arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2505.04065math.OC
收藏

通过变量与算子分裂的加速梯度方法

Accelerated Gradient Methods Through Variable and Operator Splitting

Department of Mathematics, University of California at Irvine(加州大学欧文分校数学系);National Center for Applied Mathematics in Chongqing, Chongqing Normal University(重庆师范大学重庆应用数学中心);Chongqing Research Institute of Big Data, Peking University(北京大学重庆大数据研究院);Department of Applied Mathematics, The Hong Kong Polytechnic University(香港理工大学应用数学系)

Long Chen, Luo Hao, Jingrong Wei

AI总结:

本文提出基于变量与算子分裂的统一加速梯度框架,利用强Lyapunov函数、加速超松弛和预测-校正外推等技术,有效处理凸优化、复合凸优化及双线性耦合鞍点问题。

摘要:

本文通过变量与算子分裂(VOS)引入了一个统一的加速梯度方法框架。算子分裂将优化过程解耦为更简单的子问题,更重要的是,变量分裂带来了加速效果。关键贡献包括构造强 Lyapunov 函数以分析稳定性和收敛速率,以及先进的离散化技术,如加速超松弛(AOR)和基于预测-校正方法的外推(EPC)。对于凸情形,我们引入了一个动态更新参数和一个扰动 VOS 流。该框架能有效处理广泛的优化问题,包括凸优化、复合凸优化以及具有双线性耦合的鞍点系统。

英文摘要:

This survey develops a variable and operator splitting (VOS) framework for accelerated first-order methods. Operator splitting separates the components of the optimality system, while variable splitting introduces auxiliary dynamics that generate acceleration. The strong Lyapunov property guides the stability analysis, and practical schemes follow from Accelerated Over-Relaxation (AOR) and Extrapolation by Predictor--Corrector (EPC) discretizations. For merely convex problems, a perturbed VOS flow combined with homotopy in the auxiliary scale yields accelerated rates. The framework gives a common derivation and analysis of Nesterov-type, FISTA-type, composite, mirror-descent-type, and bilinear saddle-point accelerated methods. Its main contribution is structural and expository: it organizes these methods within one Lyapunov and splitting template and explains how acceleration survives different splittings and discretizations.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑