Expectation propagation for distributed estimation in sensor networks

We show that the expectation propagation (EP) family of algorithms constitute a natural choice for distributed estimation and detection in sensor networks. In particular, random sleep strategies, which are commonly chosen to ensure robustness, equal power dissipation across the network, and ease of deployment, espouse a sparse dependence structure among the parameters to be estimated. This sparse dependence structure mimics the structure which belief and expectation propagation exploited in the decoding of turbo and low density parity check (LDPC) codes to bring the performance of physical layer communications systems to the fundamental limits set out by Shannon. We provide examples of practical sensor network tasks which fall into the framework set out in this paper. By applying extensions of the extrinsic information transfer (EXIT) chart theory to EP in these distributed estimation applications, we can predict the performance and convergence of the distributed estimation algorithm in very large networks with an easy to obtain plot.