Searcharxiv⌕ Search

arXiv · 2611.16560

When and How SignSGD Outperforms SGD: A Functional Scaling Law Perspective

Abstract

We study how the sign operator interacts with gradient noise to shape the learning dynamics and scaling behavior of sign-based optimizers. In power-law kernel regression with diagonal features, we establish a functional scaling law (FSL) for SignSGD under general learning rate schedules. The FSL describes the learning dynamics through two processes: signal learning and noise forgetting. The key mechanism is noise-induced preconditioning: gradient noise smooths the sign operator, inducing an effective preconditioned SGD. Noise-curvature alignment enables this preconditioning to transform the effective Hessian from $\mathbf{H}$ to $\mathbf{H}^{1/2}$. The resulting dynamics exhibit faster signal learning but slower noise forgetting than SGD. This tradeoff yields a regime-dependent comparison with SGD: (i) SignSGD achieves improved data efficiency in the hard-task regime; (ii) both optimizers achieve the same data efficiency in the easy-task regime, but SignSGD maintains this efficiency at larger batch sizes, requiring fewer optimization steps for a fixed data budget. Power-law kernel regression experiments support the proposed mechanism and predicted scaling behavior. LLM pre-training experiments further suggest that the SignSGD FSL better captures the batch-size dependence of loss dynamics for sign-based optimizers than the original SGD FSL.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Zilin Wang, Binghui Li, Jia-Nan Wang, Lean Wang, Jinbo Wang, Lei Wu. 2026-11-12. When and How SignSGD Outperforms SGD: A Functional Scaling Law Perspective. https://searcharxiv.com/abs/2611.16560

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Scalable Learning from Probability Measures with Mean Measure Quantization

We consider statistical learning problems in which data are observed as a set of probability measures. Optimal transport (OT) is a popular tool to compare and manipulate such objects, but its computational cost becomes prohibitive when the measures have large support. We study a quantization-based approach in which all input measures are approximated by $K$-point discrete measures sharing a common support. We establish consistency of the resulting quantized measures. We further derive convergence guarantees for several OT-based downstream tasks computed from the quantized measures. Numerical experiments on synthetic and real datasets demonstrate that the proposed approach achieves performance comparable to individual quantization while substantially reducing runtime.

stat.ML↗

Measuring and Decomposing Mode Separation via the Canonical Diffusion

Mode separation, namely how sharply a distribution fragments into barrier-separated clusters, is a fundamental geometric property of densities, difficult to quantify in high dimensions. Modes are often the meaningful alternatives in the data, such as the conformations of a molecule or the distinct semantic types among a generative model's outputs. Mode separation is distinct from dispersion, which is what standard label-free tools capture: differential entropy rises with spread regardless of fragmentation, and PCA orders directions by variance regardless of barriers. Mutual information does capture it, but only given the latent component labels, which are usually unavailable. We measure mode separation through a single stochastic process intrinsic to the density, namely the unique reversible diffusion with $f$ as its stationary distribution and a constant scalar diffusion coefficient. From its autocovariance matrix we extract two readouts. The first, SSA (Sum of Squared Autocorrelations), is a scalar barrier-sensitive measure. The second, DA (Dominant Autocorrelation directions), gives linear projections ordered by metastability rather than variance. Under an isotropic-Gaussian null, we derive the limiting spectrum of the empirical autocovariance in closed form, generalizing the Marchenko--Pastur law. Its upper edge gives a noise floor that separates metastable directions from sampling noise. Both readouts use only samples and a score function, scaling to high dimensions through pretrained score-based generative models via Tweedie's identity. We apply our framework to three settings: (i) synthetic Gaussian mixtures, where SSA tracks mutual information; (ii) SDXL text-to-image generations, where SSA and DA capture structure that entropy and PCA miss; and (iii) molecular dynamics of alanine dipeptide, where DA recovers the known slow backbone dihedrals from static samples alone.

stat.ML↗

Sharp feature-learning transitions and Bayes-optimal neural scaling laws in extensive-width networks

We study the information-theoretic limits of learning a one-hidden-layer teacher network with hierarchical features from noisy queries, in the context of knowledge transfer to a smaller student model. We work in the high-dimensional regime where the teacher width $k$ scales linearly with the input dimension $d$ -- a setting that captures large-but-finite-width networks and has only recently become analytically tractable. Using a heuristic leave-one-out decoupling argument, validated numerically throughout, we derive asymptotically sharp characterizations of the Bayes-optimal generalization error and individual feature overlaps via a system of closed fixed-point equations. These equations reveal that feature learnability is governed by a sequence of sharp phase transitions: as data grows, teacher features become recoverable sequentially, each through a discontinuous jump in overlap. This sequential acquisition underlies a precise notion of \textit{effective width} $k_c$ -- the number of learnable features at a given data budget $n$ -- which unifies two distinct scaling regimes: a feature-learning regime in which the Bayes-optimal generalization error $\varepsilon^{\rm BO}$ scales as $ n^{1/(2β)-1}$, and a refinement regime in which it scales as $n^{-1}$, where $β>1/2$ is the exponent of the power-law feature hierarchy. Both laws collapse to the single relation $\varepsilon^{\rm BO}=Θ(k_c d/n)$. We further show empirically that a student trained with \textsc{Adam} near the effective width $k_c$ achieves these optimal scaling laws (up to a small algorithmic gap), and provide an information-theoretic account of the associated scaling in model size.

stat.ML↗