A Lower Bound on Convergence of a Distributed Network Consensus Algorithm

This paper gives a lower bound on the convergence rate of a class of network consensus algorithms. Two different approaches using directed graphs as a main tool are introduced: one is to compute the "scrambling constants" of stochastic matrices associated with "neighbor shared graphs" and the other is to analyze random walks on a sequence of graphs. Both approaches prove that the time to reach consensus within a dynamic network is logarithmic in the relative error and is in worst case exponential in the size of the network.