Formalize fairness in federated learning under unknown client availability

Develop a formal treatment of fairness in federated learning, including a comparison between minimum per-client accuracy and the minimax fairness notion of Mohri et al.

Background

The paper evaluates fairness empirically through minimum per-client accuracy and reports that FedeRage improves this metric under restricted and imbalanced client availability. However, the analysis does not provide a formal fairness framework or establish how this operational metric relates to alternative notions of fairness in federated learning.

In particular, the authors identify comparison with the minimax formulation of Mohri et al. as unresolved. Addressing this problem would clarify the theoretical fairness guarantees of FedeRage and distinguish its risk-averse treatment of client heterogeneity from minimax optimization over client distributions.

References

The experiments show that fairness, modeled here through the minimum per-client accuracy, is improved; a formal treatment, together with a comparison to the minimax notion of, remains open.

— FedeRage: Provably Convergent Agnostic Federated Learning under General Client Drift  (2609.21057 - Rahimi et al., 17 Sep 2026) in Section Conclusion and Future Work, subsection “Future Directions,” bullet “Fairness in FL”