Work

Code for "Forgetting matters: when subspace optimization accelerates without the strong convexity parameter"

Year: 2026

Type: software

Abstract: Code and computed results for the numerical experiments of the paper “Forgetting matters: when subspace optimization accelerates without the strong convexity parameter” by I. Kuruzov, A. Vyguzov, I. Dorofeev and A. Gasnikov (arXiv:2611.00727). The package fmcode implements exact subspace methods (SESOP, the plane form of conjugate gradients, windows, the activation schedule of filter pairs and the smooth adaptive pair), Nesterov’s method with and without restarts and L-BFGS on regularized logistic regression and on one- and two-dimensional capillary energies, together with the hard instances for SESOP. The folder experiments contains one script per table and figure of Section 6, results contains the outputs reported in the paper, and checks contains numerical checks of several lemmas and theorems. Requirements: Python 3.10 or newer, NumPy, SciPy and Matplotlib. See README.md for installation and usage.

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.23916707

Authors: Kuruzov, Ilya, Vyguzov, Alexander, Dorofeev, Ivan, Gasnikov, Alexander

Institutions: Innopolis University, AIRI - Artificial Intelligence Research Institute

Language: English

Topic: Stochastic Gradient Optimization Techniques

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

OpenAlex ID: W7225196068

Cites:

Cited by:

Related to: 10