Nonlinear Residual Minimization by Iteratively Reweighted Least Squares

TL;DR

This paper develops and analyzes a generalized IRLS algorithm for solving nonlinear residual minimization problems, demonstrating convergence and effectiveness in phase retrieval and sparse recovery scenarios.

Contribution

It extends IRLS analysis to nonlinear problems, providing convergence results and practical algorithms for non-convex, non-smooth residual minimization.

Findings

IRLS converges under certain conditions for nonlinear residual problems.

The algorithm effectively solves phase retrieval and sparse recovery tasks.

Numerical experiments validate theoretical convergence and robustness.

Abstract

We address the numerical solution of minimal norm residuals of {\it nonlinear} equations in finite dimensions. We take inspiration from the problem of finding a sparse vector solution by using greedy algorithms based on iterative residual minimizations in the $\ell_p$-norm, for $1 \leq p \leq 2$. Due to the mild smoothness of the problem, especially for $p \to 1$, we develop and analyze a generalized version of Iteratively Reweighted Least Squares (IRLS). This simple and efficient algorithm performs the solution of optimization problems involving non-quadratic possibly non-convex and non-smooth cost functions, which can be transformed into a sequence of common least squares problems, which can be tackled more efficiently.While its analysis has been developed in many contexts when the model equation is {\it linear}, no results are provided in the {\it nonlinear} case. We address the convergence and the rate of error decay of IRLS for nonlinear problems. The convergence analysis is based on its reformulation as an alternating minimization of an energy functional, whose variables are the competitors to solutions of the intermediate reweighted least squares problems. Under specific conditions of coercivity and local convexity, we are able to show convergence of IRLS to minimizers of the nonlinear residual problem. For the case where we are lacking local convexity, we propose an appropriate convexification.. To illustrate the theoretical results we conclude the paper with several numerical experiments. We compare IRLS with standard Matlab functions for an easily presentable example and numerically validate our theoretical results in the more complicated framework of phase retrieval problems. Finally we examine the recovery capability of the algorithm in the context of data corrupted by impulsive noise where the sparsification of the residual is desired.