Abstract
We study min-max algorithms to solve zero-sum differential games onRiemannian manifold. Based on the notions of differential Stackelbergequilibrium and differential Nash equilibrium on Riemannian manifold, weanalyze the local convergence of two representative deterministic simultaneousalgorithms $\tau$-GDA and $\tau$-SGA to such equilibria. Sufficient conditionsare obtained to establish the linear convergence rate of $\tau$-GDA based onthe Ostrowski theorem on manifold and spectral analysis. To avoid strongrotational dynamics in $\tau$-GDA, $\tau$-SGA is extended from the symplecticgradient-adjustment method in Euclidean space. We analyze an asymptoticapproximation of $\tau$-SGA when the learning rate ratio $\tau$ is big. In somecases, it can achieve a faster convergence rate to differential Stackelbergequilibrium compared to $\tau$-GDA. We show numerically how the insightsobtained from the convergence analysis may improve the training of orthogonalWasserstein GANs using stochastic $\tau$-GDA and $\tau$-SGA on simplebenchmarks.