Reference [8] in the NIPS 2005 Paper
O. Patrick Kreidl
The NIPS
2005 paper refers the reader to a publication titled "Posterior
assignment in directed graphical models with minimal online
communication." This publication has not been produced, however, but
the referenced results can be viewed as a special case of the results
presented in the CDC
2006 paper. For the reader of the earlier NIPS 2005 paper, it is
worth emphasizing that Proposition 4 turns out to hold even without
the stated assumption that the global hidden process X is
Markov on the same graph that defines the constrained
communication. Specifically, as discussed in the CDC 2006 paper, it
suffices to assume that each node knows the joint probability
mass function of the hidden variables local to itself and its parents.