Asymptotics of unitary and orthogonal matrix integrals (Q2389241)

The authors treat unitary and orthogonal matrix integrals of the form: \[ I_N(V, A_i^N) := \int_{ {\mathcal U}_N^m } e^{ N \text{Tr}( V( U_i, U_i^*, A_i^N, 1 \leqslant i \leqslant m ))} d U_1 \cdots d U_m, \tag{1} \] where \(( A_i^N, 1 \leqslant i \leqslant m )\) are \(N \times N\) deterministic uniformly bounded Hermitian matrices, \(d U\) denotes the Haar measure on the unitary group \({\mathcal U}_N\) normalized such that \(\int_{ {\mathcal U}_N } d U = 1\), \(V\) is a polynomial function in the non-commutative variables \(( U_i, U_i^*, A_i^N, 1 \leqslant i \leqslant m)\), and Tr denotes the usual trace on \(N \times N\) matrices given by \(\text{Tr}(A) =\sum_{i=1}^N A_{ii}\). They study the first order asymptotics of the matrix integrals (1) when the joint distribution of the \(( A_i^N\), \(1 \leqslant i \leqslant m)\) converges, namely, for all polynomial function \(P\) in \(m\) non-commutative indeterminates \[ \lim_{N \to \infty} \frac{1}{N} \text{Tr} ( P( A_i^N, 1 \leqslant i \leqslant m )) = \tau(P) \tag{2} \] for some linear functional \(\tau\) on the set of polynomials. The purpose of the paper is to answer affirmatively the following three open problems: {\parindent=9mm \begin{itemize}\item[(P1)] Does the limit of the matrix integrals of the type (1) exist for small parameters \(z\)? \item[(P2)] Does the power series \(\sum_k \frac{z^k}{k!} f_k(V, \tau)\) have a strictly positive radius of convergence? \item[(P3)] Is the limit of the matrix integral (1) equal to the sum of the power series? \end{itemize}} Here is the first result, which indicates that in small parameter regions, arbitrary unitary matrix integrals of the form (1) converge in the large \(N\) limit and match their formal expansion. Theorem A. Assume that the spectral radius of the matrices \(( A_i^N\), \(1 \leqslant i \leqslant m\), \(N \in {\mathbb N})\) is uniformly bounded by \(M\). Then there exists \(\varepsilon\) \(=\) \(\varepsilon(M,V)\) \(> 0\) such that, for \(z \in [- \varepsilon, \varepsilon]\), the limit \[ F_{V, \tau}(z) = \lim_{N \to \infty} \frac{1}{N^2} \log \int_{ {\mathcal U}_N^m} e^{ z N \text{Tr}( V( U_i, U_i^*, A_i^N, 1 \leqslant i \leqslant m ))} d U_1 \cdots d U_m \] exists. Moreover, \(F_{V, \tau}(z)\) is an analytic function of \(z \in {\mathbb C} \cap B(0, \varepsilon):=\{ z \in {\mathbb C}: \, | z | \leqslant \varepsilon \}\) and for all \(k \in {\mathbb N}\), \[ \left. \frac{\partial^k}{\partial z^k} F_{V, \tau}(z) \right|_{z=0} = f_k(V, \tau). \] The above latter result implies that \(F_{V, \tau}(z)\) has a positive radius of convergence. Secondly, the authors give a combinatorial model for the matrix integral asymptotics and investigate examples related to free probability and the HCIZ integral: \[ HCIZ(A, B) = \int_{ U \in {\mathcal U}_N } e^{ N \text{Tr} ( U^* A U B )} d U. \] This spherical integral is an important example of unitary matrix integral (1), which was originally studied by \textit{X. Harish-Chandra} [Am. J. Math. 79, 87--120 (1957; Zbl 0072.01901)] and also by Itzykson and Zuber, cf. \textit{B. Collins} and \textit{P. Śniady} [Commun. Math. Phys. 264, No.~3, 773--795 (2006; Zbl 1108.60004)] and \textit{A. Guionnet} and {M. Maïda} [J. Funct. Anal. 222, No.~2, 435--490 (2005; Zbl 1065.60023)]. Theorem B. Assume that \(( A_N, B_N)\) is a sequence of matrices with spectral radius bounded by one, such that \(\text{Tr} ( U_N^* A_N U_N B_N )\) is real for any \(N \times N\) unitary matrix \(U_N\) and the joint moments of \(A_N\), \(A_N^*\) and \(B_N\), \(B_N^*\) converge: \[ \lim_{N \to \infty} \frac{1}{N} \text{Tr} ( P( A_n, A_N^*)) = \tau_A(P) \] and \[ \lim_{N \to \infty} \frac{1}{N} \text{Tr} ( P( B_n, B_N^*)) = \tau_B(P). \] Then there exists \(\beta_0 > 0\) such that for all \(\beta \in [0, \beta_0]\), \[ \frac{1}{N^2} \log HCIZ( \beta A_N, \beta B_N) \] converges as \(N\) goes to infinity to a limit \(I( \tau_{\beta A}, \tau_{ \beta B} )\). Let \(V_t = \sum_i t_i q_i\) be a polynomial and \({\mathcal T}\) be the set of tracial states on the algebra generated by the variables \(( U_i, U_i^*, A_i )_{1 \leqslant i \leqslant m}\), namely, the set of linear forms on \[ {\mathbb C}_A^* \equiv {\mathbb C} \langle ( U_i, U_i^*, A_i )_{1 \leqslant i \leqslant m} \rangle \] such that for all \(P, Q \in {\mathbb C}_A^*\), \[ \mu( P P^*) \geq 0, \quad \mu(PQ) = \mu(QP), \quad \mu(1)=1. \] \({\mathcal M}\) denotes the subset of \({\mathcal T}\) such that \[ \mu( ( A_i)^{2n} ) \leqslant 1, \quad \forall n \in {\mathbb N}, \quad \forall i \in \{ 1, \dots, m \} \] for \(\mu \in {\mathcal T}\). The set \({\mathcal M}\) is endowed with weak topology, so \({\mathcal M}\) is the unit ball of \({\mathcal T}\) for the \(\text{weak}^*\) topology, and \({\mathcal M}\) becomes a compact metric space, by virtue of Banach-Alaoglu's theorem. Then the notation \({\mathcal M} |_{ ( A_i)_{1 \leqslant i \leqslant m} }\) stands for the set of tracial states of \({\mathcal M}\) restricted to the algebra generated by the \(( A_i )_{1 \leqslant i \leqslant m}\). In particular, the limiting distribution \(\tau\) given by (2) belongs to this apace \({\mathcal M} |_{ ( A_i)_{1 \leqslant i \leqslant m} }\). Then a tracial power state \(\mu \in {\mathcal M}\) is said to satisfy the Schwinger-Dyson equation \(SD^f[V_t, \tau]\) if and only if for all \(P \in\) \({\mathbb C} \langle (A_i)_{ 1 \leqslant i \leqslant m} \rangle\), \(\mu(P)\) \(=\) \(\tau(P)\) holds, and for all \(P \in\) \({\mathbb C} \langle ( U_i, U_i^*, A_i)_{1 \leqslant i \leqslant m} \rangle\) and for all \(i \in\) \(\{ 1, \dots, m \}\), the equation \[ \mu \otimes \mu ( \partial_i P) + \mu( D_i V_t P ) = 0 \] holds in the algebra \({\mathbb C}[ [t]]\) of power series in the variables \(t_1, \dots, t_n\). When \({\mathcal H}_N\) denotes the subset of Hermitian matrices of \({\mathcal M}_N\), then the microstates \(\Gamma_R =\) \(\Gamma_R( \mu, \varepsilon, k)\) is given by \[ \begin{multlined} \Gamma_R(\mu, \varepsilon, k) = \{ X_1, \dots, X_m \in {\mathcal H}_N : \\ \left| \frac{1}{N} \text{Tr} ( X_{i_1} \cdots X_{i_p}) - \mu( X_{i_1} \dots X_{i_p} ) \right| < \varepsilon, \\ p \leqslant k, i_{\ell} \in \{ 1, \dots, m \}, \| X_i \|_{\infty} \leqslant R \}. \end{multlined} \] Voiculescu's microstates free entropy \(\chi\) is defined as asymptotic volume of matrices whose empirical distribution approximates sufficiently well a tracial state. In fact, up to a Gaussian factor, the free entropy \(\chi\) is given by \[ \chi(\mu) = \limsup _{\substack{ \varepsilon \downarrow 0 \\ k \uparrow \infty, R \uparrow \infty }} \limsup_{N \to \infty} \frac{1}{N^2} \log \mu_N^{ \otimes m} ( \Gamma_R ). \] Moreover, under some additional assumptions, it can be shown that \[ \begin{aligned} \chi(\mu) & = \liminf _{\substack{ \varepsilon \downarrow 0 \\ k \uparrow \infty, R \uparrow \infty }} \liminf_{N \to \infty} \frac{1}{N^2} \log \mu_N^{ \otimes m} ( \Gamma_R(\mu, \varepsilon, k) ) \\ & = \sum_{i=1}^m I(\mu_i) - \mu(\nu) + F(V, \mu_i ). \end{aligned} \] Here \(I(\mu)\) is given by \(\int\int \log | x - y | d \mu(x) d \mu(y) - \frac{1}{2} \int x^2 d \mu(x) + Constant\), and \(\mu(\nu)\) can be written as an absolutely convergent power series like \[ \sum_k \prod_{i=1}^n \frac{t_i^{k_i}}{k_i !} \mu^k(V) \] and \(F(V, \mu_i)\) is given by a limit \(\lim_{N \to \infty} N^{-2} \log I_N ( V_t, A_i)\). It is interesting to note that these convergence results also lead to a new result of regularity of microstates. Finally, the authors generalize their approach to integrals over the orthogonal group. For other related works, see \textit{B. Collins} [Int. Math. Res. Not. 2003, No.~17, 953--982 (2003; Zbl 1049.60091)] for the Itzykson-Zuber integral; and [\textit{A. Guionnet} and \textit{O. Zeitouni}, J. Funct. Anal. 188, No.~2, 461--515 (2002; Zbl 1002.60021)] for large deviation asymptotics for spherical integrals.