Achieving information bounds in non and semiparametric models (Q756324)

If a parametric model \({\mathbb{P}}=\{P_{(\theta,\eta)}:\theta\in R\), \(\eta \in R^ m\}\) for a random variable X with a \(\mu\)-density \(p(x)=[dP_{(\theta,\eta)}/d\mu](x)=p_{(\theta,\eta)}(x)\) is regular in some sense, then \(\theta\) can be estimated at the rate \(n^{-1/2}\) with smallest asymptotic variance \(I^{11}(\theta,\eta)\) given by the first element of the inverse of Fisher's information matrix I(\(\theta\),\(\eta\)). In non or semi-parametric cases, this bound may be totally misleading even if \(I>0\). In fact, two examples are given where no estimators converge at a rate \(n^{-\alpha}\) for any \(\alpha >0:\) (1) Estimation of the integral \(\theta (p)=\int p^ 2(x)d\mu (x)\). (2) Estimation of \(\theta\) in the model \(Y=\theta W+t(Z)+\epsilon\) considered by \textit{R. F. Engle}, \textit{C. W. J. Granger}, \textit{R. Rice} and \textit{A. Weiss} [J. Am. Stat. Assoc. 81, 310-320 (1986)] with \(\epsilon \sim N(0,\sigma^ 2).\) The proofs proceed by a Bayesian construction of suitable estimators for p(\(\cdot)\), \(\theta\), t(\(\cdot)\), respectively. A positive result is obtained for the semiparametric model \({\mathbb{P}}=\cup^{\infty}_{m=0}{\mathbb{P}}_ m\) with nice regular families \({\mathbb{P}}_ m=\{P_{(\theta,\eta^ m)}:\theta\in \Theta\), \(\eta^ m=(\eta_ 1,...,\eta_{d(m)})\}\) where \({\mathbb{P}}_ m\subseteq {\mathbb{P}}_{m+1}\) and the dimension d(m) is generally increasing: Conditions are given for which the lower information bound of the form \[ I^{-1}(P_ 0,\theta)=\| 4^{-1}(\dot s_ 1-\prod (\dot s_ 1| {\dot \zeta}_ 2(\theta_ 0,\eta^ m_ 0))\|^{-2} \] is achieved at a given \(P_ 0=(\theta_ 0,\eta^ m_ 0)\) (with \(\|\) the \(L_ 2\)-norm, \(\prod\) the projection operator in \(L_ 2\), \(\dot s_ 1\) the first element of the Fréchet derivative of \(s(\theta,\eta)=p^{1/2}_{(\theta,\eta)}\) etc.).