情報幾何学

Abstract

Information geometry is a new theoretical method to elucidate intrinsic geometrical structures underlying inforrnation systems. It is applicable to wide areas of information sciences including statistics, information theory, systems theory, etc. More concretely, information geometry studies the intrinsic geometrical structure of the manifold of probability distributions. It is found that the manifold of probability distributions leads us to a new and rich differential geometrical theory. Since most of information sciences are closely related to probability distributions, it gives a powerfu] method to study their intrinsic structures. A manifold consisting of, a smooth family of probability distributions has a unique invariant Riernannian metric given by the Fisher information. It admits a one-parameter family of invariant affne connections, called the a -connection, where a and a-connections are dually coupled with the Riemannian metric. The duality in afine connections is a new concept in differential geometry. When a manifold is dually flat, it admits an invariant divergence measure for which a genetalized Pythagorian theorem and a projection theorem hold. The dual structure of such manifolds can be applied to statistigal inference, multiterininal information theory, control systems theory, neural networks manifolds, etc. It has potential ability to be applied to general disciplines including physical and engineering sciences.