Globally optimal decentralized spatial smoothing for wireless sensor networks with local interactions

In most sensor network applications, the vector containing the observations gathered by the sensors lies in a space of dimension equal to the number of nodes, typically because of observation noise, even though the useful signal belongs to a subspace of much smaller dimension. This motivates smoothing or rank reduction. We formulate a convex optimization problem, where we incorporate a fidelity constraint that prevents the final smoothed estimate from diverging too far from the observations. This leads to a distributed algorithm in which nodes exchange updates only with neighboring nodes. We show that the widely studied consensus algorithm is indeed only a very specific case of our more general formulation. Finally, we study the convergence rate and propose some approaches to maximize it.