On the Superlinear Relationship between SGD Noise Covariance and Loss Landscape Curvature

• URL: http://arxiv.org/abs/2602.05600v1
• Date: Thu, 05 Feb 2026 12:35:13 GMT
• Title: On the Superlinear Relationship between SGD Noise Covariance and Loss Landscape Curvature
• Authors: Yikuan Zhang, Ning Yang, Yuhai Tu,
• Abstract summary: Gradient Descent (SGD) introduces anisotropic noise that is correlated with the local curvature of the loss landscape, thereby biasing optimization toward flat minima.<n>We show that this assumption holds only under restrictive conditions that are typically violated in deep neural networks.<n>Experiments across datasets, architectures, and loss functions validate these bounds, providing a unified characterization of the noise-curvature relationship in deep learning.
• Score: 1.6773271875801752
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Stochastic Gradient Descent (SGD) introduces anisotropic noise that is correlated with the local curvature of the loss landscape, thereby biasing optimization toward flat minima. Prior work often assumes an equivalence between the Fisher Information Matrix and the Hessian for negative log-likelihood losses, leading to the claim that the SGD noise covariance $\mathbf{C}$ is proportional to the Hessian $\mathbf{H}$. We show that this assumption holds only under restrictive conditions that are typically violated in deep neural networks. Using the recently discovered Activity--Weight Duality, we find a more general relationship agnostic to the specific loss formulation, showing that $\mathbf{C} \propto \mathbb{E}_p[\mathbf{h}_p^2]$, where $\mathbf{h}_p$ denotes the per-sample Hessian with $\mathbf{H} = \mathbb{E}_p[\mathbf{h}_p]$. As a consequence, $\mathbf{C}$ and $\mathbf{H}$ commute approximately rather than coincide exactly, and their diagonal elements follow an approximate power-law relation $C_{ii} \propto H_{ii}^γ$ with a theoretically bounded exponent $1 \leq γ\leq 2$, determined by per-sample Hessian spectra. Experiments across datasets, architectures, and loss functions validate these bounds, providing a unified characterization of the noise-curvature relationship in deep learning.

Related papers

• Privacy Loss of Noise Perturbation via Concentration Analysis of A Product Measure [9.081364242189691]
  In this paper, we propose a novel scheme to generate $mathbfn$ by leveraging the fact that the noise is typically spherically symmetric.<n>Under the same $(,)$-DP guarantee, our mechanism yields a smaller expected noise magnitude than the classic Gaussian noise in high dimensions.
  arXiv  Detail & Related papers  (2025-12-06T02:55:46Z)
• Efficient reductions from a Gaussian source with applications to statistical-computational tradeoffs [8.162867143465382]
  We leverage our technique to establish reduction-based computational lower bounds for several canonical high-dimensional statistical models under widely-believed conjectures in average-case complexity.<n>We show that computational lower bounds proved for spiked tensor PCA with Gaussian noise are universal, in that they extend to other non-Gaussian noise distributions within our class.
  arXiv  Detail & Related papers  (2025-10-08T17:16:36Z)
• Local minima of the empirical risk in high dimension: General theorems and convex examples [8.748904058015574]
  We consider a general model for high-dimensional empirical risk whereby the data vectors $mathbfxi$ are $d- minimization.<n>We derive sharps on the estimation and prediction error.
  arXiv  Detail & Related papers  (2025-02-04T03:02:24Z)
• Semidefinite programming relaxations and debiasing for MAXCUT-based clustering [1.9761774213809036]
  We consider the problem of partitioning a small data sample of size $n$ drawn from a mixture of 2 sub-gaussian distributions in $mathbbRp$.<n>We use semidefinite programming relaxations of an integer quadratic program that is formulated as finding the maximum cut on a graph.
  arXiv  Detail & Related papers  (2024-01-16T03:14:24Z)
• High-probability Convergence Bounds for Nonlinear Stochastic Gradient Descent Under Heavy-tailed Noise [59.25598762373543]
  We show that wetailed high-prob convergence guarantees of learning on streaming data in the presence of heavy-tailed noise. We demonstrate analytically and that $ta$ can be used to the preferred choice of setting for a given problem.
  arXiv  Detail & Related papers  (2023-10-28T18:53:41Z)
• A Unified Framework for Uniform Signal Recovery in Nonlinear Generative Compressed Sensing [68.80803866919123]
  Under nonlinear measurements, most prior results are non-uniform, i.e., they hold with high probability for a fixed $mathbfx*$ rather than for all $mathbfx*$ simultaneously. Our framework accommodates GCS with 1-bit/uniformly quantized observations and single index models as canonical examples. We also develop a concentration inequality that produces tighter bounds for product processes whose index sets have low metric entropy.
  arXiv  Detail & Related papers  (2023-09-25T17:54:19Z)
• Compressed and distributed least-squares regression: convergence rates with applications to Federated Learning [11.870656106069447]
  We investigate the impact of compression on gradient algorithms for machine learning.<n>We highlight differences in terms of convergence rates between several unbiased compression operators.<n>We extend our results to the case of federated learning.
  arXiv  Detail & Related papers  (2023-08-02T18:02:00Z)
• Optimal Extragradient-Based Bilinearly-Coupled Saddle-Point Optimization [116.89941263390769]
  We consider the smooth convex-concave bilinearly-coupled saddle-point problem, $min_mathbfxmax_mathbfyF(mathbfx) + H(mathbfx,mathbfy)$, where one has access to first-order oracles for $F$, $G$ as well as the bilinear coupling function $H$. We present a emphaccelerated gradient-extragradient (AG-EG) descent-ascent algorithm that combines extragrad
  arXiv  Detail & Related papers  (2022-06-17T06:10:20Z)
• Lattice partition recovery with dyadic CART [79.96359947166592]
  We study piece-wise constant signals corrupted by additive Gaussian noise over a $d$-dimensional lattice. Data of this form naturally arise in a host of applications, and the tasks of signal detection or testing, de-noising and estimation have been studied extensively in the statistical and signal processing literature. In this paper we consider instead the problem of partition recovery, i.e.of estimating the partition of the lattice induced by the constancy regions of the unknown signal. We prove that a DCART-based procedure consistently estimates the underlying partition at a rate of order $sigma2 k*
  arXiv  Detail & Related papers  (2021-05-27T23:41:01Z)
• Optimal Robust Linear Regression in Nearly Linear Time [97.11565882347772]
  We study the problem of high-dimensional robust linear regression where a learner is given access to $n$ samples from the generative model $Y = langle X,w* rangle + epsilon$ We propose estimators for this problem under two settings: (i) $X$ is L4-L2 hypercontractive, $mathbbE [XXtop]$ has bounded condition number and $epsilon$ has bounded variance and (ii) $X$ is sub-Gaussian with identity second moment and $epsilon$ is
  arXiv  Detail & Related papers  (2020-07-16T06:44:44Z)
• Agnostic Learning of a Single Neuron with Gradient Descent [92.7662890047311]
  We consider the problem of learning the best-fitting single neuron as measured by the expected square loss. For the ReLU activation, our population risk guarantee is $O(mathsfOPT1/2)+epsilon$. For the ReLU activation, our population risk guarantee is $O(mathsfOPT1/2)+epsilon$.
  arXiv  Detail & Related papers  (2020-05-29T07:20:35Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.