DECENTRALIZED SUBSPACE PROJECTION IN LARGE NETWORKS

A great number of applications in wireless sensor networks involve projecting a vector of observations onto a subspace dictated by prior information. Accomplishing such a task in a centralized fashion entails great power consumption, congestion at certain nodes, and suffers from robustness issues. A sensible alternative is to compute such projections in a decentralized fashion. To this end, recent works proposed schemes based on graph filters, which compute projections exactly with a finite number of local exchanges among sensor nodes. However, existing methods to obtain these filters are confined to reduced families of projection matrices or small networks. This paper proposes a method that can accommodate large networks and find suitable shift matrices in a much wider range of scenarios. Numerical experiments support the performance of the proposed algorithm.