Stability and Generalization Framework for Biased Stochastic Gradient Minimax Optimization

This research introduces the first framework to analyze the stability and generalization of biased stochastic gradient methods in minimax problems. The authors relax the bounded gradient assumption to a bounded second moment condition and introduce a generalized Lipschitz-type condition for bias estimators.