Information geometry and statistics

We apply the functional analytical and differential geometric results of the preceding chapters to the field of statistics and obtain very general versions of the basic classical results. In a narrower sense, the term statistic refers to a mapping from a given sample space Ω to another Ω′, and it is called sufficient for a parametric family, if the parameter can be estimated as well from that statistic as from samples taken on the original space Ω. More generally, a Markov kernel associates to a sample from Ω a probability measure on Ω′. We prove a general version of Chentsov’s theorem, saying that the Fisher metric and the Amari–Chentsov tensor are characterized by their invariance under sufficient statistics. The Cramér–Rao inequality, of which we also prove a new general version, controls the variance of an unbiased estimator from below by the inverse of the Fisher metric. When the estimator is possibly biased, additional terms enter, which we likewise handle in our framework.

DDC Class

510: Mathematik

Options

Information geometry and statistics