Large deviations of estimators (Q1082733): Difference between revisions

The authors investigate the asymptotic behaviour of estimators \(T_ n\) of g(\(\theta)\) using its inaccuracy rate \[ e(\epsilon,\theta,T_ n)=- \liminf_{n\to \infty}n^{-1} \log P_{\theta}\{\| T_ n- g(\theta)\| >\epsilon \} \] for fixed \(\epsilon >0\). In convex exponential families of distributions, it is shown that, for the maximum likelihood estimator \({\hat \theta}_ n\) of \(\theta\), g(\({\hat \theta}_ n)\) is inaccuracy rate optimal in the sense of g(\({\hat \theta}_ n)\) attaining the upper bound of \(e(\epsilon,\theta,T_ n)\) for consistent estimators of g(\(\theta)\). In the case when a family of distributions is not exponentially convex, inaccuracy rate optimal estimators do not usually exist. The optimality w.r.t. the inaccuracy rate is based on propositions by \textit{R. R. Bahadur}, \textit{S. L. Zabell} and \textit{J. C. Gupta} [Asymptotic theory of statistical tests and estimation, Proc. Int. Symp., Chapel Hill/NC 1979, 33-64 (1980; Zbl 0601.62037)] and \textit{R. R. Bahadur} [Recent advances in statistics, Pap. in Honor of H. Chernoff, 273-286 (1983; Zbl 0543.62064)]. In location parameter cases, the inaccuracy rate of M-estimators is determined, and an optimal M-estimator is given in the class of equivariant estimators. Tail-behaviour of location estimators is also discussed.