On the total variation distance of labelled Markov chains

Stefan Kiefer (University of Oxford)

When: October 5, 2pm

Where: room C417 (at the very end of the corridor)

Abstract

Labelled Markov chains (LMCs) are widely used in probabilistic verification, speech recognition, computational biology, and many other fields. Checking two LMCs for equivalence is a classical problem subject to extensive studies, while the total variation distance provides a natural measure for the ``inequivalence'' of two LMCs: it is the maximum difference between probabilities that the LMCs assign to the same event. We develop a theory of the total variation distance between two LMCs, with emphasis on the algorithmic aspects: (1) we provide a polynomial-time algorithm for determining whether two LMCs have distance 1, i.e., whether they can almost always be distinguished; (2) we provide an algorithm for approximating the distance with arbitrary precision; and (3) we show that the threshold problem, i.e., whether the distance exceeds a given threshold, is NP-hard and hard for the square-root-sum problem. We also make a connection between the total variation distance and Bernoulli convolutions. Joint work with Taolue Chen.