Computational Convergence Analysis of Distributed Gradient Descent for Smooth Convex Objective Functions.
Year: 2018
Type: preprint
Abstract: We present a computational proof on the $O(1/K)$ convergence rate of distributed gradient descent when the objective function is smooth and convex (but not strongly convex). The method is inspired by recent work on applying tools from robust control, in particular integral quadratic constraint (IQC), and dissipativity theory in analyzing optimization algorithms. We show that IQC and dissipativity theory can be used together in a unified framework, which is useful for analyzing the joint setting of distributed optimization and non-strongly convex objective functions. Our method relies on only a few analytic derivations from basic properties of convex functions, after which a numerical certificate of convergence can be automatically generated by solving a linear matrix inequality. The computational proof is found to certify convergence for a much broader range of step size than what is given by the original analytic proof for the same algorithm.
Source: arXiv (Cornell University)
Author: Shuo Han
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2894350555
