Overcoming the limitations of phase transition by higher order analysis of regularization techniques (Q1991696)

This paper establishes asymptotic properties of the \(l_q\)-regularized least squares estimators, for different values of \(q\). It focuses on the high dimensional setting with sparsity. A small error framework is built to develop the study. It demonstrates, by exploring the case of \(q\) in the interval \((1,2]\), the inadequacy of the phase-transition approach and then extends the analysis by considering second-order components of the asymptotic risk. Methods for improving the approximations when first and second-order are inaccurate are proposed. Results of simulations performed to numerically evaluate the estimators in the reality of samples of finite size are presented.