Communication-Efficient Algorithms for Distributed Nonconvex Minimax Optimization Problems
Abstract
We study stochastic nonconvex Polyak-{\L}ojasiewicz minimax problems and propose algorithms that are both communication- and sample-efficient. The proposed methods are developed under three setups: decentralized/distributed, federated/centralized, and single-agent. By exploiting second-order Lipschitz continuity and integrating communication-efficient strategies, we develop a new decentralized normalized accelerated momentum method with local updates and establish its convergence to an $\varepsilon$-game stationary point. Compared to existing decentralized minimax algorithms, our proposed algorithm is the first to achieve a state-of-the-art communication complexity of order $\mathcal{O}\Big( \frac{ \kappa^3\varepsilon^{-3}}{NK(1-\lambda)^{3/2}}\Big)$, demonstrating linear speedup with respect to both the number of agents $K$ and the number of local updates $N$, as well as the best known dependence on the level of accuracy of the solution $\varepsilon$. In addition to improved complexity, our algorithm offers several practical advantages: it relaxes the strict two-time-scale step size ratio required by many existing algorithms, simplifies the stability conditions for step size selection, and eliminates the need for large batch sizes to attain the optimal sample complexity. Moreover, we propose more efficient variants tailored to federated/centralized and single-agent setups, and show that all variants achieve best-known results while effectively addressing some key issues. Experiments on robust logistic regression and fair neural network classifier using real-world datasets demonstrate the superior performance of the proposed methods over existing baselines.