Rate region of the quadratic Gaussian CEO problem

In the so-called CEO problem, a hidden source random process is of interest to a central unit or the "CEO". But this process cannot be observed directly. L sensors or agents observe independently corrupted versions of the source. They encode their observations without cooperating with one another and send through rate constrained noiseless channels to the CEO. The problem was first studied by T. Berger et al. (1996) in the context of discrete memoryless sources. The quadratic Gaussian version of the problem was studied. The best result known to date is the characterization of the sum-rate when all the agents have the same quality of observations. Here we characterize the rate region for any number of agents without assuming that their quality of observations is the same. This is one of the few examples of multiterminal lossy source coding problems in which the rate region can be characterized completely.