An exponentially stable discrete-time primal-dual algorithm for distributed constrained optimization

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Ren, Xiaoxing, Bin, Michelangelo, Notarnicola, Ivano, Parisini, Thomas
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915189486518272
author Ren, Xiaoxing
Bin, Michelangelo
Notarnicola, Ivano
Parisini, Thomas
author_facet Ren, Xiaoxing
Bin, Michelangelo
Notarnicola, Ivano
Parisini, Thomas
contents This paper studies a distributed algorithm for constrained consensus optimization that is obtained by fusing the Arrow-Hurwicz-Uzawa primal-dual gradient method for centralized constrained optimization and the Wang-Elia method for distributed unconstrained optimization. It is shown that the optimal primal-dual point is a semiglobally exponentially stable equilibrium for the algorithm, which implies linear convergence. The analysis is based on the separation between a slow centralized optimization dynamics describing the evolution of the average estimate toward the optimum, and a fast dynamics describing the evolution of the consensus error over the network. These two dynamics are mutually coupled, and the stability analysis builds on control theoretic tools such as time-scale separation, Lyapunov theory, and the small-gain principle. Our analysis approach highlights that the consensus dynamics can be seen as a fast, parasite one, and that stability of the distributed algorithm is obtained as a robustness consequence of the semiglobal exponential stability properties of the centralized method. This perspective can be used to enable other significant extensions, such as time-varying networks or delayed communication, that can be seen as ``perturbations" of the centralized algorithm.
format Preprint
id arxiv_https___arxiv_org_abs_2503_06662
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle An exponentially stable discrete-time primal-dual algorithm for distributed constrained optimization
Ren, Xiaoxing
Bin, Michelangelo
Notarnicola, Ivano
Parisini, Thomas
Optimization and Control
65K05, 93A14, 93A16, 90C25, 90C30, 90C35, 93D05, 93D23
This paper studies a distributed algorithm for constrained consensus optimization that is obtained by fusing the Arrow-Hurwicz-Uzawa primal-dual gradient method for centralized constrained optimization and the Wang-Elia method for distributed unconstrained optimization. It is shown that the optimal primal-dual point is a semiglobally exponentially stable equilibrium for the algorithm, which implies linear convergence. The analysis is based on the separation between a slow centralized optimization dynamics describing the evolution of the average estimate toward the optimum, and a fast dynamics describing the evolution of the consensus error over the network. These two dynamics are mutually coupled, and the stability analysis builds on control theoretic tools such as time-scale separation, Lyapunov theory, and the small-gain principle. Our analysis approach highlights that the consensus dynamics can be seen as a fast, parasite one, and that stability of the distributed algorithm is obtained as a robustness consequence of the semiglobal exponential stability properties of the centralized method. This perspective can be used to enable other significant extensions, such as time-varying networks or delayed communication, that can be seen as ``perturbations" of the centralized algorithm.
title An exponentially stable discrete-time primal-dual algorithm for distributed constrained optimization
topic Optimization and Control
65K05, 93A14, 93A16, 90C25, 90C30, 90C35, 93D05, 93D23
url https://arxiv.org/abs/2503.06662