Asymptotically normal estimation of a parameter in a linear-fractional regression problem (Q1577174)

Let \(\xi_1,\dots,\xi_N\) be independent, identically distributed random variables with zero mean and unit variance. Let a sequence of independent random variables \(X_1,X_2,\dots\) be defined in the following way: \[ X_i=a_i(1+b_i\theta)^{-1}+\sigma_i\xi_i. \] The values \(a_i>0\) and \(b_i>0\) are assumed known, while the values of the parameter \(\theta\) and the variance \({\mathcal D}X_i\equiv\sigma_{i}^2\) are unknown. The authors consider the problem of estimating the unknown parameter \(\theta\) from the observations \(X_1,\dots,X_N\). They propose to use the following simple estimator: \[ \theta^*=\sum c_i(a_i-X_i)\bigl(\sum c_ib_iX_i\bigr)^{-1}. \] It turns out that this estimator is asymptotically normal under rather general assumptions on the constants \(\{c_i\}\). Moreover, in the case when some information on the behavior of the variances \(\{\sigma_i\}\) is available, the authors show how to choose functions \(\{\gamma_i(\theta)\}\) so that the ``improved'' estimator \[ \theta^{**}=\sum \gamma_i(\theta^*)(a_i-X_i)\bigl(\sum \gamma_i(\theta^*)b_iX_i\bigr)^{-1} \] becomes asymptotically efficient in some sense.