Nonconvex-Nonconcave Min-Max Optimization with a Small Maximization Domain

We study the problem of finding approximate first-order stationary points in optimization problems of the form , where the sets are convex and is compact. The objective function is smooth, but assumed neither convex in nor concave in . Our approach relies upon replacing the function with its th order Taylor approximation (in ) and finding a near-stationary point in the resulting surrogate problem. To guarantee its success, we establish the following result: let the Euclidean diameter of be small in terms of the target accuracy , namely for and for , with the constant factors controlled by certain regularity parameters of ; then any -stationary point in the surrogate problem remains -stationary for the initial problem. Moreover, we show that these upper bounds are nearly optimal: the aforementioned reduction provably fails when the diameter of is larger. For the surrogate function can be efficiently maximized in ; our general approximation result then leads to efficient algorithms for finding a near-stationary point in nonconvex-nonconcave min-max problems, for which we also provide convergence guarantees.
View on arXiv