A Model Selection Criterion for High-Dimensional Linear Regression

Statistical model selection is a great challenge when the number of accessible measurements is much smaller than the dimension of the parameter space. We study the problem of model selection in the context of subset selection for high-dimensional linear regressions. Accordingly, we propose a new model selection criterion with the Fisher information that leads to the selection of a parsimonious model from all the combinatorial models up to some maximum level of sparsity. We analyze the performance of our criterion as the number of measurements grows to infinity, as well as when the noise variance tends to zero. In each case, we prove that our proposed criterion gives the true model with a probability approaching one. Additionally, we devise a computationally affordable algorithm to conduct model selection with the proposed criterion in practice. Interestingly, as a side product, our algorithm can provide the ideal regularization parameter for the Lasso estimator such that Lasso selects the true variables. Finally, numerical simulations are included to support our theoretical findings.