Entropy, Information, and the Updating of Probabilities

This paper is a review of particular approach to the method maximum entropy as general framework for inference. The discussion emphasizes pragmatic elements in derivation. An epistemic notion information defined terms its relation Bayesian beliefs ideally rational agents. updating from prior posterior probability distribution designed through an eliminative induction process. logarithmic relative singled out unique tool that (a) universal applicability; (b) recognizes value information; and (c) privileged role played by independence science. resulting -- ME can handle arbitrary priors constraints. It includes MaxEnt Bayes' rule special cases and, therefore, it unifies entropic methods into single inference scheme. goes beyond mere selection posterior, but also addresses question how much less probable other distributions might be, which provides direct bridge theories fluctuations large deviations.