Analysis of accelerated gossip algorithms

Gossiping is a distributed process whose purpose is to enable the members of a group of n>1n>1 autonomous agents to asymptotically determine in a decentralized manner, the average of the initial values of their scalar gossip variables. This paper analyzes the accelerated gossip algorithms, first proposed in Cao, Spielman, and Yeh (2006), in which local memory is exploited by installing shift-registers at each agent. For the two-register case, the existence of the desired convergence is established under a symmetry assumption by separately studying the convergence in expectation and in mean square. In particular, the optimal rate of convergence in expectation is derived which is faster than that of the standard gossip algorithm, and a sufficient condition on the adjustable parameter for the convergence in mean square is provided. These theoretical results are validated for some classes of networks by comparison with existing empirical data. More general multi-register cases are also discussed.