Continuous-time Analysis of Adam in Min-Max Games: Local Convergence and Implicit Gradient Regularization
Lists
Only you see this
Abstract
Although Adam has been widely used in solving min-max games, its dynamical behavior in this setting has not been studied as extensively as in standard minimization. In min-max games, each player aims to minimize their own loss function, which may lead to the intuitive assumption that Adam’s properties in minimization naturally extend to min-max games with minor adjustments. In this paper, we demonstrate that this belief is incorrect. Specifically, we investigate the local convergence and implicit gradient regularization aspects of Adam in min-max games, focusing on how its first- and second-order momentum parameters influence the dynamics. We show that, in both aspects, the momentum parameters affect Adam’s behavior in min-max games in precisely the opposite manner to their effect in minimization. Our methodology is based on the continuous-time analysis techniques commonly used for studying learning algorithms. We further provide numerical experiments to support our theoretical findings.
BibTeX
@misc{feng2026continuous,
author = {Yi Feng and Weiming Ou and Xiao Wang},
title = {{Continuous-time Analysis of Adam in Min-Max Games: Local Convergence and Implicit Gradient Regularization}},
year = {2026},
url = {https://openreview.net/forum?id=XtYcnbnJ1n},
howpublished = {OpenReview},
note = {Submitted to ICLR 2026},
}Loading related content...