Almost sharp covariance and Wishart-type matrix estimation

Abstract: Let $X_1,..., X_n \in \mathbb{R}^d$ be independent Gaussian random vectors with independent entries and variance profile $(b_{ij}){i \in [d],j \in [n]}$. A major question in the study of covariance estimation is to give precise control on the deviation of $\sum{j \in [n]}X_jX_j^T-\mathbb{E} X_jX_j^T$. We show that under mild conditions, we have \begin{align*} \mathbb{E} \left|\sum_{j \in [n]}X_jX_j^T-\mathbb{E} X_jX_j^T\right| \lesssim \max_{i \in [d]}\left(\sum_{j \in [n]}\sum_{l \in [d]}b_{ij}^2b_{lj}^2\right)^{1/2}+\max_{j \in [n]}\sum_{i \in [d]}b_{ij}^2+\text{error}. \end{align*} The error is quantifiable, and we often capture the $4$th-moment dependency already presented in the literature for some examples. The proofs are based on the moment method and a careful analysis of the structure of the shapes that matter. We also provide examples showing improvement over the past works and matching lower bounds.