On the Stability of Consensus Gradient Dynamics for Regularized Distributed Deep Matrix Factorization: Explicit Bounds
IEEE Signal Processing Letters · 2026 · 33 · 2884-2888
DOI: 10.1109/lsp.2026.3711534 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von Richard Oliveira und 1 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The stability of consensus-based gradient methods remains a challenging yet crucial interrogative in dynamical systems and learning theory, particularly for non-convex optimization problems.Deep matrix factorization (DMF) concerns the decomposition of a matrix as the product of multiple latent factors, and serves as a canonical problem for such analyses in centralized settings, with main applications in computer vision. However, the distributed DMF problem has remained unaddressed. In this letter, we extend DMF to a consensus-based framework with weight decay regularization. We derive analytically tractable learning rate bounds that ensure the stability of consensus gradient dynamics around stationary points. These depend on the largest eigenvalue of the Hessian of the regularized loss and on the graph topology, ensuring stability of the iterates. Moreover, we corroborate our stability claims through numerical examples.
Zitieren
APA
Oliveira, R., & Savazzi, S. (2026). On the Stability of Consensus Gradient Dynamics for Regularized Distributed Deep Matrix Factorization: Explicit Bounds. IEEE Signal Processing Letters, 33, 2884-2888. https://doi.org/10.1109/lsp.2026.3711534
Vancouver
Oliveira R, Savazzi S. On the Stability of Consensus Gradient Dynamics for Regularized Distributed Deep Matrix Factorization: Explicit Bounds. IEEE Signal Processing Letters. 2026;33:2884-2888. doi:10.1109/lsp.2026.3711534
BibTeX
@article{Oliveira2026,
author = {Oliveira, Richard and Savazzi, Stefano},
title = {On the Stability of Consensus Gradient Dynamics for Regularized Distributed Deep Matrix Factorization: Explicit Bounds},
journal = {IEEE Signal Processing Letters},
year = {2026},
volume = {33},
pages = {2884--2888},
doi = {10.1109/lsp.2026.3711534}
}
EUVIMED