Information geometry approach to parameter estimation in hidden Markov model (Q2073215)

The author proposes a theoretical EM-algorithm approach to estimate parameters of a Hidden Markov Model (HMM) based on the geometrical properties of the transition matrices. It is assumed that the HMM has \(k-1\) memory and is parametrized by an exponential family of transition matrices with generators \(\{g_{i}\}\). Instead of working with all observations, \(\{Y_{j}\}\) the estimation procedure is based on averages of the observations \[ Y_{i}^{n} = \frac{1}{n}\sum\limits_{j=1}^{n-1}g_{i}(Y_{j+1},\ldots,Y_{j+k}). \] It is shown, under some regularity conditions, that the error of the estimator converges to a zero-mean Gaussian distribution, and that for a sufficiently large number of observations a \(\chi^{2}\) goodness-of-fit test can be used. When the Markov process has order \(k\) large enough, regularity conditions are met, and the true transition matrices come from an appropriate set, then as the number of observations goes to infinity, then the projective Fisher information matrix is invertible, and the distribution of the error of the estimator converges to a mean-zero Gaussian with variance equalling the inverse of the projective Fisher information matrix.