Gaussian fluctuation for Gaussian Wishart matrices of overall correlation

Abstract

In this note, we study the Gaussian fluctuations for the Wishart matrices $d^{-1}\mathcal{X}_{n, d}\mathcal{X}^{T}_{n, d}$, where $\mathcal{X}_{n, d}$ is a $n\times d$ random matrix whose entries are jointly Gaussian and correlated with row and column covariance functions given by $r$ and $s$ respectively such that $r(0)=s(0)=1$. Under the assumptions $s\in \ell^{4/3}(\mathbb{Z})$ and $\|r\|_{\ell^1(\mathbb{Z})}< \sqrt{6}/2$, we establish the $\sqrt{n^3/d}$ convergence rate for the Wasserstein distance between a normalization of $d^{-1}\mathcal{X}_{n, d}\mathcal{X}^{T}_{n, d}$ and the corresponding Gaussian ensemble. This rate is the same as the optimal one computed in \cite{JL15,BG16,BDER16} for the total variation distance, in the particular case where the Gaussian entries of $\mathcal{X}_{n, d}$ are independent. Similarly, we obtain the $\sqrt{n^{2p-1}/d}$ convergence rate for the Wasserstein distance in the setting of random $p$-tensors of overall correlation. Our analysis is based on the Malliavin-Stein approach.