Regularization of Discrete Ill-Conditioned Problems Done Right -- I

Abstract

When solving rank-deficient or discrete ill-posed problems by regularization methods, the choice of the regularization parameter is crucial. It is also of interest, the regularization norm used in the selection of the solution. In this work, we propose a stabilization of the existing regularization methods to address the delicate task of choosing this parameter. The analysis we carried out is independent of the chosen regularization norm. Under an unperturbed data least squares problem and of a maximal rank matrix, the stabilized-regularized method we propose provides the minimal norm solution whatever the chosen positive regularization parameter. And under a perturbed data least squares problem, this approach provides increasingly accurate and stable approximations of the minimal norm solution with respect to a refined mesh and a huge regularization parameter (over-regularization). We also investigate standard rank-deficient and ill-posed numerical examples corroborating the theoretical analysis, where the accuracy and the stability of the proposed approach is widely discussed.