Affine Invariant Covariance Estimation for Heavy-Tailed Distributions

In this work we provide an estimator for the covariance matrix of a heavy-tailed multivariate distributionWe prove that the proposed estimator admits an \textit{affine-invariant} bound of the form \[(1-\varepsilon) \mathbf{S} \preccurlyeq \widehat{\mathbf{S}} \preccurlyeq (1+\varepsilon) \mathbf{S}\]in high probability, where is the unknown covariance matrix, and is the positive semidefinite order on symmetric matrices. The result only requires the existence of fourth-order moments, and allows for where is a measure of kurtosis of the distribution, is the dimensionality of the space, is the sample size, and is the desired confidence level. More generally, we can allow for regularization with level , then gets replaced with the degrees of freedom number. Denoting the condition number of , the computational cost of the novel estimator is , which is comparable to the cost of the sample covariance estimator in the statistically interesing regime . We consider applications of our estimator to eigenvalue estimation with relative error, and to ridge regression with heavy-tailed random design.
View on arXiv