Revisiting LocalSGD and SCAFFOLD: Improved Rates and Missing Analysis

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Luo, Ruichen, Stich, Sebastian U, Horváth, Samuel, Takáč, Martin
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866909507545726976
author Luo, Ruichen
Stich, Sebastian U
Horváth, Samuel
Takáč, Martin
author_facet Luo, Ruichen
Stich, Sebastian U
Horváth, Samuel
Takáč, Martin
contents LocalSGD and SCAFFOLD are widely used methods in distributed stochastic optimization, with numerous applications in machine learning, large-scale data processing, and federated learning. However, rigorously establishing their theoretical advantages over simpler methods, such as minibatch SGD (MbSGD), has proven challenging, as existing analyses often rely on strong assumptions, unrealistic premises, or overly restrictive scenarios. In this work, we revisit the convergence properties of LocalSGD and SCAFFOLD under a variety of existing or weaker conditions, including gradient similarity, Hessian similarity, weak convexity, and Lipschitz continuity of the Hessian. Our analysis shows that (i) LocalSGD achieves faster convergence compared to MbSGD for weakly convex functions without requiring stronger gradient similarity assumptions; (ii) LocalSGD benefits significantly from higher-order similarity and smoothness; and (iii) SCAFFOLD demonstrates faster convergence than MbSGD for a broader class of non-quadratic functions. These theoretical insights provide a clearer understanding of the conditions under which LocalSGD and SCAFFOLD outperform MbSGD.
format Preprint
id arxiv_https___arxiv_org_abs_2501_04443
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Revisiting LocalSGD and SCAFFOLD: Improved Rates and Missing Analysis
Luo, Ruichen
Stich, Sebastian U
Horváth, Samuel
Takáč, Martin
Optimization and Control
Distributed, Parallel, and Cluster Computing
Machine Learning
LocalSGD and SCAFFOLD are widely used methods in distributed stochastic optimization, with numerous applications in machine learning, large-scale data processing, and federated learning. However, rigorously establishing their theoretical advantages over simpler methods, such as minibatch SGD (MbSGD), has proven challenging, as existing analyses often rely on strong assumptions, unrealistic premises, or overly restrictive scenarios. In this work, we revisit the convergence properties of LocalSGD and SCAFFOLD under a variety of existing or weaker conditions, including gradient similarity, Hessian similarity, weak convexity, and Lipschitz continuity of the Hessian. Our analysis shows that (i) LocalSGD achieves faster convergence compared to MbSGD for weakly convex functions without requiring stronger gradient similarity assumptions; (ii) LocalSGD benefits significantly from higher-order similarity and smoothness; and (iii) SCAFFOLD demonstrates faster convergence than MbSGD for a broader class of non-quadratic functions. These theoretical insights provide a clearer understanding of the conditions under which LocalSGD and SCAFFOLD outperform MbSGD.
title Revisiting LocalSGD and SCAFFOLD: Improved Rates and Missing Analysis
topic Optimization and Control
Distributed, Parallel, and Cluster Computing
Machine Learning
url https://arxiv.org/abs/2501.04443