Faster Distributed Subgradient for Constrained Non-Smooth Optimization with Hölderian Growth

The Journal of Geometric Analysis · Springer Nature · 2026-09-26 · DOI: 10.1007/s11424-026-5442-y

Аннотация

This paper develops two subgradient-based algorithms for solving distributed constrained non-smooth optimization problems under a Hölderian growth condition (HGC) with θ ∈ (0, 1]. Firstly, the authors propose a projected subgradient method with a fixed step size and demonstrate that it linearly converges to a suboptimal solution. Moreover, for $$\theta \in (0, \frac{1}{2}]$$, the authors improve the projected subgradient method with diminishing step sizes to find an ε -solution in terms of decision variables in $$\cal{O}_x(e^{-\frac{2(1-\theta)}{\theta}})$$ iterations, which is faster than that of the conventional distributed subgradient methods. Furthermore, for $$\theta \in [\frac{1}{2}, 1]$$ the authors design an epoch-based projected subgradient method and demonstrate its $$\cal{O}_x(\varepsilon^{-\frac{4\theta^2-2\theta+1}{2\theta^2}})$$ iteration complexity to find an ε -solution, which is better than existing subgradient methods. Finally, the authors present two examples to illustrate the effectiveness of the proposed algorithms.

Читать у издателя →

Похожие статьи