Distributed Nonconvex Optimization: Gradient-free Iterations and $ε$-Globally Optimal Solution

TL;DR

We address distributed constrained nonconvex optimization with univariate Lipschitz local objectives by introducing CPCA, a gradient-free framework that combines Chebyshev polynomial proxies, average consensus, and polynomial optimization to achieve $ε$-globally optimal solutions. CPCA replaces gradient or high-cost oracle queries with precomputed polynomial proxies, disseminates them across agents, and then solves an easier proxy problem to obtain a globally near-optimal solution with a provable accuracy split $ε_1+ε_2=ε$. The method provides a distributed stopping rule, analyzes zeroth-order query and communication complexities, and offers an SDP-based alternative for the polynomial optimization step, with extensions to multivariate settings discussed. Numerical results demonstrate CPCA’s efficiency in zeroth-order queries and competitive communication profiles compared to gradient-based and other zeroth-order schemes, highlighting its practical potential for large-scale distributed optimization tasks.

Abstract

Distributed optimization utilizes local computation and communication to realize a global aim of optimizing the sum of local objective functions. This article addresses a class of constrained distributed nonconvex optimization problems involving univariate objectives, aiming to achieve global optimization without requiring local evaluations of gradients at every iteration. We propose a novel algorithm named CPCA, exploiting the notion of combining Chebyshev polynomial approximation, average consensus, and polynomial optimization. The proposed algorithm is i) able to obtain $ε$-globally optimal solutions for any arbitrarily small given accuracy $ε$, ii) efficient in both zeroth-order queries (i.e., evaluations of function values) and inter-agent communication, and iii) distributed terminable when the specified precision requirement is met. The key insight is to use polynomial approximations to substitute for general local objectives, distribute these approximations via average consensus, and solve an easier approximate version of the original problem. Due to the nice analytic properties of polynomials, this approximation not only facilitates efficient global optimization, but also allows the design of gradient-free iterations to reduce cumulative costs of queries and achieve geometric convergence for solving nonconvex problems. We provide a comprehensive analysis of the accuracy and complexities of the proposed algorithm.

Figures (4)

• Figure 1: An overview of CPCA and its algorithmic flow.
• Figure 2: Comparison of CPCA, SONATA-L, Alg. 2 in tang2019distributed, and Proj-DGD for solving problem \ref{['problem:main_focus']} with \ref{['eq:local_obj_exp']} as local objectives.
• Figure 3: Comparison of CPCA, SONATA-L, Alg. 2 in tang2019distributed, and Proj-SGD for solving problem \ref{['problem:main_focus']} with \ref{['eq:local_obj_log_exp']} as local objectives.
• Figure 4: Performance of the proposed algorithm with the basic and accelerated (corresponding to "acc." in the captions) consensus methods.

Theorems & Definitions (14)

• Theorem 1
• proof
• Remark 1
• Remark 2
• Lemma 1
• proof
• Theorem 2
• proof
• Theorem 3
• proof