Model Consistency of Iterative Regularization for Low-Complexity Regularization

Regularization is a core component of modern inverse problems as it allows to establish well-posedness to the solution of interests. Popular regularization approaches include variational regularization and iterative regularization. The former one can be tackled by solving a variational optimization problem, which is the sum of a regularization term and a data-fidelity term balanced by a proper weight, while the latter one chooses a proper stopping time to avoid overfitting to the noise. In the study of regularization, an important topic is the relation between the solution obtained by regularization and the original ground truth. When the ground truth has low-complexity structure which is encoded as the "model", a sensitivity property shows that the solution obtained from proper regularization that promotes the same structure is robust to small perturbations, this is called "model consistency". For variational regularization, model consistency of linear inverse problem is studied in [1]. While, for iterative regularization, the existence of model consistency is an open problem. In this paper, based on a recent development of partial smoothness which is also considered in [1], we show that if the noise level is sufficiently small and a proper stopping time is chosen, the solution by iterative regularization also achieves model consistency and more exhibit local linear convergence behavior. Numerical simulations are provided to verify our theoretical findings.