Weighting operators for sparsity regularization

TL;DR

This paper explores alternative weighting operators for sparsity regularization to improve solution recovery, reducing computational demands and error amplification, with theoretical analysis and numerical experiments.

Contribution

It introduces broader classes of weighting schemes for regularization, extending previous methods involving ^\u221dA, and demonstrates their effectiveness through analysis and experiments.

Findings

Broader class of weighting operators can be used for regularization.

Certain weighting schemes achieve near-perfect basis recovery.

Numerical experiments validate theoretical results.

Abstract

Standard regularization methods typically favor solutions which are in, or close to, the orthogonal complement of the null space of the forward operator/matrix $\mathsf{A}$. This particular biasedness might not be desirable in applications and can lead to severe challenges when $\mathsf{A}$ is non-injective. We have therefore, in a series of papers, investigated how to "remedy" this fact, relative to a chosen basis and in a certain mathematical sense: Based on a weighting procedure, it turns out that it is possible to modify both Tikhonov and sparsity regularization such that each member of the chosen basis can be almost perfectly recovered from their image under $\mathsf{A}$. In particular, we have studied this problem for the task of using boundary data to identify the source term in an elliptic PDE. However, this weighting procedure involves $\mathsf{A}^\dagger \mathsf{A}$, where $\mathsf{A}^\dagger$ denotes the pseudo inverse of $\mathsf{A}$, and can thus be CPU-demanding and lead to undesirable error amplification. We therefore, in this paper, study alternative weighting approaches and prove that some of the recovery results established for the methodology involving $\mathsf{A}$ hold for a broader class of weighting schemes. In fact, it turns out that "any" linear operator $\mathsf{B}$ has an associated proper weighting defined in terms of images under $\mathsf{B}\mathsf{A}$. We also present a series of numerical experiments, employing different choices of $\mathsf{B}$.