Model selection with covariance matching based non-negative lasso

Abstract

We consider the problem of model selection for high-dimensional linear regressions in the context of support recovery with multiple measurement vectors available. Here, we assume that the regression coefficient vectors have a common support and the elements of the additive noise vector are potentially correlated. Accordingly, to estimate the support, we propose a non-negative Lasso estimator that is based on covariance matching techniques. We provide deterministic conditions under which the support estimate of our method is guaranteed to match the true support. Further, we use the extended Fisher information criterion to select the tuning parameter in our non-negative Lasso. We also prove that the extended Fisher information criterion can find the true support with probability one as the number of rows in the design matrix grows to infinity. The numerical simulations confirm that our support estimate is asymptotically consistent. Finally, the simulations also show that the proposed method is robust to high correlation between columns of the design matrix.