A filtering problem with counting observations:error bounds due to the uncertainty on the infinitesimal parameters (Q4354629)

The paper is devoted to the sensitivity (robustness) analysis of an (ideal) filter \(F\) whose state \(X\) is not observable, the observation \(Y\) is a counting process, while \((X,Y)\) is supposed to be a pure jump Markov process with discrete state space. This filter \(F\) is compared with the ``incorrect'' filter \(F'\) of another system which differs from the previous one only by the infinitesimal parameters and the initial distribution. The main result of the paper is a bound of the distance in variation norm \(|F-F' |\) between the two filters, expressed in terms of the distance between the initial distributions and between the infinitesimal parameters, or as a coupling of two Markov chains with different transition probabilities and initial laws. Two important consequences are entailed: (1) The discrete time approximation of the incorrect filter \(F'\) is a robust approximation of the ideal filter \(F\); (2) If a model with infinite state space is considered, under certain special assumptions, then an upper bound can be obtained for the error due to the use of a finite state space model instead of an infinite one.