Covariance Estimation: The GLM and Regularization Perspectives

TL;DR

This paper surveys methods for covariance matrix estimation, emphasizing regression-based approaches like Cholesky decomposition and regularization techniques, highlighting their advantages in high-dimensional settings and their interpretability.

Contribution

It unifies covariance estimation perspectives through regression formulations, discusses advantages of Cholesky-based methods, and reviews regularization techniques with theoretical insights.

Findings

Regression-based covariance estimation guarantees positive definiteness.

Cholesky decomposition offers interpretable, unconstrained parameterization.

Elementwise regularization methods have favorable asymptotic properties.

Abstract

Finding an unconstrained and statistically interpretable reparameterization of a covariance matrix is still an open problem in statistics. Its solution is of central importance in covariance estimation, particularly in the recent high-dimensional data environment where enforcing the positive-definiteness constraint could be computationally expensive. We provide a survey of the progress made in modeling covariance matrices from two relatively complementary perspectives: (1) generalized linear models (GLM) or parsimony and use of covariates in low dimensions, and (2) regularization or sparsity for high-dimensional data. An emerging, unifying and powerful trend in both perspectives is that of reducing a covariance estimation problem to that of estimating a sequence of regression problems. We point out several instances of the regression-based formulation. A notable case is in sparse estimation of a precision matrix or a Gaussian graphical model leading to the fast graphical LASSO algorithm. Some advantages and limitations of the regression-based Cholesky decomposition relative to the classical spectral (eigenvalue) and variance-correlation decompositions are highlighted. The former provides an unconstrained and statistically interpretable reparameterization, and guarantees the positive-definiteness of the estimated covariance matrix. It reduces the unintuitive task of covariance estimation to that of modeling a sequence of regressions at the cost of imposing an a priori order among the variables. Elementwise regularization of the sample covariance matrix such as banding, tapering and thresholding has desirable asymptotic properties and the sparse estimated covariance matrix is positive definite with probability tending to one for large samples and dimensions.