An Inexact Proximal Linearized DC Algorithm with Provably Terminating Inner Loop

TL;DR

This paper introduces tPLDCA, an inexact proximal linearized DC algorithm with a guaranteed finite inner loop termination, improving convergence guarantees over existing methods for difference-of-convex programs.

Contribution

The paper proposes a novel inexact DC algorithm, tPLDCA, with provable finite inner loop termination and a practical implementation for a class of DC functions.

Findings

tPLDCA converges with finite inner loop iterations.

The proposed surrogate simplifies implementation.

Numerical results confirm effectiveness.

Abstract

Standard approaches to difference-of-convex (DC) programs require exact solution to a convex subproblem at each iteration, which generally requires noiseless computation and infinite iterations of an inner iterative algorithm. To tackle these difficulties, inexact DC algorithms have been proposed, mostly by relaxing the convex subproblem to an approximate monotone inclusion problem. However, there is no guarantee that such relaxation can lead to a finitely terminating inner loop. In this paper, we point out the termination issue of existing inexact DC algorithms by presenting concrete counterexamples. Exploiting the notion of $\epsilon$-subdifferential, we propose a novel inexact proximal linearized DC algorithm termed tPLDCA. Despite permission to a great extent of inexactness in computation, tPLDCA enjoys the same convergence guarantees as exact DC algorithms. Most noticeably, the inner loop of tPLDCA is guaranteed to terminate in finite iterations as long as the inner iterative algorithm converges to a solution of the proximal point subproblem, which makes an essential difference from prior arts. In addition, by assuming the first convex component of the DC function to be pointwise maximum of finitely many convex smooth functions, we propose a computational friendly surrogate of $\epsilon$-subdifferential, whereby we develop a feasible implementation of tPLDCA. Numerical results demonstrate the effectiveness of the proposed implementation of tPLDCA.