Linear covariance selection model via ℓ1-penalization

This paper presents a study on an ℓ1-penalized covariance regression method. Conventional approaches in high-dimensional covariance estimation often lack the flexibility to integrate external information. As a remedy, we adopt the regression-based covariance modeling framework and introduce a linear covariance selection model (LCSM) to encompass a broader spectrum of covariance structures when covariate information is available. Unlike existing methods, we do not assume that the true covariance matrix can be exactly represented by a linear combination of known basis matrices. Instead, we adopt additional basis matrices for a portion of the covariance patterns not captured by the given bases. To estimate high-dimensional regression coefficients, we exploit the sparsity-inducing ℓ1-penalization scheme. Our theoretical analyses are based on the (symmetric) matrix regression model with additive random error matrix, which allows us to establish new non-asymptotic convergence rates of the proposed covariance estimator. The proposed method is implemented with the coordinate descent algorithm. We conduct empirical evaluation on simulated data to complement theoretical findings and underscore the efficacy of our approach. To show a practical applicability of our method, we further apply it to the co-expression analysis of liver gene expression data where the given basis corresponds to the adjacency matrix of the co-expression network.