Let $\{X_i : i = 1, 2, \dots ,n\}$ be independent random variables with finite second moments. They are not necessarily identically distributed. They do have the same mean,$\mu=E(X_i)$ for all i, but possibly different variances, $\sigma^2_i > 0$. Assume these variances are known. Let $p$ be the estimator that solves the problem $\min \Sigma[(X_i-m)^2/\sigma^2_i]$ with respect to $m$. Each squared term is weighted by the inverse of the variance. Show that $p = \Sigma(w_i X_i)$ for weights $w_i > 0$ such that $\Sigma w_i = 1$.

I know we have must find the weights. I get the weights as follows: $w_i = {(\Sigma[1/\sigma^2_i])}^{-1} / \sigma^2_i$.

Is this right?


  • $\begingroup$ Please see stats.stackexchange.com/questions/243922/…. $\endgroup$
    – whuber
    Nov 7, 2018 at 20:00
  • $\begingroup$ I did see. While it is similar, the other question relates to estimate of variance. I just want to know whether my answer is right $\endgroup$ Nov 8, 2018 at 14:54
  • $\begingroup$ Also..,how would you prove that these weights are optimal. I know we have to set up the langrangian and differentiate wrt to each wi. But I am stuck at the simplification part $\endgroup$ Nov 9, 2018 at 15:32
  • $\begingroup$ You can use elementary inequalities to demonstrate optimality, or even Euclidean geometry. I give the geometric argument at stats.stackexchange.com/a/9073/919. $\endgroup$
    – whuber
    Nov 9, 2018 at 15:46


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.