A Unified Framework for Continuous-time Unconstrained Distributed Optimization

TL;DR

This paper presents a unified continuous-time framework for distributed optimization, encompassing existing algorithms and enabling the development of new methods through a general class of control dynamics.

Contribution

It introduces a novel class of flow-tracker dynamics that unify various distributed optimization algorithms and extend their convergence analysis.

Findings

Many existing algorithms are special cases of the proposed dynamics.

The framework guarantees convergence under generalized graph conditions.

New algorithms can be designed within this unified approach.

Abstract

We introduce a class of distributed nonlinear control systems, termed as the flow-tracker dynamics, which capture phenomena where the average state is controlled by the average control input, with no individual agent has direct access to this average. The agents update their estimates of the average through a nonlinear observer. We prove that utilizing a proper gradient feedback for any distributed control system that satisfies these conditions will lead to a solution of the corresponding distributed optimization problem. We show that many of the existing algorithms for solving distributed optimization are instances of this dynamics and hence, their convergence properties can follow from its properties. In this sense, the proposed method establishes a unified framework for distributed optimization in continuous-time. Moreover, this formulation allows us to introduce a suit of new continuous-time distributed optimization algorithms by readily extending the graph-theoretic conditions under which such dynamics are convergent.