Particle filtering under communications constraints

Particle filtering is often applied to the problem of object tracking under non-Gaussian uncertainty: however, sensor networks frequently require that the implementation be local to the region of interest, eventually forcing the large, sample-based representation to be moved among power-constrained sensors. We consider the problem of successive approximation (i.e., lossy compression) of each sample-based density estimate, in particular exploring the consequences (both theoretical and empirical) of several possible choices of loss function and their interpretation in terms of future errors in inference, justifying their use for measuring approximations in distributed panicle filtering