Continuous-Time Analysis of Federated Averaging

Tom Overman, Diego KlabjanarXiv 2025 · DOI · arXiv

Federated Learning (FL) has emerged as a cornerstone for privacy-preserving machine learning, with Federated Averaging (FedAvg) serving as its primary algorithmic driver. Despite its ubiquity, the discrete-time analysis of FedAvg is notoriously difficult. The divergence of local models—often termed "client drift"—introduces mathematical complexities that make traditional proofs cumbersome. This paper addresses these challenges by shifting the analytical lens from discrete iterations to continuous-time Stochastic Differential Equations (SDEs).

Bridging Discrete Updates and Continuous Flow

To model the evolution of global model weights, the authors derive a multivariate SDE that describes how server weights change over time. This framework captures the "drift" (the average gradient across clients) and "diffusion" (the inherent stochasticity and noise of the process).

A pivotal contribution of this work is the mathematical justification of the "normality assumption." Using the Lyapunov Central Limit Theorem, the authors demonstrate that as the number of clients increases, the aggregated updates converge to a normal distribution. Crucially, this holds true even when client data is highly non-identical (non-IID), providing a rigorous foundation for applying stochastic calculus to large-scale Federated Learning systems.

Convergence in Non-Convex Landscapes

Utilizing Dynkin’s formula—a fundamental tool in stochastic calculus—the authors provide compact convergence proofs for non-convex and "weakly quasi-convex" functions. The latter is particularly significant, as weakly quasi-convex functions are more general than standard convex functions and can include locally concave regions, making them a superior proxy for the complex landscapes of actual neural networks.

A key insight from this analysis is the necessity of client-side learning rate decay. The authors prove that simply decreasing the server-side learning rate is insufficient for convergence to a stationary point in non-convex environments; the client-side learning rate must also be decayed at specific rates to ensure stability.

The Generalization Trade-off

To explore the impact of local training on generalization, the paper analyzes a quadratic loss landscape, which allows for a closed-form analytical solution to the SDE. The results reveal a fundamental trade-off associated with increasing the number of local iterations ($E$):

  • Increased Bias: The mean of the solution drifts further away from the true global minimum as local iterations increase.
  • Increased Variance: The variance of the solution increases at a rate of $O(E^2)$.
  • The Generalization Benefit: While higher variance is typically viewed as a hurdle for optimization, the continuous-time analysis suggests it serves a functional purpose. The increased stochasticity allows the model to "jump out" of sharp local minima that generalize poorly, potentially leading to better performance on unseen data by finding flatter regions of the loss landscape.

Limitations and Future Directions

While the SDE framework offers powerful insights, it relies on the assumption of small step sizes ($h$). In practical FL scenarios where communication is infrequent and steps are large, the accuracy of this continuous approximation may be challenged. Furthermore, the generalization analysis is currently limited to quadratic landscapes, which do not fully capture the complexity of deep learning.

Future research may build on this framework to study "first exit times"—calculating exactly how long it takes for FedAvg to escape a local minimum. Such work would provide even deeper insights into the algorithm's unique generalization capabilities compared to centralized SGD.

Summary updated

Abstract

Federated averaging (FedAvg) is a popular algorithm for horizontal federated learning (FL), where samples are gathered across different clients and are not shared with each other or a central server. Extensive convergence analysis of FedAvg exists for the discrete iteration setting, guaranteeing convergence for a range of loss functions and varying levels of data heterogeneity. We extend this analysis to the continuous-time setting where the global weights evolve according to a multivariate stochastic differential equation (SDE), which is the first time FedAvg has been studied from the continuous-time perspective. We use techniques from stochastic processes to establish convergence guarantees under different loss functions, some of which are more general than existing work in the discrete setting. We also provide conditions for which FedAvg updates to the server weights can be approximated as normal random variables. Finally, we use the continuous-time formulation to reveal generalization properties of FedAvg.
BibTeX
@misc{overman2025continuous,
  author = {Tom Overman and Diego Klabjan},
  title = {{Continuous-Time Analysis of Federated Averaging}},
  year = {2025},
  eprint = {2501.18870},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.18870},
}
Figure