Abstract
Abstract. Filtering deals with the optimal estimation of signals from their noisy observations. The standard setting consists of a pair of random processes (X, Y) = (Xt, Yt)t≥0, where the signal component X is to be estimated at a current time t> 0 on the basis of the trajectory of Y, observed up to t. Under the minimal mean square error criterion, the optimal estimate of Xt is the conditional expectation E(Xt|Y[0,t]). If both X and (X, Y) are Markov processes, then the conditional distribution πt(A) = P (Xt ∈ A|Y[0,t]), A ⊆ R satisfies a recursive equation, called filter, which realizes the optimal fusion of the a priori statistical knowledge about the signal and the a posteriori information borne by the observation path. The filtering equation is to be initialized by the probability distribution ν of the signal at time t = 0. Suppose ν is unknown and another reasonable probability distribution ¯ν is used to start the filter. As the corresponding solution ¯πt(·) differs from the optimal πt(·), the natural question of stability arises: what are the conditions in terms of the signal/observation parameters to guarantee limt→ ∞ ‖πt − ¯πt ‖ = 0 in an appropriate sense? The article discusses the recent progress in solving this stability problem, which turns