Asymptotics for configural location estimators (Q1082736): Difference between revisions

Let \(x_ 1,...,x_ n\) be independent real-valued observations from the distribution function (d.f.) F(x-\(\mu)\); it is assumed that F(\(\cdot)\) is symmetric around 0, \(0<F(x)<1\) for all \(x\in {\mathbb{R}}\), and has a probability density function f(\(\cdot)\) with respect to Lebesgue measure. When sampling from a symmetric situation as above, the Pitman estimator of \(\mu\) is given by \[ T_ F(x_ 1,...,x_ n)=\int^{\infty}_{- \infty}r\prod^{n}_{i=1}f(x_ i+r)dr/\int^{\infty}_{- \infty}\prod^{n}_{i=1}f(x_ i+r)dr. \] This estimator has the smallest mean-square-error among all location equivariant estimators, where T(\(\cdot)\) is location equivariant, if \(T(x_ 1+r,...,x_ n+r)=T(x_ 1,...,x_ n)+r\), \(r\in {\mathbb{R}}.\) In this paper, the main interest is centered at the behavior of the so- called bioptimal or compromise location estimators when the sample is coming from a d.f. G(\(\cdot)\) symmetric around 0. With d.f.'s \(F_ 1(\cdot)\) and \(F_ 2(\cdot)\) satisfying the same assumptions as the d.f. F(\(\cdot)\) above, a compromise estimator \(T_{F_ 1,F_ 2,\alpha}(x_ 1,...,x_ n)\) is defined by \[ T_{F_ 1,F_ 2,\alpha}(x_ 1,...,x_ n)= \] \[ \alpha w_{F_ 1}(x_ 1,...,x_ n)T_{F_ 1}(x_ 1,...,x_ n)+(1-\alpha)w_{F_ 2}(x_ 1,...,x_ n)T_ F(x_ 1,...,x_ n), \] where \(0<\alpha <1\); for \(k=1,2\), \(T_{F_ k}(x_ 1,...,x_ n)\) are given as \(T_ F(x_ 1,...,x_ n)\) above, and the weights \(w_{F_ k}(x_ 1,...,x_ n)\) are given by \[ w_{F_ k}(x_ 1,...,x_ n)= \] \[ \int^{\infty}_{- \infty}\prod^{n}_{i=1}f_ k(x_ i+r)dr/\int^{\infty}_{- \infty}\{\alpha \prod^{n}_{i=1}f_ 1(x_ i+r)+(1- \alpha)\prod^{n}_{i=1}f_ 2(x_ i+r)\}dr, \] k\(=1,2\). In many applications, \(F_ 1(\cdot)\) is chosen to be of the Gaussian shape, and, on the basis of various considerations, \(F_ 2(\cdot)\) is chosen as a heavy tailed distribution; for example, a distribution with Pareto tails. Under certain overly nice conditions, which do not include, for example, distributions with Pareto tails, the asymptotic behavior of the weights is established; namely, it is shown that \(w_{F_ 2}(x_ 1,...,x_ n)/w_{F_ 1}(x_ 1,...,x_ n)\) converges to 0 a.s. The same result is also proved under weaker conditions. It follows that the compromise estimator is equivalent to the Pitman estimator. Furthermore, expressions for the asymptotic variances of \(T_{F_ 1}(x_ 1,...,x_ n)\) and \(T_{F_ 2}(x_ 1,...,x_ n)\) are given. Finally, if the d.f. G(\(\cdot)\) is chosen to be \(G(x)=(1-\epsilon)\Phi (x)+\epsilon H(x)\), where \(\Phi\) (\(\cdot)\) is the d.f. of the standard normal distribution, \(H(x)+H(-x)=1\) for all x's and H(\(\cdot)\) puts all its mass outside the interval [-k,k], but is otherwise arbitrary, and \(0\leq \epsilon \leq 0.5\), it is shown that the asymptotic variance of the compromise estimator is \(\leq\) the asymptotic variance of Huber's minimax estimator.