English

Stochastic Variance-Reduced Prox-Linear Algorithms for Nonconvex Composite Optimization

Optimization and Control 2021-05-17 v2

Abstract

We consider minimization of composite functions of the form f(g(x))+h(x)f(g(x))+h(x), where ff and hh are convex functions (which can be nonsmooth) and gg is a smooth vector mapping. In addition, we assume that gg is the average of finite number of component mappings or the expectation over a family of random component mappings. We propose a class of stochastic variance-reduced prox-linear algorithms for solving such problems and bound their sample complexities for finding an ϵ\epsilon-stationary point in terms of the total number of evaluations of the component mappings and their Jacobians. When gg is a finite average of NN components, we obtain sample complexity O(N+N4/5ϵ1)\mathcal{O}(N+ N^{4/5}\epsilon^{-1}) for both mapping and Jacobian evaluations. When gg is a general expectation, we obtain sample complexities of O(ϵ5/2)\mathcal{O}(\epsilon^{-5/2}) and O(ϵ3/2)\mathcal{O}(\epsilon^{-3/2}) for component mappings and their Jacobians respectively. If in addition ff is smooth, then improved sample complexities of O(N+N1/2ϵ1)\mathcal{O}(N+N^{1/2}\epsilon^{-1}) and O(ϵ3/2)\mathcal{O}(\epsilon^{-3/2}) are derived for gg being a finite average and a general expectation respectively, for both component mapping and Jacobian evaluations.

Keywords

Cite

@article{arxiv.2004.04357,
  title  = {Stochastic Variance-Reduced Prox-Linear Algorithms for Nonconvex Composite Optimization},
  author = {Junyu Zhang and Lin Xiao},
  journal= {arXiv preprint arXiv:2004.04357},
  year   = {2021}
}
R2 v1 2026-06-23T14:45:07.475Z