14
6

Sharper dimension-free bounds on the Frobenius distance between sample covariance and its expectation

Abstract

We study properties of a sample covariance estimate Σ^=(X1X1++XnXn)/n\widehat \Sigma = (\mathbf{X}_1 \mathbf{X}_1^\top + \ldots + \mathbf{X}_n \mathbf{X}_n^\top) / n, where X1,,Xn\mathbf{X}_1, \dots, \mathbf{X}_n are i.i.d. random elements in Rd\mathbb R^d with EX1=0\mathbb E \mathbf{X}_1 = \mathbf{0}, EX1X1=Σ\mathbb E \mathbf{X}_1 \mathbf{X}_1^\top = \Sigma. We derive dimension-free bounds on the squared Frobenius norm of (Σ^Σ)(\widehat\Sigma - \Sigma) under reasonable assumptions. For instance, we show that Σ^ΣF2EΣ^ΣF2=O(Tr(Σ2)/n)| \|\widehat\Sigma - \Sigma\|_{\rm F}^2 - \mathbb E \|\widehat\Sigma - \Sigma\|_{\rm F}^2| = \mathcal O({\rm{Tr}}(\Sigma^2) / n) with overwhelming probability, which is a significant improvement over the existing results. This leads to a bound the ratio Σ^ΣF2/EΣ^ΣF2\|\widehat\Sigma - \Sigma\|_{\rm F}^2 / \mathbb E \|\widehat\Sigma - \Sigma\|_{\rm F}^2 with a sharp leading constant when the effective rank r(Σ)=Tr(Σ)/Σ\mathtt{r}(\Sigma) = {\rm Tr}(\Sigma) / \|\Sigma\| and n/r(Σ)6n / \mathtt{r}(\Sigma)^6 tend to infinity: Σ^ΣF2/EΣ^ΣF2=1+O(1/r(Σ))\|\widehat\Sigma - \Sigma\|_{\rm F}^2 / \mathbb E \|\widehat\Sigma - \Sigma\|_{\rm F}^2 = 1 + \mathcal O(1 / \mathtt{r}(\Sigma)).

View on arXiv
Comments on this paper