The ideas previously used (Stochastics. Vol. 5, pp. 65–92, 1981) to construct some finite-dimensional nonlinear filters also yield related new filters of finite dimension with arbitrarily large bases; this is because the finite dimensionality is not destroyed by insertion of noiseless linear differential operations on the observations.
In engineering language, the new filters are obtained from the old by smoothing the output through an n-pole linear system before adding observation noise in the usual way; this adds 2n to the Lie algebra dimension. In the simplest case (drift = tanh x, observation = x) we put x through a one-pole described by a new variable ζ, and observe ζ + noise instead of x + noise; the new Lie algebra has additional generators ζ and ∂/∂ξ besides the four from the oscillator algebra, to give dimension 6.
The filter, which gives a recursive construction of the conditional density, can be ‘derived’ by any of three (here equivalent methods: (i) direct integration of the Kallianpur-Striebel formula as a Gaussian integral; (ii) solution of a parabolic PDE with quadratic potential; and (iii) the Wei-Norman procedure.
[1]
M. Zakai.
On the optimal filtering of diffusion processes
,
1969
.
[2]
S. Marcus,et al.
On Lie algebras and finite dimensional filtering
,
1982
.
[3]
D. Ocone.
Topics in Nonlinear Filtering Theory.
,
1980
.
[4]
M. Chaleyat-Maurel,et al.
Un théorème de non-existence de filtre de dimension finie
,
1983
.
[5]
V. Benes.
Exact finite-dimensional filters for certain diffusions with nonlinear drift
,
1981
.
[6]
Michiel Hazewinkel,et al.
Preface : Stochastic systems : the mathematics of filtering and identification and applications
,
1981
.
[7]
Mark H. A. Davis.
On a multiplicative functional transformation arising in nonlinear filtering theory
,
1980
.
[8]
G. Kallianpur,et al.
Arbitrary system process with additive white noise observation errors
,
1968
.