arXiv preprint, 2023.
In brief
Combines the stochastic accelerated primal–dual method with variance reduction for strongly-convex–strongly-concave saddle-point problems of finite-sum form, so that the noise of stochastic gradients no longer caps the accuracy: the method converges linearly to the saddle point, with a rate that reflects both the acceleration and the variance reduction.
Topics
Cite
@misc{can2023vrsapd,
title = {A variance-reduced stochastic accelerated primal–dual algorithm},
author = {Bugra Can and Mert Gürbüzbalaban and Necdet Serhat Aybat},
year = {2023},
howpublished = {arXiv preprint arXiv:2202.09688},
}Relatedsame topics