Abstract: This paper addresses the problem of how one can
improve the performance of a non-optimal filter. First the theoretical question on dynamical representation for a given time correlated
random process is studied. It will be demonstrated that for a wide class of random processes, having a canonical form, there exists
a dynamical system equivalent in the sense that its output has the
same covariance function. It is shown that the dynamical approach is more effective for simulating and estimating a Markov and non-
Markovian random processes, computationally is less demanding,
especially with increasing of the dimension of simulated processes.
Numerical examples and estimation problems in low dimensional
systems are given to illustrate the advantages of the approach. A very useful application of the proposed approach is shown for the
problem of state estimation in very high dimensional systems. Here a modified filter for data assimilation in an oceanic numerical model
is presented which is proved to be very efficient due to introducing
a simple Markovian structure for the output prediction error process
and adaptive tuning some parameters of the Markov equation.
Abstract: The (sub)-optimal soolution of linear filtering problem
with correlated noises is considered. The special recursive form of
the class of filters and criteria for selecting the best estimator are
the essential elements of the design method. The properties of the
proposed filter are studied. In particular, for Markovian observation
noise, the approximate filter becomes an optimal Gevers-Kailath filter
subject to a special choice of the parameter in the class of given linear
recursive filters.