Nonlinear data fusion

Abstract
Consider the following estimation problem: The state trajectory of a random process is observed by K distinct observers, using noise-corrupted observations. Each observer processes his own observation history, to obtain the local conditional distribution of the state, as a function of time. Assume that sufficient statistics representing each local conditional distribution are communicated to a coordinator at a central location at each point in time. The coordinators's estimation problem consists of constructing the overall conditional distribution of the state, conditioned on knowing all of the observations, while using only the sufficient statistics communicated to him.