Stochastic Subgradient Descent Escapes Active Strict Saddles on Weakly Convex Functions
Abstract
In non-smooth stochastic optimization, we establish the non-convergence of the stochastic subgradient descent (SGD) to the critical points recently called active strict saddles by Davis and Drusvyatskiy. Such points lie on a manifold where the function has a direction of second-order negative curvature. Off this manifold, the norm of the Clarke subdifferential of is lower-bounded. We require two conditions on . The first assumption is a Verdier stratification condition, which is a refinement of the popular Whitney stratification. It allows us to establish a reinforced version of the projection formula of Bolte \emph{et.al.} for Whitney stratifiable functions, and which is of independent interest. The second assumption, termed the angle condition, allows to control the distance of the iterates to . When is weakly convex, our assumptions are generic. Consequently, generically in the class of definable weakly convex functions, the SGD converges to a local minimizer.
Cite
@article{arxiv.2108.02072,
title = {Stochastic Subgradient Descent Escapes Active Strict Saddles on Weakly Convex Functions},
author = {Pascal Bianchi and Walid Hachem and Sholom Schechtman},
journal= {arXiv preprint arXiv:2108.02072},
year = {2023}
}
Comments
Accepted for publication in Mathematics of Operations Research