Viterbi sequences and polytopes

A Viterbi path of length n of a discrete Markov chain is a sequence of n + 1 states that has the greatest probability of ocurring in the Markov chain. We divide the space of all Markov chains into Viterbi regions in which two Markov chains are in the same region if they have the same set of Viterbi paths. The Viterbi paths of regions of positive measure are called Viterbi sequences. Our main results are (1) each Viterbi sequence can be divided into a prefix, periodic interior, and suffix, and (2) as n increases to infinity (and the number of states remains fixed), the number of Viterbi regions remains bounded. The Viterbi regions correspond to the vertices of a Newton polytope of a polynomial whose terms are the probabilities of sequences of length n. We characterize Viterbi sequences and polytopes for two- and three-state Markov chains.

• Type: Journal
• Year: 2006
• Where: JSC
• Authors: Eric H. Kuo
• Keywords: JSC 2006, Markov Chains, Viterbi Paths, Viterbi Sequences