Moderate deviations for particle filtering (Q1774189)

A ``state space in discrete time'' model underlies the paper being reviewed. The state is a Markov chain (\({\underline{X}}_n\in \mathbb R^p\)) and the observations (\({\underline{Y}}_n\in \mathbb R^q\)) are independent given the current state. The conditional law of the current state given the preceding one and that of the current observation given the current state are each assumed to have a conditional density with respect to an a priori law. The problem is that of estimating recursively the law of the current state given current and past observations. It is usually impossible to obtain that law analytically and particle filters are means to obtain approximations of that law in the form of an average of point masses. They typically sample from a recursively defined conditional density for the state and are related to genetic algorithms. The theory then covers questions of convergence of these particle filters. The asymptotic behaviour of averages of centered and scaled functional values of the states is investigated and bounds for sequences of the following type are provided: \(\frac{1}{b_N^2}\ln [P_T(M_N^T\in \Gamma)]\), where \(M_N^T\) is the average considered, \(T\) the observation span, \(N\) the sampling size, \(\Gamma\) a Borel set and \(b_N\) a sequence increasing more slowly than \(\sqrt{N}.\) \(P_T\) represents the conditional law. The bounds are in terms of the behaviour of a ``good rate function'' on the interior and closure of \(\Gamma.\) This rate function depends on the asymptotic covariance matrix of the considered functionals of the state. Towards the end of the paper three examples are considered, but only as far as to show that the assumptions obtain indeed that are required for the results of the paper to be valid. Given the complexity of particle filters constructions, extended applications in a ``real world context'' would seem necessary for the evaluation of their scope.