Information Bounds for the Risk of Bayesian Predictions and the Redundancy of Universal Codes
A. Barren, Bertrand S. Clarke, David H. Haussler · 2005
Several diverse problems have solutions in terms of an information-theoretic quantity for which we examine the asymptotic. Let Y/sub1/, Y/sub 2/, Y/sub n/ be a sample of random variables with distribution depending on a (possibly infinite-dimensional) parameter /spl theta/. The maximum of the mutual information I/sub n/ = I(/spl theta/; Y/sub 1/, Y/sub 2/, . . . . Y/sub n/) over choices of the prior distribution of /spl theta/ provides a bound on the cumulative Bayes risk of prediction of the sequence of random variables for several choices of loss function. This same quantity is the minimax redundancy of universal data compression and the capacity of certain channels. General bounds for this mutual information are given. A special case concerns the estimation of binary-valued functions with Vapnik-Chervonenkis dimension due, for which the information is bounded by d/sub vc/ log N. For smooth families of probability densities with a Euclidean parameter of dimension d, the information bound is (d/2) log N plus a constant. The prior density proportional to the square root of the Fisher information determinant is the unique continuous density that achieves a mutual information within o(1) of the capacity for large N. The Bayesian procedure with this prior is asymptotically minimax for the cumulative relative entropy risk.