User-Friendly Covariance Estimation for Heavy-Tailed Distributions
We propose user-friendly covariance matrix estimators that are robust against heavy-tailed data. Specifically, we introduce element-wise and spectrum-wise truncation operators, as well as their M-estimator counterparts, to robustify the sample covariance matrix. Different from the classical notion of robustness which is typically characterized by the breakdown property, we focus on the tail robustness that is evidenced by the nonasymptotic deviation property of the estimator for data with only finite fourth moments. The key observation is that the robustification parameter needs to adapt to the sample size, dimensionality and moment to achieve optimal tradeoff between bias and robustness. Furthermore, to facilitate their practical use, we propose tuning-free procedures that automatically calibrate the tuning parameters. Applications to a series of structured models in high dimensions, including the bandable covariance estimation, sparse precision matrix estimation, low-rank covariance estimation, and covariance estimation and multiple testing under factor models are investigated. Numerical examples lend strong support to our proposed methodology.
READ FULL TEXT