High-dimensional limit theorems for SGD: Effective dynamics and critical scaling

Gérard Ben Arous, Reza Gheissari*, Aukosh Jagannath

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

1 Scopus citations


We study the scaling limits of stochastic gradient descent (SGD) with constant step-size in the high-dimensional regime. We prove limit theorems for the trajectories of summary statistics (i.e., finite-dimensional functions) of SGD as the dimension goes to infinity. Our approach allows one to choose the summary statistics that are tracked, the initialization, and the step-size. It yields both ballistic (ODE) and diffusive (SDE) limits, with the limit depending dramatically on the former choices. We show a critical scaling regime for the step-size, below which the effective ballistic dynamics matches gradient flow for the population loss, but at which, a new correction term appears which changes the phase diagram. About the fixed points of this effective dynamics, the corresponding diffusive limits can be quite complex and even degenerate. We demonstrate our approach on popular examples including estimation for spiked matrix and tensor models and classification via two-layer networks for binary and XOR-type Gaussian mixture models. These examples exhibit surprising phenomena including multimodal timescales to convergence as well as convergence to sub-optimal solutions with probability bounded away from zero from random (e.g., Gaussian) initializations. At the same time, we demonstrate the benefit of overparametrization by showing that the latter probability goes to zero as the second layer width grows.

Original languageEnglish (US)
Pages (from-to)2030-2080
Number of pages51
JournalCommunications on Pure and Applied Mathematics
Issue number3
StatePublished - Mar 2024

ASJC Scopus subject areas

  • General Mathematics
  • Applied Mathematics


Dive into the research topics of 'High-dimensional limit theorems for SGD: Effective dynamics and critical scaling'. Together they form a unique fingerprint.

Cite this