Inexact DC Algorithms in Hilbert Spaces with Applications to PDE-Constrained Optimization

TL;DR

This paper introduces inexact adaptive algorithms for minimizing difference-of-convex functions in Hilbert spaces, with applications to PDE-constrained optimization, providing convergence guarantees and numerical validation.

Contribution

The paper proposes I-ADCA, an inexact adaptive DC algorithm with convergence guarantees, tailored for PDE-constrained problems with nonconvex regularizers.

Findings

Convergence to stationary points under inexact evaluations

Explicit convergence rates under Polyak-Lojasiewicz property

Numerical experiments demonstrating efficiency and accuracy

Abstract

In this paper, we design and apply novel inexact adaptive algorithms to deal with minimizing difference-of-convex (DC) functions in Hilbert spaces. We first introduce I-ADCA, an inexact adaptive counterpart of the well-recognized DCA (difference-of-convex algorithm), that allows inexact subgradient evaluations and inexact solutions to convex subproblems while still guarantees global convergence to stationary points. Under a Polyak-Lojasiewicz type property for DC objectives, we obtain explicit convergence rates for the proposed algorithm. Our main application addresses elliptic optimal control problems with control constraints and nonconvex $L^{1-2}$ sparsity-enhanced regularizers admitting a DC decomposition. Employing I-ADCA and appropriate versions of finite element discretization leads us to an efficient procedure for solving such problems with establishing its well-posedness and error bound estimates confirmed by numerical experiments.