Asymptotically normal families of distributions and efficient estimation (Q1184200)

A very important role is played by the notion of local asymptotic normality (LAN). In particular, the theory of LAN establishes very general lower bounds on the accuracy of estimates. Moreover, in nonparametric estimation the corresponding lower bounds can be attained in some infinite-dimensional estimation problems. In this article the authors suggest a new definition of LAN for families of distributions \(\{P_ \theta\), \(\theta\in\Theta\}\), where the parameter space \(\Theta\) is a subset of a normal space or a smooth infinite-dimensional manifold. For such families with an infinite- dimensional parametric set \(\Theta\), satisfying the LAN condition, it is required to estimate the value \(\varphi(\theta)\) of a known (Euclidean or) Hilbert-valued function at an unknown point \(\theta\in\Theta\) on the basis of an observation from the given family. This problem is treated as a semiparametric estimation problem. Variants of some known results (convolution theorem and asymptotic minimax bound) are derived, under the LAN conditions. The latter result provides the notion of an asymptotically efficient estimator. It is difficult to construct asymptotically efficient estimators in situations where the parameter set is infinite-dimensional. The authors [see e.g., ``Statistical estimation. Asymptotic theory'' (1981; Zbl 0467.62026)] have previously suggested a method of constructing such estimators. This method is demonstrated here for a Gaussian white-noise estimation problem.