Asymptotically efficient estimation of functionals in a Gaussian white noise (Q1363254)

We consider the following statistical problem of estimation: The observed process \[ X_\varepsilon(t)= \int^t_0 \theta(u)du +\varepsilon w(t), \quad 0\leq t\leq 1, \tag{1} \] where \(w\) is a standard Wiener process, and \(\theta(t)\) is an unknown functional parameter, \(\theta\in\Theta \subseteq L_2 (0,1)\); below, a norm and a scalar product in \(L_2\) are denoted by \(|\cdot |\) and \((\cdot, \cdot)\), respectively. We suppose that the parametric set \(\Theta\) and the parameter \(\varepsilon\) are known. Let \(\Phi: L_2\to R^1\) be a known Frechet differentiable functional with the derivative \(\Phi'\). Let the value \(\Phi (\theta)\) of the functional \(\Phi\) at the point \(\theta\) be estimated by the observations of (1). Is it possible to construct asymptotically efficient estimates for \(\Phi (\theta)\) as \((\varepsilon\to 0)\)? We understand the efficiency of statistical estimates in Bahadur's sense, where the asymptotic efficiency of the estimates \(T_\varepsilon\) is measured by the value of concentration of estimates in the interval of fixed (i.e., independent of \(\varepsilon)\) length with the center at the value to be estimated \([\Phi (\theta)\), in our case], i.e., by the value of probabilities \({\mathbf P}_\theta \{|T_\varepsilon -\Phi (\theta)|>\gamma\}\), or, rather, by their asymptotics as \(\varepsilon\to 0\), or even as \(\varepsilon\to 0\) and \(\gamma\to 0\). As a rule, these probabilities decrease exponentially rapidly; therefore, it is natural to take the upper bound of the expressions \[ \varliminf_{\gamma\to 0} \varliminf_{\varepsilon\to 0} \varepsilon^2 \gamma^{-2} \ln {\mathbf P}_\theta \biggl\{ \bigl|T_\varepsilon-\Phi(\theta) \bigr|> \gamma \biggr\} \] over all \(T_\varepsilon\) as the measure of the asymptotic efficiency. Unlike the efficiency in Fisher's sense, the estimates that are asymptotically efficient in Bahadur's sense turn out to already exist if the parametric set \(\Theta\) is compact in \(L_2\). (If \(\Theta\) is not compact, then one cannot be sure even of the existence of consistent estimates).