14
11

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

Abstract

We study the problem of finding approximate first-order stationary points in optimization problems of the form minxXmaxyYf(x,y)\min_{x \in X} \max_{y \in Y} f(x,y), where the sets X,YX,Y are convex and YY is compact. The objective function ff is smooth, but assumed neither convex in xx nor concave in yy. Our approach relies upon replacing the function f(x,)f(x,\cdot) with its kkth order Taylor approximation (in yy) 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 YY be small in terms of the target accuracy ε\varepsilon, namely O(ε2k+1)O(\varepsilon^{\frac{2}{k+1}}) for kNk \in \mathbb{N} and O(ε)O(\varepsilon) for k=0k = 0, with the constant factors controlled by certain regularity parameters of ff; then any ε\varepsilon-stationary point in the surrogate problem remains O(ε)O(\varepsilon)-stationary for the initial problem. Moreover, we show that these upper bounds are nearly optimal: the aforementioned reduction provably fails when the diameter of YY is larger. For 0k20 \le k \le 2 the surrogate function can be efficiently maximized in yy; 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
Comments on this paper