A Convex-Nonconvex Framework for Enhancing Minimization Induced Penalties

TL;DR

This paper introduces a convex-nonconvex framework that enhances minimization induced penalties using a generalized Moreau envelope, improving signal component estimation while maintaining convexity for efficient optimization.

Contribution

It proposes a novel GME-MI penalty that enhances MI penalties with nonconvex features and provides conditions for convexity and a convergent algorithm.

Findings

The GME-MI penalty improves signal component estimation.

The framework guarantees convergence to a global optimum.

Numerical examples demonstrate effectiveness of the approach.

Abstract

This paper presents a novel framework for nonconvex enhancement of minimization induced (MI) penalties while preserving the overall convexity of associated regularization models. MI penalties enable the adaptation to certain signal structures via minimization, but often underestimate significant components owing to convexity. To overcome this shortcoming, we design a generalized Moreau enhanced minimization induced (GME-MI) penalty by subtracting from the MI penalty its generalized Moreau envelope. While the proposed GME-MI penalty is nonconvex in general, we derive an overall convexity condition for the GME-MI regularized least-squares model. Moreover, we present a proximal splitting algorithm with guaranteed convergence to a globally optimal solution of the GME-MI model under the overall convexity condition. Numerical examples illustrate the effectiveness of the proposed framework.