On the Number of Iterations for Dantzig--Wolfe Optimization and Packing-Covering Approximation Algorithms

Abstract

We give a lower bound on the iteration complexity of a natural class of Lagrangian-relaxation algorithms for approximately solving packing/covering linear programs. We show that, given an input with $m$ random 0/1-constraints on $n$ variables, with high probability, any such algorithm requires $\Omega(\rho \log(m)/\epsilon^2)$ iterations to compute a $(1+\epsilon)$-approximate solution, where $\rho$ is the width of the input. The bound is tight for a range of the parameters $(m,n,\rho,\epsilon)$. The algorithms in the class include Dantzig--Wolfe decomposition, Benders' decomposition, Lagrangian relaxation as developed by Held and Karp for lower-bounding TSP, and many others (e.g., those by Plotkin, Shmoys, and Tardos and Grigoriadis and Khachiyan). To prove the bound, we use a discrepancy argument to show an analogous lower bound on the support size of $(1+\epsilon)$-approximate mixed strategies for random two-player zero-sum 0/1-matrix games.