Abstract
We study zero-sum games in the space of probability distributions over theEuclidean space $\mathbb{R}^d$ with entropy regularization, in the setting whenthe interaction function between the players is smooth and stronglyconvex-strongly concave. We prove an exponential convergence guarantee for themean-field min-max Langevin dynamics to compute the equilibrium distribution ofthe zero-sum game. We also study the finite-particle approximation of themean-field min-max Langevin dynamics, both in continuous and discrete times. Weprove biased convergence guarantees for the continuous-time finite-particlemin-max Langevin dynamics to the stationary mean-field equilibrium distributionwith an explicit bias term which does not scale with the number of particles.We also prove biased convergence guarantees for the discrete-timefinite-particle min-max Langevin algorithm to the stationary mean-fieldequilibrium distribution with an additional bias term which scales with thestep size and the number of particles. This provides an explicit iterationcomplexity for the average particle along the finite-particle algorithm toapproximately compute the equilibrium distribution of the zero-sum game.