------- We now formally define the mechanisms described earlier in terms of the transformations defined in Section \[sec:transformation-rules\]. We describe each mechanism as a series of one or more transformations, and define the parameters for each transformation to obtain the correct semantics for the mechanism. Let $s$ be the Elastic Sensitivity of $Q$. Let $\gamma = s / \epsilon$. If $Q {\overset{\gamma}{\rightarrow}}Q'$, then $Q'$ is an intrinsically private query via Elastic Sensitivity. Let $s$ be the restricted sensitivity of $Q$. Let $\gamma = s / \epsilon$. If $Q {\overset{\gamma}{\rightarrow}}Q'$, then $Q'$ is an intrinsically private query via Restricted Sensitivity. Let ${i}= 1$ and let ${j}$ scale weights as described in Section \[sec:weighted-pinq\]. Let $\gamma = 1/\epsilon$. If $Q {\overset{{i},{j},{c}}{\rightarrow}}Q' {\overset{\Gamma, \Gamma'}{\rightarrow}}Q'' {\overset{\gamma}{\rightarrow}}Q'''$, then $Q'''$ is an intrinsically private query via Weighted PINQ. Let ${\mathcal{A}}$ implement private Winsorized mean [@smith2011privacy] for a desired $\epsilon$. If $Q {\overset{\mathcal{A}}{\rightarrow}}Q'$, then $Q'$ is an intrinsically private query via Sample & Aggregate. Correctness {#sec:corr-prop} ----------- The correctness criterion for a traditional query rewriting system is straightforward: the rewritten query should have the same semantics (i.e. return the same results) as the original query. For intrinsically private queries, however, the definition of correctness is less clear: enforcing differential privacy *requires* modifying the results to preserve privacy. In this section, we define *semantic preservation*, which formalizes the intuitive notion that our transformations should perturb the result attributes of the query, as required for enforcing differential privacy, but change nothing else about its semantics. Semantic preservation holds when the transformation in question preserves the size and shape of the query’s output, and additionally preserves its logical attributes. Logical attributes are those which are used as join keys or to perform filtering (i.e. the query makes decisions based on these attributes, instead of simply outputting them). Each of our transformations preserve this property. Furthermore, semantic preservation is preserved over composition of transformations, so semantic preservation holds for *any* mechanism defined using our transformations—including those defined in Section \[sec:mech-defin\]. An attribute $a$ is a logical attribute if it appears as a join key in a join expression, in the filter condition $\varphi$ of a filter expression, or in the set of grouping attributes of a ${\mbox{\emph{Count}}}$ or ${\mbox{\emph{Sum}}}$ expression. A transformation ($\rightarrow$) satisfies *semantic preservation* if for all queries $Q$ and $Q'$, if $Q \rightarrow Q'$, then (1) the schema is preserved: $Q: U \Rightarrow Q': U$; (2) the number of output rows is preserved: $|Q| = |Q'|$; and (3) logical attributes are preserved (see Definition \[def:logic-pres\]). \[def:shape-pres\] Consider a transformation $Q:U \rightarrow Q':U$. Split $U$ into two sets of attributes $\{U_k, U_a\}$, such that $U_k$ contains all of the attributes from $U$ used as logical attributes in $Q$ and $U_a$ contains all of the other attributes. Now construct $Q'_r$ by renaming each attribute $k \in U_k$ in $Q'$ to $k'$. Then $\rightarrow$ preserves logical attributes if there exists a one-to-one relation $E \subseteq Q \times Q'_r$ such that for all $e \in E$ and $k \in U_k$, $e_k = e_{k'}$. \[def:logic-pres\] If two transformations ($\rightarrow_a$ and $\rightarrow_b$) both satisfy semantic preservation, then their composition ($\rightarrow_c$) satisfies semantic preservation: for all queries $Q$, $Q'$, and $Q''$, if $Q \rightarrow_a Q' \rightarrow_b Q''$ implies that $Q \rightarrow_c Q''$, and both $\rightarrow_a$ and $\rightarrow_b$ preserve semantics, then $\rightarrow_c$ preserves semantics. \[thm:composition\_block\] Assume that $\rightarrow_a$ and $\rightarrow_b$ preserve semantics, and $Q \rightarrow_a Q' \rightarrow_b Q''$. We have that: (1) $Q:U$, $Q':U$, and $Q:U''$ ; (2) $|Q| = |Q'| = |Q''|$; and (3) $Q$, $Q'$, and $Q''$ contain the same logical attributes. Thus by Definition \[def:shape-pres\], $\rightarrow_c$ preserves semantics. Semantic preservation holds for the transformations defined in in Section \[sec:transformation-rules\], and thus for the every mechanism defined by composing these transformations. \[thm:defin-semant-pres\] See the extended version of this paper [@chorus-arxiv-personal-webpage] for proofs of Theorems \[thm:composition\_block\] and \[thm:defin-semant-pres\]. The Laplace Noise transformation (Figure \[fig:laplace-noise\]) satisfies the semantic preservation property. \[thm:lap\_block\] The rules in Figure \[fig:laplace-noise\] define only one transformation, at the top level of the query. The $\operatorname{Unif}$ function adds a column $u_x$ for each original column $x$; the $\operatorname{Lap}$ function consumes this column, replacing the original column $x$ with its original value plus a value sampled from the Laplace distribution. The outermost projection produces exactly the same set of attributes as the input query $Q$, preserving the schema; the transformation adds only projection nodes, so the number of output rows is preserved; no join or select can occur at the top level of a query, so no logical attributes are modified. The Metadata Propagation transformation (Figure \[fig:metadata\]) satisfies the semantic preservation property. \[thm:metadata\_block\] The only rules that make major modifications to the original query are those for joins and counts. The other rules add a new attribute ${\mbox{m}}$ and propagate it through the query, but do not change the number or contents of the rows of any relation. At the top level of the query (i.e. $q \in Q$), the transformation eliminates the attribute ${\mbox{m}}$. For queries without joins or subquery aggregations, the Metadata Propagation transformation is the identity transformation, so it satisfies semantic preservation. We argue the remaining cases by induction on the structure of $Q$. [**Case $r_1 \underset{A = B}{\bowtie} r_2$.**]{} Let $r = r_1 \underset{A = B}{\bowtie} r_2$. If ${j}$ does not change the query’s semantics, except to update the attribute ${\mbox{m}}$ (i.e. $r:U \Rightarrow r = \Pi_{U-m} {j}(r)$), then the semantics of $r$ are preserved. [**Case ${\mbox{\emph{Count}}}(r)$ and ${\mbox{\emph{Count}}}_{G_1..G_n}(r)$.**]{} The rule modifies the ${\mbox{m}}$ attribute, but does not modify any other attribute or change the size of the relation, so semantics are preserved. The Replace Aggregation Function transformation (Figure \[fig:aggfn\]) satisfies the semantic preservation property. \[thm:aggfn\_block\] Aggregation function replacement has the potential to modify the values of the query’s results, but not its shape or logical attributes. The transformation’s only rule allows changing one function to another, but preserves the grouping columns and number of functions. The schema, number of rows, and logical attributes are therefore preserved. If the aggregation function ${\mathcal{A}}$ aggregates over the ${\mbox{m}}$ attribute, then our Subsampling transformation (Figure \[fig:sampling\]) satisfies the semantic preservation property. \[thm:samp\_block\] We know that $Q$ has the form ${\mbox{\emph{Count}}}(r)$ or $\underset{G_1..G_n, m}{{\mbox{\emph{Count}}}}(r)$. By Theorem \[thm:metadata\_block\], we know that in either case, if $r {\overset{{i},{j},{c}}{\rightarrow}}r'$, then $r$ has the same semantics as $r'$. We proceed by cases. [**Case $Q = {\mbox{\emph{Count}}}(r)$.**]{} Let $q_1 = {\mbox{\emph{Count}}}_{{\mbox{m}}}(r')$. By the definition of the transformation, $Q' = {\mathcal{A}}(q_1)$. The query $q_1$ has exactly one row per unique value of ${\mbox{m}}$. Since ${\mathcal{A}}$ aggregates over ${\mbox{m}}$, ${\mathcal{A}}(q_1)$ has exactly one row, and therefore preserves the semantics of $Q$. [**Case $Q = \underset{G_1..G_n, m}{{\mbox{\emph{Count}}}}(r)$.**]{} Let $q_1 = \underset{G_1..G_n, m}{{\mbox{\emph{Count}}}}(r')$. By the definition of the transformation, $Q' = {\mathcal{A}}(q_1)$. The query $q_1$ has exactly one row per unique tuple $(G_1..G_n, {\mbox{m}})$. Since ${\mathcal{A}}$ aggregates over ${\mbox{m}}$, ${\mathcal{A}}(q_1)$ has exactly one row per unique tuple $(G_1..G_n)$, and therefore preserves the semantics of $Q$. Handling Histogram Queries {#sec:histogram_queries} -------------------------- SQL queries can use the `GROUP BY` operator to return a relation representing a histogram, as in the following query which counts the number of trips greater than 100 miles in each city: SELECT city_id, COUNT(*) as count FROM trips WHERE distance > 100 GROUP BY city_id This type of query presents a problem for differential privacy because the presence or absence of a particular city in the results reveals whether the count for that city was zero. The general solution to this problem is to require the analyst to explicitly provide a set of desired bins, and return a (noisy) count of zero for absent bins. Such an approach is used, for example, in PINQ [@mcsherry2009privacy], Weighted PINQ [@proserpio2014calibrating], and FLEX [@allegro]. Unfortunately, this approach impose an unfamiliar user experience and is burdensome for the analyst, who is never allowed to view the results directly. For bins with finite domain, [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}provides a superior solution by enumerating missing histogram bins automatically *in the rewritten query*. The missing bins are populated with empty aggregation values (e.g., 0 for counts) before mechanism-specific rewriting, at which point they are handled identically as non-empty bins. This allows the full results of histogram queries to be returned to the analyst without post-processing or interposition. If the domain cannot be enumerated (e.g., because it is unbounded), [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}falls back to the approach described above and does not release results directly to the analyst. This feature requires the operator to define a mapping from columns that may be used in a `GROUP BY` clause to the database field containing the full set of values from that column’s domain. This information may be defined manually or extracted from the database schema (e.g., via primary and foreign key constraints), and is provided during initial deployment. In this example, suppose the full set of city ids are stored in column `city_id` of table `cities`. Using this information, [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}generates the following intermediate query: WITH orig AS ( SELECT city_id, COUNT(*) as count FROM trips WHERE distance > 100 GROUP BY city_id ) SELECT cities.city_id, (CASE WHEN orig.count IS NULL THEN 0 ELSE orig.count END) as count FROM orig RIGHT JOIN cities ON orig.city_id = cities.city_id The `RIGHT JOIN` ensures that exactly one row exists in the output relation for each city\_id in `cities`, and the `CASE` expression outputs a zero for each missing city in the original query’s results. The query thus contains *every* city\_id value regardless of the semantics of the original query. This intermediate query is then sent to the mechanism rewriter, which adds noise to each bin as normal. Implementation & Deployment {#sec:implementation} =========================== This section describes our implementation of [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}and our experience deploying it to enforce differential privacy at [a large technology company]{}. Implementation -------------- Our implementation of [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}automatically transforms an input SQL query into an intrinsically private query. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}currently supports the four differential privacy mechanisms discussed here, and is designed for easy extension to new mechanisms. We have released [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}as an open source project [[@chorusdownloadanon]]{}. Our implementation is built on Apache Calcite [@calcite], a generic query optimization framework that transforms input queries into a relational algebra tree and provides facilities for transforming the tree and emitting a new SQL query. We built a custom dataflow analysis and rewriting framework on Calcite to support intrinsically private queries. The framework, mechanism-specific analyses, and rewriting rules are implemented in 5,096 lines of Java and Scala code. The approach could also be implemented with other query optimization frameworks or rule-based query rewriters such as Starburst [@starburst], ORCA [@GPORCA], and Cascades [@cascades]. Privacy Budget Management ------------------------- We have designed [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}to be flexible in its handling of the privacy budget, since best approach in a given setting is likely to depend on the domain and the kinds of queries posed. A complete study of approaches for managing the privacy budget is beyond the scope of this work, but we outline some possible strategies here. We describe the specific method used in our deployment in the next subsection. As described earlier, a simple approach to budget management involves standard composition. More sophisticated methods for privacy budget accounting include the *advanced composition* [@dwork2010boosting] and *parallel composition* [@dwork2006calibrating], both of which are directly applicable in our setting. For some mechanisms, the moments account [@abadi2016deep] could be used to further reduce privacy cost. #### Support for other mechanisms. Mechanisms themselves can also have a positive effect on the privacy budget, and many mechanisms have been designed to provide improved accuracy for a workload of similar queries. Many of these mechanisms are implemented in terms of lower-level mechanisms (such as the Laplace mechanism) that [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}already supports, and therefore could be easily integrated in [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}. The *sparse vector technique* [@dwork2009complexity] answers a sequence of queries, but only gives answers for queries in whose results lie above a given threshold. The technique is implemented in terms of the Laplace mechanism. The *Matrix Mechanism* [@li2010optimizing] and *MWEM* [@hardt2012simple] algorithms both answer a query workload by posing carefully chosen queries on the private database using a lower-level mechanism (e.g. the Laplace mechanism), then using the results to build a representation that can answer the queries in the workload. The *Exponential Mechanism* [@mcsherry2007mechanism] enforces differential privacy for queries that return categorical (rather than numeric) data, by picking from the possible outputs with probability generated from an analyst-provided *scoring function*. This technique can be be implemented as an intrinsically private query if the scoring function is given in SQL; the transformed query can run the function on each possible output and then pick from the possibilities according to the generated distribution. Deployment ---------- [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}is currently being deployed to enforce differential privacy for analysts that query customer data at [a large technology company]{}. The primary goals of this deployment are to protect the privacy of customers from insider attacks, and to ensure compliance with the requirements of Europe’s General Data Protection Regulation (GDPR) [@gdpr]. In the current deployment, [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}processes more than 10,000 queries per day. #### Data environment & architecture. The data environment into which [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}is deployed consists of several DBMSs (three primary databases, plus several more for specific applications), and a single central query interface through which all queries are submitted. The query interface is implemented as a microservice that performs query processing and then submits the query to the appropriate DBMS and returns the results. Our deployment involves a minimal wrapper around the [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}library to expose its rewriting functionality as a microservice. The only required change to the data environment was a single modification to the query interface, to submit queries to the [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}microservice for rewriting before execution. The wrapper around [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}also queries a policy microservice to determine the security and privacy policy for the user submitting the query. This policy informs which rewriter is used—by default, differential privacy is required, but for some privileged users performing specific business tasks, differential privacy is only used for older data. A major challenge of this deployment has been supporting the variety of SQL dialects used by the various DBMSs. The Calcite framework is intended to provide support for multiple dialects, but this support is incomplete and we have had to make changes to Calcite in order to support custom SQL dialects such as Vertica. #### Privacy budget. The privacy budget is managed by the microservice wrapper around [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}. The microservice maintains a small amount of state to keep track of the current cumulative privacy cost of all queries submitted so far, and updates this state when a new query is submitted. The current design of the [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}microservice maintains a single global budget, and uses advanced composition [@dwork2014algorithmic] to track the total budget used for the queries submitted so far. As we gain experience with the deployment, we are beginning to consider more sophisticated budget management approaches that take advantages of properties of the data and the query workload. For example, new data is added to the database continuously in this particular use case, so recent work leveraging the growth of the database to answer an unbounded number of queries [@cummings2018differential] may be directly applicable. #### Mechanism selection. Our deployment of [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}leverages a syntax-based selection procedure which aims to optimize for utility (low error). As we show in Section \[sec:selection\_experiment\], this approach performs well for this deployment. For different query workloads, other approaches may work significantly better, and [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}is designed to support extension to these cases. The syntax-based approach uses a set of rules that map SQL constructs supported by each mechanism to a heuristic scoring function indicating how likely queries using that construct will yield high utility. For example, Restricted sensitivity [@Blocki:2013:DPD:2422436.2422449] supports counting queries with joins, but does not support many-to-many joins. Elastic sensitivity [@allegro] supports a wider set of equijoins, including many-to-many joins, but generally provides slightly lower utility than Restricted sensitivity due to smoothing. Sample & aggregate [@smith2011privacy] does not support joins, but does support additional aggregation functions (including “average” and “median”). When a query is submitted, the mechanism selection engine analyzes the query to determine its syntactic properties including how many joins it has, of what type, and what aggregation functions it uses. It then applies the rules to determine which mechanisms can support the query and selects the mechanism with highest score. Note this process does not depend on the data and hence does not consume privacy budget. This approach represents a simple but effective strategy for automatic mechanism selection. In Section \[sec:selection\_experiment\], we demonstrate that our rules are effective for selecting the best mechanism on a real-world query workload. This approach is also easily extended when a new mechanism is added: the mechanism designer simply adds new rules for SQL constructs supported by the mechanism. Moreover, the scoring function can be tuned for other objectives, for example favoring mechanisms achieving low performance overhead rather than highest utility. Evaluation {#sec:evaluation} ========== This section reports results of the following experiments: - We report the percentage of queries that can be supported by each mechanism as intrinsically privacy queries using a corpus of real-world queries, demonstrating that a combination of the four evaluated mechanisms covers 93.9% of these queries. - We use [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}to conduct the first empirical study of several differential privacy mechanisms on a real-world SQL workload. We report the performance overhead and utility of each mechanism across its supported class of queries. - We demonstrate that our rule-based approach for automatic mechanism selection is effective at selecting the best mechanism for each input query. Using the simple set of rules presented earlier, our approach selects the optimal mechanism for nearly 90% of the queries in our corpus. #### Corpus. We use a corpus of [18,774]{}real-world queries containing all statistical queries executed by data analysts at [a large technology company]{} during October 2016. The corpus includes queries written for several use cases including fraud detection, marketing, business intelligence and general data exploration. It is therefore highly diverse and representative of SQL data analytics queries generally. The queries were executed on a database of data sampled from the production database. Mechanism Support for Queries {#sec:query-transf-succ} ----------------------------- ![Size and relationship of query sets supported by each evaluated mechanism.[]{data-label="fig:query_overlap"}](figures/query_overlap){width="45.00000%"} Each mechanism evaluated supports a different subset of queries. This is due to the unique limitations and supported constructs of each mechanism, as summarized in Section \[sec:ipq\]. We measured the percentage of queries from our corpus supported by each mechanism to assess that mechanism’s coverage on a real-world workload. Figure \[fig:query\_overlap\] depicts the relative size and overlap of each set of supported queries for the evaluated mechanisms. Elastic Sensitivity is the most general mechanism and can support 71.4% of the queries in our corpus, followed by Restricted Sensitivity (57.6%), WPINQ (30.3%) and Sample & Aggregate (45.4%). Elastic Sensitivity and Restricted Sensitivity support largely the same class of queries, and WPINQ supports a subset of the queries supported by these two mechanisms. In Section \[sec:disc-key-take\] we discuss limitations preventing the use of WPINQ for certain classes of queries supported by Elastic Sensitivity and Restricted Sensitivity. Sample & Aggregate supports some queries supported by other mechanisms (counting queries that do not use join), as well as a class of queries using statistical estimators (such as sum and average), that are not supported by the other mechanisms. In total, 93.9% of queries are supported by at least one of the four mechanisms. The results highlight a key advantage of our approach: different classes of queries can be simultaneously supported via selection of one or more specialized mechanisms. This ensures robust support across a wide range of general and specialized use cases, and allows incremental adoption of future state-of-the-art mechanisms. Performance Overhead {#sec:perf-overh-intr} -------------------- ------------------------ ---------- ------------------- ---------------------- **Primary cause** **Mean** **Median** **of overhead** Elastic Sensitivity   2.8   1.7 Random noise gen. Restricted Sensitivity   3.2   1.6 Random noise gen. WPINQ  50.9  21.6 Additional joins Sample & Aggregate 587 394 Grouping/aggregation ------------------------ ---------- ------------------- ---------------------- : Performance overhead of evaluated differential privacy mechanisms.[]{data-label="tbl:performance"} ![Performance overhead of differential privacy mechanisms by execution time of original query.[]{data-label="fig:scatter_plots"}](figures/performance_results.png){width="47.50000%"} We conduct a performance evaluation demonstrating the performance overhead of each mechanism when implemented as an intrinsically private query. #### Experiment Setup. We used a single HP Vertica 7.2.3 [@vertica] node containing 300 million records including trips, rider and driver information and other associated data stored across 8 tables. We submitted the queries locally and ran queries sequentially to avoid any effects from network latency and concurrent workloads. To establish a baseline we ran each original query 10 times and recorded the average after dropping the lowest and highest times to control for outliers. For every mechanism, we used [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}to transform each of the mechanism’s supported queries into an intrinsically private query. We executed each intrinsically private query 10 times and recorded the average execution time, again dropping the fastest and slowest times. We calculate the overhead for each query by comparing the average runtime of the intrinsically private query against its baseline.[^6] #### Results. The results are presented in Table \[tbl:performance\]. The average overhead and median overhead for Elastic Sensitivity are 2.82% and 1.7%, for Restricted Sensitivity these are 3.2% and 1.6%, for WPINQ 50.9% and 21.58% and for Sample & Aggregate 587% and 394%. Figure \[fig:scatter\_plots\] shows the distribution of overhead as a function of original query execution time. This distribution shows that the percentage overhead is highest when the original query was very fast (less than 100ms). This is because even a small incremental performance cost is fractionally larger for these queries. The values reported in Table \[tbl:performance\] are therefore a conservative estimate of the overhead apparent to the analyst. WPINQ and Sample & Aggregate significantly alter the way the query executes (see Section \[sec:intr-priv-quer-1\]) and these changes increase query execution time. In the case of WPINQ, the query transformation adds a new join to the query each time weights are rescaled (i.e. one new join for each join in the original query), and these new joins result in the additional overhead. Sample & Aggregate requires additional grouping and aggregation steps. We hypothesize that these transformations are difficult for the database to optimize during execution. Figure \[fig:scatter\_plots\] shows that, in both cases, the performance impact is amortized over higher query execution times, resulting in a lower relative overhead for more expensive queries. Utility of Selected Mechanisms {#sec:util-select-mech} ------------------------------ {width="99.00000%"} [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}enables the first empirical study of the *utility* of many differential privacy mechanisms on a real-world query workload. This experiment reveals innate trends of each mechanism on a common database and query workload. For each differential privacy mechanism, this experiment reports the relative magnitude of error added to results of its supported query set. We present the results as a function of query sample size, discussed below. #### Experiment Setup. We use the same setup described in the previous section to evaluate the utility of Elastic Sensitivity, Restricted Sensitivity, and Sample & Aggregate. As described in Section \[sec:weighted-pinq\], WPINQ’s output is a differentially private statistic used as input to a post-processing step, rather than an answer to the query posed, so we do not measure WPINQ’s utility. For each query, we set the privacy budget $\epsilon=0.1$ for all mechanisms. For Elastic Sensitivity, we set $\delta = n^{-\epsilon \ln n}$ (where $n$ is the database size), following Dwork and Lei [@dwork2009differential]. For Sample & Aggregate, we set the number of subsamples $\ell = n^{0.4}$, following Mohan et al. [@mohan2012gupt]. We ran each intrinsically private query 10 times on the database and report the median relative error across these executions. For each run we report the relative error as the percentage difference between the differentially private result and the original non-private result. Consistent with previous evaluations of differential privacy [@DBLP:conf/sigmod/HayMMCZ16] we report error as a proxy for utility since data analysts are primarily concerned with accuracy of results. If a query returns multiple rows (e.g., histogram queries) we calculate the mean error across all histogram bins. If the query returns multiple columns we treat each output column independently since noise is applied separately to every column. #### Query Sample Size. Our corpus includes queries covering a broad spectrum of use cases, from highly selective analytics (e.g., trips in San Francisco completed in the past hour) to statistics of large populations (e.g., all trips in the US). Differential privacy generally requires the addition of more noise to highly selective queries than to queries over large populations, since the influence of any individual’s data diminishes as population size increases. Consequently, a query’s selectivity is important for interpreting the relative error introduced by differential privacy. To measure the selectivity we calculate the *sample size* of every aggregation function in the original query, which represents the number of input records to which the function was applied. #### Results. Figures \[fig:scatter\_plots\] and \[fig:scatter\_plots\_saa\_function\] show the results of this experiment. All three mechanisms exhibit the expected inverse relationship between sample size and error; moreover, this trend is apparent for queries with and without joins. Where the other three mechanisms support only counting queries, Sample & Aggregate supports all statistical estimators. Figure \[fig:scatter\_plots\_saa\_function\] shows the utility results for Sample & Aggregate, highlighting the aggregation function used. These results indicate that Sample & Aggregate can provide high utility ($<$10% error) for each of its supported aggregation functions on approximately half of the queries. ![Utility of Sample & Aggregate, by aggregation function.[]{data-label="fig:scatter_plots_saa_function"}](figures/scatterplots_saa.png){width="47.00000%"} Automatic Mechanism Selection {#sec:selection_experiment} ----------------------------- We evaluated the effectiveness of the syntax-based automatic mechanism selection approach described in Section \[sec:implementation\].For each query in our corpus, this experiment compares the utility achieved by the mechanism selected by our rule-based approach to the best possible utility achievable by any mechanism implemented in [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}. #### Experiment Setup. We used the same corpus of queries and the same database of trips as in the other experiments. For each query, we ran all of the mechanisms that support the query and recorded the relative error (i.e. utility) of each one. We defined the *oracle utility* for each query to be the minimum error achieved by any of the four implemented mechanisms for that query. The oracle utility is intended to represent the utility that could be obtained if a perfect oracle for mechanism selection were available. We used our syntax-based mechanism selection method to select a single mechanism, and determined the utility of that mechanism. Finally, we computed the difference between the oracle utility and the utility achieved by our selected mechanism. ![Effectiveness of automatic mechanism selection.[]{data-label="fig:mechanism_selection"}](figures/mechanism_selection_bar.pdf){width="45.00000%"} #### Results. We present the results in Figure \[fig:mechanism\_selection\]. For 88% of the queries in our corpus, the automatic mechanism selection rules select the best mechanism, and therefore provide the same utility as the oracle utility. Of the remaining queries, the selected mechanism often provides nearly oracle utility: for 7% of queries, the selected mechanism is within 10% error of the oracle utility. The remaining queries (5%) represent opportunities for improving the approach—perhaps through the use of a prediction model trained on features of the query and data. Previous work [@DBLP:conf/sigmod/HayMMCZ16; @kotsogiannis2017pythia] uses such a machine learning-based approach; for range queries, these results suggest that a learning-based approach can be very effective, though the approach has not been evaluated on other types of queries. Discussion and Key Takeaways {#sec:disc-key-take} ---------------------------- #### Strengths & weaknesses of differential privacy. The mechanisms we studied generally worked best for statistical queries over large populations. None of the mechanisms was able to provide accurate results (e.g. within 1% error) for a significant number of queries over populations smaller than 1,000. These results confirm the existing wisdom that differential privacy is ill-suited for queries with small sample sizes. For large populations (e.g. more than 10,000), on the other hand, multiple mechanisms were able to provide accurate results. In addition, a large set of such queries exists in our corpus. #### Mechanism performance. Our performance evaluation highlights the variability in computation costs of differential privacy mechanisms. Approaches such as Elastic Sensitivity or Restricted Sensitivity incur little overhead, suggesting these mechanisms are ideal for performance critical applications such as real-time analytics. Given their higher performance cost, mechanisms such as WPINQ and Sample & Aggregate may be most appropriate for specialized applications where performance is less important than suitability of the mechanism for a particular problem domain. For example, WPINQ is the only evaluated mechanism that supports synthetic data generation, a task known to be highly computation-intensive. The performance of intrinsically private queries can depend on the database engine and transformations applied to the query. In this work we do not attempt to optimize the rewritten queries for performance. #### Unsupported queries. The current implementation of [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}applies a single mechanism to an entire input query. As a result, every aggregation function used by the input query must be supported by the selected mechanism, or the transformation fails. For example, consider a query with joins that outputs both a count and an average. Neither Elastic Sensitivity (which does not support average) nor Sample & Aggregate (which does not support join) can fully support this query. This issue disproportionately affects WPINQ, since our implementation of WPINQ does not support `COUNT(DISTINCT ...)` queries. It is not obvious how to do so: the weights of any record in the database only reflect the number of duplicate rows until a join is performed (and weights are re-scaled). It is possible to leverage multiple mechanisms in a single intrinsically private query by treating each output column separately. This approach would provide improved support for queries like the example above, which use several different aggregation functions. We leave such an extension to future work. Related Work {#sec:related-work} ============ #### Differential Privacy. Differential privacy was originally proposed by Dwork [@dworkdifferential2006; @dwork2006calibrating; @dwork2008differential]. The reference by Dwork and Roth [@dwork2014algorithmic] provides an overview of the field. Much recent work has focused on task-specific mechanisms for graph analysis [@hay2009accurate; @sala2011sharing; @karwa2011private; @kasiviswanathan2013analyzing; @Chen:2013:RMT:2463676.2465304], range queries [@hardt2012simple; @li2014data; @li2010optimizing; @li2015matrix; @zhang2014towards; @xiao2012dpcube; @cormode2012differentially; @qardaji2013differentially; @acs2012differentially; @xu2013differentially], and analysis of data streams [@dwork2010differential; @shi2011privacy]. As described in Section \[sec:disc-key-take\], such mechanisms are complementary to our approach, and could be implemented on top of [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}to provide more efficient use of the privacy budget. #### Differential Privacy Systems. A number of systems for enforcing differential privacy have been developed. PINQ [@mcsherry2009privacy] supports a LINQ-based query language, and implements the Laplace mechanism with a measure of global sensitivity. Weighted PINQ [@proserpio2014calibrating] extends PINQ to weighted datasets, and implements a specialized mechanism for that setting. Airavat [@roy2010airavat] enforces differential privacy for MapReduce programs using the Laplace mechanism. Fuzz [@gaboardi2013linear; @haeberlen2011differential] enforces differential privacy for functional programs, using the Laplace mechanism in an approach similar to PINQ. DJoin [@narayan2012djoin] enforces differential privacy for queries over distributed datasets. Due to the additional restrictions associated with this setting, DJoin requires the use of special cryptographic functions during query execution so is incompatible with existing databases. GUPT [@mohan2012gupt] implements the Sample & Aggregate framework for Python programs. In contrast to our approach, each of these systems supports only a single mechanism and, with the exception of Airavat, each implements a custom database engine. #### Security & Privacy via Query Rewriting. To the best of our knowledge, ours is the first work on using query transformations to implement differential privacy mechanisms. However, this approach has been used in previous work to implement access control. Stonebreaker and Wong [@stonebraker1974access] presented the first approach. Barker and Rosenthal [@barker2002flexible] extended the approach to role-based access control by first constructing a view that encodes the access control policy, then rewriting input queries to add `WHERE` clauses that query the view. Byun and Li [@byun2008purpose] use a similar approach to enforce purpose-based access control: purposes are attached to data in the database, then queries are modified to enforce purpose restrictions drawn from a policy. Conclusion ========== This paper presents [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}, a system with a novel architecture for enforcing differential privacy for SQL queries on an unmodified database. [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}works by automatically transforming input queries into intrinsically private queries. We have described the deployment of [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}at [a large technology company]{} to provide differential privacy, where it processes more than 10,000 queries per day. We make [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}available as open source [[@chorusdownloadanon]]{}. We used [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}to perform the first empirical evaluation of various mechanisms on real-world queries and data. The results demonstrate that our approach supports 93.9% of statistical queries in our corpus, integrates with a production DBMS without any modifications, and scales to hundreds of millions of records. Acknowledgments {#acknowledgments .unnumbered} =============== The authors would like to thank Om Thakkar for his helpful comments. This work was supported by the Center for Long-Term Cybersecurity, and DARPA & SPAWAR under contract N66001-15-C-4066. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes not withstanding any copyright notation thereon. The views, opinions, and/or findings expressed are those of the author(s) and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government. [^1]: Work done while at the University of California, Berkeley [^2]: We omit the name of the company to ensure anonymity for the submission. [^3]: For simplicity we consider approaches where [[<span style="font-variant:small-caps;">[Chorus]{}</span>]{}]{}stores the budget directly. However, our query rewriting approach could allow the DBMS to assist with budget accounting, for example by storing $\epsilon$ values in a separate table and referencing and updating the values within the rewritten query. [^4]: Smooth sensitivity guarantees $\epsilon,\delta$-differential privacy, and incorporates the setting of $\delta$ into the smoothed sensitivity value. [^5]: Subsampling changes the semantics of joins, since join keys in separate partitions will be prevented from matching. It also changes the semantics of aggregations in subqueries, since the aggregation is computed within a single subsample. [^6]: Transformation time is negligible and therefore not included in the overhead calculation. The transformation time averages a few milliseconds, compared with an average query execution time of 1.5 seconds.

--- abstract: | We present an efficient algorithm to compute Euler characteristic curves of gray scale images of arbitrary dimension. In various applications the Euler characteristic curve is used as a descriptor of an image. Our algorithm is the first streaming algorithm for Euler characteristic curves. The usage of streaming removes the necessity to store the entire image in RAM. Experiments show that our implementation handles terabyte scale images on commodity hardware. Due to lock-free parallelism, it scales well with the number of processor cores. Our software—CHUNKYEuler—is available as open source on Bitbucket. Additionally, we put the concept of the Euler characteristic curve in the wider context of computational topology. In particular, we explain the connection with persistence diagrams. author: - Teresa Heiss - Hubert Wagner bibliography: - 'Literature.bib' title: Streaming Algorithm for Euler Characteristic Curves of Multidimensional Images --- Introduction {#sec:introduction} ============ The Euler characteristic curve is a powerful tool in image processing [@Gonzalez_Wintz_1977]. It has been used in a variety of fields including astrophysics[^1] [@gott; @rhoads1994genus], medical image analysis [@segonne2007geometrically; @odgaard1993quantification], and image processing in general [@Richardson201499; @Snidaro20031533]. Its wide applicability stems from simplicity and efficient computability. However, with the advances in image acquisition technology, there is need to handle very large images. For example the state-of-the-art micro-CT scanner Skyscan 1272 creates images of size $14450\times14450\times2600$ with 14-bit precision. Therefore, a single scan yields more than half a trillion voxels. It is also possible to combine multiple scans of the same object which further multiplies the size of data. As loading the resulting multi-terabyte image into RAM of a commodity computer is infeasible, a streaming approach is needed. We present the first streaming algorithm for computing Euler characteristic curves.[^2] Our algorithm divides a multidimensional image into chunks that fit into RAM, calculates the Euler characteristic curves for each chunk separately and merges them in the end. Since these chunks can be made arbitrarily small, commodity hardware can be used to compute Euler characteristic curves of arbitrarily large images. The fact that the chunks are not dependent on each other makes lock-free parallelism possible. For defining the Euler characteristic curve we first need to explain what the Euler characteristic is. There are two ways to define the Euler characteristic and the Euler-Poincaré formula states that they are both equivalent. For discrete two-dimensional surfaces, like a triangulation of a sphere or a torus, the definitions are: first, the number of vertices minus the number of edges plus the number of faces; second, the number of connected components minus the number of tunnels plus the number of voids. Originally the Euler characteristic was defined for the surface of a convex polyhedron where it always equals two. To see this, consider that such a surface consists of one connected component, no tunnels and one void. The equivalence between these two definitions seems to be the reason for the usefulness of the Euler characteristic: It captures global topological structures—like holes—although it can be computed locally—by adding up vertices, edges and faces. The Euler characteristic curve of an image is the vector of Euler characteristics of consecutive thresholded images. We illustrate this for the example image of a bone[^3] in Fig. \[fig:exampleImageSubfigure\] with values ranging from 0 (black) to 255 (white). The Euler characteristic curve of this image maps each $t \in \{0,1,...,255\}$ to the Euler characteristic of the set of pixels with gray value smaller or equal to $t$. Figure \[fig:explanation\_ECC\] illustrates this process. This concept can be extended to more general settings, e.g., images with floating point gray values (see Sec. \[sec:algorithm\]).  \ Theoretical Background {#sec:background} ====================== We give basic definitions needed in Sec. \[sec:algorithm\] using the language borrowed from computational topology [@kaczynski2004computational]. With this we can provide precise definitions in arbitrary dimension and explain the connection between the Euler characteristic curve and other topological descriptors. #### Cubical Cell. A $k$-dimensional **cubical cell** (short: **cell**) $c$ of embedding dimension $d$ is defined as the Cartesian product of intervals and singletons: $$c:=I_1 \times I_2 \times \dots \times I_d$$ where exactly $k$ of the sets $(I_i)_{i\in\{1, 2, \dots,d\}}$ are intervals of the form $I_i=[a_i,a_i+1]$ with integers $a_i\in{\mathbb{Z}}$ and the remaining $d-k$ sets are singletons $I_i=\{b_i\}$ with integers $b_i\in{\mathbb{Z}}$. A zero-dimensional cell is called a **vertex**, a one-dimensional cell an **edge**, a two-dimensional cell a **square**, a three-dimensional cell a **cube**. #### Face. A cell $c_1$ of embedding dimension $d$ is called a **face** of a cell $c_2$ of embedding dimension $d$ if $c_1$ is a subset of $c_2$. #### Cubical Complex. A $p$-dimensional **cubical complex** (short: **complex**) of embedding dimension $d$ is a finite set of cubical cells of embedding dimension $d$ such that 1. The faces of each cell are also elements of the complex 2. The intersection of any two cells is also an element of the complex[^4] where $p$ is the highest dimension of all cells in the complex. The complexes that appear in our algorithm always fulfill $p=d$. Figure \[fig:example\_complex\] shows an example of a cubical complex. ![Example of a two-dimensional cubical complex of embedding dimension two with one square, eight edges and eight vertices.[]{data-label="fig:example_complex"}](example_complex.pdf){width="25.00000%"} #### Filtration of Cubical Complexes. A sequence of complexes $K_1,K_2,\dots,K_m$ is called a **filtration** if the complexes are monotonically increasing: $K_1\subseteq K_2\subseteq\dots\subseteq K_m$. #### Sublevel Set Filtration. Let $K$ be a cubical complex. A cell that is not a face of any other cell than itself is called a **maximal cell**. Let $f:M\to{\mathbb{R}}$ be a function, where $M$ is the set of maximal cells. This function $f$ can be extended to a function $\tilde{f}:K\to{\mathbb{R}}$ defined on all cells: $$\begin{aligned} \tilde{f}\colon K &\to {\mathbb{R}}\\ c &\mapsto \tilde{f}(c):= \begin{cases} f(c) & \mbox{if } c\in M\\ \displaystyle\min_{\substack{m\in M\\ c\subseteq m}}f(m) & \mbox{otherwise.} \end{cases}\end{aligned}$$ For each $t \in {\mathbb{R}}$ the **sublevel set** $\tilde{f}^{-1}\left((-\infty,t]\right)$ of this extended function is the set of cells that are a face of at least one maximal cell with $f$-value smaller or equal to $t$. As $K$ consists of only a finite number of cells, $\tilde{f}$ can only have a finite number of different function values $\{t_1, t_2,\dots,t_m\}$. The sublevel sets $\tilde{f}^{-1}\left((-\infty,t_1]\right),\dots,\tilde{f}^{-1}\left((-\infty,t_m]\right)$ form a filtration of cubical complexes—the **sublevel set filtration** induced by the function $f$. To see this, notice that the definition of $\tilde{f}$ implies that for each $t\in{\mathbb{R}}$ the sublevel set $\tilde{f}^{-1}\left((-\infty,t]\right)$ is a cubical complex: all faces of a cell $c$ belong to the same sublevel set as $c$. Furthermore the sublevel sets are monotonically increasing $\tilde{f}^{-1}\left((-\infty,t_1]\right) \subseteq \tilde{f}^{-1}\left((-\infty,t_2]\right) \subseteq \dots \subseteq \tilde{f}^{-1}\left((-\infty,t_m]\right)$. Therefore, the sublevel sets form a filtration. #### Consecutive Thresholded Images as Sublevel Set Filtrations. A $d$-dimensional gray scale image with $n_1\times n_2 \times \dots \times n_d$ voxels[^5] can be interpreted as a $d$-dimensional cubical complex $K$ of embedding dimension $d$ with a function $f$ on its maximal cells: for each voxel index $(i_1,\dots,i_d)$, $i_1 \in \{1,\dots,n_1\},\dots,i_d\in \{1,\dots,n_d\}$ the corresponding voxel position is represented by the $d$-dimensional cell $c_{i_1,\dots,i_d}$ of embedding dimension $d$: $$c_{i_1,\dots,i_d} := \left[i_1 - 1;i_1\right]\times\left[i_2 - 1;i_2\right]\times\dots\times\left[i_d - 1;i_d\right] \enspace .$$ The cubical complex $K$ is defined as the set of all these cells $c_{i_1,\dots,i_d}$ along with all their faces. The function $f$ maps each maximal cell $c_{i_1,\dots,i_d}$ to the gray value of the voxel with index $(i_1,\dots,i_d)$. The sublevel set filtration[^6] induced by the function $f$ is formed by consecutive thresholdings of the image.[^7] #### Euler Characteristic. The **Euler characteristic** $\chi$ of a $p$-dimensional complex $K$ of embedding dimension $d$ is defined as $$\chi(K):=\sum_{k=0}^{p}(-1)^k n_k$$ where $n_k$ is the number of $k$-dimensional cells in $K$. The Euler-Poincaré formula states $$\chi(K)=\sum_{k=0}^{p}(-1)^k n_k =\sum_{k=0}^{d-1}(-1)^k\beta_k$$ where $\beta_k$ is the $k$th Betti number (the number of $k$-dimensional holes). For a formal definition of cubical homology and the involved Betti numbers, see [@kaczynski2004computational]. In three-dimensional space, $\beta_0$ is the number of connected components, $\beta_1$ is the number of tunnels and $\beta_2$ is the number of voids. #### Euler Characteristic Curve. The **Euler characteristic curve** $e$ of a filtration of cubical complexes $K_1\subseteq K_2\subseteq\dots\subseteq K_m$ is the vector $$e=\left(\chi(K_1),\chi(K_2),\dots,\chi(K_m)\right)\enspace .$$ This vector can also be interpreted as a function $$\begin{aligned} e\colon \{1,2,\dots,m\} &\to {\mathbb{Z}}\\ t &\mapsto \chi(K_t) \enspace ,\end{aligned}$$ which is used to visualize the Euler characteristic curve as in Fig. \[fig:eulercurve\]. #### Euler Characteristic Curve of an Image. We already saw that for an arbitrary gray scale image the sequence of consecutive thresholded images is a filtration—the sublevel set filtration. The Euler characteristic curve of this filtration is the **Euler characteristic curve of an image**. #### Connection to Other Topological Descriptors. We want to put the above considerations in the wider context of computational topology. Two popular topological descriptors of a filtration are Betti curves and persistence diagrams [@edelsbrunner2010computational], which both capture information about holes at different thresholds. For each hole in the image the persistence diagram tracks the first and last threshold at which the hole occurs. The $k$th Betti curve, which counts the $k$-dimensional holes at each threshold, is easily computable from the persistence diagram. Furthermore, the alternating sum of the Betti curves yields the Euler characteristic curve. Therefore, the Euler characteristic curve summarizes a persistence diagram, but it can be computed locally. The usefulness of the Euler characteristic curve suggests that the two richer descriptors may also be useful in image processing. However, large images are out of reach of the currently available persistence diagram software. For now, the Euler characteristic curve remains the only feasible option. Algorithm {#sec:algorithm} ========= The input for our algorithm is a gray scale image of arbitrary dimension $d$ with $n$ voxels. The output is the Euler characteristic curve of this image, as defined in Sec. \[sec:background\]. ### Range of Values. In Sec. \[sec:background\] the function $f$ maps to ${\mathbb{R}}$. However, in practice the gray values of an image are in a predefined range, usually $\{0,1,\dots,255\}$ or $\{0,1,\dots,65535\}$. If the range contains negative numbers, it can be shifted so that it starts from zero. For this reason we focus on ranges of the form $\{0,1,\dots,m-1\}$ with a positive integer $m$. A version of our algorithm that can handle floating point values will be discussed at the end of this section. ### Tracking the Changes. It is suboptimal to compute the Euler characteristic for each threshold separately, as already noted in [@Snidaro20031533]. To avoid redundant computations we track the changes between consecutive thresholds. More precisely, we determine how each voxel contributes to the change in Euler characteristic. Therefore we first compute a vector of changes in Euler characteristic (VCEC) $\left(a_0, a_1, \dots, a_{m-1}\right)$ whose entries $a_t := \chi \left( \tilde{f}^{-1}\left(\left(-\infty,t\right]\right) \right) - \chi \left( \tilde{f}^{-1}\left(\left(-\infty,t-1\right]\right) \right)$ are the difference between the Euler characteristics of two consecutive thresholded images.[^8] The Euler characteristic curve is then: $$\begin{aligned} &\left(a_0, a_0+a_1, \dots, \sum_{t=0}^{m-1}a_t\right) =\\ =&\left( \chi \left( \tilde{f}^{-1}\left(\left(-\infty,0\right]\right) \right), \chi \left( \tilde{f}^{-1}\left(\left(-\infty,1\right]\right) \right), \dots, \chi \left( \tilde{f}^{-1}\left(\left(-\infty,m-1\right]\right) \right) \right) \enspace .\end{aligned}$$ When changing from one thresholded image $\tilde{f}^{-1}\left(\left(-\infty,t-1\right]\right)$ to the next $\tilde{f}^{-1}\left(\left(-\infty,t\right]\right)$, all voxels with gray value $t$ are included, along with all their faces that have not already been included at a previous threshold. We say that these new faces are introduced by these new voxels. More precisely, a face $c$ of a voxel $v$ is **introduced** by $v$ if all other voxels $w$ that have $c$ as a face fulfill one of the following two conditions: 1. $f(w)>f(v)$ 2. $f(w)=f(v)$ and $w\succeq v$, \[enum:order\] where $\succeq$ is any total order of the voxel positions, e.g., the lexicographical order[^9]. The output of our algorithm is independent of the chosen total order. However, it is necessary to require $w\succeq v$ in condition \[enum:order\] to ensure that each face is introduced by a unique voxel. Now we can decompose the cubical complex of the input image into blocks such that each block contributes to exactly one change of threshold. A block consists of a voxel together with all the faces it introduces. Which faces are introduced by a certain voxel is determined only by the gray values of the voxel’s $3^d-1$ neighbors. We exploit this locality in the design of our parallel streaming algorithm. Figure \[fig:blocks\] shows the decomposition of a two-dimensional example image into blocks. ### Storage. For the computations we store only the gray values of the voxels. The geometric information and the adjacency relations between cells are implicit in the voxel grid and calculated locally whenever needed. Similarly the function $\tilde{f}$ and the block decomposition it induces are never explicitly stored. Apart from storing the result vector, the memory overhead is essentially zero. ### Streaming. If the entire image does not fit into RAM, we divide it into chunks that fit into RAM. In our implementation we use a simple strategy: an image of size $n_1 \times n_2 \times \dots \times n_d$ is divided into $c$ chunks of size $\frac{n_1}{c} \times n_2 \times \dots \times n_d$ (see Fig. \[fig:cores\], left). As these correspond to contiguous memory regions, streaming the chunks from a single input file is easy. We then separately compute the VCEC for each chunk either sequentially or in parallel. ![Whenever a worker thread is free it loads one chunk along with a one voxel thick collar into its space in RAM and computes the VCEC of this chunk.[]{data-label="fig:cores"}](threads.pdf){width="86.10000%"} ### Parallel Computations. For parallelism we use a thread pool. Each worker thread is assigned memory for a single chunk and one initially empty VCEC vector. One task is to read a chunk from disk, update the worker thread’s VCEC vector by the VCEC of this chunk and discard the chunk. At any given time at most $w$ chunks reside in RAM, where $w$ is the number of worker threads. Because different worker threads work with disjoint memory regions we achieve lock-free parallelism. The underlying data structure for the collection of VCECs is a vector of vectors, called `euler_changes`. ### Processing one Chunk. Along with the chunk we read a one voxel thick collar surrounding it. This way we have access to all neighbors of the voxels in the chunk (see Fig. \[fig:cores\]). With this information we compute the VCEC of this chunk as specified in Algorithm \[alg\]. One chunk of an image along with a one voxel thick collar surrounding it and `current_thread`, the index of the worker thread processing this chunk. An updated version of the vector `euler_changes`\[`current_thread`\] which is the VCEC of all chunks this thread has processed. change $=0$ change$++$ change$--$ \[line\] remove chunk and collar from RAM ### Post-Processing. In the end, when all chunks have been processed, a single thread sums up the VCECs yielding the VCEC of the whole image, which is a vector $\left(a_0, a_1, \dots, a_{m-1}\right)$. The Euler characteristic curve is then computed as $\left(a_0, a_0+a_1, \dots, \sum_{t=0}^{m-1}a_t\right)$. ### Analysis. To analyze the complexity we remind that $d$ is the dimension of the image, $n$ is the number of voxels, $m$ is the number of gray values[^10] and $w$ is the number of worker threads. We introduce a new variable $s$, the number of voxels per chunk including the collar. Assuming perfect parallelization, the worst case running time of our algorithm is ${\mathcal{O}}(\frac{3^dn}{w}+mw)$ because for each voxel we visit all its neighbors and sequentially post-process the VCECs. We analyze the practical scaling behavior in Sec. \[sec:experiments\]. As the dimension $d$ is usually small (mostly 2 or 3), the exponential term is usually not a problem in practice. For each worker thread, we need $s$ integers of $\log_2(m)$ bits to store the gray values of a chunk. Additionally, the `euler_changes` data structure consists of $wm$ 64-bit integers. Therefore, the total storage is $\log_2(m)ws+64wm+{\mathcal{O}}(1)$ bits in RAM. By decreasing the chunk size $s$, the dominant part, $\log_2(m)ws$, can be made arbitrarily small. Because of this, our algorithm works for arbitrarily large images on commodity hardware. ### Other Ranges of Gray Values. When the range of gray values is not of the form $\{0,1,\dots,m-1\}$—for example for floating point values—one option is to use a hash map to store the `euler_changes`. If the number of different input values approaches $n$, the output size dominates the overall storage and the advantage of a streaming approach disappears. In this situation, it is preferable to transform (i.e., round, scale, shift) the input values to obtain a range of the form $\{0,1,\dots,m-1\}$. Running our standard algorithm on the transformed input yields the same result as transforming the domain of the Euler characteristic curve computed for the original data. Experiments {#sec:experiments} =========== We implemented the above algorithmic scheme in C++14. We made experiments on two different machines: a laptop with Intel core i5-5200U CPU with two physical cores clocked at 2.2GHz with 8GB of RAM and a workstation with Intel Xeon E5645 CPU with 12 physical cores clocked at 2.4GHz and 72GB of RAM. Table \[tab:constanttime\] shows the running time and memory usage for different 3D input images ran on the laptop. We use images from a standard data set[^11], see [@wagner2011efficient; @delgado2015skeletonization]. The names’ suffixes distinguish between 8- and 16-bit precision images. The last column shows that the running time is linear in $n$ and does not depend on the content of the image. name size million voxels memory\[MB\] time\[s\] time\[s\]/million voxels ------------ ------------------------------ ---------------- -------------- ----------- -------------------------- prone16 512$\times$512$\times$463 121.4 70.4 24.7 0.20 xmastree16 512$\times$499$\times$512 130.8 72.9 26.7 0.20 vertebra16 512$\times$512$\times$512 134.2 74 28.4 0.21 random8 512$\times$512$\times$512 134.2 51 28.9 0.22 random8 1024$\times$1024$\times$1024 1073.7 93.1 236.7 0.22 random8 2048$\times$2048$\times$2048 8589.9 261.6 1767.1 0.21 : Running time and memory usage.[]{data-label="tab:constanttime"} Due to the above, we show the scaling behavior using a single image. The computations were performed on the workstation for a $512\times 499\times 512$ image with 16-bit precision. We used from 1 to 12 threads taking the mean running time and standard deviation across 20 runs. Figure \[fig:parallelizationplot\] shows the speed-up gained using $w$ threads instead of one. In particular using 10 threads is 7.1 times faster than using a single thread. ![Scaling with number of threads[]{data-label="fig:parallelizationplot"}](scaling_graph.pdf){width="84.00000%"} Table \[tab:huge\] shows that the Euler characteristic curve of terabyte scale images can be computed on a single computer with limited memory. Memory usage could be further decreased by changing the chunking scheme. However, the experiments demonstrate that—for the foreseeable future—our implementation is a reasonable trade-off between performance and simplicity. size threads memory --------------------------------- --------- ----------- -------- 4096$\times$4096$\times$4096 12 1.8 hours 1.93GB 14500$\times$14500$\times$2600 24 9 hours 4.5GB 14500$\times$14500$\times$2600 12 13 hours 2.27GB 10000$\times$10000$\times$10000 8 32 hours 3.98GB : Running time and memory usage for computations on large 3D images, performed on the 12-core workstation. The voxel values are generated independently from a uniform random distribution in the range $\{0,1,\dots,250\}$. As we already showed, using other images of the same size will exhibit almost identical performance.[]{data-label="tab:huge"} Related Work and Discussion =========================== We review the work related to computing the Euler characteristic (curve) of images. We embed this in the context of computing other topological descriptors of images, particularly persistence diagrams. Algorithms related to the Euler characteristic received a lot of attention in image processing [@dyer1980computing; @saha1995new; @sossa1996computation; @ziou2002generating], starting from the seminal work by Gray [@gray1971local]. Many modern implementations aim at real-time processing of small 2D images [@Snidaro20031533]. Our goal is different, namely handling large multidimensional images. Computing other topological descriptors of images is a more recent advancement [@verri1993use; @pikaz1997efficient; @kaczynski2004computational; @wagner2011efficient; @robins2011theory; @gunther2012efficient; @delgado2015skeletonization], which has however not entered mainstream image processing. Specialized methods for computing persistence diagrams handle 3D images up to $500^3$ voxels [@delgado2015skeletonization; @gunther2012efficient] on what we consider commodity hardware. The main limitation is the storage of the entire image in memory. There exist distributed implementations [@bauer2014distributed], which alleviate the storage problem per machine, but are not specialized to image data, resulting in large overall memory overhead. The largest reported computed instances are in the range of $1000^3$ on 32 server nodes. Overall it is clear that a specialized, streaming approach is necessary for handling large images. We offer a robust implementation for the Euler characteristic curve, with possible future extensions to other topological descriptors. We expect that these more complex topological descriptors will be computable for this terabyte scale data in the future but currently we are limited to using Euler characteristic curves. Let us discuss the properties of our algorithmic scheme and mention limitations of our current implementation. The advantages of our algorithm are: - It can handle arbitrarily large images on commodity hardware. - It can handle images of arbitrary dimension. - Linear running time. - Predictable running time and memory usage. - Due to lock-free parallelism, running time scales well with increasing number of threads. - Our algorithm can be easily adapted to a massively-distributed setting using a map-reduce framework [@dean2008mapreduce]. Some limitations of the current implementation: - It uses $(3^d-1)$-connectivity. For other types (e.g., 6-connectivity for 2D images), modifications on the algorithm can be made. - For simplicity we use slices of the image as chunks. For very large images even a one voxel thick slice may not fit into memory. - For technical reasons we surround each chunk with a second collar of voxels with value $\infty$. Effectively a five voxel thick slice has to fit into memory, which may become a problem for very large images. - To include the value $\infty$ we may need a larger data type. For example if the input contains all values from 0 to 255 we use a 16-bit data type to store the original values along with an extra value for infinity. Despite the limitations, our implementation is robust and can handle even the largest data produced by state-of-the-art image acquisition technology. We released this software under the name CHUNKYEuler as open source: <https://bitbucket.org/hubwag/chunkyeuler/src>. [^1]: The astrophysics community refers to the Euler characteristic curve as the genus. [^2]: We review related work at the end of the paper. [^3]: We thank Reinhold Erben and Stephan Handschuh from Vetmeduni Vienna for providing micro-CT scans of rat vertebrae. [^4]: The second condition is implied by the first condition since we allow only consecutive integers as interval endpoints in the definition of cells. [^5]: Throughout this paper we use “voxel” as multidimensional generalization of “pixel”. [^6]: Another interpretation of voxel data is via the dual complex (voxels become vertices) using the lower star filtration. The way we use appears more natural in image processing context. The two approaches yield similar but not necessarily identical Euler characteristic curves. [^7]: Defining cells as products of *closed* intervals implies $(3^d-1)$-connectivity for the voxels of the thresholded images. This corresponds to 8-connectivity for 2D images. [^8]: where $\chi \left( \tilde{f}^{-1}\left(\left(-\infty,-1\right]\right) \right)=\chi(\emptyset)=0$. [^9]: In lexicographical order a voxel at position $(i_1,\dots,i_d)$ succeeds a voxel at position $(j_1,\dots,j_d)$ if $i_k>j_k$ for the first $k$ where $i_k$ and $j_k$ differ. [^10]: The input size is $\log_2(m)n$. [^11]: Most of the images are available at [www.byclb.com/TR/Muhendislik/Dataset.aspx](www.byclb.com/TR/Muhendislik/Dataset.aspx)

--- abstract: 'The scaling behaviour of the diffraction intensity near the origin is investigated for (partially) ordered systems, with an emphasis on illustrative, rigorous results. This is an established method to detect and quantify the fluctuation behaviour known under the term hyperuniformity. Here, we consider one-dimensional systems with pure point, singular continuous and absolutely continuous diffraction spectra, which include perfectly ordered cut and project and inflation point sets as well as systems with stochastic disorder.' address: - 'Fakultät für Mathematik, Universität Bielefeld, Postfach 100131, 33501 Bielefeld, Germany' - 'School of Mathematics and Statistics, The Open University,Walton Hall, Milton Keynes MK7 6AA, United Kingdom' author: - Michael Baake - Uwe Grimm title: | Scaling of diffraction intensities near the origin:\ Some rigorous results --- Introduction ============ While the concept of *order* is intuitive, it is surprisingly challenging to ‘measure’ the order in a system in a way that leads to a meaningful classification of different manifestations of order in Nature. Some of the most prominent measures available are based on ideas from statistical physics and crystallography, such as *entropy* which is closely related to disorder, or *diffraction* which is the main tool for determining the structure of solids. Essentially, diffraction is the Fourier transform of the pair correlation function, and hence quantifies the order in the two-point correlations of the structure. Following the discovery of quasicrystals in the early 1980s, a proper mathematical treatment of the diffraction of aperiodically ordered structures was required; see [@TAO; @Hof] and references therein for background and details. It turns out that there is a close connection between diffraction and what is known as the *dynamical spectrum* in mathematics [@BL]. The latter is the spectrum of the corresponding Koopman operator and is an important concept in ergodic theory. Aperiodically ordered structures, in particular those constructed by inflation rules, provide interesting examples of systems that exhibit a scaling (self-similarity) type of order, which differs substantially from the translational order found in periodic structures, such as conventional crystals. Starting from the idea to use the degree of ‘(hyper)uniformity’ in density fluctuations in many-particle systems [@TS] to characterise their order, the scaling behaviour of the diffraction near the origin has emerged as a measure that captures the variance of the long-distance correlations. Recently, a number of conjectures on the scaling behaviour of the diffraction of aperiodically ordered structures were made [@Josh1; @Josh2], reformulating and extending earlier, partly heuristic, results by Luck [@Luck] from this perspective; see also [@Aubry; @GL]. The purpose of this article is to link the recent interest in these questions with some of the known techniques and results from rigorous diffraction theory, as started by Hof in [@Hof] and later developed by many people; see [@TAO] and references therein for a systematic account. Our approach will make substantial use of the exact renormalisation relations for primitive inflation rules [@BFGR; @BG15; @BGM; @NM; @BaGriMa], which will allow us to establish the scaling behaviour rigorously. This is in some contrast to classic methods of finite size scaling [@VR88; @VR03; @VR05; @VR10], where such a behaviour is extrapolated and only asymptotically true. The paper is organised as follows. We begin by recalling some background material on diffraction, the projection formalism, inflation rules and Lyapunov exponents in Section \[sec:prelim\]. Then, in Section \[sec:pp\], we discuss systems with pure point spectrum, starting from the paradigmatic Fibonacci chain, which we treat in two different ways. We also discuss various generalisations, including noble means inflations and a limit-periodic system. After a brief look into substitutions with more than two letters, we close with an instructive example of number-theoretic origin, which can also be understood as a projection set. In Section \[sec:ac\], we summarise the situation for systems with absolutely continuous spectrum via a comparative exposition of random and deterministic cases. In particular, we discuss the Poisson process in comparison to the Rudin–Shapiro sequence, the classical random matrix ensembles, as well as the Markov lattice gas and binary random tilings. Finally, Section \[sec:sc\] deals with systems with singular continuous spectrum, of which the classic –Morse sequence is a paradigm with a decay behaviour faster than any power law. We then embed this into the family of generalised –Morse sequences. Preliminaries and general methods {#sec:prelim} ================================= Our approach to the scaling behaviour of the diffraction measure near the origin requires a number of different methods. In this section, we summarise key results and provide references for background and further details. Diffraction and scaling ----------------------- Throughout this article, we use the notation and results from [@TAO] on diffraction theory, as well as some more advanced results on the Fourier transform of measures from [@MS; @TAO2]. Various definitions are discussed there, and we use standard results from these sources without further reference. Note that the term ‘measure’ in this context refers to general (complex) Radon measures in the mathematical sense. They can be identified with the continuous linear functionals on the space of compactly supported continuous functions on ${\mathbb{R}}^d$. In particular, given a (usually translation bounded) measure $\omega$, we assume that an averaging sequence ${\mathcal{A}}$ of van Hove type is specified, and the autocorrelation $\gamma = \gamma^{}_{\omega} = \omega \circledast \widetilde{\omega}$ is given as the Eberlein (or volume-averaged) convolution along ${\mathcal{A}}$. The measure $\gamma$ is positive definite, and hence Fourier transformable as a measure. The Fourier transform of $\gamma$ is the *diffraction measure* $\widehat{\gamma}$, which is a positive measure. This is the measure-theoretic formulation of the *structure factor* from physics and crystallography, which is better suited for rigorous results. In particular, this approach defines the different spectral components by means of the Lebesgue decomposition $$\widehat{\gamma} \, = \, \widehat{\gamma}^{}_{\mathsf{pp}} + \widehat{\gamma}^{}_{\mathsf{sc}} + \widehat{\gamma}^{}_{\mathsf{ac}}$$ of $\widehat{\gamma}$ into its pure point, singular continuous and absolutely continuous parts; see [@TAO Sec. 8.5.2] for details. In what follows, we only consider the one-dimensional case. For the investigation of scaling properties, we follow the existing literature and define $$\label{eq:Z-def} Z (k) \, {\mathrel{\mathop:}=}\, \widehat{\gamma} \bigl( (0,k]\bigr) ,$$ which is a modified version of the distribution function of the diffraction measure. Due to the reflection symmetry of $\widehat{\gamma}$ with respect to the origin, this quantity can also be expressed as $$Z (k) \, = \, {\frac{\raisebox{-2pt}{$1$}} {\raisebox{0.5pt}{$2$}}} \Bigl(\widehat{\gamma} \bigl( [-k,k]\bigr) - \widehat{\gamma} \bigl( \{0\}\bigr) \Bigr).$$ Sometimes, it is natural to replace $Z(k)$ by $Z(k) / I(0)$, where $I(0) = \widehat{\gamma} \bigl( \{ 0 \} \bigr)$ is the intensity of the central diffraction (or Bragg) peak. Note that this just amounts to a different normalisation. It will always be clear from the context when we do so. This normalisation has no influence on the scaling behaviour of $Z (k)$ as $k\,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0$. The interest in the scaling near the origin is based on the intuition that the small-$k$ behaviour of the diffraction measure probes the long-wavelength fluctuations in the structure, which is related to the variance in the distribution of patches, and hence can serve as an indicator for the degree of uniformity of the structure [@TS]. Clearly, any periodic structure leads to $Z(k) = 0$ for all sufficiently small $k$, so that the main interest is focused on non-periodic systems, both ordered and disordered. Projection formalism -------------------- One way to produce well-ordered aperiodic systems is based on the (partial) projection of higher-dimensional (periodic) lattices. Such *cut and project sets* or *model sets* can be viewed as a generalised variant of the notion of a quasiperiodic function. We can only present a brief summary here, for details we refer to [@TAO Ch. 7]. The general setting for a model set in *physical* (direct) space ${\mathbb{R}}^{d}$ is encoded in the *cut and project scheme* (CPS) $({\mathbb{R}}^{d},H,{\mathcal{L}})$, $$\label{eq:cps} \renewcommand{\arraystretch}{1.2}\begin{array}{r@{}ccccc@{}l} & {\mathbb{R}}^{d} & \xleftarrow{\,\;\;\pi\;\;\,} & {\mathbb{R}}^{d} \, \times \, H & \xrightarrow{\;\pi^{}_{\mathrm{int}\;}} & H & \\ & \cup & & \cup & & \cup & \hspace*{-2ex} \raisebox{1pt}{\text{\footnotesize dense}} \\ & \pi({\mathcal{L}}) & \xleftarrow{\; 1-1 \;} & {\mathcal{L}}& \xrightarrow{\; \hphantom{1-1} \;} & \pi^{}_{\mathrm{int}}({\mathcal{L}}) & \\ & \| & & & & \| & \\ & L & \multicolumn{3}{c}{\xrightarrow{\qquad\quad\quad \;\; \;\star\; \;\; \quad\quad\qquad}} & {L_{}}^{\star} & \\ \end{array}\renewcommand{\arraystretch}{1}$$ where the *internal* space $H$ is a locally compact Abelian group (in many examples, $H$ will turn out to be another Euclidean space, so $H={\mathbb{R}}^{m}$), ${\mathcal{L}}$ is a lattice (co-compact discrete subgroup) in ${\mathbb{R}}^{d}\times H$, and where $\pi$ and $\pi^{}_{\mathrm{int}}$ denote the natural projections onto the physical and internal spaces. The assumption that $L=\pi({\mathcal{L}})\subset{\mathbb{R}}^{d}$ is a bijective image of ${\mathcal{L}}$ in physical space guarantees that the $\star\,$-map $x\mapsto x^{\star}$ is well defined on $L$. A *model set* for a given CPS is then a set of the form $$\label{eq:ms} {\varLambda}\, = \, {\mbox{\Large $\curlywedge$}}(W) \, = \, \bigl\{ x\in L : x^{\star} \in W \bigr\} ,$$ where the domain $W\subseteq H$ (called the *window* or *acceptance domain*) is a relatively compact set with non-empty interior. These conditions on the window guarantee that the model set ${\varLambda}\subseteq L$ is both uniformly discrete and relatively dense, so a Delone set in ${\mathbb{R}}^{d}$. If the window is sufficiently ‘nice’ (for instance, when $W$ is compact with boundary of measure $0$), the diffraction measure $\widehat{\gamma}$ of the associated Dirac comb $$\delta^{}_{\!{\varLambda}}\, {\mathrel{\mathop:}=}\sum_{x\in{\varLambda}} \delta^{}_{x}$$ is a pure point measure; see [@TAO Ch. 9] for details. Note that the dynamical spectrum is pure point as well; see [@BL] and references therein. \[ex:Fibo\] Consider a CPS with $d=1$ and $H={\mathbb{R}}$, and the planar lattice $${\mathcal{L}}\, = \, \bigg\langle\! \binom{1}{1}, \binom{\tau}{1-\tau} \!\bigg\rangle_{\!\!{\mathbb{Z}{\hspace{0.5pt}}}} \subset\, {\mathbb{R}}^{2},$$ where $\tau=\bigl(1+\sqrt{5}\, \bigr)/2$ is the golden ratio. The projection of ${\mathcal{L}}$ to physical space is the ${\mathbb{Z}{\hspace{0.5pt}}}$-module $L={\mathbb{Z}{\hspace{0.5pt}}}[\tau]= \{m+n\tau\mid m,n\in{\mathbb{Z}{\hspace{0.5pt}}}\}$, and similarly $L^{\star} = {\mathbb{Z}{\hspace{0.5pt}}}[\tau]$. The $\star$-map acts as algebraic conjugation $\sqrt{5}\mapsto -\sqrt{5}$, so that $(m+n\tau)^{\star}=m+n-n\tau$. The *Fibonacci model set* is ${\varLambda}={\mbox{\Large $\curlywedge$}}(W)$ with window $W = (-1,\tau-1]$. As indicated in Figure \[fig:goldproj\], the two different point types (left endpoints of long and short intervals, respectively) in the Fibonacci model set are obtained as model sets for the two smaller windows $W^{}_{\! L} = (-1,\tau-2]$ and $W^{}_{\! S} = (\tau-2,\tau-1]$, with $W=W^{}_{\! L}\operatorname{\dot\cup}W^{}_{\! S}$.[$\Diamond$]{} ![\[fig:goldproj\]Projection description of the Fibonacci model set from Example \[ex:Fibo\]. The horizontal line represents the physical space, the vertical line the internal space. The black points form the lattice ${\mathcal{L}}$. Points inside the strip ${\mathbb{R}}\times W$ are projected to physical space to produce the model set ${\varLambda}$.](goldproj.eps){width="80.00000%"} Inflation rules and exact renormalisation {#subsec:renorm} ----------------------------------------- Another important construction of aperiodic sequences and tilings is based on substitution and inflation rules; see [@TAO Chs. 4–5] for background. The idea is perhaps most easily phrased in terms of tilings. Starting from a (finite) set of prototiles (equivalence classes of tiles under translation), an *inflation rule* consists of a rescaling (more generally, an expansive linear transformation) of the prototiles and a subsequent dissection of the rescaled tiles into prototiles of the original shape and size. Iterating such an inflation rule produces tilings of space, which generically will not be periodic. Let us illustrate this with the example of the Fibonacci model set from above, which can also be constructed by an inflation rule (note that this is not generally the case: Cut and project tilings generically do *not* possess an inflation symmetry, and inflation tilings generally do not allow an embedding into a higher-dimensional lattice with a bounded window). \[ex:fiboinf\] Consider two prototiles (intervals) in ${\mathbb{R}}$, a long interval $a$ of length $\tau$ and a short interval $b$ of length $1$. The (geometric) Fibonacci inflation rule with inflation factor $\tau$ is $$\includegraphics[width=0.6\textwidth]{fiboinfl.eps}$$ Starting from an initial configuration of, say, two adjacent large tiles with common vertex at the origin, and iterating this rule produces a tiling of ${\mathbb{R}}$ which is a fixed point under the square of the inflation rule. Denoting the set of left endpoints of intervals of type $a$ by ${\varLambda}_{a}$ and of type $b$ by ${\varLambda}_{b}$, we have ${\varLambda}_{a,b}\subset {\mathbb{Z}{\hspace{0.5pt}}}[\tau]$ by construction, which are model sets in the description of Example \[ex:Fibo\]; see [@TAO Ex. 7.3] for details. [$\Diamond$]{} The inherent recursive structure of the inflation approach not only leads to recursive relations for the point sets constructed in this way, but also ensures the existence of a set of renormalisation equations for their two-point correlation functions; see [@BG15] and references therein for details. Here, we illustrate this for the Fibonacci case. \[ex:fiborec\] Consider the point set ${\varLambda}={\varLambda}_{a}\operatorname{\dot\cup}{\varLambda}_{b}$ constructed by the inflation of Example \[ex:fiboinf\]. Define the *pair correlation functions* (or coefficients) $\nu^{}_{i j} (z)$ with $i,j \in \{ a,b\}$ as the *relative* frequency (frequency per point of ${\varLambda}$) of two points in ${\varLambda}$ at distance $z$, subject to the condition that the left point is of type $i$ and the right point of type $j$. The pair correlation functions exist (as a consequence of unique ergodicity) and satisfy the relations $\nu^{}_{i j} (z)=0$ for any $z\not\in {\varLambda}-{\varLambda}$, $\nu^{}_{i j} (0)=0$ for $i\neq j$, and $\nu^{}_{i j} (-z) = \nu^{}_{j i} (z)$ for any $z$. The recognisability of our inflation rule then implies the following set of exact renormalisation equations [@BG15; @NM; @BGM], $$\label{eq:Fibo-rec} \begin{split} \nu^{}_{aa} (z) \, & = \, \tfrac{1}{\tau} \bigl( \nu^{}_{aa} (\tfrac{z}{\tau}) + \nu^{}_{ab} (\tfrac{z}{\tau}) +\nu^{}_{ba} (\tfrac{z}{\tau}) + \nu^{}_{bb} (\tfrac{z}{\tau}) \bigr) , \\ \nu^{}_{ab} (z) \, & = \, \tfrac{1}{\tau} \bigl( \nu^{}_{aa} (\tfrac{z}{\tau} - 1) + \nu^{}_{ba} (\tfrac{z}{\tau} - 1) \bigr), \\ \nu^{}_{ba} (z) \, & = \, \tfrac{1}{\tau} \bigl( \nu^{}_{aa} (\tfrac{z}{\tau} + 1) + \nu^{}_{ab} (\tfrac{z}{\tau} + 1) \bigr), \\ \nu^{}_{bb} (z) \, & = \, \tfrac{1}{\tau} \bigl( \nu^{}_{aa} (\tfrac{z}{\tau}) \bigr), \end{split}$$ where $z\in {\mathbb{Z}{\hspace{0.5pt}}}[\tau]$ and $\nu^{}_{ij} (z) = 0$ whenever $z\not\in {\varLambda}_{j} - {\varLambda}_{i}$. This set of equations has a unique solution if one of the single-letter frequencies $\nu^{}_{aa} (0), \nu^{}_{bb} (0)$ is given; see [@BG15] for details and further examples. The autocorrelation measure for the point set ${\varLambda}$ is of the form $$\gamma^{}_{{\varLambda}} \, = \sum_{z\in{\varLambda}- {\varLambda}} \eta (z)\, \delta_{z}$$ with autocorrelation coefficients $$\eta(z) \, {\mathrel{\mathop:}=}\, \lim_{R\to\infty} \frac{\operatorname{card}\bigl( {\varLambda}^{}_{R} \cap (z+{\varLambda}^{}_{R}) \bigr)}{2R} ,$$ where ${\varLambda}^{}_{R} {\mathrel{\mathop:}=}{\varLambda}\cap [-R,R]$. Again, these coefficients exist by unique ergodicity. The autocorrelation coefficients can be expressed in terms of the pair correlation functions as $$\label{eq:rel-nu} \frac{ \eta (z)}{\operatorname{dens}({\varLambda})} \, = \, \nu^{}_{aa} (z) + \nu^{}_{ab} (z) + \nu^{}_{ba} (z) + \nu^{}_{bb} (z) .$$ By taking Fourier transforms, this allows us to use the renormalisation approach to analyse the diffraction of ${\varLambda}$; see [@BFGR; @BaGriMa; @BGM] for details. [$\Diamond$]{} The approach of Example \[ex:fiborec\] can be applied to inflation tilings in quite a general setting, as it only requires recognisability (not necessarily local) of the inflation rule. Studying the corresponding diffraction spectrum and, in particular, its spectral components, naturally leads to the investigation of matrix iterations and their asymptotic behaviour; see [@BFGR; @BaGriMa; @BGM] for recent work in this direction, as well as [@BuSol; @BuSol2] for related work in the more general setting of matrix Riesz products for substitution systems with arbitrary choices of length scales. Lyapunov exponents for (effective) single matrix iterations {#subsec:iter} ----------------------------------------------------------- If $M \in \operatorname{Mat}(d, {\mathbb{C}})$ is a fixed matrix, the asymptotic behaviour of $ \| M^n v \|$ as $n\to\infty$ with $v\in {\mathbb{C}}^d$ can be understood in terms of the Lyapunov exponents of this iteration, via determining $$\lim_{n\to\infty} {\frac{\raisebox{-2pt}{$1$}} {\raisebox{0.5pt}{$n$}}} \log \| M^{n} v\| {\hspace{0.5pt}};$$ see [@BP; @Viana] for background. In this simple single matrix case, the *Lyapunov spectrum* of $M$ is $$\sigma^{}_{\mathrm{L}} (M) \, = \, \{ \log {\hspace{0.5pt}}\lvert \mu \rvert : \mu \in \sigma (M) \} {\hspace{0.5pt}},$$ where $\sigma (M)$ denotes the spectrum of $M$; compare [@BP]. By an expansion in the principal vectors and eigenvectors of $M$, one can also extract the explicit scaling behaviour as $n\to\infty$, for any given $v$. Slightly more interesting and relevant to us is the following situation. Consider a matrix family $B (k)$ for $k \geqslant 0$ that is smooth (in fact, analytic) in $k$, with $\lim_{k \searrow 0} B (k) = M$. Then, we are interested in the asymptotic behaviour of $$\| B(\lambda^{-n} k) B(\lambda^{-n+1} k) \cdots B(\lambda^{-1} k){\hspace{0.5pt}}v \|$$ for fixed $v$ as $n\to\infty$, which means that we consider a specific *matrix cocycle* of $k \mapsto k/\lambda$ on ${\mathbb{R}}$, with $\lambda > 1$; see [@BP; @Viana] for background. Due to the relation with $M=B(0)$, the Lyapunov spectrum of the cocycle agrees with that of $M$, so $$\label{eq:lspec} \lim_{n\to\infty} {\frac{\raisebox{-2pt}{$1$}} {\raisebox{0.5pt}{$n$}}} \, \log \| B(\lambda^{-n} k) B(\lambda^{-n+1} k) \cdots B(\lambda^{-1} k) {\hspace{0.5pt}}v \| \, \in \, \sigma^{}_{\mathrm{L}} (M) {\hspace{0.5pt}}.$$ This behaviour emerges from the observation that the cocycle, with increasing length but fixed $k$, more and more looks like a multiplication by powers of $M$. Aperiodic systems with pure point spectrum {#sec:pp} ========================================== Let us set the scene by considering a paradigmatic example in some detail, which is actually approachable by two different methods, namely by the projection formalism and by the renormalisation approach. Common to all examples in this section is the fact that $Z(k)$ is not a continuous function of $k$, which requires some extra care. The Fibonacci chain and related systems --------------------------------------- Consider the primitive substitution $\varrho^{}_{\mathrm{F}}$ on the binary alphabet ${\mathcal{A}}= \{ a,b\}$, as defined by $a \mapsto ab$ and $b \mapsto a$, which is the symbolic substitution that underlies Example \[ex:Fibo\]. The Abelianisation leads to the corresponding *substitution matrix* $$M^{}_{\mathrm{F}} \, = \, \begin{pmatrix}1 & 1 \\ 1 & 0 \end{pmatrix}.$$ Its Perron–Frobenius (PF) eigenvalue is the golden mean, $\tau$. The corresponding right and left eigenvectors code the relative frequencies of letters and the natural interval lengths for the geometric representation, respectively. There are two bi-infinite fixed points for $\varrho^{2}_{\mathrm{F}}$, where the legal seeds are $a|a$ and $b|a$, with the vertical line denoting the location of the origin; see [@TAO Sec. 4] for details and background. We choose the one with seed $a|a$ and let ${\mathbb{X}}$ define the *symbolic* hull generated by it via an orbit closure under the shift action, which is actually independent of the choice we made. Now, we turn this ${\mathbb{X}}$ into a *tiling system* by fixing natural tile (or interval) lengths according to the left PF eigenvector, namely $\tau$ for $a$ and $1$ for $b$. The left endpoints of the tiles, for the tiling that emerges from our chosen fixed point with seed $a|a$, form a Delone set with distances $\tau$ and $1$, and the orbit closure (in the local topology) under the (continuous) translation action by ${\mathbb{R}}$ gives the *geometric* hull ${\mathbb{Y}}$ we are interested in. Let ${\varLambda}\in {\mathbb{Y}}$ be arbitrary, and consider the Dirac comb $\delta^{}_{\! {\varLambda}}$, which is a translation-bounded measure with well-defined autocorrelation $\gamma$ and diffraction $\widehat{\gamma}$. The latter does not depend on the choice of ${\varLambda}$, and is given by $$\label{eq:F-diff} \widehat{\gamma} \, = \sum_{k\in L^{\circledast}} I(k) \, \delta^{}_{k} \quad \text{with} \quad I(k) \, = \, \operatorname{dens}({\varLambda})^2 \operatorname{sinc}\bigl( \pi \tau k^{\star} \bigr)^2 ,$$ where $L^{\circledast} = {\mathbb{Z}{\hspace{0.5pt}}}[\tau]/\text{\small $\sqrt{5}$}\subset {\mathbb{Q}}(\text{\small $\sqrt{5}$}\, )$ is the *Fourier module*, $k^{\star}$ is obtained from $k$ by replacing $\sqrt{5}$ with $-\sqrt{5}$, and $\operatorname{sinc}(x) = \frac{\sin (x)}{x}$; see [@TAO Sec. 9.4.1] for a derivation. Note that $L^{\circledast}$ is also the dynamical spectrum (in additive notation) of the dynamical system $({\mathbb{Y}},{\mathbb{R}})$, which is pure point with all eigenfunctions being continuous [@Boris; @Daniel]. Behind this description lies the fact that the Delone set of our special fixed point is the *regular model set* of Example \[ex:Fibo\], with CPS $({\mathbb{R}}, {\mathbb{R}}, {\mathcal{L}})$ and a half-open window of length $\tau$, where our choice of the lattice is ${\mathcal{L}}= \{ (x, x^{\star}) : x \in {\mathbb{Z}{\hspace{0.5pt}}}[\tau] \}$, which is the Minkowski embedding of ${\mathbb{Z}{\hspace{0.5pt}}}[\tau]$ as a planar lattice in ${\mathbb{R}}\times {\mathbb{R}}\simeq {\mathbb{R}}^2$. The $\star$-map then is algebraic conjugation in the quadratic field ${\mathbb{Q}}( \mbox{\small $\sqrt{5}$} \, )$, as induced by $\sqrt{5} \mapsto - \sqrt{5}$. More generally, if we use the same CPS with an interval of length $s$ as window, the diffraction measure of the resulting point set is of the same form as in Eq. , with the intensity now being $I (k) = I (0) {\hspace{0.5pt}}\operatorname{sinc}(\pi s {\hspace{0.5pt}}k^{\star})^2$. If one considers a sequence of positions $ ( k /\tau^{\ell} )$ with $k \in L^{\circledast}$, it is clear that $I ( k / \tau^{\ell} ) = {\mathcal{O}}\bigl( \tau^{-2\ell} \bigr)$ as $\ell\to\infty$, because $\operatorname{sinc}(x) = {\mathcal{O}}(x^{-1} )$ as $x \to\infty$. As we shall discuss later in Remark \[rem:generic\], this leads to an asymptotic behaviour of the form $Z(k) = {\mathcal{O}}(k^2)$ as $k\,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0$. In what follows, we shall see that the decay of $Z(k)$ from Eq.  can be faster, provided $s$ takes special values. This result is implicit in [@GL] and was recently re-derived, by slightly different methods, in [@Josh1]. \[prop:module-scale\] Let ${\varLambda}= {\mbox{\Large $\curlywedge$}}(W)$ be a regular model set in the CPS $({\mathbb{R}}, {\mathbb{R}}, {\mathcal{L}})$ of the Fibonacci chain, where the window $W\subset {\mathbb{R}}$ is an interval of length $s\in {\mathbb{Z}{\hspace{0.5pt}}}[\tau]$. Then, with $L^{\circledast} = {\mathbb{Z}{\hspace{0.5pt}}}[\tau]/\mbox{\small $\sqrt{5}$}$ as above, the diffraction measure of $\delta^{}_{\! {\varLambda}}$ is given by $\widehat{\gamma} = I(0) \sum_{k\in L^{\circledast} } \operatorname{sinc}\bigl(\pi s {\hspace{0.5pt}}k^{\star} \bigr)^2$, and the intensities along a sequence $( k/\tau^{\ell})^{}_{\ell\in{\mathbb{N}}_0}$, for any $0\ne k\in L^{\circledast}$, decay like $\tau^{-4\ell}$ as $\ell\to\infty$. The statement on the diffraction measure, $\widehat{\gamma}$, is a direct consequence of the results in [@TAO Sec. 9.4.1]. It does not depend on the position of the window, or on whether the interval is closed, open or half-open. Now, $s\in{\mathbb{Z}{\hspace{0.5pt}}}[\tau]$ means $s=a+b \tau$ with $a,b\in{\mathbb{Z}{\hspace{0.5pt}}}$. Let $0\ne k\in L^{\circledast}$, so $k = \kappa/\mbox{\small $\sqrt{5}$}$ with $\kappa = m + n \tau$ for some $m,n \in {\mathbb{Z}{\hspace{0.5pt}}}$, excluding $m=n=0$. Applying the $\star$-map then gives $$I \Bigl( {\frac{\raisebox{-2pt}{$k$}} {\raisebox{0.5pt}{$\tau^{\ell}$}}}\Bigr) \, = \, \operatorname{dens}({\varLambda})^2 \, \operatorname{sinc}\Bigl( {\frac{\raisebox{-2pt}{$\pi \tau^{\ell} s {\hspace{0.5pt}}\kappa^{\star}$}} {\raisebox{0.5pt}{$\mbox{\small $\sqrt{5}$}$}}} \Bigr)^2 ,$$ with $\ell\in{\mathbb{N}}_0$. To continue, we recall the relation $$\label{eq:sin-red} \sin (m \pi + x) \, = \, (-1)^m \sin (x) \qquad \text{for $m\in{\mathbb{Z}{\hspace{0.5pt}}}$ and $x\in{\mathbb{R}}$}.$$ Moreover, with $f_n$ denoting the Fibonacci numbers as defined by $f^{}_{0} = 0$, $f^{}_{1} = 1$ together with the recursion $f^{}_{n+1} = f^{}_{n} + f^{}_{n-1}$, we will need the well-known formula $$\label{eq:fib-form} f^{}_n \, = \, {\frac{\raisebox{-2pt}{$1$}} {\raisebox{0.5pt}{$\mbox{\small $\sqrt{5}$}$}}} \, \Bigl( \tau^n - \bigl( -1/\tau\bigr)^n \Bigr) ,$$ which holds for all $n\in{\mathbb{Z}{\hspace{0.5pt}}}$. Then, as $\ell\to\infty$, we get $$\sin \Bigl( {\frac{\raisebox{-2pt}{$\pi \tau^{\ell} s {\hspace{0.5pt}}\kappa^{\star}$}} {\raisebox{0.5pt}{$\mbox{\small $\sqrt{5}$}$}}} \Bigr)^2 \, = \: \sin \Bigl( {\frac{\raisebox{-2pt}{$\pi {\hspace{0.5pt}}\lvert s {\hspace{0.5pt}}\kappa^{\star} \rvert$}} {\raisebox{0.5pt}{$\mbox{\small $ \sqrt{5}$}$}}} \, \tau^{-\ell} \Bigr)^2 \, = \, {\frac{\raisebox{-2pt}{$\pi^2 \bigl( s {\hspace{0.5pt}}\kappa^{\star} \bigr)^2$}} {\raisebox{0.5pt}{$5$}}} \, \tau^{-2\ell} \, + {\mathcal{O}}\bigl( \tau^{-6 \ell}\bigr) ,$$ where the first step follows from using Eq.  to replace $\tau^{\ell}/\mbox{\small $\sqrt{5}$}$ and then reducing the argument via Eq. , which is possible because all Fibonacci numbers are integers. The second step uses the Taylor approximation $\sin (x) = x + {\mathcal{O}}(x^3)$ for small $x$. Inserting this expression into the formula for the intensity gives $$\label{eq:fiboint} I \Bigl( {\frac{\raisebox{-2pt}{$k$}} {\raisebox{0.5pt}{$\tau^{\ell}$}}} \Bigr) \, = \: c(s,k) \, \tau^{-4\ell} + {\mathcal{O}}\bigl( \tau^{-8 \ell} \bigr) \qquad \text{as $\ell\to\infty$},$$ where the constant derives from the previous calculation. Analysing the steps of the last proof, one finds $$c (s,k) \, = \, \operatorname{dens}({\varLambda})^2 \pi^2 s^2 (k^{\star})^2,$$ where $k^{\star}$ lies in the window and is thus bounded. So, for any fixed $s \in {\mathbb{Z}{\hspace{0.5pt}}}[\tau]$, we have $c (s,k) = {\mathcal{O}}(1)$, which allows us to sum over the inflation series of peaks as follows. Consider $$\varSigma (k) \, {\mathrel{\mathop:}=}\sum_{\ell=0}^{\infty} I \Bigl( {\frac{\raisebox{-2pt}{$k$}} {\raisebox{0.5pt}{$\tau^{\ell}$}}} \Bigr) , $$ where the previous estimates lead to the asymptotic behaviour $$\label{eq:sig-scal} \varSigma \Bigl( {\frac{\raisebox{-2pt}{$k$}} {\raisebox{0.5pt}{$\tau^{\ell}$}}} \Bigr) \, \sim \: \widetilde{c} (k) {\hspace{0.5pt}}\varSigma (k) \, \tau^{-4\ell} \qquad \text{as $\ell \to \infty$} {\hspace{0.5pt}},$$ with $\widetilde{c} (k) = {\mathcal{O}}(1)$. Now, observe that the diffraction measure satisfies $$Z(k) \, = \, \widehat{\gamma} \bigl( (0,k {\hspace{0.5pt}}] \bigr) \, = \, \sum_{\kappa \in L^{\circledast} \cap ({\hspace{-0.5pt}}\frac{k}{\tau}, k {\hspace{0.5pt}}]} \varSigma (\kappa) {\hspace{0.5pt}},$$ which then, via , implies the asymptotic behaviour $$Z \Bigl( {\frac{\raisebox{-2pt}{$k$}} {\raisebox{0.5pt}{$\tau^{\ell}$}}} \Bigr) \, \asymp \, Z(k) \, \tau^{-4\ell} .$$ The upper bound is a consequence of the summability of the geometric series in conjunction with the boundedness of $\widetilde{c}(k)$, while the lower bound follows from the existence of at least one series of non-trivial peaks with a scaling according to and the behaviour of $\operatorname{sinc}(x)$ near $x=0$. Put together, we have the following result. \[thm:fibo\] Under the assumptions of Proposition [\[prop:module-scale\]]{.nodecor}, one has $$Z (k) \, \asymp \, k^4 \qquad \text{as $k\,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0$}.$$ In particular, $Z(k)={\mathcal{O}}(k^4)$. \[rem:infl\] The cases with $s\in{\mathbb{Z}{\hspace{0.5pt}}}[\tau]$ treated in Proposition \[prop:module-scale\] are connected in the sense that the minimal components[^1] of the corresponding hulls are all mutually locally derivable (MLD) to each other, so they all have a local inflation/deflation symmetry in the sense of [@TAO]. Within this family, there is the Fibonacci chain, which can be constructed from a self-similar tiling inflation rule. It is thus natural to assume that the particularly rapid decay of $Z (k)$ in this case can also be traced back to the inflation nature, as in Section \[sec:noble\] below. [$\Diamond$]{} \[rem:generic\] For general $s$, as mentioned previously, one only has $I(k/\tau^{\ell})={\mathcal{O}}\bigl(\tau^{-2\ell}\bigr)$. Since the summation argument remains unchanged, one then only gets $Z(k)={\mathcal{O}}(k^2)$ as $k\,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0$. Unlike the situation treated in Theorem \[thm:fibo\], it is not clear when one finds $Z(k)\asymp k^2$. Consequently, it is an open question whether intermediate decay exponents are possible, for instance as a result of Diophantine approximation properties. [$\Diamond$]{} Noble means inflations {#sec:noble} ---------------------- These are generalisations of the Fibonacci case, with substitution matrices $M_p=\left(\begin{smallmatrix}p & 1 \\ 1 & 0 \end{smallmatrix}\right)$ for integer $p\in{\mathbb{N}}$, where $p=1$ is the Fibonacci example. One possible substitution rule [@TAO Rem. 4.7] is $a\mapsto a^pb$, $b\mapsto a$, but the position of the letter $b$ in the image of $a$ does not matter, as all choices produce equivalent rules, by an application of [@TAO Prop. 4.6]. All noble means substitutions result in Sturmian sequences. The inflation factor $\lambda_p$ is the Perron–Frobenius eigenvalue of $M_p$, so $$\lambda_{p}\, =\, {\frac{\raisebox{-2pt}{$1$}} {\raisebox{0.5pt}{$2$}}}\bigl(p+\sqrt{p^2+4}\,\bigr) \, =\, [p;p,p,p,\ldots]{\hspace{0.5pt}},$$ with $p=1$ giving the Fibonacci case (golden mean) discussed above. The corresponding inflation rule works with lengths $\lambda_p$ and $1$ for the intervals of type $a$ and $b$, respectively. Then, the arguments used above generalise directly to the noble means inflation case, which possess a model set description where both the physical and the internal space are one-dimensional, resulting in the same scaling exponents for the entire family. In fact, the noble means inflations can more easily be analysed using the renormalisation approach [@BFGR; @BGM] mentioned in Section \[subsec:renorm\]. The two eigenvalues of the inflation matrix $M_{p}$ are $\lambda_{p}=\frac{1}{2}\big(p+\sqrt{p^2+4}\,\bigr)$ and $-1/\lambda_{p}$. Consequently, as explained in Section \[subsec:iter\], the Lyapunov spectrum of the matrix iteration is $\sigma^{}_{\mathrm{L}} (M_p) = \{\log(\lambda_p) , -\log(\lambda_p) \}$. To continue, we need the Fourier–Bohr (FB) coefficients for the $a$- and $b$-positions separately, which we denote by $A_a (k)$ and $A_b (k)$, respectively. Note that one can obtain the FB coefficients of any weighted version of the chain by an appropriate superposition, whose absolute square is the corresponding diffraction intensity at $k$. Let $A(k) = \bigl(A_a (k), A_b (k) \bigr)^T$ be the vector of FB coefficients at a given $k$. Then, the renormalisation approach gives the relation $$\big\| A(\lambda^{-n}_{p} k)\big\| \, =\, \lambda^{-n}_{p} {\hspace{0.5pt}}\big\| B(\lambda^{-n}_{p}k) B(\lambda^{-n+1}_{p}k)\dots B(\lambda^{-1}_{p}k) {\hspace{0.5pt}}A(k)\big\| .$$ Due to the additional prefactor, by taking logarithms, the Lyapunov exponents are shifted by $-\log(\lambda_{p})$, and thus are $\{0 , -2\log(\lambda_p) \}$. The exponent $0$ is not the relevant one, because it would correspond to intensities that do not decay as we approach $k=0$, which contradicts the measure property of $\widehat{{\hspace{0.5pt}}\gamma^{}_{p}{\hspace{0.5pt}}}$. Consequently, the scaling of the amplitudes is governed by the Lyapunov exponent $-2\log(\lambda_{p})$, and the corresponding diffraction intensities scale as $$I(\lambda^{-\ell}_{p}k) \, \sim \, c(k){\hspace{0.5pt}}\lambda^{-4\ell}_{p} I(k){\hspace{0.5pt}},$$ which is the generalisation of the behaviour we observed for the Fibonacci case in Proposition \[prop:module-scale\]. On the basis of the same argument for the boundedness of the $c(k)$, we can sum the contributions and conclude as follows. For the noble means inflations with $p\in{\mathbb{N}}$, one has the asymptotic behaviour $$Z_p (k) \, = \, \widehat{{\hspace{0.5pt}}\gamma^{}_{p}{\hspace{0.5pt}}} \bigl( (0,k] \bigr) \, \asymp \, k^4 \quad \text{as $k\,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0$.}$$ In particular, $Z_p(k)={\mathcal{O}}(k^4)$. Note that this result relies on the inflation structure, and thus only applies to inflation-symmetric cases, in line with Remark \[rem:infl\]. Let us continue with further examples of inflation systems, where we state the results in a less formal manner. Limit-periodic systems and beyond --------------------------------- Consider the two-letter substitution rule $$a\mapsto ab\, , \quad b\mapsto aa \, ,$$ with substitution matrix $\left(\begin{smallmatrix}1 & 2\\1 & 0 \end{smallmatrix}\right)$. The latter is diagonalisable and has eigenvalues $2$ and $-1$. This substitution rule defines the *period doubling* system, which is a paradigm of a Toeplitz-type system with limit-periodic structure; see [@TAO Sec. 4.5] for background and [@TAO Sec. 9.4.4] for a detailed discussion of its pure point diffraction measure. As the period doubling system arises from a primitive substitution rule, which in this case is identical to the inflation rule, the renormalisation approach of Section \[subsec:renorm\] applies. Here, the exponents (taking into account the additional factor $2^{-n}$ in the amplitude renormalisation equation) are $\{0,-\log(2)\}$, of which only the negative one is relevant (for the same reason as above). Consequently, the diffraction intensities scale as $$I(2^{-\ell}k) \, \sim \, c(k){\hspace{0.5pt}}4^{-\ell} I(k) {\hspace{0.5pt}},$$ and the scaling behaviour for the period doubling system is $Z(k)\asymp k^2$ as $k \,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0$. Again, this follows via geometric series summations in conjunction with the observation that the $c(k)$ are bounded (for instance, via the existing model set description; see [@TAO Ex. 7.4]). The same type of analysis also works for limit-quasiperiodic systems. As an example, consider the two-letter substitution rule $a\mapsto aab$, $b\mapsto abab$ from [@BMS], in its proper geometric realisation. The eigenvalues of the substitution matrix are $\lambda_{\pm}=2\pm\sqrt{2}$. The standard renormalisation analysis from above for the FB coefficient now results in $$\big\| A(k/\lambda_{+})\big\| \, \sim \, {\frac{\raisebox{-2pt}{$2-\sqrt{2}$}} {\raisebox{0.5pt}{$2+\sqrt{2}$}}}\, \big\| A(k)\big\|$$ for $k\,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0$, and thus in the scaling behaviour $I(\lambda^{-n}_{+}{\hspace{0.5pt}}k) \sim c(k){\hspace{0.5pt}}\lambda_{+}^{-2n\tilde{\alpha}}$ with $$\tilde{\alpha} \, = \, 2 - \frac{\log{\hspace{0.5pt}}\lvert\det(M)\rvert}{\log(\lambda_{+})} \, = \, 2 \,\frac{\log\bigl(1+\sqrt{2}\,\bigr)} {\log\bigl(2+\sqrt{2}\,\bigr)} \, \approx\, 1.436{\hspace{0.5pt}},$$ where the $c(k)$ are once again bounded. Consequently, one finds $Z(k)\asymp k^{2\tilde{\alpha}}$ as $k\,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0$. Our parameter $\tilde{\alpha}$ is related to the hyperuniformity parameter $\alpha$ from [@Josh1] by $1+\alpha=2\tilde{\alpha}$. More generally, this formula applies to primitive binary inflation systems, with the appropriate interpretation of the case $\det(M)=0$. We discuss an example of this phenomenon in Section \[sec:TM\] below. This exceptional scaling behaviour is known from [@GL]; see also [@Josh1]. Substitutions with more than two letters ---------------------------------------- The methods described above are not restricted to the binary case. In general, one can at least obtain a bound for the behaviour of $Z(k)$ from the spectral gap of the substitution matrix. For an arbitrary substitution, one generates the corresponding inflation rule with intervals of natural length, as determined by the left eigenvector of the substitution matrix. As an example, let us consider the substitution $$a\mapsto abc\, , \quad b\mapsto ab\, , \quad c\mapsto b\, ,$$ which occurs in the description of the Kolakoski sequence with run lengths $3$ and $1$; see [@BS04] as well as [@TAO Exs. 4.8 and 7.5]. The substitution matrix has spectrum $\sigma(M)=\{\lambda,\mu,\overline{\mu}\}$, which are the roots of the polynomial $x^3-2x^2-1$. In particular, $\lambda$ is a Pisot–Vijayaraghavan number and a unit, and $\lvert\mu\rvert^2=1/\lambda$. The natural interval lengths are $\lambda(\lambda-1)$, $\lambda$ and $1$. The Lyapunov spectrum is $\sigma^{}_{\mathrm{L}}(M)=\{\log(\lambda),\log\lvert\mu\rvert\}= \{\log(\lambda),-\frac{1}{2}\log(\lambda)\}$. This leads to the relevant amplitude scaling $$\Big\| A\Bigl({\frac{\raisebox{-2pt}{$k$}} {\raisebox{0.5pt}{$\lambda^{n}$}}}\Bigr) \Big\|\, \sim \, c(k)\, \lambda^{-3n/2} \, \big\| A(k)\big\|$$ as $n\to\infty$. With the usual arguments from above, this gives $$Z(k) \, \asymp \, k^3 \quad\text{as $k\,\raisebox{2pt} {$\scriptscriptstyle \searrow$}\, 0{\hspace{0.5pt}}$.}$$ The same asymptotic behaviour for small $k$ emerges for the plastic number substitution, compare [@TAO Exs. 3.5 and 7.6], which is another ternary substitution whose characteristic polynomial has one real and a pair of complex conjugate roots. When one considers a substitution with three real eigenvalues of $M$, say $\sigma(M)=\{\lambda,\lambda^{}_{1},\lambda^{}_{2}\}$, and $\lvert\det(M)\rvert=1$, it is not immediately clear which of the two relevant Lyapunov exponents determines the scaling behaviour. When the characteristic polynomial is irreducible, it will be the larger of the two in modulus. In any case, one obtains $Z(k)={\mathcal{O}}(k^{2\tilde{\alpha}})$ as $k\,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0$, with $$\tilde{\alpha}\, =\, 1- \frac{\log\lvert\lambda_{i}\rvert}{\log(\lambda)}$$ for $i=1$ or $i=2$. The analogous formula applies to larger alphabets as well, where it remains to be determined which eigenvalue dominates the asymptotic behaviour. Square-free integers -------------------- As an example of a pure-point diffractive system of a different nature (and which, in particular, does not possess an inflation symmetry), we consider the point set of square-free integers in ${\mathbb{Z}{\hspace{0.5pt}}}$. An integer is called *square-free* if it is not divisible by the square of any prime, which for instance means that $1$, $2$, $3$, $5$ and $6$ are square-free, while $0$, $4$, $8$, $9$ and $12$ are not. Clearly, $n$ is square-free if and only if $-n$ is. Let $$V \, = \, \{ n \in {\mathbb{Z}{\hspace{0.5pt}}}: n \text{ is square-free} \}$$ denote the set of square-free integers, which has density $\operatorname{dens}(V) = \frac{6}{\pi^2}$. In fact, $V$ can be described by the projection formalism [@BMP], and is actually an example of a weak model set of maximal density [@BHS; @KR], which gives access to the full spectral information. Let us now define the (discrete) hull of $V$ as ${\mathbb{X}}_V = \overline{\{t+V : t \in {\mathbb{Z}{\hspace{0.5pt}}}\}}$, where the closure is taken in the local topology. Then, one considers the topological dynamical system $({\mathbb{X}}_V , {\mathbb{Z}{\hspace{0.5pt}}})$, which has positive topological entropy. Some people prefer to consider the flow under the continuous translation action of ${\mathbb{R}}$, which emerges via a standard suspension with a constant roof function. By slight abuse of notation, we write $({\mathbb{X}}_V, {\mathbb{R}})$ for it, although the hull is now different. One usually equips ${\mathbb{X}}_V$ with the patch frequency or *Mirsky* measure, $\nu^{}_{\text{M}}$. This system has pure point diffraction and dynamical spectrum,[^2] and the set $V$ itself is generic for $\nu^{}_{\text{M}}$. Therefore, one can calculate the diffraction measure from $V$ itself. It is given [@BMP] by $$\widehat{\gamma^{}_{V}} \, = \, \operatorname{dens}(V)^2 \sum_{\kappa \in L^{\circledast}} I(\kappa) \, \delta^{}_{\kappa} \, , \quad \text{with $\operatorname{dens}(V) = {\frac{\raisebox{-2pt}{$1$}} {\raisebox{0.5pt}{$\zeta (2)$}}} = {\frac{\raisebox{-2pt}{$6$}} {\raisebox{0.5pt}{$\pi^2$}}}$} {\hspace{0.5pt}},$$ where $\zeta(s)$ denotes Riemann’s zeta function. Here, $L^{\circledast}$ is the subset of rational numbers with cube-free denominator, which is a subgroup of ${\mathbb{Q}}$ and, at the same time, the dynamical spectrum of $({\mathbb{X}}_V, {\mathbb{R}},\nu^{}_{\mathrm{M}})$ in additive notation. For $\kappa = \frac{m}{q}$ with $m$ and $q$ coprime integers, one has $I(\kappa) = I(0) {\hspace{0.5pt}}f(q)^2$ with $I(0) = \operatorname{dens}(V)^2$ and $$\label{eq:f} f (q) \, = \, \begin{cases} \prod_{p | q} \frac{1}{p^2 -1} {\hspace{0.5pt}}, & \text{if $q$ is cube-free}, \\ 0 {\hspace{0.5pt}}, & \text{otherwise}, \end{cases}$$ where $p | q$ means that $p$ runs through all primes that divide $q$. In particular, $I (\kappa )$ only depends on the denominator of $\kappa \in {\mathbb{Q}}$, so that the diffraction measure is $1$-periodic. Let ${\mathbb{N}}^{(3)}$ denote the positive cube-free integers and consider $Z (k) {\mathrel{\mathop:}=}\widehat{\gamma^{}_{V}} \bigl( (0,k] \bigr) / I(0)$ for small $k>0$, where we get $$\label{eq:sqfreeZ} Z (k) \; = \sum_{\substack{q \in {\mathbb{N}}^{(3)} \\ qk \geqslant 1}} \operatorname{card}\{ 1 \leqslant m \leqslant qk : \gcd (m,q)=1 \} \, f (q)^2$$ because $I \bigl( \frac{m}{q} \bigr)$ with $\gcd (m,q) = 1$ does not depend on $m$. Via some standard, though slightly tricky, methods from analytic number theory, one can show [@BC] that $$\label{eq:bounds} \log\bigl(Z (k)\bigr) \, \sim \, {\frac{\raisebox{-2pt}{$3$}} {\raisebox{0.5pt}{$2$}}} \log(k) \quad \text{as $k\,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0{\hspace{0.5pt}}$}.$$ ![\[fig:squarefree\]Approximation of $R(k)=\log(Z(k))/\log(k)$ based on truncating the sum in Eq.  to cube-free numbers $q$ that are multiples of the first $2^{13}$ square-free numbers (circles) and the first $2^{20}$ square-free numbers (crosses), respectively. The blue line corresponds to the ratio $3/2$.](sfrat.eps){width="80.00000%"} The result of a numerical approximation to the sum in Eq.  is shown in Figure \[fig:squarefree\]. Here, the sum over all cube-free numbers is rewritten as a sum over all square-free numbers and their cube-free multiples which share the same value of the function $f$ of Eq. . Truncated sums including up to $2^{20}$ square-free numbers are used, showing a behaviour that is close to the asymptotic form of $Z(k)$ from Eq. . Note that, due to the truncation of a sum of positive numbers, the numerical approximations for the ratio $R(k)=\log(Z(k))/\log(k)$, where $k\,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0$, are always larger than the actual values, and apart from a couple of points at large values of $k$, all numerical estimates are above the asymptotic line. Also, since fewer terms contribute to the sums for smaller $k$, eventually the difference will become large for small $k$, as seen in Figure \[fig:squarefree\] for the data using fewer square-free numbers. A quantitive estimate of the error is difficult. For a more detailed analysis of this asymptotic behaviour, including power-free numbers for larger powers, we refer to [@BC]. Systems with absolutely continuous spectrum {#sec:ac} =========================================== The generic case for the emergence of absolutely continuous diffraction is the presence of some degree of randomness. This means that the most natural setting here is that of stochastic point processes, and quite a bit of work has recently been done; compare [@GL18] and references therein. Our goal in this section is modest in the sense that we only aim at collecting some one-dimensional results that are all well known, but rarely appear together. Moreover, we are not looking into the hyperuniform setting as in [@GL18], but rather into systems that frequently appear in practice. For this reason, we keep also this section somewhat informal. Poisson and Rudin–Shapiro ------------------------- The homogeneous *Poisson process* with intensity (or point density) $1$ on the real line has diffraction measure $\widehat{\gamma} = \delta^{}_{0} + {\lambda^{}_{\mathrm{L}}}$, where ${\lambda^{}_{\mathrm{L}}}$ denotes Lebesgue measure. This result applies to almost every realisation of this classic stochastic point process; see [@TAO Ex. 11.6]. This gives $$\label{eq:Poisson} Z^{}_{\text{P}} (k) \, = \, k$$ in this case, which can serve as a reference for the comparison with other stochastic processes. The *Bernoulli lattice gas* on ${\mathbb{Z}{\hspace{0.5pt}}}$ with occupation probability $p$ per site, see [@TAO Ex. 11.2], leads to $\widehat{\gamma} = p^2 \delta^{}_{{\mathbb{Z}{\hspace{0.5pt}}}} + p (1-p) {\lambda^{}_{\mathrm{L}}}$ for a.e. realisation, and hence to $$Z (k) \, = \, p (1-p) k {\hspace{0.5pt}}.$$ When using weights $\pm 1$ instead of $1$ and $0$, this changes to $\widehat{\gamma} = (2p-1)^2 \delta^{}_{{\mathbb{Z}{\hspace{0.5pt}}}} + 4 p (1-p) {\lambda^{}_{\mathrm{L}}}$ and thus to $$Z (k) \, = \, 4 p (1-p) k$$ by [@TAO Prop. 11.1]. In particular, $p=\frac{1}{2}$ gives $Z (k) = k$ as for the Poisson process above. Similar results emerge for more general Bernoulli systems, where the slope of $Z$ is a measure of the point (or particle) density and of the variance of the underlying random variable; see [@TAO Rem. 11.1] for more. The binary *Rudin–Shapiro chain*, when realised as a sequence with weights $\{ \pm 1 \}$, is a deterministic system with diffraction $\widehat{\gamma} = {\lambda^{}_{\mathrm{L}}}$, and thus $Z (k) = k$ exactly; see [@TAO Thm. 10.2] and references given there. In particular, this system has linear complexity, as one can obtain it as a factor of a substitution rule on four letters [@TAO Sec. 4.7.1]. Note that these results apply to every element of the Rudin–Shapiro hull, which is minimal. One can apply a simple Bernoullisation process [@Bern] to the Rudin–Shapiro chain (or to any element of its hull). Given $p\in [0,1]$, one simply changes the sign at each site independently with probability $p$; see [@TAO Sec. 11.2.2] for the general setting and further details. By an application of [@TAO Prop. 11.2], the diffraction measure almost surely does not change, which means that we get $$Z (k) \, = \, k \quad \text{for all } p \in [0,1] {\hspace{0.5pt}}.$$ This is an interesting case of *homometry*,[^3] as one can continuously change the entropy of the system from $0$ (for $p=0$) to $\log (2)$ (for $p=\frac{1}{2}$, which really gives the fair coin tossing sequence with weights $\pm 1$). In particular, the (frequency weighted) patch complexity has no influence on $Z$ in this case. This highlights the limitation of this quantity (as well as that of the full diffraction measure) as a characterisation of order. Random matrix ensembles ----------------------- ![\[fig:dyson\]Absolutely continuous part of the diffraction measure for the classic random matrix ensembles with parameters $\beta=1$ (red), $\beta=2$ (green) and $\beta=4$ (blue). ](dyson-diff.eps){width="40.00000%"} The point processes (with point density $1$) derived from the classic random matrix ensembles with parameter $\beta$, compare [@F] for background, can also be analysed for the scaling near $0$. Based on the calculations in [@BK11], see also [@TAO Thm. 11.3], the scaling behaviour emerges from integrating the Radon–Nikodym densities, which are illustrated in Figure \[fig:dyson\]. The result reads $$Z_{\beta}(k) \, = \, \begin{cases} k^2 -\tfrac{2}{3}k^3 + {\mathcal{O}}(k^4), & \text{for $\beta=1$},\\ \tfrac{1}{2} k^2, & \text{for $\beta=2$},\\ \tfrac{1}{4}k^2 +\tfrac{1}{12} k^3 + {\mathcal{O}}(k^4), & \text{for $\beta=4$}. \end{cases}$$ The general behaviour is $Z_{\beta}(k)=\frac{1}{\beta} k^2+{\mathcal{O}}(k^3)$, which significantly differs from the other cases discussed in this section. As $\beta\,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0$, the leading term diverges, which indicates that one gets a different power law in this limit. Indeed, $\beta\,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0$ gives the homogeneous Poisson process with unit density, with $Z^{}_{\text{P}}(k)=k$ as in Eq. . In contrast, for $\beta\to\infty$, the leading term vanishes. This limit corresponds to the (deterministic) point process that realises the lattice ${\mathbb{Z}{\hspace{0.5pt}}}$, where $Z(k)=0$ for all sufficiently small $k>0$. Markov lattice gas ------------------ Let us consider a simple one-dimensional Markov lattice gas on ${\mathbb{Z}{\hspace{0.5pt}}}$, as defined by $$\raisebox{-18pt}{\includegraphics[width=0.3\textwidth]{markovgraph.eps}}\qquad \text{with Markov matrix} \quad M \, = \, \begin{pmatrix} p & 1-p \\ 1-q & q \end{pmatrix},$$ where $p$ and $q$ are probabilities subject to the constraint $0<p+q<2$. The two states of the Markov chain correspond to empty (weight $0$) and occupied (weight $1$) sites. Interpreted as a lattice gas, we almost surely get a particle system of density $\varrho=\frac{1-p}{2-p-q}$. Employing the results from [@TAO Ex. 11.11], we know that the absolutely continuous part of the diffraction measure, $\widehat{\gamma}^{}_{\text{ac}}$, is almost surely represented by the Radon–Nikodym density $$g(k) \, = \, \frac{(1-p)(1-q)(1+r)} {(1-r)(1-2{\hspace{0.5pt}}r\cos(2\pi k) + r^2)}\, ,$$ where $r=p+q-1$, which satisfies $|r|<1$ by our assumption. A straightforward calculation gives $$\label{eq:MarkovZ} Z(k) \, = \, g(0) \Bigl( k - {\frac{\raisebox{-2pt}{$4{\hspace{0.5pt}}\pi^2$}} {\raisebox{0.5pt}{$3$}}}{\hspace{0.5pt}}{\frac{\raisebox{-2pt}{$r$}} {\raisebox{0.5pt}{$(1-r)^2$}}} \, k^3 + {\mathcal{O}}(k^5)\Bigr)$$ with $g(0)=\frac{(1-p)(1-q)(1+r)}{(1-r)^3}\geqslant 0$. When $r=0$, Eq.  simplifies to $$Z(k)\, =\, (1-p)(1-q){\hspace{0.5pt}}k{\hspace{0.5pt}}.$$ Note that $r<0$ and $r>0$ correspond to effectively repulsive and attractive systems, respectively; see Figure \[fig:Markov\] for an illustration of $g$ for these two situations. ![\[fig:Markov\]Absolutely continuous part of the Markov chain diffraction. The left panel shows the case $p=q=\frac{1}{4}$, which is effectively repulsive, with parameter . The right panel shows the corresponding effectively attractive case for $p=q=\frac{3}{4}$, with $r=\frac{1}{2}$.](markov.eps){width="84.00000%"} Binary random tilings on the line --------------------------------- To keep things simple, let us consider a random tiling of ${\mathbb{R}}$ that is built from just two intervals (of lengths $u$ and $v$, say) with corresponding probabilities $p$ and $q=1-p$. We place a scatterer of unit strength at the left endpoint of each interval. Then, almost surely, the diffraction comprises a pure point and an absolutely continuous component, see [@TAO Thm. 11.6], where only the latter contributes to the scaling of $Z(k)$ near $0$. By a straightforward integration, as $k\,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0$, one obtains $$Z(k) \, = \, \frac{p{\hspace{0.5pt}}q(u-v)^2}{(pu+qv)^3} \left(k+\frac{\pi^2}{9}{\hspace{0.5pt}}\frac{u^2 v^2 - 2 u{\hspace{0.5pt}}v (pu^2 + qv^2 )} {p{\hspace{0.5pt}}q(u-v)^2 - (pu^2+qv^2)}\, k^3 + {\mathcal{O}}(k^5)\right),$$ which is non-trivial only for $u\ne v$. For $u=v$, the resulting point set is always periodic, and $Z(k)=0$ for all sufficiently small $k>0$. Note that, as in all other examples of this section, $Z(k)$ is a continuous function, and one can often determine both the scaling exponent and the corresponding coefficient explicitly. While the continuity persists for singular continuous $\widehat{\gamma}$, determining the constant will generally be more complicated. Singular continuous measures {#sec:sc} ============================ Let us finally consider the case of structures with singular continuous diffraction. While this seems a rare phenomenon in practice, such systems are actually generic in some sense (in particular among substitution systems), and thus deserve more attention. The –Morse measure {#sec:TM} ------------------ The singular continuous –Morse (TM) measure $\mu^{}_{\mathrm{TM}}$ is defined as an infinite Riesz product [@Q; @TAO], $$\mu^{}_{\mathrm{TM}} \, = \prod_{\ell=0}^{\infty} \bigl( 1 - \cos(2^{\ell+1} \pi x) \bigr) ,$$ which is to be understood as a limit in the vague topology; we refer to [@TAO Ch. 10.1] and references therein for background and notation. Note that $\mu^{}_{\mathrm{TM}}$ is a probability measure on ${\mathbb{T}}={\mathbb{R}}/{\mathbb{Z}{\hspace{0.5pt}}}$, which is the viewpoint from dynamical systems. It is connected to the diffraction measure $\widehat{\gamma}^{}_{\mathrm{TM}}$ on ${\mathbb{R}}$ via $$\widehat{\gamma}^{}_{\mathrm{TM}} \, = \, \mu^{}_{\mathrm{TM}} * \delta^{}_{{\mathbb{Z}{\hspace{0.5pt}}}} {\hspace{0.5pt}},$$ under the obvious identification of ${\mathbb{T}}$ with $[0,1)$ and addition modulo $1$. For the investigation of $Z (k)$ for small $k$, we can thus simply work with $\mu^{}_{\mathrm{TM}}$. More concretely, for $n\in{\mathbb{N}}_0$, we consider the functions $$f^{}_{n} (x) \, = \prod_{\ell=0}^{n-1} \bigl( 1 - \cos(2^{\ell+1} \pi x) \bigr) ,$$ which are probability densities on the unit interval, where $f^{}_{0} (x) =1$ as usual. They satisfy $$\label{eq:f-split} f^{}_{n+m} (x) \, = \, f^{}_{n} (x) \, f^{}_{m} (2^n x)$$ for all $n,m \in {\mathbb{N}}_0$. Since $f^{}_{n} (x) = f^{}_{n} (1-x)$, one also has $$\label{eq:expect} \int_{0}^{1} \! x \, f^{}_{n} (x) {\,\mathrm{d}}x \, = \, {\frac{\raisebox{-2pt}{$1$}} {\raisebox{0.5pt}{$2$}}}$$ for all $n\in{\mathbb{N}}_0$. The corresponding distribution functions $$F^{}_{n} (k) \, = \int_{0}^{k} \! {\,\mathrm{d}}F^{}_{n} (x) \, = \int_{0}^{k} \! f^{}_{n} (x) {\,\mathrm{d}}x$$ converge uniformly to the distribution function $F (x) {\mathrel{\mathop:}=}\mu^{}_{\mathrm{TM}} \bigl( [0,k] \bigr)$, which is a continuous function on $[0,1]$ with $F (0)=0$ and $F(1)=1$. In fact, $F$ is strictly increasing and satisfies the symmetry relation $F(1-k) = 1- F(k)$. Moreover, $F$ possesses the uniformly converging Fourier series representation [@TAO Eq. 10.9] $$F (k) \, = \, k \, + \sum_{m=1}^{\infty} \frac{\eta (m)}{m \pi} {\hspace{0.5pt}}\sin (2 \pi m k) {\hspace{0.5pt}},$$ where the $\eta(m)$ are the Fourier coefficients of $\mu^{}_{\mathrm{TM}}$. Note that the $\eta(m)$ are also the pair correlation coefficients of the –Morse sequence when realised with weights $\pm 1$. They are recursively specified by $\eta(0)=1$ together with $$\eta(2m) \, =\, \eta(m)\quad\text{and}\quad \eta(2m+1) \, = \, -{\frac{\raisebox{-2pt}{$1$}} {\raisebox{0.5pt}{$2$}}} \bigl( \eta (m) + \eta(m+1)\bigr)$$ for $m\in{\mathbb{N}}_0$. In particular, this entails $\eta(1) = - \frac{1}{3}$, while all other values then follow recursively. To study $Z(k) = F(k)$ for small $k$, compare [@BGKS], it suffices to analyse the behaviour of $F$ along sequences $(2^{-n})^{}_{n\in{\mathbb{N}}}$ because $F$ is strictly increasing and continuous. Let $N\geqslant n$ and observe $$\label{eq:upper} \begin{split} F^{}_{N} (2^{-n}) \, & = \int_{0}^{2^{-n}} \!\!\! f^{}_{N} (x) {\,\mathrm{d}}x \, \stackrel{\eqref{eq:f-split}}{=} \int_{0}^{2^{-n}} \!\!\! f^{}_{n} (x) \, f^{}_{N-n} (2^n x) {\,\mathrm{d}}x \\[2mm] & \leqslant \sup \big\{ f^{}_{n} (x) : x\in [0,2^{-n}] \big\} \; 2^{-n} \! \int_{0}^{1} \! f^{}_{N-n} (y) {\,\mathrm{d}}y \, = \, 2^{-n} f^{}_{n} (2^{-n}) {\hspace{0.5pt}}, \end{split}$$ where the last step uses that $f^{}_{n}$ is increasing on $[0,2^{-n}]$. In the other direction, one gets $$\label{eq:lower} \begin{split} F^{}_{N} (2^{-n}) \, & = \, 2^{-n} \! \int_{0}^{1} f^{}_{n} ( 2^{-n} y) \, f^{}_{N-n} (y) {\,\mathrm{d}}y \\[2mm] & \geqslant \, 2^{-n} f^{}_{n} \biggl( 2^{-n} \! \int_{0}^{1} \! y \, f^{}_{N-n} (y) {\,\mathrm{d}}y \biggr) \, \stackrel{\eqref{eq:expect}}{=} \, 2^{-n} f^{}_{n} (2^{-n-1}) {\hspace{0.5pt}}, \end{split}$$ where the estimate follows from Jensen’s inequality, as $f^{}_{n} (2^{-n} x)$ is convex on $[0,1]$. \[rem:improve\] One can improve the lower bound by splitting the integral as $$F^{}_{N} (2^{-n}) \, = \, 2^{-n} \! \int_{0}^{\frac{1}{2}} f^{}_{n-1} ( 2^{-n+1} y) \, 2 {\hspace{0.5pt}}f^{}_{N-n+1} (y) {\,\mathrm{d}}y \, \geqslant \, 2^{-n} f^{}_{n-1} (2^{-n+1} \beta^{}_{N-n+1} )$$ with $\beta^{}_{N-n+1} = 2 \int_{0}^{1/2} y {\hspace{0.5pt}}f^{}_{N-n+1} (y) {\,\mathrm{d}}y$. This trick uses the fact that $2 f^{}_{n} (x)$ is a probability density on $\bigl[ 0, \frac{1}{2}\bigr]$, as a consequence of the reflection symmetry. Taking the limit as $N\to\infty$, one obtains the estimate $F (2^{-n}) \geqslant 2^{-n} f^{}_{n-1} (2^{-n+1} \beta)$ with $$\beta \, = \, {\frac{\raisebox{-2pt}{$1$}} {\raisebox{0.5pt}{$4$}}} - {\frac{\raisebox{-2pt}{$2$}} {\raisebox{0.5pt}{$\pi^2$}}} \sum_{m \geqslant 0} \frac{\eta (2 m +1)}{(2 m + 1)^2} \, \approx \, 0.309 94 {\hspace{0.5pt}},$$ as can be calculated with the Fourier series representation of $F$. [$\Diamond$]{} So far, by taking $N\to\infty$, we have the bounds $$\label{eq:TM-bounds} 2^{-n} f^{}_{n} (2^{-n-1}) \, \leqslant \, F (2^{-n}) \, \leqslant \, 2^{-n} f^{}_{n} (2^{-n}) {\hspace{0.5pt}},$$ where the lower one can be improved as explained in Remark \[rem:improve\]. For small $x$, a series expansion analysis of $f^{}_{n}$ gives $$f^{}_{n} (x) \, = \, 2^{n^2} (\pi x)^{2n} \Bigl( 1 - {\frac{\raisebox{-2pt}{$4^n - 1$}} {\raisebox{0.5pt}{$9$}}} (\pi x)^2 + {\frac{\raisebox{-2pt}{$11\cdot 16^n - 25 \cdot 4^n + 14$}} {\raisebox{0.5pt}{$2015$}}} (\pi x)^4 + {\mathcal{O}}\bigl(x^6 \bigr) \Bigr) .$$ With a bit more work, for $x=2^{-n}$, one can now establish the asymptotic behaviour $$\label{eq:f-upper} 2^{-n} f^{}_{n} (2^{-n}) \, \sim \, c \, 2^{-n^2} \Bigl({\frac{\raisebox{-2pt}{$\pi^2$}} {\raisebox{0.5pt}{$2$}}} \Bigr)^{n} \qquad \text{as $n\to\infty$},$$ with $c \approx 0.306 {{\hspace{0.5pt}}} 663$. The constant was calculated numerically from the quickly converging sequence of quotients. Note that the right-hand side of Eq.  decays faster than any power of $2^{-n}$, which means that we will not get a power law for the scaling of $F(k)$ as $k\,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0$. Similarly, one obtains the asymptotic behaviour $$\label{eq:f-lower} 2^{-n} f^{}_{n} (2^{-n-1}) \, \sim \, {\frac{\raisebox{-2pt}{$\pi^2 c$}} {\raisebox{0.5pt}{$4$}}} \, 2^{-n^2} \Bigl( {\frac{\raisebox{-2pt}{$\pi^2$}} {\raisebox{0.5pt}{$8$}}} \Bigr)^{n} \qquad \text{as $n\to\infty$},$$ with the same constant $c$, hence $\frac{\pi^2 c}{4} \approx 0.756 {{\hspace{0.5pt}}} 660$. It remains to express the result in terms of $k = 2^{-n}$, via $\log (k) = -n \log (2)$. From Eqs.  and , in conjunction with the continuity and monotonicity of $F$, one obtains the following result. \[thm:tm\] The distribution function $F$ of the –Morse measure satisfies the asymptotic bounds $$c^{}_{1} k^{2+\alpha} \exp \Bigl( - {\frac{\raisebox{-2pt}{$\log (k)^2$}} {\raisebox{0.5pt}{$\log (2)$}}} \Bigr) \, \leqslant \, F (k) \, \leqslant \, c^{}_{2} {\hspace{0.5pt}}k^{\alpha} {\hspace{0.5pt}}\exp \Bigl( - {\frac{\raisebox{-2pt}{$\log (k)^2$}} {\raisebox{0.5pt}{$\log (2)$}}} \Bigr)$$ for small $k>0$, with exponent $$\alpha \, = \, - \frac{\log \bigl( \frac{\pi^2}{2} \bigr)}{\log(2)} \, \approx \, -2.302 {{\hspace{0.5pt}}} 992$$ and suitable constants $c^{}_{1}$, $c^{}_{2}$. A similar estimate, with the same leading term, was previously given in [@GL]. Some numerical experiments indicate that the upper bound is better then the lower one (which can be improved by Remark \[rem:improve\]), and that the true behaviour is closer to a prefactor of the form $k^{\alpha+s}$ with $s$ near $\frac{1}{2}$, which deserves further analysis. Generalised –Morse (gTM) sequences ---------------------------------- Let us finally consider the family of binary, primitive, constant-length substitutions $\varrho^{}_{p,q} \! : a \mapsto a^p b^q, b \mapsto b^{{\hspace{0.5pt}}p} a^q$ with $p,q\in{\mathbb{N}}$; see [@BGG] for a general exposition and [@TAO Ex. 10.1] for a short summary. The case $p=q=1$ is the classic –Morse substitution considered above. In what follows, we suppress the explicit dependence on $p$ and $q$ for ease of notation. ![Distribution functions $F(k)$ of the gTM measures for $(p,1)$ with $p\in\{1,3,5\}$ and the behaviour at small $k$ for $p\in\{1,2,3,4,5\}$ (inset).\[fig:TM\] Dashed lines indicate the scaling behaviour from Eq.  for $p\in\{2,3,4,5\}$.](tmgenmeas.eps){width="90.00000%"} Here, the spectral measure is given by the Riesz product $\mu = \prod_{m\geqslant 0} \vartheta\bigl((p+q)^m x\bigr)$ with the non-negative trigonometric polynomial $$\label{eq:thetadef} \vartheta(x) \, = \, 1 + {\frac{\raisebox{-2pt}{$2$}} {\raisebox{0.5pt}{$p+q$}}} \sum_{r=1}^{p+q-1} \alpha (p,q,r) \, \cos(2\pi rx) {\hspace{0.5pt}},$$ where $\alpha (p,q,r)=p+q-r-2{\hspace{0.5pt}}\min(p,q,r,p+q-r)$. For $x\to 0$, $\vartheta$ behaves as $$\label{eq:thetaexp} \vartheta(x) \, = \, \frac{(p-q)^2}{p+q} + \frac{12{\hspace{0.5pt}}p^2 q^2 + (p-q)^2 - (p-q)^4}{3(p+q)} {\hspace{0.5pt}}(\pi x)^2 + {\mathcal{O}}(x^4){\hspace{0.5pt}}.$$ Whenever $p=q$, we have $\vartheta(0)=0$, and we are in a situation similar to that of the classic –Morse sequence. Otherwise, $\vartheta(0)>0$, where this value can be less than $1$, precisely $1$ or larger than $1$. Nevertheless, all these cases can be treated by the same approach. Defining $f^{}_{n}(x)=\prod_{m=0}^{n-1}\vartheta\bigl((p+q)^mx\bigr)$, one has $$f^{}_{n+m}(x) \, = \, f^{}_{n}(x)\, f^{}_{m}\bigl((p+q)^n x\bigr)$$ for arbitrary $m,n\in{\mathbb{N}}_{0}$. In particular, this gives $$f^{}_{n+m}\Bigl({\frac{\raisebox{-2pt}{$x$}} {\raisebox{0.5pt}{$(p+q)^n$}}}\Bigr) \, = \, \vartheta\Bigl({\frac{\raisebox{-2pt}{$x$}} {\raisebox{0.5pt}{$(p+q)^n$}}}\Bigr) \, f^{}_{n+m-1}\Bigl({\frac{\raisebox{-2pt}{$x$}} {\raisebox{0.5pt}{$(p+q)^{n-1}$}}}\Bigr).$$ Since $\lim_{n\to\infty}\vartheta\bigl((p+q)^{-n}x\bigr)=\vartheta(0)$, we see that, for any fixed $m$, the limit $$\lim_{n\to\infty} \vartheta(0)^{-n} f^{}_{n+m}\bigl((p+q)^{-n}x\bigr)$$ exists. By standard arguments, this implies the asymptotic behaviour $$F\bigl((p+q)^{-n}\bigr) \, \sim \, \biggl(\frac{p+q}{\vartheta(0)}\biggr)^{\! -n}\qquad \text{as $n\to\infty$.}$$ Since $F$ is continuous and increasing, using $\vartheta(0)=(p-q)^2/(p+q)$, one finds the scaling behaviour $$\label{eq:gtmscaling} Z(k) \, = \, F(k) \, \sim \, k^{2-2\frac{\log|p-q|}{\log(p+q)}}\qquad \text{as $k\,\raisebox{2pt}{$\scriptscriptstyle \searrow$}\, 0 {\hspace{0.5pt}}$,}$$ which is in agreement with the scaling argument from [@GL] and Conjecture 3 in [@Josh2]. For $p\ne q$, the exponent can be $2$, $1$ or take positive values below $1$ that can become arbitrarily small. The corresponding behaviour for small $k$ matches well with the examples from [@BGG] and [@TAO Fig. 10.2], as illustrated in Figure \[fig:TM\] for some characteristic choices of the parameters. Concluding remarks ================== The situation for one-dimensional systems is in reasonably good shape, though a better understanding is still desirable. This is particularly so for multi-type systems, where the existing approach gives bounds but not necessarily the exact asymptotic behaviour. Less clear is the appropriate approach to analyse tilings and point processes in higher dimension, where one could single out special directions, but also consider the total diffraction into regions near the origin. This should be possible within the realm of both inflation tilings and projection point sets. Acknowledgements {#acknowledgements .unnumbered} ================ It is our pleasure to thank Michael Coons, Franz Gähler, Philipp Gohlke, Neil Mañibo and Joshua Socolar for valuable discussions and suggestions. We thank the MATRIX Institute in Creswick and the Department of Mathematics and Physics at the University of Tasmania in Hobart for hospitality, where this manuscript was completed. This work was supported by the German Research Foundation (DFG) through CRC 1283 and by EPSRC through grant EP/S010335/1. This paper is dedicated to the memory of Vladimir Rittenberg, who was our PhD supervisor. We owe a lot to Vladimir, who always freely shared his views and insights with his students, but also strongly supported them in pursuing their own ideas. [99]{} Alcaraz F C, Baake M, Grimm U and Rittenberg V, Operator content of the XXZ chain, *J. Phys. A: Math. Gen.* **21** (1988) L117–L120. Alcaraz F C and Rittenberg V, Shared information in stationary states at criticality, *J. Stat. Mech.* (2010) P03024:1–28; `arXiv:0912.2963`. Aubry S, Godrèche C and Luck J M, Scaling properties of a structure intermediate between quasiperiodic and random, *J. Stat. Phys.* **51** (1988) 1033–1074. Baake M and Coons M, Scaling of the diffraction measure of $k$-free integers near the origin, *preprint* (2019); `arXiv:1904.00279`. Baake M, Frank N P, Grimm U and Robinson E A, Geometric properties of a binary non-Pisot inflation and absence of absolutely continuous diffraction, *Studia Math.* **247** (2019) 109–154; `arXiv:1706.03976`. Baake M and Gähler F, Pair correlations of aperiodic inflation rules via renormalisation: Some interesting examples, *Topol.  & Appl.* **205** (2016) 4–27; `arXiv:1511.00885`. Baake M, Gähler F and Grimm U, Spectral and topological properties of a family of generalised –Morse sequences, *J. Math. Phys.* **53** (2012) 032701:1–24; `arXiv:1201.1423`. Baake M, Gähler F and Mañibo N, Renormalisation of pair correlation measures for primitive inflation rules and absence of absolutely continuous diffraction, *Commun. Math. Phys.*, in press; `arXiv:1805.09650`. Baake M, Gohlke P, Kesseböhmer M and Schindler T, Scaling properties of the –Morse measure, *Discr. Cont. Dynam. Syst. A* **39** (2019) 4157–4185; `arXiv:1810:06949`. Baake M and Grimm U, Kinematic diffraction is insufficient to distinguish order from disorder, *Phys. Rev. B* **79** (2009) 020203(R):1–4 and *Phys. Rev. B* **80** (2009) 029903(E); `arXiv:0810.5750`. Baake M and Grimm U, *Aperiodic Order. Vol. $1$: A Mathematical Invitation*, Cambridge University Press, Cambridge (2013). Baake M and Grimm U (eds.), *Aperiodic Order. Vol. $2$: Crystallography and Almost Periodicity*, Cambridge University Press, Cambridge (2017). Baake M, Grimm U and Mañibo N, Spectral analysis of a family of binary inflation rules, *Lett. Math. Phys.* **108** (2018) 1783–1805; `arXiv:1709.09083`. Baake M, Huck C and Strungaru N, On weak model sets of extremal density, *Indag. Math.* **28** (2017) 3–31; `arXiv:1512.07129`. Baake M and Kösters H, Random point sets and their diffraction, *Philos. Mag.* **91** (2011) 2671–2679; `arXiv:1007.3084`. Baake M and Lenz D, Spectral notions of aperiodic order, *Discr. Cont. Dynam. Syst. S* **10** (2017) 161–190; `arXiv:1601.06629`. Baake M, Moody R V and Pleasants P A B, Diffraction of visible lattice points and $k$th power free integers, *Discr. Math.* **221** (2000) 3–42; `arXiv:math.MG/9906132`. Baake M, Moody R V and Schlottmann M, Limit-(quasi)periodic point sets as quasicrystals with $p$-adic internal spaces, *J. Phys. A: Math. Gen.* **31** (1998) 5755–5765; `arXiv:math-ph/9901008`. Baake M and Sing B, Kolakoski-$(3,1)$ is a (deformed) model set, *Can. Math. Bull.* **47** (2004) 168–190; `arXiv:math.MG/0206098`. Barreira L and Pesin Y, *Nonuniform Hyperbolicity*, Cambridge University Press, Cambridge (2007). Bufetov A and Solomyak B, On the modulus of continuity for spectral measures in substitution dynamics, *Adv. Math.* **260** (2014) 84–129; `arXiv:1305.7373`. Bufetov A and Solomyak B, A spectral cocycle for substitution systems and translation flows, *preprint* `arXiv:1802.04783`. de Gier J, Nichols A, Pyatov P and Rittenberg V, Magic in the spectra of the XXZ quantum chain with boundaries at $\Delta=0$ and $\Delta =-1/2$, *Nucl. Phys. B* **729** (2005) 387–418; `arXiv:hep-th/0505062`. de Gier J, Nienhuis B, Pearce P A and Rittenberg V, Stochastic processes and conformal invariance, *Phys. Rev. E* **67** (2003) 016101:1–4; `arXiv:cond-mat/0205467`. Forrester P, *Log-Gases and Random Matrices*, Princeton University Press, Princeton (2010). Ghosh S and Lebowitz J L, Generalized stealthy hyperuniform processes:  Maximal rigidity and the bounded holes conjecture, *Commun. Math. Phys.* **363** (2018) 97–110; `arXiv:1707.04328`. Godrèche C and Luck J M, Multifractal analysis in reciprocal space and the nature of the Fourier transform of self-similar structures, *J. Phys. A: Math. Gen.* **23** (1990) 3769–3797. Hof A, On diffraction by aperiodic structures, *Commun. Math. Phys.* **169** (1995) 25–43. Keller G and Richard C, Periods and factors of weak model sets, *Israel J. Math.* **229** (2019) 85–132; `arXiv:1702.02383`. Lenz D, Continuity of eigenfunctions of uniquely ergodic dynamical systems and intensity of Bragg peaks, *Commun. Math. Phys.* **287** (2009) 225–258; `arXiv:math-ph/0608026`. Luck J M, A classification of critical phenomena on quasi-crystals and other aperiodic structures, *Europhys. Lett.* **24** (1993) 359–364. Mañibo N, Spectral analysis of primitive inflation rules, *Oberwolfach Rep.* **14** (2017) 2830–2832. Moody R V and Strungaru N, Almost periodic measures and their Fourier transforms, in [@TAO2], pp. 173–270. Oğuz E C, Socolar J E S, Steinhardt P J and Torquato S, Hyperuniformity of quasicrystals, *Phys. Rev. B* **95** (2017) 054119:1–10; `arXiv:1612:01975`. Oğuz E C, Socolar J E S, Steinhardt P J and Torquato S, Hyperuniformity and anti-hyperuniformity in one-dimensional substitution tilings, *Acta Cryst. A* **75** (2019) 3–13; `arXiv:1806.10641`. Pleasants P A B and Huck C, Entropy and diffraction of the $k$-free points in $n$-dimensional lattices, *Discr. Comput. Geom.* **50** (2013) 39–68; `arXiv:1112.1629`. Queffélec M, *Substitution Dynamical Systems — Spectral Analysis*, 2nd ed., LNM 1294, Springer, Berlin (2010). Solomyak B, Dynamics of self-similar tilings, *Ergod. Th. & Dynam. Syst.* **17** (1997) 695–738 and *Ergod. Th. & Dynam. Syst.* **19** (1999) 1685 (erratum). Torquato S and Stillinger F H, Local density fluctuations, hyperuniformity, and order metrics, *Phys. Rev. E* **68** (2003) 041113:1–25 and *Phys. Rev. E* **68** (2003) 069901 (erratum); `arXiv:cond-mat/0311532`. Viana M, *Lectures on Lyapunov Exponents*, Cambridge University Press, Cambridge (2013). [^1]: Recall that a dynamical system $({\mathbb{X}},G)$ is *minimal* when the $G$-orbit of every $x\in{\mathbb{X}}$ is dense in ${\mathbb{X}}$. [^2]: Note that the Kolmogorov–Sinai entropy with respect to $\nu^{}_{\mathrm{M}}$ vanishes, as it must for a system with pure point spectrum. In particular, the Mirsky measure is not a measure of maximal entropy; see [@Huck] and references therein for more. [^3]: Recall that two systems are called *homometric* when they possess the same autocorrelation and hence the same diffraction measure; see [@TAO Sec. 9.6] for background.

--- abstract: | Let $G$ be a locally compact group. We describe elements of $KK^G (A,B)$ by equivariant homomorphisms, following Cuntz’s treatment in the non-equivariant case. This yields another proof for the universal property of $KK^G$: It is the universal split exact stable homotopy functor. To describe a Kasparov triple $(\Hilm, \phi, F)$ for $A,B$ by an equivariant homomorphism, we have to arrange for the Fredholm operator $F$ to be equivariant. This can be done if $A$ is of the form $\Comp(L^2G) \otimes A'$ and more generally if the group action on $A$ is proper in the sense of Exel and Rieffel. address: | SFB 478—Geometrische Strukturen in der Mathematik\ Universität Münster\ Hittorfstraße 27\ 48149 Münster\ Germany author: - Ralf Meyer bibliography: - 'kk.bib' date: 'April 30, 2000' title: Equivariant Kasparov theory and generalized homomorphisms --- Introduction {#sec:intro} ============ In this article, we carry over the description of Kasparov theory in terms of generalized homomorphisms to the equivariant case. Let us first recall the well-known situation for Kasparov theory without group actions. The existence and associativity of the Kasparov product mean that we can define a category $KK$ whose objects are the separable algebras and whose morphisms from $A$ to $B$ are the elements of $KK (A,B)$. In [@cuntz:generalizedhom] and [@cuntz:newlook], Cuntz relates elements of $KK(A,B)$ with trivially graded algebras $A$ and $B$ to ordinary homomorphisms. He defines a certain ideal $\littleQ A$ in the free product $A \ast A$ and constructs a natural bijection between $KK (A,B)$ and the set $[\littleQ A, \Comp \otimes B]$ of homotopy classes of homomorphisms $\littleQ A \to \Comp \otimes B$, where $\Comp$ denotes the algebra of compact operators on a separable Hilbert space. Skandalis [@skandalis:KKsurvey] remarks that we also have $KK (A,B) \cong [\Comp \otimes \littleQ A, \Comp \otimes \littleQ B]$. The Kasparov product becomes simply the composition of homomorphisms in this picture. Cuntz’s description of $KK (A,B)$ is used by Higson [@higson:characterizeKK] to characterize Kasparov theory by a universal property: The canonical functor from separable algebras to $KK$ is the universal split exact stable homotopy functor. For graded algebras, Haag [@haag:gradedKK] describes $KK (A,B)$ in a similar way as the set of homotopy classes of grading preserving homomorphisms from $\littleX A$ to $\hat{\Comp} \hot B$ for a suitable graded algebra $\littleX A$. He shows that $KK (A,B) \cong KK^{\Z_2} (\hat{S} \hot A,B)$, where $KK^{\Z_2}$ is the equivariant Kasparov theory for trivially graded algebras and $\hat{S}$ is $\CVI(\R)$ graded by reflection at the origin. Furthermore, Haag identifies the Kasparov product for graded algebras in this setting [@haag:algebraic]. It is straightforward to carry over these results to $KK^G$ for a *compact* group $G$. However, new ideas are necessary if $G$ is merely locally compact. The only result in that generality I am aware of is due to Thomsen [@thomsen:KKGuniversal]. He shows that $KK^G$ can still be characterized as the universal split exact stable homotopy functor. However, he does not obtain a description of $KK^G$ by equivariant homomorphisms. Let $A$ and $B$ be algebras. Let $\Comp$ be the algebra of compact operators on the direct sum of infinitely many copies of $L^2G$. We would like to associate a equivariant homomorphism $\littleQ A \to \Comp \otimes B$ (or $\littleX A \to \hat{\Comp} \hot B$ in the graded case) to a Kasparov triple $(\Hilm, \phi, F)$ for $A,B$. This may be impossible for two reasons: The operator $F$ need not be equivariant, and there may be no equivariant embedding $\Hilm \subset L^2(G, B)^\infty$. It is surprisingly easy to overcome these problems: We just have to replace $A$ by $\Comp(L^2G) \otimes A$. If the map $\phi \colon \Comp(L^2G) \otimes A \to \Adj(\Hilm)$ is *essential* in the sense that $\phi(\Comp(L^2G) \otimes A) \cdot \Hilm$ is dense in $\Hilm$, then $\Hilm = L^2(G, \Hilm')$ for some Hilbert module $\Hilm'$. Hence $\Hilm$ can be embedded in $L^2(G, B)^\infty$. Moreover, the additional copy of $L^2G$ gives us enough freedom to replace $F$ by a *compact perturbation* $F'$ that is equivariant (Lemma \[lem:equi\_connection\]). Once we have that $F$ is equivariant and $\Hilm \subset L^2(G, B)^\infty$, we can proceed as in the non-equivariant case. We show that we get the same groups if we restrict to Kasparov triples and homotopies $(\Hilm, \phi, F)$ with a equivariant symmetry $F$ and $\Hilm \subset L^2(G,B)^\infty$ (Proposition \[pro:special\_aep\]). *Symmetry* means that $F = F^\ast$ and $F^2 = 1$. This yields a bijection between $KK^G( \Comp(L^2G) \otimes A, B)$ and the set of homotopy classes of equivariant homomorphisms $\littleQ (\Comp(L^2G) \otimes A) \to \Comp \otimes B$ (Proposition \[pro:KKGs\_XQ\]). In addition, we obtain an analogous statement for graded algebras and show that we may tensor $\littleQ (\Comp(L^2G) \otimes A)$ with $\Comp(\Hils)$ for any Hilbert space $\Hils$. The universal property of $KK^G$ for trivially graded separable algebras is an immediate consequence of this description of $KK^G$ because $F(\littleQ A) \cong F(A)$ for any split exact stable homotopy functor $F$. For graded algebras, we prove that $$KK^G(A, B) \cong KK^{G \times \Z_2} (\hat{S} \hot A, B),$$ where $KK^G$ and $KK^{G \times \Z_2}$ denote the Kasparov theories for graded algebras and trivially graded algebras, respectively. We describe the Kasparov product in this setting. In addition, we show that we can obtain $KK^G (A,B)$ using only Kasparov triples $(\Hilm, \phi, F)$ with equivariant $F$ and $\Hilm \subset L^2(G, B)^\infty$ if $A$ is *proper* in the sense of Exel [@exel:MoritaSpectral] and Rieffel [@rieffel:pre1]. This notion of properness is quite general and covers both algebras of the form $\Comp(L^2G) \otimes A$ and the proper algebras of [@guentner-higson-trout]. Our key result concerning proper group actions is that a countably generated Hilbert module $\Hilm$ satisfies $\Hilm \oplus L^2(G,A)^\infty \cong L^2(G,A)^\infty$ if and only if $\Comp(\Hilm)$ is a proper algebra. This is not surprising in view of Rieffel’s treatment of square-integrable representations of groups on Hilbert space [@rieffel:pre1]. Notation and Conventions {#sec:prep} ======================== For the convenience of the reader, we recall the definitions of Hilbert modules, Kasparov triples, and connections. Moreover, we fix some notation. Let $G$ be a locally compact, compact topological group. Let $dg$ be a left invariant Haar measure on $G$ and let $L^2G = L^2(G,dg)$. The *left regular representation* of $G$, defined by $\lambda_g(f) (g') \defeq f(g^{-1} g')$ for $f \in L^2G$ and $g,g' \in G$, is a strongly continuous unitary representation of $G$ on $L^2G$. We always equip the algebra $\Comp(L^2G)$ of compact operators on $L^2G$ with the action induced by $\lambda$. That is, $\lambda_g(T) = \lambda_g \circ T \circ \lambda_g^{-1}$ for all $g \in G$, $T \in \Comp(L^2G)$. Let $\Ztwo$ be the group with two elements and let $G_2 \defeq G \times \Ztwo$. A *graded algebra* or, briefly, algebra is a algebra with a strongly continuous action of $G_2$. Recall that a grading is nothing but a action. We always write $\alpha_g$ and $\beta_g$ for the actions of $g \in G_2$ on the algebras $A$ and $B$, respectively. Let $\Mat_2$ and $\hat{\Mat}_2$ be the algebra of matrices with the trivial grading and with the off-diagonal grading, respectively. That is, the off-diagonal terms in $\hat{\Mat}_2$ are odd. Let $\Cl_1$ be the first Clifford algebra, that is, the universal algebra generated by an odd symmetry. Hilbert modules {#sec:Hilbert_modules} --------------- Let $B$ be a algebras. A *graded Hilbert module* or, briefly, *Hilbert module* is a Hilbert module $\Hilm$ with valued inner product $\5{\blank}{\blank}_B$ that is equipped with a strongly continuous linear action $\gamma_g$ of $G_2$ satisfying $ \gamma_g(\xi \cdot b) = \gamma_g(\xi) \beta_g(b) $ and $ \beta_g (\5{\xi}{\eta}_B) = \5{\gamma_g\xi}{\gamma_g\eta}_B $ for all $g \in G_2$, $\xi, \eta \in \Hilm$, $b \in B$. We call $\Hilm$ *full* iff the linear span of $\5{\Hilm}{\Hilm}_B$ is dense in $B$. We write $\Adj(\Hilm)$ and $\Comp(\Hilm)$ for the algebras of *adjointable* and *compact* operators on $\Hilm$. The latter is generated by the *rank one operators* $\ket{\xi}\bra{\eta}$ defined by $$\ket{\xi}\bra{\eta} (\zeta) \defeq \xi \cdot \5{\eta}{\zeta}_B \qquad \text{for all $\xi,\eta,\zeta \in \Hilm$.}$$ We always endow $\Adj(\Hilm)$ with the induced action, $\gamma_g(T) \defeq \gamma_g \circ T \circ \gamma_g^{-1}$. This action is strongly continuous on $\Comp (\Hilm)$ but usually not on $\Adj (\Hilm)$. We call $T \in \Adj (\Hilm)$ *continuous* iff the map $g \mapsto \gamma_g(T)$ is norm continuous. We denote graded and ungraded spatial *tensor products* of graded algebras and Hilbert modules by $\hot$ and $\otimes$, respectively. If $A$ is trivially graded, then there is no difference between $A \hot B$ and $A \otimes B$. ### Standard Hilbert modules {#sec:standard_modules} Let $\ell^2(\N)$ be the separable Hilbert space with trivial action. Let $\ell^2(\Ztwo\N)$ be the graded Hilbert space $\ell^2(\N)^\even \oplus \ell^2(\N)^\odd$. Let $L^2(G\N) \defeq L^2G \otimes \ell^2(\N)$ and $L^2(G_2\N) \defeq L^2G \otimes \ell^2(\Ztwo\N)$. We abbreviate $\Comp(G) \defeq \Comp(L^2G)$, $\Comp(\N) \defeq \Comp\bigl(\ell^2(\N)\bigr)$, etc. Moreover, we write $\Comp(\cdots) A$ instead of $\Comp(\cdots) \hot A \cong \Comp(\cdots) \otimes A$. Let $\Hilm$ be a Hilbert module. Define $\Hilm^\infty \defeq \ell^2(\N) \hot \Hilm$, $L^2(G, \Hilm) \defeq L^2G \hot \Hilm$, and $L^2(G_2, \Hilm) \defeq L^2(G \times \Ztwo) \hot \Hilm$. Let $$\Hilsg_B \defeq L^2(G_2, B)^\infty \cong L^2(G, B \oplus B^\opp)^\infty, \qquad \Hilsu_B \defeq L^2(G, B)^\infty.$$ The Hilbert module $\Hilsu_B$ is important only if $B$ is trivially graded. ### Isometric embeddings {#sec:iso_embed} Let $B$ be a algebra and let $\Hilm$ and $\Hilm[F]$ be Hilbert modules. A map $\iota \colon \Hilm \to \Hilm[F]$ is called an *isometric embedding* iff it is a linear, equivariant module map and satisfies $\5{\iota(\xi)} {\iota(\eta)}_B = \5{\xi} {\eta}_B$ for all $\xi,\eta \in \Hilm$. Hence $\iota$ is injective and $\iota (\Hilm) \subset \Hilm[F]$ is a closed invariant submodule. We do not require $\iota$ to be adjointable. This happens iff $\iota (\Hilm)$ is *complementable*, that is, $\iota (\Hilm) \oplus \iota (\Hilm)^\bot = \Hilm[F]$. We write $\Hilm \subset \Hilm[F]$ iff there is an isometric embedding $\Hilm \to \Hilm[F]$. Hilbert bimodules {#sec:Hilbert_bimodules} ----------------- Let $A$ and $B$ be algebras. A *graded Hilbert bimodule* or, briefly, *Hilbert bimodule* is a Hilbert module $\Hilm$ with a equivariant homomorphism $\phi \colon A \to \Adj(\Hilm)$. We often use module notation for the action of $A$ on $\Hilm$, writing $a \xi$ instead of $\phi(a) (\xi)$. The equivariance of $\phi$ means that $\gamma_g (a\xi) = \alpha_g (a) \gamma_g (\xi)$ for all $g \in G_2$, $a \in A$, $\xi \in \Hilm$. Let $A \cdot \Hilm \subset \Hilm$ be the subset of all elements of the form $a \xi$ with $a \in A$, $\xi \in \Hilm$. The Cohen-Hewitt factorization theorem [@cohen:factorization], [@hewitt-ross:harmonicII] implies that $A \cdot \Hilm$ is a closed linear subspace. We call $\Hilm$ *essential* iff $A \cdot \Hilm = \Hilm$. Let $\Mult(A)$ be the *multiplier algebra* of $A$. If $\Hilm$ is essential, then there is a unique extension of $\phi$ to a equivariant homomorphism $\phi \colon \Mult(A) \to \Adj(\Hilm)$. The extension is defined by $\phi(m) (a \cdot \xi) = (m\cdot a) \cdot \xi$ for all $m \in \Mult(A)$, $a \in A$, $\xi \in \Hilm$. If $\Hilm_1$ is a Hilbert module and $\Hilm_2$ is a Hilbert module, then the tensor product $\Hilm_1 \hot_B \Hilm_2$ over $B$ is defined as in [@kasparov88:equivariantKK]. It is a Hilbert bimodule. If $B$ acts on $\Hilm_2$ via $\phi \colon B \to \Adj(\Hilm_2)$, then we also use the more precise notation $\Hilm_1 \hot_\phi \Hilm_2$ of [@blackadar98:Ktheory]. Imprimitivity bimodules {#sec:imprimitivity} ----------------------- Let $A$ and $B$ be algebras. A Hilbert bimodule $(\Hilm, \phi)$ is called an *imprimitivity bimodule* iff it is full and $\phi$ is an isomorphism onto $\Comp(\Hilm)$ [@rieffel74:induced]. We call $A$ and $B$ *Morita-Rieffel equivalent* iff there is an imprimitivity bimodule for them. This is an equivalence relation. Especially, if $\Hilm$ is an imprimitivity bimodule for $A,B,G_2$, then there is a *dual* imprimitivity bimodule $\Hilm^\ast$ for $B,A,G_2$. It satisfies $\Hilm^\ast \hot_A \Hilm \cong B$ as Hilbert bimodules and $\Hilm \hot_B \Hilm^\ast \cong A$ as Hilbert bimodules. A concrete model for $\Hilm^\ast$ is $\Comp(\Hilm, B)$\[dual\_imprimitivity\]. The algebras $\Comp(\Hilm) \cong A$ and $\Comp(B) \cong B$ operate on $\Comp(\Hilm, B)$ by composition. The valued inner product is defined by $\5{T_1}{T_2} \defeq T_1^\ast T_2$ for all $T_1, T_2 \in \Comp(\Hilm, B)$. Kasparov triples {#sec:Kasparov_triples} ---------------- Let $A$ and $B$ be unital algebras. A *Kasparov triple for $A,B$* is a triple $(\Hilm,\phi,F)$, where $(\Hilm, \phi)$ is a countably generated Hilbert bimodule and $F \in \Adj(\Hilm)$ is odd with respect to the grading and satisfies $$\label{eq:KK_relations} [F,\phi(a)],\ (1-F^2)\phi(a),\ (F-F^\ast)\phi(a),\ (\gamma_g(F) - F) \phi(a) \in \Comp(\Hilm)$$ for all $a \in A$, $g \in G$. The expression $[F,\phi(a)]$ in  is a *graded commutator*. In the following, all commutators will be graded. The Kasparov triple is called *degenerate* iff all the terms in  are zero. Thomsen [@thomsen:extensions] shows that  implies that the operators $F \cdot \phi(a)$ are continuous for all $a \in A$. Hence this additional requirement of Kasparov [@kasparov88:equivariantKK] is redundant. Two Kasparov triples $(\Hilm_t, \phi_t, F_t)$, $t = 0,1$, are *unitarily equivalent* iff there is a equivariant unitary $U \colon \Hilm_0 \to \Hilm_1$ with $\phi_1(a) U = U \phi_0(a)$ for all $a \in A$ and $F_1 U = U F_0$. Up to unitary equivalence, Kasparov triples are functorial for equivariant homomorphisms in both variables. If $f \colon B_1 \to B_2$ is a equivariant homomorphism and $(\Hilm, \phi, F)$ is a Kasparov triple for $A,B_1$, then $$f_\ast(\Hilm, \phi, F) \defeq (\Hilm \hot_f B_2, \phi \hot 1, F \hot 1).$$ Let $B[0,1] \defeq C([0,1]; B)$ with the pointwise action of $G_2$ and let $\ev_t \colon B[0,1] \to B$ be the evaluation homomorphism at $t \in [0,1]$. A *homotopy* between two Kasparov triples $T_0$ and $T_1$ is a Kasparov triple $\bar{T} = (\bar{\Hilm}, \bar{\phi}, \bar{F})$ for $A, B[0,1]$ such that $\bar{T}|_t \defeq (\ev_t)_\ast (\bar{\Hilm}, \bar{\phi}, \bar{F})$ is unitarily equivalent to $T_t$ for $t = 0,1$. The *Kasparov group* $KK^G(A,B)$ is defined as the set of homotopy classes of Kasparov triples for $A,B$. Let $(\Hilm, \phi, F)$ be a Kasparov triple for $A,B$. We call $F' \in \Adj(\Hilm)$ a *compact perturbation* of $F$ iff $$(F' - F)\phi(a) \in \Comp(\Hilm) \quad \text{and} \quad \phi(a)(F' - F) \in \Comp(\Hilm) \qquad \text{for all $a \in A$.}$$ If $F'$ is a compact perturbation of $F$, then $(\Hilm, \phi, F')$ is a Kasparov triple as well. The triples $(\Hilm, \phi, F)$ and $(\Hilm, \phi, F')$ are *operator homotopic* via the obvious path $F_t \defeq (1-t)F + tF'$, and therefore also homotopic. Connections {#sec:connections} ----------- Let $\Hilm_1$ be a Hilbert module and let $\Hilm_2$ be a Hilbert bimodule. Let $\Hilm_{12} \defeq \Hilm_1 \hot_A \Hilm_2$. For $\xi \in \Hilm_1$, define an adjointable operator $T_\xi \colon \Hilm_2 \to \Hilm_{12}$ by $T_\xi(\eta) \defeq \xi \hot \eta$ and $T_\xi^\ast( \eta \hot \zeta) \defeq \5{\xi}{\eta}_A \cdot \zeta$. For $\xi \in \Hilm_1$, $F_2 \in \Adj(\Hilm_2)$, and $F_{12} \in \Adj(\Hilm_{12})$, define adjointable operators on $\Hilm_2 \oplus \Hilm_{12}$ by $$\tilde{T}_\xi \defeq \begin{pmatrix} 0 & T_\xi^\ast \\ T_\xi & 0 \end{pmatrix} \qquad \text{and} \qquad F_2 \oplus F_{12} \defeq \begin{pmatrix} F_2 & 0 \\ 0 & F_{12} \end{pmatrix}.$$ The operator $F_{12}$ is called an *connection* iff $[F_2 \oplus F_{12}, \tilde{T}_\xi] \in \Comp(\Hilm_2 \oplus \Hilm_{12})$ for all $\xi \in \Hilm_1$. Assume that $F_2$ and $F_{12}$ are odd and self-adjoint and denote the grading automorphism on $\Hilm_1$ by $\tau$. Then $F_{12}$ is an connection iff $$F_{12} T_\xi - T_{\tau\xi} F_{2} \in \Comp(\Hilm_2, \Hilm_{12}) \qquad \text{for all $\xi \in \Hilm_1$}.$$ We will freely use the standard properties of connections [@blackadar98:Ktheory 18.3]. Equivariant connections and special Kasparov triples {#sec:equi_connections} ==================================================== Let $A$ and $B$ be unital algebras and let $\Hils$ be a separable Hilbert space. A Kasparov triple $(\Hilm, \phi, F)$ for $A,B$ is called *special* iff (i) $F$ is a equivariant symmetry; and (ii) $\Hils \hot \Hilm \subset \Hilsg_B$. An *special homotopy* is given by an special Kasparov triple for $A, B[0,1]$. We let $KK^G_{s,\Hils} (A,B)$ be the set of special Kasparov triples modulo special homotopy. If $\Hils = \C$, we omit the $\Hils$ and talk about *special triples*, *special homotopies*, and $KK^G_s (A,B)$. We are mostly interested in the cases $\Hils = \C$ and $\Hils = L^2(G_2\N)$. In the latter case, the condition $\Hils \hot \Hilm \subset \Hilsg_B$ becomes tautological. The additional flexibility of choosing $\Hils$ is useful in connection with Proposition \[pro:KKGs\_XQ\]. A special triple is automatically special because $\Hils \hot \Hilsg_B \cong \Hilsg_B$. Hence there are canonical maps $KK^G_s (A,B) \to KK^G_{s, \Hils} \to KK^G(A,B)$. Usually, these maps fail to be isomorphisms. For instance, if the unit element of $KK^G(\C,\C)$ comes from an element of $KK^G_s (\C,\C)$, then $G$ has to be compact. In this section, we show that $KK^G_s (A,B) \cong KK^G_{s,\Hils} (A,B) \cong KK^G (A,B)$ if $A$ has the property AE that is defined below. We verify that algebras of the form $\Comp(L^2G) A$ have this property. In Section \[sec:proper\], we will see that proper algebras have property AE as well. \[lem:equi\_connection\] Let $A$ and $B$ be unital algebras. Let $(\Hilm, \phi, F)$ be an essential Kasparov triple for $A,B$. Let $\Hilm' \defeq L^2(G,A) \hot_\phi \Hilm \cong L^2(G, \Hilm)$. There is a equivariant connection $F'$ on $\Hilm'$. Even more, we can achieve that $F'$ is a equivariant self-adjoint contraction. Let $\CC(G, \Hilm)$ be the space of continuous functions $G \to \Hilm$ with compact support. The inner product $\5{f_1}{f_2}_B \defeq \int_G \5{f_1(g)}{f_2(g)}_B \,dg$ turns $\CC(G, \Hilm)$ into a pre-Hilbert module. Its completion is $L^2(G, \Hilm)$. We have $L^2(G,A) \hot_\phi \Hilm \cong L^2(G, \Hilm)$ because $\phi$ is essential. We may assume that $F$ is a self-adjoint contraction by [@blackadar98:Ktheory 17.4.3]. Define $F' \colon \CC(G, \Hilm) \to \CC(G, \Hilm)$ by $$(F'f)(g) = \gamma_g(F) f(g) = \gamma_g\bigl( F \gamma_g^{-1} f(g) \bigr) \qquad \text{for all $g \in G$, $f \in \CC(G, \Hilm)$.}$$ It is straightforward to check that $F'$ is equivariant and odd and extends to a self-adjoint contraction $F' \colon L^2(G, \Hilm) \to L^2(G, \Hilm)$. We claim that $F'$ is an connection. Denote the grading automorphisms on $A$ and $L^2(G,A)$ by $\tau$. We have to check that $K \defeq T_\xi F - F' T_{\tau \xi} \in \Comp(\Hilm, \Hilm')$ for all $\xi \in L^2(G,A)$. We may restrict to $\xi$ of the form $\xi(g) = f(g) a$ with $f \in \CC(G)$, $a \in A$, because such elements span a dense subspace of $L^2(G,A)$. We have $$(K \eta)(g) = f(g) \phi(a) F\eta - f(g) \gamma_g(F) \phi\tau(a) \eta = K_g(\eta)$$ for all $\eta \in \Hilm$, where $$K_g \defeq f(g) \phi(a) F - f(g) \gamma_g(F) \phi\tau(a) = f(g) [\phi(a), F] + f(g) \bigl(F - \gamma_g(F) \bigr) \phi\tau(a).$$ Since $(\Hilm, \phi, F)$ is a Kasparov triple and $f$ has compact support, $K_g$ is a norm continuous compactly supported function $G \to \Comp(\Hilm)$. Using a partition of unity, we can approximate the function $g \mapsto K_g$ uniformly by finite sums of functions $g \mapsto \psi(g) T$ with $\psi \in \CC(G)$, $T \in \Comp(\Hilm)$. Approximating $T$ by sums of finite rank operators, we can approximate $K$ in norm by finite sums of operators of the form $\eta \mapsto \psi \hot \ket{\xi} \bra{\zeta} (\eta)$. Hence $K \in \Comp( \Hilm, \Hilm')$, so that $F'$ is an connection. We say that a algebra $A$ has *property AE* iff: For all unital algebras $B$ and all *essential* Kasparov triples $(\Hilm, \phi, F)$ for $A,B$, there is a equivariant compact perturbation $F'$ of $F$ and there is an isometric embedding $\Hilm \subset \Hilsg_B$. The letters AE are an abbreviation for “automatic equivariance”. \[pro:stable\_aep\] Let $A$ and $B$ be unital algebras and let $(\Hilm, \phi, F)$ be an *essential* Kasparov triple for $\Comp(G) A, B$. Then we can find a equivariant compact perturbation of $F$ and an isomorphism $$\Hilm \oplus \Hilsg_B \cong \Hilsg_B$$ of Hilbert modules. Thus $\Comp(G) A \defeq \Comp(L^2G) \hot A$ has property AE. Let $$\psi \colon \Comp(G) A \congto \Comp\bigl( L^2(G,A) \bigr)$$ be the canonical isomorphism. Thus $(L^2(G,A), \psi)$ is an imprimitivity bimodule. Let $(L^2(G,A)^\ast, \psi^\ast)$ be the corresponding dual imprimitivity bimodule. That is, $L^2(G,A)^\ast$ is a Hilbert module and $\psi^\ast$ is an isomorphism between $A$ and $\Comp( L^2(G,A)^\ast)$ such that $$L^2 (G, A) \hot_{\psi^\ast} L^2 (G, A)^\ast \cong \Comp(G) A$$ as Hilbert bimodules. Let $$\Hilm_0 \defeq L^2(G, A)^\ast \hot_\phi \Hilm, \qquad \phi_0 \defeq \psi^\ast \hot 1 \colon A \to \Adj(\Hilm_0).$$ Let $F_0 \in \Adj(\Hilm_0)$ be an connection. Then $(\Hilm_0, \phi_0, F_0)$ is an essential Kasparov triple for $A,B$. It is a Kasparov product of $(L^2(G,A)^\ast, \psi^\ast, 0)$ and $(\Hilm, \phi, F)$. Since $\phi$ is essential, we have $\Comp(G) A \hot_\phi \Hilm \cong \Hilm$ and hence $$\Hilm \cong L^2(G, A) \hot_{\psi^\ast} L^2(G, A)^\ast \hot_\phi \Hilm \cong L^2(G, A) \hot_{\phi_0} \Hilm_0$$ as Hilbert modules. We have $\phi = \psi \hot 1 \colon \Comp(G) A \to \Adj(L^2(G,A) \hot_{\phi_0} \Hilm_0)$. By Lemma \[lem:equi\_connection\], there is a equivariant connection $F'$ on $\Hilm$. The operator $F'$ is an connection on $\Comp(G) A \hot_\phi \Hilm$ by [@blackadar98:Ktheory 18.3.4.f]. Thus $F - F'$ is a connection. This means that $F'$ is a compact perturbation of $F$ by [@blackadar98:Ktheory 18.3.2.c]. As a result, $F'$ is a equivariant compact perturbation of $F$. The equivariant stabilization theorem [@mingo-phillips:triviality Theorem 2.5] for the compact group $\Ztwo$ yields $\Hilm_0 \oplus (B \oplus B^\opp)^\infty \cong (B \oplus B^\opp)^\infty$ as graded Hilbert modules. Hence $$\Hilm \oplus \Hilsg_B \cong L^2(G, \Hilm_0 \oplus (B \oplus B^\opp)^\infty) \cong L^2(G, (B \oplus B^\opp)^\infty) = \Hilsg_B$$ as Hilbert modules by [@mingo-phillips:triviality Lemma 2.3]. Thus $\Hilm \subset \Hilsg_B$. It is a well-known fact that any Kasparov triple is homotopic to an essential triple [@blackadar98:Ktheory 18.3.6]. We need a more explicit construction of the homotopy. \[lem:make\_essential\] Let $A$ and $B$ be unital algebras. Let $(\Hilm, \phi, F)$ be a Kasparov triple for $A,B$. Let $\Hilm_\ess \defeq \phi(A) \cdot \Hilm \cong A \hot_\phi \Hilm$ and define $\phi_\ess \colon A \to \Adj(\Hilm_\ess)$ by $\phi_\ess(a) = a \hot_\phi \ID_{{\Hilm}}$ for all $a \in A$. Let $F_\ess$ be an connection on $\Hilm_\ess$. Then $(\Hilm_\ess, \phi_\ess, F_\ess)$ is a Kasparov triple. There is a canonical homotopy $(\bar{\Hilm}, \bar{\phi}, \bar{F})$ between $(\Hilm, \phi, F)$ and $(\Hilm_\ess, \phi_\ess, F_\ess)$. We have $\bar{\Hilm} \subset (\Hilm \oplus \Hilm)[0,1]$. The operator $\bar{F}$ is a equivariant self-adjoint contraction if both $F$ and $F_\ess$ are equivariant self-adjoint contractions. Define maps $\phi_{11} \colon A \to \Adj(\Hilm)$, $\phi_{12} \colon A \to \Adj(\Hilm_\ess, \Hilm)$, $\phi_{21} \colon A \to \Adj(\Hilm, \Hilm_\ess)$, $\phi_{22} \colon A \to \Adj(\Hilm_\ess)$ by $\phi_{ij}(a) \xi \defeq \phi(a) \xi$ for all $\xi$ in the appropriate source $\Hilm$ or $\Hilm_\ess$. These maps combine to a equivariant homomorphism $$\phi_\ast \defeq \begin{pmatrix} \phi_{11} & \phi_{12} \\ \phi_{21} & \phi_{22} \end{pmatrix} \colon \Mat_2(A) \to \begin{pmatrix} \Adj(\Hilm, \Hilm) & \Adj(\Hilm_\ess, \Hilm) \\ \Adj(\Hilm, \Hilm_\ess) & \Adj(\Hilm_\ess, \Hilm_\ess) \end{pmatrix} = \Adj(\Hilm \oplus \Hilm_\ess).$$ We claim that $T \defeq (\Hilm \oplus \Hilm_\ess, \phi_\ast, F \oplus F_\ess)$ is a Kasparov triple for $\Mat_2(A)$ and $B$. We have $\phi_{11} = \phi$ and $\phi_{22} = \phi_\ess$. If $a \in A$, then $\phi_{21} (a)$ and $\phi_{12} (a^\ast)$ are the operators named $T_a$ and $T_a^\ast$ in the definition of a connection in Section \[sec:connections\]. Hence $[F \oplus F_\ess, \phi_\ast(x)] \in \Comp( \Hilm \oplus \Hilm_\ess)$ if $x$ is off-diagonal. Using $A \cdot A = A$, we can extend this to arbitrary $x \in \Mat_2(A)$. The other conditions for a Kasparov triple like $(1- (F \oplus F_\ess)^2) \phi_\ast(x) \in \Comp(\Hilm \oplus \Hilm_\ess)$ follow easily from the standard properties of connections [@blackadar98:Ktheory 18.3.4] if $x$ is diagonal. We can extend this to off-diagonal $x$ using once again that $A \cdot A = A$. Hence $T$ is a Kasparov triple as asserted. Let $\iota_t \colon A \to \Mat_2(A)[0,1]$ be the rotation homotopy $$\iota_t(a) \defeq \begin{pmatrix} (1-t^2)a & t \sqrt{1-t^2} a \\ t\sqrt{1-t^2} a & t^2 a \end{pmatrix}.$$ We have $$(\iota_0)_\ast (T) = (\Hilm, \phi, F) \oplus (\Hilm_\ess, 0, F_\ess) \quad\text{and}\quad (\iota_1)_\ast (T) = (\Hilm, 0, F) \oplus (\Hilm_\ess, \phi_\ess, F_\ess).$$ Thus up to degenerate triples $(\Hilm \oplus \Hilm_\ess, \phi_\ast \circ \iota_t, F \oplus F_\ess)$ is a homotopy between $(\Hilm, \phi, F)$ and $(\Hilm_\ess, \phi_\ess, F_\ess)$. Using also the canonical homotopy between a degenerate triple and zero [@blackadar98:Ktheory 17.2.3], we obtain an explicit homotopy $(\bar{\Hilm}, \bar{\phi}, \bar{F})$ between $(\Hilm, \phi, F)$ and $(\Hilm_\ess, \phi_\ess, F_\ess)$. Clearly, $\bar{\Hilm}$ and $\bar{F}$ have the desired properties. \[pro:special\_aep\] Let $A$ and $B$ be unital algebras and let $\Hils$ be a separable Hilbert space. Assume that $A$ has property AE. Then the canonical maps $KK^G_s (A,B) \to KK^G_{s,\Hils} (A,B) \to KK^G (A,B)$ are bijective. That is, any Kasparov triple for $A,B$ is homotopic to a special triple; if two special triples are homotopic, then there is an special homotopy between them. Let $(\Hilm, \phi, F)$ be a Kasparov triple for $A,B$. We may replace the operator $F_\ess$ in Lemma \[lem:make\_essential\] by an arbitrary compact perturbation [@blackadar98:Ktheory 18.3.2.c]. Hence we may select a connection $F_\ess$ that is a equivariant self-adjoint contraction by property AE and [@blackadar98:Ktheory 17.4.2–3]. A standard trick [@blackadar98:Ktheory 17.6] allows us replace $F_\ess$ by a symmetry. First add the degenerate triple $(\Hilm_\ess^{\opp}, 0, -F_\ess)$. The operator $$\tilde{F} \defeq \begin{pmatrix} F_\ess & \sqrt{1-F_\ess^2} \\ \sqrt{1-F_\ess^2} & -F_\ess \end{pmatrix} \in \Adj(\Hilm_\ess \oplus \Hilm_\ess^\opp)$$ is a equivariant symmetry and a compact perturbation of $F_\ess \oplus -F_\ess$. Property AE implies that $\Hilm_\ess \oplus \Hilm_\ess^\opp \subset \Hilsg_B \oplus \Hilsg_B^\opp \cong \Hilsg_B$. Thus $$\Psi(\Hilm, \phi, F) \defeq (\Hilm_\ess \oplus \Hilm_\ess^\opp, \phi_\ess \oplus 0, \tilde{F})$$ is a special Kasparov triple that is homotopic to $(\Hilm_\ess, \phi_\ess, F_\ess)$ and hence to $(\Hilm, \phi, F)$ by Lemma \[lem:make\_essential\]. The Kasparov triple $\Psi(\Hilm, \phi, F)$ is not quite well-defined because we have to choose a equivariant connection $F_\ess$. Since $F_\ess$ is determined uniquely up to a compact perturbation, $\Psi(\Hilm, \phi, F)$ is well-defined up to special homotopy. We have $\Psi \circ \Psi(\Hilm, \phi, F) = \Psi(\Hilm, \phi, F)$ because the essential part of $\Psi(\Hilm, \phi, F)$ is equal to $(\Hilm_\ess, \phi_\ess, F_\ess)$. Assume that two special Kasparov triples of the form $\Psi(T_0)$ and $\Psi(T_1)$ are homotopic. If we apply $\Psi$ to a homotopy between them, we obtain a special homotopy between representatives of $\Psi \circ \Psi(T_j) = \Psi(T_j)$, $j = 0,1$. Hence if Kasparov triples of the form $\Psi(T)$ are homotopic, then they are specially homotopic and a fortiori specially homotopy. The proof will be finished if we show that if $T = (\Hilm, \phi, F)$ is an special Kasparov triple, then there is an special homotopy between $T$ and $\Psi(T)$. By Lemma \[lem:make\_essential\], there is a homotopy $(\bar{\Hilm}, \bar{\phi}, \bar{F})$ between $T$ and $(\Hilm_\ess, \phi_\ess, F_\ess)$ such that $\bar{F}$ is a equivariant self-adjoint contraction and $\bar{\Hilm} \subset (\Hilm \oplus \Hilm)[0,1]$. Thus $\Hils \hot \bar{\Hilm} \subset \Hilsg_B [0,1] = \Hilsg_{B[0,1]}$ because $\Hils \hot \Hilm \subset \Hilsg_B$. Replacing $\bar{F}$ by a symmetry as above, we obtain an special homotopy between $T \oplus (\Hilm^\opp, 0, -F)$ and $\Psi(T)$. The canonical homotopy between $T$ and $T \oplus (\Hilm^\opp, 0, -F)$ [@blackadar98:Ktheory 17.2.3] is special as well. Isometric embeddings of Hilbert modules {#sec:embed_Hilbert} ======================================= In this section, we provide some techniques to deal with not necessarily adjointable embeddings of Hilbert modules. Although the group action does not create any additional difficulty here, we give complete proofs because the corresponding arguments in [@cuntz:generalizedhom], [@cuntz:newlook], and [@haag:gradedKK] are rather sketchy. Let $B$ be a algebra and let $\Hilm$ and $\Hilm[F]$ be Hilbert modules. Let $\iota \colon \Hilm \to \Hilm[F]$ be an isometric embedding as defined in Section \[sec:iso\_embed\]. Let $$\begin{aligned} \Adj_{\Hilm[F]}(\Hilm) & \defeq \{ T \in \Adj(\Hilm[F]) \mid \text{$T(\Hilm[F]) \subset \iota(\Hilm)$ and $T^\ast(\Hilm[F]) \subset \iota(\Hilm)$} \}, \\ \Comp_{\Hilm[F]}(\Hilm) & \defeq \Adj_{\Hilm[F]}(\Hilm) \cap \Comp(\Hilm[F]).\end{aligned}$$ Clearly, $\Adj_{\Hilm[F]} (\Hilm)$ and $\Comp_{\Hilm[F]} (\Hilm)$ are subalgebras of $\Adj(\Hilm[F])$. If $T_1,T_2 \in \Adj_{\Hilm[F]} (\Hilm)$, then $T_1 \Adj (\Hilm[F]) T_2 \subset \Adj_{\Hilm[F]} (\Hilm)$. Thus $\Adj_{\Hilm[F]} (\Hilm)$ and $\Comp_{\Hilm[F]} (\Hilm)$ are hereditary subalgebras of $\Adj (\Hilm[F])$. \[lem:embedding\_functorial\] For $T \in \Adj_{\Hilm[F]} (\Hilm)$, define $\rho(T) \colon \Hilm \to \Hilm$ by $\rho(T)(\xi) \defeq \iota^{-1} (T\iota\xi)$ for all $\xi \in \Hilm$. This yields a equivariant isometric homomorphism $\rho \colon \Adj_{\Hilm[F]} (\Hilm) \to \Adj (\Hilm)$. Its restriction to $\Comp_{\Hilm[F]} (\Hilm)$ is an isomorphism onto $\Comp (\Hilm)$. Let $\Comp(\iota) \colon \Comp(\Hilm) \to \Comp_{\Hilm[F]} (\Hilm) \subset \Comp(\Hilm[F])$ be the inverse of $\rho|_{\Comp_{\Hilm[F]}(\Hilm)}$. Then $\Comp(\iota)$ is the unique homomorphism satisfying $$\label{eq:comp_iota} \Comp(\iota) (\ket{\xi} \bra{\eta}) = \ket{\iota\xi} \bra{\iota\eta} \qquad \text{for all $\xi,\eta \in \Hilm$.}$$ Clearly, $\rho(T)$ is adjointable for all $T \in \Adj_{\Hilm[F]}(\Hilm)$, with adjoint $\rho(T^\ast)$. Thus $\rho$ is a homomorphism $\Adj_{\Hilm[F]} (\Hilm) \to \Adj (\Hilm)$. If $\rho(T) = 0$, then $T$ vanishes on $\iota(\Hilm) \supset \Ran T^\ast$, so that $T \circ T^\ast = 0$ and hence $T = 0$. Thus $\rho$ is isometric. Since $\rho$ is natural, it is equivariant. If $\xi,\eta \in \Hilm$, then $\ket{\iota\xi} \bra{\iota\eta} \in \Comp_{\Hilm[F]} (\Hilm)$ and $\rho( \ket{\iota\xi} \bra{\iota\eta}) = \ket{\xi} \bra{\eta}$. Thus $\rho\bigl( \Comp_{\Hilm[F]} (\Hilm) \bigr)$ contains $\Comp(\Hilm)$ and $\Comp(\iota)$ satisfies . It remains to show $\rho\bigl( \Comp_{\Hilm[F]} (\Hilm) \bigr) \subset \Comp (\Hilm)$. It suffices to verify $\rho(TT^\ast) \in \Comp(\Hilm)$ for all $T \in \Comp_{\Hilm[F]} (\Hilm)$. Evidently, $\rho(T \ket{\xi} \bra{\eta} T^\ast) = \rho( \ket{T\xi} \bra{T\eta})$ is a rank one operator for all $\xi, \eta \in \Hilm[F]$ because $T\xi, T\eta \in \iota (\Hilm)$. Therefore, $\rho(T u T^\ast) \in \Comp(\Hilm)$ for all $u \in \Comp(\Hilm[F])$. If we let $u$ run through an approximate unit for $\Comp(\Hilm[F])$, we get $\rho(TT^\ast) \in \Comp(\Hilm)$. By the way, if $\rho(T) = 1$, then $T^\ast T \colon \Hilm[F] \to \iota(\Hilm)$ is a projection onto $\iota(\Hilm)$, so that $\iota(\Hilm)$ is complementable. Hence $\rho$ is surjective iff $\iota(\Hilm)$ is complementable. As an immediate consequence, we obtain the following result of Combes and Zettl [@combes-zettl]. \[cor:hereditary\_comp\] Let $B$ be a algebra and $\Hilm[F]$ a Hilbert module. Let $H \subset \Comp (\Hilm[F])$ be a hereditary subalgebra. Then $H = \Comp_{\Hilm[F]} (H \cdot \Hilm[F]) \cong \Comp (H \cdot \Hilm[F])$. Thus the hereditary subalgebras of $\Comp (\Hilm[F])$ correspond bijectively to the not necessarily complementable Hilbert submodules of $\Hilm[F]$. Since $H$ is hereditary, $\ket{\xi} \bra{\eta} \in H$ for all $\xi, \eta \in H \cdot \Hilm[F]$. By Lemma \[lem:embedding\_functorial\], these operators generate $\Comp_{\Hilm[F]} (H \cdot \Hilm[F])$. Thus $\Comp_{\Hilm[F]} (H \cdot \Hilm[F]) \subset H$. Obviously, $H \subset \Comp_{\Hilm[F]} (H \cdot \Hilm[F])$. Two isometric embeddings $\iota_0, \iota_1 \colon \Hilm \to \Hilm[F]$ are *homotopic* iff they can be connected by a continuous path of isometric embeddings $\iota_t \colon \Hilm \to \Hilm[F]$, $t \in [0,1]$. Such a path $\iota_t$ gives rise to an isometric embedding $h \colon \Hilm{}[0,1] \to \Hilm[F][0,1]$, $(hf)(t) = \iota_t\bigl( f(t) \bigr)$. The embedding $h$ induces a map $\Comp(h) \colon \Comp(\Hilm)[0,1] \to \Comp(\Hilm[F])[0,1]$ by Lemma \[lem:embedding\_functorial\]. Composing it with the inclusion $\Comp(\Hilm) \to \Comp(\Hilm)[0,1]$ by constant functions, we obtain a equivariant homotopy between $\Comp(\iota_0)$ and $\Comp(\iota_1)$. As a result, homotopic isometric embeddings $\Hilm \to \Hilm[F]$ induce homotopic homomorphisms $\Comp(\Hilm) \to \Comp(\Hilm[F])$. \[lem:stabilization\_homotopy\] Let $B$ be a algebra and let $\Hilm$ and $\Hilm[F]$ be Hilbert modules. Then any two isometric embeddings $\Hilm \to \Hilm[F]^\infty$ are homotopic. Let $\iota_0, \iota_1 \colon \Hilm \to \Hilm[F]^\infty$ be two isometric embeddings. It is well-known that $\Hilm[F]^\infty \oplus \Hilm[F]^\infty \cong \Hilm[F]^\infty$ as Hilbert modules, and that the inclusions of the direct summands $j_0,j_1 \colon \Hilm[F]^\infty \to \Hilm[F]^\infty$ are homotopic to the identity map. These homotopies may be chosen equivariant. Hence $\iota_0$ is homotopic to $\iota'_0 \defeq j_0 \circ \iota_0$ and $\iota_1$ is homotopic to $\iota'_1 \defeq j_1 \circ \iota_1$. By construction, $\iota'_0$ and $\iota'_1$ have orthogonal ranges, that is, $\5{\iota'_0(\xi)} {\iota'_1(\eta)}_B = 0$ for all $\xi, \eta \in \Hilm$. Hence $\iota'_t \defeq \sqrt{1-t^2} \iota'_0 + t\iota'_1 \colon \Hilm \to \Hilm[F]^\infty$ is an isometric embedding for all $t \in [0,1]$. Thus $\iota'_0$ and $\iota'_1$ are homotopic. The following lemma generalizes the observation of Skandalis [@skandalis:KKsurvey] that a degenerate Kasparov triple is homotopic to zero. It is also related to [@cuntz:generalizedhom Lemma 5.1]. \[lem:degenerate\_homotopy\] Let $(\Hilm, \phi, F)$ be a Kasparov triple for $A,B$. Let $E$ be the subalgebra of $\Adj(\Hilm)$ generated by $\phi(A)$ and the operators $\gamma_g(F)$ for $g \in G$. Let $J \subideal E$ be the smallest invariant ideal containing the operators $$[F,\phi(a)],\ (1-F^2) \phi(a),\ (F-F^\ast) \phi(a),\ (\gamma_g(F) - F) \phi(a)$$ for all $a \in A$, $g \in G$. These are precisely the operators in  whose compactness (or vanishing) is required for a (degenerate) Kasparov triple. Let $\Hilm' \defeq J \cdot \Hilm$. Then $\Hilm' \subset \Hilm$ is a closed, invariant submodule and $E (\Hilm') \subset \Hilm'$. Hence restriction to $\Hilm'$ yields a well-defined equivariant homomorphism $\rho \colon E \to \Adj(\Hilm')$. Let $F' \defeq \rho(F)$, $\phi' \defeq \rho \circ \phi$. Then $(\Hilm', \phi', F')$ is a Kasparov triple for $A,B$. There is a canonical homotopy $(\bar{\Hilm}, \bar{\phi}, \bar{F})$ between $(\Hilm, \phi, F)$ and $(\Hilm', \phi', F')$. If $(\Hilm, \phi, F)$ is an special Kasparov triple, then $(\Hilm', \phi', F')$ and $(\bar{\Hilm}, \bar{\phi}, \bar{F})$ are special Kasparov triples as well. Since $J \subideal E$ is an ideal, $E(\Hilm') \subset \Hilm'$. If $T \in \Adj(\Hilm)$ satisfies $T(\Hilm') \subset \Hilm'$ and $T^\ast(\Hilm') \subset \Hilm'$, then the restriction of $T$ to $\Hilm'$ is an adjointable operator $\rho(T) \colon \Hilm' \to \Hilm'$. This yields the desired map $\rho \colon E \to \Adj(\Hilm')$. Since $(\Hilm, \phi, F)$ is a Kasparov triple, $J \subset \Comp (\Hilm)$. We have defined $\Hilm'$ so that even $J \subset \Comp_{\Hilm} (\Hilm')$. Hence $\rho(J) \subset \Comp(\Hilm')$ by Lemma \[lem:embedding\_functorial\]. This means that $(\Hilm', \phi', F')$ is a Kasparov triple. Let $\bar{\Hilm}$ be the Hilbert module $\{ f\in \Hilm{}[0,1] \mid f(1) \in \Hilm' \}$. Define $\bar{F} \in \Adj(\bar{\Hilm})$ and $\bar{\phi} \colon A \to \Adj(\bar{\Hilm})$ by $(\bar{F}f)(t) \defeq F f(t)$ and $(\bar{\phi}(a) f)(t) \defeq \phi(a) f(t)$ for all $a \in A$, $f \in \bar{\Hilm}$, $t \in [0,1]$. An argument similar to the proof that $(\Hilm', \phi', F')$ is a Kasparov triple shows that $(\bar{\Hilm}, \bar{\phi}, \bar{F})$ is a Kasparov triple for $A,B[0,1]$. It provides the desired homotopy between $(\Hilm, \phi, F)$ and $(\Hilm', \phi', F')$. Clearly, $(\Hilm', \phi', F')$ and $(\bar{\Hilm}, \bar{\phi}, \bar{F})$ are special if $(\Hilm, \phi, F)$ is special. Some universal algebra {#sec:universal} ====================== In this section, we recall the definitions and some elementary properties of the algebras $\littleQ A$ and $\littleX A$ introduced by Cuntz [@cuntz:newlook] and Haag [@haag:gradedKK]. We examine their relationship to special Kasparov triples and utilize this to describe $KK^G(A,B)$ as a set of homotopy classes of equivariant homomorphisms. The algebras $\littleX A$, $\midX A$, and $\bigX A$ {#sec:X_def} --------------------------------------------------- Let $A$ be a algebra. Define $\bigX A$ as the universal (unital) algebra generated by $A$ and a symmetry [@haag:gradedKK]. That is, we have a homomorphism $j_A \colon A \to \bigX A$ and a symmetry $F_A \in \bigX A$ such that for all triples $(B,\phi,F)$ consisting of a unital algebra $B$, a homomorphism $\phi \colon A \to B$, and a symmetry $F \in B$, there is a unique unital homomorphism $(\phi,F)_\ast \colon \bigX A \to B$ satisfying $(\phi,F)_\ast \circ j_A = \phi$ and $(\phi, F)_\ast (F_A) = F$. The construction of $\bigX A$ is clearly functorial. Hence if $A$ is a algebra, then there is an induced action of $G$ on $\bigX A$. This action is uniquely determined by the requirement that $j_A$ be equivariant and $F_A$ be invariant. Since non-commutative polynomials in $j_A(a)$, $a \in A$, and $F_A$ are dense in $\bigX A$, the action on $\bigX A$ is strongly continuous. If $A$ is graded, then we endow $\bigX A$ with the unique grading $\tau$ for which $j_A$ is equivariant and $F_A$ is *odd*, that is, $\tau(F_A) = -F_A$. If $\phi \colon A \to B$ is a equivariant homomorphism, then the induced map $\bigX \phi \colon \bigX A \to \bigX B$ is a equivariant homomorphism as well. Let $\littleX A \subideal \bigX A$ be the ideal generated by the *graded* commutators $[j_A(a),F]$ with $a \in A$. The ideal $\littleX A$ is invariant and essential. Thus $\bigX A \subset \Mult( \littleX A)$. The quotient $\bigX A / \littleX A$ is the universal unital algebra generated by $A$ and a symmetry that graded commutes with $A$. Thus $\bigX A / \littleX A \cong \Cl_1 \hot \Unse{A}$, where $\Unse{A}$ is the algebra obtained by adjoining a unit to $A$, with $\Unse{A}/ A = \C$. Let $\midX A \subideal \bigX A$ be the ideal generated by $j_A (A)$. It follows that $\midX A / \littleX A \cong \Cl_1 \hot A$, so that we have a canonical extension of algebras $$\label{eq:ext_xX} \littleX A \injto \midX A \prto \Cl_1 \hot A.$$ It is shown in the proof of [@haag:gradedKK Theorem 3.6] that this extension has a natural—hence equivariant—completely positive section. Roughly speaking, $\midX A$ is the universal algebra generated by $A$ and a symmetry in the *multiplier algebra* $\Mult(\midX A)$. Let $A$ and $B$ be algebras. There is a canonical map $\midX (A \hot B) \to \midX A \hot B$ that restricts to a map $\littleX (A \hot B) \to \littleX A \hot B$. It is defined by the homomorphism $j_A \hot \ID_B \colon A \hot B \to \midX A \hot B$ and the symmetry $F_A \hot 1 \in \Mult(\midX A \hot B)$. For $B = \NBC([0,1])$, we obtain that $\midX$ and $\littleX$ are *homotopy functors*. That is, if $f_0,f_1 \colon A \to A'$ are homotopic, then $\littleX f_0, \littleX f_1 \colon \littleX A \to \littleX A'$ are homotopic as well. For $A = \C$, we obtain canonical maps $\littleX B \to (\littleX \C) \hot B$ and $\midX B \to (\midX \C) \hot B$. Our next goal is to show that these maps are equivalences. We follow arguments in the proof of [@haag:gradedKK Proposition 3.8] in the non-equivariant case. \[pro:midX\_homotopy\] Let $A$ be a algebra. Then the canonical homomorphism $\ID \hot j_A \colon \Comp(\Z_2\N) A \to \Comp(\Z_2\N) \midX A$ is a homotopy equivalence. We call a map of the form $x \mapsto \bigl( \begin{smallmatrix} x & 0 \\ 0 & 0 \end{smallmatrix} \bigr)$ an *upper left corner embedding*. We will exhibit a canonical homomorphism $f \colon \midX A \to \hat{\Mat}_2 A$ such that $f \circ j_A$ and $(\ID_{\hat{\Mat}_2} \hot j_A) \circ f$ are both homotopic to the upper left corner embeddings $A \to \hat{\Mat}_2 A$ and $\midX A \to \hat{\Mat}_2 \midX A$. It follows that $\ID_{\Comp(\Z_2\N)} \hot f$ is a homotopy inverse for $\ID \hot j_A$. The homomorphism $f$ is defined by requiring $f \circ j_A$ to be the upper left corner embedding and $f(F_A)$ to be the *standard symmetry* $$\label{eq:def_S} S \defeq \left( \begin{smallmatrix} 0 & 1 \\ 1 & 0 \end{smallmatrix} \right).$$ By definition, $f \circ j_A \colon A \to \hat{\Mat}_2 A$ is equal to the upper left corner embedding. The symmetries $S$ and $F' \defeq F_A \oplus -F_A$ in $\hat{\Mat}_2 \Mult( \midX A)$ anti-commute. Hence $t \mapsto \sqrt{1-t^2} \cdot S + t F'$ is a path of equivariant symmetries in $\hat{\Mat}_2 \Mult(\midX A)$ connecting them. This path yields a homotopy between $(\ID_{\hat{\Mat}_2} \hot \midX A) \circ f \colon \midX A \to \hat{\Mat}_2 \midX A$ and the upper left corner embedding $\midX A \to \hat{\Mat}_2 \midX A$. Hence the canonical map $\midX A \to (\midX \C) \hot A$ is invertible in $KK^{G_2} (\midX A, (\midX \C) \hot A)$. \[pro:split\_littleX\] Let $A$ be a separable algebra. Then the canonical map $\littleX A \to (\littleX \C) \hot A$ is invertible in $KK^{G_2} (\littleX A, (\littleX \C) \hot A)$. This canonical map is part of a morphism of extensions from $\littleX A \injto \midX A \prto \Cl_1 \hot A$ to $(\littleX \C) \hot A \to (\midX \C) \hot A \prto \Cl_1 \hot A$, where the map $\midX A \to (\midX \C) \hot A$ is a equivalence by Proposition \[pro:midX\_homotopy\] and the map $\Cl_1 \hot A \to \Cl_1 \hot A$ is the identity map. Since the two extensions have completely positive equivariant sections, the long exact sequences in theory are available. The Five Lemma yields that the map $\littleX A \to (\littleX \C) \hot A$ is a equivalence as well. The algebras $\littleQ A$ and $\bigQ A$ {#sec:Q_def} --------------------------------------- Let $\bigQ A \defeq A \ast A$ be the free product of two copies of $A$ [@cuntz:newlook]. Thus there are two homomorphisms $\iota^+_A, \iota^-_A \colon A \to \bigQ A$ such that for any triple $(B,\phi^+,\phi^-)$ consisting of a algebra $B$ and a pair of homomorphisms $\phi^+,\phi^- \colon A \to B$, there is a unique homomorphism $\phi^+ \ast \phi^- \colon \bigQ A \to B$ satisfying $(\phi^+ \ast \phi^-) \circ \iota^{\pm}_A = \phi^{\pm}$. Let $\littleQ A \subideal \bigQ A$ be the ideal that is generated by the differences $\iota^+(a) - \iota^-(a)$ with $a \in A$. Alternatively, we can describe $\littleQ A$ as the kernel of the homomorphism $\ID_{A} \ast \ID_{A} \colon \bigQ A \to A$. Thus we obtain an extension of algebras $\littleQ A \injto \bigQ A \prto A$. The homomorphisms $\iota^\pm_A \colon A \to \bigQ A$ are sections for $\ID_{A} \ast \ID_{A}$. There is a natural homomorphism $\pi_A \defeq (\ID_{A} \ast 0)|_{\littleQ A} \colon \littleQ A \to A$. If $A$ is a algebra, then there is a unique strongly continuous action on $\bigQ A$ for which the homomorphisms $\iota^{\pm}_A$ are equivariant. The ideal $\littleQ A$ is invariant. The maps $\iota^\pm_A$, $\pi_A$, and $\ID_{A} \ast \ID_{A}$ above are equivariant. The functor $A \mapsto \littleQ A$ is a homotopy functor. \[pro:Q\_stable\_homotopy\] Let $A$ and $B$ be algebras. Let $\iota_1 \colon A \to A \oplus B$ and $\iota_2 \colon B \to A \oplus B$ be the standard inclusions. The homomorphism $\ID_{\Comp(\N)} \otimes (\iota_1 \ast \iota_2) \colon \Comp(\N) (A \ast B) \to \Comp(\N) (A \oplus B)$ is a homotopy equivalence. In particular, $\Comp(\N) \bigQ A$ is homotopy equivalent to $\Comp(\N) (A \oplus A)$. The stable homotopy inverse for $\iota_1 \ast \iota_2$ is the map $f \colon A \oplus B \to \Mat_2(A \ast B)$, $$f(a,b) \defeq \left( \begin{smallmatrix} a & 0 \\ 0 & b \end{smallmatrix} \right) \qquad \text{for $a \in A$, $b \in B$.}$$ The compositions $f \circ (\iota_1 \ast \iota_2)$ and $\bigl(\ID_{{\Mat_2}} \otimes (\iota_1 \ast \iota_2) \bigr) \circ f$ are homotopic to the upper left corner embeddings $A \ast B \to \Mat_2(A \ast B)$ and $A \oplus B \to \Mat_2(A \oplus B)$ in a natural way. Roughly speaking, the homotopies leave $a$ fixed and rotate $b$ to the upper left corner. Consequently, $\ID_{\Comp(\N)} \otimes (\iota_1 \ast \iota_2)$ is a homotopy equivalence. The occurring homotopies are natural and therefore equivariant. Universal algebras and Kasparov triples {#sec:uni_KK} --------------------------------------- \[pro:KKGs\_XQ\] Let $A$ and $B$ be unital algebras and let $\Hils$ be a separable Hilbert space. There are natural bijections $$KK^G_s (A,B) \cong [\littleX A, \Comp(G_2\N) B], \qquad KK^G_{s,\Hils} (A,B) \cong [\Comp(\Hils) \littleX A, \Comp(G_2\N) B].$$ If $A$, $B$, and $\Hils$ are trivially graded, then there are natural bijections $$KK^G_s(A,B) \cong [\littleQ A, \Comp(G\N) B], \qquad KK^G_{s,\Hils} (A,B) \cong [\Comp(\Hils) \littleQ A, \Comp(G\N) B].$$ All the sets $KK^G_s (A,B)$, $[\littleX A, \Comp(G_2\N) B]$, etc., in the proposition are functorial for equivariant homomorphisms $A \to A'$, $B' \to B$. Naturality means that the isomorphisms are compatible with this functoriality, so that we have isomorphisms of functors, not just of sets. Since special Kasparov triples are nothing but special Kasparov triples, it suffices to prove the assertions about $KK^G_{s,\Hils}$. Let $T \defeq (\Hilm, \phi, F)$ be an special Kasparov triple. The pair $(\phi, F)$ defines a equivariant homomorphism $(\phi, F)_\ast \colon \bigX A \to \Adj(\Hilm)$ whose restriction to $\littleX A$ has values in $\Comp (\Hilm)$. Hence we get a map $(\phi, F)_\ast^{\Hils} \defeq \ID_{\Comp(\Hils)} \hot (\phi, F)_\ast \colon \Comp(\Hils) \littleX A \to \Comp(\Hils) \Comp(\Hilm) \cong \Comp(\Hils \hot \Hilm)$. Since the Kasparov triple $T$ is special, there is an isometric embedding $\iota \colon \Hils \hot \Hilm \to \Hilsg_B$. Let $\Psi(T) \colon \littleX A \to \Comp(G_2\N) B$ be the composition $\Comp(\iota) \circ (\phi, F)_\ast^{\Hils}$. The homomorphism $\Psi(T)$ is determined uniquely up to homotopy by Lemma \[lem:stabilization\_homotopy\]. Since we can apply $\Psi$ to special homotopies as well, it descends to a map on homotopy classes $\Psi \colon KK^G_{s,\Hils} (A, B) \to [\Comp(\Hils) \littleX A, \Comp(G_2\N) B]$. It is straightforward to verify that $\Psi$ is natural. That is, if $f \colon A' \to A$ and $g \colon B \to B'$ are equivariant homomorphisms, and $T \in KK^G_{s, \Hils} (A,B)$, then $\Psi\bigl( f^\ast(T) \bigr) = \Psi(T) \circ (\ID_{\Comp(\Hils)} \hot \littleX f)$ and $\Psi\bigl( g_\ast(T) \bigr) = (\ID_{\Comp(G_2\N)} \hot g) \circ \Psi(T)$—even if $g$ is not essential. Conversely, let $f \colon \Comp(\Hils) \littleX A \to \Comp(G_2\N)B \cong \Comp(\Hilsg_B)$ be a equivariant homomorphism. Let $\Hilm_1 = f(\Comp(\Hils) \littleX A) \cdot \Hilsg_B \subset \Hilsg_B$ and let $\iota \colon \Hilm_1 \to \Hilsg_B$ be the inclusion mapping. By construction, $f(\Comp(\Hils) \littleX A) \subset \Comp_{\Hilsg_B} (\Hilm_1)$. Hence Lemma \[lem:embedding\_functorial\] yields $f = \Comp(\iota) \circ f_1$ for a equivariant essential homomorphism $f_1 \colon \Comp(\Hils) \littleX A \to \Comp(\Hilm_1)$. We claim that $\Hilm_1 \cong \Hils \hot \Hilm_2$ and $f_1 \cong \ID_{\Comp(\Hils)} \hot f_2$ for a Hilbert module $\Hilm_2$ and an essential equivariant homomorphism $f_2 \colon \littleX A \to \Comp(\Hilm_2)$. This is trivial if $\Hils = \C$. Consider the dual $(\Hils^\ast \hot \littleX A, \psi^\ast)$ of the imprimitivity bimodule $\Hils \hot \littleX A$. Thus $(\Hils \hot \littleX A) \hot_{\psi^\ast} (\Hils^\ast \hot \littleX A) \cong \Comp(\Hils) \littleX A$. Let $$\Hilm_2 \defeq (\Hils^\ast \hot \littleX A) \hot_{f_1} \Hilm_1 \qquad \text{and} \qquad f_2 \defeq \psi^\ast \hot 1.$$ Since $\psi^\ast$ is essential, so is $f_2$. Since $f_1$ is essential as well, we have $\Hils \hot \Hilm_2 \cong (\Hils \hot \littleX A) \hot_{f_2} \Hilm_2 \cong \Hilm_1$. Under this isomorphism, $f_1$ corresponds to $\ID_{\Comp(\Hils)} \hot f_2$. Since $f_1(\Comp(\Hils) \littleX A) \subset \Comp(\Hilm_1)$, it follows that $f_2(\littleX A) \subset \Comp(\Hilm_2)$. We may extend $f_2$ to $\bigX A \subset \Mult(\littleX A)$. By the universal property of $\bigX A$, this extension is of the form $(\phi, F)_\ast \colon \bigX A \to \Comp(\Hilm_2)$ for some equivariant homomorphism $\phi \colon A \to \Adj(\Hilm_2)$ and some invariant symmetry $F \in \Adj(\Hilm_2)$. The triple $\Psi^{-1}(f) \defeq (\Hilm_2, \phi, F)$ is a Kasparov triple because $(\phi,F)_\ast (\littleX A) \subset \Comp(\Hilm_2)$. It is special because $F$ is a equivariant symmetry and $\Hils \hot \Hilm_2 \cong \Hilm_1 \subset \Hilsg_B$. Evidently, $\Psi^{-1}$ descends to a map $[\Comp(\Hils) \littleX A, \Comp(G_2\N)B] \to KK^G_{s, \Hils} (A,B)$. By construction, $$\Comp(\iota) \circ (\ID_{\Comp(\Hils)} \hot (\phi, F)_\ast) = \Comp(\iota) \circ (\ID_{\Comp(\Hils)} \hot f_2) = \Comp(\iota) \circ f_1 = f.$$ That is, $\Psi \circ \Psi^{-1}$ is the identity map on $[\Comp(\Hils) \littleX A, \Comp(G_2\N) B]$. Let $(\Hilm, \phi, F)$ be an special Kasparov triple. Going through the above constructions, we find that $\Psi^{-1} \circ \Psi(\Hilm, \phi, F)$ is the Kasparov triple that is called $(\Hilm', \phi', F')$ in Lemma \[lem:degenerate\_homotopy\]. Therefore, $[\Psi^{-1} \circ \Psi (\Hilm, \phi, F)] = [(\Hilm, \phi, F)]$ in $KK^G_{s,\Hils} (A,B)$. The proof of the isomorphism $KK^G_{s,\Hils} (A,B) \cong [\Comp(\Hils) \littleX A, \Comp(G_2\N) B]$ is finished. Suppose now that $A$, $B$, and $\Hils$ are trivially graded. Let $(\Hilm, \phi, F)$ be an special Kasparov triple for $A$, $B$. The even and odd part $\Hilm^+$ and $\Hilm^-$ of $\Hilm$ are Hilbert modules as well. We may use $F$ to identify $\Hilm^+ \cong \Hilm^-$. Then $F$ becomes the standard symmetry $S \in \Adj(\Hilm^+ \oplus \Hilm^+)$ of . Since $A$ is trivially graded, we have $\phi = \phi_+ \oplus \phi_-$ for certain homomorphisms $\phi^\pm \colon A \to \Adj (\Hilm^+)$. The condition $[F, \phi(a)] \in \Comp (\Hilm)$ becomes $\phi_+(a) - \phi_-(a) \in \Comp (\Hilm^+)$ for all $a \in A$. Thus special Kasparov triples correspond bijectively to equivariant homomorphisms $f \colon \bigQ A \to \Adj(\Hilm)$ with $f(\littleQ A) \subset \Comp(\Hilm)$ and $\Hils \otimes \Hilm \subset \Hilsu_A$. Copying the argument above with $\littleQ A \subideal \bigQ A$ instead of $\littleX A \subideal \bigX A$, we obtain the desired bijection $KK^G_{s, \Hils} (A, B) \cong [\Comp(\Hils) \littleQ A, \Comp(G\N)B]$ if $A$ and $B$ are trivially graded. If $K = \Comp(G\N)$ or $K = \Comp(G_2\N)$, let $[A,B]_K$ be the set of homotopy classes of equivariant homomorphisms from $K \hot A$ to $K \hot B$. Let $$\littleX_s A \defeq \littleX (\Comp(G_2\N) A) \qquad \text{and} \qquad \littleQ_s A \defeq \littleQ (\Comp(G\N) A).$$ \[the:KKG\_XQ\] Let $G$ be a locally compact, compact topological group. Let $A$ and $B$ be unital algebras. Let $\Hils_1$ and $\Hils_2$ be separable Hilbert spaces. There are natural bijections $$KK^G(A,B) \cong [\Comp(\Hils_1)\, \littleX (\Comp(L^2G \hot \Hils_2)A),\, \Comp(G_2\N) B] \cong [\littleX_s A, B]_{\Comp(G_2\N)}.$$ If $A$, $B$, $\Hils_1$, and $\Hils_2$ are trivially graded, then there are natural bijections $$KK^G(A,B) \cong [\Comp(\Hils_1)\, \littleQ (\Comp(L^2G \otimes \Hils_2) A),\, \Comp(G\N) B] \cong [\littleQ_s A, B]_{\Comp(G\N)}.$$ The sets $KK^G(A,B)$, etc., occurring in the Theorem are functorial for equivariant homomorphisms $A' \to A$, $B \to B'$. The naturality of the isomorphisms means that they are compatible with this functoriality. Since Morita-Rieffel equivalent algebras are equivalent, there are natural isomorphisms $$KK^G(A,B) \cong KK^G(\Comp(G) A, B) \cong KK^G(\Comp(G) \Comp(\Hils_2) A, B).$$ By Proposition \[pro:stable\_aep\], $\Comp(G) A$ has the property AE. Hence Proposition \[pro:special\_aep\] yields $KK^G(\Comp(G) A, B) \cong KK^G_s(\Comp(G) A, B) \cong KK^G_{s, \Hils_1} (\Comp(G) A,B)$. A similar statement holds for $\Comp(L^2G \hot \Hils_2) A$ instead of $\Comp(G) A$. Therefore, Proposition \[pro:KKGs\_XQ\] yields the assertions. The universal property of equivariant Kasparov theory {#sec:Kasparov_universal} ===================================================== In this section, we formulate and establish the universal property of equivariant Kasparov theory for trivially graded separable algebras. Let $\GCalg$ be the category of separable algebras with equivariant homomorphisms as morphisms. Let $\SHo$ be the stable homotopy category, whose objects are the separable algebras and whose set of morphisms from $A$ to $B$ is $[A,B]_s \defeq [A,B]_{\Comp(G\N)}$. Let $KK^G$ be the category whose objects are the separable algebras and whose set of morphisms from $A$ to $B$ is $KK^G(A,B)$. The Kasparov product yields the composition of morphisms in $KK^G$. We rely on Kasparov’s work [@kasparov88:equivariantKK] and assume that the Kasparov product exists and is associative. We do not attempt an alternative definition of the Kasparov product as in [@cuntz:newlook]. It is clear that $KK^G$ is an additive category. There are obvious functors $\GCalg \to \SHo$ and $\GCalg \to KK^G$. Let $\Cat$ be a category. A functor $F \colon \GCalg \to \Cat$ is called a *homotopy functor* iff $F(f_0) = F(f_1)$ whenever $f_0$ and $f_1$ are equivariantly homotopic. A functor $F \colon \GCalg \to \Cat$ is called *stable* iff the map $F(\Comp(\Hils) A) \to F(\Comp(\Hils \oplus \Hils') A)$ induced by the inclusion $\Hils \subset \Hils \oplus \Hils'$ is an isomorphism for all separable Hilbert spaces $\Hils,\Hils'$ and all separable algebras $A$. \[pro:sho\_universal\] The functor $\GCalg \to \SHo$ is a stable homotopy functor. A functor $F \colon \GCalg \to \Cat$ is a stable homotopy functor iff it can be factored through the functor $\GCalg \to \SHo$. This factorization is automatically unique. In other words, $\SHo$ is the universal stable homotopy functor. It is left to the reader to check that the canonical functor $\GCalg \to \SHo$ is a stable homotopy functor. Thus any functor $\GCalg \to \Cat$ that factors through it is a stable homotopy functor as well. Conversely, let $F \colon \GCalg \to \Cat$ be a stable homotopy functor. Let $\Hils = \C \oplus L^2(G\N)$ and let $j_1^A \colon A \to \Comp(\Hils) A$ and $j_2^A \colon \Comp(G\N) A \to \Comp(\Hils) A$ be the canonical inclusions. Since $F$ is stable, $F(j_1^A)$ and $F(j_2^A)$ are isomorphisms. Thus $\sigma_A \defeq F(j_2^A)^{-1} \circ F(j_1^A)$ is a natural isomorphism $F(A) \congto F(\Comp(G\N) A)$. Define $$F_\ast \colon [A,B]_s \to \Mor_{\Cat}\bigl( F(A), F(B) \bigr), \qquad F_\ast [\phi] \defeq \sigma_B^{-1} \circ F(\phi) \circ \sigma_A.$$ It is left to the reader to check that this defines a functor $F_\ast \colon \SHo \to \Cat$ that extends $F$ and that the functor $F_\ast$ is determined uniquely. A homotopy functor $F \colon \GCalg \to \Cat$ is stable iff $F(A) \cong F(\Comp(G\N) A)$ naturally. The proof of Proposition \[pro:sho\_universal\] shows that a natural isomorphism $F(A) \cong F(\Comp(G\N) A)$ allows us to factor $F$ through $\SHo$. Our definition of a stable homotopy functor is equivalent to the definitions in [@guentner-higson-trout] and in [@thomsen:KKGuniversal]. A functor $F \colon \GCalg \to \Cat$ into an additive category $\Cat$ is called *split exact* iff $\bigl( F(i), F(s) \bigr) \colon F(A) \oplus F(C) \to F(B)$ is an isomorphism for all extensions $$0 \to A \xrightarrow{i} B \xrightarrow{p} C \to 0$$ of algebras that split by a equivariant homomorphism $s \colon C \to B$. \[pro:KKG\_properties\] The canonical functor $\GCalg \to KK^G$ is a split exact stable homotopy functor. Clearly, $KK^G$ is a stable homotopy functor. Split exactness is a straightforward consequence of the associativity of the Kasparov product. The argument in [@cuntz:newlook Proposition 2.1] carries over without change. Since $A \mapsto \littleQ A$ is a homotopy functor, $A \mapsto \littleQ_s A$ descends to a functor from $\SHo$ to itself. The map $\pi_{\Comp(G\N) A} \colon \littleQ_s A \to \Comp(G\N) A$ gives rise to a natural morphism $\pi^s_A \in [\littleQ_s A, A]_s$. \[lem:split\_q\] Let $F \colon \GCalg \to \Cat$ be a split exact stable homotopy functor and let $F_\ast \colon \SHo \to \Cat$ be the unique extension of $F$. Then $F_\ast (\pi^s_A)$ is invertible for all $A$. Split exactness applied to the extension $A \injto A \oplus B \prto B$ yields that the canonical map $F(A \oplus B) \to F(A) \oplus F(B)$ is an isomorphism. That is, $F$ is *additive*. Proposition \[pro:Q\_stable\_homotopy\] yields $F(A \ast B) \cong F(A) \oplus F(B)$. Split exactness applied to the extension $\littleQ A \injto \bigQ A \prto A$ implies that $F(\pi_A) \colon F(\littleQ A) \to F(A)$ is an isomorphism for all $A$. This implies that $F_\ast(\pi^s_A)$ is an isomorphism as well. By Proposition \[pro:KKG\_properties\] and Proposition \[pro:sho\_universal\], the canonical functor $\GCalg \to KK^G$ factors through a functor $\natural \colon \SHo \to KK^G$. Lemma \[lem:split\_q\] implies that $\natural(\pi^s_A) \in KK^G(\littleQ_s A, A)$ is invertible for all $A$. \[the:KKG\_doubleQ\] Let $A$ and $B$ be separable algebras. The map $$[\littleQ_s A, \littleQ_s B]_s \to KK^G(A,B), \qquad f \mapsto \natural(\pi^s_B) \circ \natural(f) \circ \natural(\pi^s_A)^{-1},$$ is a natural isomorphism. Hence the Kasparov product on $KK^G$ corresponds to the composition of homomorphisms. Since $\pi^s_B$ induces an isomorphism $KK^G(A, \littleQ_s B) \cong KK^G(A, B)$, it suffices to verify that the isomorphism $[\littleQ_s A, B]_s \to KK^G(A,B)$ of Theorem \[the:KKG\_XQ\] is given by $f \mapsto \natural(f) \circ \natural(\pi^s_A)^{-1}$. By naturality, it suffices to check this for the identity map in $[\littleQ_s A, \littleQ_s A]_s$. Composing with the invertible element $\pi^s_A$, we can reduce the theorem to the following claim: The isomorphism of Theorem \[the:KKG\_XQ\] maps $\pi^s_A \in [\littleQ_s A, A]_s$ to the unit in $KK^G (A,A)$, represented by the Kasparov triple $(A, \ID_{A}, 0)$. The proof of this claim is made somewhat messy by stabilizations, but otherwise straightforward. Therefore, we omit it. \[the:universal\_KKG\] The functor $\GCalg \to KK^G$ is the universal split exact stable homotopy functor in the following sense. An additive functor $F \colon \GCalg \to \Cat$ into an additive category $\Cat$ can be extended to a functor $F_\ast \colon KK^G \to \Cat$ iff it is a split exact stable homotopy functor. The extension is necessarily unique. Let $F \colon \GCalg \to \Cat$ be a split exact stable homotopy functor. By Proposition \[pro:sho\_universal\], we may assume that $F$ is a functor $F \colon \SHo \to \Cat$. Split exactness implies that $F(\pi^s_A)$ is an isomorphism for all $A$. If $f \in [\littleQ_s A, \littleQ_s B]_s$, define $F_\ast(f) \defeq F(\pi_s^B) \circ F(f) \circ F(\pi_s^A)^{-1}$. By Theorem \[the:KKG\_doubleQ\], this yields a functor $KK^G \to \Cat$. Evidently, this is the unique functor extending $F$. It is clear that any additive functor that factors through $KK^G$ is a split exact stable homotopy functor. The case of graded algebras {#sec:graded} =========================== Following Haag [@haag:gradedKK], we write $\Ex^G (A,B) \defeq KK^{G_2} (A,B)$ for the equivariant theory for trivially graded algebras. We show $KK^G (A,B) \cong \Ex^G (\hat{S} \hot A, B)$ and describe the Kasparov product in $KK^G$ in terms of the product in $\Ex^G$. We redefine $KK^G$ to be the category whose objects are the graded separable algebras and whose set of morphisms from $A$ to $B$ is $KK^G (A,B)$. Let $\GCalg[G_2]$ be the category of separable algebras and let $\SHo[G_2]$ be the stable homotopy category, as defined in the previous section. We redefine $\littleQ_s A \defeq \littleQ (\Comp(G_2\N) A)$, so that $\Ex^G (A, B) \cong [\littleQ_s A, B]_s$ by Theorem \[the:KKG\_XQ\]. The canonical functor $\GCalg[G_2] \to KK^G$ is still a split exact stable homotopy functor. By Theorem \[the:universal\_KKG\], we may extend it to a functor $\alpha \colon \Ex^G (A,B) \to KK^G (A,B)$. The functor $\alpha$ can be computed as follows. As in [@haag:gradedKK p. 15], the homomorphism $\iota^+ \oplus \iota^- \colon A \to \hat{\Mat}_2 (\bigQ A)$ and the symmetry $S$ of  yield a canonical map $$\alpha_0 \defeq (\iota^+ \oplus \iota^-, S)_\ast \colon \littleX A \to \hat{\Mat}_2 ( \littleQ A)$$ We view $\alpha_0$ as an element of $[\littleX A, \littleQ A]_s$. Replacing $A$ by $\Comp(G_2\N) A$, we obtain $\alpha_0 \in [\littleX_s A, \littleQ_s A]_s \cong KK^G (A, \littleQ_s A)$ by Theorem \[the:KKG\_XQ\]. \[lem:alpha0\] $\alpha_0 \in KK^G (A, \littleQ_s A)$ is the inverse of $\pi^s_A \in [\littleQ_s A, A]_s$. Lemma \[lem:split\_q\] implies that the image of $\pi^s_A$ in $KK^G (\littleQ_s A, A)$ is invertible. It remains to prove that $(\pi^s_A)_\ast (\alpha_0)$ is the identity element of $KK^G (A,A)$. We may suppose $\Comp(G_2\N) A \cong A$, so that we may omit the stabilizations and work with the map $\alpha_0 \colon \littleX A \to \hat{\Mat}_2 (\littleQ A)$. It corresponds to the Kasparov triple $(\littleQ A \oplus (\littleQ A)^\opp, \iota^+ \oplus \iota^-, S)$. Since $\pi_A \circ \iota^+ = \ID_A$, $\pi_A \circ \iota^- = 0$, we have $$(\pi_A)_\ast (\alpha_0) = (A \oplus A^\opp, \ID_A \oplus 0, S).$$ The right hand side represents the identity element of $KK^G (A,A)$. \[cor:alpha\] Let $A$ and $B$ be separable algebras. Using the isomorphisms of Theorem \[the:KKG\_XQ\], we obtain a map $$[\littleQ_s A, B]_s \cong \Ex^G (A,B) \xrightarrow{\alpha} KK^G (A,B) \cong [\littleX_s A, B]_s.$$ This map is equal to composition with $[\alpha_0] \in [\littleX_s A, \littleQ_s A]_s$. Let $f \in [\littleQ_s A,B]_s$, then the image of $f$ in $KK^G(A,B)$ is $f_\ast (\pi^s_A)^{-1} = f_\ast [\alpha_0]$. This is mapped to $f \circ [\alpha_0] \in [\littleX_s A, B]_s$. There is a canonical Kasparov triple $(\littleX A, j_A, F_A)$ for $A, \littleX A$. Replacing $A$ by $\Comp(G_2\N) A$, we obtain a canonical element $i_A \in KK^G (A, \littleX_s A)$. The isomorphism $KK^G (A, \littleX_s A) \to [\littleX_s A, \littleX_s A]_s$ maps $i_A$ to the identity map. The naturality of the isomorphism $[\littleX_s A, B] \to KK^G(A,B)$ of Theorem \[the:KKG\_XQ\] implies that it maps $f \mapsto f_\ast(i_A)$ for all $f \in [\littleX_s A,B]_s$. \[lem:ExlittleX\_KK\] Let $A$ and $B$ be separable algebras. The canonical map $$\Ex^G (\littleX_s A, B) \xrightarrow{\alpha} KK^G(\littleX_s A, B) \xrightarrow{i_A^\ast} KK^G(A,B)$$ is an isomorphism. Theorem \[the:KKG\_XQ\] yields canonical isomorphisms $\Ex^G (\littleX_s A, B) \cong [\littleQ_s \littleX_s A, B]_s$ and $KK^G (A, B) \cong [\littleX_s A, B]_s$. We are going to show that $\pi \defeq \pi^s_{\littleX_s A} \colon \littleQ_s \littleX_s A \to \littleX_s A$ is invertible in $\SHo[G_2]$. Therefore, $[\littleQ_s \littleX_s A, B]_s \cong [\littleX_s A, B]_s$. It is straightforward to show that the corresponding isomorphism $\Ex^G (\littleX_s A, B) \cong KK^G (A,B)$ is equal to the map in the statement Lemma \[lem:ExlittleX\_KK\]. The homotopy inverse for $\pi$ is constructed as a Kasparov product. Let $i = i_A \in KK^G(A, \littleX_s A)$ be as above. Let $j \in KK^G(\littleX_s A, \littleQ_s \littleX_s A)$ be the inverse of $\pi$. Let $h \in KK^G(A, \littleQ_s \littleX_s A) \cong [\littleX_s A, \littleQ_s \littleX_s A]_s$ be the Kasparov product of $i$ and $j$. The associativity of the Kasparov product implies $\pi \circ h = i$ in $KK^G(A, \littleX_s A) = [\littleX_s A, \littleX_s A]_s$. Since $\pi$ is invertible in $\Ex^G$, composition with $\pi$ is an isomorphism $$\Ex^G (\littleX_s A, \littleQ_s \littleX_s A) \congto \Ex^G (\littleX_s A, \littleX_s A).$$ Hence the equality $\pi \circ h \circ \pi = \pi$ in $[\littleQ_s \littleX_s A, \littleX_s A]_s$ implies $h \circ \pi = \ID$ in $[\littleQ_s \littleX_s A, \littleQ_s \littleX_s A]_s$. Thus $h$ is inverse to $\pi$ in $\SHo[G_2]$. Let $\hat{S}$ be the algebra $\CVI(\R)$ graded by $\tau f(x) = f(-x)$ for all $x \in \R$, $f \in \CVI(\R)$ and with trivial action. It is shown in the proof of [@haag:gradedKK Proposition 3.8] that $\littleX \C \cong \hat{\Mat}_2 \hat{S}$, so that $\hat{S}$ and $\littleX \C$ are Morita-Rieffel equivalent. Together with Proposition \[pro:split\_littleX\], we obtain a canonical isomorphism in $\Ex^G (\littleX_s A, \hat{S} \hot A)$. Let $e \in KK^G(\C, \hat{S}) \cong [\littleX_s \C, \hat{S}]_s$ be represented by the isomorphism $\littleX \C \to \hat{\Mat}_2 \hat{S}$. It is easy to verify that $e$ is homotopic to the Kasparov triple $(\hat{S}, 1, x/\sqrt{1+x^2})$, where $1 \colon \C \to \Adj(\hat{S}) \cong \BC(\R)$ is the unique unital homomorphism and $x/ \sqrt{1+x^2}$ denotes the bounded function $x \mapsto x/ \sqrt{1+x^2}$ on $\R$. \[the:graded\_equi\_KK\] Let $G$ be a locally compact compact topological group and let $A$ and $B$ be separable algebras. The composition $$\sigma \colon \Ex^G (\hat{S} \hot A, B) \xrightarrow{\alpha} KK^G (\hat{S} \hot A, B) \xrightarrow{(e \hot \ID_A)^\ast} KK^G (A,B)$$ is an isomorphism. Here $(e \hot \ID_A)^\ast$ denotes the Kasparov product with the exterior product $e \hot \ID_A \in KK^G (A, \hat{S} \hot A)$. The isomorphism $KK^G(A, \littleX_s A) \cong KK^G(A, (\littleX \C) \hot A)$ induced by the canonical map $\littleX_s A \to (\littleX \C) \hot \Comp(G_2\N) A$ maps $i_A$ to the exterior product $i_\C \hot \ID_A$. Hence the isomorphism $KK^G (A, \littleX_s A) \to KK^G(A, \hat{S} \hot A)$ maps $i_A$ to $e \hot \ID_A$. If we compose the isomorphism $\Ex^G (\littleX_s A, B) \to KK^G (A,B)$ of Lemma \[lem:ExlittleX\_KK\] with the the isomorphism $\Ex^G(\hat{S} \hot A, B) \to \Ex^G (\littleX_s A,B)$ induced by the equivalence $\hat{S} \hot A \to \littleX_s A$, we obtain that $\sigma$ is an isomorphism. We have to compute the exterior product $e \hot e \in KK^G(\C, \hat{S} \hot \hat{S})$. Since the action on $\littleX \C$ and $\hat{S}$ is trivial, we may forget about the actions. Therefore, we briefly resort to the case of trivial $G$. Theorem \[the:KKG\_XQ\] implies $$KK (\C, B) = [\littleX \C, \Comp(\Z_2\N) B] \cong [\hat{S}, \Comp(\Z_2\N) B].$$ We claim that $e \hot e \in KK (\C, \hat{S} \hot \hat{S})$ belongs to the homomorphism $\hat{S} \to \hat{S} \hot \hat{S}$ that is called $l$ by Haag [@haag:algebraic p. 87] and $\Delta$ by Higson and Kasparov [@higson-kasparov:actHilbert]. To verify this elementary claim, it is convenient to describe $KK (A,B)$ by unbounded operators following Baaj and Julg [@baaj-julg:unbounded] because in this picture exterior products are straightforward to compute. The unbounded picture of $KK(\C, B)$ is also nicely related to the isomorphism $KK(\C, B) \cong [\hat{S}, \Comp(\Z_2\N) B]$. The essential, grading preserving homomorphisms $\hat{S} \to \Adj(\Hilm)$ correspond bijectively to odd, self-adjoint, possibly unbounded multipliers of $\Hilm$ via $f \mapsto f(\ID_\R)$ for $f \colon \hat{S} \to \Adj(\Hilm)$. Since $e$ belongs to the unbounded multiplier $\ID_\R$ of $\hat{S}$, the exterior product $e \hot e$ belongs to the unbounded multiplier $\ID_\R \hot 1 + 1 \hot \ID_\R$ of $\hat{S} \hot \hat{S}$. Thus $e \hot e$ is represented by the map $\Delta$ of [@higson-kasparov:actHilbert]. It is easy to check that the concrete formula for $l$ in [@haag:gradedKK] yields nothing but $\Delta$. \[the:graded\_product\] Let $A$, $B$, and $C$ be algebras and let $x \in \Ex^G (\hat{S} \hot A,B)$, $y \in \Ex^G (\hat{S} \hot B,C)$. The Kasparov product of $\sigma(y) \in KK^G (B,C)$ and $\sigma(x) \in KK^G (A,B)$ is mapped by $\sigma^{-1}$ to the composition $$\hat{S} \hot A \xrightarrow{\Delta \hot \ID_A} \hat{S} \hot \hat{S} \hot A \xrightarrow{\ID_{\hat{S}} \hot y} \hat{S} \hot B \xrightarrow{x} C$$ in $\Ex^G$. Recall the definition of $\sigma$ in Theorem \[the:graded\_equi\_KK\] and that $\alpha$ is multiplicative. Moreover, it is easy to check that $\alpha$ is compatible with exterior products, so that $\alpha(\ID_{\hat{S}} \hot x) \cong \ID_{\hat{S}} \hot \alpha(x)$. Hence we compute $$\begin{gathered} \sigma(y) \circ \sigma(x) = \alpha(y) \circ (e \hot \ID_B) \circ \alpha(x) \circ (e \hot \ID_A) \\ = \alpha(y) \circ \bigl( e \hot \alpha(x) \bigr) \circ (e \hot \ID_A) = \alpha(y) \circ \bigl(\ID_{\hat{S}} \hot \alpha(x) \bigr) \circ (e \hot \ID_{\hat{S} \hot A}) \circ (e \hot \ID_A) \\ = \alpha(y) \circ \alpha(\ID_{\hat{S}} \hot x) \circ (e \hot e \hot \ID_A) = \sigma\bigl(y \circ (\ID_{\hat{S}} \hot x) \circ (\Delta \hot \ID_A) \bigr). \end{gathered}$$ We used that the Kasparov product is compatible with exterior products. Proper actions and square-integrable Hilbert modules {#sec:proper} ==================================================== Exel [@exel:MoritaSpectral] and Rieffel [@rieffel:pre1] define the concept of a proper action of a locally compact group on a algebra. Furthermore, Rieffel relates proper actions on the algebra $\Comp(\Hils)$ to square-integrable representations of $G$. It is very illuminating to consider also square-integrable group actions on Hilbert modules. The main result is that a countably generated Hilbert module is square-integrable iff it is a direct summand of $\Hilsu_A$. We conclude that proper algebras have property AE. Concerning questions of properness, we may ignore gradings whenever convenient. Since the group $\Ztwo$ is compact, a algebra is proper iff it is proper as a algebra. Let $A$ be a algebra and let $\Hilm$ be a Hilbert module. We denote the actions on $A$ and $\Hilm$ by $\alpha$ and $\gamma$, respectively. We frequently view $A$ as a right Hilbert module. Let $(K_n)_{n \in \N}$ be a sequence of compact subsets of $G$ such that $K_{n+1}$ is a neighborhood of $K_n$ for all $n$ and $G = \bigcup K_n$. Let $(\kappa_n)_{n \in \N}$ be an increasing sequence of functions $\kappa_n \colon G \to [0,1]$ with $\kappa_n|_{K_n} = 1$ and $\kappa_n|_{G \setminus K_{n+1}} = 0$. A continuous function $f \colon G \to A$ is called *square-integrable* iff the sequence $\int_G f^\ast(g) f(g) \kappa_n(g) \,dg$ is a norm Cauchy sequence in $A$. Equivalently, the sequence of integrals $\int_{K_n} f^\ast(g) f(g) \,dg$ is norm convergent. Observe that these sequences are increasing sequences of positive elements and that the notion of square-integrability does not depend on the choice of the sets $K_n$ or the functions $\kappa_n$. It is easy to check that $f$ is square-integrable iff the sequence $(f \cdot \kappa_n)_{n \in \N}$ is a Cauchy sequence with respect to the norm $\norm{h} \defeq \norm{\int_G h^\ast(g) h(g) \,dg}^{1/2}$ on $\CC(G,A)$. Since the completion of $\CC(G,A)$ with respect to this norm is precisely $L^2(G,A)$, we can view square-integrable continuous functions as elements of $L^2(G,A)$. If $\xi, \eta \in \Hilm$, then we define the *coefficient function* $c_{\xi\eta} \colon G \to A$ by $$c_{\xi\eta}(g) \defeq \5{\gamma_g(\xi)} {\eta}_A \qquad \text{for all $g \in G$.}$$ In the special case $\Hilm = A$, we have $c_{ab}(g) \defeq \alpha_g(a)^\ast b$. We call $\xi \in \Hilm$ *square-integrable* iff the function $c_{\xi\eta}$ is square-integrable for all $\eta \in \Hilm$. The Hilbert module $\Hilm$ is called *square-integrable* iff the set of square-integrable elements is dense in $\Hilm$. A algebra $A$ is called *proper* iff it is square-integrable as a right Hilbert module. Let $A_+ \subset A$ be the cone of positive elements. We call $a \in A_+$ *integrable* iff $a^{1/2}$ is square-integrable. By definition, $a \in A_+$ is integrable iff the integrals $\int_{K_n} b^\ast \alpha_g(a) b \,dg$ form a Cauchy sequence with respect to the norm topology for all $b \in A$. Moreover, $a \in A$ is square-integrable iff $a a^\ast$ is integrable. Hence $A$ is proper iff the set of integrable elements is dense in $A_+$. The above definition of properness is equivalent to Rieffel’s definition in [@rieffel:pre1] and thus also to Exel’s definition in [@exel:MoritaSpectral]. \[lem:square\_integrable\] Let $\Hilm$ be a Hilbert module and let $\xi, \eta, \zeta \in \Hilm$. (i) If $\xi$ is square-integrable, then the map $$\Gamma_\xi \colon \Hilm \to L^2(G,A), \qquad \Gamma_\xi(\eta) \defeq c_{\xi\eta},$$ is adjointable. The adjoint $\Gamma_\xi^\ast \colon L^2 (G,A) \to \Hilm$ satisfies $$\label{eq:gamma_ast_def} \Gamma_\xi^\ast(f) \defeq \int_G \gamma_g(\xi) \cdot f(g) \,dg \qquad \text{for all $f \in \CC(G, A)$.}$$ (ii) The operators $\Gamma_\xi$ and $\Gamma_\xi^\ast$ are equivariant. (iii) The closure of the range of $\Gamma_\xi^\ast$ is the smallest invariant Hilbert submodule of $\Hilm$ containing $\xi$. In particular, $\xi$ is contained in the closure of $\Ran \Gamma_\xi^\ast$. (iv) If $\xi$ and $\zeta$ are square-integrable, then the sequence $$\int_G \gamma_g(\xi) \cdot \5{\gamma_g(\zeta)} {\eta}_A \,\kappa_n(g) \,dg, \qquad n \in \N,$$ in $\Hilm$ is norm convergent. Its limit is $\Gamma_\xi^\ast \Gamma_\zeta (\eta)$. (v) $\xi$ is square-integrable iff $\ket{\xi} \bra{\xi} \in \Comp(\Hilm)$ is integrable. The Banach-Steinhaus theorem yields that $\Gamma_\xi$ is bounded. We can define an operator $\Gamma_\xi^\ast \colon \CC(G,A) \to \Hilm$ by . For $f \in \CC(G,A)$, we compute $$\label{eq:Txi_adjoint} \5{\Gamma_\xi^\ast(f)}{\eta}_A = \int_G \5{\gamma_g(\xi) f(g)}{\eta}_A \,dg = \int_G f(g)^\ast \cdot c_{\xi\eta}(g) \,dg = \5{f}{\Gamma_\xi (\eta)}_A.$$ Hence $\norm{ \5{\Gamma_\xi^\ast(f)}{\eta}_A } \le \norm{f}_{L^2(G,A)} \norm{\eta} \cdot \norm{\Gamma_\xi}$. Since $\eta$ is arbitrary, it follows that $\norm{\Gamma_\xi^\ast(f)} \le \norm{f}_{L^2(G,A)} \cdot \norm{\Gamma_\xi}$. Thus we may extend $\Gamma_\xi^\ast$ to $L^2(G,A)$. Equation  shows that $\Gamma_\xi^\ast$ is adjoint to $\Gamma_\xi$. Straightforward computations show that $\Gamma_\xi$ and $\Gamma_\xi^\ast$ are equivariant. Assertion (iii) follows easily once we know that $\xi$ is contained in the closed range of $\Gamma_\xi^\ast$. Choose $\epsilon >0$. There is $u \in A$ with $0 \le u \le 1$ and $\norm{ \xi\cdot u - \xi} \le \epsilon/2$. There is a compact neighborhood $U$ of $1 \in G$ with $\norm{\gamma_g(\xi) - \xi} < \epsilon/2$ for all $g \in U$. Let $f \colon G \to \R_+$ be a continuous function with support $U$ and $\int_G f(g) \,dg = 1$. Then $\norm{\Gamma_\xi^\ast (f \otimes u) - \xi} \le \epsilon$. Hence $\xi$ is contained in the closure of $\Ran \Gamma_\xi^\ast$. We compute $$\label{eq:Gxi_Gxi_ad} \Gamma_\xi^\ast \Gamma_\zeta(\eta) = \Gamma_\xi^\ast (c_{\zeta\eta}) = \lim_{n\to \infty} \Gamma_\xi^\ast (c_{\zeta\eta}\kappa_n) = \lim_{n\to \infty} \int_G \gamma_g(\xi) \5{\gamma_g(\zeta)}{\eta}_A \, \kappa_n(g) \,dg.$$ The boundedness of $\Gamma_\xi^\ast$ implies that the sequence is norm convergent. Equation  implies that the sequence $$I_n \defeq \int_G \ket{\gamma_g(\xi)} \bra{\gamma_g(\xi)} \, \kappa_n(g) \,dg = \int_G \gamma_g(\ket{\xi} \bra{\xi}) \, \kappa_n(g) \,dg \in \Comp(\Hilm)$$ is bounded and converges strongly (that is, pointwise on $\Hilm$) towards $\Gamma_\xi^\ast \Gamma_\xi$. Therefore, the sequences $(I_n \cdot T)$ and $(T \cdot I_n)$ converge in norm for all $T \in \Comp(\Hilm)$. This means that $\ket{\xi} \bra{\xi} \in \Comp(\Hilm)$ is integrable. Conversely, if $\ket{\xi} \bra{\xi} \in \Comp(\Hilm)$ is integrable, then the sequence $\5{\eta} {I_n(\eta)}_A$ is norm convergent for all $\eta \in \Comp(\Hilm) \cdot \Hilm = \Hilm$. Since $$\5{\eta} {I_n(\eta)}_A = \int_G \5{\gamma_g(\xi)}{\eta}_A^\ast \5{\gamma_g(\xi)}{\eta}_A \,\kappa_n(g) \,dg = \int_G c_{\xi\eta}(g)^\ast c_{\xi\eta}(g) \, \kappa_n(g) \,dg,$$ this means that $\xi$ is square-integrable. If $\Gamma_\xi^\ast \colon \CC(G,A) \to \Hilm$ extends to an *adjointable* map $L^2(G,A) \to \Hilm$, then $\xi$ is square-integrable. The map $\Gamma_\xi^\ast$ extends to a bounded operator $L^2(G,A) \to \Hilm$ if and only if the function $g \mapsto c_{\xi\eta}(g)^\ast c_{\xi\eta}(g)$ is *order-integrable* in the sense of Rieffel [@rieffel:pre1] for all $\eta \in \Hilm$. Hence it may happen that $\Gamma_\xi^\ast$ extends to a bounded operator on $L^2(G,A)$ that is not adjointable. \[pro:si\_proper\] Let $A$ be a algebra and let $\Hilm$ be a Hilbert module. Then $\Hilm$ is square-integrable iff $\Comp(\Hilm)$ is proper. If $\Hilm$ is square-integrable, then the linear span of the integrable elements is dense in $\Comp(\Hilm)$ by Lemma \[lem:square\_integrable\].(v). Therefore, $\Comp(\Hilm)$ is proper. Conversely, assume that $\Comp(\Hilm)$ is proper. Let $T \in \Comp(\Hilm)_+$ be square-integrable. If $\xi \in \Hilm$, then $\ket{T\xi} \bra{T\xi} = T \ket{\xi} \bra{\xi} T^\ast \le \norm{\xi}^2 TT^\ast$ is integrable because $TT^\ast$ is integrable and the set of integrable elements is a hereditary cone in $\Comp(\Hilm)_+$ [@rieffel:pre1]. Hence $T\xi \in \Hilm$ is square-integrable by Lemma \[lem:square\_integrable\].(v). Since $\Comp(\Hilm)$ is proper, the set of elements of $\Hilm$ of the form $T\xi$ with square-integrable $T \in \Comp(\Hilm)$ is dense in $\Hilm$. Thus $\Hilm$ is square-integrable. \[pro:proper\_functorial\] Let $A$ and $B$ be algebras, let $\Hilm$ be a Hilbert module, and let $\phi \colon A \to \Adj(\Hilm)$ be an essential equivariant homomorphism. If $A$ is proper, then $\Hilm$ is square-integrable. Identify $\Adj(\Hilm) \cong \Mult\bigl( \Comp(\Hilm) \bigr)$. By [@rieffel:pre1 Theorem 5.3], we conclude that $\Comp(\Hilm)$ is proper. Thus $\Hilm$ is square-integrable by Proposition \[pro:si\_proper\]. \[the:square\_integrable\] Let $A$ be a algebra and let $\Hilm$ be a countably generated Hilbert module. Then the following assertions are equivalent: (i) $\Hilm$ is square-integrable; (ii) $\Comp(\Hilm)$ is proper; (iii) there is a equivariant unitary $\Hilm \oplus \Hilsu_A \cong \Hilsu_A$; (iv) $\Hilm$ is a direct summand of $\Hilsu_A$. Proposition \[pro:si\_proper\] asserts that (i) and (ii) are equivalent. It is trivial that (iii) implies (iv). It remains to show that (iv) implies (i) and that (i) implies (iii). We prove that (iv) implies (i). It is straightforward to show that $\Comp(L^2G)$ is a proper algebra. Equivalently, $L^2G$ is a square-integrable Hilbert space. Let $\Hilm[F]$ be an arbitrary Hilbert module. The canonical homomorphism $\Comp(L^2G) \to \Adj(L^2G \otimes \Hilm[F])$, $T \mapsto T \otimes 1$, is essential and equivariant. Hence $L^2(G, \Hilm[F])$ is square-integrable by Proposition \[pro:proper\_functorial\]. Especially, $\Hilsu_A$ is square-integrable. A direct summand of a square-integrable Hilbert module is square-integrable as well because the projection onto the direct summand maps square-integrable elements to square-integrable elements. Hence any direct summand of $\Hilsu_A$ is square-integrable. That is, (iv) implies (i). The proof that (i) implies (iii) is very similar to the proof of the stabilization theorem by Mingo and Phillips [@mingo-phillips:triviality]. Suppose that $\Hilm$ is square-integrable. Hence there is a sequence $(\xi_n)_{n \in \N}$ of square-integrable elements of $\Hilm$, whose linear span is dense in $\Hilm$. Let $\Gamma_n \defeq \Gamma_{\xi_n}$ be as in Lemma \[lem:square\_integrable\]. We may assume that $\norm{\Gamma_n} \le 1$ for all $n \in \N$ and that each $\xi_n$ is repeated infinitely often. An element of $\Hilsu_A$ can be viewed as a sequence $(f_n)_{n \in \N}$ with $f_n \in L^2(G,A)$. We formally write $\sum f_n \delta_n$ for this sequence. Define an adjointable operator $T \colon \Hilsu_A \to \Hilm \oplus \Hilsu_A$ by $$\begin{gathered} T\biggl( \sum_{n=1}^\infty f_n \delta_n \biggr) \defeq \sum_{n=1}^\infty 2^{-n} \Gamma_n^\ast(f_n) \oplus \sum_{n=1}^\infty 4^{-n} f_n \delta_n, \\ T^\ast|_{\Hilm} (\eta) \defeq \sum_{n=1}^\infty 2^{-n} \Gamma_n(\eta) \delta_n, \qquad T^\ast|_{\Hilsu_A} \biggl( \sum_{n=1}^\infty f_n \delta_n \biggr) \defeq \biggl( \sum_{n=1}^\infty 4^{-n} f_n \delta_n \biggr). \end{gathered}$$ Lemma \[lem:square\_integrable\].(ii) implies that $T$ is equivariant. Evidently, $T^\ast$ has dense range. We claim that $T$ has dense range as well. Let $\Hilm[F] \subset \Hilm \oplus \Hilsu_A$ be the closure of the range of $T$. Let $f \in L^2(G,A)$. Since each $\Gamma_n^\ast$ is repeated infinitely often, we have $\Gamma_n^\ast(f) \oplus 2^{-k}f \delta_k \in \Ran T$ for infinitely many $k \in \N$. Hence $\Gamma_n^\ast(f) \oplus 0 \in \Hilm[F]$ for all $f \in L^2(G,A)$. By Lemma \[lem:square\_integrable\].(iii), this implies $\xi_n \oplus 0 \in \Hilm[F]$ for all $n$ and hence $\Hilm \subset \Hilm[F]$. Finally, we get $0 \oplus f \delta_n \in \Hilm[F]$ for all $f \in L^2(G,A)$ and thus $\Hilm[F] = \Hilm \oplus \Hilsu_A$. Since both $T$ and $T^\ast$ have dense range, the composition $T^\ast T$ has dense range. Thus $\abs{T} \defeq (T^\ast T)^{1/2}$ has dense range because $\abs{T} (\Hilm) \supset \abs{T} (\abs{T} \Hilm) = T^\ast T (\Hilm)$. Since $\5{\abs{T}\eta}{\abs{T}\eta}_A = \5{T^\ast T\eta}{\eta}_A = \5{T\eta}{T\eta}_A$, the formula $U(\abs{T} \eta) \defeq T\eta$ well-defines an isometry $U$ from $\Ran \abs{T}$ onto $\Ran T$. Extending $U$ continuously, we obtain the desired unitary $U \colon \Hilsu_A \to \Hilm \oplus \Hilsu_A$. Since $T$ is equivariant, so is $U$. Thomsen [@thomsen:extensions] calls a algebra $A$ *proper* iff $\Hilm \oplus \Hilsu_A \cong \Hilsu_A$ for all Hilbert bimodules $\Hilm$. Theorem \[the:square\_integrable\] implies that $A$ is proper in Thomsen’s sense iff all algebras that are Morita-Rieffel equivalent to $A$ are proper in our sense. For instance, the algebra $\Comp(G)$ is not proper, unless $G$ is compact, because it is Morita-Rieffel equivalent to the improper algebra $\C$. \[pro:proper\_aep\] All unital proper algebras have property AE. Let $A$ and $B$ be unital algebras and let $(\Hilm, \phi, F)$ be an essential Kasparov triple for $A,B$. Suppose that $A$ is proper. Proposition \[pro:proper\_functorial\] implies that $\Hilm$ is square-integrable. Hence $\Hilm \subset \Hilsg_B$ by Theorem \[the:square\_integrable\]. Let $\Hilsg_A^\ast$ be the imprimitivity bimodule dual to $\Hilsg_A$. As remarked in Section \[sec:imprimitivity\], we have $\Hilsg_A^\ast = \Comp(\Hilsg_A, A)$ with a canonical Hilbert bimodule structure. Since $A$ is proper and unital, Theorem \[the:square\_integrable\] yields a equivariant, adjointable isometry $T \colon A \to \Hilsg_A$. Composition with $T$ gives rise to an adjointable isometry $T_\ast \colon \Comp(\Hilsg_A, A) \to \Comp(\Hilsg_A, \Hilsg_A)$. Thus $\Hilsg_A^\ast$ is a direct summand in $\Comp(\Hilsg_A)$. Lemma \[lem:equi\_connection\] yields a equivariant connection $\bar{F}$ on $L^2(G_2, \Hilm)^\infty = \Hilsg_A \hot_\phi \Hilm$. If we view $\bar{F}$ as an operator on $\Comp(\Hilsg_A) \hot_{\Comp(\Hilsg_A)} L^2(G_2, \Hilm)^\infty \cong L^2(G_2, \Hilm)^\infty $, then we obtain an connection. Hence the compression $$F' \defeq (T_\ast \hot_{\Comp(\Hilsg_A)} 1)^\ast \cdot \bar{F} \cdot (T_\ast \hot_{\Comp(\Hilsg_A)} 1)$$ of $\bar{F}$ to $\Hilsg_A^\ast \hot_{\Comp(\Hilsg_A)} \Hilsg_A \hot_\phi \Hilm \cong A \hot_\phi \Hilm \cong \Hilm$ is an connection as well. By [@blackadar98:Ktheory 18.3.4.f], $F'$ is an connection. Another connection is $F$ itself. Therefore, $F'$ is a compact perturbation of $F$. Since $\bar{F}$ and $T$ are equivariant, so is $F'$. \[the:proper\_KK\] Let $A$ and $B$ be unital algebras. If $A$ is proper, then $$KK^G (A,B) \cong KK^G_s (A,B) \cong [\littleX A, \Comp(G_2\N) B] \cong [\littleX A, B]_{\Comp(G_2\N)}.$$ If $A$ is proper and $A$ and $B$ are trivially graded, then $$KK^G (A,B) \cong KK^G_s (A,B) \cong [\littleQ A, \Comp(G\N) B] \cong [\littleQ A, B]_{\Comp(G\N)}.$$ By Proposition \[pro:proper\_aep\], $A$ has property AE. Hence the assertions follow from Proposition \[pro:special\_aep\] and Proposition \[pro:KKGs\_XQ\]. Other equivariant theories {#sec:generalize} ========================== The arguments above generalize to other more general versions of equivariant Kasparov theory. First of all, we did not even care to specify whether we work with complex or real algebras: The theory above goes through in both cases. In the real case, we only have to interpret $\C$ as the algebra $\R$ of real numbers everywhere above. Hence $KK^G$ is the universal split exact stable homotopy functor also for real algebras. We may also treat “real” algebras as in [@kasparov:KK]. Our results carry over to Kasparov’s functor $\RKK^G$ with some obvious changes. We first define the functors $\RKK^G$ and $RKK^G$. Let $G$ be, as usual, a locally compact compact topological group and let $X$ be a locally compact compact space. An *algebra* is a algebra $A$ equipped with a equivariant *essential* homomorphism from $\CVI(X)$ into the *center* of $\Mult(A)$. Let $A,B$ be algebras. A homomorphism $\phi \colon A \to \Mult(B)$ is called *equivariant* iff it is equivariant and in addition satisfies $\phi(f \cdot a) = f \cdot \phi(a)$ for all $f \in \CVI(X)$, $a \in A$. If $B$ is an algebra and $\Hilm$ is a Hilbert module, then $\Comp(\Hilm)$ is an algebra as well. The homomorphism $\CVI(X) \to \Mult\bigl( \Comp(\Hilm) \bigr) = \Adj(\Hilm)$ is defined by $f \cdot (\xi \cdot b) \defeq \xi \cdot (f \cdot b)$ for all $f \in \CVI(X)$, $\xi \in \Hilm$, $b \in B$. By the Cohen-Hewitt factorization theorem, all elements of $\Hilm$ are of the form $\xi \cdot b$ for suitable $\xi \in \Hilm$, $b \in B$. Computing the inner products $\5{f \cdot (\xi \cdot b)}{\eta \cdot c}$ shows that $f \cdot (\xi \cdot b)$ is well-defined and defines a homomorphism from $\CVI(X)$ into the center of $\Adj(\Hilm)$. Kasparov [@kasparov88:equivariantKK] defines the functor $\RKK^G(X; A,B)$ for algebras using Kasparov triples $(\Hilm, \phi, F)$ for $A,B$ with the additional assumption that $\phi \colon A \to \Adj(\Hilm)$ be equivariant. If $A$ and $B$ are just algebras, then he puts $RKK^G(X; A,B) \defeq \RKK^G\bigl(X; \CVI(X,A), \CVI(X,B) \bigr)$. Hence $RKK^G$ is a special case of $\RKK^G$. The presence of the central homomorphism $\CVI(X) \to \Adj(\Hilm)$ creates no problems in Sections \[sec:equi\_connections\] and \[sec:embed\_Hilbert\]. In Section \[sec:universal\], we have to modify the definitions of the universal algebras $\bigX A$ and $\bigQ A$ because, as they are defined there, they are not algebras. Thus we replace them by algebras with analogous universal properties in the category of algebras. This amounts to dividing out the relations $\iota^\pm(fa) \iota^\pm(b) = \iota^\pm(a) \iota^\pm(fb)$ for all $a,b \in A$, $f \in \CVI(X)$ in $\littleQ A$ and the relation $j_A(fa) F_A j_A(b) = j_A(a) F_A j_A(fb)$ for all $a,b \in A$, $f \in \CVI(X)$ in $\littleX A$. The quotients by the ideals generated by these relations carry a canonical algebra structure. Once this modification is made, the results of Sections \[sec:universal\], \[sec:Kasparov\_universal\], and \[sec:graded\] carry over without change. Especially, $\RKK^G(X; A,B)$ is the universal split exact stable homotopy functor for trivially graded separable algebras. Theorem \[the:proper\_KK\] remains valid as well. The functor $\RKK^G(X; A,B)$ is a special case of the equivariant Kasparov theory for groupoids developed in [@legall:KK_groupoids]. I expect that the arguments above carry over to the case of locally compact groupoids with Haar system. However, I have not checked the details. Some work has to be done to carry over the proof of Lemma \[lem:equi\_connection\]. Moreover, to carry over the results of Section \[sec:proper\], one first has to define properness in the sense of Rieffel for actions of groupoids.

--- abstract: 'Angle-resolved photoemission spectroscopy allows direct visualization and experimental determination of the electronic structure of crystals in the momentum space, including the precise characterization of the Fermi surface and the superconducting order parameter. It is thus particularly suited for investigating multiband systems such as the Fe-based superconductors. In this review, we cover several aspects of these recently discovered materials that have been addressed by this technique, with a special emphasis on their superconducting gap and their Fermi surface topology. We provide sufficient experimental evidence to support the reliability and the consistency of the angle-resolved photoemission spectroscopy measurements over a wide range of material compositions.' address: - '1 Beijing National Laboratory for Condensed Matter Physics, and Institute of Physics, Chinese Academy of Sciences, Beijing 100190, China' - '2 Department of Physics, Tohoku University, Sendai 980-8578, Japan' - '3 TRiP, Japan Science and Technology Agency (JST), Kawaguchi 332-0012, Japan' - '4 WPI Research Center, Advanced Institute for Materials Research, Tohoku University, Sendai 980-8577, Japan' author: - 'P. Richard$^1$, T. Sato$^{2,3}$, K. Nakayama$^2$, T. Takahashi$^{2,4}$, H. Ding$^1$' bibliography: - 'biblio\_en.bib' title: 'Fe-based superconductors: an angle-resolved photoemission spectroscopy perspective' --- Introduction ============ Struggling for more than twenty years with a problem that it could not resolve despite tremendous efforts, the scientific community received in 2008 the announcement of the discovery of superconductivity at 26 K in a Fe-layered material [@Kamihara] as a powerful electro-shock. The monopoly of copper oxides (cuprates) on high-$T_c$ superconductivity was over. And yet, the future proved to be full of further surprises to those who started racing once more for the keys leading to higher $T_c$’s. Less than 4 years later, an impressive amount of data has been collected on these materials, accompanied by significant progress in our understanding of their exotic properties. Nevertheless, a consensus has not been reached on most aspects and it becomes very important to organize the pieces of the puzzle to get a more effective overall view. Up to now, there has been several review papers dealing with experiments on Fe-based superconductivity [@JohnstonAdv_Phys2010; @GasparovJETP2010; @PaglioneNPhys2010; @StewartReview2011; @LumsdenJPhys2010; @MizuguchiJPSJ2010]. In the present paper, we do not intend to repeat the content of previous reviews, but rather to provide a different perspective based on a particular experimental technique, namely angle-resolved photoemission spectroscopy (ARPES). ARPES is a powerful tool to access *directly* the momentum-resolved electronic structure of crystals, and thus it occupies a unique position when trying to understand their electronic behaviour. In this review, we will cover some important aspects of the Fe-based superconductors that have been addressed by ARPES, with a special emphasis on the superconducting (SC) gap. The review is organized as follows. We first provide basic knowledge to understand simply what ARPES is and what are its advantages and limitations. In Section \[section\_structure\], we describe the electronic structure of the Fe-based superconductors, from core levels to the electronic states lying close to the Fermi level ($E_F$). The SC gap and its relationship to the electronic band structure will be discussed in Section \[section\_gap\]. We describe two models commonly used to explain superconductivity in these materials: the quasi-nesting model, which dominated the scene until only recently, and the local antiferromagnetic exchange pairing model, which now after the discovery of high-$T_c$ superconductivity in the 122-chalcogenides many consider a better candidate to unify the superconducting pairing mechanism in the Fe-based superconductors. Finally, before concluding this review, we will discuss the ARPES results in a broader context, where we compare ARPES with transport measurements. Angle-resolved photoemission spectroscopy {#section_ARPES} ========================================= Basic principles ---------------- The idea of photoemission spectroscopy, which is simply based on the conservation of energy, is to extract information on the electronic structure of crystals by measuring the kinetic energy of electrons ejected from the surface of a sample due to the interaction with a photon flux of known energy h$\nu$ and vector potential ${\mathbf}{A}$. ARPES is a sophisticated method among a variety of photoemission spectroscopies. In addition to the conservation of energy, ARPES takes advantage of the conservation of the in-plane momentum. Hence, the direction of emission of the photo-excited electrons is recorded as well as their energy. This procedure, illustrated in Figure \[ARPES\_process\], allows the determination of the electronic band dispersions. Although ARPES measurements require much more time than angle-integrated photoemission spectroscopy (PES), which provides information on the electronic density of states (DOS) only, they give access to a more comprehensive determination of the electric band structure. More specifically, the ARPES signal $I({\mathbf}{k},E,{\mathbf}{A},h\nu)$ is proportional to the one-particle spectral weight $A$(${\mathbf}{k}$, $E$), which is the probability to have an electron in the sample with momentum ${\mathbf}{k}$ and energy $E$, times the Fermi-Dirac distribution $f$($E$,$T$): $$I({\mathbf}{k},E,{\mathbf}{A},h\nu) = M({\mathbf}{k},E,{\mathbf}{A},h\nu)A({\mathbf}{k},E)f(E,T)$$ where $M$ represents the photoemission matrix element determined by the photoemission process itself and carries no direct information on the band dispersion. However, $M$ contains precious information on the nature of the electronic states probed. ![\[ARPES\_process\] (Colour online) Schematic diagram explaining ARPES measurements. Under a flux of photons of energy h$\nu$, electrons are ejected from the surface of a sample (orange slab) at an angle ($\theta$,$\varphi$) from the normal direction and detected with a high-resolution analyzer (in gray). The process implies the conservation of energy and in-plane momentum (${\mathbf}{k}_{||}$).](ARPES_process_low){width="8cm"} The ARPES experimental data can be represented by colour images, where the colour scale gives the photoemission intensity, which is generally displayed as a function of energy and momentum. The energy is usually referenced to the leading edge of a metallic film spectrum, typically of gold, in electric contact with the sample. Although colour images give a more intuitive and natural representation of the electronic dispersion, they fail to reveal the precise spectral lineshape. For this reason, it is also convenient to represent the ARPES data with a series of curves of photoemission intensity as a function of energy at a fixed momentum $k$, or energy distribution curves (EDCs). Alternatively, one can display a series of curves of photoemission intensity as a function of momentum at a fixed energy $E$, or momentum distribution curves (MDCs). Main advantages and limitations {#advantages_limitations} ------------------------------- As with any experimental techniques, ARPES has advantages and limitations. Its main advantage is quite obvious: it allows a *direct* determination of the electronic structure of materials, including the Fermi surface and the band dispersion in the momentum space. This is crucially important when studying multi-band systems like the Fe-based superconductors. The experimental information is irreplaceable since it is not obtained from a fitting procedure of several parameters on various models, which may turn out to be inadequate. Moreover, ARPES is suitable for relatively “dirty" systems such as the copper oxides or the Fe-based superconductors, where doping is introduced through partial chemical substitution or the incorporation of interstitial dopants and vacancies. Such doping process induces disorder, making determination of Fermi surfaces from quantum oscillation measurements difficult or even impossible under low magnetic field. In contrast to transport techniques, ARPES is not limited to the Fermi surface, and it is sensitive to states away from the Fermi level, which is very useful when the physical properties are partly determined by high-energy states. Moreover, ARPES can be used to determine the orbital characters of the electronic states by taking advantage of selection rules involved in the photoemission process. The main disadvantage of ARPES, which can be also regarded as an advantage when dealing with surface phenomena as in the topological insulators, is its surface sensitivity. To avoid surface contamination, single-crystals of complex compounds are usually cleaved *in situ* and measured in a vacuum better than 10$^{-9}$ Torr, the finite lifetime of the samples increasing as vacuum improves. Hence, the lifetime of the samples can vary from a few hours to a few days. To maximize the sample lifetime, vacuum in the 10$^{-11}$ Torr range are routinely achieved using modern ultra-high vacuum systems such as in the Institute of Physics of the Chinese Academy of Sciences or in Tohoku University. Strictly speaking, ARPES measures the surface of materials, not the bulk. Nevertheless, the surface is *always* to some extend related to the bulk of materials. In other words, *surface* = *bulk* + $\delta$. The reliability of the ARPES measurements when interpreting the physical properties of the bulk is thus determined by the size of $\delta$, which is compound-dependent. There are several ways to check if $\delta$ is qualitatively small or large. $\delta$ can be considered small when: 1. Low energy electron diffraction (LEED) pictures do not show obvious surface reconstruction; 2. The surface carrier doping, as determined from the Luttinger theorem, is consistent with that of the bulk; 3. The Fermi surface evolves smoothly with doping. 4. The electronic structure (band dispersion, gap size, etc...) varies with $k_z$, in sharp contrast to pure surface states; 5. The SC gap observed by ARPES closes at the bulk $T_c$; 6. The core levels of the relevant elements are not doubled; 7. The band dispersions are similar, albeit for some renormalization, to LDA predictions; 8. No unexpected band folding is observed by ARPES; 9. Gaps measured by ARPES are consistent with gaps measured from bulk-sensitive probes. It is to note that ARPES bulk-sensitivity is highly enhanced by the use of very low photon energies ($h\nu<9$ eV) [@KissPRL2005; @SoumaRSI2007] or high photon energies ($h\nu$ &gt; 500 eV) [@KamakuraEPL2004]. Fortunately, several Fe-based superconductors fulfill these requirements. Although the influence of the surface cannot be predicted systematically, the crystal structure symmetry of layered systems guides us to determine which compound is susceptible to present a surface problem. When the cleaving plane is unique and occurs between two symmetric layers, the two cleaved pieces are perfectly symmetric. The surface is thus usually non-polar and surface effects are expected to be minimal. For example, that case corresponds to the 11 and 111 structural phases of Fe-based superconductors. On the other hand, when the structure presents more than one equivalent cleaving planes leading to more than one possible surface exposed, the surface may be highly polar. The surface doping in this case is not representative of the bulk. YBa$_2$Cu$_3$O$_{7-\delta}$ is one of the most famous structures belonging to this category, in which the 1111-pnictide structure falls as well. Nevertheless, ARPES can be used in this case to get a reasonable idea of the band structure, exact doping and $k_z$ dispersion put aside. Another important situation is the one when cleaving occurs exactly at a symmetry plane occupied by one layer of atoms. To account for electrostatic stability, only half of the atoms remain on the cleaved surface. The surface may reconstruct or not. The experiment alone and the use of the criteria described above can determine whether the experimental results are representative of the bulk for the electronic states of interest. A special attention need to be given to the 122 family of Fe-pnictides, which belongs to this category. Surface effects have been reported in SrFe$_2$As$_2$ [@Hsieh]. Interestingly, such effect is not observed in hole-doped Ba$_{1-x}$K$_x$Fe$_2$As$_2$, which satisfies the criteria enumerated above, in agreement with a LEED and STM study reporting the absence of surface reconstruction in BaFe$_2$As$_2$ [@Nascimento_PRL2009]. Furthermore, as shown below, the 122-pnictides share similar electronic structures and SC gap functions with other Fe-based superconductors with ideal cleaving planes, suggesting that even supposing that a small distortion occurs in the (Ba, K) layer at the cleaved surface, the effect on the Fe electronic states are negligible. Brillouin zone notation ----------------------- Unfortunately, the notation used in ARPES studies of Fe-based superconductors to describe high-symmetry points in the momentum space is not homogeneous, and two main representations are usually found in the literature. In the first one, only the Fe atoms are considered. The Fe atoms form a square lattice of parameter $a$, which corresponds to a square Brillouin zone (BZ) of size $2\pi/a$. In this *1 Fe/unit cell* or *unreconstructed* description, the zone center $\Gamma$ and the M point correspond to $(0,0)$ and $(\pi/a,0)$, respectively, while the X point is defined as $(\pi/2,\pi/2)$. The $\Gamma$-M direction is associated to momentum along the Fe-Fe bounding. Alternatively, one can prefer to consider the “real" BZ and take into account the effect of the alternative distribution of chalcogen and pnictogen atoms up and down the Fe plane. The unit cell parameter becomes $a'=a\sqrt{2}$. In this *2 Fe/unit cell* or *reconstructed* description, M = $(\pi/a',\pi/a')$ and X = $(\pi/a',0)$. To add to the confusion, some authors rather define M = $(\pi/a',0)$ and X = $(\pi/a',\pi/a')$. Unless specifically indicated, we use the 1 Fe/unit cell representation by default in the present review. Electronic structure {#section_structure} ==================== The electronic properties of a material are governed by its own electronic structure, which is determined by the composition and the arrangement of the atoms from which it is made. All Fe-based superconductors are characterized by the presence of Fe layers. It is thus quite natural that they share similar electronic structure. Since it has been studied in more details, we begin this chapter by presenting ARPES results on optimally-doped [Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$]{}as a typical Fe-based superconductor. We will then review ARPES results obtained on several other Fe-based materials. Optimally-doped Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$ ----------------------------------------------- The core level spectrum displayed in Figure \[core\_levels\_OP\](a) is consistent with the chemical composition of [Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$]{}. From high to low binding energies, we can distinguish unambiguously core levels associated with Fe$3p$ (52.4, 53.0 eV), As$3d$ (40.4, 41.1 eV), K$3s$ (33.0 eV), Ba$5s$ (29.7 eV), K$3p$ (17.8 eV) and Ba$5p$ (14.2, 16.2 eV) [@Ding_J_Phys_Condens_Matter2011]. We also observe a weak peak around 12 eV, whose origin remains unclear. Although it could come from As$4s$ states, we note that several divalent Fe compounds, such as FeO, exhibit a peak at similar energy attributed to a satellite state of the Fe$3d^5$ configuration [@Lad_PRB1989]. The strongest core level peaks are those corresponding to the As$3d_{5/2}$ and As$3d_{3/2}$ levels. Although As atoms are located just below the cleaved surface, it is worth mentioning that unlike GaAs, for which a set of surface components are observed 0.4 eV below the peak energies, the As $3d$ levels in [Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$]{}do not show obvious additional component that could result from a surface reconstruction. It is thus reasonable to assume that the electronic states of Fe atoms, which are located below the first As layer, are not affected significantly by the surface termination. ![\[core\_levels\_OP\] (Colour online) (a) Wide range EDC near $\Gamma$ showing shallow core levels marked by vertical bars above the $x$-axis. The inset magnifies the valence band and a possible satellite peak at $\sim$ 12 eV, and highlights the difference between spectra taken at 100 and 21.2 eV. (b) Valence band near $\Gamma$ measured at different photon energies (46 - 66 eV). All EDCs are normalized by the photon flux. (c) Intensity plot of second derivatives of spectra along $\Gamma$-X and X-M. LDA bands (red lines) are plotted for comparison. (d) Photon energy dependence of the EDC intensity shown in (b) obtained at binding energies 0.1, 7, and 12 eV. From Ding *et al.*, J. Phys.: Condens. Matter., **23**, 135501 (2011) [@Ding_J_Phys_Condens_Matter2011], copyright ©(2011) by IOP Publishing.](Core_levelsOP_low){width="8cm"} The nature of the electronic states within about 8 eV below $E_F$ can be determined by investigating their photon energy dependence. The inset of Figure \[core\_levels\_OP\](a) compares the spectra of [Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$]{}at 21.2 eV and 100 eV. While the former photon energy is more sensitive to As$4s$ states, the latter enhances Fe$3d$ states. We thus conclude that the sharp peak within 1 eV below $E_F$ originates mainly from Fe$3d$ states and the states at higher biding energy mainly from As$4s$ orbitals, in agreement with LDA band calculations [@Ma_Front_phys_China]. This assessment is reinforced by the detailed photon energy evolution of the low energy states across the Fe$3p\rightarrow$Fe$3d$ resonance at 56 eV, which is illustrated in Figure \[core\_levels\_OP\](b). The difference of the spectra recorded at 56 eV (at the resonance) and at 52 eV (below the resonance) exhibits two features: a sharp peak near $E_F$ and a broad peak centered around 7 eV, which can be regarded as the coherent and incoherent parts of the Fe$3d$ states, respectively. While the photoemission intensity near $E_F$ shows an anti-resonance profile, the incoherent part at 7 eV has a Fano-like resonance profile (see Figure \[core\_levels\_OP\](d)), which is similar to the Fe$3d$ states in FeO [@Lad_PRB1989]. The Fe$3d$ and As$4s$ levels within 8 eV below $E_F$ are dispersive, as illustrated in Figure \[core\_levels\_OP\](c) by the intensity plot of second derivative, for the $\Gamma$-M and $\Gamma$-X high symmetry lines. When renormalized by a factor 2, LDA band calculations [@G_Xu_EPL2008] at this doping capture some overall features observed experimentally. This indicates that correlations in this material are not negligible. Interestingly, band renormalization seems more important at lower binding energy, as discussed below. We now turn our attention to the states closer to the Fermi energy, which mainly govern the electronic properties of materials. The Fermi surface of [Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$]{}recorded with the He I$\alpha$ resonance line is displayed in Figure \[FS\_OP\](a) [@Ding_EPL], and a schematic version is given in Figure \[FS\_OP\](b). It illustrates the multi-band nature of the Fe-based superconductors. ARPES intensity plots along cuts passing through or close to the $\Gamma$ point allow us to identify at least two holelike bands crossing $E_F$. Hereafter, we call $\alpha$ and $\beta$ the bands giving rise to the small and large $\Gamma$-centered FS pockets, respectively. These band are quite clear in the intensity plot given in Figure \[FS\_OP\](c) (below $T_c$). Similarly, the ARPES intensity plots along cuts passing near the M point reveal two distinct electronlike FS pockets, $\gamma$ and $\delta$, as shown in Figure \[FS\_OP\](d). Their origin can be understood from the hybridization of ellipses elongated towards the $\Gamma$-M direction and folded across the $\Gamma$-X plane due to the BZ reconstruction induced by the alternative positions of As atoms below and above the Fe layer. ![\[FS\_OP\] (Colour online) (a) FS contour determined by plotting the ARPES spectral intensity integrated within $\pm 10$ meV with respect to $E_F$. (b) Schematic view of the four FS sheets with a definition of the FS angle ($\theta$). (c) Intensity plot near $\Gamma$ measured at $T=15$ K. (d) Intensity plot near M measured at 15 K. Dots are EDC peak positions. (e) Second derivative plot of the dispersion along Z-R ($k_z\pi$) measured using 32 eV photons. Three hole-like bands ($\alpha$ (inner), $\alpha$’ (middle) and $\beta$ (outer)) are observed. (a) is from Ding *et al.*, Europhy. Lett., **83**, 47001 (2008) [@Ding_EPL], copyright ©(2008) by the European Physical Society. (b) is from Nakayama *et al.*, Europhy. Lett., **85**, 67002 (2009) [@Nakayama_EPL2009], copyright ©(2009) by the European Physical Society. (c) and (d) are from Ding *et al.*, J. Phys.: Condens. Matter., **23**, 135501 (2011) [@Ding_J_Phys_Condens_Matter2011], copyright ©(2011) by IOP Publishing. (e) is from Xu *et al.*, Nature Phys., **7**, 198 (2011) [@YM_Xu_NPhys2011], copyright ©(2011) by Macmillan Publishers Ltd.](FS_OP_low){width="8cm"} Although very high energy resolution can be achieved in ARPES measurements performed with a He discharging lamp, the results are confined to a single $k_z$ value. Using the tunability of synchrotron radiation, ARPES experiments can reveal a small warping in the band dispersion along $k_z$ [@YM_Xu_NPhys2011; @Y_Zhang_PRL2010]. More importantly, an additional holelike band, which is almost degenerate with the $\alpha$ band at the zone center, is resolved as $k_z$ increases towards Z $=(0,0,\pi/c)$ (where $c=6.5$ Åis the lattice parameter of the primitive unit cell and the distance between two Fe layers), as illustrated in Figure \[FS\_OP\](e). The synchrotron data indicate that He I$\alpha$ radiation corresponds approximately to $k_z=0$, explaining why this third band was not observed in the earliest ARPES experiments. The results obtained at different values of $k_z$ have several significant consequences: (i) They are consistent with LDA band calculations, which all predict 3 $\Gamma$-centered holelike bands and 2 M-centered electronlike bands; (2) They also confirm that the surface of [Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$]{}does not carry extra charge, unlike [YBa$_2$Cu$_3$O$_{7-\delta}$]{}. Indeed, assuming a double degeneracy for the $\alpha$ band, one can show that the algebraic FS area satisfies the Luttinger theorem with the bulk doping [@Ding_EPL; @Nakayama_EPL2009]; (iii) Finally, the $k_z$ warping of the electronic structure proves that the electronic states measured by ARPES *cannot* be pure surface states and that ARPES captures at least the essence of the bulk electronic properties. It is interesting to note that the Fermi velocities determined from ARPES differ from the values predicted by LDA calculations by a factor greater than the renormalized factor of 2 obtained for the overall band structure, indicating that low-energy states are further renormalized [@Ding_J_Phys_Condens_Matter2011]. Interestingly, such observation has also been done for other strongly correlated electron systems like the cuprates [@Valla_Science1999] and the cobaltates [@HB_YangPRL2005]. For practical purposes, the band structure can be parametrized using a tight-binding-like model [@Ding_J_Phys_Condens_Matter2011]: $$E^{\alpha,\beta}(k_x,k_y)= E_0^{\alpha,\beta}+t_1^{\alpha,\beta}(\cos k_x+\cos k_y)+t_2^{\alpha,\beta}\cos k_x\cos k_y$$ and $$E^{\gamma,\delta}(k_x,k_y)= E_0^{\gamma,\delta}+t_1^{\gamma,\delta}(\cos k_x+\cos k_y)+t_2^{\gamma,\delta}\cos k_x/2\cos k_y/2$$ Since ARPES can reveal the unoccupied part of the band structure for only a few $k_BT$’s, mild constrains must be used for the fit. The results obtained with this model allow an estimation of the effective masses and reveal a total renormalization factor of about 3-4 for the near-$E_F$ bands, in addition of providing essential parameters to calculate some thermodynamical parameters such as the Sommerfeld coefficient [@Ding_J_Phys_Condens_Matter2011]. [@ccccc]{} &$\alpha$&$\beta$&$\gamma$&$\delta$\ $E_0$&-0.24&-0.025&0.7&0.7\ $t_1$&0.16&0.013&0.38&0.38\ $t_2$&-0.052&0.042&0.8&-0.8\ Magnetic parent compounds ------------------------- The parent compounds of both the cuprates and the Fe-based superconductors have magnetically ordered states, with superconductivity emerging away from the magnetic ordered phase. The study of these materials is therefore of prior interest. Unlike the cuprates though, the parent compounds of the Fe-based superconductors are generally metallic, which is an important advantage for ARPES studies. Hence, ARPES investigations have been reported on the parent compounds of the 122-pnictide [@Hsieh; @VilmercatiPRB2009; @LiuPRL2009; @RichardPRL2010; @FinkPRB2009; @Malaeb_JPSJ2009; @M_YiPRB2009_2; @LX_YangPRL2009; @GD_LiuPRB2009; @Brouet_PRB2009; @Kondo_PRB2010; @M_YiPNAS2011; @Y_KimPRB2011], 11-chalcogenide [@Y_XiaPRL2009; @Y_Zhang_PRB2010] and 111-pnictide [@C_HePRL2010] systems. ![\[fig\_parent\] (Colour online) (a) FS mapping (25 K) obtained by integrating the photoemission intensity in a 20 meV window centered at $E_F$. The FS is described in terms of the unreconstructed BZ. The double arrow indicates light polarization. (b) Corresponding second derivative intensity plot. $\widetilde{Q}_{SDW}$ is the in-plane projection of the SDW wave vector. From Richard *et al.*, Phys. Rev. Lett., **104**, 137001 (2010) [@RichardPRL2010], copyright ©(2010) by the American Physical Society.](fig_parent_low){width="8cm"} Magnetic ordering in the parent compounds is accompanied by a reduction of the BZ and subsequent band folding. The relationship between the unreconstructed, reconstructed and 2D magnetic BZ in the 122 phase is described in Figure \[fig\_parent\](a), which also shows the Fermi surface of BaFe$_2$As$_2$ well below the SDW transition $T_{{SDW}}=138$ K. The intensity mapping indicates Fermi surfaces of equivalent sizes around the $\Gamma$ and M points. The FS contours are better represented by the second derivative of the intensity map shown in Figure \[fig\_parent\](b), which reveals almost identical patterns around $\Gamma$ and M, confirming band folding across the magnetic BZ boundaries. However, these patterns are quite unusual. In particular, strong photoemission intensity spots are observed near $\Lambda=(0, 0.75)$ and equivalent symmetry points, away from high symmetry points [@RichardPRL2010]. As temperature increases, these high intensity spots remain quite clear below $T_{{SDW}}$. Above that critical temperature, the spots are hard to identify, suggesting that they originate from the magnetic state. Interestingly, the intensity spots at the $\Lambda$ point cannot emerge from a single band. A careful analysis shows that two bands cross the Fermi level around the $\Lambda$ point. Even more surprising is the almost conical band structure at that particular momentum point, which cannot be produced by simple band folding. Using the broadening of the Fermi cutoff at higher temperature, one can access states above $E_F$ within a range of a few $k_BT$’s. The spectra recorded at 100 K, which has been divided by the Fermi function convoluted by the energy resolution function, is displayed in Figure \[fig\_cone\](a). Along with its second derivative intensity plot \[Figure \[fig\_cone\](b)\], it reveals a X-like pattern rather than a simple $\Lambda$-like one. This strongly suggests the presence of a Dirac cone [@RichardPRL2010], which is further confirmed by the absence of hybridization gap at the cone apex, as illustrated by the EDC’s in Figure \[fig\_cone\](c). ![\[fig\_cone\] (Colour online) (a) ARPES intensity spectra of BaFe$_2$As$_2$ at the $\Lambda$ point ($\theta\sim 90^{\circ}$) recorded at 100 K, after division by the Fermi-Dirac function convoluted with the instrumental resolution function. (b) Corresponding second derivative intensity plot. Long dashed lines are guides for the eye. The corresponding EDCs are given in (c), where the bold EDC refers to the $\Lambda$ point. (d) Polar representation of $v_F$ around the $\Lambda$ point (25 K). Open and close circles represent data measured and data obtained by reflection with respect to the $\Gamma$-M symmetry line, respectively. The large filled circle represents the average value of $v_F$ while the thick line is a fit of the data to the two-ellipse model described in the text, with parameters $a$, $b$ and $c$. (e) Contour plot of the electronic dispersion below $E_F$ around the $\Lambda$ point, as calculated from our model. The small filled circle represents the FS associated with the cone. From Richard *et al.*, Phys. Rev. Lett., **104**, 137001 (2010) [@RichardPRL2010], copyright ©(2010) by the American Physical Society.](fig_cone_low){width="8cm"} As shown in Figure \[fig\_cone\](d), the analysis of the low-temperature band structure of BaFe$_2$As$_2$ suggests that the Fermi velocity $v_F$ varies slightly around the Dirac cone, from 290 $\cdot$ Åto 360 meV$\cdot$ Å, with an average of 330 $\pm$ 60 meV$\cdot$Å, which is compatible with the symmetry of the $\Lambda$ point [@RichardPRL2010]. The apex is found at 1 $\pm$ 5 meV above $E_F$, implying a very small hole Fermi surface pocket covering only around 10$^{-3}$% of the first Brillouin zone, as suggested by Figure \[fig\_cone\](e). Interestingly, bulk-sensitive quantum oscillation measurements led to similar results for the Fermi velocity [@Harrison], as well as for the presence of small Fermi surface pockets [@Sebastian; @Analytis]. The access to $v_F$ around the Dirac cone allows a 3D reconstruction of the Dirac cone, which is given in Figure \[cone3D\](a). Although such Dirac cones in the magnetic phase of these materials were not commonly expected, they were predicted by Ran *et al.* [@Ran], who argued that even in the presence of perfect nesting, the degeneracy of the band structure at the $\Gamma$ and M points must lead to nodes in the spin-density-wave gap function. Indeed, their calculations indicate the formation of Dirac cones at the SDW gap nodes. The SDW gap increases away from the Dirac cone, in agreement with ARPES measurements of the leading edge shift around the M point (as defined in Figure \[cone3D\](b)), which are shown in Figure \[cone3D\](c). Figures \[cone3D\](d) and (e) confirm that the electronic states are gapped away from the Dirac cone. Interestingly, ARPES measurements indicate that the electronlike bands have gaps of $\sim 30$ and $\sim 50$ meV. In a first approximation, considering that the Dirac cone apex is almost at the Fermi level, the corresponding full SDW gaps can be estimated at $\sim 60$ and $\sim 100$ meV, which is not far from the values of 45 and 110 meV determined from optical data [@WZ_HuPRL2008]. The presence of Dirac cones in the Fe-based superconductors goes much beyond superconductivity and connects several materials which appear completely different at the first sight [@Hassan_physics2010]. For example, time-reversal symmetry and $C_3$ crystal structure induce Dirac cones in graphene, a single layer of carbon atoms on a triangular lattice. Dirac fermions also emerge at the surface of the newly discovered topological insulators as a consequence of spin-orbit coupling. Finally, the cuprates, which have a Mott insulating parent compound, display a Dirac cone at the nodal point of their $d$-wave superconducting gap once they are doped with impurities. The massless dispersion common to all these materials is believed to play a major role in the next generation of electronic and spintronic devices. Not only Fe-based superconductors now join this category of materials, they constitute the first illustration by ARPES of an anisotropic Dirac cone dispersion. In addition, this discovery shows the importance that orbital characters may play in Fe-based superconductivity [@Hassan_physics2010]. ![\[cone3D\] (Colour online) (a) 3D representation of the Dirac cone at $\Lambda$. The colour scale indicates the distance from $\Lambda$. (b) Schematic FS around the M point. The folded $\alpha$ band, which is barely touching $E_F$, is not indicated. Shaded areas indicate gapped regions and the dashed arrow indicates the orientation of the ARPES cut associated with (d) and (e). (c) Minimum gap location of EDCs (25 K) as a function of the angle $\omega$ defined in (b). The inset shows the leading edge gap (LEG) as a function of $\omega$ after a 3.5 meV shift (see the text). (d) Symmetrized EDCs (25 K) along the cut indicated in (b). (e) Second derivative intensity plot (25 K) along the cut indicated in (b). The vertical dashed line indicates the M point. From Richard *et al.*, Phys. Rev. Lett., **104**, 137001 (2010) [@RichardPRL2010], copyright ©(2010) by the American Physical Society.](cone3D_low){width="8cm"} Recent DMFT calculations in the magnetic state of BaFe$_2$As$_2$ show a very nice consistency with experimental data [@Yin_ZP]. Unlike most ARPES experiments performed on twinned samples, the calculations exhibit the expected 2-fold symmetry rather than a 4-fold symmetry. To reveal the in-plane anisotropy of the band structure from ARPES, samples have to be untwined *in situ* before the experiment [@M_YiPNAS2011; @Y_KimPRB2011]. Contrary to optimally-doped [Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$]{}, where only a small warping of the band structure is reported [@YM_Xu_NPhys2011], a strong variation of the electronic band structure along the $k_z$ direction is found experimentally [@VilmercatiPRB2009; @LiuPRL2009; @FinkPRB2009; @Malaeb_JPSJ2009; @M_YiPRB2009_2; @Brouet_PRB2009; @Kondo_PRB2010; @M_YiPNAS2011]. However, only the $\alpha$ band shows a significant $k_z$-dependence, with a periodicity of 4$\pi/c$ due to the periodicity of the primitive unit cell. While a 3D $\alpha$ pocket centered at $\Gamma$ has been reported for CaFe$_2$As$_2$ [@LiuPRL2009], some reports on BaFe$_2$As$_2$ suggest that the $\alpha$ band crosses $E_F$ at every $k_z$ [@Malaeb_JPSJ2009]. In contrast to the $\alpha$ band, the bands forming the Dirac cone, as well as the electronlike bands at the M point, are only weakly sensitive to $k_z$. Interestingly, one report on CaFe$_2$As$_2$ clearly associates the $k_z$ variations to the SDW state [@LiuPRL2009]. Above $T_{{SDW}}$, the band structure does not show obvious modulations along $k_z$. The 11-chalcogenide system has the simplest structure of all Fe-based superconductors. Interestingly, its magnetic ground state is also different. Unlike BaFe$_2$As$_2$, the AF wave vector of FeTe points in the $\Gamma$-X direction. Consequently, the electronic band structure folds with respect to different magnetic zone boundaries below the ordering temperature. This effect has been observed directly by ARPES [@Y_XiaPRL2009; @Y_Zhang_PRB2010], where bands are folded to the X point, as expected. Doping evolution of the electronic states in the 122-pnictide system {#section_doping} -------------------------------------------------------------------- The 122-pnictides have been the most studied among the Fe-based superconductors and we already have a rough overview of the phase diagram as seen by ARPES. The schematic phase diagram of these materials is illustrated in Figure \[Phase\_diagram\](a). Both the electron-doped and the hole-doped sides show superconductivity and antiferromagnetism. A complication occurs in the electron-doped side, where the spin-density-wave transition does not coincide with the structural transition that accompanies the spin-density-wave transition in the hole-doped case. However, the main difference between the two sides of the phase diagram is the respective sizes of the antiferromagnetic and SC regions. The SC critical temperature reaches a maximum of about 37 K in the hole-doped case, which is obtained at a doping of about 0.2 doping hole per Fe, whereas it tops only around 25 K in the electron-doped materials, with an optimal doping that is roughly 0.08 electron dopant per Fe. Moreover, the SC dome is wide for hole doping and a SC tail survives up to the end of the phase diagram with KFe$_2$As$_2$. It is likely that the reason for such behaviour lies in the electronic band structure itself. ![\[Phase\_diagram\] (Colour online) (a) Phase diagram of the hole- and electron-doped Ba-122 systems taken from Refs. [@Rotter_Angew2008] and [@PrattPRL2009], respectively. $T_c$, $T_{SDW}$, and $T_{str}$ refer to the SC, the SDW, and the tetragonal to orthorombic structural transitions, respectively. (b) Doping dependence of the Lindhard function $\chi_0$ at the M point (quasi-nesting wave vector) normalized by its value at the zone center. The Lindhard function was obtained by using LDA calculations. (c)-(e) Experimentally determined $k_F$ points of the $\alpha$, $\beta$, and $\gamma/\delta$ bands \[red (gray), blue (dark gray), and green (light gray) circles, respectively\]. The $k_F$ points of the $\gamma/\delta$ FSs are shifted by ${\mathbf}{Q}=(-\pi,0)$. (a) and (b) are from Neupane *et al.*, Phys. Rev. B, **83**, 094522 (2011) [@Neupane_PRB2011], copyright ©(2011) by the American Physical Society. (c)-(e) are from Nakayama *et al.*, Phys. Rev. B, **83**, 020501(R) (2011) [@Nakayama_PRB2011], copyright ©(2011) by the American Physical Society.](Phase_diagram_low){width="8cm"} With the electronic carrier concentration varying, the Fermi surface of bulk states must evolve according to the Luttinger theorem, with the size of electronlike Fermi surface pockets increasing (decreasing) as we add electrons (holes), and the size of holelike Fermi surface pockets behaving exactly the other way. An illustration of the size evolution of the Fermi surface pockets in hole-doped Ba$_{1-x}$K$_{x}$Fe$_2$As$_2$ around optimal doping is given in Figures \[Phase\_diagram\](c)-(e). For convenience, the M-centered electronlike FS pockets have been shifted to the $\Gamma$ point. As expected, the size of the $\alpha$ and $\beta$ holelike FS pockets increases with the hole carrier concentration increasing from underdoped Ba$_{0.75}$K$_{0.25}$Fe$_2$As$_2$ [@YM_Xu_UD] to overdoped Ba$_{0.3}$K$_{0.7}$Fe$_2$As$_2$ [@Nakayama_PRB2011] while the size of the electronlike pockets is reduced. Precise analysis of the algebraic FS area is consistent with the bulk concentration in each case. To characterize the similarity between the size of the $\alpha$ Fermi surface and that of the M-centered electronlike Fermi surface pockets, here we extend the notion of nesting beyond its strict meaning. In 2D, if two circular electronlike and holelike pockets overlap completely under translation by the momentum ${\mathbf}{Q}$, the bare static spin susceptibility $\chi_0({\mathbf}{q}, E=0)$ diverges logarithmically as ${\mathbf}{q}$ approaches ${\mathbf}{Q}$. This singularity defines perfect nesting with the important consequence that the random phase approximation (RPA) susceptibility diverges as well, *i.e.* the system becomes unstable towards an SDW with wave vector ${\mathbf}{Q}$ for arbitrarily weak interactions. Although $\chi_0({\mathbf}{q}, E=0)$ no longer diverges at ${\mathbf}{Q}$ if the size and shape of the Fermi surfaces start to deviate slightly from each other, the susceptibility can remained peaked at ${\mathbf}{Q}$, and thus moderate interactions can still drive the system into a SDW or CDW transition in the presence of fluctuations near the wave vector ${\mathbf}{Q}$. The precise behaviour of the susceptibility for what becomes natural to call *near-nesting* or *quasi-nesting* can be analytically understood in terms of a cutoff for the divergence arising from the deviation to perfect nesting [@Cvetkovic_EPL2009]. Since the Fermi surfaces are simply the constant energy contours of the energy dispersions at zero energy, the electronlike and holelike pockets can be even better nested by low-energy contours close to the Fermi level. This implies that quasi-nesting is even more robust for dynamical fluctuations, for which $\chi_0({\mathbf}{q}, E)$ is very much enhanced, producing strong low-energy fluctuations with wave vector ${\mathbf}{Q}$, as indicated by the numerical renormalization group approach [@F_WangEPL2009]. In practice, we will say that the holelike and electronlike Fermi surfaces are quasi-nested if for large portions of the holelike Fermi surface, we can find a small wave vector $\delta{\mathbf}{q}_i$ and a small energy $\delta E_i$ so that the Fermi wave vector ${\mathbf}{k}_{F,i}$ on the holelike Fermi surface can be connected by the wave vector ${\mathbf}{Q}+\delta{\mathbf}{q}_i$ to one point on the $\delta E_i$ energy contour of the electronlike dispersion. Within the extended notion of nesting described above, the $\alpha$ Fermi surface is well quasi-nested with the M-centered electronlike Fermi surface pockets in the optimally-doped system. Unavoidably, the quasi-nesting conditions must evolve while doping the system. With underdoping, the size of the $\beta$ Fermi surface shrinks but is still far to match the Fermi surface at the M point. In the meantime, the quasi-nesting of the M-centered electron Fermi surfaces and the $\alpha$ Fermi surface deteriorates only slightly, mainly along the $\Gamma-M$ direction. In the slightly overdoped region, the quasi-nesting conditions are more severely affected. Yet, each Fermi surface survives and antiferromagnetic scattering between $\Gamma$ and M is still expected. This statement is no longer true when doping the system even more with holes. Sato *et al.* reported a detailed analysis of the band structure of fully K-substituted Ba$_{1-x}$K$_x$Fe$_2$As$_2$ [@Sato_PRL2009]. The main finding of this report is the total absence of electronlike Fermi surface at the M point in KFe$_2$As$_2$, which is equivalent to a suppression of the quasi-nesting condition between holelike and electronlike Fermi surface pockets. Instead, four small Fermi surfaces, elongated along the $\Gamma$-M direction, emerge away from M, as indicated in Figure \[mu\_shift\](a). Interestingly, this dramatic change in the Fermi surface topology is not caused by a reconstruction of the band structure, but simply by a $\sim 20$ meV shift of the chemical potential as compared to the optimally-doped compound (see Figure \[mu\_shift\](b)). The optimally electron-doped BaFe$_{1.85}$Co$_{0.15}$As$_2$ system ($T_c=25.5$ K) has first been studied by Terashima and co-workers [@Terashima_PNAS2009]. Due to the chemical potential shift that accompanies electron doping, the $\alpha$ Fermi surface pocket is totally absent in this material. Instead, only the $\beta$ Fermi surface survives around the $\Gamma$ point. Nevertheless, its size is significantly reduced. On the other hand, the M-centered electronlike Fermi surface pockets expand, as expected, and the size of the inner electronlike pocket at M becomes similar to the size of the $\beta$ pocket, suggesting that antiferromagnetic scattering can become important for this composition. As with its holelike counterpart, this is not the case when the system is extremely electron-doped. Sekiba *et al.* demonstrated the absence of holelike Fermi surface pocket and thus the disappearance of quasi-nested Fermi surfaces in BaFe$_{1.7}$Co$_{0.3}$As$_2$, in which $T_c$ vanishes [@Sekiba_NJP2009]. Indeed, even the $\beta$ band sinks completely below $E_F$ at that high electron doping. In contrast, the M-centered electronlike Fermi surfaces become larger than in BaFe$_{1.85}$Co$_{0.15}$As$_2$, as required by the Luttinger theorem. We note that this change in the Fermi surface topology as compared to the optimally-doped system is induced by a chemical potential shift of about 20 meV due to electron-doping and that the overall band structure is preserved [@Sekiba_NJP2009]. Unlike the hole-doped side of the phase diagram, traces of what can be attributed to long-range antiferromagnetism survives at quite high electron doping in the underdoped regime, at least when using normalized doping $x/x_c$. Indeed, the bright spots attributed to the band-folding-induced Dirac cone in the parent compound [@RichardPRL2010] are observed until the onset of superconductivity is reached [@C_LiuNaturePhys2010]. For the electron-doped compounds though, the bright spots evolve into small holelike pockets [@C_LiuNaturePhys2010]. We note that their elongated shape is consistent with the anisotropy of the Dirac cone [@RichardPRL2010]. ![\[mu\_shift\] (Colour online) (a) Comparison (in the 2 Fe/unit cell notation) of experimentally determined $k_F$ points between overdoped KFe$_2$As$_2$ ($T_c=3$ K) and optimally doped Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$ ($T_c=37$ K) [@Ding_EPL] (blue and red circles, respectively). The $k_F$ points are symmetrized by assuming a fourfold symmetry with respect to the $\Gamma$ and M points. (b) Experimental band dispersion in the vicinity of $E_F$ for two high-symmetry lines determined by tracing the peak position of the ARPES spectra. The chemical potential of the KFe$_2$As$_2$ sample is shifted downward with respect to that of the Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$ sample. (c) Fits of the $\alpha$ and $\beta$ bands for cuts passing through the $\Gamma$ point. The bands have been shifted to match the band slopes. (d) The blue dots are the valence band shifts shown in panel (c) and the red dash curve is the LDA calculated values of the chemical potential divided by a factor of 4. The black dots correspond to the relative shifts derived from panel (c), with the shift of the BaFe$_{1.7}$Co$_{0.3}$As$_2$ compound fixed arbitrarily on the renormalized LDA curve. (a) and (b) are from Sato *et al.*, Phys. Rev. Lett., **103**, 047002 (2009) [@Sato_PRL2009], copyright ©(2009) by the American Physical Society. (c) and (d) are from Neupane *et al.*, Phys. Rev. B, **83**, 094522 (2011) [@Neupane_PRB2011], copyright ©(2011) by the American Physical Society.](mu_shift_low){width="8cm"} A critical observation in the 122-pnictides is that, besides the band folding that is present in the antiferromagnetic regime, the band structure of these materials does not vary significantly with doping. Instead, the electronic states near the Fermi level are tuned by a simple shift of the chemical potential with doping. Neupane *et al.* tracked the position of the bottom of the M-centered electronlike bands as a function of doping on both side of the phase diagram [@Neupane_PRB2011]. Similarly, the authors also estimated the shift in the position relative to $E_F$ of the $\Gamma$-centered holelike bands by matching their electronic dispersions, as illustrated in Figure \[mu\_shift\](c). Interestingly the shift of the electronic band structure at the $\Gamma$ and M points coincide, at least at the first order. More importantly, Figure \[mu\_shift\](d) shows that the chemical potential shift determined by this method is in good agreement with the one estimated from LDA calculations after renormalization of the band width by a factor of 4 [@Neupane_PRB2011] to take into account the stronger renormalization observed for the electronic states very near $E_F$ [@Ding_J_Phys_Condens_Matter2011]. This conclusion has two important consequences. First, it confirms experimentally that the system is really doped through the substitution of Fe by Co, and that this substitution does not affect the effective masses of the system significantly, in contrast to speculations from a theoretical model that suggest that the Co substitution does not dope the system [@WadatiPRL2010]. Second, it suggests that renormalized LDA band calculations can be used to predict various electronic behaviours of these compounds, at least at the first order and away from long-range antiferromagnetic ordering. Starting from this assumption, Neupane *et al.* calculated the doping evolution of the Lindhard function at the antiferromagnetic wave vector [@Neupane_PRB2011]. The result, displayed in Figure \[Phase\_diagram\](b) shows a surprising similarity with the phase diagram, if we disregard the antiferromagnetic region for reasons mentioned above. For example, the maximum value of the Lindhard function is higher on the hole-doped side than in the electron-doped side, and the position of that maximum is also found at higher doping. Furthermore, the Lindhard function keeps a relatively high value for a wider range of doping in the hole-doped side. Overall, this surprising result suggests some correlation between superconductivity and antiferromagnetic scattering near-$E_F$ in the 122-pnictides. Other materials {#Other_materials} --------------- Except for the 122-chalcogenides, which will be described at the end of this section, all Fe-based superconductors share the same fermiology for most part of their phase diagram: circular to slightly squarish holelike Fermi surface pockets centered at $\Gamma$ and a pattern formed by the hybridization of two ellipse electronlike Fermi surface pockets at the antiferromagnetic wave vector $M=(\pi,0)$. However, the quasi-nesting conditions between these two types of Fermi surfaces vary from one compound to another. For example, relatively good quasi-nesting is found in the electron-doped 111-pnictide NaFe$_{0.95}$Co$_{0.05}$As [@ZH_LiuPRB2011]. In contrast, the quasi-nesting conditions, as defined in Section \[section\_doping\], are much weaker[^1] in SC LiFeAs ($T_c=18$ K) [@BorisenkoPRL2010], even though slightly inelastic antiferromagnetic scattering near the Fermi surface remains possible . It is also interesting to point out that all predicted holelike bands are observed in the 111-pnictides near the $\Gamma$ point [@C_HePRL2010; @ZH_LiuPRB2011; @BorisenkoPRL2010] with much less ambiguity than in the 122-pnictides. As with these later materials, the band structure of the 111-pnictides is also renormalized significantly near $E_F$ compared to band structure calculations. A renormalization factor of 3 has been reported for LiFeAs [@BorisenkoPRL2010] while He *et al.* report renormalization by factors varying between 4.3 and 5.4 on different bands in the parent compound NaFeAs [@C_HePRL2010]. Despite an obvious surface problem induced by the absence of appropriate cleaving plane that results in excess hole doping at the surface layer, ARPES results on the 1111-pnictides [@DH_LuNature2008; @Kondo_PRL2008; @C_LiuPRB2010; @HY_LiuPRL2010] also indicate essentially the same fermiology as their pnictide cousins. In addition, experimental data indicate a renormalization of the band structure in LaOFeP of 2.2 [@DH_LuNature2008], which is similar to the other pnictides. Unlike for the other Fe-based superconductors though, no variation of the band structure is observed by ARPES when changing the photon energy [@C_LiuPRB2010], which is easily understood since the state probed is a real surface state induced by the cleaving problem for this material. Among all pnictides, a special attention must be devoted to Sr$_2$VFeAsO$_3$, for which there exists, to our knowledge, only one ARPES report [@Qian_PRB2011]. In addition to Fe3$d$ states, most LDA band calculations predicted the presence of V3$d$ states at the Fermi level that would completely modify the fermiology of this system [@KW_LeeEPL2010; @MazinPRB2010_21311; @Shein_JSNM2009; @G_WangPRB2009]. In fact, one of these papers mentioned that if the Fermi surface is derived only from Fe states, it would have a topology comparable to that of the other pnictides [@MazinPRB2010_21311]. This is precisely what is found experimentally [@Qian_PRB2011]. Using samples smaller than $0.2\times 0.2$ mm$^2$, Qian *et al.* determined that the Fermi surface of this system is composed by a small and a large $\Gamma$-centered holelike Fermi surfaces (called $\alpha$ and $\beta$, respectively) and by electronlike Fermi surfaces with elongated ellipse shape at the M point. The electron counting from the algebraic Fermi surface area leads to a carrier concentration compatible with the Luttinger theorem. Comparing their results with LDA+U calculations, the authors found a good agreement with the overall Fe3$d$ band structure when considering a normalization factor of 1.6, as shown in Figure \[V\_paper\](a). However, the states closer to the Fermi level fit well the calculation only for a renormalization factor of about 3.3 [@Qian_PRB2011]. The main discrepancy between the experimental results and the other LDA calculations relates to the position of the V states. From their resonance profile as a function of incident photon energy, illustrated in Figures \[V\_paper\](b)-(d), the V3$d$ states are found around 1 eV below $E_F$, a result that is reproduced only when incorporating an effective Hubbard energy U of 6 eV in the LDA+U calculations [@Qian_PRB2011]. It is interesting to note that with V$3d$ electrons in a Mott state and Fe$3d$ electrons at the Fermi level in the absence of Fe-V hybridization, Sr$_2$VFeAsO$_3$ can be viewed as a perfect orbital selective Mott transition [@AnasimovEPJB2002; @LiebschPRL2003] system, such as in Ca$_{1.8}$Sr$_{0.2}$RuO$_4$, which has also been characterized by ARPES [@NeupanePRL2009]. ![\[V\_paper\] (Colour online) (a) Second derivative plot of along the $\Gamma$-X-M-$\Gamma$ high symmetry lines of Sr$_2$VFeAsO$_3$. LDA+U bands are also plotted for comparison. The Fe $3d$ bands (red/gray lines) are renormalized by a factor of 1.6, whereas the V $3d$ bands (black lines) are not. The Fe 3$d_{xy}$ band (black dashed line) is renormalized by a factor of 3.3 to reproduce the experimental band near $E_F$. (b) Valence band (VB) at the BZ center measured at different photon energies (34-80 eV). All the spectra are normalized by the photon flux. The fitting curves for the spectrum at $h\nu=80$ eV are also shown. (c) Photon energy dependence of the intensities of the two peaks obtained from the fitting curves and the spectral intensity at $E_B=0.01$ eV. (d) Direct comparison of the VB measured at $h\nu=53$ and 80 eV, along with the VB of BaFe$_2$As$_2$ ($h\nu=80$ eV). From Qian *et al.*, Phys. Rev. B, **83**, 140513 (2011) [@Qian_PRB2011], copyright ©(2011) by the American Physical Society.](V_paper_low){width="8cm"} The Fe-chalcogenides have been less studied than the pnictides by ARPES, even the 11-chalcogenides that have been synthesized in the early days of Fe-based superconductivity. Nevertheless, existing data of doped 11-chalcogenides show spectral intensity at the $\Gamma$ and M$(\pi,0)$ points [@Nakayama_PRL2010; @Tamai_PRL2010; @Miao2011]. Although up to three bands can be identified at $\Gamma$, the situation remains more nebulous at the M point, making a careful electron counting analysis difficult. As with pnictides, some high energy features can be well approximated by LDA calculations renormalized by a factor around 2 [@Nakayama_PRL2010]. However, additional renormalization much stronger than in the pnictides is found near $E_F$ [@Nakayama_PRL2010; @Tamai_PRL2010], with renormalization factors as large as 20 [@Tamai_PRL2010]. Recently, the landscape of Fe-based superconductivity changed abruptly with the discovery of high-$T_c$ ($T_c=29$ K) in heavily-electron-doped K$_x$Fe$_2$Se$_2$ [@JG_Guo_PRB2010] and related materials. The input of ARPES to characterize and explain the electronic properties of these materials has been very helpful. Unlike all the other Fe-based superconductors, this system does not show any holelike Fermi surface [@Qian_PRL2011; @XP_WangEPL2011; @Y_Zhang_NatureMat2011; @D_MouPRL2011], as illustrated in Figure \[new122\_structure\](a). In contrast, a large electronlike Fermi surface is observed at the M point. Although the zone center is free of holelike Fermi surfaces, there is evidence for additional electronlike pockets whose origin are not quite clear at the moment. A small 3D electronlike pocket centered at the Z point has been reported in a synchrotron-based study on A$_x$Fe$_2$Se$_2$ (A = K, Cs) [@Y_Zhang_NatureMat2011]. Traces of these pockets have also been detected in (Tl,K)Fe$_{1.78}$Se$_2$ [@XP_WangEPL2011] and Tl$_{0.58}$Rb$_{0.42}$Fe$_2$Se$_2$ [@D_MouPRL2011] using He discharge lamps. These laboratory-based experiments also reveal an additional electronlike pocket with a size more or less similar to the one centered at the M point [@XP_WangEPL2011; @D_MouPRL2011], as we can see from the comparison of Figures \[new122\_structure\](c) and (d). ![\[new122\_structure\] (Colour online) (a) Momentum-resolved photoemission intensity mapping of Tl$_{0.63}$K$_{0.37}$Fe$_{1.78}$Se$_2$ recorded in the normal state (35 K) and integrated over a 10 meV window centered at $E_F$. The small red circles indicate the FS obtained from the momentum distribution curve (MDC) peak position at $E_F$. (b) EDCs along several high symmetry directions in K$_{0.8}$Fe$_{1.7}$Se$_2$ recorded with the He I resonance line ($h\nu=21.218$ eV). Blue (dark gray) curves correspond to high symmetry points. (c) ARPES intensity plot along a cut passing through the M point (cut1 from (a)). (d) ARPES intensity plot along a cut passing through the $\Gamma$ point (cut3 from (a)). (a), (c) and (d) are from Wang *et al.*, Europhy. Lett., **93**, 57001 (2011) [@XP_WangEPL2011], copyright ©(2011) by the European Physical Society. (b) is from Qian *et al.*, Phys. Rev. Lett., **106**, 187001 (2011) [@Qian_PRL2011], copyright ©(2011) by the American Physical Society.](new122_structure_low){width="8cm"} As with the other Fe-based superconductors, the 122-chalcogenides have their overall band width renormalized by a factor of about 2.5 [@Qian_PRL2011]. What is different though is the energy distribution of the spectral weight, which is closer to the one observed in the cuprates. While the spectral weight of the Fe3$d$ electrons is much larger in the other Fe-based superconductors, the near-$E_F$ bands in K$_{0.8}$Fe$_{1.7}$Se$_2$ have much weaker intensity. On the other hand, the intensity of the valence band is more important in this materials. In fact, Figure \[new122\_structure\](b) shows that the weight below 2 eV is mainly concentrated in a broad peak located around 800 meV below $E_F$, which does not show significant dispersion. That peak, which has been attributed to the incoherent part of the Fe3$d$ states [@Qian_PRL2011], shows an important energy shift at a temperature $T^*$ corresponding to a bump in the resistivity curves [@XP_WangEPL2011]. Superconducting gap {#section_gap} =================== The order parameter describing the SC phase transition is characterized by a complex function represented by the momentum distribution of the gap developing at $E_F$ below the SC critical temperature $T_c$. The amplitude and phase of the order parameter are directly determined by the electronic structure and by the mechanism responsible for the pairing of charge carriers. Although ARPES is not directly sensitive to the phase, it allows precise measurement of the momentum dependence of the SC gap size, which serves as a powerful tool to validate or invalidate theoretical models. The previous statement is particularly true for multi-band systems, for which the interpretation of transport measurements, which may be momentum-sensitive but not momentum-resolved, becomes difficult or even impossible. ![\[Gap\_OP\] (Colour online) Superconducting gap of [Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$]{}. (a)-(c) Temperature dependence of the EDCs on the $\alpha$, $\beta$ and $\delta$ Fermi surfaces, respectively. The $k_F$ position of the EDCs is given on the schematic FS (in the 2 Fe/unit cell representation) displayed in the inset of panel (a). (d)-(f) Corresponding symmetrized EDCs. (g)-(i) Momentum dependence of symmetrized EDCs at 15 K along the $\alpha$, $\beta$ and $\delta$ Fermi surfaces, respectively. The $k_F$ position are given on the FS displayed in panel (j) in the 1 Fe/unit cell representation. Squares, circles and triangles correspond to the $k_F$ positions on the $\alpha$, $\beta$ and $\delta$ Fermi surfaces, respectively. Filled and empty symbols refer to real and symmetrized data, respectively. Adapted from Ding *et al.*, Europhy. Lett., **83**, 47001 (2008) [@Ding_EPL], copyright ©(2008) by the European Physical Society.](Gap_OP_low){width="8cm"} The first ARPES determinations of the SC gap in Fe-based superconductors were obtained on optimally-doped [Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$]{}[@Ding_EPL; @L_Zhao]. Figures \[EPL\_final\](a)-(c) show the temperature evolution of EDCs from Ding *et al.* recorded with the He I$\alpha$ resonance line on the $\alpha$, $\beta$ and $\gamma$ Fermi surfaces, respectively [@Ding_EPL]. As temperature decreases below $T_c$, a coherence peak develops on each FS, indicating the opening of a superconducting gap, which is better visualized from the corresponding symmetrized EDCs displayed in Figures \[EPL\_final\](d)-(f). The full superconducting gap size $2\Delta$ is given by the distance between the two coherence peaks in the symmetrized EDCs. The superconducting gap is also characterized by its momentum dependence on each FS, which are shown in Figures \[EPL\_final\](g)-(i). The $k_F$ points where these data were collected are indicated in Figure \[EPL\_final\](j). Figure \[EPL\_final\] summarizes the data from Ding *et al.* [@Ding_EPL]. Unlike the SC gap in the cuprates, these experiments show clearly the absence of nodes along all the Fermi surface pockets, ruling out any order parameter with a $d$-wave symmetry. The invariance of the coherence peak positions in the symmetrized EDCs shown in Figures \[EPL\_final\](g)-(i), as well as further refined measurements with the He I$\alpha$ line [@Nakayama_EPL2009], indicate that the gaps are indeed isotropic. Interestingly though, the gap size is not unique and varies from band to band. More precisely, the gap size is around 12 meV along the quasi-nested FSs, yielding to $2\Delta/k_BT_c$ ratios of about 7.5, while the gap size along the non-quasinested $\beta$ FS pocket drops to 6 meV. We note that although the SC gap seems to disappear above $T_c$, the temperature evolution of the SC gap size does not follow the BCS function and shows a much steeper drop while approaching $T_c$. Isotropic gaps have been reported not only in the 122-pnictides [@Ding_EPL; @L_Zhao], but also in the 111-pnictide [@ZH_LiuPRB2011; @BorisenkoPRL2010], the 1111-pnictide [@Kondo_PRL2008] and the 11-chalcogenide [@Nakayama_PRL2010] systems. ![\[EPL\_final\] (Colour online) Three-dimensional plot of the SC gap size ($\Delta$) measured at 15 K on three FS sheets (shown at the bottom as an intensity plot) and their temperature evolutions (inset). From Ding *et al.*, Europhy. Lett., **83**, 47001 (2008) [@Ding_EPL], copyright ©(2008) by the European Physical Society.](EPL_final_low){width="8cm"} To explain these intriguing results, two main models have been proposed for the pairing mechanism: the *FS quasi-nesting model* and the *local AF exchange pairing model*. These two models are radically different. Essentially, the basic question to discriminate one from the other is: Does pairing occur due to interactions in the momentum-space or in the real-space? Historically, the former model gained popularity very quickly, partly due to ARPES results. We thus describe it first, along with the reasons for its initial success, among which are the evidence for the presence of interband scattering in the pnictides and the weakening of the quasi-nesting conditions accompanying the disappearance of superconductivity at very high doping in these materials. Paradoxically, recent ARPES measurements on the 122-chalcogenides are main indicators announcing the dusk of that model to describe the Fe-based superconductors. Nowadays, the local AF exchange pairing model described afterwards appears as a better prospect in claiming for universality of the pairing mechanism in the Fe-based superconductors. Yet, the community is still far from ready to unanimously award laurel wreath to any model candidate. Quasi-nesting model {#section_QN} ------------------- Until the recent discovery of superconductivity in the chalcogenide version of the 122 phase, the quasi-nesting model was the most popular way to account for superconductivity in Fe-based superconductors. The idea is indeed very simple and illustrated in Figure \[fig\_nesting\]: when a holelike and an electronlike FS pockets have similar but not exact size and shape, the system can avoid long-range ordered CDW or SDW states but yet maintain important scattering induced by short-range fluctuations through the “quasi-nesting" wave vector (in this case the antiferromagnetic wave vector) that connects these two FS pockets. Despite the absence of perfect nesting, a peak of enhanced magnetic susceptibility is still expected at the quasi-nesting wave vector, in which case the kinetic process by which a zero momentum pair formed on one FS is scattered onto another one can be enhanced, whereby increasing the pairing amplitude [@Ding_EPL; @MazinPhysicaC2009; @Graser_NJP2009]. ![\[fig\_nesting\] (Colour online) Comparison of energy bands between BaFe$_{1.85}$Co$_{0.15}$As$_2$ and BaFe$_{1.7}$Co$_{0.3}$As$_2$ samples. The interband scattering is dramatically suppressed in the non-SC BaFe$_{1.7}$Co$_{0.3}$As$_2$ compound since the holelike $\alpha$ and $\beta$ bands at the $\Gamma$ point are basically occupied. From Sekiba *et al.*, New J. Phys., **11**, 025020 (2009) [@Sekiba_NJP2009], copyright ©(2009) by IOP Publishing and Deutsche Physikalische Gesellschaft.](fig_nesting_low){width="8cm"} In practice, the quasi-nesting model can be tested by checking the evolution of the SC gap size as a function of the quasi-nesting conditions. The latter are easily tuned with doping since the doping evolution of holelike and electronlike FS pockets with carrier injection have opposite trends. The most drastic effect is observed when extra-doping the system until one type of FS pocket, either electronlike or holelike, disappears. As mentioned previously, the purely K-substituted 122 compound (KFe$_2$As$_2$) is extremely over-hole-doped and does not exhibit any electronlike FS [@Sato_PRL2009]. In agreement with the quasi-nesting model, its $T_c$ drops to a few Kelvins only. Unfortunately, the gap size becomes too small to be extracted. Similarly, the disappearance of $\Gamma$-centered holelike FS pockets in heavily electron-doped BaFe$_{1.7}$Co$_{0.3}$As$_2$ is accompanied by a total suppression of superconductivity [@Sekiba_NJP2009]. Another experimental fact consistent with the quasi-nesting model in the 122 system is the switch of the pairing strength on the $\beta$ band from “weak coupling" to “strong coupling" as we compare the optimally-hole-doped and optimally-electron-doped compounds (see Figure \[polar\]). While the $\beta$ FS is quite larger than the M-centered electronlike FS pockets in the hole-doped compound, its size becomes compatible with the FS pockets at the M point in the optimally-electron-doped system, thus promoting AF interband scattering [@Terashima_PNAS2009]. ![\[polar\] (Colour online) Polar representation of the SC gap amplitude for various bands and materials. The data on Ba$_{0.3}$K$_{0.7}$Fe$_2$As$_2$ are from Nakayama *et al.*, Phys. Rev. B, **83**, 020501(R) (2011) [@Nakayama_PRB2011], copyright ©(2011) by the American Physical Society. The data on Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$ are from Nakayama *et al.*, Europhy. Lett., **85**, 67002 (2009) [@Nakayama_EPL2009], copyright ©(2009) by the European Physical Society. The data on BaFe$_{1.85}$Co$_{0.15}$As$_2$ are from Terashima *et al.*, Proc. Natl. Acad. Sci. USA, **106**, 7330 (2009) [@Terashima_PNAS2009], copyright ©(2009) by the National Academy of Sciences of the United States of America. The data on Tl$_{0.63}$K$_{0.37}$Fe$_2$Se$_2$ are from Wang *et al.*, Europhy. Lett., **93**, 57001 (2011) [@XP_WangEPL2011], copyright ©(2011) by the European Physical Society.](polar_low){width="8cm"} Whatever it is or not the cause for Fe-based superconductivity, ARPES provides sufficient evidence for interband scattering with the quasi-nesting wave vector. A first evidence is the smaller EDC linewidth of the unnested $\beta$ band in [Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$]{}as compared to the quasi-nested bands [@Ding_EPL]. It indicates that the lifetime of the quasi-particles associated with the $\beta$ band is longer due the absence of interband scattering. A second evidence is the observation of an anomaly, or *kink*, in the dispersion of bands associated with quasi-nested FS pockets [@RichardPRL2009]. At 15 K, this anomaly is found around 25 meV, as can been seen from Figures \[fig\_kink\](a)-(c). To confirm this result, a self-energy analysis was performed (see Figure \[fig\_kink\](e)) by approximating the bare band dispersion by the dispersion determined at 150 K, way above the critical temperature. This choice of bare band approximation, justified by the absence of anomaly at 150 K , is critical since the Fe-based superconductors are multi-band systems. In contrast, the cuprates are single-band systems studied not far from half filling, where the band dispersion near $E_F$ can be reasonably well approximated by a straight line. By dividing out the ARPES intensity plot recorded at 150 K by the Fermi-Dirac function, one can reveal the top of the $\alpha$ band about 25 meV above $E_F$, as illustrated in Figure \[fig\_kink\](d), indicating that the band dispersion in the energy region of the kink is naturally curved and that a straight line cannot be used to approximate the bare band dispersion. In the SC state, both the imaginary and real parts of the self-energy deviate around 25 meV from the quite linear behaviour observed at higher energy. ![\[fig\_kink\] (Colour online) (a) ARPES intensity plot in the SC state (15 K) along a cut crossing the $\alpha$ band. The inset shows the schematic FS of Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$ with the location of the cut (red) in the 2 Fe/unit cell notation. (b) Corresponding EDCs. (c) Corresponding MDCs, in the 0-60 meV binding energy range. Grey dots indicate the maximum position of the peaks. (d) ARPES intensity plot at 150 K divided by a Fermi-Dirac function, recorded along the same cut as in (a). (e) Real and imaginary parts of $\Sigma(\omega)$. Fade colours are used for binding energies smaller than 17 meV since $\Sigma(\omega)$ in this range is complicated by particle-hole mixing due to superconductivity. The inset compares the partial DOS along the cuts measured at 15 K (blue) and 150 K (red), respectively. (f) Temperature dependence of the real part of the self-energy referred from the MDC fit dispersion at 150 K. (g) Maximum value of the real part of the self-energy (blue) plotted as a function of temperature. The ARPES results are compared to the neutron scattering intensity of the 14 meV spin resonance (red) located at the AF wave vector [@Christianson]. The dashed line indicates the critical temperature. From From Richard *et al.*, Phys. Rev. Lett., **102**, 047003 (2009) [@RichardPRL2009], copyright ©(2009) by the American Physical Society.](fig_kink_low){width="8cm"} The coupling of an Einstein mode of energy $\Omega$ is expected to give rise to an anomaly at $\Omega+\Delta$ [@Norman_PRB1998], where the SC gap $\Delta$ is well known for each band. Subtracting the 12 meV energy corresponding to the gap on the $\alpha$ FS from the kink energy yields $\Omega=13\pm 2$ meV. While an anomaly is also found in the quasi-nested $\gamma$ band, the $\beta$ band shows an anomaly neither at 25 meV nor at 18 meV ($\Delta_{\beta}+\Omega=6+13$ meV). This band-selectivity of the anomaly is in apparent contradiction with a conventional electron-phonon coupling and point towards an electronic origin [@RichardPRL2009]. Interestingly, the temperature dependence of the kink anomaly, shown in Figure \[fig\_kink\](f), coincides with the temperature dependence of a spin resonance mode observed by inelastic neutron scattering at 14 meV [@Christianson], as well illustrated in Figure \[fig\_kink\](g). As with the ARPES kink, the spin resonance, detected at the antiferromagnetic wave vector, disappears above $T_c$, suggesting that the two phenomena are related. Similarly to the cuprates, antiferromagnetic scattering increases while under-doping Fe-based superconductors [@Ning_PRL2010]. This affects not only the width of the SC quasi-particle peaks, but also their weight. While the SC quasi-particle peak associated with the unnested $\beta$ band in Ba$_{0.75}$K$_{0.25}$Fe$_2$As$_2$ ($T_c=26$ K) remains sharp and coherent, the peaks associated with the nested $\alpha$ and $\gamma/\delta$ bands lose their integrity compared to the optimally-doped material [@YM_Xu_UD]. Indeed, this dichotomy between the behaviour of SC quasi-particle peaks on the unnested and quasi-nested FS pockets upon under-doping is quite similar to the one observed in cuprates between the nodal and antinodal regions. Unlike the quasi-particle peak in the nodal region, the quasi-particle peak at the antinode is quite suppressed. At least for the 122-pnictides, which has been more deeply studied, the doping evolution of the quasi-nesting conditions is in qualitative agreement with the variation of $T_c$. A good illustration of this statement is provided by the doping dependence of the renormalized Lindhard function at the antiferromagnetic wave vector given in Figure \[mu\_shift\](b), which shows a strong electron-hole asymmetry reflecting the asymmetric band structure with respect to electron- or hole-doping [@Neupane_PRB2011]. While the Lindhard function keeps a high value for a wide hole-doping range before starting to decrease, it decreases monotonically on the electron-doped side, albeit for a small shoulder around x = 0.24. Although these calculations performed with the non-magnetic LDA band structure (renormalized by a factor of 4) are not suitable for the antiferromagnetic region near 0 doping, it qualitatively reproduces the size, height and shape of the SC dome away from 0 doping. Fair agreement with the quasi-nesting model is also obtained for electron-doped NaFe$_{0.95}$Co$_{0.05}$As ($T_c=18$ K), for which Liu *et al.* reported almost identical SC gaps (6.8 *vs* 6.5 meV) on the $\Gamma$-centered $\alpha^{\prime}$ FS pocket and on the M-centered electron FS pockets, which all have similar size [@ZH_LiuPRB2011]. This gap size leads to a $2\Delta/k_BT_c\sim 8$ ratio, indicating that the system is in a strong coupling regime. Using the leading edge shift rather than the SC coherent peak position to identify the gap size, Borisenko *et al.* found a $2\Delta/k_BT_c$ ratio of about 3.5 in LiFeAs [@BorisenkoPRL2010], which is more consistent with the BCS regime. Even though the leading edge method is necessarily an underestimation of the SC gap size, at least by the half-width at half maximum of the SC coherent peak, LiFeAs shows smaller gap size than NaFe$_{0.95}$Co$_{0.05}$As. Interestingly, the quasi-nesting conditions are also poorer in LiFeAs, which has one very large and one very small holelike $\Gamma$-centered FS pockets, along with intermediate size M-centered electronlike FS pockets. This reduction of both $T_c$ and the $2\Delta/k_BT_c$ ratio is thus compatible with the quasi-nesting model. As explained above, LDA calculations predicted the failure of the quasi-nesting model to explain $T_c=37$ K superconductivity in Sr$_2$VFeAsO$_3$ [@KW_LeeEPL2010; @MazinPRB2010_21311; @Shein_JSNM2009; @G_WangPRB2009]. The arguments were mainly related to the presence of V states at $E_F$ in the calculated band structure. Although SC gaps were never measured by ARPES due to the small size of the samples, the experimental Fermi surface is quite similar to that of the other pnictides and thus consistent with the quasi-nesting scenario [@Qian_PRB2011]. A more intriguing case is that of the 11-chalcogenides. In contrast to the pnictides, the magnetic ground state of their parent compound is not described by the $(\pi,0)$ wave vector, but rather by a wave vector pointing in the $\Gamma-X$ direction. Yet, the FS topology of SC 11-chalcogenides is quite similar to that of the pnictides, and ARPES results by Nakayama *et al.* suggest that the SC gap of the holelike band at $\Gamma$ is also isotropic and in the strong coupling regime [@Nakayama_PRL2010]. Unfortunately, the low spectral intensity at the M point did not allow the determination of the SC gap there until very recently [@Miao2011]. In any case, these surprising similarities between SC pnictides and 11-chalcogenides indicate the importance of the $(\pi,0)$ scattering, even though it does not necessarily proves the validity of the quasi-nesting model in explaining high-$T_c$ superconductivity in these materials. In fact, some neutron scattering experiments suggest that AF scattering at $(\pi,0)$ is important in this system and even evolves into a resonance for doped samples [@IikuboJPSJ2009; @YM_QiuPRL2009; @TJ_LiuNMAT2010]. Local antiferromagnetic exchange pairing model {#section_local} ---------------------------------------------- Enthusiasm for the quasi-nesting scenario in the pnictides was just too strong in the early days to allow the local pairing model to really take off. Nevertheless, Wray *et al.* [@Wray_PRB2008] and Nakayama *et al.* [@Nakayama_EPL2009] pointed out that the SC gap measured by ARPES in the 122-pnictides is qualitatively consistent with an order parameter that takes the form $\Delta_0\cos(k_x)\cos(k_y)$. Interestingly, such a formula can be derived from a picture where the pairing occurs in a short distance in real space, for example by considering antiferromagnetic exchange between nearest ($J_1$) and second-nearest ($J_2$) neighbors. Figure \[J\_model\](a) displays the detailed comparison between this formula and the observed gap values obtained by Nakayama *et al.* [@Nakayama_EPL2009]. Apart from small deviations, the same formula applies also to the 111-pnictide NaFe$_{0.95}$Co$_{0.05}$As, where similar $\Delta_0$ values of 6.8 and 6.5 meV are found for the $\Gamma$-centered holelike FS pockets and M-centered electronlike FS pockets, respectively [@ZH_LiuPRB2011]. ![\[J\_model\] (Colour online) (a) Theoretical SC gap value $|\Delta (k) = \Delta_0|\cos k_x \cos k_y|$ with $\Delta_0=13.5$ meV as a function of the two-dimensional wave vector [@SeoPRL2008; @KorshunovPRB2008]. (b) Extracted values of the SC gap (defined as the half value of peak-to-peak positions in the symmetrized EDCs) on the $\alpha$ FS at different photon energies. The dots (squares) are obtained from the left (right) side of $k_F$ on the $\alpha$ FS. (c) The SC gap values on the $\alpha$ FS (red dots), $\beta$ FS (blue dots), $\gamma/\delta$ FS (green dots) and $\alpha$’ FS (brown dots) as functions of the 3D gap function $|\Delta(k_x,k_y,k_z)|=|\Delta_1\cos k_x\cos k_y+(\Delta/2)(\cos k_x+\cos k_y)\cos k_z|$, with $\Delta_1 = 12.3$ meV and $\Delta_2=2.07$ meV, to fit all the SC gaps. The error bars are standard deviations of the measured SC gaps. Panel (a) is Nakayama *et al.*, Europhy. Lett., **85**, 67002 (2009) [@Nakayama_EPL2009], copyright ©(2009) by the European Physical Society. (b) and (c) are from Xu *et al.*, Nature Phys., **7**, 198 (2011) [@YM_Xu_NPhys2011], copyright ©(2011) by Macmillan Publishers Ltd.](J_model_low){width="8cm"} The presence of a cosine oscillation along the $k_z$ direction of the gap value associated with the $\alpha$ band, well illustrated by Figure \[J\_model\](b) justifies modifications of the $\Delta=\Delta_0\cos(k_x)\cos(k_y)$ formula. Xu *et al.* [@YM_Xu_NPhys2011] started from the simple formula $$\Delta_{3D}(k_x,k_y,k_z)=\Delta_{2D}(k_x,k_y)(1+\eta\cos(k_z))$$ which is a generalization to layered superconductors of the BCS expression for a superconductor with an isotropic in-plane gap function and where $\eta$ is a measure of the interlayer coupling strength. With specific considerations applying to the pnictides, they then derived the generalized $s$-wave gap function [@YM_Xu_NPhys2011]: $$\begin{aligned} \Delta=\Delta_1\cos(k_x)\cos(k_y)+\nonumber\\ (\Delta_2/2)(\cos(k_x)+\cos(k_y))\cos(k_z)+\Delta_3(\cos(k_x)\cos(k_y))\cos(k_z)\end{aligned}$$ where the gap parameters $\Delta_1$, $\Delta_2$ and $\Delta_3$ can be orbital-dependent or band-dependent in the most general case. Neglecting the last term which is experimentally much smaller than the $\Delta_2$ term due to the vanishingly small $\eta$ of the M-centered electronlike FS pockets, the experimental data obtained on optimally-doped Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$ can be well fitted with $\Delta_1=12.3$ meV and $\Delta_2=2.07$ meV [@YM_Xu_NPhys2011]. Figure \[J\_model\](c) shows the fit of SC gap data on all the Fermi surfaces of Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$ using a single gap function. Interestingly, the $\Delta_1/\Delta_2=5.9$ ratio coincide almost with the $J_{ab}/J_c=6$ ratio of the in-plane and out-of-plane magnetic coupling constants determined from spin-wave dispersions obtained by neutron scattering experiments on the parent compound [@J_Zhao_PRL2008]. The validity of the gap function given above has been checked for over-doped Ba$_{0.3}$K$_{0.7}$Fe$_2$As$_2$ ($T_c=22$ K) with the He I$\alpha$ resonance line (21.218 eV). According to the $k_z$ dispersion of hole-doped 122-pnictides [@YM_Xu_NPhys2011], the corresponding photon energy is close to the $\Gamma$ point, and thus $\cos(k_z)\simeq 1$. The experimental results yield $\Delta_1=8.3$ meV and $\Delta_2=0.7$ meV [@Nakayama_PRB2011]. It is worth nothing that within this model the electronlike and the holelike FS pockets may carry different gaps even if their size is the same since the term $\cos(k_x)+\cos(k_y)$ is 2 and 0 for the $\Gamma$ and M points, respectively. Interestingly, the 3$^{rd}$ nearest neighbor antiferromagnetic exchange coupling $J_3$ in the $J_1$-$J_2$-$J_3$ model is very strong in the chalcogenides and must be taken into account [@LipscombePRL2011]. As a consequence, the formula $|\Delta (k) = \Delta_2 |\cos k_x \cos k_y|$ does not apply and must be replaced by $|\Delta (k)| = |\Delta_2 \cos k_x \cos k_y- \Delta_3(\cos 2k_x -\cos 2k_y)/2|$. The later fits pretty well the data of FeTe$_{0.55}$Se$_{0.45}$ ($T_c = 14$ K) and the 4.2 meV and 2.6 meV gaps found on the M-centered $\gamma$ FS and $\Gamma$-centered $\alpha$ FS assuming $\Delta_2 = 3.55$ meV and $\Delta_3 = 0.95$ meV [@Miao2011]. A noticeable feature is the stronger gap at the M point than at the $\Gamma$ point, which is not the case for the pnictides and that comes out naturally from the $J_1-J_2-J_3$ model. It also worth mentioning that the $\Delta_2/\Delta_3$ ratio is almost the same as the $J_2/J_3$ ratio determined experimentally from neutron scattering experiments [@LipscombePRL2011]. The revival of the local pairing picture is mainly due to the recent discovery of superconductivity at high-temperature in over-electron-doped 122-chalcogenides. As mentioned in Section \[Other\_materials\], these systems are completely free of holelike FS pockets at the $\Gamma$ point and thus prevent $(\pi,0)$ scattering in the electron-hole channel. Unless this new class of materials is the host of a different pairing mechanism, which is quite unlikely due to the structural and electronic similarities between the chalcogenides and the pnictides, the local pairing picture seems more robust as a universal model to explain superconductivity in the Fe-based superconductors. The SC gap of the 122-chalcogenides has been measured for (Tl,K)Fe$_{1.78}$Se$_2$ [@XP_WangEPL2011], A$_x$Fe$_2$Se$_2$ [@Y_Zhang_NatureMat2011] and (Tl$_{0.58}$Rb$_{0.42}$)Fe$_{1.78}$Se$_2$ [@D_MouPRL2011]. All data suggest an isotropic SC around the M-centered electronlike FS pocket. In Tl$_{0.63}$K$_{0.37}$Fe$_2$Se$_2$ ($T_c=29$ K), a 8.5 meV gap is found, as indicated in Figure \[polar\], which leads to $2\Delta/k_BT_c\sim 7$, clearly in the strong coupling limit [@XP_WangEPL2011]. SC gaps are also reported for electronlike FSs centered at the $\Gamma$ point. Hence, Wang *et al.* reported a 8 meV for the SC gap on the large $\Gamma$-centered electronlike FS pocket in Tl$_{0.63}$K$_{0.37}$Fe$_2$Se$_2$ [@XP_WangEPL2011]. Zhang *et al.* reported a 7 meV on the small 3D electronlike pocket centered at the Z point in K$_{0.8}$Fe$_2$Se$_2$, as compared to 10.3 meV at the M point [@Y_Zhang_NatureMat2011]. Finally, Mou *et al.* recorded a 12 meV gap on the large $\Gamma$-centered electronlike FS pocket in Tl$_{0.58}$Rb$_{0.42}$Fe$_{1.72}$Se$_2$ ($T_c=32$ K), in contrast to the 15 meV SC gap found at the M point [@D_MouPRL2011]. Following the recent developments on the SC gap of the 11-chalcogenides and the 122-chalcogenides as determined from ARPES measurements, the local antiferromagnetic exchange pairing model has been further developed to emphasize not only on the local exchange couplings, but also on the fermiology of the different materials [@Hu_universal]. More precisely, it is argued that high-temperature superconductivity occurs when the Fermi surface topology matches the form factor of the pairing symmetry favored by the local antiferromagnetic exchange interactions. The result extends to cuprates as well. Temperature dependence of the superconducting gap ------------------------------------------------- It is has been noticed in an early ARPES report on the 122-pnictides that the temperature dependence of the SC gap was not the same as in the BCS superconductors [@Ding_EPL]. Unfortunately, the $\alpha$ band shows a “shoulder" in the SC state that makes a precise characterization difficult. In contrast, the 111-pnictides constitutes an ideal system to investigate this issue since no shoulder is observed and the natural cleaved surfaces are expected to be non-polar and free of reconstruction. Liu *et al.* performed such a study on electron-doped NaFe$_{0.95}$Co$_{0.05}$As [@ZH_LiuPRB2011]. From their data analysis, it is possible to see that the SC gap size $\Delta$, given by half the distance between the two SC coherent peaks, does not change much with temperature increasing. Instead, the gap is filling up as temperature increases. A more quantitative approach consists in fitting the experimental EDCs using the self-energy function $\Sigma(k,\omega)$ suggested by Norman *et al.* for quasiparticles in the SC state [@Norman_PRB1998]: $$\Sigma(k,\omega)=-i\Gamma+\frac{\Delta^2}{(\omega+i0^+)+\epsilon(k)}$$ where $\Gamma$ is the single-particle scattering rate, which is here assumed to be independent of $\omega$. Assuming a polynomial background, this function shows a good agreement with the experiments at all temperatures until the SC peak vanishes. As with the scattering rate and the SC gap size, the normalized coherent area $Z_A$ defined as $Z_A=\int A_{coh}(k,\omega)d\omega/\int A(k,\omega)d\omega$ as in earlier studies on cuprates [@DingPRL2001; @FengScience2000] can also be calculated. While the gap size and the scattering rate seem almost unaffected by temperature, this later parameter exhibits a significant decrease as temperature increases towards $T_c$ [@ZH_LiuPRB2011]. This results suggests that the unconventional behaviour of the SC gap, with the gap filling rather than closing with temperature increasing, is associated with the lost of coherence of the quasiparticle at high-temperature, a characteristic also shared with many cuprates. Scaling of the superconducting gap size with doping --------------------------------------------------- A quite trivial but yet very important remark concerning the SC gap that can be made at least for hole-doped 122-pnictides is the scaling of the SC gap size with $T_c$ as doping varies [@Nakayama_EPL2009; @YM_Xu_UD], as verified for all Fermi surfaces for $T_c\geq 22$ K. This result is opposite to the trend reported in some ARPES studies on cuprates and more consistent with recent works suggesting that the SC gap size in underdoped cuprates at the tip of the nodal arc scales with $T_c$ [@HufnerRPP2008]. Essentially, this suggests that the SC gap size is controlled by the pairing amplitude. However, such observation does not easily extend to the electron-doped side of the 122-pnictides. Indeed, a switch in the pairing amplitude ($2\Delta/k_BT_c$) from the $\alpha$ to the $\beta$ Fermi surfaces has been reported [@Terashima_PNAS2009]. The relative gap amplitude between the $\beta$ FS and the electronlike FSs at the M point is also different in electron-doped 122-pnictides compared to their hole-doped counter parts. Within the local antiferromagnetic exchange pairing picture, this could be explained by the fact that the gap size is not determined for each Fermi surface separately but rather depends on their momentum locations, as defined by global gap parameters determined by local antiferromagnetic exchange constants. In the relatively narrow doping range for which the scaling of the gap size with $T_c$ has been checked, the Fermi surface areas vary (and thus the SC gap size) but perhaps not enough to affect the scaling. Pseudogap --------- One of the major problems in the study of cuprates investigated by ARPES is the presence of a pseudogap [@DingNature1996]. Indeed, there is a suppression of density of states near the Fermi level that persists at the antinode up to a temperature $T^*$ which is much higher than $T_c$. Such behaviour is not observed near the nodes. Interestingly, a pseudogap behaviour is also observed in Ba$_{0.75}$K$_{0.25}$Fe$_2$As$_2$ [@YM_Xu_UD]. The corresponding feature is detected at 18 meV on the quasi-nested $\alpha$ FS. As shown in Figure \[PG\](a), the pseudogap feature exhibits some momentum dependence along the $\alpha$ band. While the size of the pseudogap is isotropic, the feature is more pronounced along the $\Gamma$-X direction, where the Fermi surface topology suggests a stronger antiferromagnetic scattering due to a better quasi-nesting between the $\alpha$ and $\delta$ bands, as illustrated in Figure \[PG\](f). Consistently, the pseudogap is absent on the unnested $\beta$ FS (see Figure \[PG\](e)). ![\[PG\] (Colour online) (a) Symmetrized EDCs of Ba$_{0.75}$K$_{0.25}$Fe$_2$As$_2$ pnictide measured below and above $T_c$ at two different locations on the $\alpha$ FS (points no. 1 and no. 2 in panel (b)). The red arrow indicates that the PG is $\sim18$ meV, and the dashed vertical line shows that the SC gap on the $\alpha$ FS is $\sim8$ meV. (b) Schematic FS plot near the $\Gamma$ point indicating the measurement locations of spectra presented in (a, c and e). (c) $T$ dependence of the symmetrized EDCs of Ba$_{0.75}$K$_{0.25}$Fe$_2$As$_2$ measured at point no. 1 on the $\alpha$ FS above $T_c$. The vertical dashed line indicates the energy scale of the PG (18 meV). The shaded regions represent the spectral weight loss in the PG state. It is obtained by subtracting the symmetrized EDCs from a quadratic background. (d) T dependence of the relative weight loss (normalized by the one obtained at $T$ = 45 K) in the PG state of Ba$_{0.75}$K$_{0.25}$Fe$_2$As$_2$. The error bars represent the uncertainty in calculating the relative weight loss. (e) Similar as (a) but for the $\beta$ FS (points no. 3, no. 4 and no. 5). The dashed line shows the SC gap on the $\beta$ FS (4 meV). (f) Electron-like FS contours of Ba$_{0.75}$K$_{0.25}$Fe$_2$As$_2$ shifted to the $\Gamma$ point by the antiferromagnetic wave vector. Note that this figure is described in the 2 Fe/unit cell notation. From Xu *et al.*, Nat. Commun., **2**, 392 (2011) [@YM_Xu_UD], copyright ©(2011) by Macmillan Publishers Ltd.](PG_low){width="8cm"} As illustrated in Figure \[PG\](a), the pseudogap is already present in the SC state. With temperature increasing, spectral weight lost is slowly recovered until temperature reaches $T^*=115$ K, above which no obvious change in the spectrum is detected, as shown in Figure \[PG\](c). The precise temperature evolution of the spectral weight loss is given in Figure \[PG\](d). Combined with the evolution of the spectral lineshape with underdoping, the results of Xu *et al.* on Ba$_{1-x}$K$_{x}$Fe$_2$As$_2$ indicate strong similarities with the cuprates that could origin from the importance of antiferromagnetic scattering in these two systems. In particular, the authors can identify two types of regions: i) regions where there is a smaller SC gap, no pseudogap surviving way above $T_c$ and no loss of integrity of the quasiparticle peak upon underdoping. These regions correspond to the unnested Fermi surfaces in the Fe-based superconductors and to the nodal region in the cuprates; 2) regions where the SC gap amplitude is larger and where a pseudogap surviving at high temperature is found, along with a loss of integrity of the quasiparticle peaks with underdoping. These regions correspond with regions associated with stronger antiferromagnetic scattering, namely the quasi-nested Fermi surfaces in the Fe-based superconductors and the antinodal region in the cuprates, both of which are connected by the antiferromagnetic wave vector. Discussion {#section_discussion} ========== With its ability to *resolve directly* the electronic structure, the Fermi surface and the SC gap amplitude in the momentum space, there is no doubt that ARPES is a unique and powerful tool to investigate the electronic properties of multi-band systems such as the Fe-based superconductors. Even though strictly speaking ARPES measures the electronic states in the very few layers below the surface, we presented in this review sufficient consistent results on different crystal structures and cleaved surfaces to prove empirically that ARPES is at the very least a good first order representation of the electronic states of most bulk Fe-based superconductors. For instance, the momentum dependence observed along the direction perpendicular to the cleaved surface is a strong indication that the probe states are not confined to the surface. Nevertheless, ARPES is not the only experimental technique that can be used to get some insight on the electronic behaviour of these materials and at this point we would like to make a comparison regarding the symmetry of the order parameter between ARPES and a bulk transport probe, namely thermal conductivity. The ARPES results on the SC order parameter are consistent, as suggested by Figure \[polar\] for a few materials (we could also include the 111-pnictides [@ZH_LiuPRB2011] and the 11-chalcogenides [@Nakayama_PRL2010; @Miao2011]): (i) the SC gap size is Fermi surface dependent; (ii) the SC gap size is in the strong coupling regime, with typical $2\Delta/k_BT_c$ ratios of 5 to 8; (iii) the SC gap size is isotropic or in the worst case nearly isotropic along each Fermi surface taking separately, except perhaps for a small warping along the $k_z$ direction, as observed in optimally-doped Ba$_{1-x}$K$_{x}$Fe$_2$As$_2$ [@YM_Xu_NPhys2011]. In other words, all the Fermi surfaces are fully gapped, without any nodes. It is generally accepted by now that the SC state in the Fe-based superconductors competes with long-range antiferromagnetism. This widespread belief is reinforced by the proximity of the SC and magnetic regimes in the phase diagram. In addition, several theoretical models propose that superconductivity is mediated by short-range antiferromagnetic fluctuations. Consequently, it is very natural to expect that the application of an external magnetic field might not simply suppress superconductivity like in conventional superconductors, but also be accompanied by unwanted secondary effects such as the modulation of the antiferromagnetic fluctuations and the modification of the Fermi surface topology. For this simple reason, we prefer to first avoid to compare ARPES with experiments that necessitate the use of an external magnetic field. Thermal conductivity ($\kappa(T)$) is an experimental probe that does not necessitate the use of an external magnetic field. It usually contains a term linear in temperature that has an electronic origin, and a higher power law term attributed to phonons and other excitations. Near absolute zero temperature, all carriers are paired. Since Cooper pairs cannot carry entropy, the electronic term of the thermal conductivity must vanish, unless some Fermi surfaces are partly or entirely not gapped, in which case $\kappa(T=0)\neq 0$. We recall that this condition requires that the samples measured have high quality and are not phase separated. With temperature increasing, quasiparticles thermally excited above the full SC energy gap $2\Delta$ can eventually contribute to the thermal conductivity. Although thermal conductivity does not allow to directly locate nodes or non-gapped Fermi surfaces in the momentum space in the case of multiple Fermi surface sheets, it is an efficient tool to reveal the existence of such features. It is also important to add that as with ARPES, thermal conductivity in zero field does not rely on any model to claim on the presence of nodes and it is immune to localized states induced by impurities, which is a huge advantage over several other transport techniques that enhances the reliability of the conclusions. In agreement with the result first demonstrated by ARPES experiments [@Ding_EPL; @L_Zhao], thermal conductivity shows a negligible $\kappa(T\rightarrow 0)/T$ term in optimally-doped Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$, indicating that all Fermi surfaces are fully gapped [@XG_LuoPRB2009]. In fact, that result is still valid down to $x=0.16$ for in-plane and out-of-plane thermal conductivity in zero field [@Reid2011], range that extends below the lowest $x$ value for which ARPES measurements have been performed so far. Below that range, Reid *et al.* found a sudden increase in the zero field residual $\kappa(T\rightarrow 0)/T$ [@Reid2011]. It is still debated whether this strange behaviour is intrinsic or caused by phase separation at these low dopings. Unfortunately, considering that the gap size becomes small in this region, that the critical temperatures are low compared to typical temperatures used in ARPES and the the quasiparticle coherence weakens with underdoping [@YM_Xu_UD], it is unlikely that ARPES can soon provide a definite answer on this issue. It is also not possible for ARPES to check the claim of nodes in the low-$T_c$ compound KFe$_2$As$_2$ [@JK_DongPRL2010] due to energy resolution and temperature limitations. However, the Fermi surface topology measured by ARPES for this system is quite different at the M point from that of the other pnictides [@Sato_PRL2009], and from these results it is conceivable that this particular Fermi surface topology leads to a gap structure at the M point that includes nodes and strong anisotropy. Thermal conductivity measurements were also performed on the electron-doped side of the 122-pnictides. As with their hole-doped counterparts, the electron-doped 122-pnictides exhibit a negligible in-plane $\kappa(T\rightarrow 0)/T$ term [@L_DingNJP2009; @TanatarPRL2010; @ReidPRB2010], in agreement with the nodeless SC gaps determined from ARPES measurements [@Terashima_PNAS2009]. The situation is different for the out-of-plane thermal conductivity, which suggests the presence of nodes [@ReidPRB2010]. Unfortunately, gap measurements by ARPES have never been reported so far along the $k_z$ direction for the electron-doped 122-pnictides, and a direct comparison is thus impossible. Yet, the ARPES results at a fixed $k_z$ indicate that if there are nodes in Ba$_{1}$Fe$_{2-x}$Co$_{x}$As$_2$, they must be confined to a limited $k_z$ range. Fe-based superconductors other than the 122-pnictides have also been investigated by thermal conductivity. For instance, Tanatar *et al.* measured the in-plane and out-of-plane thermal conductivity of LiFeAs [@Tanatar2011]. They find no important residual $\kappa(T\rightarrow 0)/T$ term in zero field, suggesting the absence of nodes in the gap structure, in agreement with ARPES gap measurements in LiFeAs [@BorisenkoPRL2010] and NaFe$_{0.95}$Co$_{0.05}$As [@ZH_LiuPRB2011]. Similarly, Dong and coworkers found only a very small $\kappa(T\rightarrow 0)/T$ residual term in zero field measurements of FeSe$_x$ $(x\sim 1)$ by thermal conductivity [@JK_DongPRB2009]. The authors conclusion supports multi-gap $s$-wave superconductivity. Although the sample composition is different from that used by ARPES in Nakayama *et al.* [@Nakayama_PRL2010] and Miao *et al.* [@Miao2011], the results are at least not contradictory. At this point we can summarize this discussion by saying that whenever ARPES and zero field thermal conductivity data have been recorded in similar conditions, the results are compatible. Now what happens when we add the magnetic field that we deliberately avoided so far to prevent any possible misinterpretation of the experimental data? Unfortunately, the interpretation of thermal conductivity in magnetic field is rarely straightforward for a multi-band system. One can compare the field dependence with that of other materials, but comparisons may be awkward if the systems have different Fermi surface topologies, which is typically the case, or if a magnetic ground state competes with superconductivity in some cases and not in the others. Alternatively, one can base his interpretation on theoretical models, with the strong assumption that these models are valid. As a rule of thumb though, it is reasonable to admit that the presence of weakly gap states will imprint a fast rising field dependence in the low temperature thermal conductivity, as these states get more easily excited. Although no strong SC gap anisotropy is claimed from thermal conductivity measurements in magnetic field on LiFeAs [@Tanatar2011] and FeSe [@JK_DongPRB2009], the situation is more ambiguous for the 122-pnictides. No strong anisotropy is claimed in Ba$_{0.6}$K$_{0.4}$Fe$_2$As$_2$ for $x>0.16$, whereas nodes are suggested below that critical doping [@Reid2011], which has not been investigated by ARPES. For the electron-doped side, while magnetic field dependent data on Ba$_{1}$Fe$_{1.9}$Ni$_{0.1}$As$_2$ [@L_DingNJP2009] and underdoped BaFe$_{2-x}$Co$_{x}$As$_2$ [@TanatarPRL2010] indicate fully or nearly isotropic gaps, the rapid increase in $\kappa(H)$ in the overdoped regime rather suggests a deep gap minimum somewhere on the Fermi surface [@TanatarPRL2010]. In the ARPES measurements, the concepts of gap isotropy and gap anisotropy are well defined, the gap being measured directly in the momentum space around each Fermi surface taken separately. On the other hand, transport measurements do not have a real momentum resolution. It is possible, for example, to explain the thermal conductivity results in the over-doped regime of BaFe$_{2-x}$Co$_{x}$As$_2$ by claiming that the gaps around all Fermi surfaces remain isotropic while one of them, most likely associated to the vanishingly small $\beta$ Fermi surface pocket in this regime, becomes very small in a uniform fashion. Moreover, it worths to recall that unlike ARPES, transport techniques are rather sensitive to the overall gap structure around the Brillouin zone center, and depend not only on the gap size on the various Fermi surface sheets, but also on their momentum distribution with respect to the $\Gamma$ point. Hence, even though ARPES indicates isotropic SC gaps on each Fermi surface, the overall SC gap distribution with respect to the $\Gamma$ point, as measured by ARPES, is anisotropic in the Fe-based superconductors due to the Fermi surface topology. In addition, the interactions of the electronic states with an external magnetic field *must* obey selection rules that reflect the point group symmetry of the crystal and that of the overall Fermi surface, which is anisotropic. Before concluding this section, we would like to discuss another possible source of apparent discrepancy between ARPES and some interpretations of transport measurements. As mentioned in this review, the gap $\Delta$ in ARPES measurements is usually defined as half of the full gap in the electronic dispersion below the SC transition. This gap is the same that defines the strength of the pairing interaction, and the same that characterizes the precise Bogoliubov dispersion below $T_c$. In that sense, it can arguably be called the “true gap". However, the lifetime of the SC quasiparticles in the Fe-based superconductors, characterized by a scattering rate $\Gamma_k$, is quite short compared to that of conventional superconductors, which also means that $\Gamma_k$ is large. Consequently, there is an in-gap residual density of states whose extension may be roughly represented by the leading edge position $\delta_{LE}(k)$ of the ARPES EDCs. Assuming infinite resolution and neglecting thermal broadening, $\delta_{LE}(k)=\Delta-\Gamma_k$, which illustrates that the leading edge is necessarily an underestimation of the “true gap" size. If ever transport measurements are sensitive to this residual density of states near $E_F$, the symmetry of the order parameter determined by these techniques would reflect the symmetry of $\delta_{LE}(k)$ rather than that of the “true gap". Since $\Gamma_k$ is momentum-dependent and depends, for instance, on the nearly-elastic components of the antiferromagnetic scattering, we should expect a non-negligeable momentum dependence for $\delta_{LE}(k)$, especially in the presence of an applied magnetic field. Conclusion {#section_conclusion} ========== As compared with the high-$T_c$ cuprates, the progress of our understanding of the ARPES measurements on the Fe-based superconductors during the past 4 years has been extremely fast. Interestingly, there is also much less discrepancies found in literature among work from different groups. Even though there is still no consensus on the nature of superconductivity in these materials, the contribution to the current knowledge provided by ARPES is significant and includes the precise determination of the Fermi surface topology of different compounds, the observation of electronic renormalization and band dispersion anomalies, as well as the characterization of the size and the symmetry of the SC order parameter. Thus, we can draw some very important conclusions among which are: i) the Fermi surface of the Fe-based superconductors is composed of several Fermi surface pockets; ii) the electronic dispersion of these materials is more three-dimensional than most copper oxide superconductors; iii) the band structure exhibits a non-negligible renormalization compared to LDA calculations that is more pronounced for the low-energy states near the Fermi level; iv) the SC gap of these materials is FS sheet dependent but isotropic or nearly isotropic on each FS; v) the pairing strength determined from the SC gap size indicates a pairing in the strong coupling regime; vi) antiferromagnetic scattering plays an important role in these materials. Furthermore, we provided evidence for consistency between ARPES and other types of measurement. Hopefully, the accumulation of experimental data from ARPES and other probes will soon pave the way to a global understanding of the key features leading to the exotic electronic behaviour of the Fe-based superconductors. Acknowledgments {#acknowledgments .unnumbered} =============== P.R. and H. D. are thankful to the Chinese Academy of Sciences (grant No. 2010Y1JB6), the Ministry of Science and Technology of China (grants No. 2010CB923000, No. 2011CBA00101), and the Nature Science Foundation of China (grants No. 10974175, No. 11004232, and No. 11050110422). T.S, K.N. and T.T. are also thankful to the MEXT of Japan, JSPS, and TRIP-JST. References {#references .unnumbered} ========== [^1]: We note that Borisenko *et al.* use the word “nesting" in reference [@BorisenkoPRL2010] in a much stricter sense than the concept of quasi-nesting defined in Section \[section\_doping\], and thus claim that there is no nesting at all in LiFeAs.

--- abstract: 'Inclusive transverse momentum spectra of primary charged particles in Pb–Pb collisions at $\sqrt{s_{_{NN}}}$ = 2.76 TeV have been measured by the ALICE Collaboration at the LHC. The data are presented for central and peripheral collisions, corresponding to 0–5% and 70–80% of the hadronic Pb–Pb cross section. The measured charged particle spectra in $|\eta|<0.8$ and $0.3 < p_T < 20$ GeV/$c$ are compared to the expectation in pp collisions at the same $\sqrt{s_{_{NN}}}$, scaled by the number of underlying nucleon–nucleon collisions. The comparison is expressed in terms of the nuclear modification factor $R_{AA}$. The result indicates only weak medium effects ($R_{AA} \approx $ 0.7) in peripheral collisions. In central collisions, $R_{AA}$ reaches a minimum of about 0.14 at $p_T=6$–$7$ GeV/$c$ and increases significantly at larger $p_T$. The measured suppression of high–$p_T$ particles is stronger than that observed at lower collision energies, indicating that a very dense medium is formed in central Pb–Pb collisions at the LHC.' title: 'Suppression of Charged Particle Production at Large Transverse Momentum in Central Pb–Pb Collisions at $\sqrt{s_{_{NN}}} = 2.76$ TeV ' ---       Acknowledgements {#acknowledgements .unnumbered} ================ [9]{} L. Evans and P. Bryant (editors), JINST [**3**]{} (2008) S08001. I. Arsene [*et al.*]{} \[BRAHMS Collaboration\], Nucl. Phys. A [**757**]{} (2005) 1. B.B. Back [*et al.*]{} \[PHOBOS Collaboration\], Nucl. Phys. A [**757**]{} (2005) 28. J. Adams [*et al.*]{} \[STAR Collaboration\], Nucl. Phys. A [**757**]{} (2005) 102. K. Adcox [*et al.*]{} \[PHENIX Collaboration\], Nucl. Phys. A [**757**]{} (2005) 184. E. Wang and X. N. Wang, Phys. Rev. Lett.  [**89**]{} (2002) 162301. R. Baier and D. Schiff, JHEP [**0609**]{} (2006) 059. S. Wicks, W. Horowitz, M. Djordjevic and M. Gyulassy, Nucl. Phys.  A [**784**]{} (2007) 426. I. Vitev, Phys. Lett.  B [**639**]{} (2006) 38. K. J. Eskola, H. Honkanen, C. A. Salgado and U. A. Wiedemann,’ Nucl. Phys.  A [**747**]{} (2005) 511. M. Miller, K. Reygers, S. Sanders, and P. Steinberg, Ann. Rev. Nucl. Part. Sci. [**57**]{} (2007) 205. K. Aamodt [*et al.*]{} \[ALICE Collaboration\], JINST [**3**]{} (2008) S08002. J. Alme [*et al.*]{} \[ALICE Collaboration\], Nucl. Instrum. Meth. A [**622**]{} (2010) 316. K. Aamodt [*et al.*]{} \[ALICE Collaboration\], [*accepted for publication in Phys. Rev. Lett.*]{}, arXiv:1011.3916 \[nucl-ex\] and K. Aamodt [*et al.*]{}, \[ALICE Collaboration\] [*to be published in Phys. Rev. Lett.*]{}. H. de Vries, C.W. De Jager and C. de Vries, Atomic Data and Nuclear Tables, [**36**]{} (1987) 495. X.-N. Wang and M. Gyulassy, Phys. Rev. D [**44**]{} (1991) 3501; W.-T. Deng, X.-N. Wang, and R. Xu, (2010), arXiv:1008.1841 \[hep-ph\]. R. Brun [*et al.*]{}, CERN Program Library Long Write-up, W5013, GEANT Detector Description and Simulation Tool (1994). R. Brun [*et al.*]{}, \[ALICE Collaboration\], Nucl. Instrum. Meth. A [**502**]{} (2003) 339. K. Aamodt [*et al.*]{} \[ALICE Collaboration\], Phys. Lett. B [**693**]{} (2010) 53. ALICE Collaboration, [*to be published*]{}. R. Hagedorn, Riv. Nuovo Cim. [**6**]{} (1983) 1. R. Engel, J. Ranft, S. Roesler, Phys. Rev. D[**52**]{} (1995) 1459. T. Sjöstrand, S. Mrenna, P. Skands, J. High Energy Phys. 0605 (2006) 026. M. Albrow et al., Tevatron-for-LHC Conference Report of the QCD Working Group, Fermilab-Conf-06-359, hep-ph/0610012; T. Sjöstrand and P. Skands, Eur. Phys. J. C [**39**]{} (2005) 129. P. Skands, Contribution to the 1st International Workshop on Multiple Partonic Interactions at the LHC, Perugia, Italy, Oct. 2008, Fermilab-Conf-09-113-T, arXiv:0905.3418 \[hep-ph\]. T. Aaltonen [*et al.*]{} \[CDF Collaboration\], Phys. Rev. D [**79**]{} (2009) 112005. R. Sassot, P. Zurita, and M. Stratmann, Phys. Rev. D [**82**]{} (2010) 074011. B. I. Abelev [*et al.*]{} \[STAR Collaboration\], Phys. Rev. Lett.  [**97**]{} (2006) 152301. S. S. Adler [*et al.*]{} \[PHENIX Collaboration\], Phys. Rev.  C [**69**]{} (2004) 034909. S. S. Adler [*et al.*]{} \[PHENIX Collaboration\], Phys. Rev.  C [**69**]{} (2004) 034910. J. Adams [*et al.*]{} \[STAR Collaboration\], Phys. Rev. Lett.  [**91**]{} (2003) 172302. R. J. Fries and B. Muller, Eur. Phys. J.  C [**34**]{} (2004) S279. The ALICE Collaboration {#app:collab} =======================

--- abstract: 'Local temperature defined by a local canonical state of the respective subsystem, does not always exist in quantum many body systems. Here, we give some examples of how this breakdown of the temperature concept on small length scales might be observed in experiments: Measurements of magnetic properties of an anti-ferromagnetic spin-1 chain. We show that those magnetic properties are in fact strictly local. As a consequence their measurement reveals whether the local (reduced) state can be thermal. If it is, a temperature may be associated to the measurement results, while this would lead to inconsistencies otherwise.' author: - 'M. Hartmann' - 'G. Mahler' title: Measurable Consequences of the Local Breakdown of the Concept of Temperature --- With the advent of nanotechnology, the applicability of thermodynamics on small length scales has become an important subject of current research. There has been substantial progress in the fabrication and processing of nanostructures, e.g. carbon nanotubes [@Dresselhaus2001], and measurements of thermal properties on the corresponding scales are becoming possible [@Gao2002]. However, despite this amazing experimental advances, the theoretical foundation of thermodynamics on small scales remains unclear [@Cahill2003; @Allahverdyan2002]. First attempts to generalize thermodynamics such that it would apply to small systems date back to the sixties and are receiving increasing attention today [@Hill2001c; @Rajagopal2004]. Recently [@Hartmann2003; @Hartmann2003b; @Hartmann2004a], Hartmann, Mahler and Hess gave a first quantitative estimate of the length scales, below which the standard concept of temperature ceases to exist. They considered a large chain of particles with nearest neighbor interactions, which was assumed to be in a thermal, i.e. canonical, state. In this scenario, subgroups of $N$ adjacent particles are in a canonical state if $N$ is larger than a threshold value $N_{\text{min}}$ which depends on the global temperature, $N_{\text{min}} = N_{\text{min}}(T)$. The reduced density matrix of a subgroup deviates from the canonical form if $N < N_{\text{min}}$. The temperature dependence of $N_{\text{min}}$ is as follows: The lower the temperature, the larger $N_{\text{min}}$, i.e. the larger the groups need to be. In other words, for each group size $N$, there is a threshold temperature above which the groups are in a canonical state and below they are not. The predicted length scales depend on the definition of temperature used. In [@Hartmann2003b] and [@Hartmann2004a], local temperature has been defined to exist if the respective part of the system is in a canonical state. Why should we care about the non-existence of local temperature? There are at least three situations for which this possibility needs special attention: One obvious scenario refers to the limit of spatial resolution on which a temperature profile could be defined. However a spatially varying temperature calls for non-equilibrium - a complication which we will exclude here. A second application deals with partitions on the nanoscale: If a modular system in thermal equilibrium is partitioned into two pieces, say, the two pieces need no longer be in a canonical state, let alone have the same local temperature. Temperature is always measured indirectly via observables, which, in quantum mechanics, are represented by hermitian operators. Usually, one is interested in measuring the temperature of a system in a stationary state. The chosen observable should therefore be a conserved quantity, i.e. its operator should commute with the Hamiltonian of the system. A conventional technique, e.g., is to bring the piece of matter, the temperature $T$ of which is to be measured, in thermal contact with a box (volume $V$) of an ideal gas (number of particles $n$) and to measure the pressure $p$ of the gas, which is related to its temperature by $n \, k_B \, T = p \, V$ ($k_B$ is Boltzmann’s constant). Since the gas is in thermal equilibrium with the considered piece of matter, both substances have the same temperature. A measurement of $p$ for constant $V$ allows to infer the global temperature $T$ of the piece of matter. Such a thermometer functions as long as it does not significantly perturb the measured object (weak coupling), irrespective of whether its coupling to the object is local or not. One might thus doubt whether a local temperature can be measured at all. System bath models show that the system always relaxes to a canonical state with the global temperature of the bath, no matter how localized the system bath interaction might be [@Weiss1999]. However, there are examples, where a local application of Boltzmann-Gibbs thermostatistics is known to fail due to strong correlations [@Netz; @Garcia-Morales]. Here, we consider observables of the object itself, which can be used to measure local temperatures $T_{\text{loc}}$, i.e. temperatures of subsystems, provided the subsystems are in a canonical state. In turn, if the respective subsystems are not in a canonical state, this fact should modify the measurement results for those observables. For systems composed of weakly interacting subsystems, which are in a thermal state, the total state factors with respect to the subsystems and the subsystems themselves are also in thermal states, i.e. local temperature exists. For strongly interacting subsystems, however, local temperatures cease to exist due to correlations between subsystems. Note that, if the system reached its thermal state via interaction with a bath, the coupling to the latter must be weak, otherwise the state of the entire system can, in general, not be thermal [@Allahverdyan2002; @Weiss1999]. Pertinent systems, for which such effects can easily by studied, are magnetic materials. As we will demonstrate below, properties of single spins can be infered from measurements of even macroscopic magnetic observables. These materials thus allow to study the existence of temperature, as defined by the existence of a canonical state, on the most local scale possible, i.e. for single spins. As our model, we consider a homogeneous chain of spin-1 particles interacting with their nearest neighbors. For the interactions, we assume a Heisenberg model. The Hamiltonian of this system reads [@vanVleck1945]: $$\label{eq:1} H = B \, \sum_{j=1}^n \sigma_j^z + J \, \sum_{j=1}^n \vec{\sigma}_j \cdot \vec{\sigma}_{j+1} ,$$ where $\vec{\sigma}_j = (\sigma_j^x, \sigma_j^y, \sigma_j^z)$ and $\vec{\sigma}_j \cdot \vec{\sigma}_{k} = \sigma_j^x \sigma_k^x + \sigma_j^y \sigma_k^y + \sigma_j^z \sigma_k^z$. $\sigma_j^x$, $\sigma_j^y$ and $\sigma_j^z$ are the spin-$1$ matrices. $B$ is an applied magnetic field, $J$ the coupling and $n$ the number of spins. The coupling $J$ is taken to be positive, $J > 0$. The spins thus tend to align anti-parallel and the material is anti-ferromagnetic. The local Hamiltonian of subsystem $j$ is $H_j = B \, \sigma_j^z$. The system has periodic boundary conditions and is thus translation invariant. We assume that the system (\[eq:1\]) is in a thermal state, $$\label{eq:6} \rho = \frac{1}{Z} \, \exp \left(- \, \frac{H}{k_B \, T} \right) ,$$ where $T$ is the temperature and $Z$ the partition sum. The interaction of this system to a possible heat bath acting as a thermostat must be weak [@Allahverdyan2002; @Weiss1999]. Note that the temperature $T$ is a global property of the system since it is defined in the eigenbasis. The eigenstates do not factorize with respect to single spins and therefore have non-local properties. The reduced density operator of subsystem $j$ may be represented in the eigenbasis of the respective subsystem Hamiltonian, $H_j$; let its diagonal matrix elements be $p_{\alpha}$. It is convenient to introduce a spectral temperature $T_{\text{spec}}$, which would coincide with $T_{\text{loc}}$ if the local state was canonical but which formally exists for any state: $$\label{specdef} \frac{1}{T_{\text{spec}}} \equiv - k_B \, \sum_{{\alpha} > 0} \frac{p_{\alpha}}{1 - p_0} \: \frac{\ln (p_{{\alpha}}) - \ln (p_{0})}{E_{{\alpha}} - E_0} .$$ Here the $E_{\alpha}$ are the spectrum of the isolated subsystem; $E_0$ denotes the ground state. The factor $\left(1 - p_0\right)^{-1}$ is the normalization. The local state may be characterized by $T_{\text{spec}}$ and a parameter $\Delta$ describing mean relative square deviations of the occupation probabilities $p_{\alpha}$ from those of a canonical state, $p_{\alpha}^c$, with $T_{\text{loc}} = T_{\text{spec}}$, $$\label{deltadef} \Delta^2 \equiv \sum_{\alpha} \, p_{\alpha} \: \left( \frac{p_{\alpha} - p_{\alpha}^c}{p_{\alpha}} \right)^2 ,$$ where $p_{\alpha}^c = \exp \left(- E_{\alpha} / k_B \, T_{\text{spec}}\right) / \sum_{\alpha} \exp \left(- E_{\alpha} / k_B \, T_{\text{spec}}\right)$. Note that $T_{\text{spec}}$ and $\Delta$ depend on the global temperature $T$ [**and**]{} the type and strength of subsystem interactions. $T_{\text{spec}}$ and $\Delta$ do, of course, not fully characterize the local state (this would require 8 real numbers), they merely classify it in a pertinent way. Figure \[fig:2\] shows the spectral temperature $T_{\text{spec}}$, the global temperature $T$ and the deviations $\Delta$ of the local state from a canonical state with $T_{\text{spec}}$ as a function of $T$ for a spin-$1$ chain of 4 particles with the Hamiltonian (\[eq:1\]). While the deviations $\Delta$ are small at high $T$, they become larger for low $T$, where the spectral temperature $T_{\text{spec}}$ starts to rise again as $T$ is lowered further. For $T = 0$, $\Delta$ vanishes, since a completely mixed state corresponds to a canonical one with $T_{\text{spec}} \rightarrow \infty$. A signature of these local deviations from a canonical state ($\Delta \not= 0$) can be measured. As an example of such an experiment, we will now consider two different magnetic observables of a spin-$1$ system with the Hamiltonian (\[eq:1\]). The first observable is the magnetization in the direction of the applied field, $m_z$, which we define to be the total magnetic moment per particle: $$\label{eq:2} m_z \equiv \frac{1}{n} {\left\langle}\sum_{j=1}^n \sigma_j^z {\right\rangle},$$ where ${\left\langle}\mathcal{O} {\right\rangle}$ is the expectation value of the operator $\mathcal{O}$, i.e. ${\left\langle}\mathcal{O} {\right\rangle}= \text{Tr}(\rho \, \mathcal{O})$. In the translation invariant state $\rho$, the reduced density matrices of all individual spins are equal, and the magnetization (\[eq:2\]) can be written as $$\label{eq:3} m_z = {\left\langle}\sigma_k^z {\right\rangle},$$ for any $k = 1, 2, \dots, n$. The magnetization, although defined macroscopically, is thus actually a property of a single spin, i.e. a strictly local property. As our second observable we choose the occupation probability, $p$, of the $s_z = 0$ level (averaged over all spins), $$\label{eq:5} p \equiv \frac{1}{n} {\left\langle}\sum_{j=1}^n \left{\vert}0_j {\right\rangle}{\left\langle}0_j \right{\vert}{\right\rangle}= {\left\langle}\left{\vert}0_k {\right\rangle}{\left\langle}0_k \right{\vert}{\right\rangle},$$ where the second equality holds for the same reasons as for $m_z$ for any $k = 1, 2, \dots, n$. $p$ is thus strictly local, too. Now, if each single spin was in a canonical state with $\Delta = 0$ and a temperature $T_{\text{loc}} = T_{\text{spec}}$, $m_z$ and $p$ would both have to be monotonic functions of $T_{\text{loc}}$. Consequently, $T_{\text{loc}}$ could, after calibration, be infered from measurements of $m_z$ or $p$. Note, that $m_z$ is proportional to the local energy, the average energy of one subsystem. Figure \[fig:1\], shows $m_z$ and $p$ as a function of the global temperature $T$ for a spin-$1$ chain of 4 particles with the Hamiltonian (\[eq:1\]) for weak interactions, $J = 0.1 \times B$. Both quantities are monotonic functions of each other. The situation changes drastically when the spins are strongly coupled. In this case the concept of temperature breaks down locally due to correlations of each single spin with its environment. Figure \[fig:3\] shows $m_z$ and $p$ as a function of temperature $T$ for a spin-$1$ chain of 4 particles with the Hamiltonian (\[eq:1\]) for strong interactions $J = 2 \times B$. Both quantities are non-monotonic functions of $T$ and therefore no mapping between $m_z$ and $p$ exists. Note that, although the present plots are done for $4$ spins only, the characteristics do not change for larger numbers of spins (The exact quantitative values are not needed for our reasoning). We have studied this with chains of $4$, $5$ and $6$ spins. In the case of $m_z$, the characteristics have even been already observed in experiments [@Avenel1992]. How could a local observer determine whether the system he observes, a single spin, is in a thermal state and can therefore be characterized by a temperature? The local observer needs to compare two situations: In the first situation, the spin is weakly coupled to a larger system, the heat bath. In this situation, the local observer could measure $m_z$ and $p$ as functions of the temperature of the heat bath and would get a result similar to figure \[fig:1\]. This result would not be sensitive to the details of the coupling to the heat bath. The local observer would thus recognize this situation as a particular one and might term it the “thermal” situation. The second situation is fundamentally different. The spin is now strongly coupled to its surrounding. If the local observer again measures $m_z$ and $p$ as functions of the temperature of the surrounding, he gets the result of figure \[fig:3\]. The observer can tell the difference between both situations, even if he has no access to the temperature $T$ of the surrounding. In the first case he can construct a mapping from say $m_z$ to $p$ or vice versa, in the second he cannot: Here the concept of a local temperature breaks down at least on the level of individual particles, since temperature measurements via different local observables would contradict each other. Finally, we address the question of whether the effects described here could be observed in real experiments. Indeed, pertinent experiments are available and have partly already been carried out: A realization of a quasi one dimensional anti-ferromagnetic spin-1 Heisenberg chain is the compound CsNiCl$_3$ [@Kenzelmann2001; @Affleck1989; @Avenel1992]. Here the coupling is $J \approx 2.3$ meV. To achieve a detectable modulation of $m_z$ and $p$, the spins should be significantly polarized for $T > 0$. Therefore a sufficiently strong applied magnetic field is needed. For CsNiCl$_3$, a field of roughly $9.8$ Tesla would correspond to $J = 4 \times B$. The magnetization in an applied field can be measured with high precision with a SQUID [@Lipa1981]. The occupation probability of the $s_z = 0$ states, on the other hand, is accessible via neutron scattering experiments [@Kenzelmann2002a; @Ma1995]. The differential cross section for neutron scattering of spin-systems is given by [@Lovesey1989] $$\label{eq:8} \frac{\partial^2 \sigma}{\partial \Omega \, \partial E_f} \sim \left| f(\vec{q})^2 \right| \frac{\left| \vec{k}_f \right|}{\left| \vec{k}_i \right|} \sum_{a, b} \left( \delta_{a, b} - \frac{\vec{q}_a}{\left|\vec{q}_a\right|} \cdot \frac{\vec{q}_b}{\left|\vec{q}_b\right|} \right) S^{a b} (\vec{q}, \omega) ,$$ where $E_i$, $E_f$, $\vec{k}_i$ and $\vec{k}_f$ are the initial and final energies and momenta of the scattered neutrons. $a, b = x, y, z$, $\omega = E_f - E_i$, $\vec{q} = \vec{k}_f - \vec{k}_i$ and $f(\vec{q})$ is the magnetic form factor, which can be found tabulated [@Brown1999]. The dynamic structure factor $S^{a b} (\vec{q}, \omega)$ is the Fourier transform of the spin-spin correlation function, $$S^{a b} (\vec{q}, \omega) = \frac{1}{2 \pi} \int dt \sum_{\vec{r}, \vec{r}'} e^{i \vec{q} (\vec{r} - \vec{r}') - i \omega t} {\left\langle}\sigma^a_{\vec{r}}(0) \, \sigma^b_{\vec{r}'}(t) {\right\rangle}.$$ If we only consider events, where the difference in momentum is along the z-axis, $\vec{q} = q_z \vec{e}_z$, only $S^{x x}$, $S^{x y}$, $S^{y x}$ and $S^{y y}$ contribute in equation (\[eq:8\]). Since the applied field $B$ is along the z-axis, $S^{x y}$ and $S^{y x}$ are zero due to symmetry reasons. Summing up over all $\vec{q}$ and all $\omega$ and using our knowledge of $\vec{k}_i$, $\vec{k}_f$ and $f(\vec{q})$, we can obtain information about the quantity $$\frac{1}{n} \, \sum_{\vec{r}} {\left\langle}\sigma^x_{\vec{r}}(0) \, \sigma^x_{\vec{r}}(0) + \sigma^y_{\vec{r}}(0) \, \sigma^y_{\vec{r}}(0) {\right\rangle}= 1 + p$$ from the measurement data. Therefore, $p$ is measurable in neutron scattering experiments. Such experiments or a combination thereof could thus be used to demonstrate the non-existence of local temperature. We note in passing that such “local” measurements are also an interesting tool to study local and non-local features like entanglement of thermal quantum states [@Jordan2003; @oConnor2001]. In summary, we have discussed possible “local” temperature measurements via magnetic properties of the considered material. We have studied a magnetic structure in a thermal state and analyzed whether its “local” temperature could be infered from the measurement of magnetic quantities. For a large structure, for which boundary effects can be neglected and thus a model with periodic boundary conditions is a valid description, the magnetization is a strictly local property implying that a temperature measurement based on it is strictly local, too. Local properties, in general, need not be monotonic functions of the global temperature and thus cannot serve as temperature indicators. The global temperature is not accessible by a local measurement. For networks of spin-1 or higher dimensional subsystems, two local but different quantities like the magnetization and the occupation probability of the $s_z = 0$ state, cannot be mapped onto each other. As a consequence, using these two quantities for temperature measurements would yield contradictive results. These measurements thus show, that the local state cannot be canonical and that therefore a local temperature does not exist for single spins. The existence of non-thermal properties within a modular system in thermal equilibrium has so far rarely been recognized. The popularity of a thermal description and the notion of temperature is based on the fact that various properties of the system uniquely scale with temperature. If these mappings do no longer exist, the concept of temperature looses its meaning. This fact might become relevant in future nanotechnologies, in so far as the behavior of a structure can no longer be predicted from a temperature that would characterize it. We thank M. Henrich, C. Kostoglou, M. Michel, H. Schmidt, M. Stollsteimer and F. Tonner for fruitful discussions. M. Hartmann wants to thank Shuichi Wakimoto for comments on the experimental techniques, Prof. Ulrich Schollwöck for comments on some pertinent experimental materials and, in particular, Prof. Ortwin Hess for discussions and support, which made this work possible. [21]{} . . .

--- abstract: 'A new algebraic multilevel algorithm for computing the second eigenvector of a column-stochastic matrix is presented. The method is based on a deflation approach in a multilevel aggregation framework. In particular a square and stretch approach, first introduced by Treister and Yavneh, is applied. The method is shown to yield good convergence properties for typical example problems.' author: - Lukas Polthier bibliography: - 'literatur.bib' nocite: '[@*]' title: 'Algebraic Multilevel Methods for Markov Chains[^1]' --- algebraic multigrid; smoothed aggregation; Markov chains; second eigenvector 65C40, 65F10, 65F15, 60J22 Introduction ============ The eigenvectors of Markov chains contain information about the important processes of a stochastic system. For the class of irreducible and aperiodic Markov chains, the eigenvector corresponding to the eigenvalue $1$ is unique and contains the limit information of the Markov chain called stationary distribution. Eigenvectors corresponding to eigenvalues close to $1$ encode the slow structural transitions, i.e. the essential information about the dynamics. Homogeneous, finite-state, discrete-time Markov chains are described by a column-stochastic transition matrix which can be analyzed by tools from linear algebra. Let $B \in \mathbb{R}^{n \times n}$ be a column-stochastic matrix, i.e. $1^TB = 1^T$, $B_{ij}\ge 0$. The problem of finding the second eigenvector can be written algebraically as $$\begin{aligned} Bx = \lambda_2 x, \label{Hauptgleichung}\end{aligned}$$ where we assume that the eigenvalues of $B$ fulfill $$\begin{aligned} 1 = \lambda_1 > \lambda_2 > |\lambda_k| \label{AssumptionOnEigenvalues}\end{aligned}$$ for all $k \ge 3$ and $\lambda_2 \in \mathbb{R}$. Note that $\lambda_2$ is not known a priori. In particular we can not simply solve a linear system of equations but must exploit the eigenvalue structure of $B$ to compute the second eigenvector. In relevant applications, Markov chains typically have a large number of variables which makes direct solvers to obtain the eigenvectors inapplicable. Instead one uses iterative solvers which tend to converge rather slowly. For many years, it is well known that algebraic multilevel methods can be used to drastically improve the convergence of iterative solvers [@BMR84]. However the classical theory of algebraic multilevel methods is applicable only to a certain type of matrices, e.g. symmetric positive-definite matrices [@RS87]. For Markov matrices, classical multilevel aggregation methods to compute the invariant measure converge rather slowly in numerical experiments. In [@DMM10] and [@TY10] more sophisticated approaches that can drastically speed-up the convergence of the multilevel aggregation method for the first eigenvector are proposed. Based on these ideas we present a new algorithm to compute the second eigenvector of a Markov matrix using a multilevel approach combined with a Wielandt deflation. Aggregation Multilevel for the Invariant Measure {#chapterAggregationMultilevelMethods} ================================================ In this section we recall aggregation multilevel methods for Markov chains to compute the invariant measure. These methods have been used in the literature [@DMM08; @HL94; @TY10]. The problem of finding the invariant measure, can be written as $$\begin{aligned} Bx = x\end{aligned}$$ with $x \neq 0$. For $A:=I-B$ we may equivalently solve $$\begin{aligned} Ax = 0 \label{EineGleichung}\end{aligned}$$ with $|x| = 1$. In fact we are looking for a non-trivial vector in the kernel of $A$. Here $A$ is an irreducible, singular M-matrix that has a strictly positive solution to (\[EineGleichung\]) that is unique up to scaling. This property is important for the well-posedness of the multilevel framework. Relaxation Methods {#sectionRelaxationMethods} ------------------ Two common relaxation methods for (\[EineGleichung\]) are Jacobi and power iteration. Assume that $B$ has eigenvalues $\lambda_i$ with right eigenvectors $v^{(i)}$ that are ordered as in (\[AssumptionOnEigenvalues\]). In the following we analyze how the eigenvectors govern the efficiency of power and Jacobi iteration. The former is given by the iteration $$\begin{aligned} x^{(k+1)} = Bx^{(k)}.\end{aligned}$$ Suppose the initial probability vector $x^{(0)}$ is spanned by the eigenvectors $v^{(i)} \in \mathbb{R}^n$ of $B$ for some coefficients $\alpha_i$. $$x^{(0)} = \sum_{i=1}^n \alpha_i v^{(i)}$$ Then the $k$-th iterate of power iteration can be expressed by $$\begin{aligned} \label{eigenvaluesIterationMatrix} x^{(k)} = B^k x^{(0)} = \sum_{i=1}^n \alpha_i B^k v^{(i)} = \sum_{i=1}^n \alpha_i \lambda_i^k v^{(i)} = \left[ \alpha_1 v_1 + \sum_{i=2}^n \alpha_i \lambda_i^k v^{(i)} \right].\end{aligned}$$ If the initial iterate $x^{(0)}$ was chosen such that $\alpha_1 \neq 0$, the power method will converge as $|\lambda_i| < 1$ for $i=2, \cdots, n$. Conversely, Jacobi iteration aims at solving the linear system $A\pi = 0$ rather than $B\pi = \pi$. We consider a splitting of the coefficient matrix $$A = D - (L + U)$$ into diagonal part $D$ and lower respectively upper triangular part $L, U$. In particular $D,L,U$ are non-negative matrices and we assume that $D$ is non-singular.\ Then, up to rescaling, Jacobi iteration with damping parameter $\omega \in (0,1]$ is given by $$\begin{aligned} x^{(k+1)} &= x^{(k)} - \omega D^{-1}Ax^{(k)},\end{aligned}$$ with iteration matrix $G_{dJac} = I - \omega D^{-1}A$. It is the eigenvalues of $G_{dJac}$ that determine the rate of convergence of the method. In particular we can think of Jacobi iteration as a power iteration with iteration matrix $G_{dJac}$.\ Smoothing Property ------------------ Particularly enlightening is the effect of the two iterative methods to the eigenvectors of the coefficient matrix. In equation (\[eigenvaluesIterationMatrix\]) we see how the eigenvalues define to which extend the iterative method is effective in reducing a particular error.\ This is also what we observe in the numerical experiments: In the left plot of figure \[figureSmoothErrors\] we see that damped Jacobi iteration is very effective in reducing errors which correspond to eigenvalues of $B$ close to $-1$. On the other hand, errors corresponding to eigenvalues of $B$ close to $1$ are hardly affected by the relaxation. In the right plot of figure \[figureSmoothErrors\] we see that power iteration is very effective in reducing errors corresponding to eigenvalues of $B$ with small absolute value. The eigenmodes corresponding to eigenvalues of $B$ with large absolute value are hardly affected by power iteration.\ Altogether the iterative methods do not reduce all error modes equally effective. That is why we speak of *smoothing* or *relaxation* methods in the sense that they “smooth” out some errors while leaving others largely unchanged. For symmetric, positive definite matrices, this effect called *smoothing property* can be defined more precisely [@Stu99 chapter 3]. The basic idea of multilevel methods is to reduce the smooth error components on a smaller linear system. Aggregation Multilevel Methods ------------------------------ As the exact solution $x$ to (\[EineGleichung\]) has strictly positive components, it is reasonable to assume strict positivity of an approximation $x^{(i)}$ as well. Then the multiplicative error of $x^{(i)}$ to $x$ is defined by $$\begin{aligned} e^{(i)} := \left(\text{diag}(x^{(i)}) \right)^{-1} x.\end{aligned}$$ Then a multiplicative error formulation of (\[EineGleichung\]) has the form $$\begin{aligned} A \text{ diag}(x^{(i)})e^{(i)} = 0, \label{fineLevelProbEq}\end{aligned}$$ where we have $e^{(i)} = 1 \in \mathbb{R}^n$ at convergence. From the theory on algebraic multilevel methods for positive definite matrices and the general experience with multilevel methods it turns out (see [@RS87]) that smooth errors $e$ have small variation along strong algebraic couplings. That means $e_i \approx e_j$ if $|A_{ij}|/A_{ii}$ is large. Let $n_c < n$ and assume there is a partition of $\{1, \cdots, n\}$ into $n_c$ aggregates $\varphi_j$. These aggregates shall be chosen by some aggregation algorithm that aims at aggregating strongly connected elements. Then, at least approximately, smooth errors are piecewise constant along the aggregates. In [@TY10] a bottom-up aggregation approach is proposed, which seems to be a reasonable choice for the numerical example problems and will be used throughout this paper. The choice of a particular aggregation method is typically motivated by heuristics and is not subject of this paper. In fact the choice of aggregates may depend on the current iterate. Once reasonable aggregates have been determined, the aggregation matrix $Q\in \{0,1\}^{n\times n_c}$ defined by $$Q_{ij} = \bigg \{ \begin{matrix} 1 & \text{if } i \in \varphi_j\\ 0 & \text{if } i \not\in \varphi_j. \end{matrix}$$ contains the information which fine level variables belong to which aggregate. In particular $Q$ has full rank and up to reordering it has the following form. $$\begin{aligned} Q = \left[\begin{array}{c | c | c} 1 & 0 & \cdots\\ \vdots & \vdots & \cdots\\ 1 & 0 & \cdots\\ \hline 0 & 1 & \cdots\\ \vdots & \vdots & \cdots\\ 0 & 1 & \cdots\\ \hline \vdots & \vdots & \ddots \end{array}\right]\end{aligned}$$ On the coarser level we then solve for the error-components that are piecewise constant on each aggregate $\varphi_i$. If we replace $e^{(i)}$ by $Qe_c$ and multiply equation (\[fineLevelProbEq\]) from the left by $Q^T$ we obtain the multiplicative error formulation on the coarse level $$\begin{aligned} Q^TA ~\text{diag}(x^{(i)}) Qe_c = 0. \label{coarseLevelErrEq}\end{aligned}$$ $e_c$ is a coarse level approximation of the (unknown) fine level multiplicative error $e^{(i)}$. I.e. $e_c$ is the multiplicative error of the restriction $Q^Tx^{(i)}$ of the fine level iterate to the (unknown) exact coarse level solution $x_c$. $$\begin{aligned} x_c = \text{diag}(Q^Tx^{(i)}) e_c\end{aligned}$$ Together this gives the coarse level probability equation $$\begin{aligned} Q^TA ~\text{diag}(x^{(i)}) Q \left(\text{diag}(Q^Tx^{(i)})\right)^{-1} x_c = 0 \label{coarseLevelProbEq}\end{aligned}$$ with exact solution $x_c$. For simplicity of notation we define the prolongation and restriction operators by $$\begin{aligned} R &:= Q^T \in \mathbb{R}^{n_c \times n}\\ P &:= \text{diag}(x^{(i)})Q \in \mathbb{R}^{n \times n_c}.\end{aligned}$$ In particular the current iterate $x^{(i)}$ is in the image of $P$. In terms of grid-transfer operators $R, P$ we define the coarse level operator $$\begin{aligned} A_c := RAP \label{coarseLevelOperatorDefinition}\end{aligned}$$ and coarse level stochastic matrix $$\begin{aligned} B_c := RBP \left (\text{diag}(Rx^{(i)}) \right )^{-1}. \label{DefBc}\end{aligned}$$ Now the coarse level multiplicative error equation (\[coarseLevelErrEq\]) can be formulated by $$\begin{aligned} A_c e_c = 0, \label{coarseLevelErrEq2}\end{aligned}$$ and the coarse level probability equation (\[coarseLevelProbEq\]) has the form $$\begin{aligned} A_c \left(\text{diag}(Rx^{(i)}) \right)^{-1} x_c = 0. \label{coarseLevelProbEq2}\end{aligned}$$ [@DMM10 theorem 3.1] The coarse level matrix $A_c$ is again an irreducible, singular M-matrix. In particular (\[coarseLevelErrEq2\]) has a unique solution. We also get $$\begin{aligned} \begin{split} A_c \left( \text{diag}(Rx^{(i)}) \right)^{-1} &= (R I P - R B P ) \left( \text{diag}(R x^{(i)}) \right)^{-1} \\ &= I_c - B_c \end{split} \label{ABI}\end{aligned}$$ where $I_c \in \mathbb{R}^{n_c\times n_c}$ denotes the identity matrix on the coarse level. Hence by going from the fine level to the coarse level each aggregate $\varphi_j$ gets assigned the sum of its fine level weights, while going from coarse to fine level we weight the fine level variables of an aggregate according to the relative proportion of the approximated weights $x^{(i)}$.\ By applying this approach in a recursive manner, we obtain the multilevel V-cycle described in algorithm AM (algebraic aggregation for Markov chains). \[thmFixedPointAggregation\] [@DMM10 theorem 3.2] The exact solution $x$ is a fixed point of the multilevel V-cycle, i.e. $AM(A,x) = x$. Deflated Square and Stretch Aggregation Multilevel for the Second Eigenvector ============================================================================= In this chapter we introduce the new method to compute the second eigenvector of a stochastic matrix using a multilevel aggregation approach. I.e. we want to solve (\[Hauptgleichung\]) where we assume that the eigenvalues fulfill (\[AssumptionOnEigenvalues\]). Recall that $\lambda_2$ is not known a priori. Wieland Deflation {#chapterWielandDeflation} ----------------- Consider the deflated matrix $$\begin{aligned} B_1 := B - \mu v^{(1)} u^T, \label{deflatedMatrix}\end{aligned}$$ where $u \in \mathbb{R}^n$ is some (for now arbitrary) deflation vector with $(v^{(1)},u)_{\mathbb{R}^n} = 1$ and $\mu \in \mathbb{R}$ is some (for now arbitrary) shift. Then the eigenvalues of $B_1$ are the same as those of $B$, except for the eigenvalue $\lambda_1$ that gets shifted by $\mu$. \[thmWielandtDeflation\] [@Saa11 theorem 4.2] Let $\{ v^{(1)}=\pi, v^{(2)}, \cdots , v^{(n)} \}$ denote the right eigenvectors corresponding to the eigenvalues $\{ \lambda_1, \lambda_2, \cdots, \lambda_n \}$ of a column-stochastic matrix $B$. Then the deflated matrix $B_1$ has the spectrum $$\begin{aligned} \sigma(B_1) = \{ \lambda_1 - \mu, \lambda_2 , \cdots ,\lambda_n \}.\end{aligned}$$ Moreover $B_1$ has the same left eigenvectors $u^{(i)}$ for $i=2,\cdots, n$ as $B$ and right eigenvectors of the form $$\begin{aligned} \tilde{v}^{(i)} = v^{(i)} - \gamma_i v^{(1)} $$ for $i=2, \cdots, n$ with $\gamma_i = \gamma_i(u) := \frac{u^Tv^{(i)}}{1-(\lambda_1-\lambda_i)/\mu}$ depending on the deflation vector $u$. Note that $v^{(1)} u^T \in \mathbb{R}^{n\times n}$ is in general a non-sparse matrix. But we can still employ matrix-vector multiplication by $B_1x = Bx - v^{(1)} (u^T x)$ using one sparse matrix-vector multiplication and one inner product evaluation. For the special choice of a constant vector $u$, the right eigenvectors $v^{(2)}, \cdots, v^{(n)}$ are also preserved by the deflation. This particular choice $u \parallel 1$ for the deflation vector is also called Hotelling’s deflation. Inspired by the deflation idea we can adapt the Jacobi iteration as follows. First assume $\lambda >0$ such that the diagonal $D_1 = diag(E)$ of $E:=\lambda I -B_1$ is non-singular. Then Jacobi iteration with $E$ has the form $$\begin{aligned} \begin{split} x^{(k+1)} &:= x^{(k)} - \omega D_1^{-1}(E x^{(k)}) \\ &~= x^{(k)} - \omega D_1^{-1} \left( \lambda x^{(k)} - B x^{(k)} + v^{(1)} ( u , x^{(k)} )_{\mathbb{R}^n} \right) \end{split} \label{JacobiIterationWielandt}\end{aligned}$$ In a similar fashion we can combine power iteration with Hotelling’s deflation: $$\begin{aligned} x^{(k+1)} &:= B_1 x^{(k)} = B x^{(k)} - v^{(1)}(u^Tx^{(k)}) \label{PowerIterationWielandt}\end{aligned}$$ For both methods we need to normalize after each iteration. For Hotelling’s deflation with deflation vector $u \parallel 1$ and $\mu = \lambda_1$, the iteration (\[PowerIterationWielandt\]) converges to the largest eigenvector of $B_1$ for any initial vector $x^{(0)} = \sum_{i=1}^n \alpha_i v^{(i)}$ with $\alpha_2 \neq 0$. The speed of convergence depends on $\frac{|\lambda_3|}{\lambda_2}$. By lemma \[thmWielandtDeflation\], Hotelling’s deflation preserves the right eigenvectors and eigenvalues $\lambda_2, \cdots, \lambda_n$. Then the convergence follows by a decomposition of the iterate into eigenvectors similar to the elaboration on classical power iteration in chapter \[sectionRelaxationMethods\]. In particular, as Hotelling’s deflation preserves the eigenvalues $\lambda_2, \cdots,\lambda_n$, the smoothing heuristic of deflated power iteration and deflated Jacobi iteration coincides with the heuristic for classical power and Jacobi iteration respectively. Hence we can use the same multilevel heuristics as for the invariant measure. For the invariant measure, the corresponding eigenvalue was known a priori, namely $\lambda_1 = 1$. For the second eigenvector, the fine level problem has the form $$\begin{aligned} B_1x = \lambda_2 x.\end{aligned}$$ However $\lambda_2$ is not known a priori and we can not simply solve a linear system but instead we have to exploit that $\lambda_2$ is the largest eigenvalue of $B_1$. That means we are seeking the eigenvector of $B_1$ corresponding to the eigenvalue with largest absolute value. Also, the deflated matrix $B_1$ is not stochastic and even a full (i.e. non-sparse) matrix in general. We circumvent these difficulties by using the stochastic matrix $B$ for the multilevel setting and then solve for the second largest eigenvector of $B$ with deflated power iteration (\[PowerIterationWielandt\]). The second eigenvector contains both positive and negative entries. This might cause a cancellation depending on the choice of aggregates: If positive and negative nodes in an aggregate sum to $0$, $\left(\text{diag}(Q^Tx^{(i)}) \right)^{-1}$ would not be well-defined. Hence it is essential that we only aggregate elements that have the same sign in the approximation $x^{(i)}$ to the second eigenvector. Then the matrix $$\begin{aligned} \text{diag}(x^{(i)}) Q \left(\text{diag}(Q^Tx^{(i)}) \right)^{-1}\end{aligned}$$ used to prolongate the coarse level approximation to the improved fine level iterate is non-negative. In particular the coarse level probability matrix $$\begin{aligned} B_c := Q^T B \text{ diag}(x^{(i)}) Q \left(\text{diag}(Q^Tx^{(i)}) \right)^{-1}. \label{coarseLevelMatrixSecondEV}\end{aligned}$$ remains a non-negative stochastic matrix even though the prolongation operator $$\begin{aligned} P = \text{diag}(x^{(i)}) Q\end{aligned}$$ itself has both positive and negative entries. From an intuitive viewpoint and for the exact solution, the coarse level problem should then preserve the second eigenvector as only nodes with the same sign in the second eigenvector are aggregated and weighted according to ratio in the second eigenvector. Hence the aggregation respects the slowest process of the Markov chain and does not distort it. Our intuition is indeed consistent with the following theorem. \[thmcoarseLevelSecondEigenvectorPreserved\] Let $x$ be an eigenvector of the stochastic fine-level matrix $B$ with the eigenvalue $\lambda_2$. Moreover let the aggregates $\varphi_p$ be chosen such that $x_k x_l > 0$ for all $k,l \in \varphi_p$ for all aggregates $\varphi_p$. Then, for the exact solution $x^{(i)} = x$, the coarse level matrix $B_c$ defined in (\[coarseLevelMatrixSecondEV\]) is a stochastic matrix with right eigenvector $Q^Tx$ corresponding to the eigenvalue $\lambda_2$. By the preceding remarks, the assumption on the aggregates ensures that $B_c$ is indeed a non-negative matrix. Moreover $$\begin{aligned} 1_c^T B_c &= 1_c^T Q^T B \text{ diag}(x) Q \left(\text{diag}(Q^Tx) \right)^{-1}\\ &= 1^T \text{ diag}(x) Q \left(\text{diag}(Q^Tx) \right)^{-1}\\ &= x^T Q \left(\text{diag}(Q^Tx) \right)^{-1}\\ &= (Qx)^T \left(\text{diag}(Q^Tx) \right)^{-1} = 1_c.\end{aligned}$$ Now simple calculations show that $Q^Tx$ is indeed an eigenvector. $$\begin{aligned} B_c (Q^Tx) &= Q^T B \text{ diag}(x) Q \left(\text{diag}(Q^Tx) \right)^{-1} (Q^Tx)\\ &= Q^T B \text{ diag}(x) Q 1_c\\ &= Q^T B \text{ diag}(x) 1 = Q^T B x = \lambda_2 (Q^Tx)\end{aligned}$$ The resulting multilevel DAM (Wieland deflation with algebraic aggregation for Markov chains) V-cycle is described in algorithm \[AlgoSecondEigenvectorSimpleAggregation\]. First we compute the invariant measure using a separate multilevel square and stretch algorithm from [@TY10] which is a modification of the algorithm described in section \[chapterAggregationMultilevelMethods\]. Then the invariant measure is used in the relaxation method for a Hotelling’s deflation for the modified power iteration (\[PowerIterationWielandt\]). This step is denoted by $RelaxWieland(B,x,v)$.\ The deflation happens only in the pre- and post-relaxation, while the coarse level probability matrix is still defined based on $B$. On the coarsest level, we apply deflated power iteration until the residual is close to machine precision. This step is denoted by $Solve(B,x,v)$. Theorem \[thmcoarseLevelSecondEigenvectorPreserved\] tells us that, for the exact solution, the second eigenvector $x$ on the fine level is preserved on the coarse level. However for a fixed point property of the multilevel algorithm we additionally require that $(Q^Tx)$ is also the *second* largest eigenvector on the coarse level. This would be fulfilled if coarsening does not generate a new eigenvalue $\lambda_2 < \tilde{\lambda} < 1$ of $B_c$ that dominates $\lambda_2$.\ From an intuitive viewpoint this is what we expect as aggregating some nodes should preserve the dynamics of the remaining nodes, in particular if the nodes are aggregated based on the strength of connection and the ratio in $x$. **Assumption on slow process.** If $x$ is the second eigenvector of $B$, assume that $(Rx)$ is the *second* eigenvector of $B_c$. In the numerical results we observe that the assumption holds for all test problems and aggregates chosen by the bottom-up approach. However, for arbitrary stochastic matrices and arbitrary aggregates, the assumption on slow processes does not hold in general. There are even counterexamples for small matrices ($n=5$) where the coarse level matrix has another eigenvalue that is larger than the second eigenvalue of the fine matrix. It appears that this assumption is a property of the aggregation algorithm used in the multilevel framework. Assume the assumption on slow processes holds. Then the second eigenvector $x$ of $B$ is a fixed point of DAM. This follows by the previous reasoning. For the multilevel V-cycle it remains to construct aggregates that satisfy the assumption $x_k x_l > 0$ for all $k,l\in \varphi_p$. To this end we use the same bottom-up approach as for the classical method in [@TY10] where the strength of connection is defined by $B\text{ diag}( (x^{(k)})^+)$. But when constructing the connectivity matrix $S$, we set those entries at $(k,l)$ to zero, where $x_k x_l \le 0$. **Some tricks to improve convergence.** For the relaxation method, we additionally use Chebyshev iteration to influence the smoothing effect of the deflated power iteration. For a polynomial $p \in \mathbb{R}[x]$, Chebyshev iteration computes a vector of the form $x^{(k+1)} = p(B_1)x^{(k)}$. Here the polynomial $p$ should be chosen such that the desired eigenmodes are amplified, while others are damped. Consider a splitting of the initial vector $x^{(k)} = \sum_{i=1}^n \alpha_i v^{(i)}$ into eigenvectors of $B_1$. Then it holds $$p(B_1)x^{(k)} = \sum_{i=1}^n \alpha_i p(B_1)v^{(i)} = \sum_{i=1}^n \alpha_i p(\lambda_i) v^{(i)}.$$ To this end, we seek a polynomial which is large at $\lambda_2$ and small on the set $\{ \lambda_3, \cdots, \lambda_n \}$. Without the knowledge of the eigenvalues such a polynomial is not possible to compute. In practice however we can use e.g. third order polynomials with two zeros in $[-1,0]$ that hence strongly damp the rough eigenmodes. Then coarse level correction reduces the smooth eigenmodes more effectively. To speed up the convergence in the first V-cycle iteration (only in the first!) it turns out to be more efficient if we use the matrix $B\text{ diag}(v^{(1)})$ to define the strength of connection and the prolongation operator $$\text{diag}(v^{(1)})Q \left(\text{diag}(Q^Tv^{(1)}) \right)^{-1}$$ to compute the fine level approximation from the coarse level iterate. The reason behind this is that in the first iteration the (known) first eigenvector provides a good classification of the smooth errors, while the initial approximation $x^{(1)}$ to the second eigenvector still has a large error.\ Once the iterate $x^{(k)}$ becomes sufficiently accurate, the strength of connection and the definition of the prolongation operator is based on the iterate $x^{(k)}$ as described in algorithm DAM.\ In figure \[figureWielandtMultilevelVisualization\] we see how the multilevel approach speeds up the convergence of plain vector iteration. In the first V-cycle there is a drastic improvement of the residual due to the coarse grid correction. As explained in the previous paragraph, this is due to the knowledge of the first eigenvector up to machine precision.\ In the subsequent iterations, the coarse level correction introduces new errors that are then effectively reduced by the first post-relaxations. This implies the coarse level correction improves the smooth error components, but introduces new rough error modes which then require damping by the post-relaxation steps. All in all the residual is significantly reduced due to the coarse level correction. For larger matrices however, the method becomes less efficient as the convergence rated increases for larger $n$. This was also a property of the classical AM-algorithm for the first eigenvector. Square and Stretch ================== In [@TY10; @DMM10] it was observed that the simple grid-transfer operators lead to slow convergence of AM for the invariant measure. This was explained by the observation that rough errors are generated by $P$ and not damped properly by $R$. In [@DMM10] De Sterck et al propose smoothing the prolongation and restriction operator to improve the convergence. A square and stretch approach where we can still use the same grid-transfer operators but reduce the rough eigenmodes on the coarse level has been proposed in [@TY10] by Treister and Yavneh. In this section we will apply the square and stretch approach in a similar fashion to improve the convergence of DAM.\ Regarding the first eigenvector, a matrix of the form $\hat{B} = \frac{1}{1-d}B - \frac{d}{1-d}I$ was considered where $d \in (0,1)$ is a stretching parameter. This changes the spectrum of the matrix in an advantageous way (see [@TY10]).\ Here however, the eigenvalue $\lambda_2$ of interest is smaller than $1$. Thus if there is an eigenvalue $\lambda_k$ close to $0$ we may observe $$\left| \frac{\lambda_k^2}{(1-d)} - \frac{d}{(1-d)} \right| > \left| \frac{\lambda_2^2}{(1-d)} - \frac{d}{(1-d)} \right|.$$ Hence after the square and stretch transformation, there would be a different eigenvector whose eigenvalue has largest absolute value. This would distort the method. There are two remedies: The first approach is to use pre- and post-relaxations such that the eigenvalues close to $0$ are eliminated up to machine precision. Then they do not distort the convergence.\ The second approach is to use a shifting parameter $p \ge 1-\lambda_2$ that shifts the spectrum before squaring. The spectrum of the resulting matrix $$\begin{aligned} \hat{B} := \frac{1}{1-d} \left(\frac{1}{(1+p)^2}(B + pI)^2\right) - \frac{d}{1-d}I\end{aligned}$$ is still contained in the interval $[-1,1]$ but now $\lambda_2$ remains the dominant second eigenvalue in terms of largest absolute value. Of course the shifting parameter $p$ is not known in advance. Numerically we may use the approximation of the iterate to $1-\lambda_2$ with some multiplicative error correction. Algorithm DS&SM (Wieland deflation with square and stretch for Markov chains) describes the resulting multilevel V-cycle. In figure \[figureWielandtMultilevelVisualizationSquareStretch\] we see that the square and stretch transformation drastically improves the convergence compared to the simple aggregation procedure depicted in figure \[figureWielandtMultilevelVisualization\]. On the other hand we observe that the coarse level correction now significantly introduces new errors. Hence we require more relaxation steps to damp the new errors. The steep descent of the residual in the first post-relaxation steps indicates that coarse level correction is quite effective in reducing the smooth error components and that the new errors generated by the coarse level are only rough modes. This algorithm has the same fixed point property as DAM. Numerical Results ================= The choice of test problems was motivated by the examples used in [@DMM10; @TY10]. Also the numerical parameters are chosen alike. $\gamma$ denotes the convergence rate of a V-cycle. $it$ denotes the number of steps to reduce the residual by factor of $10^{-10}$. $C_{op}$ denotes the operator complexity, i.e. the number of nonzero elements on all levels divided by the number of non-zero elements on the fine level. $\gamma_{\text{eff}} = \left((10^{-10})^{1/it} \right)^{1/C_{op}}$ denotes the effective convergence factor to measure the overall effectiveness of the method. $lev$ denotes the number of levels of a V-cycle. The parameter $\theta$ is used in the bottom-up aggregation scheme described in [@TY10]. Uniform Chain {#matrixUniformChain} ------------- The simplest test matrix we consider is the uniform chain matrix. Figure \[figUniformChain\] illustrates the graph of the 1D chain with uniform weights. The connectivity matrix is symmetric with non-zero entries just above and below the main diagonal. The resulting transition matrix has eigenvalues that are almost uniformly distributed in the interval $[0,1]$.\ Aggregates of size $s=2$ with aggregation parameter $\theta = 0.1$ are chosen. For DS&SM using $d=\min(\text{diag}(A))$ or $d=0.5$ gives the same results, as the diagonal elements are approximately $0.5$. DS&SM and DAM require a larger number of smoothing steps compared to AM in order to damp the rough error modes generated by the coarse level correction. We use 100 pre- and post-relaxations where each relaxation consists of 3 matrix-vector multiplications.\ In table \[tableUniformChain2\] we see that DS&SM performs extremely well with bounded operator complexity and low convergence factors. Just as for the invariant measure, the square and stretch transformation significantly increases the performance.\ In fact, despite the high number of relaxation steps, the good convergence factor of DS&SM is the result of coarse level improvement together with relaxations. If we consider only relaxation steps without the multilevel correction, the residual is merely reduced with a convergence factor of $0.999$ for 200 relaxation steps on a 16384x16384 uniform chain matrix. ![Left: Graph representation of the uniform chain. Right: Distribution of the eigenvalues for $n=1000$.[]{data-label="figUniformChain"}](eigenvalues-uniform-eps-converted-to.pdf) Uniform Chain with Weak Link ---------------------------- This test problem is a 1D chain with one weak link in the middle of the chain with weight $\epsilon = 0.001$ and uniform weights $\omega = 1$ on the other edges. Figure \[figUniformChainWeak\] shows a graph representation and the distribution of eigenvalues. We can use the same parameters as for the uniform chain. As for the standard uniform chain without a weak link, DS&SM performs well and has very good convergence properties. It by far outperforms DAM. ![Left: Graph representation of the uniform chain with one weak link. Right: Distribution of the eigenvalues for $n=1000$.[]{data-label="figUniformChainWeak"}](eigenvalues-uniformWeak-eps-converted-to.pdf) 2D Lattice ---------- The 2D-lattice is another simple structured problem. In figure \[fig2Dlattice\] we see the graph representation and distribution of eigenvalues. For the numerical tests we consider quadratic lattices. (Only for $n=32768$ we have a 128x256 lattice.)\ We use aggregates of size $s=4$ and chose the other parameters as for the uniform chain. Regarding the second eigenvector we observe that the method DS&SM has reasonable convergence rates that increase as $n$ grows. $C_\text{op}$ seems to be bounded. The square and stretch approach significantly accelerates the convergence compared to DAM. Random Delaunay Triangulation ----------------------------- The problems we considered so far had a simple geometric structure which implicitly prescribed the algebraically generated aggregates. The (undirected) graph for this example problem is a Delaunay triangulation for $n$ randomly chosen points in the $[0,1]^2$ unit square with uniform weights. Figure \[figRandomWalk\] shows an example of such a graph and the spectrum of the corresponding Markov chain.\ For $n=262144$ we used the aggregation parameter $\theta = 0.1$ instead of $0.25$. We use aggregates of size $4$ and besides that the standard parameters. For D&SM we use a higher number of 300 relaxation steps.\ Regarding the second eigenvector we observe that the convergence factors approach $1$ and the method becomes less effective for large $n$. Moreover DS&SM does not significantly outperform standard DAM. But still the method is superior to plain relaxation. The inferior results can be explained by the observation that for this problem the coarse level correction introduces new errors which need to be reduced by additional relaxation steps. The random grid makes the choice of good aggregates more difficult. If the aggregation algorithm is not optimal, this leads to slower convergence. Note here, that the convergence is not caused by the large number of relaxation steps, but by relaxation combined with coarse level correction. Plain relaxation alone does not significantly change the residual at all for large matrices. 1D Multi-Well Potential ----------------------- The double-well potential is a simplified academic example problem. For instance, it is used to illustrate the reaction pathway of a protein folding process. It is based on an energy potential with two wells (figure \[figureEnergyPotential\]) on which a particle is moving around driven by both diffusion and drift induced by the energy potential. We discretize by dividing the domain $[0,1]$ into equidistant intervals and use a Monte Carlo method to estimate the transition probabilities between the boxes. Note that for this test problem the number of non-zero entries in the transition matrix grows quadratically with $n$, whereas for all other test problems the number of non-zero entries is proportional to the matrix size. for $n=4096$ the matrix has about 2.3e+6 non-zero entries and for $n=16384$ the matrix has more than 3.7e+7 non-zero entries.\ We use aggregates of size $2$ and take the average of the diagonal elements as the stretching parameter.\ Note that there is a spectral gap after the second eigenvalue. Hence the second eigenvector is easier to compute. We use only 3x pre- and post-relaxation steps to emphasize the convergence effect of the coarse level improvement. Also square and stretch does not significantly improve the convergence.\ The method can also be successfully applied to a four-wells potential constructed similarly as the double-well potential. The four-well potential however has more than one eigenvalue close to $1$, hence the computation of the second eigenvalue is harder. For the multilevel V-cycle we use aggregates of size $3$ and $9$ pre- and post-relaxation steps. In table \[fourWellPotentialTable\] we can observe that the convergence factor is independent of $n$ and $C_{op}$ grows slightly with $n$. \[fourWellPotentialTable\] Complex Uniform Chain {#artificialChain} --------------------- This example is a modified uniform chain with additional transitions along the chain as depicted in figure \[figArtificial\]. The non-symmetry of the sparsity pattern of the transition matrix causes complex eigenvalues. We see that the eigenvalues with large imaginary part are contained in a ball of radius $\approx 0.5$ around zero. The eigenvalues close to $1$ remain real-valued. We choose aggregates of size $3$. The stretching parameter $d$ has to be smaller than $0.5$ due to the imaginary spectrum. Regarding the second eigenvector we observe that DS&SM outperforms DAM and has a bounded operator complexity. However the convergence factor increases as $n$ grows. \[tableArtificialChain2\] Conclusions =========== The new method DS&SM to compute the second eigenvector has superior convergence properties for matrices with few connections like the uniform chain, the uniform chain with a weak link and the complex uniform chain. Also for structured problems like the 2D-lattice the methods is applicable. For the multi-well potential it appears that the method is very scalable and has good convergence results. However the method is limited by its relatively poor convergence for the unstructured Delaunay triangulation with large $n$. A reason for this behavior might be that the aggregation method is not optimal for this type of grid. But the method is still far more effective than ordinary relaxation and improves the convergence. Also it is possible that a different aggregation method could significantly improve the convergence.\ Regarding the analysis, it would be interesting if the assumption on slow processes can be proven for certain classes of stochastic matrices and aggregation methods.\ So far the theory for smooth errors has been developed largely for symmetric positive-definite matrices. I think it would be good if one could extend the theory to arbitrary relaxation methods with a thorough framework. Finally one could extend the method to compute the first $k$ dominant eigenvectors with a multilevel approach. At least for problems with few connections like the uniform chain, this approach appears to be very promising. [^1]: This work was supported by Deutsche Forschungsgemeinschaft (DFG) through the Collaborative Research Center CRC 1114 “Scaling Cascades in Complex Systems”, Project (B03) “Multilevel coarse graining of multiscale problems”. I would like to thank my supervisor Prof. Kornhuber for his helpful advice and academic encouragement.

--- abstract: 'We consider the Goursat problem in the plane for partial differential operators whose principal part is the $p$th power of the standard Laplace operator. The data is posed on a union of $2p$ distinct lines through the origin. We show that the solvability of this Goursat problem depends on Diophantine properties of the geometry of lines on which the data is posed.' address: - 'P. Ebenfelt: Department of Mathematics, University of California, San Diego, La Jolla, CA 92093–0112, USA.' - 'H. Render: Departamento de Matemáticas y Computación, Universidad de La Rioja, Edificio Vives, Luis de Ulloa s/n., 26004 Logroño, España.' author: - Peter Ebenfelt and Hermann Render title: The Goursat problem for a generalized Helmholz operator in the plane --- [^1] [^2] Introduction ============ Let us consider in $\bR^2$ the mixed Cauchy problem $$\Label{cauchy0} \left\{ \begin{aligned} &\Delta^p u + \sum_{|\alpha|\leq k_0} a_\alpha \frac{\partial^{|\alpha|}u}{\partial x^\alpha}=f\\& P|(u-g), \end{aligned} \right.$$ where $p$ is a positive integer, $k_0$ is an integer with $0\leq k_0\leq 2p-1$, $\Delta$ denotes the standard Laplace operator in $\bR^2$ $$\Delta:=\frac{\partial^2}{\partial x^2}+\frac{\partial^2}{\partial y^2},$$ the coefficients $a_{\alpha}=a_{\alpha}(x,y)$ as well as the data functions $f=f(x,y)$ and $g=g(x,y)$ are real-analytic functions near $0$, and $P=P(x,y)$ is a homogeneous polynomial of degree $2p$. Here, the notation $P|(u-g)$ means that $P$ divides $u-g$ in the ring of germs of real-analytic functions at $0$. For instance, if $P(x,y)=L(x,y)^{2p}$ for some linear function $L(x,y)$ (which is equivalent to saying that the zero set of $P(x,y)$ consists of a single line with multiplicity $2p$), then with $k_0=2p-1$ is a standard Cauchy problem with data on the line $\{L(x,y)=0\}$ and the classical Cauchy-Kowalevsky Theorem guarantees that has a unique real-analytic solution $u$ near $0$ for every choice of data functions $f$ and $g$. In the recent paper [@EbRe06], the authors show that if $P$ is elliptic (i.e. the zero set of $P(x,y)$ consists of only the origin), then with $k_0=p$ has a unique solution $u$ for every choice of data functions $f$ and $g$. In this paper, we shall consider the case where the zero set of $P(x,y)$ is a union of $2p$ distinct lines (in which case may be called a Goursat problem). This case is much more subtle and leads to a small divisor problem. We shall give a sufficient condition (which is also necessary in the case $p=1$; see Section \[nec\]) on the divisor $P$ (see Theorem \[homodelp\] below) for the homogeneous Goursat problem $$\Label{goursatp1} \left\{ \begin{aligned} &\Delta^p u=f\\& P|(u-g) \end{aligned} \right.$$ to have a unique real-analytic solution $u$ for every real-analytic data $f$ and $g$. We shall also give a sufficient condition on $P$ (Theorem \[helmdelp\] below) for the perturbed Goursat problem $$\Label{goursatp2} \left\{ \begin{aligned} &\Delta^p u+cu=f\\& P|(u-g), \end{aligned} \right.$$ where $c=c(x,y)$ is a real-analytic function near $0$, to have a unique real-analytic solution $u$ for every data function $f$ and $g$. The conditions on $P$ in Theorems \[homodelp\] and \[helmdelp\] involve Diophantine properties of a determinant constructed from the geometry of the lines constituting the zero set of $P$. For instance, if $p=1$, so that $P$ has degree two and its zero set consists of two distinct lines, then the condition can be phrased in terms of the (acute) angle $\theta=2\pi \alpha$ between the two lines. The necessary and sufficient condition for the homogeneous Goursat problem $$\Label{goursat11} \left\{ \begin{aligned} &\Delta u=f\\& P|(u-g) \end{aligned} \right.$$ to be solvable (Corollary \[homodel1\]) is that $$\Label{dioleray} \liminf_{\mathbb Z\ni m\to\infty} \left(\inf_{n\in \mathbb Z}\left|\alpha-\frac{n}{m}\right|\right)>0,$$ a condition that is satisfied by e.g. all non-Liouville numbers. Our condition for the perturbed Goursat problem $$\Label{goursat12} \left\{ \begin{aligned} &\Delta u+cu=f\\& P|(u-g), \end{aligned} \right.$$ to be solvable (Corollary \[helmdel1\]) is more restrictive, namely there exists a constant $C>0$ such that $$\Label{diophantine-1} \left|\alpha-\frac{n}{m}\right|\geq\frac{C}{m^2} ,\quad \forall n,m\in \mathbb Z, m\neq 0 .$$ We note that every irrational number $\alpha$ that satisfies an integral quadratic equation (like $\sqrt{k/l}$ for any integers $k$ and $l$) satisfies (by Liouville’s Theorem on Diophantine approximation). We also point out that every irrational, algebraic number satisfies $$\Label{diophantine-2} \left |\alpha-\frac{n}{m}\right |\geq \frac{C_\mu}{m^\mu},\quad \forall n,m\in \mathbb Z, m\neq 0 ,$$ for some constant $C_\mu$ (that depends on $\mu$) and [*every*]{} $\mu>2$ by the Thue-Siegel-Roth Theorem [@Ro55]). However, there are algebraic numbers that do not satisfy . We also mention that it follows from our proof that has a unique formal power series solution for all $f$ and $g$ if and only if $\alpha$ is irrational. Thus, as a consequence of our results, we conclude that the family of Goursat problems , parametrized by the angle $2\pi\alpha$ between the two lines in the zero set of $P$, displays “chaotic” behavior in that the set of parameters for which is solvable is dense as is the set of parameters for which there is not even a formal solution. The homogeneous Goursat problem (i.e. with $p=1$) can be transformed, by a simple linear change of coordinates, into a Goursat problem considered by J. Leray in [@L74]. (It was also briefly considered in its present form by H. Shapiro in [@Shap89].) Leray’s main result is equivalent to our Corollary \[homodel1\]. The relationship between the two Goursat problems and Leray’s work is briefly explained in Section \[lerayequiv\] below. Leray’s work was extended to complex parameters and to higher dimensions by Yoshino in [@Y81a] and [@Y81]. Other related work on mixed Cauchy and Goursat problems include that of Gårding [@G65] (see also Theorem 9.4.2 in Hörmander [@H90]), Shapiro [@Shap89], the first author and Shapiro [@EbSh96], and the authors [@EbRe06]. Our approach to studying the Goursat problem is inspired by ideas from [@Shap89] and [@EbSh96]. The proofs are based on a new estimate for an associated Fischer operator in the real Fischer norm (Theorem \[estimate\]). The real Fischer norm was introduced in [@Rend05] and was also used in [@EbRe06]. This paper is organized as follows. We present our main results in Section \[mainresults\]. In Section \[lerayequiv\], we discuss the relation between our results in the case $p=1$ and $c\equiv 0$ and those of Leray in [@L74]. An associated Fischer operator, which is used in the proofs of the main results, is introduced in Section \[s:est\] and a crucial estimate is proved for that operator (Theorem \[estimate\]). The proof of Theorem \[homodelp\] is also given in that section. The proof of Theorem \[helmdelp\] is given in the subsequent section. In Section \[ex\], we consider the case $p=2$ and present an explicit family of examples to which Theorem \[helmdelp\] can be applied (see Theorem \[helmdel2\]). Finally, in Section \[nec\], we show that our condition in Corollary \[homodel1\] is also necessary in this case ($p=1$). Main results ============ We shall now formulate our results more precisely. We must first introduce some notation. Let $B_{R}:=\left\{ (x,y)\in \mathbb{R}^{2}:x^2 +y^2 <R^2 \right\} $ be the open disk of radius $R$ in $\mathbb{R}^{2}$ (where $0<R\leq \infty ).$ We denote by $A\left( B_{R}\right) $ the algebra of all infinitely differentiable functions $f:B_{R}\rightarrow \mathbb{C}$ such that for any compact subset $ K\subset B_{R}$ the homogeneous Taylor series $\sum_{m=0}^{\infty }f_{m}\left( x,y\right) $ converges absolutely and uniformly to $f$ on $K$; here, $f_{m}$ is the homogeneous polynomial of degree $m$ defined by the Taylor series of $f$ $$f_{m}\left( x,y\right) =\sum_{k+l =m}\frac{1}{k! l !}\frac{% \partial ^{m }f}{\partial x^{k}\partial y^l}\left( 0\right) x^{k }y^l.$$ Note that the functions in $A(B_R)$ are real-analytic. For a real number $a$, we shall define the unimodular complex number $$\Label{A} A=A(a):=\frac{a+i}{a-i}.$$ As $a$ goes from $-\infty$ to $\infty$, $A$ ranges over the unit circle (from 1 to 1 in the negative direction) and, hence, there is a unique $\beta\in (0,1)$ such that $A=e^{2\pi i \beta}$. Note that $\beta$ is rational precisely when $A$ is a root of unity. For future reference, we observe, using the fact that $2\arctan a=i\log(1-ia)/(1+ia)$, that for $a\in [0,\infty)$ the acute angle between the lines $y=0$ and $x-ay=0$ is $\pi \beta$. Now, let us fix a positive integer $p$, distinct real numbers $a_1, a_2, \ldots, a_{2p-1}$, and write $a$ for the vector $a=(a_1,\ldots,a_{2p-1})$. We shall denote by $P_a(x,y)$ the divisor $$\Label{Pa} P_{a}(x,y):=y\, \Pi_{j=1}^{2p-1}(x-a_jy).$$ If the divisor $P$ in is a homogeneous polynomial of degree $2p$ with $2p$ distinct lines as its zero set, then there is no loss of generality in assuming that $P$ is of the form , since the Laplace operator is rotationally invariant. We associate to the vector $a$ a sequence of $2p\times 2p$ matrices $\{M_{m,p,a}\}_{m=0}^\infty$, where $$\Label{Ma} M_{m,p,a}:=\begin{pmatrix} 1& 1& 1 & \ldots & 1 & 1&\ldots& 1\\ 1& A_1& A_1^2 & \ldots & A_1^{p-1} & A_1^{m+p+1}&\ldots& A_1^{m+2p}\\ 1& A_2& A_2^2 & \ldots & A_2^{p-1} & A_2^{m+p+1}&\ldots& A_2^{m+2p} \\ \vdots & \vdots & \vdots & \ddots &\vdots & \vdots & \ddots &\vdots \\ 1 & A_{2p-1} & A_{2p-1}^2 & \ldots & A_{2p-1}^{p-1} & A_{2p-1}^{m+p+1}&\ldots& A_{2p-1}^{m+2p} \end{pmatrix}.$$ Here, $A_j:=A(a_j)$ where $A(a_j)$ is given by . We shall consider the Goursat problem $$\Label{goursatp} \left\{ \begin{aligned} & \Delta^pu+cu=f\\& P_a|(u-g), \end{aligned} \right.$$ where the coefficient $c=c(x,y)$ as well as the data functions $f=f(x,y)$, $g=g(x,y)$ belong to $A(B_R)$. Our first result concerns the homogenenous problem, i.e. $c\equiv 0$. Let $p$ be a positive integer and $a_1,\ldots,a_{2p-1}$ real, distinct, non-zero numbers. Let $A_j:=A(a_j)$, for $j=1,\ldots, 2p-1$, be the unimodular complex numbers given by , $P_a(x,y)$ the homogeneous polynomial given by , and $\{M_{m,p,a}\}_{m=0}^\infty$ given by . If $\det M_{m,p,a}\neq 0$ for all integers $m\geq 0$, and $$\Label{leraycond1} \tau:=\liminf_{m\to\infty}\left(\det M_{m,p,a}\right )^{1/m}>0,$$ then the homogeneous Goursat problem $$\Label{goursatp0} \left\{ \begin{aligned} & \Delta^pu=f\\& P_a|(u-g) \end{aligned} \right.$$ has a unique solution $u\in A(B_{\tau R})$ for every $f,g\in A(B_R)$. We mention that e.g. all numbers $a_1,\ldots,a_{2p-1}$ such that $A_1,\ldots, A_{2p-1}$ are algebraic and $\det M_{m,p,a}\neq 0$ for all $m$ satisfy (see [@W00], Lemma 2.1). It will follow from our proof of Theorem \[helmdelp\] below that the Goursat problem , and hence in particular , has a unique formal solution $u$ if and only if $\det M_{m,p,a}\neq 0$ for all integers $m\geq 0$. The Diophantine condition is sufficient (and necessary for $p=1$; see Section \[nec\] below) for the formal solution to to converge. For the formal solution to the general Goursat problem to converge, we need a stronger condition. We have the following result. Let $p$ be a positive integer and $a_1,\ldots,a_{2p-1}$ real, distinct, non-zero numbers. Let $A_j:=A(a_j)$, for $j=1,\ldots, 2p-1$, be the unimodular complex numbers given by , $P_a(x,y)$ the homogeneous polynomial given by , and $\{M_{m,p,a}\}_{m=0}^\infty$ given by . If there exists a constant $C>0$ such that $$\Label{leraycond2} \det M_{m,p,a}\geq \frac{C}{m^p},$$ for all natural numbers $m \geq 1 $ then there exists $0<r\leq R$ such that the Goursat problem has a unique solution $u\in A(B_{r})$ for every $f,g\in A(B_R)$. In Section \[ex\] below, we give some explicit examples of $a_1,a_2,a_3$ such that holds for the corresponding unimodular numbers $A_1,A_2,A_3$. In the case $p=1$, the zero set of $P_a$ is the union of the two distinct lines given by $y=0$ and $x=ay$. By the rotational symmetry of $\Delta$, we may also assume that $a\geq 0$. If we denote the acute angle between the two lines by $2\pi \alpha$ and by $\beta\in (0,1/2]$ the number such that $A:=A(a)=e^{2\pi i\beta}$, then as mentioned in the beginning of this section we have $\beta=2\alpha$. As noted in Remark \[rmkmatrix\] above, we have $\det M_{m,p,a}=A^{m+2}-1$. The condition $\det M_{m,p,a}=A^{m+2}-1\neq 0$ is clearly equivalent to $\alpha$ being irrational. Since $$|A^{m+2}-1|\approx\inf_{n\in \mathbb Z} |2\pi (m+2) \beta-2\pi n|=2\pi (m+2) \inf_{n\in \mathbb Z} \left|\beta- \frac{n}{m+2}\right|,$$ where by $E_k\approx F_k$ we mean $CF_k\leq E_k\leq DF_k$ for nonzero constants $C,D$, it is not difficult to see that Theorems \[homodelp\] and \[helmdelp\], specialized to the case $p=1$, can be formulated as follows. Let $\Gamma_1$, $\Gamma_2$ be two distinct lines through the origin in $\bR^2$, and denote by $\theta=2\pi\alpha$ the acute angle between them. Suppose that $\alpha$ is irrational and satisfies the condition $$\Label{leraycond3} \tau:=\liminf_{m\to\infty}\left(\inf_{n\in\mathbb Z}\left |\alpha-\frac{n}{m}\right |\right)^{1/m}>0.$$ Then, the homogeneous Goursat problem $$\Label{goursat10} \left\{ \begin{aligned} & \Delta u=f\\& u=g \quad \text{{\rm on $\Gamma_1\cup\Gamma_2$}} \end{aligned} \right.$$ has a unique solution $u\in A(B_{\tau R})$ for every $f,g\in A(B_R)$. The condition is also necessary for the conclusion of Corollary \[homodel1\] to hold. This fact is proved in Section \[nec\] below. As mentioned in the introduction, Corollary \[homodel1\] is equivalent to the result of Leray in [@L74]. A more detailed explanation of this equivalence is given in Section \[lerayequiv\] below. We conclude this section by reformulating Theorem \[helmdelp\] in the case $p=1$. Let $\Gamma_1$, $\Gamma_2$ be two distinct lines through the origin in $\bR^2$, and denote by $\theta=2\pi\alpha$ the acute angle between them. Suppose that $\alpha$ satisfies the Diophantine condition $$\Label{diophantine} \left |\alpha-\frac{n}{m}\right |\geq \frac{C}{m^2},\quad \forall n,m\in \mathbb Z, m\neq 0$$ for some constant $C>0$. Then, for any $c\in A(B_R)$, there exists $0<r\leq R$ such that the Goursat problem $$\Label{goursatp10} \left\{ \begin{aligned} & \Delta u+cu=f\\& u=g \quad \text{{\rm on $\Gamma_1\cup\Gamma_2$}}, \end{aligned} \right.$$ has a unique solution $u\in A(B_r)$ for every $f,g\in A(B_R)$. Leray’s Goursat problem ======================= Consider the homogeneous Goursat problem $$\Label{goursat0} \left\{ \begin{aligned} &\lambda \frac{\partial^2u}{\partial x\partial y}+\Delta u=f\\& xy|(u-g), \end{aligned} \right.$$ where $\lambda$ is a real constant. It follows from the general theory of Goursat (or mixed Cauchy) problems that has a unique real-analytic solution near $0$, for all $f$ and $g$, if $|\lambda|>2$ (see Gårding [@G65]; see also Theorem 9.4.2 in Hörmander [@H90]). The case where $\lambda\in[-2,2]$ is much more subtle, and was analyzed by Leray in [@L74] (see also the work of Yoshino [@Y81a], [@Y81] for extensions to complex parameters and higher dimensions). For $\lambda\in [-2,2]$, let $\beta\in [-1/4,1/4]$ denote the angle such that $\lambda=2\sin(2\pi\beta)$. Leray showed that the unique solvability of depends on Diophantine properties of $\beta$. For instance, there is a unique formal power series solution $u$ for every $f$ and $g$ if and only if $\beta$ is irrational. Leray also gave a necessary and sufficient Diophantine condition on irrational $\beta$ quaranteeing that this formal solution $u$ converges for all convergent $f$ and $g$, $$\Label{dioleray2} \liminf_{\mathbb Z\ni m\to\infty} \left(\inf_{n\in \mathbb Z}\left|\beta-\frac{n}{m}\right|^{1/m}\right)>0.$$ Let us show that this result, for $\lambda\in(-2,2)$, is equivalent to our Corollary \[homodel1\] above. Consider the linear change of variables $$\Label{trans} x \to - \sqrt{1-\frac{\lambda^2}{4}}x+\frac{\lambda}{2}y.$$ As the reader can easily verify, this change of variables leads to the following transformation for the principal symbol of the operator $$\lambda \frac{\partial^2}{\partial x\partial y}+\Delta \to \Delta.$$ Hence, the Goursat problem is transformed into the following $$\Label{goursat01} \left\{ \begin{aligned} &\Delta u=f\\& y(x-ay)|(u-g), \end{aligned} \right.$$ where $$\Label{b} a:=\frac{\lambda/2}{\sqrt{1-(\lambda/2)^2}}.$$ If we let $\theta=2\pi\alpha$ denote the acute angle between the two lines $L_1:=\{y=0\}$ and $L_2:=\{x=by\}$ and $\beta$ the angle such that $\lambda:=2\sin(2\pi\beta)$, then we have $$\alpha=\frac{1-2\beta}{4}.$$ Clearly, we have $$\liminf_{\mathbb Z\ni m\to\infty} \left(\inf_{n\in \mathbb Z}\left|\beta-\frac{n}{m}\right|\right)^{1/m}=\liminf_{\mathbb Z\ni m\to\infty} \left(\inf_{n\in \mathbb Z}\left|\alpha-\frac{n}{m}\right|\right)^{1/m}.$$ This shows, as mentioned in the introduction, that Leray’s result, with $\lambda\in (-2,2)$, is equivalent to our Corollary \[homodel1\], with $0<a<\infty$. An estimate for an associated Fischer operator and the proof of Theorem $\ref{homodelp}$ ======================================================================================== Let $\bC[x,y]$ denote the space of polynomials in $x,y$ with complex coefficients. For each integer $m\geq 0$, we shall let $\mathcal P_m$ denote the subspace of homogeneous polynomials of degree $m$. We endow $\bC[x,y]$ with the real Fischer inner product $$\langle f,g\rangle :=\int_{\bR^2} f(x,y)\overline{g(x,y)}e^{-(x^2+y^2)}dxdy,$$ and denote by $\|\cdot\|$ the corresponding norm (see [@Rend05]). We shall fix a positive integer $p$ and distinct real numbers $a_1,\ldots, a_{2p-1}$ and consider the Fischer operator $F_a(q):=\Delta^p(P_aq)$, where $P_a$ is given by . Observe that $F_a$ is a linear operator sending $\mathcal P_m$ into $\mathcal P_m$. Our main result in this section is the following, in which the notation introduced above is used. Let $p$ be a positive integer and $a_1,\ldots,a_{2p-1}$ real, distinct, non-zero numbers. Let $A_j:=A(a_j)$, for $j=1,\ldots, 2p-1$, be the unimodular complex numbers given by and $P_a(x,y)$ the homogeneous polynomial given by . Then the Fischer operator $F_a\colon \mathcal P_m\to\mathcal P_m$, for $m\geq 0$, is a bijection if and only if $\det M_{m,p,a}\neq 0$, where $M_{m,p,a}$ is given by . Moreover, if $\det M_{m,p,a}\neq 0$, then we have the estimate $$\|P_aq\|\leq \frac{C} {|\det M_{m,p,a}|}\|\Delta^p(P_aq)\|,\quad \forall q\in \mathcal P_m,$$ for some $C\geq 0$ (independent of $m$). For the proof of Theorem \[estimate\], we shall need the following lemma. To state the lemma, we observe the well known fact that any homogeneous polynomial $f(x,y)$ of degree $m$ can be expressed in the following way $$\Label{expaninz} f(x,y)=\sum_{k+l=m}f_{kl}z^k\bar z^l,$$ where $z=x+iy$ and $\bar z=x-iy$. Let $f(x,y)$ be a homogeneous polynomial of degree $m$ given by . Then, we have $$\Label{normid} \|f\|^2=\pi m!\sum_{k+l=m}|f_{kl}|^2.$$ As in [@Rend05] (see also [@EbRe06]), we observe that for any homogeneous polynomial $f(x,y)$ of degree $m$, we have $$\Label{norm1} \|f\|^2=I_{2m+1}\int_{\mathbb T}|f(\eta)|^2ds_{\eta}$$ where $\mathbb T$ denotes the unit circle in $\bR^2$, $ds$ arclength, and $I_{k}$ the integral $$I_{k}:=\int_{0}^{\infty }e^{-r^{2}}r^{k}dr.$$ A simple substitution argument gives $$\Label{int1} I_{2m+1}=\int_{0}^{\infty }e^{-r^{2}}r^{2m+1}dr=\frac{1}{2}\int_{0}^{\infty }e^{-x}x^{m}dx=\frac{1}{2}m!.$$ Substituting in , using the parametrization $z=e^{i\theta}$ for $\mathbb T$ and the identity , yields $$\Label{norm2} \|f\|^2=\frac{1}{2}m!\sum_{k+l=m}\sum_{i+j=m}f_{kl}\overline{f_{ij}} \int_0^{2\pi}e^{i(k+j-l-i)\theta}d\theta,$$ from which readily follows. We fix $f\in \mathcal P_m$ and consider the equation $$\Label{Faisf} F_a(q):=\Delta^p(P_aq)=f,$$ for $q\in \mathcal P_m$. Note that $q\in \mathcal P_m$ solves if and only if $u=P_aq$ solves the Goursat problem $$\Label{goursat2} \left\{ \begin{aligned} & \Delta^p u=f\\ & u(x,0)=u(a_1y,y)\ldots u(a_{2p-1}y,y)=0. \end{aligned} \right.$$ We shall look for $u$ of the form $u=v+w$, where $w(x,y)=(x^2+y^2)^p s(x,y)$ for some $s\in \mathcal H_{m}$ such that $$\Label{classic} \Delta^p w(x,y)=\Delta^p((x^2+y^2)^ps(x,y))=f(x,y)$$ and $v\in \mathcal H_{m+2p}$ satifies $$\Label{goursat3} \left\{ \begin{aligned} & \Delta^p v=0\\ & v(x,0)=-w(x,0)\\ &v(a_jy,y)=-w(a_jy,y), \quad j=1,\ldots 2p-1. \end{aligned} \right.$$ It is well known that has a unique solution $w(x,y)=(x^2+y^2)^ps(x,y)$ (see e.g. [@Shap89] and references therein). Moreover, in view of the results in [@EbRe06], we have $$\|w\|\leq C_1\|f\|$$ for some constant $C_1>0$. Thus, to complete the proof of the theorem it suffices to show that has a solution $v\in \mathcal P_{m+2p}$ for every $f\in \mathcal P_m$ if and only if $\det M_{m,p,a}\neq 0$, and that, in this case, $$\Label{goal1} \|v\|\leq \frac{C}{|\det M_{m,p,a}|}\|f\|$$ for some constant $C>0$. To this end, we shall actually need the exact form of the solution to . Using $z=x+iy$ and $\bar z=x-iy$, we may write $$\Label{id1} w(x,y)=W(z,\bar z)=z^p\bar z^p\sum_{k+l=m} s_{kl}z^k\bar z^l=\sum_{k+l=m}s_{kl}z^{k+p}\bar z^{l+p}.$$ We observe that $\Delta=4\partial^2/\partial z\partial \bar z$. Thus, if we write $f(x,y)=\sum_{k+l=m} f_{kl}z^k\bar z^l$, then is equivalent to $$\Label{id2} s_{kl}=\frac{f_{kl}}{4^p(k+1)\ldots (k+p)(l+1)\ldots (l+p)},\quad \forall\ k+l=m.$$ Now, we note that every function $v(x,y)$ that satisfies $\Delta^pv=0$ is of the form $$\Label{vform} v(x,y)=\sum_{t=0}^{p-1}\left({\bar z}^t\phi_t(z)+z^t\psi_t(\bar z)\right),$$ where $\phi_t(z)$ and $\psi_t(\bar z)$ are holomorphic functions of $z$ and $\bar z$, respectively. The function $v$ is a homogeneous polynomial of degree $m+2p$ if and only if $\phi_t(z)=b_{p-1-t}z^{m+2p-t}$ and $\psi_t(\bar z)=c_t \bar z^{m+2p-t}$, for constants $b_{p-1-t}$ and $c_t$ and $t=0,\ldots, p-1$. Using this notation, equation is equivalent to finding monomials $$\Label{id3}\phi_t(z)=b_{p-1-t}z^{m+2p-t},\quad \psi_t(\bar z)=c_t \bar z^{m+2p-t},$$ for $t=0,1,\ldots,p-1$, such that $$\Label{get1} \sum_{t=0}^{p-1}\left({x}^t\phi_t(x)+x^t\psi_t(x)\right)=-W(x,x)$$ and $$\begin{gathered} \Label{get2} \sum_{t=0}^{p-1}\left({((a_j-i)y)}^t\phi_t((a_j+i)y)+ ((a_j+i)y)^t\psi_t((a_j-i)y)\right)=\\-W((a_j+i)y,(a_j-i)y),\quad j=1,\ldots, 2p-1.\end{gathered}$$ In , we use the fact that $\phi_t $ is homogeneous of degree $m+2p-t $ and divide the equation by $ (a_j -i )^{m+2p } $. With $A_j:=A(a_j)$ and $A(a)$ given by , the equation becomes $$\Label{get3} \sum_{t=0}^{p-1}\left(A_j^{m+2p-t}y^t \phi_t(y )+ A_j^ty^t \psi_t(y)\right)=-W(A_j y,y),\quad j=1,\ldots, 2p-1.$$ Substituting and in and , we obtain the following system of linear equations for the coefficients $b_0,\ldots, b_{p-1}, c_0,\ldots, c_{p-1}$ $$\Label{system1} \begin{aligned} \sum_{t=0}^{p-1}\left(b_{p-1-t}+c_t\right) &=-\sum_{k+l=m}\frac{f_{kl}} {4^p(k+1)\ldots (k+p)(l+1)\ldots (l+p)}\\ \sum_{t=0}^{p-1}\left(A^{m+2p-t}_1 b_{p-1-t}+ A_1^t c_j\right) &=-\sum_{k+l=m}\frac{f_{kl}A_1^{k} } {4^p(k+1)\ldots (k+p)(l+1)\ldots (l+p)}\\ &\vdots\\ \sum_{t=0}^{p-1}\left(A^{m+2p-t}_{2p-1} b_{p-1-t}+ A_{2p-1}^t c_j\right) &=-\sum_{k+l=m}\frac{f_{kl}A_{2p-1}^{k}} {4^p(k+1)\ldots (k+p)(l+1)\ldots (l+p)} \end{aligned}$$ If we write $d$ for the column vector of coefficients $d=(c_0,\ldots,c_{p-1},b_0,\ldots, b_{p-1})^t$ and $e$ for the column vector whose $(j+1)$th component, $j=0,\ldots,2p-1$, is given by $$-\sum_{k+l=m}\frac{f_{kl}A_j^{k}} {4^p(k+1)\ldots (k+p)(l+1)\ldots (l+p)},$$ where we let $A_0:=1$, then can be written $$\Label{matrixeq} M_{m,p,a}d=e,$$ where $M_{m,p,a}$ is given by . We conclude, as claimed above, that has a unique solution $v\in\mathcal P_{m+2p}$ for every $f\in \mathcal P_m$ if and only if $\det M_{m,p,a}\neq 0$. Let us now suppose that $\det M_{m,p,a}\neq 0$ and write $d_i$ for the $i$th component of $d$, $i=1,\ldots, 2p$. Using Cramer’s rule and the fact that $|A_j|=1$, we conclude from that $$\Label{coeffest} |d_i|\leq C_1|\det M_{m,p,a}|^{-1}\sum_{k+l=m}\frac{|f_{kl}|} {(k+1)\ldots (k+p)(l+1)\ldots (l+p)}.$$ By the Cauchy-Schwarz inequality, we conclude that $$\Label{coeffest2} |d_i|\leq C_1|\det M_{m,p,a}|^{-1}\left(\sum_{k+l=m}|f_{kl}|^2\right)^{1/2} S_m,$$ where $S_m$ denotes the sum $$\Label{Sm} S_m:=\left(\sum_{k+l=m}\frac{1}{(k+1)^2\ldots (k+p)^2(l+1)^2\ldots (l+p)^2}\right)^{1/2}.$$ By setting $l=m-k$, we obtain $$\Label{Smest} \begin{aligned} S_m^2= & \sum_{k=0}^m\left(\prod_{j=1}^p(k+j)^2(m-k+j)^2\right)^{-1}\\ \leq &\, 2\sum_{k=0}^{[m/2]+1}\left(\prod_{j=1}^p(k+j)^2(m-k+j)^2\right)^{-1}\\ = &\, 2m^{-2p}\sum_{k=0}^{[m/2]+1}\left(\prod_{j=1}^p(k+j)^2\left((1+(j-k)/m\right)^2\right)^{-1} \end{aligned}$$ Now, note that, for $j=1,\dots,p$ and $k=0,\ldots,[m/2]+1$, we have $(j-k)/m\geq -3/4$ when $m\geq 2$ and, hence, $(1+(j-k)/m)^{-2}\leq 16$. Consequently, we have $$\Label{Smest2} S_m^2\leq \frac{32}{m^{2p}} \sum_{k=0}^{[m/2]+1}\left(\prod_{j=1}^p(k+j)^2\right)^{-1}\leq \frac{32}{m^{2p}} \sum_{k=0}^{\infty}\frac{1}{(k+1)^{2p}}\leq \frac{C_2}{m^{2p}},$$ for some $C_2>0$ independent of $m$. Thus, by Lemma \[Fischer\], we obtain from and the following estimates for the functions $\tilde \phi_t(z,\bar z):=\bar z^t\phi_t(z)$, where $\phi_t$ is given by , $$\Label{estphipsi} \begin{aligned} \|\tilde \phi_t\|=&\, \sqrt{(m+2p)!}\,|b_{p-1-t}|\\ \leq&\, C_1C_2|\det M_{m,p,a}|^{-1}\sqrt{(m+1)\ldots(m+2p)}\|f\| m^{-p}\\ \leq&\, C_3 |\det M_{m,p,a}|^{-1}\|f\|. \end{aligned}$$ We obtain a similar estimate for $\tilde \psi_t(z,\bar z):=z^t\psi_t(\bar z)$. These estimates yield since $v$ is given by . This completes the proof of Theorem \[estimate\]. The arguments in the proof above also yield a proof of Theorem \[homodelp\]. We conclude this section by giving this proof. It is well known that to prove Theorem \[homodelp\] it suffices to show that the equation $$\Label{PDE} \Delta^p(Pq)=f$$ has a unique solution $q\in A(B_{\tau R})$ for every $f\in A(B_R)$ (see e.g. [@EbRe06]). As in the proof of Theorem \[estimate\], we shall look for the solution $u:=P_aq$ in the form $u=v+w$, where $w(x,y)=(x^2+y^2)^ps(x,y)$ satisfies and $v$ solves . It is well known that $w\in A(B_R)$ (see [@Shap89]; see also [@EbRe06]). Thus, to complete the proof, it suffices to show that $v\in A(B_{\tau R})$. We expand $v$ as a series $v=\sum_mv_m$, where the $v_m$ are the homogeneous Taylor polynomials of degree $m$ of $v$. Similarly, we expand $w=\sum_m w_m$ and $f=\sum_m f_m$. By homogeneity, we observe that the homogeneous polynomials $v_m$, $w_m$, $f_m$ satisfy (with $v=v_m$, $w=w_m$, and $f=f_m$). The fact that $v\in A(B_{\tau R})$ now follows easily from the definition of $\tau$, the form of $v$, and the estimate . The details are left to the reader. Proof of Theorem $\ref{helmdelp}$ ================================= We fix $a=(a_1,\ldots, a_{2p-1})$ as in the theorem. For brevity, we denote $P_a$ simply by $P$. To prove Theorem \[helmdelp\], it suffices to show that there is $0<r\leq R$ such that the equation $$\Label{PDE1} (\Delta^p+c)(Pq)=f$$ has a unique solution $q\in A(B_r)$ for every $f\in A(B_R)$. We shall look for the solution $u=Pq$ as a series $u=\sum_{m}u_m=\sum_{m}Pq_{m-2p}$, where the $u_m$ are the homogeneous Taylor polynomials of degree $m$ of $u$. To this end, we expand, similarly, both $f$ and $c$ as Taylor series $f=\sum_m f_m$ and $c=\sum_m c_m$. The equation then implies $$\Label{basic0} \Delta^p(Pq_j)=f_j,\quad j=0,1,\ldots, 2p-1,$$ and, for each $m\geq 2p$, $$\Label{basic1} \Delta^p(Pq_m)=f_m-\sum_{k=0}^{m-2p}c_{m-k-2p}Pq_{k}.$$ Since the Fischer operator $F=F_a$, given by $F(q)=\Delta^p(Pq)$, is bijective $\colon \mathcal P_m\to \mathcal P_m$ for every $m$ (by Theorem \[estimate\]), we can solve, uniquely, and inductively for $q_m$. This gives us a unique formal power series solution $u=\sum_mu_m$ with $u_m=Pq_{m-2p}$. It remains to prove that there is $r>0$ such that this series converges to a function in $A(B_{r})$. For this, we observe that Theorem \[estimate\] and the assumption implies the following estimate $$\Label{basic2} \|u_{m+2p}\|\leq Cm^{p}\|\Delta(Pq_m)\|\leq Cm^{p}\left( \|f_m\|+\sum _{k=0}^{m-2p}\|c_{m-k-2p}u_{k+2p}\|\right)$$ To prove that $u\in A(B_{r})$, we must show (see Proposition 16 in [@EbRe06]) that for every $0<\rho<r$ there is a constant $B>0$ such that $$\Label{ind} \|u_k\|\leq B\rho^{-k}\sqrt{k!}$$ for every $k\geq 0$. Let us pick $\rho<\sigma<R$. In view of Proposition 16 in [@EbRe06], we may assume that there are constants $D$ and $E$ such that $$\Label{assump} \max_{\theta \in \mathbb{T} }\left| c_{k}\left( \theta \right) \right| \leq D\sigma^{-k},\quad \left\| f_{k}\right\| \leq E \rho^{-k} \sqrt{k!},$$ for all $k\geq 0$. We shall prove by induction. Thus, assume that holds for all $k\leq m+2p-1$. We shall prove that holds also for $k=m+2p$, provided that $m$ is large enough. By using , the induction hypothesis, and Proposition 8 in [@EbRe06] (see also Proposition 7 in that paper), we conclude from the following estimate, for some constant $F>0$, $$\begin{aligned} \|u_{m+2p}\| \leq &Cm^{p}\left(E\rho^{-m}\sqrt{m!}+ \sum _{k=0}^{m-2p}F\sigma^{-(m-k-2p)}[(k+2p+1)\ldots (m-1)m]^{1/2}\|u_{k+2p}\|\right )\\ \leq &Cm^{\mu-1}\left(E\rho^{-m}\sqrt{m!}+ \sum _{k=0}^{m-2p}BF\sigma^{-(m-k-2p)}\rho^{-(k+2p)}\sqrt{m!}\right )\\=& B\rho^{-(m+2p)}\sqrt{(m+2p)!}\, T_m, \end{aligned}$$ where $$\begin{aligned} T_m: &=Cm^{p}\frac{\rho^{2p}}{\sqrt{(m+1)(m+2)}}\left (E/B+F\sum_{k=0}^{m-2p} \left(\frac{\rho}{\sigma}\right)^{m-k-2p}\right)\\ &\leq Cm^{p}\frac{\rho^{2p}}{\sqrt{(m+1)\ldots (m+2p)}}\left (E/B+F\frac{1}{1-\rho/\sigma}\right). \end{aligned}$$ Since $\rho<r$, we can make $T_m\leq 1$ for all $m$ by requiring $0<r\leq R$ small enough (and keeping $\sigma<R$ fixed). This proves Theorem \[helmdelp\]. Examples of solvable Goursat problems for $\Delta^2+c$ ====================================================== In this section, we shall consider the following one-parameter family of Goursat problems $$\Label{goursat22} \left\{ \begin{aligned} &\Delta^2 u+cu=f\\& P_t|(u-g), \end{aligned} \right.$$ where $P_t(x,y)$, for $t> 0$, denotes the divisor $$\Label{Pt} P_t(x,y):=xy(x-ty)(x-y/t).$$ Recall that $A=A(t)$ denotes the unimodular number given by (with $a=t$). Let us denote by $\beta=\beta(t)$ the number $\beta\in (0,2\pi)$ such that $A=e^{2\pi i \beta}$. We shall prove the following result. Let $t> 0$ and $\beta:=\beta(t)$ as defined above. Suppose that $\beta$ satisfies the Diophantine condition $$\Label{diophantine2} \left |\beta-\frac{n}{m}\right |\geq \frac{C}{m^2},\quad \forall n,m\in \mathbb Z, m\neq 0 ,$$ for some constant $C>0$. Then, for any $c\in A(B_R)$, there exists $0<r\leq R$ such that the Goursat problem has a unique solution $u\in A(B_r)$ for every $f,g\in A(B_R)$. Theorem \[helmdel2\] is a direct consequence of Theorem \[helmdelp\], with $p=2$, and the following proposition. Let $t>0$, $a=(a_1,a_2,a_3):=(0,t,1/t)$, and let $M_{m,p,a}$ be the matrix defined by with $p=2$. If $\beta=\beta(t)$ satisfies $$\Label{diophantine3} \left |\beta-\frac{n}{m}\right |\geq \frac{C}{m^\mu},\quad \forall n,m\in \mathbb Z, m\neq 0 ,$$ for some constant $C>0$, then $$\Label{detAk} \left|\det M_{m,p,a}\right| \geq \frac{D}{m^{2\mu -2}},$$ for some $D>0$. It is easy to check that the unimodular numbers $(A_1,A_2,A_3)$ that correspond to the vector $a$ is $(-1,A,B)$, where $AB=-1$ and, in view of the discussion preceding Corollary \[homodel1\], $$\Label{Am-1} |A^m-1|\geq \frac{C'}{m^{\mu-1}}.$$ (Of course, $A$ is given by , but only the above two facts will be needed in the proof.) To prove the proposition, it suffices, in view of Remark \[rmkmatrix\], to show that $|N_m|\geq C'/m^{2\mu-2}$, where $$\Label{newmatrix} N_{m}:=M_{m-4,2,a} =\det \left( \begin{array}{ccc} -2 & (-1)^{m-1}-1 & (-1) ^{m}-1 \\ A -1 & A ^{m-1}-1 & A ^{m}-1 \\ B -1 & B ^{m-1}-1 & B ^{m}-1 \end{array} \right) .$$ We obtain, since $AB=-1$, $$A ^{m}N_{m}=\det \left( \begin{array}{ccc} -2 & (-1)^{m-1}-1 & (-1) ^{m}-1 \\ A -1 & A ^{m-1}-1 & A ^{m}-1 \\ -A ^{m-1}-A ^{m} & A (-1)^{m-1}-A ^{m} & (-1)^{m}-A ^{m} \end{array} \right) .$$ If $m$ is even, then $$A ^{m}N_{m}=\det \left( \begin{array}{ccc} -2 & -2 & 0 \\ A -1 & A ^{m-1}-1 & A ^{m}-1 \\ -A ^{m-1}-A ^{m} & -A -A ^{m} & 1-A ^{m} \end{array} \right) .$$ A straightforward calculation shows that $$\Label{even} A^mN_M=4A(A^m-1)(A^{m-2}-1).$$ If $m$ is odd, then $$A ^{m}N_{m}=\det \left( \begin{array}{ccc} -2 & 0 & -2 \\ A -1 & A ^{m-1}-1 & A ^{m}-1 \\ -A ^{m-1}-A ^{m} & A -A ^{m} & -1-A ^{m} \end{array} \right) .$$ This time we get $$\Label{odd} A^mN_M=-2(A^{m-1}-1)^2(A^{2}+1).$$ The conclusion $|N_m|\geq C'/m^{2\mu-2}$ follows easily from and . This completes the proof of the proposition. Divergence of formal solutions when $p=1$ and $\tau=0$. ======================================================= We now show that, for $p=1$ and irrational angles $\alpha$ between the two lines $\Gamma_1$ and $\Gamma_2$, the formal solution $u$ to , with $f$ convergent and $g\equiv 0$, need not converge when $\tau$, given by , is zero. A discussion similar to the one that follows can also be found in [@Shap89]. Using the notation and setup in the proof of Theorem \[estimate\], let us choose $f$ such that for each $m$ we have, for $k+l=m$, $$f_{kl}=\left\{\begin{aligned}&R^{-m},\quad k=0\\&0,\quad k>0. \end{aligned}\right.$$ Note that $f\in A(B_R)$. Let us consider the Goursat problem with $g=0$. By following the argument in the proof of Theorem \[estimate\] above, we conclude that the formal solution is of the form $u=v+w$, where $w$ is the formal solution to and $v(x,y)$ is the formal solution to . Hence, $v$ is of the form $v(x,y)=\phi(z)+\psi(\bar z)$. It is well known that the solution $w$ to converges to a function in $A(B_R)$ (see [@Shap89]; see also [@EbRe06]). Thus, the solution $u$ to the Goursat problem converges if and only if the two power series $\phi(z)=\sum_m b_{m}z^m$ and $\psi(\bar z)=\sum_m c_{m}\bar z^m$ converge. With $p=1$, it is easy to solve the system of equations for $b_m$ and $c_m$ explicitly and we obtain $$b_m=\frac{1}{(1-A^m)}\frac{A-1}{2R^{m-2}(m-1)},$$ (A similar identity holds, of course, for $c_m$.) The radius of convergence of the series $\phi(z)=\sum_m b_m z^m$ is $$R\liminf_{m\to\infty} |1-A^m|^{1/m}=0,$$ proving the assertion above that the solution $u$ does not converge. We conclude this paper by giving an explicit example of a number $\beta$ in $A=e^{2\pi i\beta}$ such that $\tau=0$. [Let us define $$\beta:=\sum_{k=1}^\infty 10^{-p_k},$$ where $p_k$ is defined recursively by $p_1=1$ and $p_{k+1}=p_k+k\, 10^{p_k}$. Note that, for every $N$, the rational number $$r_N:=\sum_{k=1}^N10^{-p_k}=\frac{q_N}{10^{p_N}}$$ satisfies $$|\beta-r_N|\leq\frac{2}{10^{p_{N+1}}}.$$ Consider the subsequence $m_N:=10^{p_N}$ and note that $$|A^{m_N}-1|\leq C \inf_{p,q\in \mathbb Z_+}q\left |\beta-\frac{p}{q}\right|\leq 2\frac{10^{p_N}}{10^{p_{N+1}}}=\frac{2}{10^{p_{N+1}-p_N}}$$ Thus, we have $$|A^{m_N}-1|^{1/m_N}\leq \frac{C}{10^{(p_{N+1}-p_N)/10^{p_N}}}=\frac{C}{10^N}\to 0,$$ which shows that $\tau=\liminf_{k\to\infty}|A^k-1|^{1/k}=0$. ]{} [99]{} P. Ebenfelt, H. Render, *On the mixed Cauchy problem with data on singular conics,* Preprint, (2006). P. Ebenfelt, H.S. Shapiro, *The mixed Cauchy problem for holomorphic partial differential equations,* J. D’Analyse Math. 65 (1996) 237–295. L. Gårding, [*Une variante de la méthode de majoration de Cauchy. (French)*]{}, Acta Math. 114, (1965) 143–158. L. Hörmander, [*The analysis of linear partial differential operators. I. Distribution theory and Fourier analysis.*]{} Second edition. Grundlehren der Mathematischen Wissenschaften, [**256**]{}. Springer-Verlag, Berlin, 1990. J. Leray, [*Caractère non fredholmien du probl\` eme de Goursat. (French),*]{} J. Math. Pures Appl. 53 (1974), 133–136. H. Render, *Real Bargmann spaces, Fischer decompositions and Sets of Uniqueness for Polyharmonic Functions,* submitted. K. F. Roth, *Rational approximations to algebraic numbers and Corrigendum*, Mathematika, 2, (1955), 1–20 and 168. H.S. Shapiro, *An algebraic theorem of E. Fischer and the Holomorphic Goursat Problem,* Bull. London Math. Soc. 21 (1989), 513–537. M. Waldschmidt, [*Diophantine approximation on linear algebraic groups. Transcendence properties of the exponential function in several variables.*]{} Grundlehren der Mathematischen Wissenschaften, [**326**]{}. Springer-Verlag, Berlin, 2000. M. Yoshino, [*Spectral property of Goursat problems.*]{} Tokyo J. Math. 4 (1981), 55–71. M. Yoshino, [*On the solvability of Goursat problems and a function of number theory.*]{} Duke Math. J. 48 (1981), 685–696. [^1]: 2000 [*Mathematics Subject Classification*]{}. 35A10, 35J05 [^2]: The first author is supported in part by DMS-0401215. The second author is supported in part by Grant MTM2006-13000-C03-03 of the D.G.I. of Spain.

--- abstract: 'The self-assembly of submonolayer amounts of Au on the densely stepped Si(553) surface creates an array of closely spaced “atomic wires” separated by 1.5 nm. At low temperature, charge transfer between the terraces and the row of silicon dangling bonds at the step edges leads to a charge-ordered state within the row of dangling bonds with $\times 3$ periodicity. Interactions between the dangling bonds lead to their ordering into a fully two-dimensional (2D) array with centered registry between adjacent steps. We show that as the temperature is raised, soliton defects are created within each step edge. The concentration of solitons rises with increasing temperature and eventually destroys the 2D order by decoupling the step edges, reducing the effective dimensionality of the system to 1D. This crossover from higher to lower dimensionality is unexpected and, indeed, opposite to the behavior in other systems.' author: - 'B. Hafke' - 'C. Brand' - 'T. Witte' - 'B. Sothmann' - 'M.' - 'S.C. Erwin' title: 'Thermally-induced crossover from 2D to 1D behavior in an array of atomic wires: silicon dangling-bond solitons in Si(553)-Au' --- [^1] Physical phenomena associated with low dimensionality are suppressed when the temperature is raised. For example, the 2D fractional quantum Hall effect [@Tsui:PRL48; @Laughlin:PRL50] and the 1D Tomonaga-Luttinger liquid [@Tomonaga:PoTP5; @Luttinger:JoMP4; @Bockrath:Nature397] are only observed at low temperature. In 1D atomic wire systems at low temperatures, Peierls distortions or more general symmetry breakings can open a gap at the Fermi level and lower the total energy by forming a charge density wave (CDW) [@AhnYeom:PRL93; @Frigge:Nature544; @SnijdersWeitering:RevModPhys82] or spin-density wave (SDW) [@Gruener:RevModPhys66; @Andrieux:JPL42; @Mortensen:SSC44; @Sassa:JoESaRP224]. Excitations generally wash out the effects of this anisotropy and hence suppress low-dimensional behavior. The resulting crossover to higher dimensionality at increased temperatures is exhibited by many systems. Recent examples include the atomic wire systems Pt(110)-Br and Si(557)-Pb. In these systems, structural changes are accompanied by a delicate interplay between CDW correlations and short-range interactions of the adsorbate atoms [@Duerrbeck:PRB98] and by correlated spin-orbit order that triggers a metal-to-insulator transition, respectively [@Brand:NatComm6; @Das:JPCM28; @Tegenkamp:PRL95; @Block:PRB84]. The resulting dimensional crossover from 1D to 2D is typical for atomic wire systems. In this Letter we demonstrate the opposite case: a system of coupled atomic wires exhibiting 2D order at low temperatures in which thermal excitations at higher temperatures induce a dimensional crossover to 1D behavior. We identify the mechanism driving this crossover to be the creation of phase solitons and antisolitons [@RiceMele:PRL49; @SuSchrieffer:PRL46], which leads to an reversible order-disorder transition at higher temperatures [@AhnYeom:PRL95]. We track the crossover across its characteristic temperature (approximately 100K) using a combination of a quantitative high resolution spot profile analyzing-low energy electron diffraction (SPA-LEED) study, density-functional theory (DFT) calculations, Monte Carlo statistical simulations, and an exactly solvable analytical model. We studied the self-organized Si(553)-Au atomic wire surface consisting of Au double-atom rows on (111)-oriented Si terraces separated by bilayer steps \[Fig. \[fig:Introductory\]\]. Charge transfer from the terraces leads to incomplete filling of the dangling $sp^3$ orbitals at the Si step edge [@Crain:PRL90; @SnijdersWeitering:PRL96; @AhnYeom:PRL95]. The low-temperature ground state consists of a charge-ordered state with $\times 3$ periodicity along the step edges, which is observed in scanning tunneling microscopy (STM) [@AhnYeom:PRL95; @SnijdersWeitering:PRL96; @ShinYeom:PRB85; @Song:ACSNano9; @Aulbach:PRL111; @Polei:PRL111; @Aulbach:PRB96; @Hafke:PRB94; @Dudy:JPCM29] and LEED experiments [@Hafke:PRB94; @Yeom:NJoP16; @AhnYeom:PRL95; @Dudy:JPCM29]. The $\times 3$ periodicity along the wires represents the simplest way to distribute the available electrons among the row of dangling bonds while maximizing the number of fully saturated dangling bonds (electron lone pairs) [@Aulbach:NanoLett16]. Angle-resolved photoemission spectroscopy measurements [@Crain:PRL90; @Song:ACSNano9; @Yeom:NJoP16; @Krawiec:ASS373] and DFT calculations [@Erwin:NatComm1; @Song:ACSNano9; @Krawiec:PRB81; @Krawiec:ASS373] reveal that the dangling-bond states do not cross the Fermi level. Hence, all the dangling-bond orbitals have integer electron occupancies of 0, 1, or 2. We will refer to orbitals with occupancy 2 as saturated dangling bonds (SDBs) and to those less than 2 as unsaturated dangling bonds (UDBs). Figure \[fig:Introductory\] depicts the arrangement of UDBs and SDBs schematically. The ordering of the Si dangling bond structure is mediated by Coulomb interaction of the UDBs (large spheres) with approximately equal spacing within and across the rows. The SDBs merely provide a compensating background charge to balance the reduced electron occupancy of the UDBs. ![Ground state structure and low-energy excitation of the Si(553)-Au atomic wire system. The underlying substrate consists of Si(111) terraces separated by steps. Each terrace contains a dimerized double row of Au atoms (gold) and a row of Si dangling bonds (gray spheres) at the step edge. The electron occupancy of these dangling bonds creates a ground state with tripled periodicity: for every two saturated dangling bonds (SDBs, small spheres) there is a third, unsaturated dangling bond (UDB, large spheres). At finite temperatures, some of these UDBs (blue) become mobile and hop to adjacent sites (red). This excitation creates a soliton-antisoliton pair that can subsequently dissociate. \[fig:Introductory\]](Fig0.png){width="0.96\columnwidth"} The actual number of electrons in the UDBs has been previously investigated using DFT. The result is sensitive to the choice of the exchange-correlation functional. The original prediction [@Erwin:NatComm1], which used the functional of Perdew, Burke, and Ernzerhof (PBE) [@Perdew:PRL77], was that five electrons are shared among three dangling bonds with an electron configuration (2,2,1) having $\times 3$ periodicity and one unpaired spin [@Hafke:PRB94]. More recent work [@Braun:PRB98], based on the revised PBEsol functional [@Perdew:PRL100], predicted that only four electrons are shared among three dangling bonds, implying the configuration (2,2,0) with no unpaired spins. At present it is not possible to distinguish between these scenarios on the basis of experimental data. Here we consider both possibilities and show that they lead to very different estimates of the order-disorder transition temperature. The experiment was performed under ultra-high vacuum (UHV) conditions at a base pressure lower than $1 \times 10^{-10}$mbar. The Si substrate was cut from an *n*-type Si(553) wafer (phosphorus doped, $0.01\,\Omega$cm). Prior to Au deposition, the sample was cleaned in several short flash-anneal cycles by heating via direct current to $1250\,^\circ$C. Next, 0.48ML (monolayer, referred to the atomic density of a Si(111) surface, i.e. $1\,\text{ML} = 7.83 \times 10^{14}$atoms per cm$^2$) Au was deposited from an electron-beam-heated graphite crucible [@Kury:RSI76] at a substrate temperature of $650\,^\circ$C, followed by a post-annealing step at $850\,^\circ$C for several seconds [@Aulbach:PRL111] and subsequent cooling to 60 K on a liquid helium cryostat. The temperature was measured by an ohmic sensor (Pt100) directly clamped to the back of the sample. ![SPA-LEED pattern of Si(553)-Au at an electron energy of 150 eV and temperatures (a) 60 K and (b) 180 K. The $\times 2$ streaks between the rows of sharp integer-order spots arise from dimerized Au double rows on the (111)-oriented terraces. The rows of elongated spots at $\times 3$ positions indicate the tripled periodicity and long-range order of the UDBs at the Si step edge. The intensity of the $\times 2$ streak is nearly unaffected at higher temperature, while the $\times 3$ features fade away. In (a) the unit cells of the Au (blue) and Si (green) sublattices as well as the directions of the line profiles LP$_\parallel$ and LP$_\perp$ (Fig. \[fig:SPALEED\_curves\]) are indicated (for more details see [@Hafke:PRB94]). (c) Surface structural model showing Si step-edge atoms (gray) and Au atoms (gold). The unit cells are depicted with the same color coding as in (a).\[fig:Diff.Pattern\]](Fig1.jpg){width="0.96\columnwidth"} ![FWHM of the $\times 3$ diffraction spots (red data points) as function of temperature in (a) $\left[\overline{3}\,\overline{3}\,10\right]$ direction and (b) $\left[1\,\overline{1}\,0\right]$ direction, as indicated in Fig. \[fig:Diff.Pattern\](a). Results from Monte Carlo simulations based on DFT interactions are shown in (b) by blue squares for the (2,2,1) configuration and triangles for the (2,2,0) configuration. The result from the analytical model in Eq.  is shown in black. The increase of the FWHM in (a) indicates loss of inter-wire correlation, while in (b) it indicates a decreasing correlation length along the steps. At low temperatures the FWHM is constant in $\left[\overline{3}\,\overline{3}\,10\right]$ direction up to $\sim$90 K. Insets in (a,b): Line profiles for both directions at various temperatures (shifted for better visibility). (c) Structural model of creation and separation of a soliton-antisoliton pair. Charge is transferred from an SDB to a UDB, generating a hop of the UDB to a neighboring site and creating a soliton-antisoliton pair. If this pair separates then a phase-shifted domain with $\times 3$ periodicity is formed.\[fig:SPALEED\_curves\]](Fig2.jpg){width="0.96\columnwidth"} At 60 K, the SPA-LEED pattern \[Fig. \[fig:Diff.Pattern\](a)\] reveals spots at $\times 3$ positions and streaks at $\times 2$ positions in the $\left[1\,\overline{1}\,0\right]$ direction. The latter indicates the formation of Au atomic wires. The spacing of the $\times 1$ spots corresponds to the reciprocal lattice constant $a_\parallel^* = 2\pi/(3.84\,\text{\AA})$ of the Si substrate along the steps. The $\times 3$ spots arise from ordering of the UDBs within the rows; hence, the UDBs have an intra-row separation of $3\times a_\parallel = 11.5\,\text{\AA}$. The UDBs in different rows are in registry: recent investigations by SPA-LEED and STM reveal a centered $p(1 \times 3)$ arrangement \[Figs. \[fig:Diff.Pattern\](a,c)\] [@Hafke:PRB94]. In the $\left[\overline{3}\,\overline{3}\,10\right]$ direction the reciprocal step-to-step distance is $a_\perp^* = 2\pi/(14.8\,\text{\AA})$ [@Crain:PRL90] and thus the separation between the UDB rows is 14.8 Å. Hence, at low temperatures the UDBs in Si(553)-Au are arranged in rows in a fully ordered 2D array with approximately equal spacing within (11.5 Å) and across (14.8 Å) the rows. The $\times 2$ streaks are attributed to the dimerized double row of Au atoms on the (111) terrace of the surface \[gold spheres in Fig. \[fig:Diff.Pattern\](c); the unit cell is shown by the blue-shaded areas in Figs. \[fig:Diff.Pattern\](a,c)\] [@Krawiec:PRB81; @KrawiecJalochowski:PRB87]. We did not detect any $\times 6$ periodicity in the $\left[1\,\overline{1}\,0\right]$ direction, indicating that the Au atoms and Si step-edge atoms are structurally decoupled [@Hafke:PRB94; @Aulbach:PRB96; @Braun:PRB98]. At 180 K, the intensity of the $\times 3$ spots fades markedly \[Fig. \[fig:Diff.Pattern\](b)\], in agreement with an earlier study [@AhnYeom:PRL95], while the intensity of the $\times 2$ streaks is nearly unaffected. To analyze the evolution of the diffraction pattern between 60 K to 190 K (heating rate 0.13K/s), we recorded a series of line profiles \[Fig. \[fig:SPALEED\_curves\](a,b)\] through the $\times 3$ spots, in two orthogonal directions: $\left[\overline{3}\,\overline{3}\,10\right]$ (LP$_\perp$, across the steps) and $\left[1\,\overline{1}\,0\right]$ (LP$_\parallel$, along the steps). The $\times 3$ diffraction spots \[insets of Fig. \[fig:SPALEED\_curves\](a,b)\] of each of the line profiles were best fitted by a series of equidistant Lorentzian functions. No Gaussian-like central spike is found and the positions of the $\times 3$ diffraction spots do not shift with temperature. Across the steps, the full width at half maximum (FWHM) $\kappa_\perp$ steadily increases as the temperature is raised from 60K to 130K. Eventually, the spots merge into streaks, consistent with the vanishing of the $\times 3$ periodicity reported earlier [@AhnYeom:PRL95]. Along the steps, the FWHM $\kappa_\parallel$ is relatively constant up to about 100 K and then steadily increases as the temperature is raised further. This broadening of the $\times 3$ diffraction spots is due to increasing disorder in the arrangement of UDBs. This type of disorder originates from a simple microscopic process in which an electron (or two electrons, for the (2,2,0) configuration) hops from an SDB onto a neighboring UDB \[middle panel of Fig. \[fig:SPALEED\_curves\](c)\]. As long as these electron hops do not bring neighboring UDBs closer than $2a_\parallel$, the configuration is metastable. We used DFT to determine the formation energy $E_0$ of this elementary excitation, which can be viewed as a soliton-antisoliton bound pair. The calculations were performed in a 1$\times$6 cell of Si(553)Au with four silicon double layers plus the reconstructed surface layer and a vacuum region of 10Å. All atomic positions were relaxed except the hydrogen-passivated bottom double layer. Total energies and forces were calculated using the generalized-gradient approximation of Perdew, Burke, and Ernzerhof (PBE) for the (2,2,1) configuration and the PBEsol revision of that functional for the (2,2,0) configuration, with projector-augmented wave potentials as implemented in VASP [@Kresse:PRB47; @Bloechl:PRB50; @Perdew:PRL100]. The plane-wave cutoff was 250 eV and the sampling of the surface Brillouin zone was $6 \times 6$. For the (2,2,1) ground-state configuration we find $E_0=30$ meV, suggesting these defects will be numerous at temperatures above $\sim$300 K, which is consistent with our experimental data. For the (2,2,0) configuration we find $E_0=85$ meV, implying a much higher temperature scale of $\sim$1000 K. To investigate the concentration and distribution of defects as a function of temperature, we used the Metropolis Monte Carlo method to sample the steady-state arrangement of UDBs in an infinite array of dangling-bond wires with the Si(553)-Au geometry. We performed 10$^7$ trial hops at each temperature and computed the diffraction intensity from the positions of the UDBs. The spectra were convolved with a Gaussian to account for instrumental broadening in the experimental data. For the (2,2,1) configuration, the resulting FWHM of the $\times$3 peaks is constant up to $\sim$100 K and then increases gradually with temperature, in agreement with experiment but with smaller values \[blue squares in Fig. \[fig:SPALEED\_curves\](b)\]. For the (2,2,0) configuration the FWHM is flat up to temperatures about three times higher (blue triangles), as expected from the $q^2$ scaling of the Coulomb energy. See Supplementary Material for additional details. Even though the geometry of our model is 2D, the energy scale of Coulomb interactions across the wires is only 0.1 meV, three orders of magnitude smaller than $E_0$. Hence, the interactions in the Monte Carlo simulations are essentially 1D. Our DFT calculations, however, reveal a much stronger interaction across the wires of order 1 meV. These may arise from the interaction of strain fields from the UDBs but other sources may contribute as well. Regardless of their origin, we turn now to investigating their role in the order-disorder transition. We show next that by including these 2D interactions, the FWHM at all temperatures is brought into quantitatively excellent agreement with experiment. We constructed an exactly solvable Potts model Hamiltonian describing the dynamics of coupled wires and the resulting steady-state FWHM of the $\times$3 peaks as a function of temperature: $$\label{eq:Sothmann-Hamiltonian} \mathcal{H} = \sum_i \left[-b\delta_{u_i,u_{i+1}}-a\delta_{u_i,\text{c}}\right],$$ where $\delta_{i,j}$ denotes the Kronecker delta. A single UDB can take three positions within each unit cell $i$: left, center and right, $u_i=\{\text{l},\text{c},\text{r}\}$. The first term, with parameter $b$, describes the energy needed to displace neighbouring UDBs relative to each other: specifically, the energy needed to create a soliton-antisoliton pair within one wire is $2b$. The second term, with parameter $a$, favors the occupation of the central position and arises from the coupling of the wire to neighboring wires. See Supplementary Material for additional details. The model fits best to our experimental data for $a = 2.1$ meV and $b = 21$ meV. These fitted values are also consistent with our DFT results: $a$ should be equal to the calculated energy difference per UDB, $2.1$ meV, between (2,2,1) configurations in staggered and centered alignments, and $b$ corresponds to $E_0/2 = 15$ meV. In the Supplementary Material, we derive analytical expressions for the profiles and FWHM of the $\times 3$ peaks as a function of temperature. The resulting FWHM, convolved as above with a Gaussian, is now in excellent agreement with our experimental results \[black curve in Fig. 2(b)\]. This improved agreement demonstrates that 2D coupling between neighboring wires indeed plays an important, central role in the order-disorder transition. We turn now to the crossover from 2D to 1D behavior. At temperatures above $\sim$120 K, the $\times3$ diffraction spots are well described by a standard Lorentzian. At temperatures below $\sim$90 K, the 2D character of the diffraction is more pronounced and hence the spot profiles are described by a Lorentzian to the power 3/2 [@Wollschlaeger:PRB44]. To characterize the transition between these two limits, we fit the spot profiles to a generalized Lorentzian, $$\label{eq:Lorentzian} \mathcal{L}(k_\parallel) = \frac{\Gamma(\nu)}{\sqrt{\pi} \Gamma(\nu - 1/2)} \cdot \frac{\kappa_\parallel^{2\nu - 1}}{\left[ (k_\parallel - k_0)^2 + \kappa_\parallel^2 \right]^\nu},$$ where $k_\parallel$ is the reciprocal space coordinate in the $\left[1\,\overline{1}\,0\right]$ direction, $k_0$ is the position of the $\times 3$ diffraction spot, and $\Gamma(x)$ is the Gamma function. The parameter $\nu = (d + 1)/2$ characterizes the dimensionality $d$ of the system: $\nu = 3/2$ describes 2D systems while $\nu = 1$ describes 1D systems [@Wollschlaeger:PRB44; @Lent:SurfSci139; @Pukite:SurfSci161]; we constrained $\nu$ to lie in this range. We find that $\nu$ exhibits a well-defined transition from 1.5 to 1.0 between $T_{-}=93$K and $T_{+}=128$K (Fig. \[fig:2d-1d-transition\]). The transition begins at about the temperature for which the FWHM $\kappa_{\parallel}$ along the steps begins to increase \[Fig. 2(b)\]. Fitting the spot profiles without allowing $\nu$ to vary leads to significantly worse fits (insets to Fig. \[fig:2d-1d-transition\]). The transition is completed at $T_{+}$, where the FWHM $\kappa_\perp$ across the wires exceeds the size of the surface Brillouin zone \[Fig. 2(a)\], reflecting the complete loss of long-range order across the wires. The underlying origin of this dimensional crossover is subtle but simple: the approximate geometrical isotropy of the 2D array of UDBs is broken by the strong anisotropy of the energy scales for creating disorder across and within the UDB wires. At temperatures above $T_{-}$ soliton defects are still rare, but a rapidly growing fraction of the wire rows undergoes registry shifts with respect to each other and hence the 2D low-temperature state begins to behave like a collection of uncoupled 1D wires. As the temperature approaches $T_{+}$ this crossover becomes nearly complete. See the Supplementary Material for additional discussion, modeling, and analysis. ![Temperature dependence of the exponent $\nu$ of the generalized Lorentzian of Eq.  in the $\left[1\,\overline{1}\,0\right]$ direction. The exponent drops from $1.5$ at $T_{-}=93$K to 1.0 at $T_{+}=128$K, indicating a crossover from 2D to 1D behavior. Insets: Experimental and fitted spot profiles at 60 K and at 128 K. At 60 K the profile is best fit by $\nu = 3/2$ (2D behavior), while at $T_{+}$ the best fit is $\nu = 1$ (1D behavior).\[fig:2d-1d-transition\]](Fig3.jpg){width="0.96\columnwidth"} To summarize, we have shown that silicon dangling-bond solitons in Si(553)-Au are created by thermal excitation. These defects interact via Coulomb forces within each step-edge atomic wire and via another mechanism, probably strain, across the wires. As the temperature is raised, the resulting disorder destroys the $\times$3 positional long-range order of the UDBs within each wire and their registry across the wires. The nature of the interactions and their respective energy scales create a dimensional crossover of the order-disorder transition from 2D at low temperature to 1D at high temperature. The generality of this crossover can readily be investigated—both experimentally and using our theoretical methods—in other atomic wire systems in the Ge/Si($hhk$)-Au family, where differences in the surface morphology and ground-state electron configuration may lead to further expanding our understanding of low-dimensional systems. We acknowledge fruitful discussions with J. Aulbach, F. Hucht, J. König and J. Schäfer. This work was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – Projektnummer 278162697 – SFB 1242 and through Projektnummer 194370842 – FOR1700. B.S. acknowledges financial support from the Ministry of Innovation NRW via the “Programm zur Föderung der Rückkehr des hochqualifizierten Forschungsnachwuchses aus dem Ausland”. This work was partly supported by the Office of Naval Research through the Naval Research Laboratory’s Basic Research Program (SCE). Computations were performed at the DoD Major Shared Resource Center at AFRL. [^1]: Corresponding author

--- abstract: 'The coincidence between polynomial neural networks and matrix Lie maps is discussed in the article. The matrix form of Lie transform is an approximation of the general solution of the nonlinear system of ordinary differential equations. It can be used for solving systems of differential equations more efficiently than traditional step-by-step numerical methods. Implementation of the Lie map as a polynomial neural network provides a tool for both simulation and data-driven identification of dynamical systems. If the differential equation is provided, training a neural network is unnecessary. The weights of the network can be directly calculated from the equation. On the other hand, for data-driven system learning, the weights can be fitted without any assumptions in view of differential equations. The proposed technique is discussed in the examples of both ordinary and partial differential equations. The building of a polynomial neural network that simulates the Van der Pol oscillator is discussed. For this example, we consider learning the dynamics from a single solution of the system. We also demonstrate the building of the neural network that describes the solution of Burgers’ equation that is a fundamental partial differential equation.' author: - Andrei Ivanov - Sergei Andrianov title: Matrix Lie Maps and Polynomial Neural Networks for Solving Differential Equations --- Introduction ============ Traditional methods for solving systems of differential equations imply a numerical step-by-step integration of the system. For some problems, this integration leads to time-consuming algorithms because of the limitations on the time interval that is used to achieve the necessary accuracy of the solution. From this perspective, neural networks as universal function approximation can be applied for the construction of the solution in a more performance way. In the article [@ref1], the method to solve initial and boundary value problems using feedforward neural networks is proposed. The solution of the differential equation is written as a sum of two parts. The first part satisfies the initial/boundary conditions. The second part corresponds to a neural network output. The same technique is applied for solving Stokes problem in [@ref2; @ref3] and implemented in code in [@ref4]. In the article [@ref5], the neural network is trained to satisfy the differential operator, initial condition, and boundary conditions for the partial differential equation (PDE). The authors in [@ref6] translate a PDE to a stochastic control problem and use deep reinforcement learning for an approximation of derivative of the solution with respect to the space coordinate. Other approaches rely on the implementation of a traditional step-by-step integrating method in a neural network basis [@ref7; @ref8]. In the article [@ref8], the author proposes such an architecture. After fitting, the neural network produces an optimal finite difference scheme for a specific system. The backpropagation technique through an ordinary differential equation (ODE) solver is proposed in [@ref9]. The authors construct a certain type of neural network that is analogous to a discretized differential equation. This group of methods requires a traditional numerical method to simulate dynamics. Polynomial neural networks are also widely presented in the literature [@ref10; @ref11; @ref12]. In the article [@ref10], the polynomial architecture that approximates differential equations is proposed. The Legendre polynomial is chosen as a basis function of hidden neurons in [@ref11]. In these articles, the polynomial architectures are used as black box models, and the authors do not explain its connection to the theory of ODEs. In all the described approaches, the neural networks are trained to consider the initial conditions of the differential equations. This means that the neural network should be trained each time when the initial conditions are changed. The above-described techniques are applicable to the general form of differential equations but are able to provide only a particular solution of the system. In the article, we consider polynomial differential equations. Such nonlinear systems arise in different fields such as automated control, robotics, mechanical and biological systems, chemical reactions, drug development, molecular dynamics, and so on. Moreover, often it is possible to transform a nonlinear equation to a polynomial view with some level of accuracy. For polynomial differential equations, it is possible to build a polynomial neural network that is based on the matrix Lie transform and approximates the general solution of the system of equations. Having a Lie transform–based neural network for such a system, dynamics for different initial conditions can be estimated without refitting of the neural network. Additionally, we completely avoid numerical ODE solvers in both simulation and data-driven system learning by describing the dynamics with maps instead of step-by-step integrating. Proposed Neural Network {#sec2} ======================= The proposed architecture is a neural network representation of a Lie propagator for dynamical systems integration that is introduced in [@ref13] and is commonly used in the charged particle dynamics simulation [@ref13; @ref14]. We consider dynamical systems that can be described by nonlinear ordinary differential equations, $$\label{odesystem} \frac{d}{dt}{\mathbf{X}}= {\mathbf{F}}(t, {\mathbf{X}}) = \sum_{k=0}^{\infty} P^{1k}(t){\mathbf{X}}^{[k]},$$ where $t$ is an independent variable, ${\mathbf{X}}\in R^n$ is a state vector, and ${\mathbf{X}}^{[k]}$ means $k$-th Kronecker power of vector ${\mathbf{X}}$. There is an assumption that function ${\mathbf{F}}$ can be expanded in Taylor series with respect to the components of ${\mathbf{X}}^{[k]}$. The solution of (\[odesystem\]) in its convergence region can be presented in the series [@ref15; @ref16], $$\label{Lie} {\mathbf{X}}(t|t_0) = \mathcal M(t|t_0) \circ {\mathbf{X}}_0 = \sum_{k=0}^{\infty} M^{1k}(t|t_0){\mathbf{X}}_0^{[k]},$$ where ${\mathbf{X}}_0 = {\mathbf{X}}(t_0)$ In [@ref14], it is shown how to calculate matrices $M^{1k}$ by introducing new matrices $P^{ij}$. The main idea is replacing (\[odesystem\]) by the equation $$\label{Lie_map} \frac{d}{dt} M^{ik}(t|t_0) = \sum_{j=i}^{k} P^{ij}(t)M^{jk}(t|t_0),\;1\leq i < k.$$ This equation should be solved with initial condition $M^{kk}(t_0) = I^{[k]},\;M^{jk}(t_0) = 0, j\neq k$, where $I$ is the identity matrix. Theoretical estimations of accuracy and convergence of the truncated series in solving of ODEs can be found in [@ref17]. The transformation $\mathcal{M}$ can be considered as a discrete approximation of the evolution operator of (\[odesystem\]) for initial time $t_0$ and interval $\Delta t$. This means that the evolution of the state vector ${\mathbf{X}}_0 = {\mathbf{X}}(t_0)$ during time $\Delta t$ can be approximately calculated as ${\mathbf{Y}}= \mathcal{M} \circ {\mathbf{X}}_0$. Hence, instead of solving the system of ODEs numerically, one can apply a calculated map and avoid a step-by-step integrating. Neural Network Representation of Matrix Lie Transform ----------------------------------------------------- The proposed neural network implements map $\mathcal M : {\mathbf{X}}\rightarrow {\mathbf{Y}}$ in form of $$\label{nnmap} {\mathbf{Y}}= W_0 + W_1\,{\mathbf{X}}+W_2\,{\mathbf{X}}^{[2]}+\ldots+W_k\,{\mathbf{X}}^{[k]},$$ where ${\mathbf{X}}, {\mathbf{Y}}\in R^n$, $W_i$ are weight matrices, and ${\mathbf{X}}^{[k]}$ means $k$-th the Kronecker power of vector ${\mathbf{X}}$. For a given system of ODEs (\[odesystem\]), one can compute matrices $W_i = M^{1k}$ in accordance with (\[Lie\_map\]) up to the necessary order of nonlinearity. Fig. \[fig1\] presents a neural network for map (\[nnmap\]) up to the third order of nonlinearities for a two-dimensional state. In each layer, the input vector ${\mathbf{X}}= (x_1, x_2)$ is consequently transformed into ${\mathbf{X}}^{[2]} = (x_1^2, x_1x_2, x_2^2)$ and ${\mathbf{X}}^{[3]} = (x_1^3, x_1^2x_2, x_1x_2^2$, $x_2^3)$ where weighted sum is applied. The output Y equals to the sum of results from every layer. In the example, we reduce Kronecker powers for decreasing of weights matrices dimension (e.g., ${\mathbf{X}}^{[2]} = (x_1^2, x_1x_2, x_2x_1, x_2^2) \rightarrow (x_1^2, x_1x_2, x_2^2)$). ![Neural network representation of third order matrix Lie map.[]{data-label="fig1"}](_Fig_1.pdf){width="55.00000%"} Fitting Neural Network ---------------------- If the differential equation is provided, the training of neural network is not necessary. The weights of the network can be calculated directly from the equation following the relation (\[Lie\_map\]). On the other hand, for data-driven system learning, the weights in form of (\[nnmap\]) can be fitted without any assumptions on view of differential equations. To fit a proposed neural network, the training data is presented as a multivariate time series (table \[table1\]) that describes the evolution of the state vector of the dynamical system in a discrete time. In a general case, each step $t_i \rightarrow t_{i+1}$ should be described as map $\mathcal M_i(t_i): {\mathbf{X}}_i \rightarrow {\mathbf{X}}_{i+1}$, but if the system (\[odesystem\]) is time independent, then weights $W_i$ depends only on time interval $\Delta t = t_{i+1} - t_i$. $t_0$ $t_1$ $\ldots$ $t_{m-1}$ $t_m$ --------------------- ------------------------------ ---------- -------------------------------------------- ------------ $x_0(t_0)$ $x_0(t_1)$ $\ldots$ $x_0(t_{m-1})$ $x_0(t_m)$ $x_1(t_0)$ $x_1(t_1)$ $\ldots$ $x_1(t_{m-1})$ $x_m(t_m)$ $\ldots$ $\ldots$ $x_n(t_0)$ $x_n(t_1)$ $\ldots$ $x_n(t_{m-1})$ $x_n(t_m)$ INPUT $\rightarrow$ $\mathcal M_1$ $\rightarrow$ $\ldots$ $\rightarrow$ $\mathcal M_m$ $\rightarrow$ OUTPUT : Discrete states of a dynamical system for training the proposed network.[]{data-label="table1"} Ordinary Differential Equations {#sec3} =============================== In this section, we consider the Van der Pol oscillator. The equation is widely used in the physical sciences and engineering and can be used for the description of the pneumatic hammer, steam engine, periodic occurrence of epidemics, economic crises, depressions, and heartbeat. The equation has well-studied dynamics and is widely used for testing of numerical methods (e.g., [@ref18]). Simulation of the Van der Pol Oscillator ---------------------------------------- The Van der Pol oscillator is defined as the system of ODEs $x'' = x' - x - x^2x'$ that can be presented in the form of $$\label{vdp_ode} \begin{aligned} x' &= y,\\ y' &= y-x-x^2y. \end{aligned}$$ The results of numerical integration of the system with the implicit Adams method of eighth order with the maximum time step $\Delta t = 0.01$ are presented in Fig. \[fig2\] with red lines. The four different particular solutions with initial conditions, $(-2, 4), (1, 2), (2, -2)$, and $(-3, -3)$, were calculated. ![Simulation of the Van der Pol oscillator. Red lines for the implicit Adams method of eighth order, blue dots for matrix Lie map of third order.[]{data-label="fig2"}](Fig_2.png){width="65.00000%"} Another method for simulating the dynamics is mapping approach. The weights of the matrix Lie map can be calculated up to the necessary order of nonlinearity based on the equation (\[Lie\_map\]). For instance, for the third order and the same time interval, it yields weight matrices $$\begin{aligned} &W_0 = \begin{pmatrix} 0\\ 0\\ \end{pmatrix};\; W_1 = \begin{pmatrix} 0.99995067& 0.01004917\\ -0.01004917& 1.00999984\\ \end{pmatrix};\; W_2 = \begin{pmatrix} 0& 0& 0\\ 0& 0& 0\\ \end{pmatrix};\\ \\ &W_3= \begin{pmatrix} 1.59504733e\text{-}7& -4.94822066e\text{-}5& -3.20576750e\text{-}7& -7.90629025e\text{-}10\\ 4.94821629e\text{-}5& -1.00975145e\text{-}2& -9.96173322e\text{-}5& -3.30168067e\text{-}07\\ \end{pmatrix}. \end{aligned}$$ The corresponding polynomial neural network implements transformation $${\mathbf{X}}_{i+1} = \mathcal M\circ {\mathbf{X}}_i = W_0 + W_1 \begin{pmatrix} x_i\\ y_i\\ \end{pmatrix} +W_2 \begin{pmatrix} x_i^2\\ x_iy_i\\ y_i^2 \end{pmatrix} +W_3 \begin{pmatrix} x_i^3\\ x_i^2y_i\\ x_iy_i^2\\ y_i^3 \end{pmatrix}.$$ The results of the numerical integration of the system with the neural network are presented in Fig. \[fig2\] with blue dots. Note that for the matrix Lie maps, the accuracy of the truncation of the series (order of nonlinearity of the transformation) and the accuracy of weights calculation should be considered separately. The theory of the accuracy and convergence of the truncated series (\[Lie\]) in solving ODEs can be found in [@ref13; @ref17]. From a practical perspective, the accuracy of the simulation provided by a polynomial neural network can be estimated with respect to the traditional numerical solver. For example, the mean relative errors between the predictions of the Lie map–based networks of the third, fifth, and seventh orders of nonlinearity with respect to the numerical solution calculated with the Adams method of eighth order are equal to $0.0110$, $0.0004$, and $4.7\cdot10^{-6}$, respectively. Learning of the Van der Pol Oscillator -------------------------------------- In the previous section, we described how weights for the proposed polynomial neural network can be calculated based on the equation. On the other hand, when the equation is not known, but a particular solution is provided, the weights can be fitted by a neural network without any assumptions in view of differential equations. A particular solution $\{{\mathbf{X}}_i\}_{i=1;n}$ of the system with the initial condition ${\mathbf{X}}_0 = (-2, 4)$ can be generated by numerically integrating system (\[vdp\_ode\]) with time step $\Delta t = 0.01$ during time $T = 10$. Having this training data set, the proposed neural network can be fitted with the mean squared error (MSE) as a loss function based on the norm $$||{\mathbf{X}}_{i+1} - \mathcal M\circ {\mathbf{X}}_i ||= || \begin{pmatrix} x_{i+1}\\ y_{i+1}\\ \end{pmatrix} -W_0 - W_1 \begin{pmatrix} x_i\\ y_i\\ \end{pmatrix} -W_2 \begin{pmatrix} x_i^2\\ x_iy_i\\ y_i^2 \end{pmatrix} -W_3 \begin{pmatrix} x_i^3\\ x_i^2y_i\\ x_iy_i^2\\ y_i^3 \end{pmatrix} ||.$$ We implemented the above-described technique in Keras/TensorFlow and fitted a third-order Lie transform–based neural network with an Adamax optimizer. ![Training data for the neural network (red dots) and provided predictions (lines).[]{data-label="fig3"}](Fig_3a.png "fig:"){width="40.00000%"} ![Training data for the neural network (red dots) and provided predictions (lines).[]{data-label="fig3"}](Fig_3b.png "fig:"){width="40.00000%"}\ a) particular solution as the training set b) prediction for new initial conditions The generalization property of the network can be investigated by examining prediction not only from the training data set but also for new initial conditions. Fig. \[fig3\] (a) shows the training set as a particular solution of the system with initial conditions $(-2, 4)$. Fig. \[fig3\] (b) demonstrates predictions that are calculated starting at both the same initial condition and for the new points $(1, 2), (2, -2)$, and $(-3, -3)$. For the prediction starting from the training initial condition, the mean relative error of the predictions is $4.8\cdot10^{-5}$. For the new initial conditions, the mean error is $9.8\cdot 10^{-6}$. Partial Differential Equations {#sec4} ============================== Burgers’ equation is a fundamental partial differential equation that occurs in various areas, such as fluid mechanics, nonlinear acoustics, gas dynamics, and traffic flow. This equation is also often used as a benchmark for numerical methods. For example, one of the problems proposed in the Airbus Quantum Computing Challenge [@ref19] is building a neural network that solves Burgers’ equation with at least the same level of accuracy and higher computational performance as the traditional numerical methods. In the article [@ref20], a feedforward neural network is trained to satisfy Burgers’ equation and certain initial conditions, but the computational performance of the approach is not estimated. In this section, we demonstrate how to build a Lie transform–based neural network that solves Burgers’ equation. The Finite Difference Method for Burgers’ Equation -------------------------------------------------- Burgers’ equation has a form $$\label{burgers_de} \frac{\partial u(t,x)}{\partial t} + u(t,x)\frac{\partial u(t,x)}{\partial x} = \nu \frac{\partial^2 u(t,x)}{\partial x^2}.$$ Following the [@ref19] for benchmarking, we use an analytic solution $$u(t,x) = -2\frac{\nu}{\phi(t,x)}\frac{d\phi}{dx}+4,\;\phi(t,x) = exp\frac{-(x-4t)^2}{4\nu(t+1)} + exp\frac{-(x-4t-2\pi)^2}{4\nu(t+1)},$$ and a traditional numerical method $$\label{FDM} \frac{u^{n+1}_i - u^{n}_i}{\Delta t} + u^{n}_i\frac{u^{n}_i - u^{n}_{i-1}}{\Delta x} = \nu \frac{u^n_{i+1} - 2u^n_i + u^n_{i-1}}{\Delta x^2},$$ where $n$ stands for the time step, and $i$ stands for the grid node. The equation (\[FDM\]) presents a finite difference method (FDM) that consists of an Euler explicit time discretization scheme for the temporal derivatives, an upwind first-order scheme for the nonlinear term, and finally a centered second-order scheme for the diffusion term. The time step for benchmarking is fixed to $\Delta t = 2.5\cdot10^{-4}$ with the uniform spacing of $\Delta x = 2\pi/1000$. Thus, for the numerical solution for times from $t = 0$ to $t = 0.$5 on $x\in[0, 2\pi]$, the method requires the mesh with 1000 steps on space coordinate $x$ and 2000 time steps. It is indicated in [@ref19] that the FDM introduces a dispersion error in the solution (see Fig. \[fig4\], a). Such error can be reduced by increasing the mesh resolution, but then the time step should be decreased to respect the stability constraints of the numerical scheme. ![A benchmark (a) on mesh $1000\times 2000$ and Lie transform–based neural network (b) on mesh $1000\times500$.[]{data-label="fig4"}](Fig_4a.png "fig:"){width="40.00000%"} ![A benchmark (a) on mesh $1000\times 2000$ and Lie transform–based neural network (b) on mesh $1000\times500$.[]{data-label="fig4"}](Fig_4b.png "fig:"){width="40.00000%"}\ a) FDM, $\Delta t = 2.5\cdot10^{-4}$ b) matrix Lie map, $\Delta t = 1.25\cdot10^{-3}$ Lie Transform–Based Neural Network ---------------------------------- Though there are Lie group methods that directly apply Lie theory to PDEs [@ref21; @ref22], we utilize a different approach. We convert the equation (\[burgers\_de\]) to a system of ODEs and build a matrix Lie map in accordance with Section \[sec2\] for this new system. Assuming that the right-hand side of the equation (\[burgers\_de\]) can be approximated by a function $f(x, u(t,x))$ and considering this approximated equation as a hyperbolic one, it is possible to derive the system of ODEs $$\label{burgers_ode} \frac{d}{dt} \begin{pmatrix} {\mathbf{X}}\\ {\mathbf{U}}\end{pmatrix} = \begin{pmatrix} {\mathbf{U}}\\ f(x, {\mathbf{U}}) \end{pmatrix},$$ where ${\mathbf{U}}=(u_1, \ldots, u_{1000})$, $u_i(t) = u(t, x_i)$, and ${\mathbf{X}}=(x_1, \ldots, x_{1000})$ is vector of discrete stamps on space. This transformation from PDE to ODE is well known and can be derived using the method of characteristics and direct method [@ref23]. If $f(x, u(t,x))$ is the same discretization as in (\[FDM\]), then the equation (\[burgers\_ode\]) leads to the system of 2000 ODEs $$\begin{aligned} x'_i &= u_i,\\ u'_i &= f(\nu, u_{i+1}, u_i, u_{i-1}, x_{i+1}, x_i, x_{i-1}), \end{aligned}$$ which can be easily expanded to the Taylor series with respect to the ${\mathbf{X}}$ and ${\mathbf{U}}$ up to the necessary order of nonlinearity. Using this system of ODEs, we have built a Lie transform–based neural network for a time interval $\Delta t = 1.25\cdot10^{-3}$. This time step is five times larger than that used in the benchmarking (see Fig. \[fig5\]). The numerical solution provided by the neural network is presented in Fig. \[fig4\] (b), and the accuracy and performance are compared in Table \[table2\]. --------------------- -------------------- ------------------ ------------- ------------------- **Method** **Time** **Mesh** **Elapsed** **MSE for** **step** **size** **time** $u(0.5,x)$ FDM $2.5\cdot10^{-4}$ $1000\times2000$ 0.055 sec $8.0\cdot10^{-2}$ Lie transform–based neural network $1.25\cdot10^{-3}$ $1000\times500$ 0.016 sec $5.5\cdot10^{-3}$ --------------------- -------------------- ------------------ ------------- ------------------- : Comparison of simulation by FDM and matrix Lie map.[]{data-label="table2"} The built polynomial neural network provides better accuracy with less computational time. If the FDM scheme is adjusted to a higher accuracy, the computational time will be increased even more. Accuracy is calculated as the MSE metric between the numerical solution and its analytic form at final time $t = 0.5$. ![Numerical schemes for FDM and Lie transform–based neural network for (\[burgers\_de\]).[]{data-label="fig5"}](Fig_5a.png "fig:"){width="45.00000%"} ![Numerical schemes for FDM and Lie transform–based neural network for (\[burgers\_de\]).[]{data-label="fig5"}](Fig_5b.png "fig:"){width="45.00000%"} a) FDM, $\Delta t = 2.5\cdot10^{-4}$ b) matrix Lie map, $\Delta t = 1.25\cdot10^{-3}$ Code ==== The implementation of the Lie transform–based neural network in Keras/TensorFlow and the algorithm for map building for autonomous systems are provided at the GitHub repository: <https://github.com/andiva/DeepLieNet>. The notebook <https://github.com/andiva/DeepLieNet/tree/master/demo/VanderPol.ipynb>\ corresponds to Section \[sec3\] and consists of simulation, the definition of metrics for accuracy estimation, neural network configuration, and fitting. The notebook <https://github.com/andiva/DeepLieNet/tree/master/demo/Burgers.ipynb> reproduces the results presented in Section \[sec4\]. Conclusion ========== In the article, we demonstrate the solving of differential equations with polynomial neural networks that are based on matrix Lie maps. Since the weights of the proposed neural network can be directly calculated from the equations, it does not require fitting with respect to the initial condition. Built at once, the neural network can be considered as a model of the system and can be used for simulation with different initial conditions. In the case of large time steps for map calculating, the proposed approach can significantly outperform traditional numerical methods. For Burgers’ equation, the computational performance is increased several times with the same level of accuracy. For some problems in the charged particle dynamics simulation, the performance is increased a thousand times with an appropriate accuracy in comparison with the traditional step-by-step integrating [@ref24; @ref25]. The proposed neural network can be used for data-driven identification of the systems. It may provide a high level of generalization when learning dynamical systems from data. As shown with the Van der Pol oscillator, learning the dynamics of the system with only a particular solution is possible. The neural network presented in Section \[sec4\] can be additionally fitted to satisfy the initial conditions. In this sense, the training will provide an optimal numerical approach with respect to certain initial conditions. The limitations of the data-driven approach for large-scale systems, optimal network configuration, and noisy data consideration should be examined in further research. [25]{} Lagaris, I.E., Likas, A., Fotiadis, D.I.: Artificial neural networks for solving ordinary and partial differential equations. Tech. rep. (1997), <https://arxiv.org/pdf/physics/9705023.pdf>, last accessed 2019/03/03. Baymani, M., Kerayechian, A., Effati, S.: Artificial Neural Networks Approach for Solving Stokes Problem. Applied Mathematics, 1, 288–292 (2010). Chiaramonte, M., Kiener, M.: Solving differential equations using neural networks (2013), <http://cs229.stanford.edu/proj2013/ChiaramonteKiener-SolvingDifferentialEquationsUsingNeuralNetworks.pdf>, last accessed 2019/03/03. Sharma, A.: NeuralNetDiffEq.jl: A Neural Network solver for ODEs, <https://julialang.org/blog/2017/10/gsoc-NeuralNetDiffEq>, last accessed 2019/03/03. Sirignano, J., Spiliopoulos, K.: DGM: A deep learning algorithm for solving partial differential equations. Journal of Computational Physics (2018). Weinan, E., Han, J., Jentzen, A.: Deep learning-based numerical methods for high-dimensional parabolic partial differential equations and backward stochastic differential equations. Tech. rep. (2017), <https://arxiv.org/pdf/1706.04702.pdf>, last accessed 2019/03/03. Anastassi, A.: Constructing Runge-Kutta Methods with the Use of Artificial Neural Networks. Tech. rep. (2013), <https://arxiv.org/pdf/1106.1194.pdf>, last accessed 2019/03/03. Wang, Y., Lin, C.: Runge-Kutta neural network for identification of dynamical systems in high accuracy. IEEE Transactions on Neural Networks, vol. 9, no. 2, 294–307 (1998). Chen, R., Rubanova, Y., Bettencourt, J., Duvenaud, D.: Neural ordinary differential equation, <https://arxiv.org/pdf/1806.07366.pdf>, last accessed 2019/03/03. Zjavka, L.: Differential polynomial neural network. Journal of Artificial Intelligence, 4 (1), 89–99 (2011). Yang, Y., Hou, M., Luo, J.: A novel improved extreme learning machine algorithm in solving ordinary differential equations by Legendre neural network methods. Advances in Difference Equations, 469 (2018). Schetinin, V.: Polynomial neural networkslearnt toclassify EEG signals. Tech. rep. (1997), <https://arxiv.org/ftp/cs/papers/0504/0504058.pdf>, last accessed 2019/03/03. Dragt, A.: Lie methods for nonlinear dynamics with applications to accelerator physics (2011), <http://inspirehep.net/record/955313/files/TOC28Nov2011.pdf>, last accessed 2019/03/03. Andrianov, S.: A role of symbolic computations in beam physics. Computer Algebra in Sc, Comp., Lecture Notes in Computer Science, 6244, 19–30 (2010). Andrianov, S.: Symbolic Computation of Approximate Symmetries for Ordinary Differential Equations. Mathematics and Computers in Simulation, vol. 57, N 3-5, 147–154 (2001). Andrianov, S.: A matrix representation of the Lie transformation. In: Proceedings of the Abstracts of the International Congress on Computer Systems and Applied Mathematics, 14 (1993). Andrianov S. The convergence and accuracy of the matrix formalism approximation. In: Proceedings of ICAP2012, Rostock, Germany, 93–95 (2012). Pan, S., Duraisamy, K. Long-time predictive modeling of nonlinear dynamical systems using neural networks. Hindawi Complexity, 4801012 (2018). Airbus Quantum Computing Challenge, <https://www.airbus.com/innovation/tech-challenges-and-competitions/airbus-quantum-computing-challenge.html>, last accessed 2019/03/03. Hayati, M., Karami, B.: Feedforward neural network for solving partial differential equations. Journal of Applied Sciences 7(19), 2812–2817 (2007). Casas. F.: Solution of linear partial differential equations by Lie algebraic methods. Journal of Computational and Applied Mathematics, 76, 159-170 (1996). Oliveri, F.: Lie symmetries of differential equations: direct and inverse problems. Note di Matematica, 23, n. 2, 195–216 (2004/2005). Evans, L.C.: Partial Differential Equations. Providence, R.I.: American Mathematical Society (2010). Senichev, Y., Lehrach, A., Maier, R., Zyuzin, D., Berz, M., Makino, K., Andrianov, S., Ivanov, A.: Storage ring EDM simulation: methods and results. In: Proceedings of ICAP2012, Rostock, Germany, 99–103 (2012). Senichev, Y., Ivanov, A., Lehrach, A., Maier, R., Zyuzin, D., Andrianov, S.: Spin tune parametric resonance investigation. In: Proceedings of the Particle Accelerator Conference, 3020–3022 (2014).

--- abstract: 'Dendrite growth is a long-standing challenge that has limited the applications of rechargeable lithium metal electrodes. Here, we have developed a grand potential based nonlinear phase-field model to study the electrodeposition of lithium as relevant for a lithium metal anode, using open-source software package MOOSE. The dynamic morphological evolution under large/small overpotential is studied in 2-dimensions, revealing important dendrite growth/stable deposition patterns. The corresponding temporal-spatial distributions of ion concentration, overpotential and driving force are studied, which demonstrate an intimate, dynamic competition between ion transport and electrochemical reactions, resulting in vastly different growth patterns. Given the importance of morphological evolution for lithium metal electrodes, wide-spread applications of phase-field models have been limited in part due to in-house or proprietary software. In order to spur growth of this field, we utilize an open-source software package and make all files available to enable future studies to study the many unsolved aspects related to morphology evolution for lithium metal electrodes.' author: - Zijian Hong - Venkatasubramanian Viswanathan bibliography: - 'refs.bib' title: 'Phase-field simulations of lithium dendrite growth with open-source software' --- {width="9cm" height="3.5cm"} There is a need for increasing the energy density of batteries for enabling a complete electrification of transportation.[@Sripad01012017; @doi:10.1021/acsenergylett.7b00432; @doi:10.1021/acsenergylett.7b01022] Lithium metal anodes represent one of the most promising near-term solutions enabling increased energy density. [@kerman2017practical] The main challenge with lithium metal anodes stems from dendrite growth during electrodeposition, which leads to deposits that could penetrate through the separator, resulting in safety issues and coloumbic inefficiency.[@AURBACH2002405; @doi:10.1021/acs.chemrev.7b00115; @C3EE40795K; @Lin2017] A comprehensive understanding of the morphology evolution during the electrodeposition process is of paramount importance for the design of next-generation batteries with Lithium metal anode. Most of the previous theoretical investigations of dendrite formation during electrodeposition is based on linear stability analysis, originally pioneered by Mullins and Sekerka [@MullinsSekerka63; @MullinsSekerka64], where the thermodynamic stability of a small perturbation is studied in terms of a change in the chemical potential[@monroe2003-dendrite; @Monroe2004Effect; @Monroe2005Impact; @Tikekar2016; @ahmad2017stability]. This simple approach has revealed the impact of some key physical parameters (elastic properties, molar volume, etc.[@Tikekar2016; @ahmad2017stability; @PhysRevMaterials.1.055403]). However, the linear stability analysis approach is limited to the understanding of dendrite initiation and does not take into account effects beyond nucleation, dimensionality of growth and dynamics of the growth process. In particular, the key features related to morphology evolution in three-dimensions are missed within this approach. Phase-field modeling is a mesoscale simulation tool that enable the quantitative understanding of phase transitions, phase transformations and microstructure evolutions. [@doi:10.1146/annurev.matsci.32.112001.132041; @doi:10.1146/annurev.matsci.32.101901.155803] To date, several phase-field models have been built to model the dynamic evolution of dendrite growth during the electrodeposition process, incorporating the nonlinear electrochemical reaction dynamics [@CHEN2015376; @PhysRevE.92.011301; @Liang2014; @Liang2012; @ELY2014581; @doi:10.1021/ar300145c]. However, these studies typically utilize custom-built software or proprietary code which could hinder the wide-spread applications of phase-field models, while also makes it extremely difficult for follow-on works and reproducible science.[@Kitchin:2015aa] Herein, we developed a hybrid grand potential based nonlinear phase field model [@PhysRevE.92.011301; @CHEN2015376; @PhysRevE.84.031601], and numerically solving the multiphysics coupled equations using the fully *open source* MOOSE (Multiphysics Object-Oriented Simulation Environment) framework [@GASTON20091768]. This enables a quantitative and comprehensive understanding of the morphological evolution during the electrodeposition process. The well tested and documented MOOSE package allows for fast, parallel and customized numerical modeling, which is easy to follow and simple to implement and validate. Most importantly, it has an already built-in, well-benchmarked phase-field module that could be utilized for further implementations [@TONKS201220; @YURKIV2018609]. In this work, the dynamic morphological evolution for lithium deposition is studied, where both stable deposition and dendrite growth are captured at low and high applied overpotentials, respectively. The corresponding spatial-temporal distributions of concentration, over-potential and driving force are studied in details, which reveals an intimate, dynamic competition between lithium-ion transport (from both ion diffusion and electrical migration) and electrochemical reaction, resulting in complete different growth patterns. A concept of “compositionally graded solid electrolyte” is proposed, which could provide an approach to suppress dendrite initiation based on the insights from this study. We believe that this study will spur on the application of phase-field models that can handle more complicated problems incorporates other effects (elastic effect, solid electrolyte interfaces, etc.). *Phase field model.* The phase field variable $\xi$ is used as the non-conserved order parameter which is defined such that $\xi=1$ and $\xi=0$ for the pure electrode and electrolyte phases, respectively. Following the derivations by Chen et al.[@CHEN2015376], the temporal evolution of the order parameter $\xi$ can be expressed as: $$\frac{\partial\xi}{\partial t}=-L_\sigma(g'(\xi)-k\nabla^2\xi)-L_\eta h'(\xi) \Bigg\{\exp\Big[\frac{(1-\alpha)nF\eta_{\alpha}}{RT}\Big]-\frac{c_{Li^+}}{c_0}\exp\Big[\frac{-\alpha nF\eta_{\alpha}}{RT}\Big]\Bigg\} .$$ where $L_\sigma$ and $L_\eta$ are the interfacial mobility and electrochemical reaction kinetic coefficient, respectively. $g(\xi)$ is the double well function, expressed by $W\xi^2(1-\xi)^2$. $W$ is related to the switching barrier [@CHEN2015376]. $k, \alpha, n, F, R, t$ and $T$ are the gradient coefficient, charge transfer coefficient, number of electrons transfered, Faraday constant, gas constant, evolution time and temperature, respectively. The switching barrier and gradient coefficient are further related to the surface tension $\gamma$ and interfacial thickness $\delta$, i.e. $W=\frac{3\gamma}{\delta}$ and $k=6\gamma \delta$ [@ELY2014581; @doi:10.1146/annurev.matsci.32.101901.155803]. Whereas $h(\xi)=\xi^3(6\xi^2-15\xi+10)$ is the interpolation function [@CHEN2015376]. $\eta_{\alpha}$ is the activation overpotential $\eta_{\alpha}=\phi-E^\theta$,$E^\theta$ is the standard equilibrium half cell potential, $\phi$ is the applied overpotential. $c_{Li^+}$ and $c_0$ are the local and initial lithium-ion molar ratio. In a two phase model, the local lithium ion molar-fraction can be written as [@PhysRevE.92.011301]: $$c_{Li^+}=c^l (1-h(\xi))=\frac{exp[\frac{(\mu-\epsilon^l)}{RT}]}{1+exp[\frac{(\mu-\epsilon^l)}{RT}]} (1-h(\xi)) .$$ where $c^l$ is the molar fraction of lithium in the electrolyte phase, $\mu$ is the chemical potential of lithium. $\epsilon^l=\mu^{0l}-\mu^{0N}$ is the differences in the chemical potential of lithium and neutral components on the electrolyte phase at initial equilibrium state. The chemical potential can be obtained by solving the modified diffusion equation [@PhysRevE.92.011301] (detailed derivation is given in Supplementary Information): $$\frac{\partial \mu}{\partial t}=\frac{1}{\chi}[\nabla.\frac{Dc_{Li^+}}{RT}(\nabla\mu+nF\nabla\phi)-\frac{\partial h(\xi)}{\partial t} (c^s \frac{C_m^s}{C_m^l}-c^l)]$$ where $C_m^s$ and $C_m^l$ are the site density of the electrode and electrolyte phases (inverse of molar volume), respectively. $c^s$ is the molar fraction of lithium in the electrode phase, while $\chi=\frac{\partial{c^l}}{\partial \mu}[1-h(\xi)]+\frac{\partial{c^s}}{\partial \mu}h(\xi)\frac{C_m^s}{C_m^l}$ . The spatial distribution of the electrical overpotential $\phi$ can be obtained by solving the conduction equation: $$\nabla\sigma\nabla\phi=nFC_m^s\frac{\partial\xi}{\partial t} .$$ The effective conductivity $\sigma$ is related to the conductivity of the electrode phase $\sigma^s$ and electrolyte phase $\sigma^l$ , i.e., $\sigma=\sigma^s h(\xi)+\sigma^l [1-h(\xi)]$ *Numerical simulations.* Equations 1-4 are solved using the open source MOOSE framework. A two-dimensional mesh of 200$\times$200 is used, with each grid representing 1 $\mu$m. The simulation temperature is 300 K. The Preconditioned Jacobian Free Newton Krylov (PJFNK) method is set as the solve type, with a bdf2 scheme and Single Matrix Preconditioning (SMP). A timestep interval (dt) of 0.02 s is set, with a maximum simulation time of 400 s. The key simulation parameters before and after normalization are listed in Supplementary table 1. The normalization constants for length, time and energy scales are set as 1 $\mu$m, 1 s and $2.5 \times 10^6 $J/m$^3$. The standard equilibrium half cell potential is set to be 0 for the sake of convenience. A Langevin noise level of 0.04 is added (corresponding to 0.056% change in magnitude) to the order parameter $\xi$ to stimulate the thermal fluctuation in the system. The Neumann boundary condition is applied at the $X$-dimension, such that the chemical potential is fixed to be zero at both sides (indicating a constant Lithium molar ratio at the two ends), while the order parameter is set as 1 (electrode phase) and 0 (electrolyte phase) at the left and right sides, respectively. The electric overpotential is fixed to the applied overpotential at the electrode side and zero at the electrolyte side to give a constant activation overpotential across the electrode/electrolyte interface. The no flux boundary condition is applied along the $Y$-dimension for all primary variables ($\xi, \mu$ and $\phi$). A set of applied electric overpotential ranging from -0.3 V to -0.5 V is studied. *Results.* The lithium growth kinetics under an applied overpotential of -0.45 V is shown in Figure \[fig:example1\], assuming a Li$^+$ concentration of 1 M in a standard carbonate-based electrolyte. The initial structure is shown in Figure \[fig:example1\](a), an electrode with a thickness of 20 $\mu$m is separated with the electrolyte by a smooth interface. No initial nuclei are added, the dendrite initiation and growth is a purely intrinsic process in this study. This differs from several previous studies where the simulation results might be sensitive to the preset artificial nuclei size, shape and orientation [@CHEN2015376; @YURKIV2018609; @ELY2014581]. After an electrodeposition process of 57 s (Figure \[fig:example1\]b), the interface moves towards the electrolyte by $\sim$20 $\mu$m. The interfacial instability is discovered, with a modulation of surface roughness for a vertical period of $\sim$10 $\mu$m. Further growth of these nuclei leads to the formation of long needle-like dendrites (Figure \[fig:example1\]c) with a length of $\sim$50 $\mu$m, while the diameter of the dendrites only increases by $\sim$10 $\mu$m, indicating a quasi-directional growth. A short video of the dendrite growth process under this condition is given in Supplementary video 1. It is worthwhile noting that although $\sim$20 nuclei is formed initially, only a few of them can grow into dendrites. ![Dendrite growth and concentration evolution with applied overpotential of -0.45 V. (a)-(c) Morphology evolution after Lithium electrodeposition of 0 s, 57 s and 108 s. (d)-(f) Corresponding evolution of Lithium-ion molar ratio.[]{data-label="fig:example1"}](figure1.png){width="1.0\linewidth"} The corresponding evolution of Li$^+$ ion molar fraction is shown in Figures \[fig:example1\](d)-(f). It is revealed that a broad interfacial region (shown in green) is formed to separate the electrode (with zero Li$^+$ molar ratio) and the electrolyte (with a homogeneous Li$^+$ molar ratio of $\sim$0.067). With the formation of small nuclei at the interface (Figure \[fig:example1\]e), a small fluctuation of ion concentration is also formed accordingly where the tip region has a slightly higher concentration than the valley region. The further growth of dendrites leads to a higher Li$^+$ concentration surrounding the dendrite tips, with a lower ion concentration formed in the valley regions. The enrichment of Li$^+$ molar ratio in the vicinity of the dendrite tips is due to the shortened diffusion path compared to the valley regions. In order to take a deeper understanding of the dendrite growth mechanism, the spatial-temporal distribution of electric potential and driving force is plotted in Figure \[fig:example2\]. As shown in Figure \[fig:example2\](a), the electric potential is nearly constant in the electrode region, followed by a gradual, linear increase from the electrode/electrolyte interface to the electrolyte region (see Figure S2d). After 57 s (Figure \[fig:example2\]b), as the interface is moving towards the electrolyte region, the gradient of electric potential (electric field) increases. The electric field is still more or less homogeneous inside the electrolyte region. With the growth of dendrite (Figure \[fig:example2\]c), a lower electric potential is found near the tip as compared to the surrounding valley regions, thus, we expect not only a large electric field along the $X$-dimension, but also an electric field towards the dendrite tip along the $Y$-dimension as well (shown in Figure S3). This could further facilitate Li$^+$ transport from both the source and the surrounding valley regions to the dendrite tips, leading to an “entrainment” phenomenon. The driving force (defined as $L_\eta h'(\xi) \Bigg\{\exp\Big[\frac{(1-\alpha)nF\eta_{\alpha}}{RT}\Big]-\frac{c_{Li^+}}{c_0}\exp\Big[\frac{-\alpha nF\eta_{\alpha}}{RT}\Big]\Bigg\}$) for the electrochemical reaction is directly related to the interface growth velocity $\frac{\partial\xi}{\partial t}$. As depicted in Figure 2(d), there’s a large driving force at the electrode/electrolyte interface, which can be understood since the electrochemical reaction only occurs at the interface where $h'(\xi)$ is non-zero. After 57 s (Figure 2e), with the continuous electrochemical deposition, the driving force at the interface increases, while forming a zig-zag like pattern, mimicking the modulation of the solid/liquid interface. With the growth of the dendrites after 108 s (Figure 2f),it is clearly shown that the driving force at the dendrite tips are orders of magnitude larger than the interface of the valley regions. Whereas the longer the dendrite, the larger the driving force, which further facilitates growth of the longer dendrites. Figure 2(g) and (h) show the line plots for the time evolution of Li-ion molar fraction along $X$-direction cutting through the valley and dendrite regions, respectively. It can be seen that up until 57 s where the dendrites are yet to grow, the concentration profile in the two regions are very similar, the interface region has a lower Li$^+$ ion molar fraction as compared to the bulk. As the dendrites grow longer (e.g., 96 s), we see an enrichment/depletion of ion concentration at the dendrite/valley regions, resulting in a widely broadened interface for the valley region which can hardly move thereafter. The line plots of the driving force for the two regions are shown in figure 2(i) and (j). Spikes are found at the interfaces in accordance to the concentration profiles. After 57 s, as the interface is moving towards the electrolyte, the electric field increases, which increases the driving force. Eventually, the growth of dendrites will lead to a huge increase in the driving force which can be partly attributed to the increase of ion concentration surrounding the dendrite tip; whereas the valley region gets an order of magnitude suppression in the driving force, resulting from the depletion of Li$^+$ ion. ![Spatial/temporal distribution of electric potential and driving force. (a)-(c) Evolution of electric potential at 0 s, 57 s and 108 s. (d)-(f) The corresponding spatial distribution of the driving force. (g)-(h) The line plots of Li-ion molar ratio as a function of time in the Valley and Dendrite regions. (i) and (j) The corresponding Line plots of the driving forces in the two regions.[]{data-label="fig:example2"}](figure21.png "fig:"){width="0.88\linewidth"} ![Spatial/temporal distribution of electric potential and driving force. (a)-(c) Evolution of electric potential at 0 s, 57 s and 108 s. (d)-(f) The corresponding spatial distribution of the driving force. (g)-(h) The line plots of Li-ion molar ratio as a function of time in the Valley and Dendrite regions. (i) and (j) The corresponding Line plots of the driving forces in the two regions.[]{data-label="fig:example2"}](figure22.png "fig:"){width="0.88\linewidth"} The morphology evolution at low applied overpotential, e.g., -0.32 V is further studied to make a detailed comparison (Figure \[fig:example3\]). A stable electrochemical deposition is observed with no obvious surface modulation (Figure 3a). Interestingly, near the electrode/electrolyte interface, an enrichment in the Li$^+$ ion is discovered (Figure 3b). This can be understood since the deposition velocity is greatly reduced (exponentially) at lower overpotential, while the ion transport only depends on the gradient of the potential distribution which roughly decreases linearly with decreasing applied overpotential. The temporal evolution of the maximum Li$^+$ concentration near the interface is plotted in Figure 3(c). An oscillation of the maximum Li$^+$ ion molar ratio is captured, indicating a dynamic competition between the ion transport and electrochemical reaction. The line plot of the average Li$^+$ concentration profile after passing the same total current is plotted for different overpotentials (Figure 3d), a gradual shift from surface concentration enrichment to surface concentration depletion is observed from lower to higher overpotentials (0.32 V-0.5 V). And the threshold overpotential for dendrite growth occurs exactly at the transition point from interface concentration enrichment to depletion ($\sim$0.37 V). ![Influence of the applied overpotential. (a) Stable Lithium deposition after 200 s with an applied overpotential of -0.32 V. (b) The corresponding spatial distribution of ion concentration. (c) The temporal evolution of the maximum Li-ion concentration. The dashed line shows the bulk concentration. (d) Average Li-ion concentration profile at different applied overpotential after passing the same total current. []{data-label="fig:example3"}](figure3.png){width="1.0\linewidth"} *Discussions.* The schematics of the two scenarios are given below for a discussion. At lower applied overpotential, initially, the reaction rate is much smaller than the ion transport rate (including diffusion and migration), resulting in an accumulation of ion concentration at the interface, which would in turn boost the reaction rate (since it is directly proportional to the ion concentration) while decreasing the transport rate (due to a large negative concentration gradient). With the increase of reaction rate and decrease of transport rate, the transport rate will be surpassed by the reaction rate , giving rise to a decrease in surface ion concentration. And as it happens, the decrease of surface ion concentration would lead to an increase in ionic transport and decrease of electrochemical reaction. Thus, an iterative feedback system is achieved to guarantee the stable electrodeposition. Importantly, in this case the maximum surface ion concentration is always higher than the concentration of the bulk electrolyte. While under higher overpotential, initially, the electrochemical reaction is faster than the ion transport, resulting in a depletion of ions at the interface. The formation of peaks (dendrites) due to the inhomogeneity of the system will boost the transport of Li$^+$, due to not only a larger concentration gradient, but also a higher migration rate from both the source and the surrounding valley regions. This results in a higher transport rate than the reaction rate, leading to a further increase in the surface ion concentration. The increase in surface ion concentration in return boosts the electrochemical reaction and interfacial growth velocity which further shortens the diffusion length. In other surrounding valley regions, the ions are “sucked” by the peak region, the ion concentration keeps dropping which hinders the electrochemical reaction. This effect is purely related to the dimensionality and cannot be captured within 1-dimensional models. The key feature that distinguishes the two scenarios is that the average ion concentration at the interface is always larger than the bulk electrolyte for the lower overpotential case, while vice versa for the higher overpotential. This indicates that even when a surface instability is formed in the first scenario, it will quickly be recovered due to a lower reaction rate with lower concentration at the tip of the surface modulation. Based on the above understanding, we proposed a new concept “Compositionally graded solid electrolyte” which could potentially suppress the dendrite initiation (given in Supplementary Figure S6). This electrolyte is designed such that a higher Li-ion concentration at the Li-metal anode side and a lower Li-ion concentration at the cathode side. This can be achieved by tuning the polymer/ceramics ratio, or by varying the vacancy/doping concentrations in the solid solutions. Under this design, the initiation of dendrite would be hindered at a reasonable current density since the tips (dendrite) would have a lower concentration than the valley regions, which would decrease the electrochemical reaction rate of the dendrite. In conclusion, we have developed a hybrid grand potential based nonlinear phase-field model for electrodeposition using fully *open source* MOOSE package. The dynamic morphological evolution during the lithium electrodeposition process under large and small overpotential is studied, revealing dendrite growth/stable deposition under the two conditions. The concentration, overpotential, driving force distribution is further studied for both cases, showing the critical role of the intimate competitions between ion transport and electrochemical reactions for the formation of dendrites. We further proposed a “compositionally graded solid electrolyte” that could potentially lead to the suppression of dendrites. We hope that open-sourcing phase-field models will benefit the community as a whole, and stimulate more theoretical/experimental follow-up studies. Z.H. and V.V. gratefully acknowledge support from the U.S. Department of Energy, Energy Efficiency and Renewable Energy Vehicle Technologies Office under Award No. DE-EE0007810. All input files for MOOSE simulations are made freely available. Contents in the supporting information: Derivation of the modified diffusion equation; Parameters used in the simulations; Nonlinear dendrite growth kinetics; Branching growth kinetics under high overpotential; Design of “compositionally graded solid electrolyte”; Supplementary videos for the morphology and concentration evolution under applied overpotential of -0.45 V and -0.32 V.

--- abstract: 'Four-fermion processes with a particle lost in the beam pipe are studied at to perform precision tests of the electroweak theory. Leading higher-order corrections to such processes are analyzed within the framework of the Structure Functions () approach. The energy scale entering the is determined by inspection of the soft and collinear limit of the $O(\alpha)$ radiative corrections to the four-fermion final states, paying particular attention to the process of single-$W$ production. Numerical predictions are shown in realistic situations for experiments and compared with existing results. A Monte Carlo event generator, including exact tree-level matrix elements, vacuum polarization, higher-order leading corrections and anomalous trilinear gauge couplings, is presented.' address: - | Dipartimento di Fisica Nucleare e Teorica, Università di Pavia\ and INFN, Sezione di Pavia, via A. Bassi 6, 27100, Pavia, Italy - | INFN, Sezione di Pavia and Dipartimento di Fisica Nucleare e Teorica,\ Università di Pavia, via A. Bassi 6, 27100, Pavia, Italy - | Dipartimento di Fisica, Università di Ferrara\ and INFN, Sezione di Ferrara, Ferrara, Italy author: - 'G.Montagna' - 'M.Moretti' - 'O.Nicrosini' - 'A.Pallavicini' - 'F.Piccinini' title: 'Higher-order QED corrections to single-$W$ production in electron-positron collisions' --- by 7 by 3 by 2 electron-positron collision, four fermions, single-$W$ process, radiative corrections, structure functions, Monte Carlo.\ [pacs]{}: 02.70.Lq,12.15.Lk,13.85.Hd \[intro\] Introduction ====================== Four-fermion final states are of special interest for the physics programme of and future high-energy electron-positron colliders, being entangled with electroweak gauge boson production and decay [@lepyr]. In particular, the process considered in the present paper, [*i.e.*]{} $$\label{wwe} e^+ e^- \rightarrow e^- (e^+) \bar \nu_e (\nu_e) \, q^\prime \bar q$$ is peculiar among all the possible four-fermion final states because the bulk of its cross section is due to two sub-processes, [*i.e.*]{} $W$-boson pair production and decay $$\label{wws} e^+ e^- \rightarrow W^* W^* \rightarrow 4 \ {\rm fermions}$$ and the radiation of an almost on shell $t$-channel photon from the electron (positron), with subsequent production of a $W$-boson and a neutrino $$\label{sw} e^+ e^- \rightarrow \gamma^* e^+ (e^-) \rightarrow W^* \bar \nu_e (e^-) \rightarrow 4 \ {\rm fermions}$$ Despite, strictly speaking, both sub-processes (\[wws\],\[sw\]) always occur simultaneously and are indistinguishable, channel (\[wws\]) dominates if the electron is emitted at large angle, whereas channel (\[sw\]) dominates if the electron is emitted in the very forward region, because of the presence of a quasi-real $t$-channel photon. In this paper the process (\[sw\]) will be addressed, by restricting the analysis to the kinematical range of forwardly emitted electrons. This signal is usually referred to as single-$W$ production, since only the two final jets are detected [@tanaka]. The importance of this process has been emphasized since long time. In the energy range it is fundamental in order to study the self-interaction of the gauge bosons, together with the process (\[wws\]), whereas in the energy regime of future colliders at the TeV scale it becomes the dominant electroweak process. In refs. [@realw; @wwrealw] cross sections and distributions were calculated in the approximation of real $W$-boson production, either by [@realw] studying the reaction $e^+e^-\rightarrow e^- \bar\nu_e W^+$, or by [@wwrealw; @trig] employing the Weizsäcker-Williams [@ww] equivalent-photon approximation for the $t$-channel photon. In refs. [@realw; @wwrealw; @trig] it was pointed out the relevance of this process for the study of trilinear gauge boson couplings and some assessment of the sensitivity has been given. The first full four-fermion calculation, including the crucial effect of fermion masses, has been presented in ref. [@grace], where the sensitivity to anomalous gauge couplings has been studied. Since then, other complete four-fermion calculations of the single-$W$ process have appeared in the literature and implemented in computational tools for data analysis [@grace; @ifl; @passarino; @dub; @bpp]. In most of these calculations the effect of fermion masses is exactly accounted for in the dynamics and kinematics for the whole four-fermion phase-space [@grace; @ifl; @dub; @bpp], while in the approach of ref. [@passarino] the Weizsäcker-Williams approximation is employed in the very forward, collinear region and matched with a massless four-fermion computation outside it. An up-to-date inventory of the present theoretical status is under preparation by the four-fermion working group of the workshop at  [@proposal]. Measurements of the single-$W$ cross section and the corresponding bounds on anomalous gauge couplings have been recently reported by the collaborations [@tanaka]. Because the foreseen accuracy of final data is of the order of $1$-$2\%$ [@proposal], accurate theoretical predictions for cross section and distributions are required. The calculation of the cross section for single-$W$ processes poses several non-trivial theoretical problems [@proposal]. For a realistic account of gauge bosons properties it is mandatory to include the gauge boson width in the propagator. In general this mixes a fixed order calculation with an all order resummation of a class of Feynman diagrams and introduces a violation of the Ward identities of the theory. This issue is of special importance here since, due to the $t$-channel photon exchange, even a tiny violation of Ward identities is enhanced by a factor of $s/m_e^2$. This is indeed the case if a running width is used in the calculation. This problem has been extensively studied [@fudge; @fwidth; @floop], and several options to address it have been explored. The most theoretically appealing procedure is the fermion loop scheme [@fwidth; @floop], which preserves both $U(1)$ and $SU(2)$ Ward identities. Recently, this scheme has been generalized to the case of massive external fermions, both in its minimal version, which considers the imaginary parts of the fermionic loops () [@ifl], and in its full realization with real and imaginary parts [@rfl0; @rfl1]. In particular, in ref. [@ifl] a detailed numerical investigation has been performed, showing no significant difference between the and the fixed width scheme, even in the region most sensible to $U(1)$ gauge invariance. For this reason, the fixed width scheme is adopted in the present calculation. Another delicate issue is the so called resolved-photon component of the cross section. The quasi-real $t$-channel photon can split into a pair of almost massless quarks, leading to a situation where the partonic picture of hadrons is inadequate and both perturbative and non-perturbative corrections become relevant. This issue is widely discussed in the literature [@stein], where the standard approach to this problem is also described and to which the reader is referred for details. However, for single-$W$ like events the resolved-photon component does not constitute a severe limitation: once a hard $q{\bar q}$ invariant mass cut is imposed, as done in realistic data analysis, the bulk of the signal is kept, whereas the resolved photon contribution becomes almost negligible [@passarino]. A further relevant issue is given by radiative corrections due to photon radiation. Because exact $O(\alpha)$ electroweak corrections to single-$W$ production are still unknown, in most of the theoretical and experimental studies presented insofar only the large contribution of initial-state radiation () has been taken into account, generally by using collinear structure function (hereafter denoted as ) and assuming $s = 4E^2$ as the proper scale for radiation. Due to the dominance of the quasi-real $t$-channel photon exchange, this can be expected not to be a suitable choice in the present case. On the other hand, it has been recently proposed to correct only the $s$-channel contributions to the single-$W$ signature, fixing the radiation scale in the usual manner, and to neglect the photonic corrections to the $t$-channel contributions [@dub]. Following previous investigations [@bhabha; @sisr; @babayaga; @bdk; @bbv] of the pattern of photonic radiation in and electroweak processes, some theoretical arguments to determine the appropriate energy scale entering the are presented and compared with existing results. The analysis here described elucidates the theoretical details and provides further numerical results of a contribution by the authors [@apt] to the activity of the four-fermion working group of the workshop at  [@proposal]. Ideas similar to those adopted in the present work have recently appeared in ref. [@kuriharat] and there applied to the two-photon process $e^+ e^- \to e^+ e^- \mu^+ \mu^-$. The paper is organized as follows. After a short review of the approach to leading log () corrections in Sect. \[brems\], Sect. \[analytic\] collects the analytical results valid for soft and collinear corrections to a generic scattering process. By comparing the results of Sect. \[brems\] and Sect. \[analytic\], the radiation scales for single-$W$ production are determined in Sect. \[scale\]. Sect. \[alphar\] deals with the problem of taking into account the effect of the photon vacuum polarization in the single-$W$ process, while Sect. \[numeric\] shows the numerical results of the present study obtained with a Monte Carlo (hereafter ) code for the single-$W$ signature, including also the effect of anomalous trilinear gauge couplings. Conclusions and prospects are given in Sect. \[ending\]. \[brems\] Structure Function approach to photon radiation ========================================================= Since in high-energy processes the corrections due to the emission of soft and collinear radiation are quite large, the contribution must be calculated at every perturbative order. A common technique to achieve this goal is the structure function approach [@sf], which consists in convoluting the hard-scattering cross section with appropriate “parton” densities. As well known, these convolution factors, [*i.e.*]{} the structure functions, include, by construction, both the real and virtual part of the photon correction, in order to ensure the cancellation of the infrared singularities. If a generic Born-level prediction $d\sigma_0$ is considered, the cross section $d\sigma$ including radiative corrections is obtained, by virtue of factorization theorems, according to the following general formula [@sf] $$\label{eq:sf} d \sigma = \prod_i \int dx_i D(\Lambda^2,x_i) \; d \sigma_0$$ where $1-x_i$ are the energy fractions carried away by the radiated photons from the $i$-th leg, $\Lambda$ is the characteristic scale of the $D(\Lambda^2,x_i)$, whose evolution is driven by the Dokshitzer-Gribov-Lipatov-Altarelli-Parisi (DGLAP) equation [@dglap] and is dependent on $\Lambda$. It is worth noticing that the choice of the scale $\Lambda$ is not dictated by general arguments and it is therefore rather arbitrary. Equation (\[eq:sf\]) can be rewritten by stressing the possibility of different scales for each as follows $$\label{eq:sfgen} d \sigma = \prod_i \int dx_i D(\Lambda_i^2,x_i) \; d \sigma_0$$ In particular, if the integrated hard-scattering cross section is a smooth function of the centre of mass (c.m.) energy, once the integrations over the energy fractions $x_i$ are performed in the soft-photon approximation, the $O(\alpha)$ double-log expansion of eq. (\[eq:sfgen\]) can be written as follows $$\label{eq:lle} d \sigma = d \sigma_0 \left ( 1 + \sum_i \frac{\alpha}{\pi} \log \frac{\Delta E}{E} L(\Lambda_i^2) \right )$$ where $\Delta E/ E$ is the maximum energy for undetected photons, to be identified with finite energy resolution of the photon detector, and $L (\Lambda_i^2) \equiv \log(\Lambda_i^2/m^2)$ is the collinear logarithm. Since the functional form of the is accurately known [@sf], the main problem in evaluating eq. (\[eq:sfgen\]) is to fix the process scales $\Lambda_i$. A generally adopted attitude is given by the comparison of eq. (\[eq:lle\]) with a perturbative calculation, which can be performed within any approximation, provided it reproduces the correct double-log contribution of the $O(\alpha)$ correction. This issue is addressed in the next Section. \[analytic\] Analytical results =============================== The double-log contribution to photon radiation traces back to soft and collinear bremsstrahlung and its virtual counterpart [@yfs], and, in the case of a calorimetric measurement of the energy of the final-state () particles, to hard radiation collinear to the particles themselves [@calo; @gatto; @jj; @cacci]. At each perturbative order, the leading contribution can be expanded in terms of infrared and collinear logs. For example, when the photons are emitted from the initial-state () particles only, such an expansion can be arranged in terms of double-log contributions of the form $\alpha^n l^n L_{s}^n$, where $l \equiv \log (\Delta E/E)$ is the infrared log and $L_{s} \equiv \log (s/m^2)$ is the collinear log. This is the reason why $\Lambda^2 = s$ is the “natural” energy scale to be used for in the presence of only. When also radiation is considered, the collinear log, $L$, is in general modified by additional factors coming from the angular integration over the photon variables. A typical example is given by the radiation emitted from one leg in the $t$-channel contribution to Bhabha scattering [@bhabha]. In the soft-photon approximation the radiation cones, one from the electron and one from the electron, have a half-opening which is determined by the angle between the emitting particles, because of a destructive interference. As a consequence, the energy scale $s$, which appears in $L_{s}$, transforms into $|t| \simeq s(1-\cos\theta)$, where $\theta$ is the electron scattering angle, and, therefore, $L_{s} \to L_{t}$. Hence the perturbative expansion contains collinear logs which are modified because of the angular ordering introduced by the radiation cones. In the presence of large scattering angles, for which $|t| \simeq s$, the above modification is numerically small, but it becomes more and more important in the forward angular range, which is the dynamically favourite region by $t$-channel Bhabha scattering and where $t \ll s$. The net result is a numerically significant depletion of radiation effects just in the most important part of the hard-scattering $t$-channel dynamics. Actually, when using to evaluate corrections to small-angle Bhabha scattering, the energy scale $\Lambda^2 = |t|$ is employed in all phenomenological applications [@bhabha]. More in general, in order to take into account dominant initial-final-state interference effects in addition to initial- and final-state leading terms, $s$ and $t$ contributions to Bhabha scattering can be corrected in terms of a unique combination of Mandelstam invariants given by $s t / u$, as discussed in refs. [@sisr; @babayaga]. Therefore, the energy scale $\Lambda^2 = s t/u$ turns out to be a suitable choice for the evaluation in terms of of LL corrections to Bhabha scattering, as demonstrated, in comparison with the exact $O(\alpha)$ calculation, in ref. [@babayaga]. Similar arguments for an appropriate choice of the energy scale for QED radiation, based on the inspection of the soft and collinear limit of the $O(\alpha)$ correction, have been also advocated in ref. [@bdk] for the reaction $e^+ e^- \to W^+ W^-$ and, very recently, in ref. [@kuriharat] for the process $e^+ e^- \to e^+ e^- \mu^+ \mu^-$. Comparisons performed in refs. [@bdk; @kuriharat] with available exact $O(\alpha)$ calculations explicitly exhibit the validity of such a strategy, which is therefore pursued in the present analysis. As already remarked, the result for corrections in the presence of a calorimetric detection of particles must include the contribution of photons which, regardless of their energy, can not be discriminated from closely collinear fermions, as a consequence of the finite angular resolution of the calorimeters. The role of such hard photons collinear to the particles becomes therefore unavoidable in the case of a calorimetric measurement of the energy of the particles, as discussed in the following. \[soft\] Soft-photon contribution --------------------------------- In this Section the contribution of photons, too soft to be detected in the calorimeter, will be computed for a generic process with $n$ ingoing-legs. [^1] The following approximations are understood $$\label{eq:spa} \left \{ \begin{array}{ll} q_i \gg k \\ s_{ij} \gg m_i^2,\,m_j^2 \\ \end{array} \right .$$ where $q_i$ is the momentum of a particle of mass $m_i$ emitting a real photon of momentum $k$, and $s_{ij}\equiv(q_i+q_j)^2$ is the invariant mass of the pair $ij$. By following the standard derivation of the eikonal factors due to soft bremsstrahlung and by generalizing it to particles with different masses and charges, the differential cross section, dressed by soft-photon emission, can be cast into the following factorized form [@yfs] $$\label{eq:soft} d\sigma_{\rm soft} = d\sigma_0 \frac{d\omega}{\omega} \frac{2\alpha}{\pi} \sum_{i>j}^n Q_iQ_j \log \frac{s_{ij}}{m_im_j}$$ where $\omega$ is the photon energy and $Q_i$ is the charge of the $i$-th particle. It is worth noticing that in the limit $s_{ij} \ll m_i^2,\,m_j^2$, provided the first inequality in eq. (\[eq:spa\]) still holds, the logarithmic behaviour present in eq. (\[eq:soft\]) disappears, leaving a power law which can be simply obtained by means of the substitution [@yfs] $$\label{eq:pancheri} \log \frac{s_{ij}}{m_im_j} \longrightarrow \frac{1}{3} \frac{s_{ij}}{m_im_j}$$ Notice that, since the goal is to determine the scale entering the , only the contribution of real photons is explicitly calculated, because the virtual corrections, in order to preserve the cancellations of infrared singularities, must share the same leading collinear structure of the real part itself. By including the virtual part needed to cancel the infrared singularity and integrating eq. (\[eq:soft\]) over the photon energy $\omega$ in the soft region $0 \leq \omega \leq \Delta E$, one gets $$\label{eq:sv} d\sigma_{\rm S+V} = d\sigma_0 \log \frac{\Delta E}{E} \frac{2\alpha}{\pi} \sum_{i>j}^n Q_iQ_j \log \frac{s_{ij}}{m_im_j}$$ \[hard\] Hard radiation collinear to the final-state particles -------------------------------------------------------------- In the case of a calorimetric set-up, which is the realistic situation for single-$W$ production at , photons collinear to the detected particles can not be distinguished from the the emitting particles themselves, independently of the photon energy. Therefore, in order to obtain the correct structure of double-log corrections for such an event selection, the effect due to the emission of unresolved hard radiation collinear to the particles must be taken into account in addition to soft+virtual corrections. To this end, let us re-consider the previous process with $n$ ingoing legs and think $m$ of them to be changed into outgoing legs at the end of the calculation (see the previous footnote). Then, the contribution of photons collinear to the particles can be cast into a gauge invariant form as follows [@gatto; @jj; @cacci] $$\label{eq:hard} d\sigma_{\rm hard} = d\sigma_0 \frac{d\omega}{\omega} \frac{2\alpha}{\pi} \sum_i^m Q_i^2 \log \frac{E_i\delta}{m_i}$$ where $E_i$ is the energy of the $i$-th particle, and $\delta$ is the half-opening angle of the calorimetric resolution. By integrating eq. (\[eq:hard\]) over the photon energy $\omega$ in the range $\Delta E \leq \omega \leq E$, the integrated correction due to hard photons collinear to the particles is given by $$\label{eq:hardi} d\sigma_{\rm hard} = d\sigma_0 \log \frac{E}{\Delta E} \frac{2\alpha}{\pi} \sum_i^m Q_i^2 \log \frac{E_i\delta}{m_i}$$ \[master\] The master formula ----------------------------- Equations (\[eq:sv\]) and (\[eq:hardi\]) give the leading double-log contribution which must be compared to (\[eq:lle\]), the $O(\alpha)$ perturbative expansion of eq. (\[eq:sfgen\]), in order to fix the process scales $\Lambda_i$. Summing the contributions of eq. (\[eq:sv\]) and eq. (\[eq:hardi\]), the analytical cross section is in conclusion given by $$\begin{aligned} \label{eq:master} d\sigma_{\rm S+V} + d\sigma_{\rm hard} & = & d\sigma_0 \frac{2\alpha}{\pi} \log \frac{\Delta E}{E} \left \{ \sum_{i=m+1}^n Q_i^2 \log \frac{E_i}{m_i} \; + \right . \\ \nonumber & & - \; \left . \sum_{i>j}^n Q_iQ_j \log 2(1-c_{ij}) - \sum_i^m Q_i^2 \log \delta \right \}\end{aligned}$$ where $c_{ij}$ is the cosine of the angle between particles $i$ and $j$. Three different kinds of logarithms occur in eq. (\[eq:master\]). The first term contains the mass and energy logarithms of the particles only, since, as expected, the energies and the masses of the particles disappeared, in agreement with the [kln]{} theorem [@kln]. The second term includes angular terms due to radiation interference, while the third one comes from the requirement of calorimetric measurement. These terms must be compared with the collinear logarithms of eq. (\[eq:lle\]) in order to fix the scales $\Lambda_i$ of the . In the following Section this task is accomplished in detail for the single-$W$ process. \[scale\] Fixing the radiation scales in the single-$W$ process =============================================================== Let us consider, for definiteness, the process $e^+e^- \rightarrow e^-{\bar\nu}u{\bar d}$ with the electron lost in the beam pipe (single-$W$ process). In this event selection (hereafter ) the leading contribution comes from $\gamma^* e^+$ scattering with the virtual photon emitted from the electron line. The leading dynamics is given by the $t$-channel Feynman diagrams shown in Fig. \[fusion\]. ![\[fusion\] The fusion and bremsstrahlung diagrams are the leading Feynman graphs for the single-$W$ signature.](fig1.ps "fig:") ![\[fusion\] The fusion and bremsstrahlung diagrams are the leading Feynman graphs for the single-$W$ signature.](fig2.ps "fig:") If a calorimetric measurement of the energies of the particles is performed, only the legs need to be corrected by the . Furthermore, since the electron is scattered in the very forward region, the interference between the electron line and the rest of the process is very small. This allows a natural sharing of the logarithms coming from eq. (\[eq:master\]) between the two associated to the colliding electron and positron. Hence the formula (\[eq:master\]), when compared with eq. (\[eq:lle\]), translates into the two following scales ($\Lambda_-$ refers to attached to the electron line, while $\Lambda_+$ to the attached to the positron line), $$\label{eq:wscales} \Lambda_-^2 = 4E^2 \frac{(1-c_-)^2}{\delta^2} \;,\quad \Lambda_+^2 = 2^{\frac{14}{9}}E^2 \frac{\big((1-c_{\bar d})(1-c_u)^2\big)^\frac{2}{3}} {\big((1-c_{u{\bar d}})^2\delta^5\big)^\frac{2}{9}}$$ where $E$ is the beam energy, $c_-$ is the cosine of the electron scattering angle, $c_u$ and $c_{\bar d}$ are the cosine of the quark scattering angles with respect to the initial positron, and $c_{u{\bar d}}$ is the cosine of the relative angle between the quarks. It is worth noticing that in the numerical implementation, whenever one of the two scales is less than a small cut-off ($\Lambda^2_{\rm cut-off} = 4 m_e^2$, where $m_e$ is the electron mass), the radiation from the corresponding leg is switched off, in accordance with eq. (\[eq:pancheri\]). It was carefully tested that variations of the cut-off do not alter the numerical results. Owing to the presence of a resonant $W$ boson, some modifications to the previous results may come from finite-width effects and from radiation decoherence [@mg]. Finite-width corrections of the form of $E_\gamma/\Gamma_W$ arise when the unstable particle propagator goes off its mass-shell, but this is not the present case, since the multi-fermion final state can accommodate a resonant $W$. Radiation decoherence is present whenever a resonance occurs and its effect is to cancel the angular dependence from the scale. As a consequence the scale $\Lambda_+$ should be modified by dropping the angular interference factors in eq. (\[eq:wscales\]) when the emitted photons have $E_\gamma \sim \Gamma_W$. Yet in the present case the effect is tiny, since the effects due to angular interference for the scale $\Lambda_+$ are already small by themselves. It is also possible to make a naive ansatz for the radiation scales without a detailed calculation, by thinking of the graphs of Fig. \[fusion\] in terms of the Weizsäcker-Williams approximation [@ww], [*i.e.*]{} in terms of a convolution of the process $e^+\gamma \rightarrow \nu_e W^*$ with an equivalent photon spectrum plus a real electron line. This leads to assigning two different scales to the single-$W$ process: one scale for the electron current and one for the positron current. The former scale is the proper one for a $t$-channel process, e.g. $t$-channel Bhabha scattering, so it is simply $|q^2_{\gamma^*}|$, where $|q^2_{\gamma^*}|$ is the squared momentum transfer in the $ee\gamma^*$ vertex. The latter is the sum of an $s$-channel electron exchange and a $t$-channel $W$ exchange (see Fig. \[fusion\]). Assuming that the $t$-channel dominates, its natural cut-off is given by the $W$-boson mass, $M_W$. Hence, the following ansatz follows $$\label{eq:naive} \Lambda_{-,{\rm naive}}^2 = |q^2_{\gamma^*}| \;,\quad \Lambda_{+,{\rm naive}}^2 = M_W^2$$ where $M_W$ is the mass of the $W$ boson. The comparison between the scales given by eq. (\[eq:wscales\]) and these naive scales, which will be performed numerically in the following Section, provides a useful cross-check of the analytical results derived by inspection with the soft/collinear limit of the $O(\alpha)$ correction. A discussion of other possible approaches to the treatment of photonic corrections to single-$W$ production can be found in the four-fermion working group report of the Workshop [@proposal]. \[alphar\] The running of the electromagnetic coupling constant =============================================================== Besides the higher-order corrections discussed in the previous Sections, other large logarithmic contributions to the single-$W$ cross section arise from the running of the electromagnetic coupling constant $\alpha$. Since in the case under study the dominant configurations come from the Feynman diagrams with an almost on-shell photon exchange, the appropriate scale for the evaluation of the electromagnetic coupling relative to the $t$-channel photon in the $ee\gamma^*$ vertex is the squared momentum transfer $q^2_{\gamma^*}$ defined above. However, because $G_F$, $M_W$ and $M_Z$ are the typically adopted input parameters for electroweak processes at , the electromagnetic coupling is fixed at tree-level to a high energy value as, for example, $$\alpha_{G_F} = 4 \sqrt{2} {{G_F M_W^2 s_W^2} \over {4 \pi}}\;, \quad \quad {\rm with} \quad \quad s_W^2 = 1 - {{M_W^2}\over {M_Z^2}} \;.$$ On the other hand, the single-$W$ production is a $q^2_{\gamma^*} \simeq 0$ dominated process and therefore the above high-energy evaluation of $\alpha$, $\alpha_{G_F}$, needs to be rescaled to its correct value at small momentum transfer. To this end, a gauge-invariant “reweighting” procedure can be adopted, by rescaling the differential cross section $d\sigma/dt$ ($t \equiv q^2_{\gamma^*}$) in the following way $$\label{eq:running} {{d\sigma}\over{dt}} \rightarrow {{\alpha^2(t)}\over{\alpha^2_{G_F}}} {{d\sigma}\over{dt}} \;,$$ where $\alpha(t)$ is the running coupling constant computed at virtuality $q^2_{\gamma^*}$. A detailed analysis of the effect of the running couplings in single-$W$ production has been recently performed within the massive fermion-loop scheme in ref. [@rfl1], where the couplings are automatically running in the calculation. As shown in ref. [@rfl1], the relative difference between the above reweighting prescription and the complete results of the fermion-loop scheme is at the 1%-2% level [^2], and it is therefore in the expected range of theoretical uncertainty due to missing full one-loop electroweak corrections. \[numeric\] Numerical results ============================= In this Section the code, developed to simulate the single-$W$ process, is described and a sample of numerical results obtained by means of it is shown and commented, with particular emphasis on the effects of higher-order corrections to single-$W$ production at energies. \[mc\] The Monte Carlo code --------------------------- A program, named [SWAP]{}, was developed to calculate cross sections and differential distributions for the single-$W$ signature. As already emphasized, the main feature of this process is the fact that the $t$-channel photon of Fig. \[fusion\] becomes quasi-real. In the limit of massless fermions, the photon propagator becomes singular in the forward direction and the cross section develops a logarithmic singularity. Indeed, whenever the final electron is lost in the beam pipe, its mass becomes a natural cutoff for the very-forward singularities, compelling to build a massive matrix element and phase-space. The phase-space integration is performed in [SWAP]{} with the aid of a multi-channel importance sampling with stratification. The main peaking structures for the single-$W$ process are given by the dynamics depicted by the fusion and bremsstrahlung graphs of Fig. \[fusion\]. They are the resonant $W$-boson invariant mass, treated with a Breit-Wigner weight, and the $t$-channel “singularity” of the quasi-real photon, treated with a $1/|t|$ weight. Moreover, the program can deal with the singularities of the sub-leading $t$-channel CC$20$ diagrams shown in Fig. \[other\], by means of the multi-channel approach. ![\[other\] The multiperipheral diagram is the main sub-leading Feynman graph for the single-$W$ signature.](fig3.ps) The exact hard-scattering matrix element is computed by means of the code [@alpha] for the automatic evaluation of Born scattering amplitudes. Fermion masses are exactly taken into account and the fixed-width scheme is adopted as gauge-restoring approach, by taking the massive gauge boson propagator as follows: $$\Pi^{\mu \nu} = {-i\left( g^{\mu\nu}-{k^\mu k^\nu\over M^2 - i\Gamma M}\right)\over k^2 - M^2 + i\Gamma M} \,,\quad\Gamma = {\rm ~cost.}$$ It is known [@ifl; @fwidth; @floop] that this scheme preserves $U(1)$ gauge invariance but still violates $SU(2)$ Ward identities. However, at least in the unitary gauge employed here, it is indistinguishable from other fully gauge-invariant schemes [@ifl; @fwidth; @floop]. The contribution of anomalous gauge couplings is also accounted for in [SWAP]{}. The anomalous gauge boson couplings $\Delta k_{\gamma}$, $\lambda_{\gamma}$, $\delta_Z$, $\Delta k_Z$ and $\lambda_{\gamma}$ are implemented in the [ALPHA]{} code according to the parameterization of refs. [@gg; @hhpz]. Photon radiation is implemented via formalism, according to the discussion of Sect. \[scale\]. It is worth noticing that, since the incoming electron/positron are required to be on-shell massive fermions, a naive four-momentum rescaling, due to photon emission, such as $\hat{p}_{\pm} = x \, p_{\pm}$ leads to potentially dangerous gauge violations, according to what previously discussed. Therefore, the rescaled incoming four-momenta are implemented as $\hat{p}_{\pm} = (x \, E, 0, 0 , \pm \sqrt{x^2 E^2 - m_e^2})$, by interpreting $x$ as the energy fraction after photon radiation, as motivated in ref. [@babayaga]. If required, $p_\perp / p_L$ effects can be provided in the treatment of ISR, by means of either $p_\perp$-dependent SF [@nunugpv] or a QED Parton Shower algorithm [@babayaga; @jap]. The effect of vacuum polarization is taken into account as described by eq. (\[eq:running\]), by including the contribution of leptons, heavy quarks and light quarks, the latter according to the parameterization of ref. [@vpol]. The program supports realistic and it can be employed either as a cross-section calculator or as a event generator, with both weighted and unweighted generation available. The technical precision of the event generator [SWAP]{} has already been carefully proved in ref. [@proposal], by means of detailed tuned comparisons between the predictions of independent codes. Perfect agreement was found, both at the level of integrated cross sections and distributions, also for purely leptonic final states. \[numdisc\] Discussion of the numerical results ----------------------------------------------- --------------------------------- -------------------------- ----------------------------------------- electron angular acceptance $|\cos\theta_e| > 0.997$ $|\cos\theta_e| > 0.997$ quarks angular acceptance 1\. no cut 2\. $|\cos\theta_{q, {\bar q}}| < 0.95$ calorimetric half-opening angle $5.00^\circ$ $5.00^\circ$ quark-antiquark invariant mass $45.0~{\rm GeV}$ $45.0~{\rm GeV}$ --------------------------------- -------------------------- ----------------------------------------- : \[calo\] The adopted for the calculations shown in the present paper for the signature $e^+ e^- \rightarrow e^- \bar \nu_e u \bar d$, according to ref. [@proposal]. The numerical simulations are elaborated according to the reviewed in Tab. \[calo\], with the electroweak input parameters shown in Tab. \[scheme\]. ----------------------------------------------- $G_F = 1.16637 \times 10^{-5}~{\rm GeV}^{-2}$ $M_Z = 91.1867~{\rm GeV}$ $M_W = 80.35~{\rm GeV}$ $\Gamma_Z = 2.49471~{\rm GeV}$ ----------------------------------------------- : \[scheme\] The adopted electroweak input parameters, according to ref. [@proposal]. All other parameters are calculated by means of the tree-level relations. In Figs. \[varying\]-\[varying1\] the numerical impact of different choices of the $\Lambda^2$-scale on the cross section of the single-$W$ process $e^+e^-\rightarrow e^-\bar{\nu}u\bar{d}$ in the energy range is shown. Since the energy scale $\Lambda_+$ of eq. (\[eq:wscales\]) depends on the quark scattering angles, two different quark angular acceptances are considered, namely no cut (Fig. \[varying\]) and $|\cos\vartheta_{u, {\bar d}}| < 0.95$ (Fig. \[varying1\]). The marker $\bullet$ represents the Born cross section, $\scriptstyle\bigcirc$ represents the correction given by $\Lambda^2_{\pm} = s$ scale for both IS , $\scriptstyle\diamondsuit$ represents the correction given by $\Lambda^2_{\pm} = |q^2_{\gamma^*}|$ scale for both IS , $\scriptstyle\triangle$ the correction given by the scales of eq. (\[eq:wscales\]), $\scriptstyle\square$ the correction given by the naive scales of eq. (\[eq:naive\]). It can be seen that neither the $s$ scale, as implemented in some computational tools used for the analysis of the single-$W$ process, nor the $|q^2_{\gamma^*}|$ scale are able to reproduce the effects due to the scales of eq. (\[eq:wscales\]) and eq. (\[eq:naive\]). These two scales are in good agreement and both predict a lowering of the Born cross section of about 8-9%, almost independent of the c.m. energy and quark angular acceptance. This fact can be understood by looking at Fig. \[scales\], where it is shown the single-$W$ differential cross section with respect to the scales $\Lambda_{\pm}$ of eq. (\[eq:wscales\]). On the left, $\Lambda_{+}$ exhibits a broad peak not far from $M_W$, while, on the right, the other scale $\Lambda_{-}$ peaks, as expected, at very small momentum transfer. ![\[scales\] The differential cross sections of the single-$W$ process $e^+e^-\rightarrow e^-\bar{\nu}u\bar{d}$ with respect to the two scales of eq. (\[eq:wscales\]) at $\sqrt{s} = 189$ GeV.](fig6.ps) Figure \[figal\] shows the effects of the reweighting procedure of eq. (\[eq:running\]) for the evaluation of the running coupling constant. The marker $\scriptstyle\triangle$ represents the relative difference between the integrated cross section computed in terms of $\alpha_{G_F}$ and the cross section computed in terms of $\alpha(0)$, while the marker $\scriptstyle\diamondsuit$ is the relative difference between the integrated cross section computed in terms of $\alpha_{G_F}$ and the cross section computed in terms of $\alpha(t)$. As can be seen, the rescaling from $\alpha_{G_F}$ to $\alpha(t)$ introduces a negative correction of about 5-6% in the energy range. The difference between $\scriptstyle\triangle$ and $\scriptstyle\diamondsuit$, which is about 2-3%, is a measure of the running of $\alpha_{QED}$ from $q^2_{\gamma^*} = 0$ to $q^2_{\gamma^*} = t$. As an illustrative example of the effect of anomalous couplings on single-$W$ differential distributions, in Fig.\[mqq\] the distribution of the $q{\bar q}$ invariant mass, around the peak of the $W$-boson resonance, and the distribution of the angle of the quarks with the line of flight of the reconstructed $W$-boson in the $W$-boson rest-frame are shown. The dashed lines correspond to the simulation as obtained by means of [SWAP]{} for the anomalous coupling $\Delta\kappa_\gamma = 0.1$, while the solid lines represent the Standard Model prediction. The effect of the anomalous coupling $\Delta\kappa_\gamma$ at energies is just an overall rescaling of the total cross section. Therefore the sensitivity to $\Delta\kappa_\gamma$ in single-$W$ events depends crucially on the accuracy of the theoretical evaluation of the total cross section. ![\[mqq\] The single-$W$ differential cross sections with respect to the quark-antiquark invariant mass (left side), and with respect to the angle between a quark and the line of flight of the reconstructed $W$-boson in the frame of rest of the $W$-boson (right side). The dashed line represents the distribution in the presence of an anomalous gauge coupling $\Delta\kappa_\gamma=0.1$, while the solid line is the Standard Model prediction. The c.m. energy is $\sqrt{s}$ = 189 GeV.](fig8.ps) \[ending\] Conclusions ====================== The process of single-$W$ production in high-energy $e^+ e^-$ collisions is relevant at for the determination of the non-abelian self-couplings of the $W$ boson, and of primary importance at future Linear Colliders at the TeV scale, its cross section being dominant at very high energies with respect to other four-fermion processes. In order to give a contribution to the reduction of the theoretical uncertainty presently associated to the calculation of the single-$W$ cross section, the issue of higher-order photonic corrections has been carefully investigated within the standard technique. Theoretical and phenomenological arguments for the choice of the energy scale entering the have been proposed. Two possible solutions for the scale of radiation have been obtained. The former has been derived by means of general arguments concerning the soft and collinear limit of the $O(\alpha)$ corrections coming from the radiation of external legs. The latter, which can be considered as a naive ansatz, has been driven by thinking of the single-$W$ process in terms of the Weizsäcker-Williams approximation. Numerical calculations show that the typically adopted choice of the center-of-mass energy of the reaction, as radiation scale for the process, can lead to over-estimate the radiative correction by a factor of $1.5$, implying an under-estimate of the cross section of about $4\%$. Also the choice of fixing the scale to the momentum transfer $t$ in the $ee\gamma^*$ vertex for both the leads to an under-estimate of the photon correction of about $4\%$. The difference between the predictions given by the two set of scales of eq. (\[eq:wscales\]) and eq. (\[eq:naive\]) is at the per mille level in the energy range. Therefore, the naive scales of eq. (\[eq:naive\]) provide a good ansatz for the energy scale of QED radiation in the single-$W$ process, which could be simply implemented in MC tools for data analysis and further corroborated by the comparison with the results of other approaches. The method here described for the energy scale determination in the can be simply generalized to other four-fermion process dominated by non-annihilation channels, such as single-$Z$ production. In order to provide adequate phenomenological predictions for precision experiments, also the running of the electromagnetic coupling constant has been accounted for in an effective way, [*i.e.*]{} by rescaling the differential cross section for the ratio of the electromagnetic coupling constant, evaluated at the typical scale of the process, to the same coupling evaluated from the input parameters according to tree-level relations. The effect of such rescaling amounts to a negative correction of about $5$-$6\%$, in agreement with recent findings [@rfl1], as far as the effect of $\alpha_{\rm QED}$ is concerned. In the light of the experimental precision for the single-$W$ process, the corrections considered in the present paper are phenomenologically relevant. According to the theoretical approach described in the present paper, an original programme [SWAP]{} has been developed, including exact tree-level matrix elements with finite fermion masses effects, anomalous couplings, vacuum polarization and higher-order corrections. The code is available for experimental analysis. The authors wish to thank the members of the four-fermion working group of the Monte Carlo Workshop (), in particular Y. Kurihara, G. Passarino, R. Pittau and M. Verzocchi, for useful discussions on the subject. A. Pallavicini is grateful to the [infn]{}, Sezione di Pavia, for having provided computer facilities. [99]{} , G. Altarelli, T. Sjöstrand and F. Zwirner eds., ( [**96-01**]{}, Geneva, 1996); G. Montagna, O. Nicrosini and F. Piccinini, [*Riv. Nuovo Cim.*]{} [**21**]{}, n. 9 (1998) 1, [hep-ph/9802302]{}, and references therein. ALEPH Coll., R. Barate et al., [*Phys. Lett. B*]{} [**462**]{} (1999) 389; L3 Coll., M. Acciarri et al., [*Phys. Lett. B*]{} [**436**]{} (1998) 417; DELPHI Coll., P. Abreu et al., [*Phys. Lett. B*]{} [**423**]{} (1998) 194; R. Tanaka, [hep-ex/9811039]{}, Proceedings of the 29th International Conference on High-Energy Physics (ICHEP98), Vancouver, Canada, 23-29 July 1998. J. Ellis and M. K. Gaillard, Report [**76-18**]{} (1976) 21; A. Donnachie and K. J. F. Gaemers, [*Z. Phys. C*]{} [**4**]{} (1980) 37; K. Neufeld, [*Z. Phys. C*]{} [**17**]{} (1983) 145. E. Gabrielli, Mod. Phys. Lett. [**A1**]{} (1986) 465. U. Baur, B.A. Kniehl, J.A.M. Vermaseren and D. Zeppenfeld, Proceedings of Large Hadron Collider Workshop, ( [**90-10**]{}, Geneva, 1990) Vol. II, p. 956; U. Baur, J.A.M. Vermaseren and D. Zeppenfeld, [*Nucl. Phys. B*]{} [**375**]{} (1992) 3; E.N. Argyres and C.G. Papadopoulos, [*Phys. Lett. B*]{} [**263**]{} (1991) 298. C.F. Weizsäcker, [*Z. Phys.*]{} [**88**]{} (1934) 612;\ E.J. Williams, [*Phys. Rev.*]{} [**45**]{} (1934) 729.\ See also S. Frixione, M. Mangano, P. Nason and G. Ridolfi, [*Phys. Lett. B*]{} [**319**]{} (1993) 339. T. Tsukamoto and Y. Kurihara, [*Phys. Lett. B*]{} [**389**]{} (1996) 162. E. Accomando, A. Ballestrero and E. Maina, [*Phys. Lett.B*]{} [**479**]{} (2000) 209. G. Passarino, The single-$W$ production case, [hep-ph/9810416]{} E. E. Boos and M. N. Dubinin, Single-$W$ production at Linear Colliders, [hep-ph/9909214]{}. F.A. Berends, C.G. Papadopoulos and R. Pittau, Four-fermion production in electron-positron collisions with NEXTCALIBUR, [hep-ph/0002249]{}. M. Grünewald, G. Passarino et al., Report of the four-fermion working group of the workshop (), in preparation. See also [*http://www.to.infn.it/$\sim$giampier/lep2.html*]{}. Y. Kurihara, D. Perret-Gallix and Y. Shimizu, [*Phys. Lett. B*]{} [**349**]{} (1995) 367; U. Baur, D. Zeppenfeld, [*Phys. Rev. Lett.*]{} [**75**]{} (1995) 1002. E.N. Argyres et al., [*Phys. Lett.B*]{} [**349**]{} (1995) 367. W. Beenakker et al. [*Nucl. Phys. B*]{} [**500**]{} (1997) 255. G. Passarino, Unstable particles and non-conserved currents: a generalization of the fermion-loop scheme, [hep-ph/9911482]{} G. Passarino, Single-$W$ production and fermion-loop scheme: numerical results, [hep-ph/0001212]{}. G.A. Schuler and T. Sjöstrand, [*Z. Phys. C*]{} [**73**]{} (1997) 677 and references therein; P. Aurenche, G.A. Schuler et al., in [*Physics at LEP2*]{}, G. Altarelli, T. Sjöstrand and F. Zwirner eds., ( [**96-01**]{}, Geneva, 1996), Vol. 1, p. 291, [hep-ph/9601317]{} and references therein. See, for example, G. Montagna, O. Nicrosini and F. Piccinini, [*Phys. Lett. B*]{} [**358**]{} (1996) 348; S. Jadach, M. Melles, B.F.L. Ward and S.A. Yost, Acta Phys.Polon. [**B30**]{} (1999) 1745. M. Greco and O. Nicrosini, [*Phys. Lett. B*]{} [**240**]{} (1990) 219. C.M. Carloni Calame, C. Lunardini, G. Montagna, O. Nicrosini and F. Piccinini, Large-angle Bhabha scattering and luminosity at flavour factories, [hep-ph/0003268]{}. W. Beenakker, F.A. Berends et al., in [*Physics at LEP2*]{}, G. Altarelli, T. Sjöstrand and F. Zwirner eds., ( [**96-01**]{}, Geneva, 1996), Vol. 1, p. 79, [hep-ph/9602351]{}; W. Beenakker and A. Denner, Standard Model predictions for $W$-pair production in electron-positron collisions, DESY 94-051. W. Beenakker, F.A. Berends and W.L. van Neerven, Proceedings of Radiative Corrections for $e^+ e^-$ collisions, J.H. Kühn ed., (Springer-Verlag, 1989), p. 3. G. Montagna et al., On photon radiation in single-$W$ process, talk given by A. Pallavicini at the meeting of the workshop, , 12/10/1999, see [*http://www.to.infn.it/$\sim$giampier/lep2.html*]{}. Y. Kurihara et al., QED radiative corrections to the non-annihilation processes using the Structure Function and the Parton Shower, [hep-ph/9912520]{}. E.A. Kuraev and V. Fadin, [*Sov. J. Nucl. Phys.*]{} [**41**]{} (1985) 466 ; G. Altarelli and G. Martinelli, in [*Physics at LEP*]{}, J. Ellis and R. Peccei eds., (86-02, Geneva, 1986), Vol. 1 p. 47; O. Nicrosini and L. Trentadue, [*Phys. Lett. B*]{} [**196**]{} (1987) 551, [*Z. Phys. C*]{} [**39**]{} (1998) 479; F.A. Berends, G. Burgers and W.L. van Neerven, [*Nucl. Phys. B*]{} [**297**]{} (1988) 429; S. Jadach and M. Skrzypek, [*Z. Phys. C*]{} [**49**]{} (1991) 577; M. Skrzypek, [*Acta Phys. Pol. B*]{} [**23**]{} (1992) 135; M. Cacciari, A. Deandrea, G. Montagna and O. Nicrosini [*Europhys. Lett. *]{} [**17**]{} (1992) 123; G. Montagna, O. Nicrosini and F. Piccinini, [*Phys. Lett. B*]{} [**406**]{} (1997) 243, A.B. Arbuzov, [*Phys. Lett. B*]{} [**470**]{} (1999) 252. V.N. Gribov and L.N. Lipatov, Sov. J. Nucl. Phys. [**15**]{} (1972) 298; G.Altarelli and G.Parisi, Nucl. Phys. [**B126**]{} (1977) 298; Y.L. Dokshitzer, Sov. Phys. JETP [**46**]{} (1977) 641. See, for example, D.R. Yennie, S.C. Frautschi and H. Suura, Ann. Phys. [**13**]{} (1961) 379; G. Pancheri, Phys. Lett. [**B315**]{} (1993) 477. G. Sterman and S. Weinberg, Phys. Rev. Lett. [**39**]{} (1977) 1436. M. Caffo, R. Gatto and E. Remiddi, [*Nucl. Phys. B*]{} [**252**]{} (1985) 378, [*Phys. Lett. B*]{} [**139**]{} (1984) 439. J. Fleischer and F. Jegerlehner, [*Z. Phys.*]{} [**C26**]{} (1985) 629. M. Cacciari, G. Montagna and O. Nicrosini, [*Phys. Lett. B*]{} [**274**]{} (1992) 473. F. Block and A. Nordsieck, [*Phys. Rev.*]{} [**52**]{} (1937) 54; T. Kinoshita, [*J. Math. Phys.*]{} [**3**]{} (1962) 650; T. D. Lee and M. Nauenberg, [*Phys. Rev. B*]{} [**133**]{} (1964) 1549. M. Greco, [*Riv. Nuovo Cim.*]{} [**11**]{} (1988) 1. F. Caravaglios and M. Moretti, [*Phys. Lett. B*]{} [**358**]{} (1995) 332. G. Montagna, M. Moretti, O. Nicrosini and F. Piccinini, [*Nucl. Phys.*]{} [**541**]{} (1999) 31; G. Montagna, O. Nicrosini and F. Piccinini, [*Comp. Phys. Commun.*]{} [**98**]{} (1996) 206. J. Fujimoto, T. Munehisa and Y. Shimizu, [*Prog. Theor. Phys.*]{} [**90**]{} (1993) 177; Y. Kurihara, J. Fujimoto, T. Munehisa and Y. Shimizu, [*Prog. Theor. Phys.*]{} [**95**]{} (1996) 375. K. Gaemers and G. Gounaris, [*Z. Phys. C*]{} [**1**]{} (1979) 259. K. Hagiwara, K. Hikasa, R.D. Peccei and D. Zeppenfeld, [*Nucl. Phys. B*]{} [**282**]{} (1987) 253. S. Eidelman and F. Jegerlehner, [*Z. Phys. C*]{} [**67**]{} (1995) 585. [^1]: \[legs\] This choice fixes our conventions. Outgoing particles will appear as ingoing ones with momentum and charge according to crossing symmetry. [^2]: Actually, for the single-$W$ final state under examination here and for realistic event selections, the differences between the two procedures are confined below the 1% level.

--- abstract: 'We present a Green’s function based framework for modeling the scanning tunneling spectrum from the normal as well as the superconducting state of complex materials where the nature of the tunneling process$-$ i.e. the effect of the tunneling ’matrix element’, is properly taken into account. The formalism is applied to the case of optimally doped Bi$_2$Sr$_2$CaCu$_2$O$_{8+\delta}$ (Bi2212) high-Tc superconductor using a large tight-binding basis set of electron and hole orbitals. The results show clearly that the spectrum is modified strongly by the effects of the tunneling matrix element and that it is not a simple replica of the local density of states (LDOS) of the Cu-$d_{x^2-y^2}$ orbitals with other orbitals playing a key role in shaping the spectra. We show how the spectrum can be decomposed usefully in terms of tunneling ’channels’ or paths through which the current flows from various orbitals in the system to the scanning tip. Such an analysis reveals symmetry forbidden and symmetry enhanced paths between the tip and the cuprate layers. Significant contributions arise from not only the CuO$_2$ layer closest to the tip, but also from the second CuO$_2$ layer. The spectrum also contains a longer range background reflecting the non-local nature of the underlying Bloch states. In the superconducting state, coherence peaks are found to be dominated by the anomalous components of Green’s function.' author: - Jouko Nieminen - Ilpo Suominen - 'R.S. Markiewicz' - Hsin Lin - 'A. Bansil' date: - Version of - Version of title: 'Spectral decomposition and matrix element effects in scanning tunneling spectroscopy of Bi$_2$Sr$_2$CaCu$_2$O$_{8+\delta}$.' --- Introduction ============ High resolution scanning tunneling spectroscopy (STS) together with other highly resolved spectroscopies such as angle resolved photoemission (ARPES), is making it possible to obtain a comprehensive mapping of the electronic spectrum of the high-temperature superconductors (HTSs) in both real and reciprocal space over a wide range of dopings and temperatures. These studies are providing insight into the rich phase diagrams of the HTSs, and are leading thus to an understanding of the ’missing links’ for developing a definitive theory of how high superconducting transition temperatures arise in these unconventional materials. In STS experiments, the focus to date has been on hole doped cuprates, especially on Bi$_2$Sr$_2$CaCu$_2$O$_{8+\delta}$ (Bi2212), which has been the subject of an overwhelming amount of experimental work, see, e.g, Refs. . Bi2212 is a typical cuprate material, which is an antiferromagetic insulator in the strongly underdoped regime, but exhibits a superconducting phase over a wide range of hole doping. STS can be applied to a substantial part of the doping and temperature spanned phase space of HTS materials. The superconducting (SC) phase is observed around optimal hole doping (OP), while the pseudogap (PG) phase is found within the underdoped regime (UD). As a practical limitation, STS requires a conducting sample, but the deeply underdoped regime is insulating and hence unreachable by STS. However, under experimental conditions the samples are not homogeneously doped. Rather, there is a strong spatial variation in doping, which makes observation of a continuum from the PG to the SC phase possible within one sample. Although these spatial variations in STS generally appear irregular, quite recently a more ordered coexistence of PG and SC phases has been observed [@Kohsaka]. The physics of the cuprates is dominated by the cuprate layers, which are usually not exposed to the tip of the apparatus. For example, in Bi2212, the quasiparticle tunneling takes place through insulating BiO and SrO layers. The conventional interpretation of the spectra is based on the assumption that the STS spectrum is directly proportional to the LDOS of the CuO$_ 2$ layer, especially the LDOS of the $d_{x^2-y^2}$ orbitals, thus neglecting the effects of the tunneling process in modifying the spectrum in the presence of the insulating overlayers and multiple orbitals. The motivation for this simplification is an attempt to reduce the quasiparticle structure to few band models, which are amenable to theoretical treatment of strong correlation effects in the presence of superconducting and antiferromagnetic order. Notably, there have been attempts to take the effect of the overlayers into account by assuming a ‘tunneling matrix element’ or a ‘filter function’ [@Balatsky; @hoogenboom; @Fischer]. With this background, our recent work on STS[@NLMB] of Bi2212 provides a significant advance in realistic material-specific modeling of the STS spectrum. We invoke a Green’s function approach where a [*large*]{} number of orbitals is included, and all tunneling paths to the tip in the semi-infinite solid are taken into account. We showed clearly that instead of being a simple reflection of LDOS of the Cu-$d_{x^2-y^2}$ orbitals, the STS signal represents a very complex mapping of the electronic structure of the system. In this study we extend our approach by decomposing the tunneling current in terms of regular and anomalous matrix elements of the spectral function in an atomic orbital basis. As in Ref. , we concentrate on Bi2212 as the canonical HTS material. We start by reformulating the well-established methods to model tunneling current in nanostructures into a more transparent form for interpreting tunneling in the superconducting state. Our derivation is based on the conventional Todorov-Pendry [@Todorov; @Pendry] approach (TP), which is closely related to the more common Tersoff-Hamann [@Tersoff] method (TH). TP and TH methods both employ a calculation of the LDOS, but TP is more naturally written in terms of Green’s functions. We will show, in fact, that TP decomposes into matrix elements of the spectral function, giving very detailed information concerning the origin of various features in the tunneling spectrum. We thus demonstrate how the contribution of different atomic orbitals to the total current can be extracted from the calculations. Our spectral decomposition also naturally distinguishes between the electron and hole nature of the quasiparticles in the superconducting state. In addition, it leads to a multiband generalization of filtering function by Martin et al [@Balatsky] and a clarification of selection rules governing tunneling through filtering layers. This information is important, e.g., in determining how a dopant or impurity atom alters the spectrum, and how the effect of such a perturbation is seen in real space. In order to gain a handle on the effects of filtering layers, we derive a consistent form of a filter function through Green’s function manipulations. This rigorous form for the filtering effects is useful for determining the relation between the tunnel current and the LDOS of the CuO$_2$ layers. We show that this relation is nontrivial in that some channels are ‘first-order forbidden’. Thus our new approach shows that no direct regular signal from $d_{x^2-y^2}$ orbitals of the Cu directly below the STM tip reaches the microscope. Instead the $d_{x^2-y^2}$ orbitals of the four neighboring Cu atoms give a major contribution to the tunneling signal. Although we concentrate on pristine systems in the present work, the results have important implications for inhomogeneous situations – e.g., the relationship between the observed features in the spectrum of an impurity atom and the underlying LDOS. This decomposition also allows treatment of the regular and anomalous propagation of quasiparticles in a superconductor, and on this basis we show that the coherence peaks result from the anomalous electron-hole propagation. The paper is organized as follows. The model for the geometrical structure and the electronic structure is introduced in Sections II.A and B, respectively. The methods to calculate the Green’s function in the normal and the superconducting state are derived in Sections II.B and C, respectively. The Todorov-Pendry equation for the tunneling current is decomposed into regular and anomalous terms to show not only the proper form of the matrix element but also the partial current terms for any chosen orbital in Section II.D. The formalism is applied to discuss STM topographic maps in Section III.A, and the STS spectrum of Bi2212 in Section III.B. The spectrum is then analyzed in terms of tunneling matrix elements and partial currents in Section III.C. Further comments on symmetry analysis are made in Section IV.A., and remarks on electron extraction/injection are made in Section IV.B. Finally, conclusions are drawn and future applications sketched in Section V. Relevant technical details of the form of boson-electron coupling assumed in the tunnel spectra and of the superconducting state calculations are given in the two appendices. ![(color online) (a) Side view of the simulation cell used to compute the tunneling spectrum of Bi2212. Tunneling signal from the conducting CuO$_2$ layers reaches the tip after passing through the filtering layers of SrO and BiO. (b) Cuprate layer showing the supercell consisting of eight primitive cells. (c) Top view of the surface showing the arrangement of various atoms.[]{data-label="geometric"}](fig1.pdf){width="50.00000%"} Description of the model ======================== Our theoretical framework involves three distinct steps. First, we choose a three-dimensional geometrical model of atoms with a sufficiently large simulation cell with periodic boundary conditions in the horizontal directions to treat a semi-infinite solid surface. Second, we attach a basis set of atomic orbitals to each atom. At this stage, the one-particle Hamiltonian is constructed and the corresponding Green’s function tensor is formed. Third, we apply our Green’s function formalism to evaluate the tunneling current. The technical details of these three steps are outlined in the following three subsections. Sample geometry --------------- We model the Bi2212 sample as a slab of seven layers [@footnote1] in which the topmost layer is BiO, followed by layers of SrO, CuO$_2$, Ca, CuO$_2$, SrO, and BiO, as shown in Fig. \[geometric\](a). The tunneling computations are based on a $2\sqrt{2} \times 2\sqrt{2}$ real space supercell consisting of 8 primitive surface cells with a total of 120 atoms (see Fig. \[geometric\](b)). The coordinates are taken from the tetragonal crystal structure of Ref. . For STS simulations, the STM tip is modeled as an orbital with an s-wave symmetry at the assumed position of the apex of the tip. This tip is allowed to scan across the substrate for generating the topographic maps such as those in Fig. \[exptheo\], or held fixed on top of a surface Bi atom for the computed spectra presented for example in Fig. \[pristine\]. Construction of the uncorrelated normal state Hamiltonian --------------------------------------------------------- In order to construct a realistic framework capable of describing the tunneling spectrum of the normal as well as the superconducting state of the cuprates, we start with the normal state Hamiltonian for the semi-infinite solid in the form $$\hat{H}_1 = \sum_{\alpha\beta\sigma} \left[\varepsilon_{\alpha}c^{\dagger}_{\alpha \sigma} c_{\alpha \sigma}+ V_{\alpha \beta} c^{\dagger}_{\alpha \sigma} c_{\beta\sigma}\right],\label{H1}$$ which describes a system of tight-binding orbitals created (or annihilated) via the real-space operators $c^{\dagger}_{\alpha \sigma}$ (or $c_{\alpha \sigma}$). Here $\alpha$ is a composite index denoting both the type of orbital (e.g. Cu-$d_{x^2-y^2}$) and the site on which this orbital is placed, and $\sigma$ is the spin index. $\varepsilon_\alpha$ is the on-site energy of the $\alpha^{th}$ orbital. $\alpha$ and $\beta$ orbitals interact with each other through the potential $V_{\alpha\beta}$ to create the energy eigenstates of the entire system. The specific electron and hole orbital sets used for various atoms are: ($s,p_x,p_y,p_z$) for Bi, Ca and O; $s$ for Sr; and ($4s,d_{3z^2-r^2},d_{xy},d_{xz},d_{yz}, d_{x^2-y^2}$) for Cu atoms. This yields 58 electron or hole orbitals in a primitive cell and a total of $2 \times 464$ orbitals in the $2\sqrt{2} \times 2\sqrt{2}$ simulation supercell. The number of [**k**]{}-points used in the computations depends on whether we do band calculations or solve the Green’s function. For band calculations, we use a dense set of [**k**]{}-values to produce smooth bands for directions $\Gamma \rightarrow M \rightarrow X \rightarrow \Gamma$ as seen for example in Fig. \[norbands\]. In the case of Green’s function calculations, we use $N_k=256$ [**k**]{}-points for the supercell Brillouin zone. This corresponds to $8 \times 256 = 2048$ [**k**]{}-points for a primitive cell. $v_{\alpha \beta m} (eV)$ --------------------------- ---------------- ---------------- ------------- ---------------- ---------------- ------------- ---------------- ------------- ---------------- $v_{ss\sigma}$ $v_{sp\sigma}$ $v_{pp\sigma}$ $v_{pp\pi}$ $v_{sd\sigma}$ $v_{pd\sigma}$ $v_{pd\pi}$ $v_{dd\sigma}$ $v_{dd\pi}$ $v_{dd\delta}$ -0.28 0.94 1.23 -0.13 -0.62 -2.81 1.16 -9.00 12.60 -2.29 : Slater-Koster prefactors, $v_{\alpha \beta m},$ and onsite energies $\varepsilon_{\alpha}.$ The $v_{\alpha \beta m}$ are used to construct the Hamiltonian overlap matrix elements $V_{\alpha\beta}$ as described in Ref. [@Slater]. \ $\varepsilon_{\alpha} (eV)$ ----------------------------- --------- --------- --------- --------- --------- --------- s/Bi p/Bi s/O(Bi) p/O(Bi) s/Sr s/Ca p/Ca -12.200 1.800 14.700 -2.400 7.819 5.631 13.335a s/O(Sr) p/O(Sr) s/Cu d/Cu s/O(Cu) p/O(Cu) -15.270 -2.353 5.001 -2.962 -18.560 -3.825 : Slater-Koster prefactors, $v_{\alpha \beta m},$ and onsite energies $\varepsilon_{\alpha}.$ The $v_{\alpha \beta m}$ are used to construct the Hamiltonian overlap matrix elements $V_{\alpha\beta}$ as described in Ref. [@Slater]. \ The Slater-Koster formalism [@Slater; @Harrison; @Shi] is used to fix the angular dependence of the tight binding overlap integrals. The onsite energies and the prefactors are fitted to the LDA band structure of Bi2212 that underlies for example the extensive angle-resolved photointensity computations of Refs.  . In Table I, we show the specific values of the $v_{\alpha \beta m}$ prefactors used for computing the Slater-Koster hopping integrals. Notably, we have shifted the bottom of the BiO conduction band to agree with experiments, which do not observe the Bi-bands at least within $1eV$ above the Fermi-level. This choice is also supported by calculations of Ref. , which show the sensitivity of the position of the Bi-band with respect to impurities and doping. The absence of the bottom of the BiO band in the STS spectra may also be due to a voltage gradient across the insulating filter layers (BiO and SrO layers) when applying a bias voltage between the tip and the sample. If so, the absolute value of the voltage within these layers is less than the bias voltage $V_b$, and thus the apparatus would need to apply a bias which would be significantly larger than $V_b$ to locally see states that are strictly at $E_F+eV_b.$ The tight-binding parameters of the normal state Hamiltonian of Eq. produce the detailed band structure of Bi2212 shown in Fig. \[norbands\]. While the tight-binding band structure is in reasonable agreement with the LDA band structure of Ref. , in order to carry out spectroscopic computations, one must additionally make sure that the underlying wavefunctions are described correctly including their symmetries. Our procedure based on the the use of Koster-Slater matrix elements not only fits the band stuctures, but the symmetries and phases of the associated wavefunctions are also described correctly. ![(color online) (a)-(d): Normal state band structure of Bi2212 for the tight-binding Hamiltonian and from first-principles LDA computations. Weights of Cu-$d_{x^2-y^2}$ and Cu-$d_{z^2}$ contribution to the bands are shown using a colorscale where red denotes high and blue low values (see colorbar). Note that the tight-binding calculations are done for a slab, so that the the tight-binding bands do not display the splitting of Bi-O bands seen in LDA results. The LDA bands have been calculated using Virtual Crystal Approximation (VCA) with 24% Pb doping to set the bottom of the BiO band. (e) and (f): Quasiparticle band structure in the superconducting state based on the Hamiltonian of Eq. is shown in (e). Panel (f) zooms in on the gap region of (e) which is shaded grey. Electron character of quasiparticles is shown in red and the hole character in blue. Notice, that the quasiparticles differ significantly from being electrons or holes only in the close neighborhood of the superconducting gap around the M-point. []{data-label="norbands"}](fig2.pdf){width="50.00000%"} Figs. \[norbands\] (a)-(b) show the normal state tight-binding band structure based on our $58$ orbital Hamiltonian of Eq. . The main cuprate bands, with predominantly Cu-$d_{x^2-y^2}$ character, are seen in panels (a) and (b) to follow the corresponding LDA calculations in panels (c) and (d). Note that in our tight-binding modeling, we have adjusted the positions and bilayer splitting of the two van Hove singularities (VHSs) to approximately match the experimental photoemission and STS findings for the optimal doping (OP) region with hole concentration $p \approx 0.16$ (Ref. ). In addition to Cu-$d_{x^2-y^2}$, Cu-$d_{z^2}$ is seen in panels (b) and (d) to give a significant spectral weight to this band, especially at energies below the Fermi-level. The complicated ‘spaghetti’ region has large contributions from the $d_{z^2}$ of Cu and horizontal $p_x(p_y)$ orbitals of the oxygens within the cuprate layer as well as the vertical $p_z$ orbital of the apical oxygen. Concerning the filter layers, the bottom of the BiO-like conduction band (or bismuth pocket) along the $M(\pi ,0)$ direction carries the character of the horizontal p-orbitals of the surface oxygens $O(Bi)$ (see Fig.  \[norbands\] (a)). In tunneling calculations, we directly evaluate the Green’s function instead of diagonalizing the Hamiltonian. For this purpose, the normal state Green’s function is solved first by starting from the orbital matrix elements of the Green’s function: $$g^{\pm}_{\alpha \beta} = \frac{\delta_{\alpha \beta}}{\varepsilon - \varepsilon_{\alpha} - \Sigma^{\pm}_{\alpha}(\varepsilon)}, \label{independent}$$ where $\varepsilon_{\alpha}$ is the onsite energy of the orbital $\alpha.$ At this point, a diagonal self-energy $\Sigma^{\pm}_{\alpha} =\Sigma_{\alpha}{'} \pm i \Sigma_{\alpha}{''}$ can be included straightforwardly. The simplest self-energy is a constant broadening of the states in the form of a convergence factor $\Sigma^{\pm}_{\alpha} = \mp i\eta.$ Appendix A (Eq. ) presents a more general self-energy which we use to model electron-boson coupling. The total Green’s function $G$ is constructed by solving Dyson’s equation $$G = g + gVG, \label{Dyson}$$ where $V_{\alpha \beta}$ are the off-diagonal overlap integrals of Eq. Dyson’s equation is exactly solved using the method described in Ref. [@Nieminen], which is suitable for tunneling calculations [@footfourier]. Pairing interaction and the superconducting state Hamiltonian ------------------------------------------------------------- Superconductivity is included by adding a pairing interaction term $\Delta$ in the Hamiltonian of Eq. as follows $$\hat{H} = \hat{H}_1 + \sum_{\alpha \beta \sigma} \left[\Delta_{\alpha \beta} c^{\dagger}_{\alpha \sigma} c^{\dagger}_{\beta -\sigma} + \Delta_{\beta \alpha}^{\dagger} c_{\beta -\sigma} c_{\alpha \sigma} \right] \label{hamiltonian}$$ A gap parameter value of $\vert\Delta\vert = 0.045 eV$ is chosen to model a typical experimental spectrum[@McElroy] for the illustrative purposes of this study. We take $\Delta$ to be non-zero only between $d_{x^2 - y^2}$ orbitals of the nearest neighbor Cu atoms, and to possess a d-wave form, i.e., $\Delta_{d (d \pm x)} = +\vert \Delta \vert$ and $\Delta_{d (d \pm y)} = -\vert \Delta \vert,$ where $d$ denotes the $d_{x^2-y^2}$ orbital at a chosen site, and $d \pm x/y$ the $d_{x^2-y^2}$ orbital of the neighboring Cu atom in x/y-direction. In momentum space, the corresponding $\Delta$ is given by $$\Delta_k = \frac{\Delta}{2} \left[\cos{k_x a} - \cos{k_y a} \right],$$ where $a$ is the in-plane lattice constant. The pairing interaction of Eq.  allows electrons of opposite spins to combine to produce superconducting pairs such that the resulting superconducting gap is zero along the nodal directions $k_x=\pm k_y$, and is maximum along the antinodal directions. This choice of pairing interaction follows, e.g., the one-band formalism given in Ref. . ![(color online) [*Main:*]{} LDOS (or the diagonal elements $\rho_{\alpha \alpha}$ of the density matrix; see Appendix B for details) of $d_{x^2-y^2}$ (green) and $d_{z^2}$ (blue) orbitals of Cu. \[Note $d_{z^2}$ curve is scaled up by a factor of 4.5 to compare the shapes of the two LDOSs.\] Oscillations at high positive or negative energies ( above $\pm 0.5$ eV) are artifacts due to the use of a sparse mesh of [**k**]{}-points in the computation. [*Inset:*]{} Anomalous density matrix term $\rho^{eh}_{\alpha \beta}$ discussed in Appendix B, where $\alpha$ and $\beta$ denote $d_{x^2-y^2}$ orbitals of two neighboring Cu atoms.[]{data-label="ldos400403"}](fig3.pdf){width="50.00000%"} For treating the superconducting case, we employ the tensor (Nambu-Gorkov) Green’s function ${\cal G}$ (see Ref. ) with the corresponding Dyson’s equation: $${\cal G} = {\cal G}^0 + {\cal G V G}^0, \label{dyson1}$$ where $${\cal G} = \left( \begin{array}{cc} G_{e}& F\\ F^{\dagger}& G_{h} \end{array} \right)~\textrm{and}~ {\cal V} = \left( \begin{array}{cc} 0& \Delta\\ \Delta^{\dagger}& 0 \end{array} \right)$$ where $G_{e}$ and $G_{h},$ denote the Green’s functions for the electrons and holes, respectively. The normal state electron Green function $G_{e}$ can be used to derive the hole Green function $G_{h}$. It can be shown by, e.g., the equation of motion method, that $$G^{\pm}_{h,\alpha \beta}(E) = -G^{\mp}_{e, \beta \alpha}(-E)$$ It is straightforwardly shown then that $$\begin{aligned} G_{e}& =& G_{e}^0 + F\Delta^{\dagger} G_{e}^0 \nonumber\\ F&= & G_{e} \Delta G_{h}^0 \label{fdelta}\end{aligned}$$ The quasiparticle Green’s function projected onto electron degrees of freedom is then written in the form $$G_{e} = G_{e}^0 + G_{e} \Sigma^{BCS} G_{e}^0, ~\textrm{where}~ \Sigma^{BCS} = \Delta G_{h}^0 \Delta^{\dagger}. \label{bcsself}$$ We also need the self-energy term $\Sigma^{h}_{\alpha}$ for holes. Since the transformation from electron to holes follows that of the Green’s function, we obtain the general form $$\Sigma^{h}_{\alpha}(\varepsilon) = - \Sigma^{e*}_{\alpha}(-\varepsilon) = -\Sigma^{'}_{\alpha}(-\varepsilon) + i \Sigma^{''}_{\alpha}(-\varepsilon).$$ In our particular case, we use a self-energy with an odd real part and an even imaginary part as discussed in Appendix A (see Eq. ) Our self-energy is thus invariant under electron-hole transformation. Figs. \[norbands\](e) and (f) show the modifications of the normal state band structure from the introduction of the pairing interaction. Only the region within $\pm 500 meV$ of the Fermi level is shown in panel (e), as the remainder of the bands are unchanged from the normal state results of panels (a) and (b). The superconducting state dispersion in panels (e) and (f) clearly displays a d-wave gap with a maximum in the antinodal region near the $M$ point and zero gap along the nodal direction near $(\pi /2,\pi /2)$. Note that both bonding and antibonding VHSs possess gaps of similar magnitude. Fig. \[norbands\](e) also shows the relative electron/hole character of the quasiparticles. As expected, the quasiparticles are very distinctly either electron- or hole-like almost everywhere except within a very narrow energy range at the top and bottom of the SC gap. Fig. \[ldos400403\] further shows that mixing of the electron and hole features gives rise to coherence peaks in the LDOS of Cu-$d_{x^2-y^2}$ and to a lesser extent in the LDOS of Cu-$d_{z^2}$. The effects of electron-hole mixing are however most pronouned in the anomalous matrix element of the quasiparticle Green’s function (inset to Fig. \[ldos400403\] and Fig. \[offdiag\]). In fact, the off-diagonal matrix element between an up-spin $d_{x^2-y^2}$ electron orbital and a down-spin $d_{x^2-y^2}$ hole orbital of two neighboring Cu atoms gives the most important term in the anomalous part of the Green’s function. This term has d-wave symmetry, which manifests itself as a change in sign each time we make a rotation of $\frac{\pi}{2}$ around the central Cu site. In addition to the coherence peaks, the anomalous density matrix inherits features from the VHSs in the regular part of the density matrix, which in view of electron-hole symmetry are reflected on both sides of the Fermi energy. Additionally, strong hybridization between up-spin Cu-$d_{x^2-y^2}$ electron orbitals and down-spin orbitals of O $p_{x}$ holes (and vice versa) takes place as shown in Fig. \[offdiag\]. This term is comparable in strength to the Cu-$d~-~$Cu-$d$ terms and changes sign in rotations of $\pi$ for reasons explained in the special case (3) of the following paragraph. Fig. \[offdiag\] also shows a small onsite contribution from the up-spin electron and down-spin hole of the $p_x$-orbital on the oxygen between two neighboring Cu atoms. It is notable that these matrix elements strictly follow the d-wave symmetry in rotations around the central Cu atom. ![(color online) [*Main:*]{} Matrix elements of the anomalous Green’s function for onsite $p_x$-orbital of an intermediate oxygen atom (green lines), $d_{x^2-y^2}$ orbitals of two neighboring Cu atoms (red lines), and between Cu-$d_{x^2-y^2}$ and a $p_x$ orbital of a neighboring oxygen (black lines). [*Inset:*]{} The directional dependence of the sign of the off-diagonal element $F_{dp}$. For details see special case (3) in the text.[]{data-label="offdiag"}](fig4.pdf){width="50.00000%"} These transformation properties follow consistently from Eq. . Let us, for example, look at the equation in the $x$-direction: $F_{\alpha \beta} = G_{e,\alpha d}\Delta_{d (d\pm x)} G^{0}_{h,(d \pm x)\beta},$ where $d$ is a shorthand notation for $d_{x^2-y^2}$ of a chosen Cu atom, and $d \pm x$ stands for the $d_{x^2-y^2}$ orbital of the neighboring Cu atom in the positive/negative x-direction and consider several specific cases as follows.\ (1) For $\alpha=d$ and $\beta=d\pm x$, both $G_{e,\alpha d}$ and $G^{0}_{h,(d \pm x)\beta}$ are onsite matrix elements, and thus their sign remains invariant when changing from one Cu to another. Hence the term $\Delta_{d(d\pm x)}$ is decisive, and the sign can change only in going from x- to y- direction;\ (2) For $\alpha= \beta=$ O-$p_x,$ we have to first look at the term $G_{e,p_x d} G^{0}_{h,(d \pm x) p_x}$. Since the relative phases of the off-diagonal matrix elements of the Green’s function are proportional to the sign of the overlap of the two orbitals, it is straightforward to see from the signs of the lobes of the $d$ and $p$ orbitals that this product is invariant to change in direction as well as in going from x to y. Therefore, $\Delta_{d (d\pm x)}$ again gives the d-wave symmetry of these terms;\ (3) For $\alpha = d$ and $\beta =$ O-$p_{x}$, $G_{e,\alpha d}$ is diagonal and thus invariant. Considering the overlaps, one sees that $$G^{0}_{h,(d - x) \beta} = -G^{0}_{h,(d + x) \beta},$$ and $$G^{0}_{h,(d + y) \beta} = -G^{0}_{h,(d - y) \beta} = -G^{0}_{h,(d + x) \beta}.$$ But, since $\Delta_{d (d\pm x)} = -\Delta_{d (d\pm y)}$, $$F_{d p_{x}(-)} = -F_{d p_{x}(+)} = -F_{d p_{y}(+)}= F_{d p_{y}(-)},$$ as shown in the inset to Fig. \[offdiag\]. Eq. of Appendix B shows that $F_{\alpha \beta}\propto\langle c_{\alpha\uparrow}c_{\beta \downarrow} \rangle.$ Hence, case (3) of the last paragraph indicates that there is a significant pairing $\langle c_{d_{x^2-y^2} \uparrow}c_{\phi \downarrow} \rangle$ when $$\vert \phi \rangle \propto \vert p_{x}(+) \rangle + \vert p_{y}(+) \rangle - \vert p_{x}(-)\rangle - \vert p_{y}(-) \rangle.$$ Recall that we introduced superconductivity in Hamiltonian of Eq. on the Cu-$d_{x^2-y^2}$ orbitals. Thus we see that within our model the strong Cu-O hybridization automatically induces pairing on the oxygen orbitals. This pairing is analogous to the concept of Zhang-Rice singlets (ZRS) in the low doping limit [@Zhang], where pair states $$\vert d_{x^2-y^2} \uparrow \rangle \vert \phi \downarrow\rangle - \vert d_{x^2-y^2} \downarrow \rangle \vert \phi \uparrow\rangle$$ are formed. Note, however, that ZRS is a concept related to doping levels in the ‘normal’ phase, and is not directly concerned with superconductivity. Nevertheless, the preceding considerations indicate that our model is in accord with the ZRS scenario of the normal state [@onsitep]. Green’s function formulation of tunneling current ------------------------------------------------- We turn now to consider the formulation of the tunneling spectrum. For this purpose, we apply the conventional form of the Todorov-Pendry expression [@Todorov; @Pendry] for the differential conductance $\sigma$ between orbitals of the tip ($t,t'$) and the sample ($s,s'$), which in our case is straightforwardly shown to yield $$\sigma = \frac{dI}{dV} = \frac{2 \pi e^2 }{ \hbar} \sum_{t t' s s'} \rho_{tt'}(E_F)V_{t's} \rho_{ss'}^{}(E_F+eV)V_{s't}^{\dagger}, \label{conductance}$$ where the density matrix $$\rho_{s s'} = -\frac{1}{\pi}Im[G_{s s'}^{+}] = \frac{1}{2\pi i} \left( G^{-}_{s s'} - G^{+}_ {s s'} \right) , \label{spectralfunctiona}$$ is given in terms of the retarded electron Green function or propagator $G_{s s'}^{+}$. Eq. differs from the more commonly used Tersoff-Hamann approach[@Tersoff] in that it takes into account the details of the symmetry of the tip orbitals and how these orbitals overlap with the surface orbitals. Since electrons are not eigenparticles in the presence of the pairing term, Dyson’s equation needs to be applied to the Green’s function tensor: $${\cal G}^{-} = {\cal G}^{+} + {\cal G}^{+}({\mathbf \Sigma^{-}}-{\mathbf\Sigma^{+}}){\cal G}^{-} = {\cal G}^{+} - 2i{\cal G}^{+}{\mathbf \Sigma^{''}}{\cal G}^{-} \label{spectaltensor}$$ After extracting the electron part from Eq. and applying Eq. , the spectral function can be written as: $$\rho_{s s'} = -\frac{1}{\pi}\sum_{\alpha} (G_{s \alpha}^{+} \Sigma{''}_{\alpha} G_{\alpha s'}^{-} + F_{s \alpha}^{+} \Sigma{''}_{\alpha} F_{\alpha s'}^{-}), \label{spectralfunction}$$ Using Eq. , the tunneling current of Eq. can be recast into the form $$\sigma = \sum_{t \alpha} T_{t \alpha}, \label{transition}$$ where $$%\begin{align} T_{t \alpha} = -\frac{2 e^2 }{ \hbar}\sum_{t' s s'} \rho_{tt'}(E_F) V_{t's}(G^{+}_{s\alpha}\Sigma{''}_{\alpha}G^{-}_{\alpha s'} + F^{+}_{s\alpha}\Sigma{''}_{\alpha}F^{-}_{\alpha s'})V_{s't}^{\dagger}, \label{partial} %\end{align}$$ and the Green’s function and the self-energy are evaluated at energy $E = E_F + e V_b.$ Eqs. and are an extension of the Landauer-Büttiker formula for tunneling across nanostructures (see, e.g., Ref. ), and represent a reformulation of Refs.  and . By comparing Eqs. and , we see that if the tip makes contact with only a single surface atom orbital, e.g., a Bi-$p_z$ orbital, then the tunneling current is directly proportional to the LDOS [*of that orbital*]{}. In particular, the tunneling current bears in general no such simple relationship to the quantity of most interest, namely, the LDOS on the CuO$_2$ plane. Obviously, the tunneling formalism of Eq. must be further elaborated in order to find the relation between the interesting LDOSs and the tunneling spectrum. ### Tunneling channels, filter function and tunneling matrix element The experimental STM spectra in the cuprates have to date been mostly compared to the electronic LDOS of the superconducting cuprate layer, especially the LDOS of the Cu-$d_{x^2-y^2}$ orbital. The discrepancies between the spectra and the LDOS are then ascribed to ‘tunneling matrix elements’ or ‘filtering functions’ [@Balatsky]. The former refers to the general problem of modeling spectroscopies, where the signal is distorted by the spectroscopic process, and may even vanish due to the presence of selection rules. The latter term refers to how the states of electrons (or quasiparticles) from the initial state within the superconducting layers are modified when traveling through the oxide overlayers before reaching the tip. Eq. above accounts fully for the tunneling process, and it can be reformulated to reveal, for example, the filtering effect more clearly. For this purpose, it is convenient to the denote various orbitals as follows: $s$ and $s'$ for the orbitals of the sample surface, which overlap with the tip orbital $t$; $f$ and $f'$ for the orbitals of the filter layers, BiO and SrO; $c$ and $c'$ for orbitals in the cuprate layer; and, $\alpha$ for any orbital that is singled out, which in our case usually will be an orbital in the cuprate layer. Denoting the Green’s function for the filter layers decoupled from the rest of the system by $G^{0+}_{sf}$, and the matrix elements within the cuprate layer in the coupled system by $G^{+}_{c\alpha}$, application of Dyson’s equation to $G^{+}_{s\alpha}$ yields $$G^{+}_{s\alpha} = G^{0+}_{sf}V_{fc}G^{+}_{c\alpha}~~ \textrm{and}~~ F^{+}_{s\alpha} = G^{0+}_{sf}V_{fc}F^{+}_{c\alpha}$$ Hence, Eq. can be written as $$%\begin{align} %\label{fpartial} T_{t \alpha} = -\frac{2 e^2 }{ \hbar}\sum_{t' c c'} \rho_{tt'}(E_F)M_{t'c}(G^{+}_{c\alpha}\Sigma{''}_{\alpha}G^{-}_{\alpha c'}+F^{+}_{c\alpha}\Sigma{''}_{\alpha}F^{-}_{\alpha c'})M_{c't}^{\dagger} \label{cupraspectral} %\end{align}$$ where $$M_{tc} = V_{ts}G^{0+}_{sf}V_{fc}, \label{filter}$$ which gives the filtering amplitude between the cuprate layer and the tip, and constitutes a multiband generalization of filtering function of Ref. . Similarly, the matrix element of the density of states operator $\rho_{cc'}$ within the cuprate plane can be recovered in terms of the spectral function: $$%\begin{align} %\label{fpartial} %T_{t \alpha} = \sigma = \frac{2 \pi e^2 }{\hbar}\sum_{tt' c c'} \rho_{tt'}(E_F)M_{t'c}\rho_{cc'}(E_F+eV)M_{c't}^{\dagger}, \label{cupraldos} %\end{align}$$ Eqs. - show a number of interesting aspects of the tunneling process as follows.\ (1) Since applying the [*filtering matrix element*]{} $M_{tc}$, which describes the effect of the BiO and SrO overlayers, involves $M$ and $M^{\dagger}$, interference effects will occur between various paths to the tip from the cuprate layers through the filter layer;\ (2) The partial current terms in under the summation are proportional to elements of the density matrix confined to the cuprate layer. Only orbitals with a notable overlap with the $p_z$ orbital of the apical oxygen on the SrO layer will give a significant contribution to the total current;\ (3) The partial elements of the [*spectral function*]{} $$\rho_{c c' \alpha}= -\frac{1}{\pi} (G^{+}_{c\alpha}\Sigma{''}_{\alpha}G^{-}_{\alpha c'}+F^{+}_{c\alpha}\Sigma{''}_{\alpha}F^{-}_{\alpha c'}) \label{singlespec}$$ extracted from Eq. show which orbitals $\alpha$ contribute to the chosen element of the density matrix $\rho_{cc'}.$ Furthermore, the current contribution $T_{t \alpha}$ between the tip can be divided into regular and anomalous terms $T^{R}_{t \alpha}$ and $T^{A}_{t \alpha},$ respectively [@footdecomposition]. Since the filter layers are insulating at low energies, these layers will give little structure to the spectrum at low bias voltages, so that the structure of the spectrum is mainly controlled by the matrix elements $\rho_{cc'}$, and in this sense the spectrum is a filtered mapping of the LDOS of the cuprate orbitals. We will show however that the Cu-$d_{x^2-y^2}$ orbitals right below the tip do not enter the spectrum through Eq.  since their overlap with the relevant orbitals of the SrO layer is zero. Instead, Cu-$d_{z^2}$ has a large overlap with $p_{z}$ of the apical oxygen and hence these orbitals of the Cu atoms play a dominant role in the tunneling spectrum. The detailed contribution of any specific orbital $\alpha$ can be extracted from Eq. . The regular and anomalous matrix elements of the spectral function, $G^{+}_{c\alpha}\Sigma{''}_{\alpha}G^{-}_{\alpha c'}$ and $F^{+}_{c\alpha}\Sigma{''}_{\alpha}F^{-}_{\alpha c'}$, describe propagation of electrons or holes within the cuprate layer from orbital $\alpha$ to the orbitals $c$ and $c'$. The latter orbitals act as “gates” between the cuprate layer and the filter layer. For example, if $\alpha$ is $d_{x^2-y^2}$ of a Cu atom and $c$ and $c'$ are $d_{z^2}$ orbitals, which strongly overlap with the filter layer, the matrix element filtered by $M$ and $M^{\dagger}$ gives the contribution of a specific $d_{x^2-y^2}$ orbital to the total tunneling spectrum. Note that in the superconducting state the anomalous matrix elements of the spectral function must also be considered. $F_{\alpha \beta}(\tau)$ involves the creation of an electron with spin up coupled to the annihilation of a hole with spin down given by $\langle c^{\dagger}_{\beta \downarrow}(\tau) c^{\dagger}_{\alpha \uparrow}(0) \rangle,$ and thus describes the formation and breakup of Cooper pairs as shown in Appendix B. The decomposition of Eqs. - are, in fact, a generalization of the tunneling channel approach to transport through one-molecule electronic components [@Magoga] and STM of adsorbate molecules [@Sautet; @Niemi]. In the present context, the “tunneling path” analysis gives us the “origin” of the signal, since $G^{+}_{c\alpha}\Sigma{''}_{\alpha} G^{-}_{\alpha c'}$ gives the probability of propagation between orbitals $\alpha$ and $c.$ Results ======= Topographic maps ---------------- ![ (a) Typical experimental topographical STM map after Ref. . (b) The computed corrugation of two STM line scans and (c) theoretically predicted topographic map. The two paths are shown in (c) by arrows. []{data-label="exptheo"}](fig5.pdf){width="50.00000%"} We discuss first the topographic STM map, i.e., the constant current surface for a tip scanning across the sample surface. The computed topographic map is very robust against changes in measuring parameters such as the bias voltage or the tip-surface distance. Figure \[exptheo\] compares the calculated and typical experimental results. Furthermore, corrugation along two paths of line scan is shown in Fig. \[exptheo\] (b). The Bi atoms are seen as bright spots, while the surface oxygens are dark due to very low current coming through these surface atoms. We will see in connection with the analysis of the tunneling channels below that the apical oxygens act as the primary gate for passing electrons from the CuO$_2$ layers up to the surface BiO layer. Accordingly, the Bi atoms appear bright because there exists an easy channel between the surface Bi atoms and the apical oxygens below via the Bi $p_z$ orbitals. On the other hand, the oxygens in the surface layer are dark because the $p_{x,y}$ orbitals of O(Bi) are orthogonal to the (assumed) $s$-symmetry of the tip, while the O(Bi) $p_{z}$ orbitals are relatively weakly coupled to the $p_{z}$ of the apical oxygen as discussed below in connection with Fig. \[channels08\]. Tunneling spectra ----------------- Fig. \[pristine\] (a) compares a typical experimental (red line) STS spectrum[@McElroy] to the calculated one (black line). The overall agreement between theory and experiment is seen to be good, although the VHSs are seen as separate structures in the calculated curve [@footnote2; @MSB]. The agreement also extends to the low energy region shown in Fig. \[pristine\](b), where the width and positions of the coherence peaks is reproduced reasonably well.[@footkpoints] The tendancy for increasing intensity towards negative bias is seen in both measurements and computations. This is in sharp contrast to the shape of the LDOS of Cu-$d_{x^2-y^2}$ orbital (green curve). As emphasized in Ref. [@NLMB], this remarkable asymmetry of the spectrum between positive and negative bias voltages reflects the opening up of channels other than Cu-$d_{x^2-y^2}$, especially of Cu-$d_{z^2}$, as one goes to high negative bias. This asymmetry thus appears naturally within our conventional picture and cannot be taken to be a hallmark of strong correlation effects as has been thought to be the case. ![(color online) (a) A typical experimental tunneling spectrum (red line) from Bi2212 (after Ref. ) is compared with the calculated spectrum (black). The green curve shows the LDOS of the Cu-$d_{x^2-y^2}$. (b) Expanded view of the experimental and calculated spectrum in the low energy region. (c) Comparison of the model self-energy (Eq. ) assumed for the Cu-$d_{x^2-y^2}$ orbitals and the self-energy from the convolution of a Debye-type phonon spectrum and the LDOS of Cu-$d_{x^2-y^2}$ (Eq. ) as discussed in Appendix A. []{data-label="pristine"}](fig6.pdf){width="50.00000%"} There has been considerable interest in understanding the coupling of electrons to bosonic modes in the so-called ‘low-energy kink’ region within $\sim\pm 100$ meV of the Fermi level. In particular, the peak-dip-hump structure seen in the experimental spectrum in Fig. \[pristine\](b) is generally believed to be the result of the coupling of electronic degrees of freedom to a collective mode (Refs. ). Fig. \[pristine\](b) shows that the peak-dip-hump feature can be described by our simple self-energy correction discussed in Appendix A. This point however requires further study, including an analysis of how this feature evolves with doping. Selection rules --------------- The filter function $M_{tc}$ controls [*selection rules*]{} dictated by matching of the symmetry properties of the cuprate layer, filter layers and the tip. A closer examination of $M_{tc}$ reveals that strong tunneling through the apical oxygen layer is associated with a matching of the symmetry of the cuprate layer wave function to that of the apical O-$p_z$. The key is the relative symmetry of the wave functions with respect to the axis of tunneling: An ‘odd’ wave function, e.g., the Cu-$d_{x^2-y^2}$ has zero overlap with an ‘even’ wave function such as O-$p_z$. In contrast, two orbitals with the same symmetry couple more strongly. Accordingly, the $p_z$ of the apical oxygen and the Cu-$d_{z^2}$ possess large overlap, while Cu-$d_{x^2-y^2}$ has zero overlap with any $s$- or $p$-orbital of the apical oxygen. This is the reason that direct tunneling is forbidden between Cu-$d_{x^2-y^2}$ and the $s$-wave symmetric tip through the filter layer. Hence, $M_{tc}$ functions here are consistent with the filter function of Ref. . Similarly, coupling between an $s$-wave tip and the $p_x$ and $p_y$ orbitals of the Bi atom lying directly below the tip is forbidden. Therefore, within the filter layer, the main ‘vertical’ overlap is between the $p_z$ orbitals of Bi and apical oxygen, and these orbitals indeed are found to provide the main channel through the filter layers as depicted in Fig. \[channels08\](a). We find additional relatively small contributions from the on-site Bi-$s$-orbital and $p$-orbitals of the surrounding Bi and O(Bi) atoms, but such ’background’ contributions to the current do not seem to be dominated by any particular channel. Figure \[channels08\](b) illustrates another example of a symmetry-forbidden tunneling path, where the tip is centered between two surface Bi’s, i.e. on the top of an oxygen of the cuprate layer. Since we assume an $s$-wave tip with negative hopping integrals to the nearby Bi atoms, when we follow either path up to the Cu-$d_{z^2}$ orbitals, the signs of the hopping integrals are identical. However, the O-$p_{x}$ orbital between the two Cu atoms changes sign from one Cu to the other. This gives the two paths from O-$p_x$ to the $s$-wave tip an opposite phase leading to destructive interference between the paths, making the O atom invisible. However, if the $s$-wave tip is replaced by one with, e.g, $p_x$ symmetry, the oxygen would become visible and a weaker signal would appear from the neighboring $d_{x^2-y^2}$ orbitals. Experimentally, this could be accomplished by functionalizing the tip by attaching a suitable molecule to the tip. A similar procedure has been used to obtain a contrast inversion for CO molecules adsorbed on a Cu surface [@Rieder; @Niemi1]. ![(color online) (a) Dominant tunneling channel from the cuprate layer, from Cu $d_{x^2-y^2}$ orbitals through the neighboring Cu $d_{z^2}$ to Bi $p_{x}$ to the tip. (b) An oxygen atom in the cuprate layer is invisible to a STM tip right above, since the paths through Cu$_{1}$ and Cu$_{2}$ interfere destructively. []{data-label="channels08"}](fig7.pdf){width="50.00000%"} Tunneling channels ------------------ The origin of the current from the cuprate layer can be understood by inspecting the individual terms of Eq. , which we refer to as ’tunneling channels’, i.e., from the regular and anomalous elements $G_{c \alpha}^{+} \Sigma{''}_{\alpha} G_{\alpha c'}^{-}$ and $F_{c \alpha}^{+} \Sigma{''}_{\alpha} F_{\alpha c'}^{-}$, of the Green’s function. \[Although tunneling channels are a normal state property, the anomalous matrix elements play an important role in generating the coherence peaks and thus are relevant more generally.\] For simplicity, we assume that the tip is right above a Bi atom. The dominant element of the filter function $M_{t,c}$ is then between the tip orbital and the $d_{z^2}$ orbital of the upper layer Cu atom lying beneath the surface Bi atom, so we take $c=c'=$ Cu-$d_{z^2}$ in results shown in Figs. \[partspectra1\] and \[partspectra2\]. Fig. \[partspectra1\] shows the relative contributions of the regular and anomalous matrix elements. The near Fermi energy current is primarily associated with the $d_{x^2-y^2}$ matrix elements. While the regular matrix elements of Cu-$d_{x^2-y^2}$ are almost solely responsible for the spectrum at energies around the VHSs, the [*anomalous*]{} elements determine the features around the gap region, especially the [*coherence peaks*]{}. Fig. \[partspectra1\] shows that coherence peaks are inherited from the anomalous and not the regular part of the Green’s function, and reflect physically the effects of non-conservation of the number of electrons near the gap region. ![(color online) Partial spectrum with $c=c'=d_{z^2}$ in $M_{tc}.$ The regular (red line) and anomalous (green line) components are shown together with the total contribution of the two parts (solid black). Blue curve shows the corresponding regular Cu-$d_{x^2-y^2}$ contributions. []{data-label="partspectra1"}](fig8.pdf){width="50.00000%"} In Fig. \[partspectra2\], the current of $d_{x^2-y^2}$ character is further broken down into contributions from various neighbors of the central Cu atom of the first and second CuO$_2$ layer away from the free surface. We see in panel (a) that the upper CuO$_2$ layer is more important than the lower one, but that the upper layer is by no means dominant. It seems that the coupling between the tip and the lower layer is strengthened via the relatively large overlap between the $d_{z^2}$ orbitals of the central Cu atoms of the two layers, which opens an important interlayer channel. The $d_{x^2-y^2}$ orbitals of the two layers mix not only to induce the well-known bilayer splitting in Bi2212, but also play a significant role in the flow of current to the tip from the lower cuprate layer. ![(color online) (a) Various contributions to tunneling spectrum from the regular matrix elements (assuming $c=c'=d_{z^2}$ in $M_{tc}$), $T^{R}_{t \alpha}$, of Cu-$d_{x^2-y^2}$ orbitals of upper and lower CuO$_2$ layer. Contributions from the nearest neighbor (nn) Cu atoms in the upper and lower layer are shown. (b) Same as (a), except this panel refers to the contributions from the anomalous matrix elements, $T^{A}_{t \alpha}$. []{data-label="partspectra2"}](fig9.pdf){width="50.00000%"} It can be seen from Fig. \[partspectra2\] (a) that the $d_{x^2-y^2}$ orbitals of the four nearest-neighbor Cu atoms of the central Cu give a significant contribution to the total spectrum, but that this amounts to only about one third of the contribution from all $d_{x^2-y^2}$ terms from the upper layer. Due to the non-local nature of the Bloch-states within the cuprate layers, it is clear then that the total signal involves long range contributions, and attributing the spectrum merely to the four nearest neighbor Cu atoms provides only a rough approximation. Anomalous contributions are considered in Fig. \[partspectra2\] (b). Here, the upper and lower layers give an almost equally large contribution, indicating that coherence peaks also are not all that local in character. Notably, we find a finite onsite [*anomalous*]{} contribution of $d_{x^2-y^2}$ even though the regular term is zero. This can be understood with reference to Eq. . Consider the term $$F_{z^2 d}= G_{e,z^2 (d+x_i)}^{0}\Delta_{(d+x_i) d} G_{h,d d},$$ where $d$ is shorthand for $d_{x^2-y^2}$ of the central Cu and $d+x_i$ is $d_{x^2-y^2}$ of the neighboring Cu in either $x$- or $y$-direction. Clearly, $G_{e,z^2 (d+x_i)}^{0}$ transforms under rotations of $\frac{\pi}{2}$ in the same way as $\Delta_{(d+x_i) d},$ and since $G_{h,d d}$ is an onsite term, the combination is invariant. Hence the four terms in the sum over the neighbors are equal, yielding a non-zero onsite term. We emphasize that the [*anomalous contribution*]{} of the four neighboring Cu atoms is quite small. Let us consider the term $$F_{z^2 (d+x_{i})}= G_{e,z^2d}^{0}\Delta_{d (d+x_i)} G_{h,(d+x_i) (d+x_i)}.$$ Due to symmetry, $G_{e,z^2d}^{0} = 0,$ and thus this term vanishes. However, there are terms like $$F_{z^2 (d+x_{i})}=G_{e,z^2 (d+2x_i)}^{0}\Delta_{(d+2x_i) (d+x_i)} G_{h,(d+x_i) (d+x_i)}$$ which do not vanish, but are very small, since $G_{e,z^2 (d+2x_i)}^{0}$ is a relatively small term. A similar analysis can be carried out for the second and third neighbors. The second nearest neighbors, which lie along the nodal direction in k-space, give the largest single contribution, although this contribution is not dominant. The third neighbor contribution is a little larger than the onsite contribution. Further Comments ================ Symmetry Analysis ----------------- The selection rules can be formalized using group theoretical arguments related to the filtering function. [@Balatsky]. For example, in order to explain the dominance of the $d_{x^2-y^2}$ orbitals of the four neighboring Cu atoms, considering representations of the two-dimensional $C_{4v}$ group, the d-orbitals $\vert d_{x^2-y^2},i\rangle$ of the site $i$ participate in eigenfunctions of the system as a linear combination $$\displaystyle{\sum_{i}} e^{-i\mathbf{k}\cdot \mathbf{R}_i} \vert d_{x^2-y^2},i\rangle.$$ This combination of the four neighboring orbitals at $(0,\pm \pi)$ and $(\pm \pi, 0)$ belongs to the same representation of $C_{4v}$ as the $4s$ and $d_{z^2}$ orbitals of the central Cu atom (see Fig. \[phases\]), as well as the $p_z$ orbitals of the apical oxygen and the surface Bi atom. At this k-point, the phase difference between the lattice sites causes all the d-orbital lobes pointing towards the central atom to have the same sign. Hence, this combination yields a large off-diagonal element overlap with the surface $p_z$-orbital, and a dominant tunneling contribution around the gap. Similar arguments can be applied to understand contributions from other farther out atoms. An example was given in Fig. \[channels08\](b) above where the position of the tip and the symmetry of the relevant orbital strongly influence the visibility of an atom. ![(color online) Relative phases of the central $d_{z^2}$ orbital and the neighboring $d_{x^{2}-y^{2}}$ orbitals at the $\Gamma$ point (a) and at the $M$ point (b).[]{data-label="phases"}](fig10.pdf){width="45.00000%"} Electron extraction/injection ----------------------------- To relate the tunneling current to the LDOS of the cuprate layer, we have introduced the concept of tunneling paths through Eq. , which implies that each path [*begins or ends on a particular atomic orbital*]{}. This non-intuitive concept requires some comment. In reality, the current flows through the sample with each electron ejected to the tip being replaced by an electron from a distant counterelectrode. For a simple system, such as a nanostructure, non-equilibrium Green’s function formalism with two ‘leads’ closing a current circuit have been invoked (see, e.g., Ref. ). Tersoff-Hamann (TH) or Todorov-Pendry (TP) approach, on the other hand, assumes that the current is composed of a series of tunneling events[@THTPfoot], and that the replacement of electrons at the counterelectrode has a negligible effect on the tunneling process. Since the current in STS is of the order of $10-100pA$, there is only about one electron each $1-10ns$ which flows across the sample, justifying the assumptions underlying TH/TP approach. Both TH and TP are based on calculating individual tunneling events in a LEED-like formalism[@ShenRMP]. Due to the finite $\Sigma"$, an electron created on a particular atom will have only a finite probability of escaping to the tunneling tip, and Eq. shows how to add up the contribution of all these tunneling processes in terms of the equilibrium LDOS of the sample. Conclusions =========== We have presented a comprehensive framework for modeling the STS spectra from the normal as well as the superconducting state of complex materials in a material-specific manner. Our formulation makes transparent the connection between the LDOS and the STS spectrum or the nature of the tunneling ’matrix element’, and it is cast in a form that reveals the filtering effect of the overlayers separating the tip and the layers of interest. Our decomposition of the tunneling current into contributions from individual local orbitals allows us to identify important ’tunneling channels’ or paths through which current reaches the STM tip in the system. Our analysis highlights the importance of anomalous terms of the Green’s function, which account for the formation and breaking up of Cooper pairs, and how such terms affect the STS spectrum. We apply the formalism to the specific case of Bi2212. Mismatch of symmetry between orbitals on adjacent atoms, or between the tip and the sample orbitals, is shown to severely restrict the corresponding contribution to the tunneling current. For these reasons, the contribution from Cu-$d_{x^2-y^2}$ orbitals comes not directly from the Cu-atom lying right below the Bi atom, but from a fourfold symmetric indirect route involving the four nearest-neighbors of the central Cu as well as longer range background from farther out Cu-$d_{x^2-y^2}$ orbitals. In the superconducting state, the coherence peaks of the spectrum are shown to be dominated by the anomalous spectral terms, which also are found not to be all that localized around the central Cu atom. In particular, we find a small anomalous on-site term and a practically vanishing first nearest neighbor contribution, with most of the anomalous contribution arising from the second neighbors and beyond. We have concentrated in this study on the large hole doping regime of the cuprates where a homogeneous electronic Fermi liquid phase is consistent with most experiments. The fact that we have obtained good overall agreement between our computations and the measurements, especially with respect to the pronounced asymmetry of the spectrum between positive and negative bias voltages, indicates that this remarkable asymmetry can be understood more or less within our conventional picture without the need for invoking exotic mechanisms. At lower dopings, strong correlation effects including the possible presence of competing orders or inhomogeneous electronic states (nanoscale phase separation) would need to be taken into account. However, the present framework can be extended fairly straightforwardly through the addition of Hubbard terms in the Hamiltonian to provide a viable scheme for investigating the tunneling response throughout the phase diagram of the cuprates and other complex materials, including the modeling of effects of impurities and dopant atoms in the system. [**Acknowledgments**]{} This work is supported by the US Department of Energy, Office of Science, Basic Energy Sciences contract DE-FG02-07ER46352, and benefited from the allocation of supercomputer time at NERSC, Northeastern University’s Advanced Scientific Computation Center (ASCC), and the Institute of Advanced Computing, Tampere. RSM’s work has been partially funded by the Marie Curie Grant PIIF-GA-2008-220790 SOQCS. I.S. would like to thank the Wihuri Foundation for financial support. Conversations with Jose Lorenzana and Matti Lindroos are gratefully acknowledged. Boson-electron coupling ======================= In the vicinity of the Fermi energy, dispersion anomalies are found in ARPES spectra arising from coupling of electronic degrees of freedom to phonons and/or magnetic modes, often giving the appearance of a peak-dip-hump feature[@kinks]. These boson-electron couplings also strongly affect the STS spectrum[@VHS]. This appendix discusses a model self-energy for describing such anomalies. A significant contribution to the electron-phonon coupling is associated with modulation of the electronic hopping integrals by the phonons. The generalized coordinate of atomic displacement in $q$-basis is quantized in the standard way: $$Q_{q}= \sqrt{\frac{\hbar}{2 \Omega_q}}\left(a_q + a^{\dagger}_q \right),$$ where $a_{q} (a^{\dagger}_{q})$ is the annihilation (creation) operator of the phonon mode $q$, and $\Omega_q$ is the frequency of the mode. However, the most natural way to couple this to real-space tight-binding basis is to make a transformation to the basis of real space displacement of atom $\mu$ in the following way: $$\hat{u}_{\mu} = \langle \mu \vert q \rangle Q_{q},$$ where Einstein summation over phonon modes $q$ is implicit. Note, that $\mu$ is a composite index denoting both the index of an atom and the direction of displacement. Consequently, in tight-binding basis, this gives rise to a term in the Hamiltonian of the form $$H^{el-vib} = \frac{1}{\sqrt{m_{\mu}}} \frac{\partial V_{\alpha \delta}}{\partial R_{\mu}} \hat{u}_{\mu} c^{\dagger}_{\alpha} c_{\beta}= \Gamma_{\mu}^{\alpha \delta} \hat{u}_{\mu} c^{\dagger}_{\alpha} c_{\beta}$$ where $V_{\alpha \delta}$ is the hopping integral between orbitals $\alpha$ and $\delta$, $ R_{\mu}$ is the coordinate of atom $\mu$. This coupling can be embedded into the electronic Hamiltonian as an energy dependent self-energy. Following the arguments of Ref. , the general form of self-energy is written as: $$\begin{aligned} \Sigma^{\pm}_{\alpha \beta} (\varepsilon) = \frac{\hbar}{2} \Gamma_{\mu}^{\alpha \delta} \Gamma_{\nu}^{\gamma \beta} \int d \Omega \frac{1}{\Omega} g_{\mu \nu}(\Omega) \nonumber \\ ( (1 - f(\varepsilon-\hbar \Omega) + n_{b}(\Omega)) G_{ \delta \gamma}^{\pm} (\varepsilon-\hbar \Omega) \nonumber \\ + ( f(\varepsilon + \hbar \Omega) + n_{b}(\Omega) ) G_{\delta \gamma}^{\pm} (\varepsilon+\hbar \Omega) ), \label{self}\end{aligned}$$ where $g_{\mu \nu}(\Omega) = \sum_{q} \langle \mu \vert q \rangle \delta(\Omega - \Omega_q) \langle q \vert \nu \rangle$ is an element of the vibration mode density matrix. Note again that we use Einstein summation convention, so that summation is implied over orbital indices $\gamma$ and $\delta$ and the phonon polarization indices $\mu$ and $\nu.$ For simplicity, we now assume that: (i) The bosonic coupling only affects the Cu-$d_{x^2-y^2}$ orbitals, where we include a diagonal self-energy of the form, $g(\Omega) = g \Omega^2$ when $\Omega \le \Omega_{d}$ and it is 0 when $\Omega > \Omega_{d}$. For a Debye spectrum of phonons, $\Omega_{d}$ is the Debye cut-off frequency, and the normalization factor is $g=3/\Omega_{d}^{3}$; (ii) $\rho_{\delta \gamma} = -\frac{1}{\pi} Im[G^{+}_{\delta \gamma}] = \rho$ is approximately a constant. This amounts to assuming that the electronic density of states is smoothly varying within the range of the phononic spectrum; (iii) Take $\Gamma_{\mu}^{\alpha \delta} = \Gamma$, a constant parameter. Using these assumptions, the final form for the self-energy is $$\Sigma^{+} = -\frac{A}{\pi} \left( (2z+i\pi) + \left(z^2 - 1\right) \ln{\left(\frac{z - 1} {z + 1} \right)} \right), \label{sigmaeinstein} %\end{align*}$$ where $z=(\varepsilon + i \eta)/(\hbar \Omega_d)$, $A = \frac{3 \hbar}{4 \Omega_d} \Gamma^2 \rho,$ and $\eta$ is a convergence parameter. Although we have derived the preceding form for coupling to a 3D Debye spectrum of phonons, the results are not too sensitive to details of the spectrum, and we would expect a similar result for an Einstein phonon or the magnetic resonance mode[@magres]. It is interesting to consider the asymptotic forms of self-energy as follows. If $\hbar \Omega_d \ll \varepsilon$, $$\Sigma(\varepsilon) \approx -A \left( \frac{2}{\pi z} +i \right)$$ For large boson energies, i.e., $\hbar \Omega_d \gg \varepsilon$, we obtain $$\Sigma(\varepsilon) \approx -A \left( \frac{4}{\pi}z +i z^2 \right). \label{asympt}$$ While Eq. gives a general form of phononic self-energy for any pair of orbitals, in the present calculations, we adopt a few simplifications. First, we assume only diagonal terms of self-energy to make the model more tractable. Second, we apply Eq. to Cu-$d_{x^2-y^2}$ orbitals using parameters $\hbar \Omega_d = 80meV$ and $A = 60meV.$ The former value gives the best fit to the peak-dip-hump structure, and the latter controls the smoothness of the spectrum. In Fig. \[pristine\](c) we make a comparison between the more general form of Eq. with the accurate density of states of Cu-$d_{x^2-y^2}$ orbitals. For the remaining orbitals we mimic a Fermi-liquid type self-energy, which can be modeled with a $\Sigma'' \propto \varepsilon^2$ and $\Sigma' \propto \varepsilon$; here we employ the asymptotic form of Eq. , choosing parameters $\hbar \Omega_d = 2.0eV$ (to ensure the correct asymptotic form for whole the energy range) and $A = 100meV.$ In this way, the need for a Kramers-Kronig transformation is avoided. We can straightforwardly include in the self-energy the effect of magnon scattering[@MSB] responsible for the high energy kink[@HEK]. This will broaden the spectrum in the vicinity of the VHS peaks, thereby improving agreement with experiment in Fig. \[pristine\](a). It should be noted, however, that a more accurate modeling of the self-energy will be required both for the bosonic coupling and the Fermi-liquid term for treating the underdoped system. Bogoliubov quasiparticles in tight-binding basis ================================================ This appendix discusses aspects of the Bogoliubov transformation within a tight-binding basis. The Bogoliubov transformation is not explicitly carried out in the present calculations since the Green’s function tensor is obtained directly from Dyson’s equation. Nevertheless, understanding the relation between the transformation and the Green’s function tensor in the tight-binding basis is necessary for interpreting some of our results. In particular, our analysis of pairing symmetry is based on the relation between $F_{\alpha \beta}$ and $\langle c_{\alpha \uparrow} c_{\beta\downarrow} \rangle .$ The Bogoliubov transformation [@bogoliubov] is conventionally carried out in a combined basis of spin-up electrons and spin-down holes: $$\mathbf{c}_{k} = \left( \begin{array}{c} c_{k \uparrow}\\ c^{\dagger}_{-k \downarrow}. \end{array} \right)$$ These $c$’s diagonalize the one-particle Hamiltonian of Eq. via the transformations $$c_{\alpha \uparrow} = \langle \alpha \vert k \rangle c_{k \uparrow}$$ and $$c^{\dagger}_{\alpha \downarrow} = \langle -k \vert \alpha \rangle c^{\dagger}_{-k \downarrow} = \langle \alpha \vert k \rangle c^{\dagger}_{-k \downarrow},$$ or in a more compact form: $$\mathbf{c}_{\alpha} = \left( \begin{array}{c} c_{\alpha \uparrow}\\ c^{\dagger}_{\alpha \downarrow}. \end{array} \right) =\left( \begin{array}{cc} \langle \alpha \vert k \rangle & 0\\ 0 & \langle \alpha \vert k \rangle \end{array} \right) \mathbf{c}_{k} = B_{\alpha k} \mathbf{c}_{k},$$ with inverse $\mathbf{c}_{k} = B_{ k \alpha} \mathbf{c}_{\alpha}.$ This change of basis diagonalizes the one-particle Hamiltonian: $$\varepsilon_{k} = \langle k \vert \alpha \rangle H_{1,\alpha \beta} \langle \beta \vert k \rangle$$ (with summation over $\alpha$ and $\beta$). In this basis the Hamiltonian of Eq.  becomes $$H = \varepsilon_{k} c^{\dagger}_{k \uparrow}c_{k \uparrow} + \varepsilon_{k} (1-c_{-k \downarrow} c^{\dagger}_{-k \downarrow}) + \Delta_{k} c^{\dagger}_{k \uparrow}c^{\dagger}_{-k \downarrow} + \Delta^{\dagger}_{k}c_{-k \downarrow} c_{k \uparrow},$$ now with summation over $k.$ After shifting this by a constant energy, it assumes the simple form $$H^{eff} = \mathbf{c}^{\dagger} \hat{H} \mathbf{c},$$ where $$\hat{H} = \left( \begin{array}{cc} \varepsilon_{k} & \Delta_{k}\\ \Delta^{\dagger}_{k} & -\varepsilon_{k} \end{array} \right),$$ which can be diagonalized into $$H^{eff} = \mathbf{c}^{\dagger}U^{-1} U \hat{H}U^{-1} U \mathbf{c},$$ where $$U = \left( \begin{array}{cc} u^{*}_{k} & v_{k}\\ -v^{*}_{k} & u_{k} \end{array} \right)~~\textrm{and}~~ U^{-1} = \left( \begin{array}{cc} u_{k} & -v_{k}\\ v^{*}_{k} & u^{*}_{k} \end{array}. \right)$$ The coefficients are chosen in the standard way in order to obtain a diagonal matrix $$U \hat{H}U^{-1} = \left( \begin{array}{cc} E_{k} & 0\\ 0 & -E_{k} \end{array} \right),$$ with $E_{k} = \sqrt{\varepsilon_k^2 + \vert \Delta_k \vert^2}.$ This Bogoliubov transformation introduces the quasi-particle basis $$\mathbf{a} = \left( \begin{array}{c} a_{k}\\ b^{\dagger}_{-k} \end{array} \right) = U \mathbf{c}.$$ Since we are working in the tight-binding basis, we end up with $$\left( \begin{array}{c} a_{k}\\ b^{\dagger}_{-k} \end{array} \right) = \left( \begin{array}{cc} u^{*}_{k} \langle k \vert \alpha \rangle & v_{k} \langle k \vert \beta \rangle\\ -v^{*}_{k} \langle k \vert \alpha \rangle & u_{k} \langle k \vert \beta \rangle \end{array} \right) \left( \begin{array}{c} c_{\alpha \uparrow}\\ c^{\dagger}_{\beta \downarrow} \end{array} \right)$$ (summation over $\alpha$ and $\beta$) or inversely $$\left( \begin{array}{c} c_{\alpha \uparrow}\\ c^{\dagger}_{\beta \downarrow} \end{array} \right) = \left( \begin{array}{cc} \langle \alpha \vert k \rangle u_{k} & -\langle \alpha \vert k \rangle v_{k}\\ \langle \beta \vert k \rangle v^{*}_{k} & \langle \beta \vert k \rangle u^{*}_{k} \end{array} \right) \left( \begin{array}{c} a_{k}\\ b^{\dagger}_{-k} \end{array} \right)$$ (summation over $k$). We are particularly interested in writing the expectation values of electron and hole densities, $\langle c^{\dagger}_{\alpha \sigma}c_{\beta \sigma}\rangle,$ and $\langle c_{\alpha \sigma}c^{\dagger}_{\beta \sigma} \rangle,$ and pairing amplitudes $\langle c^{\dagger}_{\alpha \uparrow}c^{\dagger}_{\beta \downarrow} \rangle,$ and $\langle c_{\beta \downarrow} c_{\alpha \uparrow} \rangle$ in terms of the Green’s function tensor. For this purpose, we start with a $2\times 2$ tensor $$\langle \mathbf{c}_{\alpha} \mathbf{c}^{\dagger}_{\beta} \rangle = \langle B_{\alpha k} \mathbf{c}_{k} \mathbf{c}^{\dagger}_{k}B_{k \beta} \rangle = \langle B_{\alpha k} U^{-1}\mathbf{a}\mathbf{a}^{\dagger} U B_{k \beta}\rangle. \label{Ntensor}$$ Using the fact that $\langle a_k a^{\dagger}_k \rangle = 1-f(E_k)$ and $\langle b^{\dagger}_k b_k \rangle = f(E_k)$, we evaluate each element of the tensor $ \langle \mathbf{c}_{\alpha} \mathbf{c}^{\dagger}_{\beta} \rangle $ separately as follows:\ (1) The number density $$\langle c^{\dagger}_{\alpha \uparrow}c_{\beta \uparrow} \rangle = \langle \beta \vert k \rangle \left( \vert u \vert^2 f(E_k) + \vert v \vert^2 (1-f(E_k)) \right) \langle k \vert \alpha \rangle$$ Now we use a trick following Ref. where $$\langle \beta \vert k \rangle \vert u \vert^2 f(E_k) \langle k \vert \alpha \rangle = \int d\varepsilon f(\varepsilon) \langle \beta \vert k \rangle u \delta(\varepsilon - E_{k}) u^{*} \langle k \vert \alpha \rangle$$ and $$\delta(\varepsilon - E_{k}) \approx -\frac{1}{\pi}Im(\frac{1}{\varepsilon -E_{k} +i \eta}).$$ Hence $$\langle \beta \vert k \rangle \vert u \vert^2 f(E_k) \langle k \vert \alpha \rangle = \int d\varepsilon f(\varepsilon) \rho^{e}_{\beta \alpha}(\varepsilon),$$ where $$\rho^{e}_{\beta \alpha}(\varepsilon) = -\frac{1}{\pi}Im(G^{+}_{e, \beta \alpha}(\varepsilon)),$$ where $G^{+}_{e, \beta \alpha}$ refers to the electron part of the Green’s function, $$G^{+}_{e, \alpha \beta}(\varepsilon) = \frac{\langle \alpha, e \vert k \rangle \langle k \vert e, \beta \rangle}{\varepsilon - E_k + i \eta} = \frac{\langle \alpha \vert k \rangle \vert u_{k} \vert^{2} \langle k \vert \beta \rangle} {\varepsilon - E_k + i \eta}.$$ It is straightforward to show that $$\langle c^{\dagger}_{\alpha \uparrow}c_{\beta \uparrow} \rangle = \int d\varepsilon [f(\varepsilon) \rho^{e}_{\beta \alpha}(\varepsilon) + (1-f(\varepsilon)) \rho^{h}_{\beta \alpha}(\varepsilon)],$$ where $\rho^{h}_{\alpha \beta}$ is the hole density matrix. The first part of the integral, in fact, gives the number of electrons with a chosen spin. The latter part gives the same result as the former since the Bogoliubov transformation reflects the electron bands to hole bands with respect to the Fermi energy;\ (2) The pairing amplitude $$\langle c_{\alpha \uparrow}c_{\beta \downarrow} \rangle = \langle \alpha \vert k \rangle \left( u (f(E_k) - (1-f(E_k)) v \right) \langle k \vert \beta \rangle .$$ Using the trick of Ref. again gives us the formula $$\langle c_{\alpha \uparrow}c_{\beta \downarrow} \rangle = -\int d\varepsilon (1-2f(\varepsilon)) \rho^{eh}_{\alpha \beta}(\varepsilon), \label{ccF}$$ where $$\rho^{eh}_{\alpha \beta}(\varepsilon) = -\frac{1}{\pi}Im(F^{+}_{\alpha \beta}(\varepsilon)),$$ and $$F^{+}_{\alpha \beta}(\varepsilon) = \frac{\langle \alpha, e \vert k \rangle \langle k \vert h, \beta \rangle}{\varepsilon - E_k + i \eta} = \frac{\langle \alpha \vert k \rangle u_{k} v_{k} \langle k \vert \beta \rangle}{\varepsilon - E_k + i \eta}$$ In the same manner, one can see that $$\langle c^{\dagger}_{\alpha \uparrow}c^{\dagger}_{\beta \downarrow} \rangle = -\int d\varepsilon (1-2f(\varepsilon)) \rho^{eh\dagger}_{\beta \alpha}(\varepsilon), \label{ccF1}$$ where $$\rho^{eh\dagger}_{\beta \alpha}(\varepsilon) = -\frac{1}{\pi}Im((F^{+})^{\dagger}_{\beta \alpha}(\varepsilon)),$$ Equations and also reveal how the anomalous part of the Green’s function tensor is related to the pairing amplitude $\langle c_{\alpha \uparrow}c_{\beta \downarrow} \rangle$ in a tight-binding basis, or equivalently how the anomalous part of the current is related to the making and breaking of Cooper pairs. In particular, symmetry properties of $F_{\alpha \beta}$ are seen to be related directly to those of $\langle c_{\alpha \uparrow}c_{\beta \downarrow} \rangle$. [99]{} Ø. Fischer, M. Kugler, I. Maggio-Aprile, and Chr. Berthod, and Chr. Renner, [*Rev. Mod. Phys.*]{} [**79**]{}, 353 (2007). K. McElroy, Jinho Lee, J.A. Slezak, D.-H. Lee, H. Eisaki, S. Uchida, and J.C. Davis, [*Science*]{} [**309**]{}, 1048 (2005). E.W. Hudson, K.M. Lang, V. Madhave, S.H. Pan, H. Eisaki, S. Uchida, and J.C. Davis, [*Nature*]{} [**411**]{}, 920 (2001). S.H. Pan, E.W. Hudson, K.M. Lang, H. Eisaki, S. Uchida, and J.C. Davis, [*Nature*]{} [**403**]{}, 746(2000). A.N. Pasupathy, A. Pushp, K.K. Gomes, C.V. Parker, J. Wen, Z. Xu, G. Gu, S. Ono, Y. Ando, and A. Yazdani, [*Science*]{} [**320**]{}, 196 (2008). A.V. Balatsky, , A. V., Vekhter, I., and Zhu, J.-X., [*Rev. Mod. Phys.*]{} [**78**]{}, 373 (2006). Y. Kohsaka, C. Taylor, K. Fujita, A. Schmidt, C. Lupien, T. Hanaguri, M. Azuma, M. Takano, H. Eisaki, H. Takagi, S. Uchida, and J.C. Davis, [*Science*]{} [**315**]{}, 1380 (2007). I. Martin, A.V. Balatsky, and J. Zaanen, [*Phys. Rev. Lett.*]{} [**88**]{}, 097003 (2002). B.W. Hoogenboom, C. Berthod, M. Peter, Ø.  Fischer, and A.A. Kordyuk, [*Phys. Rev. B*]{} [**67**]{}, 224502(2003). J.A. Nieminen, H. Lin, R.S. Markiewicz, and A. Bansil, [*Phys. Rev. Lett.*]{} [**102**]{},037001 (2009). T.N. Todorov, G.A.D. Briggs and A.P. Sutton, [*J.Phys.: Condens. Matter*]{} [**5**]{}, 2389 (1993). J.B. Pendry, A.B. Prêtre and B.C.H. Krutzen, [*J.Phys.: Condens. Matter*]{} [**3**]{}, 4313 (1991). J. Tersoff and D.R. Hamann, [*Phys. Rev. B*]{} [**31**]{}, 805 (1985). Note that the tunneling signal decays exponentially with layer distance from the tip, and therefore, we expect the results presented in this article to be essentially the same as for a semi-infinite solid. V. Bellini, F. Manghi, T. Thonhauser, and C. Ambrosch-Draxl, [*Phys. Rev. B*]{} [**69**]{}, 184508(2004). J.C. Slater and G.F. Koster, [*Phys. Rev.*]{} [**94**]{}, 1498 (1954). W.A. Harrison, [*Electronic Structure and Properties of Solids.*]{} Dover, New York (1980). L. Shi and D. A. Papaconstantopoulos, [*Phys. Rev. B*]{} [**70**]{}, 205101 (2004). A. Bansil and M. Lindroos, [*Phys. Rev. Lett.*]{} 83, 5154(1999). M. Lindroos, S. Sahrakorpi and A. Bansil, [*Phys. Rev. B*]{} [**65**]{}, 054514(2002) A. Bansil, M. Lindroos, S. Sahrakorpi, and R.S. Markiewicz, [*Phys. Rev. B*]{} [**71**]{}, 012503(2005). R.S. Markiewicz, S. Sahrakorpi, M. Lindroos, Hsin Lin, and A. Bansil, [*Phys. Rev. B*]{} [**72**]{}, 054519(2005). M.C. Asensio, J. Avila, L. Roca, A. Tejeda, G. D. Gu, M. Lindroos, R. S. Markiewicz, and A. Bansil, [*Phys. Rev. B*]{} [**67**]{}, 014519(2003). A. Bansil and M. Lindroos, [*Journal of Physics and Chemistry of Solids*]{} [**59**]{}, 1879(1998). H. Lin, S. Sahrakorpi, R.S. Markiewicz, and A. Bansil, [*Phys. Rev. Lett.*]{} [**96**]{}, 097001 (2006). K. K. Gomes, A. N. Pasupathy, A. Pushp, S. Ono, Y. Ando, and A. Yazdani, Nature [**447**]{}, 569-572(2007). A. Kaminski et al., [*Phys. Rev. B*]{} [**73**]{}, 174511(2006). J.A. Nieminen and S. Paavilainen, [*Phys. Rev. B*]{} [**60**]{}, 2921 (1999). In practice, we calculate the Green’s function $G^{e}_{\mathbf{k} \alpha \beta}$ for each k-point separately to produce the site-dependent Green’s function by inverse Fourier-transformation: $G^{e}_{i\alpha, j\beta} = \frac{1}{N_k} \sum_{k}G^{e}_{\mathbf{k} \alpha \beta} \exp{(-i\mathbf{k} \cdot \mathbf{R}_{ij})}.$ Here the shorthand notation, $G_{e,\alpha \beta} = G^{e}_{i\alpha, j\beta}$, is used in that indices $\alpha$ and $\beta$ implicitly contain the simulation cell index. Note also, that the inverse transformation must not be done until solving the whole Green’s function tensor. J.-M. Tang and M. E. Flatté, [*Phys. Rev. B*]{} [**66**]{}, 060504(R) (2002); J.-M. Tang and M. E. Flatté, [ *Phys. Rev. B*]{} [**70**]{}, 140510(R) (2004). A.L. Fetter and J.D. Walecka, [*Quantum Theory of Many-Particle Systems.*]{} Dover (2003). F.C. Zhang and T.M. Rice, [*Phys. Rev. B*]{} [**37**]{}, 3759(1988). The d-wave and ZRS symmetries are not uniquely determined by the present choice of pairing. For instance, we could choose an onsite pairing at the oxygen $p_{x/y}$ orbitals with $\Delta_{xx} = -\Delta_{yy}.$ In that case, Eq. could be written in the $x$-direction as, $$F_{ \alpha \beta} = G_{e,\alpha x} \Delta_{xx} G^{0}_{h,x\beta},$$ with a corresponding expression in the $y$-direction. If $\alpha = d$ and $\beta = d\pm x$, the sign of $G_{e,\alpha x}G^{0}_{h,x \beta}$ is totally determined by the product of the lobes of the $d$ orbitals of the neighboring Cu atoms. This is, however, invariant under rotation by $\pi/2$, and thus, $F_{d,d\pm x /d\pm y}$ follows the symmetry of $\Delta_{xx / yy}.$ Y. Meir and N.S. Wingreen, [*Phys. Rev. Lett.*]{} [**68**]{}, 2512 (1992). H. Ness and A.J. Fisher, [*Phys. Rev. B*]{} [**56**]{}, 12469 (1997). T. Frederiksen, M. Paulsson, M. Brandbyge, and A.-P. Jauho, [*Phys. Rev. B*]{} [**75**]{}, 205413 (2007). In the present study, the decomposition of Eq. into tunneling channels has been significantly elaborated beyond our recent work in Ref. . The most significant improvement is the explicit formulation of anomalous tunneling channels. Furthermore, we can obtain the total contribution of any chosen orbital over the whole infinite slab via the relation $ \sum_{i=0}^{\infty} G^{+}_{0 c i \alpha}\Sigma{''}_{i \alpha}G^{-}_{i \alpha 0 c'} = \frac{1}{N_k} \sum_{\mathbf{k}} G^{+}_{\mathbf{k} c \alpha}\Sigma{''}_{\alpha}G^{-}_{\mathbf{k} \alpha c'}$ to regular and anomalous terms of Eq. . Note that the simulation cell indices $0$ and $i$ are explicitly written in the left hand side. M. Magoga and C. Joachim, [*Phys. Rev. B*]{} [**59**]{}, 16011 (1999). P. Sautet, [*Surf. Sci.*]{} [**374**]{}, 374 (1997). E. Niemi and J. Nieminen, [*Chem. Phys. Lett.*]{} [**397**]{}, 200 (2004). The distinct VHS peaks in the computed spectrum are expected to be broadened to yield a smooth hump much like the experimental spectrum due to self-energy corrections resulting from magnetic response of the electron gas in the -400 meV range. These self-energy corrections are not included in the present calculations. R.S. Markiewicz, S. Sahrakorpi, and A. Bansil, [*Phys. Rev. B*]{} [**76**]{}, 174514 (2007). Note that in solving the Dyson’s equation, the initial Green’s function for each k-point is a diagonal matrix of complex Lorentzians. An infinite number of k-points would be required for a final Green’s function without any artificial “shoulders”, although the imaginary part of the self-energy acts to smooth away the unwanted structures. A. Lanzara et al., [*Nature*]{} (London) [**412**]{}, 510 (2001); X. J. Zhou et al., [*Phys. Rev. Lett.*]{} [**95**]{}, 117001 (2005). A. Kaminski et al., [*Phys. Rev. Lett.*]{} [**86**]{}, 1070 (2001); P. D. Johnson et al., ibid. [**87**]{}, 177007 (2001); S. V. Borisenko et al., ibid. [**90**]{}, 207001 (2003); A. D. Gromko et al., [*Phys. Rev. B*]{} [**68**]{}, 174520 (2003). G. Levy de Castro, Chr. Berthod, A. Piriou, E. Giannini, and Ø. Fischer, [*Phys. Rev. Lett.*]{} [**101**]{}, 267004 (2008). L. Bartels, G. Meyer and K.-H. Rieder. [*Appl. Phys. Lett.*]{} 71, 213 (1997). J. Nieminen, E. Niemi, K.-H. Rieder, [*Surface Science*]{} [**552**]{}, L47-L52 (2004). Although TP and TH are formulated in terms of LDOS, both approaches can, in principle, be decomposed into a spectral function form since the density matrix can be decomposed in this way regardless of the basis set employed. A.P. Shen, [*Rev. Mod. Phys.*]{} [**4**]{}, 382 (1971). M. Paulsson, T. Frederiksen, and M. Brandbyge, [*Phys. Rev. B*]{} [**72**]{}, 201101(R) (2005). Z.-X. Shen and J.R. Schrieffer, [*Phys. Rev. Lett.*]{} [**78**]{}, 1771 (1997); M.R. Norman and H. Ding, [*Physical Review B*]{} [**57**]{}, R11089 (1998); S. LaShell, E. Jensen, and T. Balasubramanian, [*Phys. Rev. B*]{} [**61**]{}, 2371 (2000) F. Ronning, K.M. Shen, N.P. Armitage, A. Damascelli, D.H. Lu, Z.-X. Shen, L.L. Miller, and C. Kim, [*Phys. Rev. B*]{} [**71**]{}, 094518 (2005); J. Graf, G.-H. Gweon, K. McElroy, S.Y. Zhou, C. Jozwiak, E. Rotenberg, A. Bill, T. Sasagawa, H. Eisaki, S. Uchida, H. Takagi, D.-H. Lee, and A. Lanzara, [*Phys. Rev. Lett.*]{} [**98**]{}, 067004 (2007). M. Tinkham, [*Introduction to Superconductivity.*]{} McGraw-Hill International Editions (1996). A.P. Horsfield, A.M. Bratkovsky, M. Fearn, D.G. Pettifor, and M. Aoki, [*Phys. Rev. B*]{} [**53**]{}, 12694(1996).

--- abstract: | Long-lived high-energy ($>100 {\rm MeV}$) emission, a common feature of most Fermi-LAT detected gamma-ray burst, is detected up to $\sim10^2$ s in the short GRB 090510. We study the origin of this long-lived high-energy emission, using broad-band observations including X-ray and optical data. We confirm that the late $> 100$ MeV, X-ray and optical emission can be naturally explained via synchrotron emission from an adiabatic forward shock propagating into a homogeneous ambient medium with low number density. The Klein-Nishina effects are found to be significant, and effects due to jet spreading and magnetic field amplification in the shock appear to be required. Under the constraints from the low-energy observations, the adiabatic forward shock synchrotron emission is consistent with the later-time ($t \gtrsim 2{\rm s}$) high-energy emission, but falls below the early-time ($t<2 {\rm s}$) high energy emission. Thus we argue that an extra high energy component is needed at early times. A standard reverse shock origin is found to be inconsistent with this extra component. Therefore, we attribute the early part of the high-energy emission ($t \lesssim 2 \rm s$) to the prompt component, and the long-lived high energy emission ($t\ga 2\rm s$) to the adiabatic forward shock synchrotron afterglow radiation. This avoids the requirement for an extremely high initial Lorentz factor. author: - 'Hao-Ning He, Xue-Feng Wu, Kenji Toma, Xiang-Yu Wang and Peter' title: On the High Energy Emission of the Short GRB 090510 --- INTRODUCTION ============ Gamma-ray bursts (GRBs) are the most luminous explosions in the universe. Their basic scenario based on the emission from extremely relativistic outflows with bulk Lorentz factors $~10^{2}-10^{3}$ and isotropic energies of $10^{48}-10^{55} \rm{erg}$ has been tested, although many questions remain open. The Fermi satellite has advanced our knowledge of GRBs significantly, while raising some new puzzles. During its first $\sim2 {\rm yr}$ of operation as of July 27th, 2010, Fermi has observed 19 GRBs with photons detected in the LAT (Large Area Telescope) instrument. These observations reveal three new properties [@Granot2010; @Abdo080825C; @Abdo080916C; @Abdo081024B; @Abdo090510; @Abdo090902B; @Ackermann090217A; @Ackermann090510; @Ackermann090926A]: i) A delayed high energy emission, e.g., in GRB 080916C, GRB 081024B, GRB 090510, GRB 090902B and GRB 090926A[^1]. ii) A temporally extended high energy emission: at least 10 of the first 19 Fermi LAT GRBs have long-lived high energy emission, lasting much longer than the burst duration in the sub-MeV band (which declines very rapidly); in 4 out of 10 GRBs, the long-lived LAT light curves have a relatively steeper slope, for example, $-1.33\pm0.08$ for GRB 080916C, $-1.70\pm0.08$ for GRB 090510, $ -1.40\pm0.06$ for GRB 090902B, $-2.05\pm0.14$ for GRB 090926A according to @Zhang2010catalog. iii) A deviation from a pure Band spectral function, showing an extra component in GRB 090510, GRB 090902B, GRB 090926A. These are among the brightest bursts, but the observations are compatible with the hypothesis of having such a component also in the other, less bright bursts, where it is harder to detect [[@Granot2010]]{}. Among these 19 Fermi LAT GRBs, GRB 090510 is a short, hard burst, with a duration $T_{90}=0.30\pm0.07\rm s$[@DePasquale2010; @Ukwatta2009], located at a redshift $z=0.903\pm0.003$ [@McBreen2010; @Rau2009]. It has been detected by Fermi [@Guiriec2009; @Ackermann090510], AGILE [@Longo2009], Swift [@Hoversten2009], Konus-Wind [@Golenetskii2009] and Suzaku [@Ohmori2009]. Thus, a large amount of high-quality broadband information is available on this burst, including optical, X-ray, MeV and GeV emission. The Swift BAT instrument triggered on GRB 090510 at $T_0^{\rm BAT}=00:23:00.4$ UT, May 10th, 2009 [@Hoversten2009], while the GBM instrument onboard Fermi triggered on at $T_0^{\rm F}=00:22:59.97$ UT, May 10th, 2009 [@Abdo090510]. Thus, there is a deviation between the two trigger times, which is $\Delta T_0=T_0^{\rm BAT}-T_0^{\rm F} =0.43 {\rm s}$. Hereafter we adopt the BAT trigger time $T_0^{\rm BAT}$ as a natural start time $T_0$ for computing the afterglow evolution, this being the onset of the main burst. The high energy emission of GRB 090510 has all three of the new features we summarized above: i) the bulk of the photons above 30 MeV arrive $253\pm 34 \rm ms$ later than those below $1\rm MeV$ [@Abdo090510]; ii) the high energy emission above $100\rm MeV$ shows a simple power law decay lasting $200 {\rm s}$ with a temporal decay index $\alpha_{\rm LAT}=1.38\pm0.07$ [@DePasquale2010] [^2]; iii) the time-integrated spectrum from $T_0+0.07 \rm s$ to $T_0+0.57\rm s$ is best fit by a Band function and a power-law spectrum [@Abdo090510]; the extra power-law component photon index of $-1.62\pm 0.03$ can fit the data well up to the highest-energy ($31 \rm GeV$) photon [@Abdo090510]. The XRT observations alone give a spectral index $0.57\pm 0.08$ [@Hoversten2009], while a detailed analysis of the temporal XRT emission combined with the LAT emission indicates a spectral index $\beta_{\rm X}$ ranging from $0.51$ to $0.81$[@DePasquale2010], and a temporal decay index $\alpha_{{\rm X},1}=0.74\pm0.03$ before a break time $t_{\rm X,b}=1.43\rm ks$, which subsequently steepens to $\alpha_{{\rm X},2}=2.18\pm0.10$. The optical emission initially rises with a temporal index $\alpha_{{\rm opt},1}=-0.5^{+0.11}_{-0.13}$, and after a break time $t_{\rm opt,b}=1.58^{+0.46}_{-0.37}\rm{ks}$ it decays with a temporal index $\alpha_{{\rm opt},2}=1.13^{+0.11}_{-0.10}$ [@DePasquale2010]. Most of the models aimed at explaining the long-lived high energy emission of GRB 090510 have favored the view that the high energy photons arise from the afterglow emission, being generated via synchrotron emission in the external forward shock [e.g., @Kumar2009080916C; @Kumar20093GRBs; @corsi2010; @DePasquale2010; @Gao2009; @Ghirlanda2010; @Ghisellini2010; @Wang2010; @Razzaque2010] . This explanation is fairly natural, since an external forward shock model can account for, at least in its gross features, not only for the observed delay of the $>100 \rm{MeV}$ photons, which corresponds to the deceleration time-scale of the relativistic ejecta, but also for the long lasting $>100{\rm MeV}$ emission, which can be attributed to the power-law decay of the synchrotron external forward shock emission [@Kumar2009080916C; @Kumar20093GRBs]. However, various of the above cited authors use somewhat different readings of the publicly available spectral and temporal slope data, which lead them to favor different explanations for the rapid decay of the long-lived high energy emission, falling into five different classes of models as follows. One set of models interprets the LAT emission as synchrotron emission of electrons accelerated in a standard adiabatic ISM forward shock [@Kumar2009080916C; @Kumar20093GRBs; @corsi2010; @DePasquale2010]. These authors argue that they can explain the LAT, X-ray, and optical data with plausible parameter values (and we revisit these arguments below). However, these authors did not [perform]{} a complete enough study to confirm whether the whole LAT data including the first $1$ second can be explained by this type of models. @Kumar2009080916C [@Kumar20093GRBs] fit the late LAT data as $F_{\nu } \propto t^{-1.2}$, but do not give much significance to the early-time LAT data. @corsi2010 fit the whole LAT data as $t^{-1.3}$, which seems to explain the early-time LAT data but exceeds the $2\sigma$ upper limits at late times ($\sim 100-1000s$). @DePasquale2010 suggest a steeper decay slope of the high energy emission (evolving as $t^{-1.38}$) for the whole LAT data by taking a larger electron distribution index $p=2.5$, but they don’t take into account in sufficient depth the early-time LAT data ($t<1{\rm s}$), which is necessary in order to conclude what is the origin of the entire LAT emission (see more discussion in Section 6). Other models, e.g., @Neamus2010 attribute the high-energy photons to synchrotron-self-Compton scattering (SSC) from an adiabatic forward shock propagating into a wind-like medium; this, however, requires an extremely small magnetic energy fraction $\epsilon_B=10^{-10}$. Another, different adiabatic forward shock model analyses in greater detail the Klein-Nishina (KN) effects on the high energy inverse Compton process [@Wang2010]. For some reasonable parameters, the KN effect, as it weakens in time, results in the synchrotron high energy emission being increasingly suppressed by the SSC cooling, which steepens the synchrotron high energy emission decay slope by a factor as large as $0.5$. A fourth model views the high-energy emission as decaying proportional to $t^{-1.5}$, which is interpreted as being caused by synchrotron emission of electrons accelerated in a forward shock in the radiative, rather then adiabatic, regime. For this, the electron population must be significantly enriched, which is attributed to pair production between back-scattering photons and prompt outward-going photons [@Ghisellini2010]. This model explains the high energy emission without considering the constraints from lower energy, e.g., GBM band, XRT band and UVOT band emission, and the pair formation becomes inefficient at shock radii larger than $10^{16}{\rm cm}$, while $R_{\rm dec}\gg10^{16}{\rm cm}$ for this burst (see §4). A fifth type of model for the afterglow of this burst (and others) is a hadronic model [@Razzaque2010], which explains the high energy emission as proton synchrotron emission, while attributing the low energy emission to electron synchrotron emission from a forward external shock. This requires a large total kinetic energy $E_{\rm k,iso}=2\times10^{55}\rm erg$[^3], with a low radiation efficiency and an extremely small fraction of electron energy $\epsilon_e=10^{-4}$. In this article, we re-examine the first set of models (the standard adiabatic forward shock model) in significantly greater detail than hitherto. We present detailed arguments indicating that it is most likely that the forward shock synchrotron emission can only explain the LAT emission from $\sim2-3$ sec. This conclusion disfavors the external shock origin of the early-time LAT emission [@Ghisellini2010; @Kumar2009080916C], but supports the suggestions that it is related to the prompt emission [@corsi2010; @DePasquale2010]. In $\S$ 2, we examine the XRT and UVOT observations, and set up a model of the long-lived emission based on synchrotron emission from electrons accelerated in a forward shock in a uniform ambient environment, including these X-ray and optical/UV observations. We discuss the impact of the Klein-Nishina effects on the high energy emission, under the constraints imposed by the lower energy observations, which are found to be significant for suppressing the SSC cooling. In $\S 3$ we use a semi-analytical model to calculate the development of the dynamical quantities of the forward shock across the deceleration time and into the self-similar phase, and use this to calculate the radiation properties of the long-lived high energy emission produced by synchrotron emission. We find a reasonable set of parameters which can explain most of the late afterglow, except for the six earliest LAT data points in the light curve. In $\S 4$ we check several possibilities for the origin of this early high energy emission at times $t<2-3 \rm s$. In $\S 5$ we discuss the possibility of the line-of-sight prompt emission as the origin of the early-time high-energy emission. In $\S 6$ we discuss our conclusions concerning the most probable origin of the high energy emission from the short GRB 090510. Forward shock model {#sec:FS} =================== Constraints from low energy emission of GRB 090510 {#sec:lowen-cons} -------------------------------------------------- The afterglow emission of GRB is generally well explained by synchrotron emission from electrons accelerated by the shock produced during a spherical relativistic shell colliding with an external medium. From the spectral index $\beta_{\rm X} \sim 0.51 - 0.81$ and the light curve slope $\alpha_{\rm X,1}=0.74\pm 0.03$ [@DePasquale2010], the closure relation for the X-ray afterglow $\alpha_{\rm X,1}-1.5\beta_{\rm X}$ ranges from $-0.51$ to $0.01$, suggesting a slow-cooling ISM external forward shock model with $\nu_{\rm m}^{\rm f}<\nu_{\rm X}<\nu_{\rm c}^{\rm f}$ [^4], which implies that the decay index of the X-ray light curve before the jet break is $\alpha_{\rm X,1}=3(p-1)/4$ and the X-ray spectral index is $\beta_{\rm X}=(p-1)/2$ [@Sari1998afterglow].[^5] From the spectral index we can get a constraint on $p$ which is $p \sim 2.0 - 2.6$. In the external shock model, the break seen in X-ray light curve can be explained as a jet break. We assume that the jet expands sideways (Sari et al. 1999). In this case the X-ray light curve slope steepens gradually from $-3(p-1)/4$ to $-p$, so that we have $3(p-1)/4 < \alpha_{\rm X,2} < p$, which is reduced to $2.2 < p < 3.9$ by taking $\alpha_{\rm X,2} \sim 2.2$. Combining the above two constraints we have $2.2 \lesssim p \lesssim 2.6$. Then the X-ray decay slope before the break should be $0.9 \lesssim \alpha_{\rm X,1} \lesssim 1.2$, which can be acceptable if we take into account the observed fluctuation of the X-ray flux, as the light curves we will show below. For the rising portion of the optical light curve before the break time, it is natural to assume that the optical band is below $\nu_{\rm m}^{\rm f}$, which induces an optical light curve slope of $t^{1/2}$ [@Sari1998afterglow; @Kumar20093GRBs]. The predicted spectral slope $\nu^{1/3}$ is consistent with the observations within the large error bars [@DePasquale2010]. Thus, one can try to explain both the X-ray emission and the optical emission before the break time with synchrotron emission of electrons accelerated in an adiabatic external forward shock with the assumptions $\nu_{\rm opt}<\nu_{\rm m}^{\rm f}$ and $\nu_{\rm m}^{\rm f}<\nu_{\rm X}<\nu_{\rm c}^{\rm f}$. Then, assuming that the optical band is in the regime of $\nu_{\rm m}^{\rm f}<\nu_{\rm opt}< \nu_{\rm c}^{\rm f}$ after the break time, the slope of the optical post-break light curve is the same as that of the X-ray light curve. In the analytical calculations of this section, we assume that the self-similar phase conditions have been established, and for simplicity we neglect the structure of the shock wave, considering a spherical shock with a total isotropic energy $E$ and a Lorentz factor $\Gamma$. At late times, the adiabatic dynamical evolution of the spherical shock is in the Blandford $\&$ Mckee self-similar phase, where $E=16\pi\Gamma^2R^3nm_pc^2/17$ is constant and the scaling law of the shock wave is $\Gamma\propto R^{-3/2}$ [@Blandford1976]. The shock propagates a distance $\delta R\sim 2\Gamma^2c\delta t/(1+z)$ during the small observing time $\delta t$ [@Sari1997dyn], and integrating this and using the scaling law, one obtains $t=(1+z)R/8\Gamma^2 c$. According to @Sari1998afterglow, the cooling and minimum Lorentz factors of the electrons in the forward shock depend on the total isotropic kinetic energy $E$, the number density of the external environment $n$, and the fraction of the electron energy and magnetic field energy $\epsilon_{e\rm f}$ and $\epsilon_{B\rm f}$, which can be expressed as $$\label{gammac} \gamma_{\rm c}^{\rm f}=\frac{6\pi(1+z)m_ec}{\sigma_{\rm T}B'^2\Gamma t}=2.0\times10^8\epsilon_{B\rm f,-4}^{-1}E_{53.5}^{-3/8}n_{-4}^{-5/8}t_2^{1/8}\left[{1+Y(\gamma_{\rm c}^{\rm f})}\right]^{-1}$$ and $$\label{gammam} \gamma_{\rm m}^{\rm f}=\epsilon_{e\rm f}\frac{p-2}{p-1}\frac{m_p}{m_e}\Gamma=9.0\times 10^4 \epsilon_{e\rm f,-0.4}E_{53.5}^{1/8}n_{-4}^{-1/8}t_2^{-3/8}g_p,$$ respectively, where $$\label{B} B'=(32\pi\epsilon_{B\rm f}\Gamma^2nm_pc^2)^{1/2}= 1.4\times10^{-2}\epsilon_{B\rm f,-4}^{1/2}E_{53.5}^{1/8}n_{-4}^{3/8}t_{2}^{-3/8}{\rm G}$$ and $g_p\equiv 3(p-2)/(p-1)$, $p$ is the power-law index of the electron energy distribution, $Y(\gamma_{\rm c}^{\rm f})$ is the Compton parameter of the electrons with Lorentz factor $\gamma_{\rm c}^{\rm f}$, and $\sigma_{\rm T}$ is Thomson cross section. The cooling and minimum frequencies of electrons are $$\label{hnuc} h\nu_{\rm c}^{\rm f}=h\frac{q_eB'}{2\pi m_ec}(\gamma_{\rm c}^{\rm f})^2\Gamma=2.3\times 10^{9}\left [{1+Y(\gamma_{\rm c}^{\rm f})}\right]^{-2}\epsilon_{B\rm f,-4}^{-3/2}E_{53.5}^{-1/2}n_{-4}^{-1}t_2^{-1/2} {\rm eV}$$ and $$\label{hnumin} h\nu_{\rm m}^{\rm f}=h\frac{q_eB'}{2\pi m_ec}(\gamma_{\rm m}^{\rm f})^2\Gamma=5.0\times 10^ {2}\epsilon_{e\rm f,-0.4}^{2} \epsilon_{B\rm f,-4}^{1/2}E_{53.5}^{1/2}t_2^{-3/2}g_p^2 {\rm eV},$$ respectively. The peak flux density of the forward shock synchrotron emission is $$\label{Fmax} F_{\rm max}^{\rm f}=\frac{(1+z)N_e^{\rm f}m_e c^2\sigma_{\rm T}B'\Gamma}{12\pi q_e d_{\rm L}^2}=137 \epsilon_{B\rm f,-4}^{1/2}E_{53.5}n_{-4}^{1/2}\rm{\mu Jy},$$ where the total number of the electrons that the forward shock swept up is $N_e^{\rm f}=\frac{16}{17}\pi R^3n=1.5\times10^{51}E_{53.5}^{3/4}n_{-4}^{1/4}t_2^{3/4}$, and $d_{\rm L}$ is the luminosity distance. We get two constraints from the UVOT and XRT data as follows:\ (i) The optical flux density is about $20 {\rm \mu Jy}$ at $t\sim 100 {\rm s}$ [@DePasquale2010], which indicates that $$\label{fopt} F_{\rm opt}(t=100\rm s)\simeq F_{\rm max}^{\rm f}\left(\frac{4.5{\rm eV}}{h\nu_{\rm m}^{\rm f}}\right)^{1/3} =28.5\epsilon_{ef,-0.4}^{-2/3}\epsilon_{Bf,-4}^{1/3}E_{53.5}^{5/6}n_{-4}^{1/2}g_{p}^{-2/3}{\rm \mu Jy}\sim20 {\rm \mu Jy}.$$ (ii) The X-ray flux density is about $20 {\rm \mu Jy}$ at $t\sim100 {\rm s}$, which indicates $$\label{fxrt} F_{\rm XRT}(t=100{\rm s})\simeq F_{\rm max}^{\rm f}\left(\frac{3000{\rm eV}}{\nu_m^{\rm f}}\right)^{-(p-1)/2} =35.7\epsilon_{ef,-0.4}^{p-1}\epsilon_{Bf,-4}^{(p+1)/4}E_{53.5}^{(p+3)/4}n_{-4}^{1/2}f_{e1}^{3.1}g_P^{p-1}{\rm \mu Jy}\sim20{\rm \mu Jy}.$$ Combining the above two equations (\[fopt\]) and (\[fxrt\]) , we can express the fraction of the magnetic field energy and the number density as $$\label{eBcon} \epsilon_{B\rm f,-4}\sim 0.66\epsilon_{e\rm f,-0.4}^{-4} E_{53.5}^{-1}g_p^{-4}g_{e2}^{12} ,$$ $$\label{ncon} n_{-4}\sim 0.66\epsilon_{e\rm f,-0.4}^{4} E_{53.5}^{-1}g_p^4g_{e2}^{-8.0}$$ where $g_{e1}=e^{p-2.5}$ and $g_{e2}=e^{(p-2.5)/(3p-1)}$ and we can naturally get $\epsilon_{e\rm f}>0.04E_{53.5}^{-1/4}g_{e2}^3g_p^{-1}$ from the condition $\epsilon_{B\rm f}<1$. Inserting equations (\[eBcon\]) and (\[ncon\]) into equation (\[B\]), the downstream magnetic field strength is thus constrained to be $$B'\sim 10\epsilon_{e\rm f,-0.4}^{-1/2}E_{53.5}^{-3/4}t_{2}^{-3/8}g_{p}^{-1/2}g_{e2}^{3.0}{\rm mG}.$$ If the shocks involve a magnetic field amplification factor $f_B\geq 1$ (in addition to shock compression), the upstream magnetic field strength would be $$B_{\rm u}=B'/(4\Gamma f_{B})\sim 7 f_{B}^{-1} E_{53.5}^{-1}g_{e2}^{2.0}{\rm \mu G}.$$ Thus, the upstream magnetic field strength could be $\lesssim 5-10 \mu$G [e.g., @Kumar20093GRBs], apparently compatible with shock-compression of the typical magnetic field in the interstellar medium, with no need of magnetic field amplification. However, as we discuss in §\[sec:disc\], the external density deduced here is much below the average interstellar value, and likely so would be the external magnetic field, so that additional field amplification may be needed. Inserting equations (\[eBcon\]) and (\[ncon\]) into equations (\[hnuc\]), (\[hnumin\]) and (\[Fmax\]), the characteristic energies and the peak flux density of synchrotron emission are therefore $$h\nu_{\rm c}^{\rm f}\sim 6.6\times 10^{9}\epsilon_{e\rm f,-0.4}^{2} E_{53.5}^2t_2^{-1/2}[1+Y(\gamma_{\rm c} ^{\rm f})]^{-2}g_p^2g_{e2}^{-10}{\rm eV},$$ $$\label{numin2} h\nu_{\rm m}^{\rm f}\sim4.0\times 10^2 t_2^{-3/2} g_{e2}^{6.0}{\rm eV},$$ and $$F_{\rm max}^{\rm f}\sim90 g_{e2}^{2.0} {\rm \mu Jy},$$ respectively. The above equations show that $\nu_{\rm opt}<\nu_{\rm m}^{\rm f}$ until $t\sim 1.4{\rm ks}$, and $\nu_{\rm m}^{\rm f}<\nu_{\rm X}<\nu_{\rm c}$ under the constraint $\epsilon_{e\rm f,-0.4}E_{53.5}>1.2\times10^{-3}t_{2}^{1/4}g_p^{-1}g_{e2}^{5.0}$, which is easy to satisfy, where $Y(\gamma_{\rm c}^{\rm f})\ll1$ as discussed in §\[sec;kn\]. Thus, they are consistent with our previous assumptions. The radiative efficiency in the assumed slow cooling regime is [@Sari2001] $$\label{epsilon} \eta_r=\epsilon_{e\rm f}\left(\frac{\nu_{\rm m}^{\rm f}} {\nu_{\rm c}^{\rm f}}\right)^{\frac{p-2}{2}}=6.3\times10^{-3}\epsilon_{e\rm f,-0.4}^{3-p} E_{53.5}^{2-p}t_2^{-(p-2)/2}g_{p}^{2-p}g_{e1}^{-8.3}g_{e2}^{8.0(p-2)},$$ which is much less than unity, consistent with our previous assumption of an adiabatic forward shock model. In order to check whether the LAT emission predicted by synchrotron emission from an adiabatic forward shock can explain the LAT observations we need to study two situations. \(i) Under the condition $\epsilon_{e\rm f,-0.4}E_{53.5}>0.39g_p^{-1}g_{e2}^{5.0}$, we yield $\nu_{\rm m}^{\rm f}<1{\rm GeV}<\nu_{\rm c}^{\rm f}$ at $t\la100{\rm s}$, the average synchrotron flux density from the forward shock in the LAT band ($100{\rm MeV}$ to $4 {\rm GeV}$) is $$\label{Flat} F_{\rm LAT}^{\rm f}\sim F_{\rm max}^{\rm f} \left(\frac{1{\rm GeV}}{h\nu_{\rm m}^{\rm f}}\right)^{-\frac{p-1} {2}}\sim 1.4\times 10^{-3} t_2^{-3(p-1)/4} g_{e1}^{-6.4} {\rm \mu Jy}.$$ which is independent of the electron energy $\epsilon_{e\rm f}E$. In addition, the slope of the LAT light curve is also constrained, as $\alpha_{\rm LAT}=3(p-1)/4$, the same as that of the X-ray light curve. For $p=2.5$, $\nu_{\rm m}^{\rm f}< 1{\rm GeV}<\nu_{\rm c}^{\rm f}$ at $t\la100{\rm s}$ is satisfied if $\epsilon_{e\rm f,-0.4}E_{53.5}>0.39$, then the predicted LAT flux is $1.4\times10^{-3}{\rm \mu Jy}$ at $100$ s, consistent with the observational data within the error bars at around $100$ s. The corresponding slope of the predicted LAT light curve is around $-1.125$, which can explain the late-time data of the LAT observation. If we take a small electron index, for example, $p\sim 2.2$, we have $\nu_{\rm m}^{\rm f}<1{\rm GeV}<\nu_{\rm c}^{\rm f}$ at $t=100{\rm s}$ if $\epsilon_{e \rm f,-0.4}E_{53.5}>0.60$. Then the predicted LAT flux can be calculated by equation (\[Flat\]), approximated as $\sim 8.9\times 10^{-3}{\mu \rm Jy}$ at $100$ s, which is almost one order of magnitude smaller than the observed LAT flux. What’s more, the slope of the predicted LAT light curve is about $-0.9$, which is too shallow to explain the late-time LAT observation. Thus, in the case $\nu_{\rm m}^{\rm f}<1{\rm GeV}<\nu_{\rm c}^{\rm f}$ at $100$ s, we can exclude the $p=2.2$ model. \(ii) If the electron energy satisfies the condition $\epsilon_{e\rm f,-0.4}E_{53.5} \leq0.12g_p^{-1}g_{e2}^{5}$, the lower end of the LAT band, $100$ MeV, is above the frequency $\nu_c^{\rm f}$, and the predicted LAT flux at $t=100{\rm s}$ can be calculated as $$\label{Flatsteeper} F_{\rm LAT}^{\rm f}\sim F_{\rm max}^{\rm f} \left(\frac{h\nu_{\rm c}^{\rm f}}{h\nu_{\rm m}^{\rm f}}\right)^{-\frac{p-1}{2}} \left(\frac{1{\rm GeV}}{h\nu_{\rm c}^{\rm f}}\right)^{-\frac{p}{2}}\sim 3.6\times10^{-3}\epsilon_{ef,-0.4}E_{53.5}t_{2}^{-(3p-2)/4}g_pg_{e1}^{-7.4} g_{e2}^{3p-6}{\rm \mu Jy}.$$ Inserting the condition $\epsilon_{e\rm f,-0.4}E_{53.5}\leq0.12g_p^{-1}g_{e2}^{5}$ into equation (\[Flatsteeper\]), we see that $F_{\rm LAT}^{\rm f}\leq4.3\times 10^{-4} g_{e1}^{-8.4}t_2^{-(3p-2)/4} {\rm \mu Jy}$. For $p=2.5$, the predicted LAT flux at $100$ s is $F_{\rm LAT}^{\rm f}(100{\rm s})\leq4.3\times 10^{-4}{\rm \mu Jy}$, which is almost one order of magnitude lower than the observed LAT flux. Thus in the case $\nu_{\rm LAT}>\nu_{\rm c}^{\rm f}$ at $100$ s, we can exclude the $p=2.5$ model. For $p=2.2$, to be consistent with the observed LAT flux at $100$ s, the condition $\epsilon_{ef,-0.4}E_{53.5}=6.3\times10^{-2}$ is required, constraining the electron energy to a very small value. Here we adopt fairly standard values of $\epsilon_{e\rm f}=0.16 - 0.6$, similar to those observed in long GRBs, since these are determined by collisionless shock physics processes on microphysical scales, which should be independent of the global properties of GRBs. Then the constrained total kinetic energy could be $(0.13-0.47)\times 10^{53}{\rm erg}$. Since the isotropic energy at $10{\rm keV}-30{\rm GeV}$ energy band during $T_0+0.03{\rm s}-T_0+0.53{\rm s}$ is $E_{\rm iso}=(1.08\pm0.06)\times 10^{53}{\rm erg}$ [@Ackermann090510], the radiative efficiency of the prompt emission is in the range $70\%-89\%$. Moreover, we constrain $\epsilon_{Bf,-4}=88\epsilon_{e,-0.4}^{-3}$ and $n_{-4}=1.0\epsilon_{e,-0.4}^{5}$ by inserting the constraint $\epsilon_{ef,-0.4}E_{53.5}=6.3\times10^{-2}$ into equations (\[eBcon\]) and (\[ncon\]). The constrained $\nu_{\rm c}^{\rm f}$ is $h\nu_{\rm c}^{\rm f}=1.1\times10^7{\rm eV}$ at $100{\rm s}$, and $\nu_{\rm c}^{\rm f}<10^8{\rm eV}$ when $t>1.2{\rm s}$. Since the expected slope of the LAT light curve is $\alpha_{\rm LAT}=(3p-2)/4=1.15$ for $p=2.2$ in the $\nu_{LAT}>\nu_{\rm c}^{\rm f}$ case, similar to that expected from the $p=2.5$ model in the $\nu_{\rm m}^{\rm f}<\nu_{\rm LAT}<\nu_{\rm c}^{\rm f}$ case, this can also explain the slope of the late-time LAT light curve. Even though $\nu_{\rm m}^{\rm f}$ decreases by a factor of $1.5$ by taking $p=2.2$ rather than taking $p=2.5$, it doesn’t change the optical break time by much, and we may not rule out the $p=2.2$ model in the $\nu_{\rm LAT}>\nu_{\rm c}^{\rm f}$ case. Thus, both the $p=2.5$ model in the $\nu_{\rm m}^{\rm f}<1{\rm GeV}<\nu_{\rm c}^{\rm f}$ case and the $p=2.2$ model in the $\nu_{\rm LAT}>\nu_{\rm c}^{\rm f}$ case can explain the late-time LAT light curve, leading to the same conclusion, since the predicted LAT light curves have the similar slope. For presentation purposes, hereafter, we discuss the $p=2.5$ model in the $\nu_{\rm m}^{\rm f}<1{\rm GeV}<\nu_{\rm c}^{\rm f}$ case. Impact of Klein-Nishina effects on the constraints {#sec;kn} -------------------------------------------------- @Wang2010 have studied the Klein-Nishina (KN) effects on high-energy gamma-ray emission in the early afterglow, and find that at early times the KN suppression on the IC scattering cross section for the electrons that produce the high-energy emission is usually strong, and therefore their inverse-Compton losses are small, with a Compton parameter $Y$ of less than a few for a wide range of parameter space. This leads to a relatively bright synchrotron afterglow emission at high energies at early times. However, as the KN effects weaken with time, the inverse-Compton losses increase and the synchrotron high energy emission is increasingly suppressed, which leads to a more rapid decaying synchrotron emission. This provides a potential mechanism for the steep decay of the high-energy gamma-ray emission seen in some Fermi LAT GRBs. The Compton parameter for electrons with Lorentz factor $\gamma_e$ is defined as the ratio of the synchrotron self-inverse Compton (SSC) to the synchrotron emissivity, i.e. $$Y(\gamma_e)\equiv\frac{P_{\rm SSC}(\gamma_e)}{P_{\rm syn}(\gamma_e)}.$$ When the KN suppression on the scattering cross section is negligible, $Y({\gamma_e})= Y(\gamma_c)$ is a constant for the slow-cooling case [@Sari2001]. However, for high energy electrons with a significant KN effect, $Y(\gamma_e)$ is no longer a constant and this affects the electron radiative cooling function, as well as the continuity equation of the electron distribution. The self-consistent electron distribution is given by $$\label{n_slow} N(\gamma_e)=\left\{{} \begin{array}{ll} C_1\gamma_e^{-p} \,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\, \gamma_m< \gamma_e<\gamma_c \\ \frac{1+Y(\gamma_c)}{1+Y(\gamma_e)}C_1\gamma_c \gamma_e^{-p-1} \,\,\,\,\,\, \gamma_c<\gamma_e \end{array}\right .$$ for the slow-cooling case [@Nakar2009; @Wang2010], where $C_1$ is a constant. The high energy synchrotron photons with energy $h\nu_*$ are produced by electrons with Lorentz factor $\gamma_*$ which typically have $\gamma_*>max(\gamma_{\rm c}, \gamma_{\rm m})$. Thus, the number density of electrons of $\gamma_*$ is $$N(\gamma_*)=\frac{1+Y(\gamma_c)}{1+Y(\gamma_*)}C_1\gamma_c \gamma_*^{-p-1}=\frac{N_{\rm syn}(\gamma_*)}{1+Y(\gamma_*)},$$ where $N_{\rm syn}(\gamma_*)=C_1\gamma_{c}[1+Y(\gamma_c)]\gamma_*^{-p-1}$ is the number density of electrons of $\gamma_*$ when only the synchrotron cooling is considered [@Sari1998afterglow]. Therefore, the number density of electrons with Lorentz factor $\gamma_*$ is a factor of $1+Y(\gamma_*)$ lower than that in the case where only the synchrotron cooling is considered. Thus the synchrotron luminosity is correspondingly reduced by the same factor. We have $Y(\gamma_*)\propto t^{1/2}$ , in the slow cooling regime, as long as $\gamma_*>\Gamma m_{e} c^2/h\nu_{\rm m}^{\rm f}$ [@Wang2010]. In that case, if $Y(\gamma_*)\gg 1$ the synchrotron luminosity is suppressed by the factor $Y(\gamma_*)$ which is in proportion to $t^{1/2}$, i.e. the light curve decay of the high energy synchrotron emission could be steepened by a factor $1/2$ at most. Meanwhile, the distribution of electrons which are in the region $\gamma_{\rm m}<\gamma<\gamma_{\rm c}$ is not affected by $Y(\gamma)$, due to equation \[n\_slow\], thus the lower energy synchrotron emission decay slope is normal. The critical photon energy above which the scattering with electrons of energy $\gamma_e$ just enters the KN scattering regime is defined as $h\nu_{\rm KN}(\gamma_e) \equiv \Gamma m_{e}c^2/\gamma_e$, i.e. $$h\nu_{\rm KN}(\gamma_{\rm c}^{\rm f})=\Gamma m_{e}c^2/\gamma_{\rm c}^{\rm f}=0.52\epsilon_{e\rm f,-0.4}^{-2}E_{53.5}^{-1}t_2^{-1/2}g_p^{-2}g_{e2}^{8.0}[1+Y(\gamma_{\rm c}^{\rm f})]{\rm eV}.$$ which is much smaller than $\nu_{\rm m}^{\rm f}$ under the condition $\epsilon_{e\rm f,-0.4}^{2}E_{53.5}>0.13t_4g_p^{-2}g_{e2}^{2.0} [1+Y(\gamma_{\rm c}^{\rm f})]$. In this case the synchrotron-self Compton scattering is strongly suppressed due to KN effects, and $$\begin{aligned} Y(\gamma_{\rm c}^{\rm f})\left[1+Y(\gamma_{\rm c}^{\rm f})\right] &=&\frac{\epsilon_{e\rm f}}{\epsilon_{B\rm f}}\left(\frac{\gamma_{\rm c}^{\rm f}}{\gamma_{\rm m}^{\rm f}}\right)^{2-p}\left(\frac{\nu_{\rm m}^{\rm f}}{\nu_{\rm c}^{\rm f}}\right)^{(3-p)/2}\left(\frac{\nu_{\rm KN} (\gamma_{\rm c}^{\rm f})}{\nu_{\rm m}^{\rm f}}\right)^{4/3} \nonumber \\ &=&1.4\times 10^{-4}\left[1+Y(\gamma_{\rm c}^{\rm f})\right]^{7/3}\epsilon_{e\rm f,-0.4}^{4/3}E_{53.5}^{-4/3} t_{2}^{5/6}g_{p}^{1/3}g_{e2}^{-1.3}.\end{aligned}$$ From the above equation, we find that $Y(\gamma_{\rm c}^{\rm f})\ll1$ under the condition $\epsilon_{e\rm f,-0.4}^{-1}E_{53.5}\gg 0.023t_4^{5/8}g_p^{1/4}g_{e2}^{-0.98}$, which is easy to be satisfied before $t<10^{4}{\rm s}$. Therefore, since $Y(\gamma_{*})\leqslant Y(\gamma_c)\ll1$, the distribution of electrons which contribute to the high-energy synchrotron photons doesn’t change, according to equation (\[n\_slow\]), which means that the decay slope of synchrotron high energy emission can not be affected by $Y(\gamma_{*})$. Forward Shock Synchrotron Emission Evolution {#sec:FSsyn} ============================================ We adopt the following parameters for the calculations. From equation (\[Flat\]), we see that the flux density in the LAT range is independent of the total isotropic kinetic energy and the electron energy fraction at $t<100{\rm s}$, under the constraint $\epsilon_{e\rm f,-0.4}E_{53.5}>0.39$. Thus, we can fix the total energy of the afterglow as $E=3.0\times10^{53} {\rm erg}$, indicating a radiation efficiency as $25\%$. Even if we choose other values of the total kinetic energy under the constraint $\epsilon_{e\rm f,-0.4}E_{53.5}>0.39$, our conclusion will not change. Then the fraction of electron energy is constrained as $\epsilon_{e\rm f}>0.16$. We adopt fairly standard values of $\epsilon_{e\rm f}\geq0.16$. From the equations (\[eBcon\]) and (\[ncon\]), we can get the values of $\epsilon_{B\rm f}=2.6\times 10^{-3} - 1.3\times10^{-5}$, corresponding to an external density range $\left[ n=1.7\times10^{-6} {\rm cm}^{-3} - 3.3\times10^{-4}{\rm cm^{-3}}\right]$. These densities range from much less than the intergalactic medium (IGM), up to a low density inter-cluster medium (ICM) or possibly galactic halo baryon density. The critical Lorentz factor which is the boundary between the thin and thick shell cases [@Kobayashi2007] is $\Gamma_c=[3(1+z)^3 E/32\pi nm_pc^5T^3]^{1/8}\sim 5.2 \times 10^3 E_{53.5}^{1/8}n_{-4}^{-1/8}T_{-0.5}^{-3/8}$, where the duration of GRB 090510 is $T=0.30\pm0.07{\rm s}$. For an initial Lorentz factor which is not too large, it is reasonable to assume that the initial Lorentz factor is smaller than the critical Lorentz factor, which means the shell has given the ambient medium an energy comparable to its initial energy at the deceleration time $t_{\rm dec}=\left[{3(1+z)^3E}/{32\pi c^5nm_p\Gamma_0}\right]^{1/3}>T$, indicating the thin shell case. Then the initial Lorentz factor of the forward shock can be expressed as a function of the deceleration time, which is $\Gamma_0=3.3\times 10^3 E_{53.5}^{1/8}n_{-4}^{-1/8} t_{\rm dec}^{-3/8}$. If, for example, we take an initial Lorentz factor as suggested by previous Fermi analyses of $2.2\times 10^3 - 5.2\times 10^3$, and a not too small number density of $n \sim 10^{-4}{\rm cm^{-3}}$, we would obtain a deceleration time of $t_{\rm dec}\sim 0.3 {\rm s} - 3 {\rm s}$. To calculate the dynamics of the evolution of the blast wave including the transition from the quasi-free expansion, through the deceleration and into the self-similar phase, we use the relativistic hydrodynamics equations for the evolution of the shock radius $R$, the mass $m$ swept up by the shock, the opening half-angle of jet $\theta$ and the Lorentz factor of the shock $\Gamma$ [e.g., @Huang2000apj] (see Appendix A), and solve these equations numerically. The solution of these equations provides the dynamical quantities which we use to calculate the radiation spectrum and the light curves discussed in the following. We take into account also the jet evolution as it goes through the jet break and starts to expand sideways. The half-angle of the jet evolves as $d\theta_{\rm j}=({c_{\rm s}dt'}/{r})=[{(1+\beta)\Gamma c_{\rm s}}/{r}]dt$, with a spreading velocity in the comoving frame approximated by the sound speed $c_{\rm s}$ [@Rhoads1997; @Rhoads1999; @Huang2000mn]. After the inverse Lorentz factor becomes larger than the initial jet angle, the spreading of the jet speeds up the shock deceleration significantly, which leads to a steeper light curve. In figures 1 and 2 we present model light curves in the LAT, XRT and UVOT bands for two choices of parameters. The forward shock light curves (solid black lines) are shown, for each density choice, for two different Lorentz factors and jet opening angles. A smaller opening angle is chosen for a larger Lorentz factor in order to get a certain jet break time. In figure 1, we show a solution with a density of $n=1\times10^{-6} {\rm cm}^{-3}$, which would correspond to a sub-average density IGM environment. Under the assumptions $\nu_{\rm opt}<\nu_{\rm m}^{\rm f}$ before the optical break time and $\nu_{\rm m}^{\rm f}<\nu_{\rm X}<\nu_{\rm c}^{\rm f}$, the steep decay of the latter parts of the X-ray and optical light curves can be explained using such a forward shock synchrotron emission going through a jet spreading phase, as shown in figure 1 and 2. Note, however, that, as in other analyses too, the late-time optical data are challenging to fit. This leads to a model forward shock light curve and spectral fit to the X-ray and optical observations satisfying all the constraints of §\[sec:lowen-cons\], for a set of parameters $\epsilon_{e\rm f}=0.17$, $E=3\times 10^{53}{\rm erg}$, $p=2.5$, $\epsilon_{B\rm f}=7.0\times 10^{-3}$, and $\theta_0=0.45^{\circ}$ for $\Gamma_0=4500$, or $\theta_0=0.38^{\circ}$ for $\Gamma_0=9000$. The collimation corrected energies are $4.6\times10^{48}$ and $3.3\times10^{48}$ erg. In figure 2 we show the forward shock solutions for a more moderate density of $n=10^{-3}{\rm cm}^{-3}$, corresponding to a galactic halo or interarm medium, which satisfy the low energy constraints as well as the afterglow epoch LAT, XRT and UVOT data points. The parameters are $\epsilon_{e\rm f}=0.6$, $\epsilon_{B\rm f}=1\times10^{-5}$, $p=2.5$ and $E=3\times 10^{53}$ erg, for a choice of $\theta_{0}=1.3^{\circ}$ and $\Gamma_0=1900$, and also for a choice of $\theta_{0}=1.1^{\circ}$ and $\Gamma_0=3800$. The collimation corrected energies are $3.9\times10^{49}$ and $2.8\times10^{49}$ erg, which is a typical energy for a short GRB. Despite the larger $\epsilon_{e\rm f}$ in this fit, the forward shock is still in the adiabatic regime according to equation (\[epsilon\]). In Figures 1 and 2 the LAT energy band light curves predicted by the forward shock synchrotron model shown in black solid lines (thicker for the larger Lorentz factor and smaller opening angle choice, thinner for the smaller Lorentz factor and larger angle choice). These appear to explain well the late time LAT emission, i.e. after 3 seconds, either with the larger or smaller angle/Lorentz factor, no matter how small the deceleration time is, but they fall below the first six LAT data points. This difference between the early time observed LAT data and the synchrotron forward shock emission suggests that there may be another radiation component (whose contribution can be represented by the gray dotted lines in Figures 1 and 2, corresponding to the high latitude emission of the prompt emission with variability timescale $\Delta t=0.5{\rm s}$), in addition to the forward synchrotron contribution (shown as the black solid lines), at least for its decaying portion. The gray solid lines are the sum of the steep decay component and the forward synchrotron emission with a larger Lorentz factor (the first, lower data point could be part of either a rising portion or a variable portion of the gray dotted component). Thus, it is possible that the early-time high-energy emission is not from the afterglow forward shock synchrotron emission, instead having a different origin. In the next section we discuss possible origins for the first six data points in the LAT band around the deceleration time. Other Possible Components around the Deceleration Time {#sec:tdec} ====================================================== Besides the forward shock synchrotron emission, around the deceleration time there are several other possible emission components which could contribute to the early LAT observations. Their importance can be estimated using approximate values for the characteristic quantities at the deceleration time. At the deceleration time, the Blandford-McKee self-similar solution is not yet applicable. For this we can take the usual value of the energy as $E=\frac{4 \pi}{3}R^3\Gamma^2nm_pc^2$ and $R=2\Gamma^2 ct/(1+z)$ as the radius at the deceleration time [@Sari1998afterglow], from which we obtain the initial Lorentz factor of the shock and the deceleration radius, which are $$\Gamma_{\rm 0}=3.3\times10^3 E_{53.5}^{1/8} n_{-4}^{-1/8} t_{\rm dec}^{-3/8}$$ $$R_{\rm dec}=3.5\times10^{17}E_{53.5}^{1/4}n_{-4}^{-1/4}t_{\rm dec}^{1/4}{\rm cm}.$$ Before we examine the other possible components around the deceleration time, we confirm that the forward shock synchrotron emission cannot explain the observation at early times in more details. We can then estimate the approximate high energy emission at the deceleration time under the previous constraints provided by the low energy observations. Here we take the Lorentz factor and radius at the deceleration time as $\Gamma_{\rm d}\sim\Gamma_{\rm 0}/\sqrt{2}$ and $R_{\rm d}\sim R_{\rm dec}/\sqrt{2}$, which are close to the values indicated by the numerical evolution of the previous section. Inserting the above $\Gamma_{\rm d}$ and $R_{\rm d}$ into the first expressions of equations (\[gammac\]) $-$ (\[Fmax\]), and meanwhile adopting equations (\[eBcon\]) and (\[ncon\]), the characteristic frequencies and peak flux density of the forward shock synchrotron emission at the deceleration time are then $$h\nu_{\rm m,d}^{\rm f}\sim6.8\times 10^5 t_{\rm dec}^{-3/2}g_{e2}^{6.0}{\rm eV},$$ $$h\nu_{\rm c,d}^{\rm f}\sim 3.9\times 10^{10}\epsilon_{e\rm f,-0.4}^{2} E_{53.5}^2t_{\rm dec}^{-1/2}g_p^2 g_ {e2}^{-10}{\rm eV},$$ $$F_{\rm max,d}^{\rm f}\sim12g_{e2}^{2.0} {\rm \mu Jy}.$$ @Ghisellini2010 suggests the radiative forward shock model to explain the steep temporal decay of high energy emission, which requires $\nu_{\rm m}^{\rm f}> \nu_{\rm c}^{\rm f}$ to get the fast cooling case. However, the radiative model cannot explain the shallower decay at later time ($t>10{\rm s}$) (as seen in Figure 7 in @Ghisellini2010), which agrees better with an adiabatic model in slow cooling case at that time. According to the constrained characteristic frequencies, the fast cooling case at the deceleration time requires that $$\epsilon_{e\rm f}<1.7\times10^{-3}E_{53.5}^{-1}t_{\rm dec}^{-1/2}g_{p}^{-1}g_{e2}^{8.0},$$ which is inconsistent with their assumption of a very high energy fraction of electrons $\epsilon_{e\rm f}=0.9$. This is because @Ghisellini2010 assume the late-time emission to be also in the fast cooling case, and do not take into account the constraints from the low energy emission. In addition, using equations (\[eBcon\]) and (\[ncon\]), we get an estimate for the forward shock synchrotron emission at $1{\rm GeV}$ at the deceleration time, $$\label{F_f} F_{1{\rm GeV,d}}^{\rm f}=F_{\rm max,d}^{\rm f}\left(\frac{1{\rm GeV}}{h\nu_{\rm m,d}^{\rm f}}\right)^{- \frac{p-1}{2}}\sim 0.049 t_{\rm dec}^{-3(p-1)/4} g_{e1}^{-2.6} {\rm \mu Jy}.$$ This is almost one order of magnititude below the observed LAT flux density, for the case when the deceleration time is smaller than $3 {\rm s}$ (this is seen also in Figures 1 and 2). The synchrotron forward shock emission cannot explain the early-time high-energy emission (the first six LAT data points near the deceleration time), although it can explain very well the late LAT results (beyond the sixth LAT data point, when the afterglow can be considered as established), which are consistent with the numerical results in §\[sec:FSsyn\]. Thus, we find that there has to be another component, contributing to the early high energy emission. In the rest of this section we consider several such possibilities, such as a reverse shock synchrotron component, a reverse shock SSC component or a cross IC component(forward/reverse shock synchrotron photons scattered by electrons from reverse-shocked/forward-shocked region), a high latitude (curvature) component of the prompt high energy emission. Reverse Shock Synchrotron Emission ---------------------------------- We consider the thin shell case as is assumed in §3, the flux peaks at the crossing time of the reverse shock $t_{\times}=t_{\rm dec}>T$ with Lorentz factor $\Gamma_{\times}=\Gamma_{\rm d}$, and the ratio of the comoving number densities of the forward-shocked to that of the reverse-shocked regions is given by ${n_{\rm r}}/{n_{\rm f}}\sim\Gamma_{\rm d}\sim 2.3\times 10^3 E_{53.5}^{1/8}n_{-4}^{-1/8}t_{\rm dec}^{-3/8}$ [@Kobayashi2007]. The internal energy densities $e$ and the bulk Lorentz factors $\Gamma$ of the two regions are equal with each other [@Zhang2003]. Consequently, we have that $$\frac{\nu_{\rm m}^{\rm r}}{\nu_{\rm m}^{\rm f}}=\left(\frac{n_{\rm r}}{n_{\rm f}}\right)^{-2}\Re_B\Re_e^2, ~~ \frac{\nu_{\rm c}^{\rm r}}{\nu_{\rm c}^{\rm f}}=\Re_B^{-3}\left[1+Y(\gamma_{c,\times}^{\rm r})\right]^{-2}, ~~\frac{F_{\rm max}^{\rm r}}{F_{\rm max}^{\rm f}}=\frac{n_{\rm r}}{n_{\rm f}}\Re_B,$$ characterized by the ratios $\Re_B\equiv\left(\frac{\epsilon_{B\rm r}}{\epsilon_{B\rm f}}\right)^{\frac{1}{2}}$, $\Re_e\equiv\frac{\epsilon_{e\rm r}}{\epsilon_{e\rm f}}$. The minimum and cooling frequencies of the reverse shock synchrotron emission at the crossing time under the low energy constraints are $$h\nu_{\rm m,\times}^r=1.1\epsilon_{e\rm f,-0.4}E_{53.5}^{-1/4}t_{\rm dec}^{-3/4}\Re_{B,1}\Re_{e,0}^{2} g_pg_{e2}^{4.0}{\rm eV},$$ $$h\nu_{\rm c,\times}^r=3.9\times 10^7\epsilon_{e\rm f,-0.4}^{2}E_{53.5}^{2}t_{\rm dec}^{-1/2}\Re_{B,1}^{-3} g_p^{2}g_{e2}^{-10}\left[1+Y(\gamma_{c,\times}^{\rm r})\right]^{-2} {\rm eV}.$$ According to §2.2, we have $Y(\gamma_*^{\rm r})<Y(\gamma_c^{\rm r})\ll1$ around the deceleration time, thus the synchrotron spectrum does not change and the ratio of the reverse shock synchrotron flux to forward shock synchrotron flux can be calculated as $$\label{fratio} \frac{F_{1 \rm GeV,\times}^{\rm r}}{F_{1\rm GeV,\times}^{\rm f}}=\frac{F_{\rm max,\times}^{\rm r}}{F_{\rm max,\times}^{\rm f}}\left(\frac{\nu_{\rm m,\times}^{\rm r}}{\nu_{\rm m,\times}^{\rm f}}\right)^{\frac{p-1} {2}}\left(\frac{h\nu_{\rm c,\times}^{\rm r}}{1 \rm GeV}\right)^{\frac{1}{2}}=\left(\frac{n_{\rm r}}{n_{\rm f}} \right)^{-p+2}\Re_{e}^{p-1}\Re_{B}^{(p+1)/2}\left(\frac{h\nu_{c,\times}^{\rm r}}{1\rm GeV}\right)^{1/2},$$ Inserting equations (\[eBcon\]) and (\[ncon\]) into equation (\[fratio\]), the ratio turns to be $$\frac{F_{1 \rm GeV,\times}^{\rm r}}{F_{1\rm GeV,\times}^{\rm f}}=0.23\epsilon_{e\rm f,-0.4}^{p/2}E_{53.5}^{- p/4+3/2}t_{\rm dec}^{3p/8-1}\Re_{B,1}^{p/2-1}\Re_{e,0}^{p-1}g_{e1}^{-6.7}g_p^{p/2}g_{e2}^{-p-3.0}.$$ The ratio could be as high as $0.59$ since the total kinetic energy can be as high as $E\sim9.0\times10^{53}{\rm erg}$ by considering a reasonable radiation efficiency $\gtrsim10\%$. Considering equation (\[F\_f\]), the flux density from reverse shock synchrotron emission cannot be as high as $\sim\rm\mu Jy$ to explain the observations at early times. Thus, the reverse shock synchrotron emission is unlikely to contribute to the early-time LAT observation. The SSC and EIC Emission ------------------------ Besides the reverse synchrotron emission, we consider the other four possible IC processes, including the synchrotron self-Compton (SSC) processes in forward and reverse shocks, and two combined-IC processes (i.e. scattering of reverse-shock synchrotron photons on electrons accelerated in forward shocks and froward-shock synchrotron photons on electrons accelerated in reverse shocks). The optical depth of inverse Compton scattering in the reverse shock in the Thomson regime is $\tau^{\rm r}=\frac{1}{3}\sigma_{\rm T} n_{\rm r} R_{\rm d}/\Gamma_d=7.7\times 10^{-8} \epsilon_{e\rm f,-0.4}^{5/2}E_{53.5}^{-1/4}t_{\rm dec}^{-1/8}g_p^{5/2}g_{e2}^{-5.0}$[@Wang2001]. The optical depth to inverse Compton scattering in the forward shock in the Thomson regime is $\tau^{\rm f}=\frac{1}{3}\sigma_{\rm T} n R_{\rm d}=3.8\times 10^{-12} \epsilon_{e\rm f,-0.4}^{3}E_{53.5}^{-1/2}t_{\rm dec}^{1/4}g_p^{3}g_{e2}^{-6.0}$. Taking the peak flux of synchrotron emission at the crossing time of the reverse shock and forward shock, which are $F_{\rm max}^{\rm r}=0.30\epsilon_{e\rm f,-0.4}^{-1/2}E_{53.5}^ {1/4} t_{\rm dec}^{-3/8}\Re_{B,1}g_p^{-1/2}g_{e2}^{3.0}{\rm Jy}$ and $F_{\rm max}^{\rm f}=12 g_{e2}^{2.0} {\rm \mu Jy}$, respectively, we can get the peak flux density of the four Inverse-Compton components as follows [@Wang2001], $$F_{\rm max}^{\rm IC,rr}=\tau^{\rm r}F_{\rm max}^{\rm r}=2.2\times 10^{-8}{\rm Jy}\epsilon_{e\rm f,-0.4}^{2}t_ {\rm dec}^{-1/2}\Re_{B,1}g_{p}^2g_{e2}^{-2.0},$$ $$F_{\rm max}^{\rm IC,rf}=\tau^{\rm r}F_{\rm max}^{\rm f}=8.8\times 10^{-13}{\rm Jy}\epsilon_{e\rm f,-0.4}^ {5/2}E_{53.5}^{-1/4}t_{\rm dec}^{-1/8}g_p^{5/2}g_{e2}^{-3.0},$$ $$F_{\rm max}^{\rm IC,fr}=\tau^{\rm f}F_{\rm max}^{\rm r}=1.1\times 10^{-12}{\rm Jy}\epsilon_{e\rm f,-0.4}^ {5/2}E_{53.5}^{-1/4}t_{\rm dec}^{-1/8}\Re_{B,1}g_p^{5/2}g_{e2}^{-3.0}$$ and $$F_{\rm max}^{\rm IC,ff}=\tau^{\rm f}F_{\rm max}^{\rm f}=4.5\times 10^{-17}{\rm Jy}\epsilon_{e\rm f,-0.4}^{3} E_{53.5}^{-1/2}t_{\rm dec}^{1/4}g_p^{3}g_{e2}^{-4.0},$$ respectively, where the superscripts ‘rr’ and ‘ff’ mean the SSC emission in reverse shock and forward shock respectively, ‘fr’ and ‘rf’ mean the scattering of reverse shock photons on the electrons in forward shocks and forward shock photons on the electrons in reverse shocks. Therefore, the IC contributions (except for the SSC emission in the reverse shock) can be excluded, since even their peak fluxes are much smaller than the early-time LAT observations. Although the peak flux of the SSC emission in the reverse shock is close to the observed flux, the reverse shock SSC emission can also be excluded because its flux peaks at a low energy of about $20 {\rm keV}$, indicating that the flux in the LAT band is about $6$ orders lower than the peak flux. A low flux for these IC processes is mainly due to a very low circum-burst density in GRB 090510 inferred from the low-energy observations. The High Latitude Prompt Emission --------------------------------- The spectrum of GRB 090510 shows a power-law component in the LAT band ($>100{\rm MeV}$), whose physical origin at early times is unclear. One might consider the early part of the extended emission (shown by the gray dotted line with [a decay slope $\alpha_{\rm Lat}(0.37-3{\rm s})=2.0$]{} in Figures 1 and 2) is due to the high-latitude emission of the prompt emission [@Kumar2000]. Because photons from high latitude regions with respect to the line of sight will arrive later than that from low latitude region due to the curved front surface of the jet, one observes a fast decreasing emission rather than an abrupt stop of the emission, the so-called “curvature effect". Then the high latitude emission flux of the prompt emission evolves as $$\label{cur} F_{\nu}(t) \propto\left[\frac{t-(t_0-\Delta t)}{\Delta t}\right]^{-2-\beta}$$ according to @Toma2009cocoon(The details are shown in Appendix B), where $t_0$ is the pulse peak time. According to @Ackermann090510 and @Abdo090510, the variability timescale derived from the BGO emission ($100{\rm keV}-$few MeV) is about $14\pm2{\rm ms}$ in the time interval $0.17-0.37{\rm s}$ since the BAT trigger, but the variability timescale for the LAT emission is not determined. If we assume that the variability timescale for the high energy ( larger than $100{\rm MeV}$) emission, which is dominated by the power law component, is the same as that of BGO observation, then we adopt $\Delta t\sim0.01{\rm s}$ in equation (\[cur\]). Thus, the high latitude flux of the prompt emission at $t_0+0.1 {\rm s}$ is $F(t_0+0.1 {\rm s})=11^{-2-\beta}F(t_0)$. Since the index of the power law component in the time interval $0.17-0.37{\rm s}$ is $1.66\pm0.04$ [@Ackermann090510; @Abdo090510], we take the rough value $\beta\sim0.66$. Then the flux of the high-latitude emission at $t_0+0.1{\rm s}$ will be reduced by a factor of $\sim600$ relative to that of the line-of-sight emission at $t_0\sim 0.37$, which is too steep to explain the early high energy emission at $t=0.37-3{\rm s}$. Since the origin of the LAT emission in the prompt phase, dominated by the extra power-law component, is unclear and can be different from that of the Band component in the GBM energy range, the variability timescale in the LAT range can generally be different from $\Delta t \sim 0.01{\rm s}$. In particular, the emission radius of the extra high energy component may be much larger than that of the Band component, which may lead to a much larger pulse width in the LAT energy range than in the GBM energy range (see e.g., @Toma2010). If we take $\Delta t\sim 0.5{\rm s}$, then the high latitude flux of the prompt emission at $t_0\sim0.37{\rm s}$ evolves as $F(t)\sim(2t+0.26{\rm s})^{-2.66}F(0.37{\rm s})$, which together with the synchrotron forward shock emission with a deceleration time $t_{\rm dec}\leq(2-3){\rm s}$ can explain the early high energy emission at $t\sim0.37-3{\rm s}$, shown in Fig 1. The Line-of-Sight Prompt Emission as the Origin of the Early-Time Afterglow High-Energy Emission ================================================================================================ As a final possibility, we consider the possible influence of the tail-end of the prompt emission in the description of the early afterglow. In §3 and §4 we assumed a small GRB duration $T_{90}=0.30\pm0.07{\rm s}$, following @DePasquale2010 based on the observations in the GBM and BAT bands. Under this assumption, we discussed the forward shock synchrotron emission, the forward reverse shock synchrotron emission, the four possible crossed inverse Compton processes, and the high latitude emission of the prompt emission, and we concluded that none of them can explain the early decaying part of the high energy emission at $0.3\lesssim t \lesssim 3{\rm s}$ except for the high latitude portion of the prompt emission with a larger variability time scale plus the synchrotron forward shock emission with a deceleration time $t_{\rm dec}\leq(2-3){\rm s}$. We focus on the BAT detections around $T_0 + (1-3)$ s, shown in @DePasquale2010, which implies that the prompt emission could last until such times. We calculate the flux of the synchrotron forward shock emission at $t=3$s for $p=2.5$ in the BAT energy range, and obtain $\sim 52 {\rm \mu Jy}$ (the flux is similar for $p=2.2$). This is about one order of magnitude smaller than the observed flux. Thus these BAT detections are not from the external shock, but may be attributed to the prompt emission. The LAT prompt emission may also last until such times. @Ackermann090510 obtain no significant temporal correlation between NaI ($8{\rm keV}-260{\rm keV}$) data and the LAT data $>100{\rm MeV}$ between $t=0.17$ s and $t=0.47$ s. This does not necessarily indicate that the LAT prompt emission (the extra component) and the GBM prompt emission (Band component) have fully unrelated origins. Some theoretical models suggest that the two components may arise at different radii (e.g., @Toma2010, @Wang2006) in the same shells, or the two components may arise by the leptonic and hadronic processes in the same internal shock (e.g., @Asano2009; @Razzaque2010). Therefore it appears reasonable to consider that GRB 090510 has a duration as long as $2-3$ s and the high energy emission before $t\sim 2-3$ s is a part of the prompt emission. Under this new assumption, with a longer GRB duration $T\sim2.0{\rm s}$, one infers a critical Lorentz factor $\Gamma_c\sim 2.0\times 10^3 E_{53.5}^{1/8}n_{-3}^{-1/8} T_{0.3}^{-3/8}$. For the thin shell case with an initial Lorentz factor $\Gamma_0$ smaller than $\Gamma_c$, the shell will be decelerated efficiently by the reverse shock after the GRB main duration. The high energy light curve with the initial Lorentz factor $\Gamma_0=1900$ and the number density $n=10^{-3} {\rm cm^{-3}}$ is shown by the thin solid line in Figure 2, which can explain the late-time high energy emission well. If we assume that the LAT emission around $t=2{\rm s}$ is a flare of duration $\Delta t=1{\rm s}$, which peaks at $t_0=2{\rm s}$ with a peak flux $F(t_0)$, and having [@DePasquale2010] a spectral index $\beta=0.41^{+0.28}_{-0.31}$ at $t=1.5-2.5{\rm s}$, then the flux due to curvature effect at $t=3.5{\rm s}$ is $F(t)=0.11F(t_0)$, which is compatible with the seventh data point. In this case, the deceleration time could in principle be even larger, implying a smaller Lorentz factor, but due to the lack of precise information about the last flare, e.g., its duration and the flux peak time, we cannot use this to derive a firmer deceleration time. For a larger initial Lorentz factor $\Gamma_0>\Gamma_{c}$, the reverse shock will transition from the Newtonian phase to the relativistic phase at the time $t_{\rm N}=\left({E}/{16\pi \Delta n m_p c^4\Gamma_0^8}\right)^{1/2}=0.081E_{53.5}^{1/2} \Gamma_{0,3.7}^{-4}n_{-4}^{-1/2}{\rm s}$ [@Sari1997dyn], with the width of the shell in the comoving frame $\Delta=cT/(1+z)=3.1\times 10^{10} T_{0.3} {\rm cm}$. The deceleration time when the dissipated energy is comparable to the total kinetic energy of the shell is about $t_{\rm dec}=2T$. At the time $t_{\rm N}<t<t_{\rm dec}$, the Lorentz factor of the forward shock evolves as $\Gamma\propto t^{-1/4}$, and the radius of the shell evolves as $R\propto t^{1/2}$. Thus, the characteristic frequencies and the peak flux density evolve as $\nu_{\rm min}^{\rm r}\propto t^{-1}$, $\nu_{\rm c}^{\rm r}\propto t^{-1}$ and $F_{\rm max}\propto t$. As a result, the flux of high energy emission in the region $\nu_{\rm m}^{\rm f}<\nu_{\rm LAT}<\nu_{\rm c}^{\rm f}$ evolves as $F_{\rm LAT}\propto t^{(-p+3)/2}$; and the flux density of high energy emission in the region $\nu_{\rm LAT}>\nu_{\rm c}^{\rm f}$ evolves as $F_{\rm LAT}\propto t^{(-p+2)/2}$. This shows that the forward shock synchrotron emission in the thick shell case cannot explain the early-time high energy emission due to its flatter light curve. And the reverse shock emission in the thick shell case can also be excluded, using similar calculations as in §4. Based on the above analysis, while the forward shock synchrotron afterglow dominates the LAT emission at late times, the early times LAT emission must be attributed to a prompt emission, which can be limited from above by a high-latitude curvature emission envelope. Discussion and Conclusions {#sec:disc} ========================== We have investigated whether the photons received by the LAT can be explained solely by the forward shock afterglow emission or whether the prompt emission necessarily contributes to the early part of this emission. We have addressed this question with the help of the broad-band observations from the XRT and UVOT instruments onboard of Swift and the LAT onboard of Fermi for the case of the short GRB 090510, obtaining constraints on the extended high-energy emission based on the XRT and UVOT observations. We have obtained a good fit to the XRT and UVOT observations [(except for the late-time optical data)]{} in terms of synchrotron emission from an adiabatic forward shock in spreading jet model, which constrains the environment of this short burst, implying a low number density $n \sim 10^{-3}- 10^{-6} {\rm cm^{-3}}$, consistent with a binary progenitor scenario [e.g., @Belczynski2006], corresponding to an electron energy fraction $\epsilon_{e\rm f} \sim 0.6 - 0.2$ and a weak magnetic field with an energy fraction $\epsilon_{B\rm f}\sim 10^{-5}- 10^{-2}$. The latter could in principle be due to an upstream magnetic field of $\sim 7\mu$G just shock-compressed, without need for magnetic field amplification [@Kumar20093GRBs]. However, the magnetic field values of $\sim \mu$G derived in our galaxy correspond to an average ISM of density $n\sim 1$ cm$^{-3}$, whereas the fits for GRB 090510 indicate external densities comparable to a halo or intergalactic medium with $n \sim 10^{-3} - 10^{-6} {\rm cm^{-3}}$, where based on flux-freezing the field would be expected to be much less than $\mu$G. Thus, the magnetic field may still need to be amplified in the shock. We note also that while the electron energy fraction can be as high as $0.6$, the adiabatic condition is still satisfied, since the low energy emission constrains the radiation to be in the slow cooling case, indicating a radiation efficiency much smaller than unity, in contrast to the radiative regime assumption of @Ghisellini2010, where they did not take into account low energy constraints. In @DePasquale2010, it is assumed that the LAT data at $t=1-200s$ with slope $-1.38$ arises from synchrotron emission in the regime $\nu_{\rm LAT}>\nu_{\rm c}^{\rm f}$, i.e., $\nu_c<10^8{\rm eV}$, in contrast to our conclusion that the synchrotron emission in the regime $\nu_{\rm c}^{\rm f}>1{\rm GeV}$ evolving with slope $\sim-1.1$. Part of the difference may be due to @DePasquale2010 using the formulae of @Granot2002, while we used the formulae of @Sari1998afterglow. However, applying the constraints on $\epsilon_B$,$n$, $\epsilon_{e\rm f}E$ in @DePasquale2010 and taking $p=2.5$ in the expression for $\nu_{\rm c}^{\rm f}$ from @Granot2002, one obtains $\nu_{\rm c}^{\rm f}\simeq 6.3\times10^7{\rm eV}$ at $100$ s. Therefore, $\nu_{\rm c}^{\rm f}=6.3\times10^8{\rm eV}$ at $1{\rm s}$ since $\nu_c\propto t^{-1/2}$, falling in the LAT band, inconsistent with the @DePasquale2010 assumption. The simple power-law decay shown in their Figure 1 is not applicable due to the $\nu_c$ crossing through the LAT band, and it overestimates the LAT photon flux by a factor of $1.7$ at $t=1 {\rm s}$ and $2.1$ at $t\sim 0.4{\rm s}$ . We note also that the formulae in @Granot2002 is applicable only in the self-similar phase, which takes a few e-folding times after the deceleration time to fully develop. Our numerical calculation shows that around the deceleration time, the flux should be reduced by a factor of $\sim2-3$. Furthermore, if the deceleration time is not larger than the duration $T$, the LAT light curve will be flatter at $t<T$, which cannot explain the early-time LAT emission, as discussed in §5 and also strongly argued by @Maxham2011. The initial jet angle is found to be constrained to be $\theta_{\rm j} \lesssim 1^{\circ}$, which implies a collimation-corrected kinetic energy in the range of short GRBs. We have used for our numerical calculations the semi-analytical jet spreading model of @Huang2000apj (detailed in Appendix A), assuming that the observer line of sight is approximately along the jet axis. This leads us to identify a feature in the late XRT and UVOT light curves as symptomatic of a jet break. We note however that @VanEerten2010 argues that numerically the light curve break related to a jet edge as seen by an off-axis observer may appear hidden for a longer time (weeks) than it would be for an on-axis observer, and the alternative possibility cannot be ruled out, although this being one of the brightest bursts reduces the probability of its being observed from an off-axis direction. We have considered the origin of the high energy emission under the constraints provided by the low energy observations. Calculation of the Klein-Nishina effect on the IC scattering of synchrotron photons by electrons accelerated in the forward shock indicate that the Klein-Nishina effect suppresses the high energy IC emission significantly, which results in a Y-parameter much smaller than unity. The distribution of the high energy electrons is not affected by the Klein-Nishina effect, so that the synchrotron high energy emission cannot steepen due to a suppression of the IC scattering through Klein-Nishina effects as argued in @Wang2010. The duration of the prompt emission of this GRB is a crucial parameter for establishing the physics of the LAT emission. Based on the phenomenon that the Fermi GBM emission turns over sharply at $t\sim 0.30\pm 0.07{\rm s}$ [@DePasquale2010], we adopted a duration of $T=0.30\pm 0.07{\rm s}$, compatible with a thin shell case where the deceleration time is larger than the GRB duration. Here, however, we find that a forward shock synchrotron emission model agrees well only with the late-time ($t>3{\rm s}$) high-energy observations of the LAT, but this model falls well below the early-time high-energy emission (the first six LAT data points). It cannot contribute to the power-law high energy component in the prompt spectrum, which suggests that the early-time high-energy emission may have a different origin from the late-time high-energy emission. We excluded various other possibilities for the early-time high-energy emission, such as synchrotron emission from the reverse external shock, SSC emission from reverse/forward external shock, and IC scattering of forward/reverse synchrotron photons by electrons accelerated in reverse/forward shocks, as well as the high-latitude prompt emission with a small variability timescale. The latter is too steep to contribute to the early-time high-energy emission, due to the assumed high temporal variability of the prompt emission, unless the early-time high-energy emission variability timescale is as large as $0.5$ s. We are led to the conclusion that the early-time high-energy emission is likely due to the final portions of the prompt component, or else to the high latitude component of the prompt emission with a variability time scale as large as $0.5$ s plus the synchrotron emission with a deceleration time $t_{\rm dec}\leq(2-3){\rm s}$. This is supported by the fact that we cannot explain the first six LAT data points by synchrotron forward shock emission if we consider the combined data in the three bands (XRT, UVOT and LAT). In this case, the long-lived high-energy emission can be naturally explained as the result of the prompt emission at early times ($t<(2-3) {\rm s}$), and the afterglow emission at late times ($t>(2-3) {\rm s}$) from the adiabatic forward shock synchrotron radiation, with reasonable parameters. In this two-component model of the long-lived high energy emission, the shock deceleration time of $t_{\rm dec}=2{\rm s}$ results in a fit with an initial Lorentz factor around $\Gamma\simeq 2000$, not much larger than the lower limits $\Gamma_{\rm min} =950\pm40$ and $1220\pm60$ [@Abdo090510] obtained from the pair-production limits implied by the presence of $3.4 \rm GeV$ and $31 \rm GeV$ photons. With smaller values of $t_{\rm dec}$, very large initial Lorentz factors are needed (but see @Ioka2010). Although there is so far no statistically significant evidence for an early steeper decay, due to the sparse nature of the LAT data, it is striking that by eye there are 4 out 10 GRB long-lived LAT light curves which have relatively steeper slopes than the slopes, $\sim -1$, of typical late-time X-ray and optical afterglows, according to Table 3 and the light curve fits in the figures of @Zhang2010catalog. The steep decay slope of the LAT light curve is difficult to explain with a normal external shock model. A similar conclusion was subsequently suggested by @Liu2010, and also by @Maxham2011. Our two-component model also explains the steep high-energy temporal decay index values $\alpha_{\rm LAT} \sim 1.5$ as being the natural result of the superposition of the tail-end of the prompt regime and the start of the afterglow. While our model for the high energy afterglow is in principle applicable to other LAT bursts as well, a homogeneous sample of similarly detailed data including low energy constraints will be needed, on a large number of objects, before definite conclusions can be reached. Jet Deceleration and Spreading Dynamical Model =============================================== We solve the following equations for the evolution of the shock radius $R$, the mass $m$ swept up by the shock, the opening half-angle of jet $\theta$ and the Lorentz factor of shock $\Gamma$ [e.g., @Huang2000apj]. The evolution of the shock radius is described by $$\frac{dR}{dt}=\beta c\Gamma(\Gamma+\sqrt{\Gamma^2-1}),$$ with $t$ as the observer’s time and $\beta=\sqrt{\Gamma^2-1}/\Gamma$. The swept mass evolves as $$\frac{dm}{dR}=2\pi R^2(1-\cos\theta)nm_p,$$ with $m_p$ as the proton mass and $n$ as the number density of surrounding medium. The evolution of the opening angle considering the jet spreading is described as $$\frac{d\theta}{dt}\equiv\frac{1}{R}\frac{da}{dt}=\frac{c_{\rm s}(\Gamma+\sqrt{\Gamma^2-1})}{R},$$ where $a$ is the comoving lateral radius of the jet [@Rhoads1999; @Moderski2000] and the comoving sound speed $c_{\rm s}=\hat{\gamma}(\hat{\gamma}-1)(\Gamma-1)\frac{1}{1+\hat{\gamma}(\Gamma-1)}c^2$ with the adiabatic index $\hat{\gamma}$ as $\hat{\gamma}\approx(4\Gamma+1)/(3\Gamma)$. The conservation of the total energy can be expressed as $$\frac{d\Gamma}{dm}=-\frac{\Gamma^2-1}{M_{\rm ej}+\eta_{\rm r} m+2(1-\epsilon)\Gamma m},$$ where the radiative efficiency is $\eta_{\rm r} =0$ for the adiabatic case, and $M_{\rm ej}$ is the ejecta mass. The High-latitude emission ========================== Because photons from high latitude region with respect to the line of sight will arrive later than that from low latitude region due to the curved front surface of the jet, we will observe a fast decreasing emission instead of an abrupt cutoff of the emission, which is so called “curvature effect". We neglect the radial structure of the emitting shell for simplicity. Under this assumption we calculate the flux from the shell at radius $r_i$ which expands toward us with Lorentz factor $\Gamma_i$ by [@Granot1999; @Woods1999; @Ioka2001; @Yamazaki2003; @Dermer2004; @Toma2009cocoon] $$F_{\nu}(t) = \frac{1+z}{d_{\rm L}^2} 8\pi r_i^3 j'_{\nu'} \frac{1}{[1+\Gamma_j^2 \theta^2(t)]^2},$$ where the photon frequency in comoving frame is $$\nu' = (1+z)\nu \frac{1+\Gamma_j^2 \theta^2(t)}{2\Gamma_j}$$ with $\nu$ as the photon frequency in observing frame, and $\theta(t)$ describes the emitting point of the high-latitude emission. We have the comoving emissivity $$j'_{\nu'}\propto \nu'^{-\beta} \propto (1+\Gamma_j^2 \theta^2(t))^{-\beta} ,$$ where $\beta$ is the spectral index of the observed emission, which is consistent with the power-law spectra fit of [@Abdo090510]. By using the relationship $\theta(t) = \sqrt{2} \left[1- \frac{c}{r_i} \left(\bar{t}_i - \frac{t}{1+z}\right)\right]^{1/2}$ with $\bar{t}_i$ as the emission time in the central engine frame, we get $$1+\Gamma_j^2 \theta^2(t)=\frac{t-(t_0-\Delta t)}{\Delta t}$$ where $t_0=(1+z)(\bar{t}_i-\frac{r_i}{c})$ is the observed peak time of the high energy pulse and $\Delta t=\frac{r_i(1+z)}{2c\Gamma_i^2}$ is the observed dynamical time scale. Therefore, the observed flux evolves as time as following $$F_{\nu}(t) \propto\left[\frac{t-(t_0-\Delta t)}{\Delta t}\right]^{-2-\beta}.$$ [59]{} natexlab\#1[\#1]{} , A. A. [et al.]{} 2009, , 706, L138, 0909.2470 ——. 2009, , 462, 331, 0908.1832 ——. 2010, , 712, 558 ——. 2009, Science, 323, 1688 ——. 2009, , 707, 580, 0910.4192 , A., [Gallant]{}, Y. A., [Kirk]{}, J. G., & [Guthmann]{}, A. W. 2001, , 328, 393, arXiv:astro-ph/0107530 , M. [et al.]{} 2010, , 717, L127, 1007.3409 ——. 2010, , 716, 1178 , K., [Guiriec]{}, S., & [M[é]{}sz[á]{}ros]{}, P. 2009, , 705, L191, 0909.0306 , K., [Perna]{}, R., [Bulik]{}, T., [Kalogera]{}, V., [Ivanova]{}, N., & [Lamb]{}, D. Q. 2006, , 648, 1110, arXiv:astro-ph/0601458 , R. D., & [McKee]{}, C. F. 1976, Physics of Fluids, 19, 1130 , A., [Guetta]{}, D., & [Piro]{}, L. 2010, , 720, 1008, 0911.4453 , M. [et al.]{} 2010, , 709, L146, 0910.1629 , C. D. 2004, , 614, 284, arXiv:astro-ph/0403508 , & [Fermi-GBM collaborations]{}. 2011, ArXiv e-prints, 1101.2082 , D. L., & [Waxman]{}, E. 2001, , 547, 922, arXiv:astro-ph/9912214 , W., [Mao]{}, J., [Xu]{}, D., & [Fan]{}, Y. 2009, , 706, L33, 0908.3975 , G., [Ghisellini]{}, G., & [Nava]{}, L. 2010, , 510, L7+, 0909.0016 , G., [Ghirlanda]{}, G., [Nava]{}, L., & [Celotti]{}, A. 2010, , 403, 926, 0910.2459 , S. [et al.]{} 2009, GRB Coordinates Network, 9344, 1 , J., [for the Fermi LAT Collaboration]{}, & [the GBM Collaboration]{}. 2010, ArXiv e-prints, 1003.2452 , J., [Piran]{}, T., & [Sari]{}, R. 1999, , 513, 679, arXiv:astro-ph/9806192 , J., & [Sari]{}, R. 2002, , 568, 820, arXiv:astro-ph/0108027 , S., [Connaughton]{}, V., & [Briggs]{}, M. 2009, GRB Coordinates Network, 9336, 1 , E. A., [Krimm]{}, H. A., [Grupe]{}, D., [Kuin]{}, N. P. M., [Barthelmy]{}, S. D., [Burrows]{}, D. N., [Roming]{}, P., & [Gehrels]{}, N. 2009, GCN Report, 218, 1 , Y. F., [Dai]{}, Z. G., & [Lu]{}, T. 2000, , 316, 943, arXiv:astro-ph/0005549 , Y. F., [Gou]{}, L. J., [Dai]{}, Z. G., & [Lu]{}, T. 2000, , 543, 90, arXiv:astro-ph/9910493 , K. 2010, Progress of Theoretical Physics, 124, 667, 1006.3073 , K., & [Nakamura]{}, T. 2001, , 554, L163, arXiv:astro-ph/0105321 , S., [Zhang]{}, B., [M[é]{}sz[á]{}ros]{}, P., & [Burrows]{}, D. 2007, , 655, 391, arXiv:astro-ph/0506157 , P., & [Barniol Duran]{}, R. 2009, , 400, L75, 0905.2417 ——. 2010, , 409, 226, 0910.5726 , P., & [Panaitescu]{}, A. 2000, , 541, L51, arXiv:astro-ph/0006317 , R., & [Wang]{}, X. 2010, ArXiv e-prints, 1009.1289 , F. [et al.]{} 2009, GRB Coordinates Network, 9343, 1 , A., [Zhang]{}, B., & [Zhang]{}, B. 2011, ArXiv e-prints, 1101.0144 , S. [et al.]{} 2010, , 516, A71+, 1003.3885 , R., [Sikora]{}, M., & [Bulik]{}, T. 2000, , 529, 151, arXiv:astro-ph/9904310 , E., [Ando]{}, S., & [Sari]{}, R. 2009, , 703, 675, 0903.2557 , A. 2010, ArXiv e-prints, 1005.1051 , N. [et al.]{} 2009, GRB Coordinates Network, 9355, 1 , A., [McBreen]{}, S., & [Kruehler]{}, T. 2009, GRB Coordinates Network, 9353, 1 , S. 2010, , 724, L109, 1004.3330 , J. E. 1997, , 487, L1+, arXiv:astro-ph/9705163 ——. 1999, , 525, 737, arXiv:astro-ph/9903399 , R. 1997, , 489, L37+ , R., & [Esin]{}, A. A. 2001, , 548, 787, arXiv:astro-ph/0005253 , R., [Piran]{}, T., & [Narayan]{}, R. 1998, , 497, L17+, arXiv:astro-ph/9712005 , K., [Wu]{}, X., & [M[é]{}sz[á]{}ros]{}, P. 2009, , 707, 1404, 0905.1697 , K., [Wu]{}, X., & [Meszaros]{}, P. 2010, ArXiv e-prints, 1002.2634 , T. N. [et al.]{} 2009, GRB Coordinates Network, 9337, 1 , H., [Zhang]{}, W., & [MacFadyen]{}, A. 2010, , 722, 235, 1006.5125 , X., [He]{}, H., [Li]{}, Z., [Wu]{}, X., & [Dai]{}, Z. 2010, , 712, 1232, 0911.4189 , X., [Li]{}, Z., & [M[é]{}sz[á]{}ros]{}, P. 2006, , 641, L89, arXiv:astro-ph/0601229 , X. Y., [Dai]{}, Z. G., & [Lu]{}, T. 2001, , 546, L33, arXiv:astro-ph/0010320 , E., & [Loeb]{}, A. 1999, , 523, 187, arXiv:astro-ph/9903377 , R., [Yonetoku]{}, D., & [Nakamura]{}, T. 2003, , 594, L79, arXiv:astro-ph/0306615 , B., [Kobayashi]{}, S., & [M[é]{}sz[á]{}ros]{}, P. 2003, , 595, 950, arXiv:astro-ph/0302525 , B. [et al.]{} 2010, ArXiv e-prints, 1009.3338 [^1]: The first $>100 {\rm MeV}$ photons arrive later than the first lower energy photons detected by GBM (Gamma-ray Burst Monitor). [^2]: Here we use the convention $F_{\nu}\propto t^{-\alpha}\nu^{-\beta}$. [^3]: Similarly large energies are required for hadronic models of the prompt emission of this burst [@Asano2009]. [^4]: Hereafter we use the subscripts or superscripts $'\rm f'$ and $'\rm r'$ to represent the quantities of the forward-shocked and reverse-shocked regions, respectively, and we use the convention $Q_x=Q/10^x$ in cgs units throughout the paper. [^5]: The closure relation $\alpha_{\rm X,1}-1.5\beta_{\rm X}=-0.5$ indicates the slow/fast-cooling ISM/wind external forward shock model with $\nu_{\rm X}>\nu_{\rm c}^{\rm f}$ and the spectral index $\beta_{\rm X}=p/2$, but this implies $p=1.6<2.0$, which is not favored by numerical simulations of shock acceleration (e.g., @Achterberg2001) or by observational data of general GRB afterglows [@Freedman2001].

--- abstract: 'The emission of gravitational waves from a system of massive objects interacting on elliptical, hyperbolic and parabolic orbits is studied in the quadrupole approximation. Analytical expressions are then derived for the gravitational wave luminosity, the total energy output and gravitational radiation amplitude. A crude estimate of the expected number of events towards peculiar targets (i.e. globular clusters) is also given. In particular, the rate of events per year is obtained for the dense stellar cluster at the Galactic Center.' author: - 'Salvatore Capozziello$^1$, Mariafelicia De Laurentis$^2$' title: Gravitational waves from stellar encounters --- [*Keywords*]{}: theory of orbits, gravitational radiation, quadrupole approximation. \[intro\] Introduction ======================= Gravitational-wave (GW) science has entered a new era. Experimentally, several GW ground-based-laser-interferometer detectors ($10^{-1}kHz$) have been built in the United States (LIGO) [@Abra], Europe (VIRGO and GEO) [@Caron; @Luck] and Japan (TAMA) [@Ando], and are now taking data at design sensitivity. Advanced optical configurations capable of reaching sensitivities slightly above and even below the so-called standard-quantum-limit for a free test-particle, have been designed for second [@Meers] and third generation [@Braginsky] GW detectors. A laser-interferometer space antenna (LISA) [@LISA] ($10^{-4}\sim10^{-2} Hz$) might fly within the next decade. Resonant-bar detectors ($ \sim 1 kHz$)[@Astone] are improving more and more their sensitivity, broadening their frequency band. At much lower frequencies, $\sim10^{-17}Hz$, future cosmic microwave background (CMB) probes are devoted to detect GWs by measuring the CMB polarization [@Kamionkowski]. Millisecond pulsar timing can set interesting upper limits [@Jenet] in the frequency range $10^{-9}\sim 10^{-8} Hz$. In this frequency range, the large number of millisecond pulsars which will be detectable with the square kilometer array [@skatelescope], would provide an ensemble of clocks that can be used as multiple arms of a GW detector. From a theoretical point of view, recent years have been characterized by numerous major advances due, essentially, to the development of numerical gravity. Concerning the most promising sources to be detected, the GW generation problem has improved significantly in relation to the dynamics of binary and multiple systems of compact objects as neutron stars and black holes. Besides, the problem of non-geodesic motion of particles in curved spacetime has been developed considering the emission of GWs [@Poisson; @Mino]. Solving this problem is of considerable importance in order to predict the accurate waveforms of GWs emitted by extreme mass-ratio binaries, which are among the most promising sources for LISA [@Finn]. To this aim, searching for criteria to classify the ways in which sources collide is of fundamental importance. A first rough criterion can be the classification of stellar encounters in [*collisional*]{} as in the globular clusters and in [*collisionless*]{} as in the galaxies [@binney]. A fundamental parameter is the richness and the density of the stellar system and so, obviously, we expect a large production of GWs in rich and dense systems. Systems with these features are the globular clusters and the galaxy centers. In particular, one can take into account the stars (early-type and late-type) which are around our Galactic Center, e.g. Sagittarius $A^{*}$ ($Sgr A^{*}$) which could be very interesting targets for the above mentioned ground-based and space-based detectors. In recent years, detailed information has been achieved for kinematics and dynamics of stars moving in the gravitational field of such a central object. The statistical properties of spatial and kinematical distributions are of particular interest (see e.g. [@Genzel; @Sellgreen; @CapozIov]). Using them, it is possible to give a quite accurate estimate of the mass and the size of the central object: we have $(2.61\pm0.76)\times10^6M_{\odot})$ concentrated within a radius of $0.016 pc$ (about $30$ light-days)[@Ghez; @Thatte]. More precisely, in [@Ghez], it is described a campaign of observations where velocity measurements in the central $arcsec^{2}$ are extremely accurate. Then from this bulk of data, considering a field of resolved stars whose proper motions are accurately known, one can classify orbital motions and deduce, in principle, the rate of production of GWs according to the different types of orbits. This motivates this paper in which, by a classification of orbits in accordance with the conditions of motion, we want to calculate the GW luminosity for the different types of stellar encounters. A similar approach has been developed in [@achille] but, in that case, only hyperbolic trajectories have been considered. Following the method outlined in [@pm1; @pm2], we investigate the GW emission by binary systems in the quadrupole approximation considering bounded (circular or elliptical) and unbounded (parabolic or hyperbolic) orbits. Obviously, the main parameter is the approaching energy of the stars in the system (see also [@schutz] and references therein). We expect that gravitational waves are emitted with a “peculiar” signature related to the encounter-type: such a signature has to be a “burst” wave-form with a maximum in correspondence of the periastron distance. The problem of [*bremsstrahlung*]{}-like gravitational wave emission has been studied in detail by Kovacs and Thorne [@kt] by considering stars interacting on unbounded and bounded orbits. In this paper, we face this problem discussing in detail the dynamics of such a phenomenon which could greatly improve the statistics of possible GW sources. The paper is organized as follows: in Sec. II, the main features of stellar encounters and orbit classification are reviewed. Sec. III is devoted to the emission and luminosity of GWs from binary systems in the different kinds of orbits assuming the quadrupole approximation. A discussion of the wave-form dependence from the orbital parameters is given in Sec. IV. In Sec. V, we derive the expected rate of events assuming the Galactic Center as a target. Section VI is devoted to concluding remarks. Orbits in stellar encounters ============================ Let us take into account the Newtonian theory of orbits since stellar systems, also if at high densities and constituted by compact objects, can be usually assumed in Newtonian regime. We give here a self-contained summary of the well-known orbital types in order to achieve below a clear classification of the possible GW emissions. We refer to the text books [@binney; @landau] for a detailed discussion. A mass $m_1$ is moving in the gravitational potential $\Phi$ generated by a second mass $m_2$. The vector radius and the polar angle depend on time as a consequence of the star motion, i.e. $\textbf{r}=\textbf{r}(t)$ and $\phi=\phi(t)$. With this choice, the velocity $\textbf{v}$ of the mass $m_1$ can be parameterized as $$\textbf{v}=v_r\widehat{r}+v_{\phi}\widehat{\phi}~,$$ where the radial and tangent components of the velocity are, respectively, $$v_r=\frac{dr}{dt} ~~~~~~~~v_{\phi}=r \frac{d\phi}{dt}~.$$ In this case, the total energy and the angular momentum, read out $$\frac{1}{2}{\mu\left(\frac{dr}{dt}\right)}^{2}+\frac{\mathbb{L}^{2}}{2\mu r^{2}}-\frac{\gamma}{r}=E \label{eq:energia}$$ and $$L=r^2\frac{d\phi}{dt} \label{eq:momang1},$$ respectively, where $\mu=\frac{m_1m_2}{m_1+m_2}$ is the reduced mass of the system and $\gamma=Gm_1m_2$. We can split the kinetic energy into two terms where, due to the conservation of angular momentum, the second one is a function of $r$ only. An effective potential energy $V_{eff}$, $$V_{eff}=\frac{\mathbb{L}^{2}}{2\mu r^{2}}-\frac{\gamma}{r}\label{eq:energpot}$$ is immediately defined. The first term corresponds to a repulsive force, called the angular momentum barrier. The second term is the gravitational attraction. The interplay between attraction and repulsion is such that the effective potential energy has a minimum. Indeed, differentiating with respect to $r$ one finds that the minimum lies at $r_0=\frac{L^{2}}{\gamma\mu}$ and that $$V_{eff}^{min}=-\frac{\mu\gamma^{2}}{2L^{2}}\label{eq:enrgpotmin}\,.$$ Therefore, since the radial part of kinetic energy, $$K_{r}=\frac{1}{2}\mu\left(\frac{dr}{dt}\right)^{2}$$, is non-negative, the total energy must be not less than $V_{eff}^{min}$, i.e. $$E\geq E_{min}=-\frac{\mu\gamma^{2}}{2L^{2}}\label{eq:emin}\,.$$ The equal sign corresponds to the radial motion. For $E_{min}<E<0$, the trajectory lies between a smallest value $r_{min}$ and greatest value $r_{max}$ which can be found from the condition $E=V_{eff}$, i.e. $$r_{\{min,max\}}=-\frac{\gamma}{2E}\pm\sqrt{\left(\frac{\gamma}{2E}\right)^{2}+\frac{L^{2}}{2\mu E}}\label{eq:rminmax}$$ where the upper (lower) sign corresponds to $r_{max}$ ($r_{min}$). Only for $E>0$, the upper sign gives an acceptable value; the second root is negative and must be rejected. Let us now proceed in solving the differential equations (\[eq:energia\]) and (\[eq:momang1\]). We have $$\frac{dr}{dt}=\frac{dr}{d\phi}\frac{d\phi}{dt}=\frac{L}{\mu r^{2}}\frac{dr}{d\phi}= -\frac{L}{\mu}\frac{d}{d\phi}\left(\frac{1}{r}\right)\label{eq:diff}$$ and defining, as standard, the auxiliary variable $u=1/r$, Eq. (\[eq:energia\]) takes the form $$u'^{2}+u^{2}-\frac{2\gamma\mu}{L^{2}}u=\frac{2\mu E}{L^{2}}\label{eq:diffenerg}$$ where $u'=du/d\phi$ and we have divided by $L^{2}/2\mu$. Differentiating with respect to $\phi$, we get $$u'\left(u''+u-\frac{\gamma\mu}{L^{2}}\right)=0$$ hence either $u'=0$, corresponding to the circular motion, or $$u''+u=\frac{\gamma\mu}{L^{2}}\label{eq:moto}$$ which has the solution $$u=\frac{\gamma\mu}{L^{2}}+C\cos\left(\phi+\alpha\right)$$ or, reverting the variable, $$r=\left[\frac{\gamma\mu}{L^{2}}+C\cos\left(\phi+\alpha\right)\right]^{-1}\label{eq:solution}$$ which is the canonical form of conic sections in polar coordinates [@smart]. The constant $C$ and $\alpha$ are two integration constants of the second order differential equation (\[eq:moto\]). The solution (\[eq:solution\]) must satisfy the first order differential equation (\[eq:diffenerg\]). Substituting (\[eq:solution\]) into (\[eq:diffenerg\]) we find, after a little algebra, $$C^{2}=\frac{2\mu E}{L^{2}}+\left(\frac{\gamma\mu}{L^{2}}\right)^{2}\label{eq:C}$$ and therefore, taking account of Eq. (\[eq:emin\]), we get $C^{2}\geq 0$. This implies the four kinds of orbits given in Table I. $$$$ $C=0$ $E=E_{min}$ circular orbits -------------------------------------------- --------------- ------------------- -- -- $0<\left|C\right|<\frac{\gamma\mu}{L^{2}}$ $E_{min}<E<0$ elliptic orbits $\left|C\right|=\frac{\gamma\mu}{L^{2}}$ $E=0$ parabolic orbits $\left|C\right|>\frac{\gamma\mu}{L^{2}}$ $E>0$, hyperbolic orbits : Orbits in Newtonian regime classified by the approaching energy. $$$$ Circular Orbits --------------- Circular motion corresponds to the condition $u'=0$ by which one find $r_{0}=L^{2}/\mu\gamma$ where $V_{eff}$ has its minimum. We also note that the expression for $r_{0}$ together with Eq.(\[eq:emin\]) gives$$r_{0}=-\frac{\gamma}{2E_{min}}\label{eq:rzero}$$ Thus the two bodies move in concentric circles with radii, inversely proportional to their masses and are always in opposition. Elliptical Orbits ----------------- For $0<\left|C\right|<\mu\gamma/L^{2}$, $r$ remains finite for all values of $\phi$. Since $r(\phi+2\pi)=r(\phi)$, the trajectory is closed and it is an ellipse. If one chooses $\alpha=0$, the major axis of the ellipse corresponds to $\phi=0$. We get $$r_{\left|\phi=0\right.}=r_{min}=\left[\frac{\gamma\mu}{L^{2}}+C\right]^{-1}\label{eq:rphi}$$ and $$r_{\left|\phi=\pi\right.}=r_{max}=\left[\frac{\gamma\mu}{L^{2}}-C\right]^{-1}\label{eq:rpi}$$ and since $r_{max}+r_{min}=2a$, where $a$ is the semi-major axis of the ellipse, one obtains $$a=r_{\left|\phi=0\right.}=r_{min}=\frac{\gamma\mu}{L^{2}}\left[\left(\frac{\gamma\mu}{L^{2}}\right)^{2}+C^{2}\right]^{-1}$$ $C$ can be eliminated from the latter equation and Eq.( \[eq:C\]) and then $$a=-\frac{\gamma}{2E}\label{eq:a}$$ Furthermore, if we denote the distance $r_{\left|\phi=\pi/2\right.}$ by $l$, the so-called [*semi-latus rectum*]{} or the parameter of the ellipse, we get $$l=\frac{L^{2}}{\gamma\mu}\label{eq:latusrectum}$$ and hence the equation of the trajectory $$r=\frac{l}{1+\epsilon\cos\phi}\label{eq:traiettoria}$$ where $\epsilon=\sqrt{\frac{1-l}{a}}$ is the eccentricity of the ellipse. Parabolic and Hyperbolic Orbits ------------------------------- These solutions can be dealt together. They correspond to $E\geq 0$ which is the condition to obtain unbounded orbits. Equivalently, one has $\left|C\right|\geq\gamma\mu/L^{2}$. The trajectory is $$r=l\left(1+\epsilon\cos\phi\right)^{-1}\label{eq:traie}$$ where $\epsilon\geq1$. The equal sign corresponds to $E=0$ . Therefore, in order to ensure positivity of $r$, the polar angle $\phi$ has to be restricted to the range given by $$1+\epsilon\cos\phi>0 \label{eq:cosphi}$$ This means $\cos\phi>-1$, i.e. $\phi\in(-\pi,\pi)$ and the trajectory is not closed any more. For $\phi\rightarrow\pm\pi$, we have $r\rightarrow\infty$. The curve (\[eq:traie\]), with $\epsilon=1$, is a parabola. For $\epsilon>1$, the allowed interval of polar angles is smaller than $\phi\in(-\pi,\pi)$, and the trajectory is a hyperbola. Such trajectories correspond to non-returning objects. Gravitational wave luminosity in the quadrupole approximation ============================================================= At this point, considering the orbit equations, we want to classify the gravitational radiation for the different stellar encounters. To this aim, let us start with a short review of the quadrupole approximation for GW radiation. We add this discussion for the sake of completeness, but we send the Reader to the References [@gravitation; @shapiro; @maggiore; @thorne] for a detailed exposition. The Einstein field equations give a description of how the curvature of space-time is related to the energy-momentum distribution. In the weak field approximation, moving massive objects produce gravitational waves which propagate in the vacuum with the speed of light. In this approximation, we have $$g_{\mu\nu}=\delta_{\mu\nu}+\kappa h_{\mu\nu},\qquad\left(\left|h_{\mu\nu}\right|<<1\right),$$ whit $\kappa$ the gravitational coupling. The field equations are $$\square\bar{h}_{\mu\nu}=-\frac{1}{2}\kappa T_{\mu\nu}\label{eq:h}$$ where $$\bar{h}_{\mu\nu}=h_{\mu\nu}-\frac{1}{2}\delta_{\mu\nu}h_{\lambda\lambda}\,,$$ and $T_{\mu\nu}$ is the total stress-momentum-energy tensor of the source, including the gravitational stresses. A plane GW can be written as $$\label{wave} \bar{h}_{\mu\nu}=h_{\mu\nu}=he_{\mu\nu}\cos(\omega t-{\mbox{\boldmath $\mathbf{k}\cdot\mathbf{x}$}})$$ where $h$ is the amplitude, $\omega$ the frequency, $k$ the wave number and $e_{\mu\nu}$ is a unit polarization tensor, obeying the conditions $$e_{\mu\nu}=e_{\nu\mu},\qquad e_{\mu\mu}=0,\qquad e_{\mu\nu}e_{\mu\nu}=1.$$ Let us assume a gauge in which $e_{\mu\nu}$ is space-like and transverse; thus, a wave travelling in the *z* direction has two possible independent polarizations: $$e_{1}=\frac{1}{\sqrt{2}}(\hat{x}\hat{x}-\hat{y}\hat{y})\qquad e_{2}=\frac{1}{\sqrt{2}}(\hat{x}\hat{y}-\hat{y}\hat{x}).$$ One can now search for wave solutions of (\[eq:h\]) from a system of masses undergoing arbitrary motions, and then obtain the power radiated. The result, assuming the source dimensions very small with respect to the wavelengths (quadrupole approximation [@landau]), is that the power ${\displaystyle \frac{dE}{d\Omega}}$ radiated in a solid angle $\Omega$ with polarization $e_{ij}$ is $$$$ $$\frac{dE}{d\Omega}=\frac{G}{8\pi c^{5}}\left(\frac{d^{3}Q_{ij}}{dt^{3}}e_{ij}\right)^{2}\label{eq:P}$$ where $Q_{ij}$ is the quadrupole mass tensor $$Q_{ij}=\sum _a m_a(3x_a^ix_a^j-\delta_{ij}r_a^2)~,\label{qmasstensor}$$ $G$ being the Newton constant, $r_a$ the modulus of the vector radius of the $a-th$ particle and the sum running over all masses $m_{a}$ in the system. It has to be noted that the result is independent of the kind of stresses which are present into the dynamics. If one sums (\[eq:P\]) over the two allowed polarizations, one obtains $$\begin{aligned} \sum_{pol}\frac{dE}{d\Omega} & = & \frac{G}{8\pi c^{5}}\left[\frac{d^{3}Q_{ij}}{dt^{3}}\frac{d^{3}Q_{ij}}{dt^{3}}-2n_{i}\frac{d^{3}Q_{ij}}{dt^{3}}n_{k}\frac{d^{3}Q_{kj}}{dt^{3}}-\frac{1}{2}\left(\frac{d^{3}Q_{ii}}{dt^{3}}\right)^{2}\right.\nonumber \\ & & \left.+\frac{1}{2}\left(n_{i}n_{j}\frac{d^{3}Q_{ij}}{dt^{3}}\right)^{2}+\frac{d^{3}Q_{ii}}{dt^{3}}n_{j}n_{k}\frac{d^{3}Q_{jk}}{dt^{3}}\right]\label{eq:sommatoria}\end{aligned}$$ where $\hat{n}$ is the unit vector in the radiation direction. The total radiation rate is obtained by integrating (\[eq:sommatoria\]) over all directions of emission; the result is $$\frac{dE}{dt}=-\frac{G\left\langle Q_{ij}^{(3)}Q^{(3)ij}\right\rangle }{45c^{5}}\label{eq:dEdt}$$ where the index (3) represents the differentiation with respect to time, the symbol $<>$ indicates that the quantity is averaged over several wavelengths. This crucial point is linked with the difficulties of localizing gravitational energy so the right hand side of Eq. (\[eq:dEdt\]) cannot be viewed as an instantaneous quantity. This problem has been already faced for circular and elliptical orbits in [@Finn; @bla; @buonanno]. For hyperbolic and parabolic orbits, it is crucial to estimate the quantity in the right hand side of Eq.(\[eq:dEdt\]) in the zone where stars are slightly changing their trajectories, that means at peri-astron, while we expect no emission in asymptotic regime of stars approaching to or going away from this region. For a detailed discussion in the hyperbolic case, see [@achille]. With this formalism, it is possible to estimate the amount of energy emitted in the form of GWs from a system of massive objects interacting among them [@pm1; @pm2]. In this case, the components of the quadrupole mass tensor in the equatorial plane ($\theta=\pi/2$) are $$\begin{array}{lll} Q_{xx}=\mu r^2(3\cos{^2\phi}-1)~,\\ \\ Q_{yy}=\mu r^2(3\sin{^2\phi}-1)~,\\ \\ Q_{zz}=-\mu r^2~,\\ \\ Q_{xz}=Q_{zx}=0~,\\ \\ Q_{yz}=Q_{zy}=0~,\\ \\ Q_{xy}=Q_{yx}=3\mu r^2 \cos\phi \sin\phi~, \end{array}\label{eq:quadrupoli}$$ where the masses $m_{1}$ and $m_{2}$ have polar coordinates $\{r_{i}\cos\theta\cos\phi,\; r_{i}\cos\theta\sin\phi,\: r_{i}\sin\theta\}$ with $i=1,2$. The origin of the motions is taken at the center of mass. Such components can be differentiated with respect to the time as in Eq.(\[eq:dEdt\]). In doing so, we can use some useful relations derived in the previous Section. GW luminosity from circular and elliptical orbits ------------------------------------------------- Using Eq:(\[eq:traiettoria\]), let us derive the angular velocity equation $$\dot{\phi}=\frac{\sqrt{G l (m_{1}+m_{2})} (\epsilon \cos\phi+1)^2}{l^2} \label{eq:angularvelo}$$ and then, from (\[eq:quadrupoli\]), the quadrupolar components for the elliptical orbits $$\frac{d^{3}Q_{xx}}{dt^{3}}=\beta(24 \cos\phi+\epsilon (9 \cos2 \phi)+11)) \sin \phi$$ $$\frac{d^{3}Q_{yy}}{dt^{3}}=-\beta(24 \cos\phi+\epsilon (13+9 \cos2 \phi)) \sin\phi)$$ $$\frac{d^{3}Q_{zz}}{dt^{3}}=-2\beta \epsilon \sin\phi$$ $$\frac{d^{3}Q_{xy}}{dt^{3}}=\beta (24 \cos\phi+\epsilon (11+9 \cos2 \phi)) \sin\phi)$$ where $$\beta=\frac{G l (m_{1}+m_{2}))^{3/2} \mu (\epsilon \cos\phi+1)^2}{l^4}\,.$$ Being $$\begin{aligned} Q_{ij}^{(3)}Q^{(3)ij}=\frac{G^3}{l^5}\left[ (m_{1}+m_{2})^3 \mu ^2 (1+\epsilon \cos\phi)^4\right.\\ \left(415 \epsilon ^2+3 (8 \cos\phi+3 \epsilon \cos2 \phi) \right.\\ \left.(72 \cos\phi+\epsilon (70+27 \cos2 \phi))) \sin\phi^{2}\right] \end{aligned}$$ the total power radiated is given by $$\begin{aligned} \frac{dE}{dt}=\frac{G^3}{45c^5l^5}f(\phi)\end{aligned}$$ where $$f(\phi)=\left[ (m_{1}+m_{2})^3 \mu ^2 (1+\epsilon \cos\phi)^4\right.\\ \left(415 \epsilon ^2+3 (8 \cos\phi+3 \epsilon \cos2 \phi) \right.\\ \left.(72 \cos\phi+\epsilon (70+27 \cos2 \phi))) \sin\phi^{2}\right]$$ The total energy emitted in the form of gravitational radiation, during the interaction, is given by $$\Delta E=\int^{\infty}_0 \left|\frac{dE}{dt}\right| dt~.$$ From Eq.(\[eq:momang1\]), we can adopt the angle $\phi$ as a suitable integration variable. In this case, the energy emitted for $\phi_1<\phi<\phi_2$ is $$\Delta E(\phi_1,\phi_2) =\frac{G^3}{45c^5l^5}\int^{\phi_2}_{\phi_1}f(\phi)~d\phi~,\label{eq:integraleenergia}$$ and the total energy can be determined from the previous relation in the limits $\phi_1\rightarrow 0$ and $\phi_2\rightarrow \pi$. Thus, one has $$\Delta E=\frac{G^4 \pi (m_{1}+m_{2})^3 \mu ^2}{l^5c^5}F(\epsilon)~$$ where $F(\epsilon)$ depends on the initial conditions only and is given by $$F(\epsilon)=\frac{ \left(13824+102448 \epsilon ^2+59412 \epsilon ^4+2549 \epsilon ^6\right)}{2880}$$ In other words, the gravitational wave luminosity strictly depends on the configuration and kinematics of the binary system. GW luminosity from parabolic and hyperbolic orbits -------------------------------------------------- Also in this case, we use Eq:(\[eq:traie\]) and Eq. (\[eq:quadrupoli\]) to calculate the quadrupolar formula for parabolic and hyperbolic orbits. The angular velocity is $$\dot{\phi}=l^2 L (\epsilon \cos\phi+1)^2 \label{eq:angularvelo1}$$ and the derivative are $$\frac{d^{3}Q_{xx}}{dt^{3}}=\rho(24 \cos\phi+\epsilon (9 \cos 2 \phi+11)) \sin \phi$$ $$\frac{d^{3}Q_{yy}}{dt^{3}}=-\rho(24 \cos\phi+\epsilon (13+9 \cos2 \phi)) \sin\phi)$$ $$\frac{d^{3}Q_{zz}}{dt^{3}}=-2\rho\epsilon\sin\phi$$ $$\frac{d^{3}Q_{xy}}{dt^{3}}=-\frac{3}{2}\rho(\epsilon \cos \phi+1)^2 (5 \epsilon \cos \phi+8 \cos 2 \phi+3 \epsilon \cos3 \phi)$$ where $$\rho=l^4 L^3 \mu (\epsilon \cos\phi+1)^2\,.$$ The radiated power is given by $$\begin{aligned} \frac{dE}{dt}=-\frac{G \rho^2 \left([314 \epsilon ^2+(1152 \cos (\phi+187 \epsilon \cos 2 \phi-3 (80 \cos 3 \phi+30 \epsilon \cos4 \phi+48 \cos 5 \phi+9 \epsilon \cos6 \phi)) \epsilon -192 \cos4 \phi+576\right]}{120 c^5}\end{aligned}$$ then $$\frac{dE}{dt}=-\frac{G l^8 L^6 \mu ^2 }{120 c^5 }f(\phi)$$ where $f(\phi)$, in this case, is $$f(\phi)= \left(314 \epsilon ^2+(1152 \cos (\phi+187 \epsilon \cos 2 \phi-3 (80 \cos 3 \phi+30 \epsilon \cos4 \phi+48 \cos 5 \phi+9 \epsilon \cos6 \phi)) \epsilon -192 \cos4 \phi+576\right)$$ Then using Eq. (\[eq:dEdt\]), the total energy emitted in the form of gravitational radiation during the interaction as a function of $\phi$ is given by $$\Delta E(\phi_1,\phi_2) =-\frac{G l^8 L^6 \pi \left(1271 \epsilon ^6+24276 \epsilon ^4+34768 \epsilon ^2+4608\right) \mu ^2}{480 c^5} ~d\phi~,\label{eq:integraleenergia1}$$ and the total energy can be determined from the previous relation in the limits $\phi_1\rightarrow -\pi$ and $\phi_2\rightarrow\pi$ in the parabolic case. Thus, one has $$\Delta E=-\frac{G l^8 L^6 \pi \mu^2 }{480 c^5}F(\epsilon)~,$$ where $F(\epsilon)$ depends on the initial conditions only and is given by $$F(\epsilon)= \left(1271 \epsilon ^6+24276 \epsilon ^4+34768 \epsilon ^2+4608\right)~.$$ In the hyperbolic case, we have that the total energy is determined in the limits ${\displaystyle \phi_1\rightarrow \frac{-3\pi}{4}}$ and ${\displaystyle \phi_2\rightarrow\frac{-3\pi}{4}}$, i.e. $$\Delta E=--\frac{G l^8 L^6\mu^2}{201600 c^5}F(\epsilon)~,$$ where $F(\epsilon)$ depends on the initial conditions only and is given by $$\begin{aligned} F(\epsilon)&=& \left[315 \pi \left(1271 \epsilon ^6+24276 \epsilon ^4+34768 \epsilon ^2+4608\right)+\right.\\ & & \left. +16 \epsilon \left[\epsilon \left[\epsilon \left(926704 \sqrt{2}-7 \epsilon (3319 \epsilon ^2-32632 \sqrt{2} \epsilon +55200)\right)-383460\right]+352128 \sqrt{2}\right]\right]~. $$ As above, the gravitational wave luminosity strictly depends on the configuration and kinematics of the binary system. Gravitational wave amplitude ============================ Direct signatures of gravitational radiation are its amplitude and its wave-form. In other words, the identification of a GW signal is strictly related to the accurate selection of the shape of wave-forms by interferometers or any possible detection tool. Such an achievement could give information on the nature of the GW source, on the propagating medium, and , in principle, on the gravitational theory producing such a radiation [@Dela]. It is well known that the amplitude of GWs can be evaluated by $$h^{jk}(t,R)=\frac{2G}{Rc^4}\ddot{Q}^{jk}~, \label{ampli1}$$ $R$ being the distance between the source and the observer and $\{j,k\}=1,2$. Let us now derive the GW amplitude in relation to the orbital shape of the binary systems. GW amplitude from elliptical orbits ----------------------------------- Considering a binary system and the single components of eq.(\[ampli1\]), it is straightforward to show that $$\begin{array}{llllllll} h^{11}=-\frac{2G}{Rc^4}\frac{G (m_{1}+m_{2}) \mu (13 \epsilon \cos \phi+12 \cos2 \phi+\epsilon (4 \epsilon +3 \cos3 \phi))}{2 l}~, \\ \\ h^{22}=\frac{2G}{Rc^4}\frac{G (m_{1}+m_{2}) \mu (17 \epsilon \cos\phi+12 \cos2 \phi+\epsilon (8 \epsilon +3 \cos3 \phi))}{2 l} ~, \\ \\ h^{12}=h^{21}=-\frac{2G}{Rc^4}\frac{G (m_{1}+m_{2}) \mu (13 \epsilon \cos\phi+12 \cos2 \phi+\epsilon (4 \epsilon +3 \cos3 \phi))}{2 l} ~, \end{array}$$ so that the expected strain amplitude $h\simeq(h_{11}^2+h_{22}^2+2h_{12}^2)^{1/2}$ turns out to be $$\begin{aligned} h=\frac{G^3 (m_{1}+m_{2}) \mu^2}{c^4 Rl^2} (3 (13 \epsilon \cos\phi+12 \cos2 \phi+\epsilon (4 \epsilon +3 \cos3 \phi))^2+(17 \epsilon \cos\phi+ 12 \cos2 \phi+\epsilon (8 \epsilon +3 \cos3 \phi))^2)^{\frac{1}{2}} ~,\end{aligned}$$ which, as before, strictly depends on the initial conditions of the stellar encounter. A remark is in order at this point. A monochromatic gravitational wave has, at most, two independent degrees of freedom. In fact, in the TT gauge, we have $h_{+} = h_{11} + h_{22}$ and $h_{\times} = h_{12} + h_{21}$ (see e.g. [@bla]). As an example, the amplitude of gravitational wave is sketched in Fig. \[fig:ellisse800pc\] for a stellar encounter close to the Galactic Center. The adopted initial parameters are typical of a close impact and are assumed to be $b=1$ AU and $v_{0}=200$ Km$s^{-1}$, respectively. Here, we have fixed $M_{1}=M_{2}=1.4M_{\odot}$. The impact parameter is defined as $L=bv$ where $L$ is the angular momentum and $v$ the incoming velocity. We have chosen a typical velocity of a star in the galaxy and we are considering, essentially, compact objects with masses comparable to the Chandrasekhar limit $(\sim 1.4M_{\odot})$. This choice is motivated by the fact that ground-based experiments like VIRGO or LIGO expect to detect typical GW emissions from the dynamics of these objects or from binary systems composed by them (see e.g. [@maggiore]). ![The gravitational wave-forms from elliptical orbits shown as function of the polar angle $\phi$. We have fixed $M_{1}=M_{2}=1.4M_{\odot}$. $M_{2}$ is considered at rest while $M_{1}$ is moving with initial velocity $v_{0}=200$ Km$s^{-1}$ and an impact parameter $b=1$ AU. The distance of the GW source is assumed to be $R=8$ kpc and the eccentricity is $\epsilon= 0.2,0.5, 0.7.$ []{data-label="fig:ellisse800pc"}](ellisse800pc.eps) GW amplitude from parabolic and hyperbolic orbits ------------------------------------------------- In this case the single components of Eq.(\[ampli1\]) for a parabolic and hyperbolic orbits, are $$\begin{array}{llllllll} h^{11}=-\frac{G l^2 L^2\mu}{Rc^4}(13 \epsilon \cos \phi+12 \cos2 \phi+\epsilon (4 \epsilon +3 \cos3 \phi)) ~, \\ \\ h^{22}=\frac{Gl^2 L^2\mu}{Rc^4}(17 \epsilon \cos\phi+12 \cos2 \phi+\epsilon (8 \epsilon +3 \cos3 \phi)) ~, \\ \\ h^{12}=h^{21}=-\frac{3Gl^2 L^2\mu}{Rc^4} (4 \cos \phi+\epsilon (\cos2 \phi+3)) \sin\phi~, \end{array}$$ and then the expected strain amplitude is $$\begin{aligned} h=\frac{2 l^4 L^4 \mu ^2 }{c^4 R} \left(10 \epsilon ^4+9 \epsilon ^3\cos 3 \phi+59 \epsilon ^2 \cos2 \phi+59 \epsilon ^2+\left(47 \epsilon ^2+108\right) \epsilon\cos\phi +36\right)^{\frac{1}{2}} ~,\end{aligned}$$ which, as before, strictly depends on the initial conditions of the stellar encounter. We note that the gravitational wave amplitude has the same analytical expression for both cases and differs only for the value of $\epsilon$ which is $\epsilon=1$ if the motion is parabolic and the polar angle range is $\phi\in(-\pi,\pi)$, while it is $\epsilon>1$ and $\phi\in(-\pi,\pi)$ for hyperbolic orbits. In these cases, we have non-returning objects. The amplitude of the gravitational wave is sketched in Figs. \[fig:parabola\] and \[fig:iperbole8\] for stellar encounters close to the Galactic Center. As above, we consider a close impact and assume $b=1$ AU cm, $v_{0}=200$ Km$s^{-1}$ and $M_{1}=M_{2}=1.4M_{\odot}$. ![The gravitational wave-forms for a parabolic encounter as a function of the polar angle $\phi$. As above, $M_{1}=M_{2}=1.4M_{\odot}$ and $M_{2}$ is considered at rest. $M_{1}$ is moving with initial velocity $v_{0}=200$ Km$s^{-1}$ with an impact parameter $b=1$ AU. The distance of the GW source is assumed at $R=8$ kpc. The eccentricity is $\epsilon=1$. []{data-label="fig:parabola"}](parabola.eps) ![The gravitational wave-forms for hyperbolic encounters as function of the polar angle $\phi$. As above, we have fixed $M_{1}=M_{2}=1.4M_{\odot}$. $M_{2}$ is considered at rest while $M_{1}$ is moving with initial velocity $v_{0}=200$ Km$s^{-1}$ and an impact parameter $b=1$ AU. The distance of the source is assumed at $R=8$ kpc. The eccentricity is assumed with the values $\epsilon=1.2,1.5,1.7$ .[]{data-label="fig:iperbole8"}](iperbole8.eps) Rate and event number estimations ================================= An important remark is due at this point. A galaxy is a self-gravitating collisionless system where stellar impacts are very rare [@binney]. From the GW emission point of view, close orbital encounters, collisions and tidal interactions should be dealt on average if we want to investigate the gravitational radiation in a dense stellar system as we are going to do now. Let us give now an estimate of the stellar encounter rate producing GWs in some interesting astrophysical conditions like a typical globular cluster or towards the Galactic Center after we have discussed above the features distinguishing the various types of stellar encounters. Up to now, we have approximated stars as point masses. However, in dense regions of stellar systems, a star can pass so close to another that they raise tidal forces which dissipate their relative orbital kinetic energy. In some cases, the loss of energy can be so large that stars form binary or multiple systems; in other cases, the stars collide and coalesce into a single star; finally stars can exchange gravitational interaction in non-returning encounters. To investigate and parameterize all these effects, we have to compute the collision time $t_{coll}$, where $1/t_{coll}$ is the collision rate, that is, the average number of physical collisions that a given star suffers per unit time. For the sake of simplicity, we restrict to stellar clusters in which all stars have the same mass $m$. Let us consider an encounter with initial relative velocity $\mathbf{v}_{0}$ and impact parameter $b$. The angular momentum per unit mass of the reduced particle is $L=bv_{0}$. At the distance of closest approach, which we denote by $r_{coll}$, the radial velocity must be zero, and hence the angular momentum is $L=r_{coll}v_{max}$, where $v_{max}$ is the relative speed at $r_{coll}$. From the energy equation (\[eq:energia\]), we have $$b^{2}=r_{coll}^{2}+\frac{4Gmr_{coll}}{v_{0}^{2}}\,.\label{eq:b}$$ If we set $r_{coll}$ equal to the sum of the radii of two stars, then a collision will occur if and only if the impact parameter is less than the value of $b$, as determined by Eq.(\[eq:b\]). Let $f(\mathbf{v}_{a})d^{3}\mathbf{v}_{a}$ be the number of stars per unit volume with velocities in the range $\mathbf{v}_{a}+d^{3}\mathbf{v}_{a}.$ The number of encounters per unit time with impact parameter less than $b$ which are suffered by a given star is just $f(\mathbf{v}_{a})d^{3}\mathbf{v}_{a}$ times the volume of the annulus with radius $b$ and length $v_{0}$, that is, $$\int f(\mathbf{v}_{a})\pi b^{2}v_{0}d^{3}\mathbf{v}_{a}\label{eq:integrale}$$ where $v_{0}=\left|\mathbf{v-v}_{a}\right|$ and $\mathbf{v}$ is the velocity of the considered star. The quantity in Eq.(\[eq:integrale\]) is equal to $1/t_{coll}$ for a star with velocity $\mathbf{v}$: to obtain the mean value of $1/t_{coll}$, we average over $\mathbf{v}$ by multiplying (\[eq:integrale\]) by $f(\mathbf{v})/\nu$, where $\nu=\int f(\mathbf{v})d^{3}\mathbf{v}$ is the number density of stars and the integration is over $d^{3}\mathbf{v}$. Thus “$$\frac{1}{t_{coll}}=\frac{\nu}{8\pi^{2}\sigma^{6}}\int e^{-(v^{2}+v_{a}^{2})/2\sigma^{2}}\left(r_{coll}\left|\mathbf{v-v}_{a}\right|+ \frac{4Gmr_{coll}}{\left|\mathbf{v-v}_{a}\right|}\right)d^{3}\mathbf{v}d^{3}\mathbf{v}_{a} \label{eq:invtcoll}$$ We now replace the variable $\mathbf{v}_{a}$ by $\mathbf{V}=\mathbf{v-v}_{a}$. The argument of the exponential is then $-\left[\left(\mathbf{v}-\frac{1}{2}\mathbf{V}\right)^{2}+\frac{1}{4}V^{2}\right]/\sigma^{2}$, and if we replace the variable $\mathbf{v}$ by ${\displaystyle \mathbf{v}_{cm}=\mathbf{v}-\frac{1}{2}\mathbf{V}}$ (the center of mass velocity), then we have $$\frac{1}{t_{coll}}=\frac{\nu}{8\pi^{2}\sigma^{6}} \int e^{-(v_{cm}^{2}+V^{2})/2\sigma^{2}}\left(r_{coll}V+ \frac{4Gmr_{coll}}{V}\right)dV\,.\label{eq:invtcoll1}$$ The integral over $\mathbf{v}_{cm}$ is given by $$\int e^{-v_{cm}^{2}/\sigma^{2}}d^{3}\mathbf{v}_{cm}=\pi^{3/2}\sigma^{3}\,.\label{eq:intint}$$ Thus $$\frac{1}{t_{coll}}=\frac{\pi^{1/2}\nu}{2\sigma^{3}}\int_{\infty}^{0}e^{-V^{2}/4\sigma^{2}}\left(r_{coll}^{2}V^{3}+4GmVr_{coll}\right)dV\label{eq:invtcoll2}$$ The integrals can be easily calculated and then we find $$\frac{1}{t_{coll}}=4\sqrt{\pi}\nu\sigma r_{coll}^{2}+\frac{4\sqrt{\pi}\nu Gmr_{coll}}{\sigma}\,.\label{eq:invtcooll3}$$ The first term of this result can be derived from the kinetic theory. The rate of interaction is $\nu\Sigma\left\langle V\right\rangle$, where $\Sigma$ is the cross-section and $\left\langle V\right\rangle $ is the mean relative speed. Substituting $\Sigma=\pi r_{coll}^{2}$ and $\left\langle V\right\rangle =4\sigma/\sqrt{\pi}$ (which is appropriate for a Maxwellian distribution whit dispersion $\sigma$) we recover the first term of (\[eq:invtcooll3\]). The second term represents the enhancement in the collision rate by gravitational focusing, that is, the deflection of trajectories by the gravitational attraction of the two stars. If $r_{*}$ is the stellar radius, we may set $r_{coll}=2r_{*}$. It is convenient to introduce the escape speed from stellar surface, ${\displaystyle v_{*}=\sqrt{\frac{2Gm}{r_{*}}}}$, and to rewrite Eq.(\[eq:invtcooll3\]) as $$\Gamma=\frac{1}{t_{coll}}=16\sqrt{\pi}\nu\sigma r_{*}^{2}\left(1+\frac{v_{*}^{2}}{4\sigma^{2}}\right)=16\sqrt{\pi}\nu\sigma r_{*}^{2}(1+\Theta),\label{eq:invtcoll4}$$ where $$\Theta=\frac{v_{*}^{2}}{4\sigma^{2}}=\frac{Gm}{2\sigma^{2}r_{*}}\label{eq:safronov}$$ is the Safronov number [@binney]. In evaluating the rate, we are considering only those encounters producing gravitational waves, for example, in the LISA range, i.e. between $10^{-4}$ and $10^{-2}$ Hz (see e.g. [@Rub]). Numerically, we have $$\Gamma \simeq 5.5\times 10^{-10} \left(\frac{v}{10 {\rm km s^{-1}}}\right) \left(\frac{\sigma}{UA^2}\right) \left(\frac{{\rm 10 pc}}{R}\right)^3 {\rm yrs^{-1}}\qquad\Theta<<1\label{eq:thetamin}$$ $$\Gamma \simeq 5.5\times 10^{-10} \left(\frac{M}{10^5 {\rm M_{\odot}}}\right)^2 \left(\frac{v}{10 {\rm km s^{-1}}}\right) \left(\frac{\sigma}{UA^2}\right) \left(\frac{{\rm 10 pc}}{R}\right)^3 {\rm yrs^{-1}}\qquad\Theta>>1\label{eq:thetamagg}$$ If $\Theta>>1$, the energy dissipated exceeds the relative kinetic energy of the colliding stars, and the stars coalesce into a single star. This new star may, in turn, collide and merge with other stars, thereby becoming very massive. As its mass increases, the collision time is shorten and then there may be runaway coalescence leading to the formation of a few supermassive objects per clusters. If $\Theta<<1$, much of the mass in the colliding stars may be liberated and forming new stars or a single supermassive objects (see [@Belgeman; @Shapiro]). Note that when we have the effects of quasi-collisions in an encounter of two stars in which the minimum separation is several stellar radii, violent tides will raise on the surface of each star. The energy that excites the tides comes from the relative kinetic energy of the stars. This effect is important for $\Theta>>1$ since the loss of small amount of kinetic energy may leave the two stars with negative total energy, that is, as a bounded binary system. Successive peri-center passages will dissipates more energy by GW radiation, until the binary orbit is nearly circular with a negligible or null GW radiation emission. Let us apply these considerations to the Galactic Center which can be modelled as a system of several compact stellar clusters, some of them similar to very compact globular clusters with high emission in X-rays [@townes]. For a typical **compact stellar cluster** around the Galactic Center, the expected event rate is of the order of $2\times 10^{-9}$ yrs$^{-1}$ which may be increased at least by a factor $\simeq 100$ if one considers the number of globular clusters in the whole Galaxy eventually passing nearby the Galactic Center. If the compact stellar cluster at the Galactic Center is taken into account and assuming the total mass $M\simeq 3\times 10^6$ M$_{\odot}$, the velocity dispersion $\sigma\simeq $ 150 km s$^{-1}$ and the radius of the object $R\simeq$ 10 pc (where $\Theta=4.3$), one expects to have $\simeq 10^{-5}$ open orbit encounters per year. On the other hand, if a cluster with total mass $M\simeq 10^6$ M$_{\odot}$, $\sigma\simeq $ 150 km s$^{-1}$ and $R\simeq$ 0.1 pc is considered, an event rate number of the order of unity per year is obtained. These values could be realistically achieved by data coming from the forthcoming space interferometer LISA. As a secondary effect, the above wave-forms could constitute the “signature” to classify the different stellar encounters thanks to the differences of the shapes (see the above figures). Concluding Remarks ================== We have analyzed the gravitational wave emission coming from stellar encounters in Newtonian regime and in quadrupole approximation. In particular, we have taken into account the expected luminosity and the strain amplitude of gravitational radiation produced in tight impacts where two massive objects of $1.4M_{\odot}$ closely interact at an impact distance of $1AU$. Due to the high probability of such encounters inside rich stellar fields (e.g. globular clusters, bulges of galaxies and so on), the presented approach could highly contribute to enlarge the classes of gravitational wave sources (in particular, of dynamical phenomena capable of producing gravitational waves). In particular, a detailed theory of stellar orbits could improve the statistic of possible sources. We thank G. Covone for fruitful discussions on the topics of this paper. We thank also the Referee for the useful suggestions which allowed to improve the paper. [10]{} A. Abramovici et al., [*Science*]{} [**256**]{}, 325 (1992); http://www.ligo.org B. Caron et al., [*Class. Quant. Grav.*]{} [**14**]{}, 1461 (1997); http://www.virgo.infn.it H. Luck et al., [*Class. Quant. Grav.*]{} [**14**]{}, 1471 (1997); http://www.geo600.uni-hannover.de M. Ando et al., [*Phys. Rev. Lett.*]{} [**86**]{}, 3950 (2001); http://tamago.mtk.nao.ac.jp B.J. Meers and K. Strain, [*Phys. Rev.*]{} [**A 44**]{}, 4693 (1991); J. Mizuno et al., [*Phys. Lett.*]{} [**A 175**]{}, 273 (1993); A. Buonanno and Y. Chen, [*Phys. Rev.*]{} [**D 64**]{}, 042006 (2001). V.B. Braginsky and F.Ya. Khalili, [*Phys. Lett.*]{} [**A 147**]{}, 251 (1990); V.B. Braginsky, M.L. Gorodetsky,F.Ya. Khalili, and K.S. Thorne, [*Phys. Rev.*]{} [**A 61**]{}, 044002 (2000); P. Purdue, [*Phys. Rev.*]{} [**D 66**]{}, 022001 (2001); P. Purdue and Y. Chen, [*Phys. Rev.*]{} [**D 66**]{}, 122004 (2002); J. Harms et al. [*Phys. Rev.*]{} [**D 68**]{}, 042001 (2003); ; A. Buonanno and Y. Chen, [*Phys. Rev.*]{} [**D 69**]{} 102004 (2004). http://www.lisa-science.org P. Astone et al., [*Europhys. Lett.*]{} [**12**]{}, 5 (1990); G. Pallottino, in [*Gravitational Waves, Sources and Detectors*]{}, p 159 (Singapore, World Scientific, 1997); E. Mauceli et al., [*Phys. Rev.*]{} [**D 54**]{}, 1264 (1996); D. Blair et al., [*Phys. Rev. Lett.*]{} [**74**]{}, 1908 (1995); M. Cerdonio et al., [*Class. Quantum Grav.*]{} [**14**]{}, 1491 (1997). M. Kamionkowski, A. Kosowsky, and A. Stebbins, [*Phys. Rev.*]{} [**D 55**]{}, 7368 (1997); U. Seljak and M. Zaldarriaga, [*Phys. Rev.*]{} [**D 78**]{}, 2054 (1997); M. Kamionkowski and A. Kosowsky, [*Phys. Rev.*]{} [**D 57**]{}, 685 (1998). F.A. Jenet et al., [*Ap. J.*]{} [**653**]{}, 1571 (2006). http:// www.skatelescope.org E. Poisson, [*Living Rev. Rel.*]{} [**6**]{}, 3 (2004), and references therein. http://relativity.livingreviews.org Y.Mino, [*Prog. Theor. Phys.*]{} [**113**]{}, 733 (2005); ibid. [*Prog. Theor. Phys.*]{} [**115**]{}, 43 (2006); A. Pound, E. Poisson, and B.G. Nickel, [*Phys. Rev.*]{} [**D 72**]{}, 124001 (2005); L. Barack and C. Lousto, [*Phys. Rev.*]{} [**D 72**]{}, 104026 (2005); L. Barack and N. Sago, [*Phys. Rev.*]{} [**D 75**]{}, 064021 (2007). L. S. Finn and K. S. Thorne, [*Phys. Rev.*]{} [**D 62**]{}, 124021 (2000); A. Ori and K. Thorne, [*Phys. Rev.*]{} [**D 62**]{}, 124022 (2000); S. Babak et al. [*Phys. Rev.*]{} [**D 75**]{}, 024005 (2007); L. Barack and C. Cutler, [*Phys. Rev.*]{} [**D 70**]{}, 122002 (2004); ibid. [*Phys. Rev.*]{} [**D 70**]{}, 122002 (2004); G. Sigl, J. Schnittman, and A. Buonanno, [*Phys. Rev.*]{} [**D 75**]{}, 024034 (2007). R. Genzel and C.H. Townes, [*Ann. Rev. Astr. Astoph*]{}. [**25**]{}, 377 (1987). R. Genzel, D. Hollenbach and C.H. Townes, [*Rep. Prog. Phys.*]{} [**57**]{}, 417 (1994). K. Sellgreen, D.N.B. Hall, S.G. Kleinmann, and N.Z. Scoville, [*Ap. J.*]{} [**317**]{}, 881 (1987). S. Capozziello, G. Iovane, [*Phys. Lett.*]{} [**A 259**]{}, 185 (1999). A.M. Ghez, B.L. Klein, M. Morris, and E.E. Becklin, [*Ap. J.*]{} [**509**]{}, 678 (1998). R. Genzel, N. Thatte, A. Krabbe, H. Kroker and L.E. Tacconi-Garman, [*Ap. J.*]{} [**472**]{}, 153 (1996). S. Capozziello, M. De Laurentis, F. De Paolis, G. Ingrosso and A. Nucita, [*Mod. Phys. Lett.*]{} [**A 23**]{}, 99 (2008). P.C. Peters and J. Mathews, [*Phys. Rev.*]{} [**131**]{}, 435 (1963). P.C. Peters, [*Phys. Rev.*]{} [**136**]{}, 1224 (1964). B. F. Schutz, [*Class. Quantum Grav.*]{} [**16**]{}, 131 (1999). S.J. Kovacs and K.S. Thorne, [*Ap. J.*]{} [**217**]{}, 252 (1977). J.Binney and S.Tremaine, [*Galactic Dynamics*]{}, Princeton University Press, Princeton, New Jersey (1987). W.M. Smart, [*Textbook on Spherical Astronomy*]{}, Cambridge University Press, Cambridge (1977). L. Landau and E.M. Lifsits, [*Mechanics*]{}, Pergamon Press, New York (1973). C. W. Misner, K.S. Thorne, J.A. Wheeler, [*Gravitation*]{}, Freeman, New. York (1973). S.L. Shapiro and S.A. Teukolsy, [*Black Holes, White dwarfs and Neutron Stars*]{}, Chicago Univ. Press (Chicago) 1983. M. Maggiore, [*Gravitational Waves, Volume 1: Theory and Experiments*]{} Oxford Univ. Press (Oxford) 2007. K.S. Thorne, [*Rev. Mod. Phys.*]{} [**52**]{}, 285 (1980). L. Blanchet, [*Living Rev. Rel.*]{} [**9**]{}, 4 (2006). A. Buonanno, arXiv: 0709.4682\[gr-qc\] (2007). D. Clayton, [*Principles of Stellar Evolution and Nucleosynthesis*]{}, Chicago Univ. Press, Chicago (1983). S. Capozziello, M. De Laurentis, M. Francaviglia, [*Astrop. Phys.*]{} [**29**]{}, 125 (2008). L.J. Rubbo, K. Holley - Bockelmann, and L.S. Finn, [*Ap.J.*]{} [**649**]{}, L25 (2006). M.C. Belgeman and M.J. Rees *Mon. Not. Roy. Astron. Soc.,* [**185**]{}, 847 (1978). A.P. Lightman and S. L. Shapiro *Rev. Mod. Phys.,* [**50**]{}, 437 (1978). R. Genzel and C.H. Townes, [*Ann. Rev. Astron. Astroph.*]{} [**25**]{}, 1 (1987).

--- abstract: 'We study time evolution of entanglement between two qubits, which are part of a larger system, after starting from a random initial product state. We show that, due to randomness in the initial product state, entanglement is present only between directly coupled qubits and only for short times. Time dependence of the entanglement appears essentially independent of the specific hamiltonian used for time evolution and is well reproduced by a parameter-free two-body random matrix model.' author: - Marko Žnidarič title: 'Initial-state randomness as a universal source of decoherence' --- Introduction ============ In the last few years quantum entanglement is one of the most active research fields in quantum physics, for a review see Refs. [@entReview]. On microscopic level of coherent quantum systems entanglement can be used in various quantum protocols to perform non-classical operations. On the other hand, in the world of macroscopic objects with many degrees of freedom at high temperatures there is apparently no observable manifestation of entanglement. There have been many attempts to explain this classicality of macroscopic systems. Most notably, decoherence due to external degrees of freedom is usually credited as being responsible for the disappearance of entanglement from macroscopic superpositions. In a nutshell, the argument goes as follows: even if the system is in an entangled state at the beginning, , in a coherent superposition of two macroscopic states, time evolution will in general transform this coherent superposition into an incoherent (, classical) mixture. The reason for such decoherence is an always present residual coupling of our central system to many uncontrollable external degrees of freedom – the environment. For a review of decoherence see Ref. [@decoherence]. However, one must be aware that the evolution of the central system plus environment is still unitary and therefore, even though the final state of the central system and environment will presumably be very complex, in principle, it will be a pure state possessing some bipartite entanglement. The resolution of this apparent paradox is similar to the one with the second law of thermodynamics [@Lebowitz]. Increasing of the thermodynamic entropy with time is seemingly in contradiction with the reversibility of the underlying equations of motion. For explanation one can use two observations: (i) practicality – performing time reversal by, , reversing velocities of all particles might be close to impossible from a practical point of view; (ii) probability – initial conditions are prevalently of such form that in almost all cases the entropy will subsequently increase. Similar arguments can be used to explain the apparent lack of entanglement in quantum systems with many degrees of freedom. First, even though joint pure state representing central system and environment is bipartite entangled, the detection of entanglement might be close to impossible because it would require very complex measurements involving very many degrees of freedom. Indeed, using entanglement witnesses it has been shown that the detection of entanglement in a sufficiently complex state gets exponentially hard with increasing number of particles [@ranW]. For all practical purposes the detection of entanglement in such states is impossible. Second argument, that is the role played by initial conditions in the course of loosing entanglement by time evolution is the subject of present work. We are going to study how the entanglement between two qubits changes during hamiltonian time evolution. Hamiltonian evolution will act on a system of totally $n$ qubits, two of which will be chosen as our central system of interest while the remaining $n-2$ will act as the “environment”. The idea is to study how a general hamiltonian evolution changes entanglement of a smaller subsystem, whose degrees of freedom we presumably are able to control and therefore also measure its entanglement. Time evolution with a general hamiltonian, say quantum chaotic one, will in general produce states whose statistical properties are well described by those of random quantum states. For random quantum state on $n$ qubits one knows [@iden] that tracing out $n-2$ qubits will, for large $n$, with high probability result in a separable two qubit reduced density matrix. Therefore, sufficiently “complex” time evolution will eventually wipe out entanglement between two qubits. How are things then, for instance, in integrable systems, which in general do not generate completely random states? One point we have not touched so far is the role played by initial conditions. For integrable systems there can exist simple initial states for which entanglement will persist also for long times, nevertheless, as we will see, the majority of initial conditions is such that entanglement between two qubits will rapidly decay with time [*irrespective of the hamiltonian*]{}. This universality will be a consequence of the generic form of initial states – their randomness. The initial pure state will be chosen to be a product state on the central system (two qubits) and either random or random product state on the rest. Therefore, initially there will be no entanglement between the two chosen qubits. We are then going to study how much entanglement can be produced by various hamiltonian evolutions and how long will it take until it disappears. Entanglement will change with time due to two competing effects. One is entanglement production due to time evolution with non-separable hamiltonian, while the other is entanglement loss due to the spreading of initial state randomness throughout the system and the approach of system’s state to a random state. The net result will be the increase of entanglement at short times and a complete lack thereof after some critical time. In addition, time dependence of entanglement will turn out to be universal, that is independent of the specifics of the hamiltonian used in time evolution. There have been many studies of entanglement evolutions, let us here mention only those that are closer to the present work and deal with two qubit entanglement. Evolution of concurrence for initial product states has been studied for XY model in magnetic field in Ref. [@Monta:03], see also Ref. [@Rossignoli:07]. For sufficiently strong coupling between qubits concurrence initially increased with time after which it rapidly decreased to zero, similarly as in the present work. Evolution of entanglement for initially entangled states of two qubits (Bell states) which are weakly coupled to a generic environment has been studied in [@Pineda:07]. In such cases entanglement monotonically decreases with time from its maximal value at time zero. Evolution of entanglement for initial Bell state in an integrable XY model has been studied in [@Amico:04]. Our hamiltonian will be homogeneous in space and therefore the two qubits of the central system will be coupled. This must be contrasted to studies of the so-called environment induced entanglement generation [@envinduced], where two qubits are uncoupled. Disappearance of entanglement (on average) after finite critical time found in the present work is reminiscent of the so-called sudden death of entanglement, where initially entangled state of two uncoupled qubits becomes separable after a finite time of open system dynamics [@ESD]. A system consisting of two spins coupled to electrons has been studied for initial separable states in Ref. [@Gao:05], while time evolution of entanglement for initial thermal states in a XY model has been considered in Refs. [@Sen:05; @Huang:06]. For concurrence in a kicked Ising model see [@Arul]. In [@Sen:06] it has been found that starting from an initial non-disordered product state in a spin glass model entanglement can persist for long times. Von Neumann entropy of a block of spins in Ising model and for product initial state has been studied in [@Dur:05]. Quantifying entanglement ======================== We are going to study entanglement between two qubits which are in turn part of a larger $n$ qubit system. For two qubits positivity of partially transposed density matrix with respect to one qubit, $\rho^{T_A}$, is a necessary and sufficient condition for its separability [@PPT]. Negative eigenvalues (for two qubit states there can be at most one) therefore signal the presence of entanglement. A quantity measuring this is negativity [@neg] $N(\rho)$ which is equal to the sum of absolute values of negative eigenvalues of $\rho^{T_A}$ and can be defined as $$N(\rho)=\frac{\norm{\rho^{T_A}}_1-1}{2}, \label{eq:negdef}$$ with the trace norm $\norm{A}_1=\tr{\sqrt{A^\dagger A}}$. For two qubit density matrices it is simply $N(\rho)=|\lambda^{T_A}_{\rm min}|$ if the minimal eigenvalue $\lambda^{T_A}_{\rm min}$ is negative and $0$ otherwise. Entanglement of formation [@Fully], which quantifies quantum resources needed to create a given state, has nicer mathematical properties as negativity (or logarithmic negativity). For two qubit systems entanglement of formation $E_{\rm F}(\rho)$ can be calculated in terms of a simpler quantity called concurrence $C(\rho)$, defined as $$C(\rho)={\rm max}\,\{ 0, \lambda_1-\lambda_2-\lambda_3-\lambda_4\}, \label{eq:Cdef}$$ where $\lambda_i$ are square roots of decreasingly ordered eigenvalues of $(\rho\, \sigma^y\otimes \sigma^y \rho^* \sigma^y \otimes \sigma^y)$, calculated in standard computation basis. $E_{\rm F}(\rho)$ is then given by[@Cdef] $$E_{\rm F}(\rho)=H\left(\frac{1+\sqrt{1-C^2(\rho)}}{2}\right), \label{eq:Efdef}$$ with $H(x)=-x \log_2{x}-(1-x)\log_2{(1-x)}$ being a binary entropy. For pure states the entanglement of formation is given by the von Neumann entropy of the reduced density matrix, while it is defined by a convex roof extension (minimization over all convex realizations of $\rho$) for mixed states. A state is separable iff its concurrence or iff its negativity is zero. The third and last quantity used in measuring entanglement will be the so-called fully entangled fraction [@Fully] defined as $$f(\rho)={\rm max}\, \bracket{\psi}{\rho}{\psi}, \label{eq:fdef}$$ where maximization runs over all maximally entangled states obtained by local unitary transformations from maximally entangled state, , $\ket{\psi}=U_1 \otimes U_2 (\ket{00}+\ket{11})/\sqrt{2}$. One can distill maximally entangled singlets from an ensemble of $\rho$ using BBPSSW [@BBPSSW] protocol iff $f>1/2$. Fully entangled fraction $f$ can be used as a lower bound for the entanglement of formation [@Fully], $$E_{\rm F} \ge h(f),$$ where $h(f)$ is expressed in terms of binary entropy $H(x)$, $$h(f)= \left\{ \begin{array}{ll} H(\frac{1}{2}+\sqrt{f(1-f)}) & , f \ge \frac{1}{2} \\ 0 & , f < \frac{1}{2} \end{array} \right. .$$ In the above inequality an equal sign holds if $\rho$ is pure state. Note that $f\le 1/2$ does not necessarily mean that the entanglement of formation is zero. Fully entangled fraction also determines maximal teleportation fidelity [@maxTeleport]. Fully entangled fraction $f(\rho)$ is equal to the largest eigenvalue of the real part of the density matrix $\rho$ written in the Bell basis, in which all maximally entangled states have real expansion coefficients. If $f \ge 1/2$ this is in turn equal to [@Horodecki_fully] $$f(\rho)=\frac{1}{4}(1+\Theta(\rho)),\qquad \Theta(\rho)=\norm{T}_1=\tr{\sqrt{T^\dagger T}}, \label{eq:taudef}$$ where $T$ is a real $3\times 3$ dimensional correlation matrix given by $T_{ij}=\tr{(\rho\, \sigma^i \otimes \sigma^j)}$, with $\sigma^i$ being Pauli matrices, , $i,j \in \{x,y,z\}$. Because of its simple analytical form we are going to study $\Theta(\rho)$ (\[eq:taudef\]) rather than $f(\rho)$ (\[eq:fdef\]). They essentially give the same information in the interesting regime of $f>1/2$. If $\Theta(\rho) >1 $ then the state $\rho$ can be used in entanglement distillation. Norm of the correlation matrix $T$ can be used as a simple entanglement criterion also for many-qubit systems [@Badziag:08]. In the following we are therefore going to study negativity $N(\rho)$ (sometimes just minimal eigenvalue of the partially transposed matrix $\lambda^{T_A}_{\rm min}$), concurrence $C(\rho)$ and $\Theta(\rho)$ which is connected to the fully entangled fraction. As we will see, qualitatively all behave in the same way. From the analytical viewpoint though $\Theta(\rho)$ is the simplest quantity and is therefore the best candidate for an analytical treatment. Systems studied =============== We are going to study entanglement evolution for various one dimensional spin hamiltonians consisting of $n$ spin-1/2 particles. To verify that the results do not depend on the underlying dynamics we will use both chaotic and integrable systems. Heisenberg spin-$1/2$ model is an integrable model with hamiltonian $$H=\sum_i \sigma_i^x \sigma_{i+1}^x+\sigma_i^y \sigma_{i+1}^y+\sigma_i^z \sigma_{i+1}^z. \label{eq:heis}$$ We have checked that the results are similar for anisotropic Heisenberg model as well as for isotropic Heisenberg model in a tilted magnetic field. We can break integrability of the Heisenberg model by applying magnetic field, for instance, a staggered field in $z$-direction, $$H=\sum_i \sigma_i^x \sigma_{i+1}^x+\sigma_i^y \sigma_{i+1}^y+\sigma_i^z \sigma_{i+1}^z+\sum_i h_i \sigma_i^z, \label{eq:heisS}$$ where the strength of the magnetic field is $h_{2i}=0$ and $h_{2i+1}=-\frac{1}{2}$ on odd sites. ![Level spacing distribution for the Heisenberg model in a staggered field (\[eq:heisS\]) for $n=16$. Dashed line is Wigner-Dyson distribution holding for quantum chaotic systems.[]{data-label="fig:lsdHeisstag"}](hd10stagLSD.eps){width="3.3in"} As we can see in Fig. \[fig:lsdHeisstag\], spacing of neighboring energy levels agrees with the Wigner-Dyson distribution, $p(s)=s \pi/2 \exp{(-s^2 \pi/4)}$, which approximates distribution of spacings for gaussian orthogonal random matrix ensemble and is typical for quantum chaotic systems [@rmt]. Last model will be Ising chain in tilted magnetic field, $$H=\sum_i \sigma_i^x \sigma_{i+1}^x+\sigma_i^x+\sigma_i^z. \label{eq:TI}$$ Ising model in tilted magnetic field also displays typical signatures of quantum chaos [@TIsing], similarly as its time-dependent kicked version [@Prosen:00]. We have checked that the results are essentially the same also for integrable transverse Ising model. As we will see, all three models will display similar evolution of entanglement being in turn equal to the one for a two-body random matrix model. Therefore, our main focus will actually be on a two-body random matrix model, for which we have only nearest neighbor coupling terms, $$H=\sum_i h_{i,i+1}, \label{eq:rmt}$$ with $h_{i,i+1}$ acting nontrivially only on two qubits, for which it is a $4 \times 4$ random hermitian matrix, same for all coupled pairs and normalized as $\tr{(h_{i,i+1}^2)}=1$. A random hermitian matrix is a matrix whose matrix elements are independent random complex gaussian numbers [@rmt]. We always averaged over an ensemble of random matrices $h_{i,i+1}$. For all hamiltonians the geometry is that of an one-dimensional chain with open boundary conditions. The state at time $t$ is obtained as $\ket{\psi(t)}=\exp{(-\ii H t)} \ket{\psi(0)}$ from which we get the reduced density matrix for the two qubits between which we study entanglement, $$\rho(t)=\tr_{n-2}{\ket{\psi(t)}\bra{\psi(t)}}, \label{eq:rho}$$ where a subscript $n-2$ means tracing over $n-2$ qubits. The above $\rho(t)$ will then be used in calculating various entanglement measures. Two qubits in question will be either nearest neighbors, that is qubits directly coupled by the hamiltonian, or two qubits which are not directly coupled, , next nearest neighbors. Typically they will be chosen in the middle of the chain with the results being largely independent of their precise location. The initial pure state will be of two forms. Most of the time we are going to consider random product initial state, $$\ket{\psi(0)}=\ket{\chi}_1 \otimes \cdots \otimes \ket{\chi}_n, \label{eq:prodic}$$ where $\ket{\chi}_i$ is a random state of $i$th qubit, given by $$\ket{\chi}_i=\cos{\phi_i} \, {\rm e}^{\ii \alpha_i}\ket{0}_i+\sin{\phi_i}\, {\rm e}^{\ii \beta_i}\ket{1}_i, \label{eq:random1}$$ with $\beta_i,\alpha_i,\phi_i$ independent random numbers given as $\alpha_i=2\pi \xi$, $\beta_i=2\pi \xi$ and $\phi_i=\arcsin{\sqrt{\xi}}$ where all three $\xi$s are independent (for each qubit) uniform random numbers on interval $[0,1]$. Entanglement evolution ====================== Perturbative expansion ---------------------- Initially, at time $t=0$, our initial state is always of product form between two qubits in question and therefore there won’t be any entanglement, , $C(t=0)=0$, $\lambda^{T_A}_{\rm min}(t=0)=0$ and $\Theta(t=0)=1$. Subsequent evolution will entangle two qubits therefore one expects that the entanglement will gradually build up. For sufficiently short times, one can use perturbation theory to calculate the reduced density matrix $\rho(t)$ (\[eq:rho\]). Taking for $H$ nearest-neighbor hamiltonian with an arbitrary two-qubit coupling term $h^{(2)}$ we get after expanding propagator $\exp{(-i H t)}$ to the lowest order in time, $$\lambda^{T_A}_{\rm min}=-|\delta| t\,+{\cal O}(t^2),\qquad \delta=\bracket{\chi_A^\perp \chi_B^\perp}{h^{(2)}}{\chi_A \chi_B}. \label{eq:lminLR}$$ In the derivation of the above formula we assumed the initial product state on the two qubits in question while an arbitrary state is allowed on the remaining $n-2$ qubits, $\ket{\psi(0)}=\ket{\chi_A}\otimes\ket{\chi_B}\otimes\ket{\chi}_{n-2}$. States $\ket{\chi^\perp_{A,B}}$ are single qubit states orthogonal to $\ket{\chi_{A,B}}$ and because only absolute value of $\delta$ enters their phases do not matter. Similar calculation for concurrence (\[eq:Cdef\]) and $\Theta$ (\[eq:taudef\]) gives, $$\begin{aligned} C&=&2|\delta|t\,+{\cal O}(t^2) \nonumber \\ \Theta&=&1+4|\delta|t\,+ {\cal O}(t^2). \label{eq:CLR}\end{aligned}$$ We can see that the initial speed at which entanglement is produced depends only on a single matrix element of $h^{(2)}$ between the initial product state $\ket{\chi_A}\otimes \ket{\chi_B}$ and the corresponding orthogonal product state. Because for all quantities the initial time scale depends trivially on the value of $|\delta|$ we will measure time in rescaled dimensionless units $$\tau=t \overline{|\delta|}, \label{eq:tau}$$ with $\overline{|\delta|}$ being the average absolute value of a matrix element, where averaging is done over random initial single qubit states $\ket{\chi_{A,B}}$ (\[eq:random1\]). For the isotropic Heisenberg model (\[eq:heis\]) we get $\overline{|\delta|}=1$, for the Heisenberg model in a staggered field (\[eq:heisS\]) one gets $\overline{|\delta|}\approx 0.8882$ and for the Ising model in a tilted magnetic field (\[eq:TI\]) $\overline{|\delta|}\approx 0.6168$. For two-body random matrix model (\[eq:rmt\]) we can, instead of averaging over initial product states, average over ensemble of random matrices $h^{(2)}$, resulting in $\overline{|\delta|}=\sqrt{\pi}/4\approx 0.4431$. In all figures showing time dependence of entanglement we are going to use dimensionless time $\tau$. Numerical results ----------------- ![Average concurrence $C(\tau)$ (\[eq:Cdef\]) in Fig.(a) and average $\Theta(\tau)$ (\[eq:taudef\]) in Fig.(b) for various hamiltonians. Thick full curves are for two-body random matrix model (\[eq:rmt\]), thin dotted curves for the Ising model in a tilted magnetic field (\[eq:TI\]), short dashed curves are for isotropic Heisenberg model (\[eq:heis\]) while long dashed curves are for the Heisenberg model in a staggered field (\[eq:heisS\]). Averaging is performed over random product initial states (\[eq:prodic\]) for $n=16$. Times $\tau^*$ when $\Theta=1$ are $0.66$ for two-body random matrix model, $0.73$ for tilted Ising and $0.53$ for both Heisenberg models.[]{data-label="fig:H"}](cH.eps){width="3.3in"} ![Average concurrence $C(\tau)$ (\[eq:Cdef\]) in Fig.(a) and average $\Theta(\tau)$ (\[eq:taudef\]) in Fig.(b) for various hamiltonians. Thick full curves are for two-body random matrix model (\[eq:rmt\]), thin dotted curves for the Ising model in a tilted magnetic field (\[eq:TI\]), short dashed curves are for isotropic Heisenberg model (\[eq:heis\]) while long dashed curves are for the Heisenberg model in a staggered field (\[eq:heisS\]). Averaging is performed over random product initial states (\[eq:prodic\]) for $n=16$. Times $\tau^*$ when $\Theta=1$ are $0.66$ for two-body random matrix model, $0.73$ for tilted Ising and $0.53$ for both Heisenberg models.[]{data-label="fig:H"}](fH.eps){width="3.3in"} First, we performed numerical simulations of time evolution, calculating average $C(\tau)$, $\lambda_{\rm min}^{T_A}(\tau)$ and $\Theta(\tau)$ for different hamiltonians. Averaging has been done over random initial product states (\[eq:prodic\]). As one can see in Fig. \[fig:H\] the behavior is overall similar for all studied hamiltonians. For instance, from Fig. \[fig:H\]b we can see that for times larger than some critical $\tau^*$ the average $\Theta$ is below $1$ which means that the two qubits can not be used for distillation any more. Critical $\tau^*$ is for all models between $0.5$ and $0.75$. Similar dependence (not shown) is also obtained for the average minimal eigenvalues of partially transposed density matrix $\lambda_{\rm min}^{T_A}$, which also becomes positive at roughly the same $\tau^*$. Concurrence, seen in Fig. \[fig:H\]a, has a similar time dependence. The only difference is that for $\tau > \tau^*$ concurrence is not strictly zero but instead decays exponentially with time. This is a consequence of the fact that even though for $\tau> \tau^*$ the two qubits are on average not entangled any more there are still exponentially rare instances (product initial states) for which there is still some entanglement present. With time the probability of such entangled states decreases exponentially. In all cases dependence for small times agrees with analytical perturbative result for concurrence in eq. \[eq:CLR\] and Fig. \[fig:H\]a, for $\Theta$ in eq. \[eq:CLR\] and Fig. \[fig:H\]b, and for $\lambda_{\rm min}^{T_A}$ in eq. \[eq:lminLR\] and Fig. \[fig:lmin\]. Note that $\tau^*$ is a time when the average quantity (like $\Theta$ or $\lambda_{\rm min}^{\rm T_A}$) reaches a certain value ($1$ or $0$). It should not be confused with the average time $\bar{\tau}^*$ when $\Theta$ (or $\lambda_{\rm min}^{\rm T_A}$ or $C$) reaches $1$. For each individual initial condition time when a state gets separable, [*i.e.*]{} critical $\tau$, is of course the same as that of $C$ or $\lambda_{\rm min}^{\rm T_A}$. However, becouse distributions change with time, time $\tau^*$ is not exactly the same as the average time $\bar{\tau}^*$. The average times $\bar{\tau}^*$ are for $\lambda_{\rm min}^{\rm T_A}$ (as well as for $C$ or negativity) equal to $\bar{\tau}^*=0.72$ for two-body random matrix model and tilted Ising model and $\bar{\tau}^*=0.59$ for both Heisenberg models. The average times $\bar{\tau}^*$ for $\Theta(\tau)$ are on the other hand slightly different, $\bar{\tau}^*=0.64$ for two-body random matrix model, $\bar{\tau}^*=0.61$ for tilted Ising model and $\bar{\tau}^*=0.47$ for both Heisenberg models. Compare these $\bar{\tau}^*$ with $\tau^*$ listed in Figs. \[fig:H\]b, \[fig:f\] and \[fig:lmin\]. One can observe that $\bar{\tau}^*$ for $\Theta$ and $\lambda_{\rm min}^{\rm T_A}$ are slightly different, for instance in the case of a two-body random matrix model $0.64$ vs. $0.72$. The fact that $\bar{\tau}^*$ is for $\Theta$ smaller than for $\lambda_{\rm min}^{\rm T_A}$ is not a contradiction as $f \le 1/2$ does not necessarily mean that the entangelement of formation is zero [@Horodecki_fully]. We can see that due to randomness in the initial state the evolution of entanglement between two nearest-neighbor qubits shows universal-like behavior, that is time dependence which is to a large extend independent of the precise form of the underlying nearest-neighbor hamiltonian which generates evolution. Universality is not exact, there are still “signatures” of specific hamiltonian at intermediate times (for instance, compare curves for the Heisenberg and Ising model in Fig. \[fig:H\]), overall though the dependence is similar. Bell-like shape of entanglement is a consequence of two competing processes. On the one hand, non-separable evolution naturally tends to produce entanglement from an initially separable state while on the other hand, it will tend to destroy it because $\ket{\psi(t)}$ will approach a random state as time grows and the reduced density matrix will approach an identity matrix having zero entanglement [@iden]. The later process of destroying entanglement has two sources: first, randomness of the initial state is spread out throughout the system and second, dynamics itself will tend to produce random state. Universality is a consequence of randomness in the initial state, , of its generic separable form. For each hamiltonian there are rare specific separable initial states for which deviations from the above average behavior will be large. As we have seen in Fig. \[fig:H\] two-body random matrix model, which is a parameter-free model, describes evolution of entanglement sufficiently well also for other systems. Therefore, from now on we are going to focus on a two-body random matrix model, studying more precisely how the entanglement between two qubits evolves with time. Two-body random matrix model ---------------------------- ![Average $\Theta(\tau)$ (\[eq:taudef\]) for nearest-neighbor qubits ($r=1$), next nearest ($r=2$) and qubits separated by 2 qubits ($r=3$). For directly coupled qubits, $r=1$, states on average can not be used for distillation for times larger than $\tau^* \approx 2/3$. Chain curve (almost overlapping with the full curve for $r=1$) is $\Theta=\frac{1+4\tau}{1+6\tau^2}$. All is for two-body random matrix model with $n=16$ qubits and initial product states (\[eq:prodic\]). In the inset we show results for $r=1$ and smaller systems.[]{data-label="fig:f"}](f.eps){width="3.3in"} In Fig. \[fig:f\] we again show dependence of $\Theta$ (\[eq:taudef\]) on time for a two-body random matrix model. In addition to entanglement between nearest-neighbor qubits (denoted by distance $r=1$) we also show entanglement for next-nearest neighbors ($r=2$) and qubits separated by two other qubits ($r=3$). As one can see, for qubits which are not directly coupled by the hamiltonian, , $r=2$ and $r=3$ cases, $\Theta$ is always less than $1$. This happens because the production of entanglement depends on higher order terms, , for $r=2$ terms of form $h_{i,i+1} h_{i+1,i+2}$ are needed, whereas for nearest-neighbors a single term $h_{i,i+1}$ is already sufficient to entangle two qubits. As a consequence, entanglement production is slower the larger is the distance between qubits while entanglement destruction due to randomness is approximately independent of the distance. In all cases we show data for $n=16$ for which finite size effects are already small. For instance, the difference between $\tau^*$ for $n=16$ and $n=18$ is $0.01$ in the case of $r=1$. In the figure we also plot rational function $\Theta(\tau)=\frac{1+4\tau}{1+6\tau^2}$ which almost perfectly overlaps with the numerics for $r=1$. Note that for short times this of course agrees with our perturbative result (\[eq:CLR\]), seen as the initial line with slope $4$ in Figs. \[fig:H\] and \[fig:f\]. One would be tempted to think that the dependence beyond this short time, therefore also the above rational function, could be explained by higher order perturbation theory. Unfortunately it is not so. This rational dependence can not be explained by higher order perturbative calculation. In fact, we do not have any theoretical explanation for this almost perfect fit. Although going to perturbations of 2nd order in time will result in a rational function with the denominator and numerator being polynomials of order $2$ in $\tau$, the coefficients of polynomials are wrong. We have numerically checked that next perturbative orders also do not improve the situation. It therefore seems that the functional dependence of $\Theta$ (as well as of $C$ and $\lambda_{\rm min}^{T_A}$) for two-body random matrix model is essentially non-perturbative. This is in contrast with, for instance, purity or fidelity decay of initial pure states in the presence of weak coupling where perturbative approaches have been very successful [@echo; @Pineda:07]. ![Average minimal eigenvalue $\lambda_{\rm min}^{T_A}$ of partially transposed density operator. We show results for nearest-neighbor qubits ($r=1$), and those for qubits separated by one ($r=2$) and two ($r=3$) qubits. All is for two-body random matrix model with $n=16$ qubits and initial product states (\[eq:prodic\]). In the inset we show results for $r=1$ and smaller systems.[]{data-label="fig:lmin"}](lmin.eps){width="3.3in"} In Fig. \[fig:lmin\] we show numerical results for the average $\lambda_{\rm min}^{T_A}$. Overall, the dependence is very similar as for $\Theta(\tau)$, the only difference being that the time when $\lambda_{\rm min}^{T_A}$ becomes positive and the state gets separable is $\tau^* \approx 0.75$. In Fig. \[fig:c\] similar results are shown for concurrence and negativity. We can see that negativity and concurrence are almost the same for nearest-neighbor qubits. For next-nearest neighbor qubits ($r=2$) concurrence is this time non-zero but small as opposed to $\lambda_{\rm min}^{T_A}$ which is always positive. For qubits further apart, for instance $r=3$, concurrence is below the level of statistical fluctuations. We have checked that similar results are obtained also for other topologies of the coupling between qubits, , other than nearest-neighbor. In all situations entanglement is present for two qubit reduced density matrices between qubits directly coupled by the hamiltonian, whereas entanglement is small or zero for qubits which are not directly coupled. In all cases entanglement on average disappears after finite time. Difference from sudden death of entanglement phenomenon [@ESD] is that our two qubits are coupled and start from an initially separable state. In addition, our system is conservative, that is we have a finite “environment”. ![Average concurrence (\[eq:Cdef\]) for nearest-neighbor qubits ($r=1$) and next nearest-neighbors ($r=2$). With dotted line we also show negativity, $-2\cdot N(\rho)$ (\[eq:negdef\]), displaying essentially the same behavior as concurrence. For two-body random matrix model with $n=16$ qubits and initial product states.[]{data-label="fig:c"}](c.eps){width="3.3in"} So far we have always used product initial states, where states $\ket{\chi_i}$ on individual qubits were independent. As a final numerical calculation let us check how the results depend on the choice of an initial state. Because results are similar for all quantities studied we are going to show only $\Theta(\tau)$. Besides product initial state (\[eq:prodic\]) we used product initial state with states $\ket{\chi_i}$ on all qubits being the same. Because two-body random matrix model is invariant for single qubit rotations this is equivalent as choosing state $\ket{0\ldots 0}$ for the initial state and averaging over an ensemble of two-body random matrices. Second choice is an initial state which is of product form on the two qubits used for entanglement calculation and random on the remaining $n-2$ qubits, $$\ket{\psi(0)}=\ket{\chi}_{n-2} \otimes \ket{\chi_A}\otimes \ket{\chi_B}, \label{eq:prodr}$$ where $\ket{\chi_{A,B}}$ are random single qubit states and $\ket{\chi}_{n-2}$ is a random state of $n-2$ qubits. Numerical results are shown in Fig. \[fig:fo\]. We can see that, expectedly, entanglement decays slower for homogeneous initial state, $\ket{0\ldots 0}$, while it decays faster for the initial state having a full (nonseparable) random state on $n-2$ qubits (\[eq:prodr\]). Figuratively speaking one can say that the decay of entanglement is faster the more randomness there is in the initial state. ![Average $\Theta(\tau)$ for various choices of initial states. Full line is for initial product states (\[eq:prodic\]) (same data as in Fig. \[fig:f\]), chain curve is for initial state which is of product form just on the two qubits involved (\[eq:prodr\]), while dotted curve is for homogeneous initial state $\ket{0\ldots 0}$. All is for two-body random matrix model with $n=16$ qubits. Times when average $\Theta(\tau)<1$ are $0.42, 0.66$ and $0.86$.[]{data-label="fig:fo"}](fo.eps){width="3.3in"} Conclusion ========== We have studied how the entanglement between two qubits evolves with time when two qubits are part of a larger spin chain. Starting from an initial separable random product state entanglement first increases, reaching a maximal value, after which it decays to zero resulting in a separable state after finite time. Therefore, starting from a generic initial state possessing some randomness, two qubit reduced density matrix is on average entangled only for finite time and only for qubit pairs which are directly coupled by the hamiltonian. Time dependence of entanglement is almost independent of the specifics of the hamiltonian used for time evolution, being integrable or chaotic, and is well described by a two-body random matrix model. Results can be interpreted also in another way: it is hard to generate entanglement regardless of the dynamics if there is some randomness present in the initial state. This can be used to explain the lack of entanglement in small subsystems for generic initial conditions. The author would like to thank T. Prosen for reading the manuscript and anonymous referee for valuable suggestions and for bringing to our attention Refs. [@Sen:06] and [@Dur:05]. M. B. Plenio and S. Virmani, Quant. Inf. Comp. [**7**]{}, 1 (2007); R. Horodecki, preprint quant-ph/0702225. W. H. Zurek, Rev. Mod. Phys. [**75**]{}, 715 (2003). J. L. Lebowitz, Physica A [**263**]{}, 516 (1999). M. Žnidarič, J. Phys. A [**40**]{}, 13787 (2007). V. M. Kendon, J. Mod. Opt. [**49**]{}, 1709 (2002); V. M. Kendon, Phys. Rev. A [**66**]{}, 062310 (2002). S. Montangero, Phys. Rev. Lett [**91**]{}, 187901 (2003). R. Rossignoli and C. T. Schmiegelow, Phys. Rev. A [**75**]{}, 012320 (2007). C. Pineda and T. H. Seligman, Phys. Rev. A [**73**]{}, 012305 (2006); C. Pineda, New J. Phys. [**9**]{}, 106 (2007). L. Amico, Phys. Rev. A [**69**]{}, 022304 (2004). D. Braun, Phys. Rev. Lett. [**89**]{}, 277901 (2002); F. Benatti, Phys. Rev. Lett. [**91**]{}, 070402 (2003). L. Diosi, Lect. Notes Phys. [**622**]{}, 157 (2003); P. J. Dodd and J. J. Halliwell, Phys. Rev. A [**69**]{}, 052105 (2004); T. Yu and J. H. Eberly, Phys. Rev. Lett. [**93**]{}, 140404 (2004). Y. Gao and S.-J. Xiong, Phys. Rev. A [**71**]{}, 034102 (2005). A. Sen(Di), Phys. Rev. A [**72**]{}, 052319 (2005). Z. Huang and S. Kais, Phys. Rev. A [**73**]{}, 022339 (2006). A. Lakshminarayan and V. Subrahmanyam, Phys. Rev. A [**71**]{}, 062334 (2005). A. Sen(De), Phys. Rev. A [**74**]{}, 062309 (2006). W. D" ur, Phys. Rev. Lett. [**94**]{}, 097203 (2005). A. Peres, Phys. Rev. Lett. [**77**]{}, 1413 (1996); P. Horodecki, Phys. Lett. A [**232**]{}, 333 (1997). M. Horodecki, Phys. Lett. A [**223**]{}, 1 (1996). K.  . Zyczkowski, Phys. Rev. A [**58**]{}, 883 (1998). J. Eisert and M. B. Plenio, J. Mod. Opt. [**46**]{} 145 (1999). C. H. Bennett, Phys. Rev. A [**54**]{}, 3824 (1996). W. K. Wootters, Phys. Rev. Lett. [**80**]{}, 2245 (1998). C. H. Bennett, Phys. Rev. Lett. [**76**]{}, 722 (1996). R. Horodecki, Phys. Lett. A [**222**]{}, 21 (1996); M. Horodecki, Phys. Rev. A [**60**]{}, 1888 (1999). M. Horodecki, Phys. Rev. Lett. [**78**]{}, 574 (1997). P. Badziag, Phys. Rev. Lett [**100**]{}, 140403 (2008). M. L. Mehta, [*Random Matrices*]{}, 2nd. ed. (Academic Press, 1991). C. Mejia-Monasterio, Phys. Rev. A [**71**]{}, 062324 (2005); T. Prosen and M. Žnidarič, Phys. Rev. E [**75**]{}, 015202(R) (2007); J. Karthik, Phys. Rev. A [**75**]{}, 022304 (2007). T. Prosen, Phys. Rev. E [**65**]{}, 036208 (2002). T. Gorin, Phys. Rep. [**435**]{}, 33 (2006); I. Pižorn, Phys. Rev. B [**76**]{}, 035122 (2007).