Generalized Mirror Prox for Monotone Variational Inequalities: Universality and Inexact Oracle

Abstract

We introduce an inexact oracle model for variational inequalities (VI) with monotone operator, propose a numerical method which solves such VI's and analyze its convergence rate. As a particular case, we consider VI's with Hölder-continuous operator and show that our algorithm is universal. This means that without knowing the Hölder parameter $\nu$ and Hölder constant $L_{\nu }$ it has the best possible complexity for this class of VI's, namely our algorithm has complexity $O\left({\inf }_{\nu \in [0,1]}{\left(\frac{L_{\nu }}{\epsilon }\right)}^{\frac{2}{1+\nu }}{R}^2\right)$, where $R$ is the size of the feasible set and $\epsilon$ is the desired accuracy of the solution. We also consider the case of VI's with strongly monotone operator and generalize our method for VI's with inexact oracle and our universal method for this class of problems. Finally, we show, how our method can be applied to convex-concave saddle point problems with Hölder-continuous partial subgradients.

References in corpus (2)

• On Acceleration with Noise-Corrupted Gradients
• Universal similar triangulars method for searching equilibriums in traffic flow distribution models

Cited by in corpus (6)

• Adaptive Gradient Descent for Convex and Non-Convex Stochastic Optimization
• Mirror Descent for Constrained Optimization Problems with Large Subgradient Values
• Adaptive extra-gradient methods for min-max optimization and games
• Halpern Iteration for Near-Optimal and Parameter-Free Monotone Inclusion and Strong Solutions to Variational Inequalities
• Stochastic Saddle-Point Optimization for Wasserstein Barycenters
• Geometry-Aware Universal Mirror-Prox