Multivariate normal approximation for traces of random unitary matrices (Q2072086)

This paper studies the convergence of the real and imaginary part of the traces of the first \(m\) powers of a random unitary \(n\times n\) matrix as \(n\) and \(m\) tend to infinity, and when the unitary matrix is distributed according to the Haar measure. For fixed \(m\), it is known that the limit distribution is a multivariate Gaussian distribution, see, e.g., [\textit{P. Diaconis} and \textit{M. Shahshahani}, J. Appl. Probab. 31A, 49--62 (1994; Zbl 0807.15015)], and that in the total variation distance this convergence is superexponentially fast, see [\textit{K. Johansson}, Ann. Math. (2) 145, No. 3, 519--545 (1997; Zbl 0883.60010)]. In the present paper the parameter \(m\) is allowed to tend to infinity almost as \(\sqrt{n}\), and explicit estimates for the total variation distance and also the Kantorovich-Wasserstein distance are given. Roughly speaking, it is shown that for \(m\ll \sqrt{n}\) the order of convergence of the traces to a multivariate Gaussian is \(\Gamma\left(\frac{n}{m}+1\right)^{-1}\) times a correction. The methods of the proof include formulae for the Toeplitz determinants, change of variables, and Stein's method. A special merit of the work is that the authors kept explicitly track of the constants and the dependence on \(m\).