Accelerated gradient methods for networked optimization

This paper explores the use of accelerated gradient methods in networked optimization. Optimal algorithm parameters and associated convergence rates are derived for distributed resource allocation and consensus problems, and the practical performance of the accelerated gradient algorithms are shown to outperform alternatives in the literature. Since the optimal parameters for the accelerated gradient method depends on upper and lower bounds of the Hessian, we study how errors in these estimates influence the convergence rate of the algorithm. This analysis identifies, among other things, cases where erroneous estimates of the Hessian bounds cause the accelerated method to have slower convergence than the corresponding (non-accelerated) gradient method. An application to Internet congestion control illustrates these issues.