Information geometry (Q5925492)

This paper is based on the 23rd Takagi Lectures delivered by the author on June 8, 2019, at Research Institute for Mathematical Science, Kyoto University. It is an interesting overview on topics of Information geometry. Basically Information geometry studies statistical manifolds, that is Riemannian manifolds with a symmetric cubic tensor, or equivalently Riemannian manifolds with a pair of dual affine connections. These manifolds can be considered as Riemannian manifolds whose points correspond to probability distributions. Such structures are important in many areas: mathematics and physics but also machine learning, signal processing, neuroscience. Differential geometry plays a central role in this study. Information geometry is strictly related to asymmetric divergence functions and affine differential geometry. Indeed on a manifold with such a divergence function a Riemannian structure with a pair of dual affine connections can be defined. Also a dual pair of affine connections can be defined on affine manifolds. In particular in a Riemannian manifold with a pair of flat affine connections there is a unique canonical divergence, a generalized Pythagorean theorem and a projection theorem hold. In this framework the Wasserstein distance, which concerns the distance of probability distributions, is very interesting. In this sense the author describes a divergence function obtained from the entropy-regularized Wasserstein problem and relates it to information geometry. Finally some future perspectives are discussed.