Computable infinite-dimensional filters with applications to discretized diffusion processes

Abstract

Let us consider a pair signal-observation ((x n , y n ), n ≥ 0) where the unobserved signal (x n ) is a Markov chain and the observed component is such that, given the whole sequence (x n ), the random variables (y n ) are independent and the conditional distribution of y n only depends on the corresponding state variable x n . The main problems raised by these observations are the prediction and filtering of (x n ). We introduce sufficient conditions allowing to obtain computable filters using mixtures of distributions. The filter system may be finite or infinite dimensional. The method is applied to the case where the signal x n = X n∆ is a discrete sampling of a one dimensional diffusion process: Concrete models are proved to fit in our conditions. Moreover, for these models, exact likelihood inference based on the observation (y 0 , . . . , y n ) is feasable.MSC: primary 93E11, 60G35; secondary 62C10.

Adaptive estimation of the transition density of a particular hidden Markov chain

“…Following Chaleya-Maurel and Genon-Catalot [11], we observe that X n is the square of the Euclidean norm of a -dimensional vector whose components are linear autoregressive processes of order 1. The stationary distribution is a Gamma distribution with parameter /2 and | |/ 2 0 so that…”

Filtering the Wright-Fisher diffusion

“…In Section 2, we briefly recall some properties of the Wright-Fisher diffusion. In Section 3, we recall the filtering-prediction algorithm and the sufficient conditions of [2] to obtain computable filters. Section 4 contains our main results.…”

“…Evidently, the difficulty is to find models satisfying these conditions. Examples are given in [2] where the hidden Markov process (x(t n )) is a discretization of a diffusion process. Here, we study another diffusion process.…”

“…The property that Beta distributions are conjugate with respect to Bernoulli distributions is well known in Bayesian statistics (see e.g. [2]). Analogously, the class F is also a conjugate class for the parametric family of distributions (1.2) and (1.3) with respect to x ∈ (0, 1).…”

On the asymptotic variance in the central limit theorem for particle filters

“…Unfortunately, except for very few models, such as the Kalman filter or some other models (for instance, those presented in [3]), these recursions rapidly lead to intractable computations and exact formulae are out of reach. Moreover, the standard Monte-Carlo methods fail to provide good approximations of these distributions (see e.g.…”