Robustness of Nonlinear Filters Over the Infinite Time Interval (Q4210181)
From MaRDI portal
!
This is the item page for this Wikibase entity, intended for internal use and editing purposes. Please use the normal view instead:
scientific article; zbMATH DE number 1200820
| Language | Label | Description | Also known as |
|---|---|---|---|
| default for all languages | No label defined |
||
| English | Robustness of Nonlinear Filters Over the Infinite Time Interval |
scientific article; zbMATH DE number 1200820 |
Statements
Robustness of Nonlinear Filters Over the Infinite Time Interval (English)
0 references
21 September 1998
0 references
nonlinear filtering
0 references
model robustness
0 references
asymptotic stability
0 references
Hilbert metric
0 references
Birkhoff's contraction coefficient
0 references
initial law
0 references
initial condition
0 references
insensitivity
0 references
0.8676410913467407
0 references
0.8296810984611511
0 references
0.8235264420509338
0 references
This paper studies the sensitivity of the filter to the initial law, the transition function of the signal process, and the model for the observation: when the time interval is large, errors might accummulate. With the classical Kalman-Bucy filter, the effect of the initial condition disappears as time goes to infinity. This paper is devoted to the double robustness problem: to know that small errors in the signal model do not have serious effects on the filter output, and insensitivity to the initial law of the filter; both problems being considered over an arbitrarily large interval. Three models are studied. For instance, the first one presents the signal as a stationary Markov chain \(X\) with transition kernel \(G\) such that there exists a probability measure \(\rho\) satisfying \(C_1\rho(A) \leq\int_A G(x,y)\lambda(dy)\leq C_2\rho(A)\) for any Borel set \(A\) (mixing condition) and an observation process \(Y_n=H(X_n)+ \nu_n\), where \(\nu\) is a noise process and \(H\) a measurable function. The exact filter at time \(n\) is denoted as \(\Pi_n\), and \(\Pi_n^{(k)}\) is the filter with erroneous transitional kernel \(G^{(k)}\) and erroneous initial law \(\rho_k\). Then, if \((\log G^{(k)})\) converges to log \(G\) uniformily on \(\{(x,y):G(x,y) > 0\}\), then \[ \lim_{k\to \infty}\limsup_{n\to\infty} h(\Pi_n^{(k)},\Pi_n)=0, \] where \(h\) is the Hilbert projective distance on the space of finite nonnegative measures. If moreover \(\|\rho_k -\rho_0 \|_{TV}\) goes to 0 when \(k\) goes to infinity, then \[ \lim_{k \to\infty} \sup_n\| \Pi^{(k)}_n -\Pi_n \|_{TV}=0, \] where \(\|\cdot \|_{TV}\) is the total variation norm. The main tools used in this work are Hilbert's projective metric and Birkhoff's contraction inequality.
0 references