Analyzing the speed of convergence in nonsmooth optimization via the Goldstein subdifferential with application to descent methods

Abstract

The Goldstein $\varepsilon$-subdifferential is a relaxed version of the Clarke subdifferential which has recently appeared in several algorithms for nonsmooth optimization. With it comes the notion of $(\varepsilon,δ)$-critical points, which are points in which the element with the smallest norm in the $\varepsilon$-subdifferential has norm at most $δ$. To obtain points that are critical in the classical sense, $\varepsilon$ and $δ$ must vanish. In this article, we analyze at which speed the distance of $(\varepsilon,δ)$-critical points to the minimum vanishes with respect to $\varepsilon$ and $δ$. Afterwards, we apply our results to gradient sampling methods and perform numerical experiments. Throughout the article, we put a special emphasis on supporting the theoretical results with simple examples that visualize them.