Asymptotic optimality in sequential interval estimation (Q1085930): Difference between revisions

Let \(X_ 1\), \(X_ 2,..\). be i.i.d. normal random variables with unknown mean \(\mu\) and unknown variance \(\sigma^ 2\). Let \(\bar X{}_ n\) and \(\bar S{}^ 2_ n\) denote the sample means and variances. For h \((>0)\) let \(I_ n=[\bar X_ n-h\), \(\bar X+h]\) be the confidence intervals with confidence coefficient being at least larger than a prescribed \(\gamma\) \((0<\gamma <1)\). Let c denote the \((1+\gamma)/2\) quantile of the standard normal distribution and let t be a stopping time with respect to the sequence of sample variances. Put \(N=c^ 2\sigma^ 2/h^ 2\). (N is the hypothetical fixed sample size which would be required if \(\sigma\) were known.) Define the loss function by \(L_ h(t : \sigma)=N K(t/N)\) where \(K(x)=2ac^{-2}[1-\Phi (c\sqrt{x})]+x-[ac^{-2}(1-\gamma)+1]\), \(\Phi\) (x) denotes the standard normal distribution function, and \(a=c/\Phi '(c)\). In \(L_ h(t : \sigma)\) the first term represents the probability that \(I_ t\) does not cover \(\mu\), the second represents a cost of sampling and the third is included for normalization. In this paper, the author showed the lower bound for the average risk functional of a stopping time when the above loss function \(L_ h\) is given, i.e., \[ \liminf_{h\to 0}\inf_{t}\int^{\infty}_{0}E_{\sigma}L_ h(t : \sigma)d\pi (\sigma)\geq (1+c^ 2)/2 \] where \(\pi\) is a prior distribution satisfying some condition and the infimum extends over all stopping times (with respect to the sample variance sequence). From this result an asymptotic lower bound for the expected sample size of any sequential sampling plan with the specified confidence coefficient was obtained as \(h\to 0\).