Proposes the novel framework of Multiplayer Federated Learning and analyzes the communication-efficient Per-Player Local SGD (PEARL-SGD).

Local SGD converges faster under low second-order heterogeneity, and we prove it with tight bounds and supporting experiments.