All ERMs Can Fail in Stochastic Convex Optimization Lower Bounds in Linear Dimension

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Burla, Tal, Livni, Roi
Formato: Preprint
Publicado: 2026
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866910016376668160
author Burla, Tal
Livni, Roi
author_facet Burla, Tal
Livni, Roi
contents We study the sample complexity of the best-case Empirical Risk Minimizer in the setting of stochastic convex optimization. We show that there exists an instance in which the sample size is linear in the dimension, learning is possible, but the Empirical Risk Minimizer is likely to be unique and to overfit. This resolves an open question by Feldman. We also extend this to approximate ERMs. Building on our construction we also show that (constrained) Gradient Descent potentially overfits when horizon and learning rate grow w.r.t sample size. Specifically we provide a novel generalization lower bound of $Ω\left(\sqrt{ηT/m^{1.5}}\right)$ for Gradient Descent, where $η$ is the learning rate, $T$ is the horizon and $m$ is the sample size. This narrows down, exponentially, the gap between the best known upper bound of $O(ηT/m)$ and existing lower bounds from previous constructions.
format Preprint
id arxiv_https___arxiv_org_abs_2602_08350
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle All ERMs Can Fail in Stochastic Convex Optimization Lower Bounds in Linear Dimension
Burla, Tal
Livni, Roi
Machine Learning
We study the sample complexity of the best-case Empirical Risk Minimizer in the setting of stochastic convex optimization. We show that there exists an instance in which the sample size is linear in the dimension, learning is possible, but the Empirical Risk Minimizer is likely to be unique and to overfit. This resolves an open question by Feldman. We also extend this to approximate ERMs. Building on our construction we also show that (constrained) Gradient Descent potentially overfits when horizon and learning rate grow w.r.t sample size. Specifically we provide a novel generalization lower bound of $Ω\left(\sqrt{ηT/m^{1.5}}\right)$ for Gradient Descent, where $η$ is the learning rate, $T$ is the horizon and $m$ is the sample size. This narrows down, exponentially, the gap between the best known upper bound of $O(ηT/m)$ and existing lower bounds from previous constructions.
title All ERMs Can Fail in Stochastic Convex Optimization Lower Bounds in Linear Dimension
topic Machine Learning
url https://arxiv.org/abs/2602.08350