SPRT and CUSUM in hidden Markov models (Q1412371)

This interesting paper is devoted to sequential probability ratio tests for parametrized hidden Markov models. In particular, the performance and asymptotic optimality of these tests are investigated, and corrected Brownian approximations for the error probabilities are given. The same is done for cumulative sum procedures, as by the renewal property, these procedures may be regarded as repeated one-sided sequential probability ratio tests. Moreover, motivated by the sequential analysis of hidden Markov models, Wald's likelihood ratio identity and Wald's equation for products of Markov random matrices are also given. The main results of the paper are then applied to several classes of examples like i.i.d. hidden Markov models and certain Gaussian models which appear in applications like speech recognition.