GEOMETRICAL-STATISTICS—CLASSICAL-AND-QUANTUM1.pdfVIP

Stockholm USITP 05-3 September 2005 GEOMETRICAL STATISTICS — CLASSICAL AND QUANTUM1 Ingemar Bengtsson2 Stockholm University, Alba Nova Fysikum S-106 91 Stockholm, Sweden Abstract This is a review of the ideas behind the Fisher–Rao metric on classical prob- ability distributions, and how they generalize to metrics on density matrices. As is well known, the unique Fisher–Rao metric then becomes a large family of monotone metrics. Finally I focus on the Bures–Uhlmann metric, and dis- cuss a recent result that connects the geometric operator mean to a geodesic billiard on the set of density matrices. 1Talk at the third V¨axj¨o conference on Quantum Theory: Reconsideration of founda- tions, June 2005. 2Email address: ingemar@physto.se. Supported by VR. 1 1. Classical preliminaries How far apart are two probability distributions? An answer to this question should encapsulate the ease with which the two distributions can be distin- guished from each other by means of some kind of sampling. The answer will depend very much on what the precise rules of this game are. What kind of Riemannian metric is appropriate on a set of classical prob- ability distributions over N events, that is on a probability simplex? One way of answering the second question proceeds by taking the first question seriously, and formulating it for a large number of samplings, so that only infinitesimal distances need to be considered. In effect, with some minor changes, we will use the idea that long ago led investigators in colour theory to introduce a metric on the convex set of colours. This is, in a way, very close to the original ideas of Riemann himself [1]. But we c