Nearly Optimal Linear Convergence of Stochastic Primal-Dual Methods for Linear Programming

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Lu, Haihao, Yang, Jinwen
Format: Preprint
Published: 2021
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914624963608576
author Lu, Haihao
Yang, Jinwen
author_facet Lu, Haihao
Yang, Jinwen
contents There is a recent interest on first-order methods for linear programming (LP). In this paper,we propose a stochastic algorithm using variance reduction and restarts for solving sharp primal-dual problems such as LP. We show that the proposed stochastic method exhibits a linear convergence rate for solving sharp instances with a high probability. In addition, we propose an efficient coordinate-based stochastic oracle for unconstrained bilinear problems, which has $\mathcal O(1)$ per iteration cost and improves the complexity of the existing deterministic and stochastic algorithms. Finally, we show that the obtained linear convergence rate is nearly optimal (upto $\log$ terms) for a wide class of stochastic primal dual methods.
format Preprint
id arxiv_https___arxiv_org_abs_2111_05530
institution arXiv
publishDate 2021
record_format arxiv
spellingShingle Nearly Optimal Linear Convergence of Stochastic Primal-Dual Methods for Linear Programming
Lu, Haihao
Yang, Jinwen
Optimization and Control
Machine Learning
There is a recent interest on first-order methods for linear programming (LP). In this paper,we propose a stochastic algorithm using variance reduction and restarts for solving sharp primal-dual problems such as LP. We show that the proposed stochastic method exhibits a linear convergence rate for solving sharp instances with a high probability. In addition, we propose an efficient coordinate-based stochastic oracle for unconstrained bilinear problems, which has $\mathcal O(1)$ per iteration cost and improves the complexity of the existing deterministic and stochastic algorithms. Finally, we show that the obtained linear convergence rate is nearly optimal (upto $\log$ terms) for a wide class of stochastic primal dual methods.
title Nearly Optimal Linear Convergence of Stochastic Primal-Dual Methods for Linear Programming
topic Optimization and Control
Machine Learning
url https://arxiv.org/abs/2111.05530