Federated learning typically relies on a fixed, hand-tuned number of local training epochs before clients communicate with the server. This practice is simple, but the paper argues it ignores a fundamental trade-off: each additional epoch reduces the client's loss, yet it also increases client drift—the divergence between the local model and the global objective.
The authors recast the decision of when to stop local training as an optimal stopping problem. In this view, an extra epoch is worthwhile only while the loss it removes outweighs the drift it adds. The optimal stopping point is where the marginal benefit of further training equals its marginal cost.
The paper provides both theoretical analysis and empirical evidence for this approach, suggesting that dynamic stopping rules could outperform the common fixed-epoch baseline. The abstract does not detail the specific methods or results, but the framing offers a principled alternative to hand-tuned local epochs in federated learning.