Title: 
=======

On a role of predictor in filtering stability 

(Joint work with R.Liptser)

Abstract:
=======

The problem of filtering is to calculate the conditional distribution 
of the current state of the signal process, given the history of its 
noisy observations. For Markov processes, the conditional distribution 
satisfies a recursive filtering equation. The stability of this 
equation with respect to its initial condition is known under certain 
restrictive assumptions : typically the signal is required to be 
ergodic and to evolve on a compact state space. In this talk I will 
discuss certain predicting estimates of the signal state, which turn 
to be stable in a surprisingly general setting. The proof uses a 
change of measure and the martingale convergence and, unlike all 
other known methods, avoids direct analysis of the filtering 
equation.