Shuffling the Data, Stretching the Step-size: Sharper Bias in constant step-size SGD

TL;DR

This paper demonstrates that combining Random Reshuffling and Richardson-Romberg extrapolation in stochastic gradient methods yields superior bias reduction and convergence speed in finite-sum min-max optimization and variational inequality problems.

Contribution

It provides the first theoretical analysis of the synergy between reshuffling and extrapolation techniques, showing enhanced bias reduction and convergence guarantees.

Findings

Composition of reshuffling and extrapolation improves bias reduction beyond individual methods.

Theoretical guarantees established for structured non-monotone VIPs.

Experiments show substantial practical speedups with the combined approach.

Abstract

From adversarial robustness to multi-agent learning, many machine learning tasks can be cast as finite-sum min-max optimization or, more generally, as variational inequality problems (VIPs). Owing to their simplicity and scalability, stochastic gradient methods with constant step size are widely used, despite the fact that they converge only up to a constant term. Among the many heuristics adopted in practice, two classical techniques have recently attracted attention to mitigate this issue: \emph{Random Reshuffling} of data and \emph{Richardson--Romberg extrapolation} across iterates. Random Reshuffling sharpens the mean-squared error (MSE) of the estimated solution, while Richardson-Romberg extrapolation acts orthogonally, providing a second-order reduction in its bias. In this work, we show that their composition is strictly better than both, not only maintaining the enhanced MSE guarantees but also yielding an even greater cubic refinement in the bias. To the best of our knowledge, our work provides the first theoretical guarantees for such a synergy in structured non-monotone VIPs. Our analysis proceeds in two steps: (i) we smooth the discrete noise induced by reshuffling and leverage tools from continuous-state Markov chain theory to establish a novel law of large numbers and a central limit theorem for its iterates; and (ii) we employ spectral tensor techniques to prove that extrapolation debiases and sharpens the asymptotic behavior even under the biased gradient oracle induced by reshuffling. Finally, extensive experiments validate our theory, consistently demonstrating substantial speedups in practice.