Convergence of Steepest Descent and Adam under Non-Uniform Smoothness

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Vaswani, Sharan, Sun, Yifan, Babanezhad, Reza
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917546719969280
author Vaswani, Sharan
Sun, Yifan
Babanezhad, Reza
author_facet Vaswani, Sharan
Sun, Yifan
Babanezhad, Reza
contents Recent work has analyzed the convergence of first-order methods under non-uniform smoothness assumptions that better model the loss landscape in machine learning tasks. We generalize this assumption to objectives whose curvature is an affine function of the objective value. This property is satisfied by a broad class of problems, including logistic regression, generalized linear models with a logistic link function, softmax policy gradient in reinforcement learning, and a class of neural networks. Under this assumption and gradient domination conditions, we establish a general convergence rate for the steepest descent method, and deterministic, diagonal variants of RMSProp and Adam. Our results imply that for logistic regression on separable data and the softmax policy gradient objective, sign GD converges linearly and is provably faster than GD. Furthermore, we show that for a class of two-layer neural networks on separable data, RMSProp and Adam can converge at a linear rate with a constant step-size and momentum parameter. Finally, we present a lower bound demonstrating that, under our assumption, RMSProp and Adam are provably faster than AdaGrad, AMSGrad, gradient descent, and heavy-ball momentum.
format Preprint
id arxiv_https___arxiv_org_abs_2605_30648
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Convergence of Steepest Descent and Adam under Non-Uniform Smoothness
Vaswani, Sharan
Sun, Yifan
Babanezhad, Reza
Machine Learning
Optimization and Control
Recent work has analyzed the convergence of first-order methods under non-uniform smoothness assumptions that better model the loss landscape in machine learning tasks. We generalize this assumption to objectives whose curvature is an affine function of the objective value. This property is satisfied by a broad class of problems, including logistic regression, generalized linear models with a logistic link function, softmax policy gradient in reinforcement learning, and a class of neural networks. Under this assumption and gradient domination conditions, we establish a general convergence rate for the steepest descent method, and deterministic, diagonal variants of RMSProp and Adam. Our results imply that for logistic regression on separable data and the softmax policy gradient objective, sign GD converges linearly and is provably faster than GD. Furthermore, we show that for a class of two-layer neural networks on separable data, RMSProp and Adam can converge at a linear rate with a constant step-size and momentum parameter. Finally, we present a lower bound demonstrating that, under our assumption, RMSProp and Adam are provably faster than AdaGrad, AMSGrad, gradient descent, and heavy-ball momentum.
title Convergence of Steepest Descent and Adam under Non-Uniform Smoothness
topic Machine Learning
Optimization and Control
url https://arxiv.org/abs/2605.30648