FedPCD behavior with many local steps

Characterize the behavior of FedPCD when clients perform many local optimization steps, particularly in regimes where client pseudo-gradient directions degrade because of client drift.

Background

The evaluation uses 100 clients and only one local epoch, while the theoretical analysis imposes deterministic full participation and accounts for local drift through restrictive geometric conditions. The paper therefore does not establish how FedPCD behaves when clients take many local steps, a setting in which the pseudo-gradients may become less representative of the round-start client gradients.

References

Finally, we note that the evaluation is restricted to $N = 100$ clients and one local epoch, so the behavior of FedPCD under many local steps, where pseudo-gradient directions degrade through client drift, is not established.

— FedPCD: Priority-Constrained Federated Aggregation  (2610.13508 - Irshaid et al., 9 Oct 2026) in Conclusion