English

Open Problem: Anytime Convergence Rate of Gradient Descent

Optimization and Control 2024-06-21 v1 Machine Learning

Abstract

Recent results show that vanilla gradient descent can be accelerated for smooth convex objectives, merely by changing the stepsize sequence. We show that this can lead to surprisingly large errors indefinitely, and therefore ask: Is there any stepsize schedule for gradient descent that accelerates the classic O(1/T)\mathcal{O}(1/T) convergence rate, at \emph{any} stopping time TT?

Keywords

Cite

@article{arxiv.2406.13888,
  title  = {Open Problem: Anytime Convergence Rate of Gradient Descent},
  author = {Guy Kornowski and Ohad Shamir},
  journal= {arXiv preprint arXiv:2406.13888},
  year   = {2024}
}

Comments

COLT 2024 open problem; 5 pages

R2 v1 2026-06-28T17:12:46.318Z