Stability properties of some particle filters (Q389073): Difference between revisions
From MaRDI portal
ReferenceBot (talk | contribs) Changed an Item |
Set OpenAlex properties. |
||
Property / OpenAlex ID | |||
Property / OpenAlex ID: W3101274747 / rank | |||
Normal rank |
Latest revision as of 09:45, 30 July 2024
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Stability properties of some particle filters |
scientific article |
Statements
Stability properties of some particle filters (English)
0 references
17 January 2014
0 references
The author deals with filtering problems for hidden Markov models. He considers a hidden Markov model which is a bivariate, discrete-time Markov chain \(((X_n, Y_n);n\geq0)\), where the signal process \((X_n)\) is also a Markov chain, with a (noncompact) state-space \(X\), and each observation \(Y_n\), with values in the observation space \(Y\), is conditionally independent of the rest of the bi-variate process given \(X_n\). The spaces \(X\) and \(Y\) are assumed to be Polish spaces endowed with their respective Borel \(\sigma\)-algebras. For observations \((y_0,y_1,\dots)\), a recursive one-step-ahead prediction filter is defined as the sequence \(((\pi_n);n\geq0)\) of conditional distributions of \(X_n\) given \((Y_0,Y_1,\dots,Y_{n-1})=(y_0,y_1,\dots,y_{n-1})\) and \(((Z_n);n\geq0)\) -- the joint density of \((Y_0,Y_1,\dots,Y_{n-1})\) evaluated at \((y_0,y_1,\dots,y_{n-1})\). Hidden Markov models are simple and yet flexible models which have found countless applications. However, in many practical situations, \(((\pi_n);n\geq0)\) and \(((Z_n);n\geq0)\) are not available in closed form. Particle filters are a class of stochastic algorithms which yield approximations \((\pi_n^N)\) and \((Z_n^N)\) of \((\pi_n)\) and \((Z_n)\) using \(N\) samples. A large number of variations and extensions of this algorithm have been developed. However, there are still very few results which establish stability over time of particle filtering methods. The author of the present paper proves some stability properties of a standard particle filter under assumptions which are verifiable for some hidden Markov models with noncompact state spaces. It is known that, under some mild conditions, the error associated with particle approximation of filtering distributions satisfies a central limit theorem. The first stability property obtained by the author is a time-uniform bound on the corresponding asymptotic variance. The second stability property obtained is a linear-in-time bound on the nonasymptotic, relative variance of the particle approximations of normalizing constants. These two properties are established by first proving some multiplicative stability and exponential moment results for the Feynman-Kac formulas underlying the particle filter. The adopted approach involves Lyapunov functions, multiplicative stability ideas in a weighted \(\infty\)-norm setting, which allows treatment of a noncompact state space. The main assumptions are typically satisfied under some constraints on the observation component of the hidden Markov model and/or the observation sequence driving the filter.
0 references
sequential Monte Carlo
0 references
filtering
0 references
particle filter
0 references
approximation
0 references
stability
0 references
Markov chain
0 references
hidden Markov model
0 references
0 references
0 references
0 references
0 references
0 references
0 references