Distance function for stochastic matrices.

The ability to quantify the similarity of stochastic processes is important in a wide range of areas including communications, simulation, and operational research. Such processes are commonly modeled as Markov Chains and so a natural way to compare their similarity is to compute distances between pairs of Markov Chains. We propose a distance function on the space of stochastic matrices that draws on ideas from information geometry. We first show that the Bhattacharyya angle is an appropriate measure of distance between Markov chain sequences, drawing on examples taken from healthcare processes and deriving bounds on the convergence of the distance and mixing times. We then extend these ideas to derive a distance measure which enables direct comparison of the transition matrices themselves. Our result is a true metric which has a closed form and is efficient to implement for numerical evaluation. In the case of ergodic Markov chains, it is shown that considering either the Bhattacharyya angle on Markov sequences or our stochastic matrix distance leads to the same distance between models.