Answers are grounded in this paper and cite their sources.
Server-side adaptive optimizers in Federated Learning (FL), such as FedAdam and FedAvgM, scale global updates using historical temporal statistics but do not explicitly model instantaneous cross-client variance within a single communication round. Under severe statistical heterogeneity (non-IID data), this can leave the adaptive step size insufficiently damped precisely when local client updates conflict most. We propose FedHST (Federated Hybrid Spatial–Temporal Optimization), a server-side aggregation framework that augments the temporal second-moment estimate with an instantaneous, coordinate-wise cross-client spatial variance term. This term acts as an additional damper under high client disagreement and vanishes under consensus, at which point FedHST reduces to standard temporal-only adaptation. We establish a non-convex convergence rate of 𝒪 ( 1 / T ) , with the residual term showing explicitly how the spatial damping coefficient γ moderates the effect of statistical heterogeneity on convergence. We evaluate FedHST on FashionMNIST, CIFAR-10, and CIFAR-100 under Dirichlet-partitioned non-IID splits ( α ∈ { 0.1 , 0.3 } ), comparing against FedAdam and FedDyn — the strongest representatives of the adaptive server-optimization and dynamic client-regularization paradigms, respectively — over three independent random seeds with paired significance testing; FedAvgM, FedYogi, FedAdagrad, FedProx, and SCAFFOLD are reviewed in Related Work. FedHST attains accuracy statistically indistinguishable from, or modestly and significantly exceeding (up to + 0.5 percentage points, p < 0.05 on FashionMNIST), FedAdam, while substantially and significantly outperforming FedDyn (+ 5 to + 28 percentage points across datasets, p < 0.01), at a computational overhead within 1–4% of FedAdam. Ablation studies further show that FedHST’s advantage over FedAdam grows systematically with the number of local training epochs and with non-IID severity, consistent with the mechanism by which spatial damping is designed to activate under cross-client conflict.
No takes yet. Share an insight, caveat, or question.
Le et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: