English

Optimal Extragradient-Based Bilinearly-Coupled Saddle-Point Optimization

Optimization and Control 2022-08-15 v3 Computational Complexity Computer Science and Game Theory Machine Learning

Abstract

We consider the smooth convex-concave bilinearly-coupled saddle-point problem, minxmaxy F(x)+H(x,y)G(y)\min_{\mathbf{x}}\max_{\mathbf{y}}~F(\mathbf{x}) + H(\mathbf{x},\mathbf{y}) - G(\mathbf{y}), where one has access to stochastic first-order oracles for FF, GG as well as the bilinear coupling function HH. Building upon standard stochastic extragradient analysis for variational inequalities, we present a stochastic \emph{accelerated gradient-extragradient (AG-EG)} descent-ascent algorithm that combines extragradient and Nesterov's acceleration in general stochastic settings. This algorithm leverages scheduled restarting to admit a fine-grained nonasymptotic convergence rate that matches known lower bounds by both \citet{ibrahim2020linear} and \citet{zhang2021lower} in their corresponding settings, plus an additional statistical error term for bounded stochastic noise that is optimal up to a constant prefactor. This is the first result that achieves such a relatively mature characterization of optimality in saddle-point optimization.

Keywords

Cite

@article{arxiv.2206.08573,
  title  = {Optimal Extragradient-Based Bilinearly-Coupled Saddle-Point Optimization},
  author = {Simon S. Du and Gauthier Gidel and Michael I. Jordan and Chris Junchi Li},
  journal= {arXiv preprint arXiv:2206.08573},
  year   = {2022}
}

Comments

More polishing and clarifications; 36 pages

R2 v1 2026-06-24T11:54:41.422Z