English

Adaptive Regularization within Trust Region Methods for Stochastic Nonconvex Optimization

Optimization and Control 2026-04-30 v2

Abstract

We propose a stochastic nonconvex optimization algorithm that achieves almost sure O~(ϵ1.5)\tilde{\mathcal{O}}(\epsilon^{-1.5}) iteration complexity for problems with smooth objective functions and gradients only observable with noise. The mean-zero stochastic noise is decision-dependent and has unbounded support with subexponential tail, allowing our framework to cover a broad class of problems. The improved almost sure iteration complexity is achieved with a new variant of the adaptive sampling trust-region optimization (ASTRO) augmented with an adaptively regularized local model, which we term Reg-ASTRO. Adaptive sampling ensures that the estimation precision is aligned with a measure of stationarity, so that iterates closer to stationarity trigger higher accuracy requirement for sampling. A key analytical challenge arises because the trust-region radius and regularization are coupled and not determined prior to gradient estimation at each iteration. We further establish an almost sure O~(ϵ4.5)\tilde{\mathcal{O}}(\epsilon^{-4.5}) sample complexity for Reg-ASTRO, which improves to O~(ϵ3.5)\tilde{\mathcal{O}}(\epsilon^{-3.5}) under stronger regularity conditions and use of common random numbers, substantially outperforming first-order methods in theory and numerical experiments.

Keywords

Cite

@article{arxiv.2604.15457,
  title  = {Adaptive Regularization within Trust Region Methods for Stochastic Nonconvex Optimization},
  author = {Yunsoo Ha and Sara Shashaani and Quoc Tran-dinh},
  journal= {arXiv preprint arXiv:2604.15457},
  year   = {2026}
}
R2 v1 2026-07-01T12:13:26.634Z