A1373
Title: Leveraging one-step transitions to improve Markov chain inference
Authors: Henri Pechoux - INRIA (France) [presenting]
Romain Azais - Inria (France)
Benoit Henry - IMT Nord Europe (France)
Abstract: Markov chains play a central role in the modeling of many real phenomena arising in the social and natural sciences. The estimation of the stationary distribution of such a process when the distribution of one-step transitions is available along with observations is considered. A new estimator is proposed whose performance improves upon that of the classical empirical estimator. Subsequently, the more realistic assumption of a parametric model to describe one-step transitions is adopted, which must be estimated in order to apply the aforementioned method. Under mild assumptions, the resulting estimator is shown to improve the performance of the empirical estimator. Finally, a situation is exhibited in which the assumptions are satisfied while the natural plug-in estimator of the stationary distribution is not defined.
