Analysis of the Heavy-ball Algorithm using Integral Quadratic Constraints

Abstract

In this paper, we analyze the convergence rate of the Heavy-ball algorithm applied to optimize a class of continuously differentiable functions. The analysis is performed with the Heavy-ball tuned to achieve the best convergence rate on the sub-class of quadratic functions. We review recent work to characterize convergence rate upper bounds for optimization algorithms using integral quadratic constraints (IQC). This yields a linear matrix inequality (LMI) condition which is typically solved numerically to obtain convergence rate bounds. We construct an analytical solution for this LMI condition using a specific “weighted off-by-one” IQC. We also construct a specific objective function such that the Heavy-ball algorithm enters a limit cycle. These results demonstrate that IQC condition is tight for the analysis of the tuned Heavy-ball, i.e. it yields the exact condition ratio that separates global convergence from non-global convergence for the algorithm.