Asymptotic Bayesian estimation of a first order equation with small diffusion (Q796167)

Let \(t\to x^{\epsilon}(t)\), \(\epsilon>0\), denote a diffusion process evolving in \({\mathbb{R}}^ n\) and governed by a generator of the form \(A^{\epsilon}=f+(\epsilon /2)(g^ 2_ 1+...+g^ 2_ m)\), where \(f,g_ 1,...,g_ m\) is a set of vector fields on \({\mathbb{R}}^ n\). Suppose that \(t\to x^{\epsilon}(t)\) is observed in the presence of an independent Brownian motion, i.e. \(t\to x^{\epsilon}(t)\) and \(t\to y^{\epsilon}(t)=\int^{t}_{0}h(x^{\epsilon}(s))ds+\sqrt{\epsilon}b( t),\) are considered, where \(t\to b(t)\) denotes Brownian motion and h: \({\mathbb{R}}^ n\to {\mathbb{R}}^ P\) is a given map. The main object is to establish upper and lower large deviation estimates for the conditional distribution \(P^{\epsilon}_{x| y}\) of \(t\to x^{\epsilon}(t)\) given \(t\to y^{\epsilon}(t)\), as \(\epsilon \searrow 0\). These provide extensions of the analogous estimates for the unconditional distribution \(P_ x^{\epsilon}\), the latter case being covered by specifying \(h=0\). The proof is mainly based upon the unconditional results, one part of which is treated by a new compactness argument. An application is also briefly discussed.