9
$\begingroup$

I would like to know if the statistic $$T(X_1,\ldots,X_n)=\frac{\sum_{i=1}^n (X_i-\bar{X}_n)^2}{n-1}$$ is complete for $\sigma^2$ in a $N(\mu,\sigma^2)$ setting.

Does this depend on whether $\mu$ is previously known or not? If $T$ is complete for $\sigma^2$, then by Lehmann-Scheffé it is UMVUE. But if $\mu$ were known, we could have considered $$W(X_1,\ldots,X_n)=\frac{\sum_{i=1}^n (X_i-\mu)^2}{n},$$whose variance equals the Cramer-Rao bound $2\sigma^4/n$, and is strictly less than $2\sigma^4/(n-1)=\text{Var}[T]$, so $T$ could not be UMVUE.

$\endgroup$
  • $\begingroup$ Maybe you would agree with me that T is not unbiased when mu is known. $\endgroup$ – Michael R. Chernick Dec 11 '16 at 15:28
  • 1
    $\begingroup$ @MichaelChernick Don't you have in general that $E[T]=\sigma^2$? $\endgroup$ – user39756 Dec 11 '16 at 15:36
  • $\begingroup$ Sorry you are right T has the sample average used in the formula. I was thinking of W. $\endgroup$ – Michael R. Chernick Dec 11 '16 at 15:47
  • 1
    $\begingroup$ Hint: Did you check to see if $T$ is sufficient in the case where $\mu$ is known? $\endgroup$ – cardinal Dec 12 '16 at 3:34
5
$\begingroup$

I think I solved my own question. Comments about this answer and new answers are welcome.

If $x_1,\ldots,x_n$ are observations in a $N(\mu,\sigma^2)$ population and $\mu$ is unknown, then $$f(x_1,\ldots,x_n|\mu,\sigma^2)=\left(\frac{1}{\sqrt{2\pi\sigma^2}}\right)^ne^{-\frac{n\mu^2}{2\sigma^2}}e^{\frac{\mu}{\sigma^2}\sum_{i=1}^nx_i-\frac{1}{2\sigma^2}\sum_{i=1}^n x_i^2}$$ (this shows that the normal family is a exponential family). As the image of the map $$(\mu,\sigma^2)\in \mathbb{R}\times\mathbb{R}^+\mapsto (\frac{\mu}{\sigma^2},-\frac{1}{2\sigma^2})$$ contains an open set of $\mathbb{R}^2$, by a theorem (for instance, see page 6 here), the statistic $U=(\sum_{i=1}^n X_i,\sum_{i=1}^n X_i^2)$ is sufficient and complete for $(\mu,\sigma^2)$. As $T$ is a function of $U$ and is centered for $\sigma^2$, by Lehmann-Scheffé $T$ is UMVUE for $\sigma^2$.

Now, if $\mu=\mu_0$ is known, $\mu$ does not belong to the parametric space anymore, therefore the "new" density function is $$f(x_1,\ldots,x_n|\sigma^2)=\left(\frac{1}{\sqrt{2\pi\sigma^2}}\right)^ne^{-\frac{1}{2\sigma^2}\sum_{i=1}^n(x_i-\mu_0)^2}$$(we have a new exponential family). As the image of the map $$\sigma^2\in\mathbb{R}^+\mapsto -\frac{1}{2\sigma^2}$$ contains an open subset of $\mathbb{R}$, our statistic $W$ is sufficient and complete for $\sigma^2$. Since it is in addition centered, $W$ is UMVUE for $\sigma^2$ by Lehmann-Scheffé.

| cite | improve this answer | |
$\endgroup$
1
$\begingroup$

In your statistic $T(X_{1},\ldots,X_{n})$, $\bar{X}$ is used as the estimate of $\mu$. If you know the true value of $\mu$, then the estimator of the variance $W(X_{1},\ldots,X_{n})$ is preferable. $W$ is unbiased and has a lower variance than $T$. Thus, in the setting where $\mu$ is known, use $W$.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ I know that $W$ is preferable, but I would like to understand why $T$ is not complete when $\mu$ is known. $\endgroup$ – user39756 Dec 11 '16 at 16:09
  • 1
    $\begingroup$ I have been searching the net refreshing my memory on the Lehmann-Scheffe Theorem and Rao-Blackwell. I agree with the OP that the answer might be that T is not a complete sufficient statistic when mu is known. I think in may be that the parameter space changes. $\endgroup$ – Michael R. Chernick Dec 11 '16 at 16:44
1
$\begingroup$

W(X1,…,Xn) is not an unbiased estimator, so it is not a UMVUE. T actually is the UMVUE because it is a complete sufficient statistic and also an unbiased estimator for $\sigma^2$. Hence according to Lehmann-Scheffé, it is the UMVUE.

| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.