3
$\begingroup$

I am trying to solve Robert Hogg's mathematical anaysis 6th exercise 6.2.11. The problem says. Let $\bar{X}$ be the mean of a random sample of size n from a $N(\theta,\sigma^2)$ distribution, $-\infty<\theta<\infty$, $\sigma^2>0$. Assume that $\sigma^2$ is known. Show that $\bar{X}-\frac{\sigma^2}{n}$ is an unbiased estimator of $\theta^2$ and find its efficiency.

The first part pretty easy, the efficiency part sames very conter-intuitive.

My solution:

Since $\bar{X}$ be the mean of a random sample of size n from a $N(\theta,\sigma^2)$ by central limit theorem we know that $Var(\bar{X})=E(\bar{X}^2)-[E(\bar{X})]^2=\frac{\sigma^2}{n}$ and $E(\bar{X})=\theta$,.

Therefore, $E(\bar{X}^2)=\frac{\sigma^2}{n}+\theta^2$.

$E(\bar{X}^2-\frac{\sigma^2}{n})=E(\bar{X}^2)-\frac{\sigma^2}{n}=\frac{\sigma^2}{n}+\theta^2-\frac{\sigma^2}{n}=\theta^2$

Therefore, $\bar{X}^2-\frac{\sigma^2}{n}$ is an unbiased estimator of $\theta^2$

For the second part, first I need to calculate the variance of $\bar{X}^2-\frac{\sigma^2}{n}$

$Var(\bar{X}^2-\frac{\sigma^2}{n})=Var(\bar{X}^2)$

By CTL we know that $\frac{\bar{X}-\theta}{\frac{\sigma}{\sqrt{n}}}$ has a N(0,1) distribution. Further the square of standard normal has a $\chi^2(1)$ distribution.

Therefore,$ Var((\frac{\bar{X}-\theta}{\frac{\sigma}{\sqrt{n}}})^2)=1/2*4=2$, i.e variance of $\chi^2(1)=2$

$\therefore Var(\frac{\bar{X}^2-2\theta\bar{X}+\theta^2}{\sigma^2/n})=2$

Solve this equation:

$\frac{n^2}{\sigma^4}Var(\bar{X}^2)-\frac{4\theta^2n^2}{\sigma^4}Var(\bar{X})+0=2$, we already know what is $Var(\bar{X})$

$\therefore Var(\bar{X}^2)=\frac{2\sigma^4}{n^2}+\frac{4\theta^2\sigma^4}{n}$

By this step, I get the variance of the $\bar{X}^2-\frac{\sigma^2}{n}$ which is $\frac{2\sigma^4}{n^2}+\frac{4\theta^2\sigma^4}{n}$

Since $\bar{X}^2-\frac{\sigma^2}{n}$ is an unbiased estimator of $\theta^2$, we need to find the fisher information of $\theta^2$.

We first calculate the $I(\theta)$ for $\bar{X}$

$f(\bar{X};\theta)=\frac{1}{\sqrt{2\pi}\frac{\sigma}{n}}exp(-\frac{1}{2}\frac{(\bar{X}-\theta)^2}{\sigma^2/n})$

$logf(\bar{X};\theta)=log\frac{1}{\sqrt{2\pi}\frac{\sigma}{n}}-\frac{1}{2}\frac{(\bar{X}-\theta)^2}{\sigma^2/n}$

$\frac{\partial logf(\bar{X};\theta)}{\partial \theta}=\frac{n}{\sigma^2}(\bar{X}-\theta)$

Therefore, $I(\theta)=Var(\frac{\partial logf(\bar{X};\theta)}{\partial \theta})=Var(\frac{n}{\sigma^2}(\bar{X}-\theta))=Var(\frac{n}{\sigma^2}\bar{X})=\frac{n^2}{\sigma^4}Var(\bar{X})=\frac{n^2}{\sigma^4}.\frac{\sigma^2}{n}=\frac{n}{\sigma^2}$

Next we will use formular $I(g(\theta))=\frac{I(\theta)}{(g'(\theta))^2}$ to calculate $I(\theta^2)$

$I(\theta^2)=\frac{I(\theta)}{((\theta^2)')^2}=\frac{\frac{n}{\sigma^2}}{4\theta^2}=\frac{n}{4\theta^2\sigma^2}$

Efficiency=$\frac{\frac{1}{nI(\theta^2)}}{Var(\bar{X}^2-\frac{\sigma^2}{n})}=\frac{4\theta^2\sigma^2}{n^2}.\frac{n^2}{2\sigma^4+4\theta^2\sigma^2n}=\frac{1}{\frac{2\sigma^2}{4\theta^2}+n}$

We can see when sample becomes bigger efficiency becomes smaller, this is very counter-intuitive to me.

Please help me to find which part is wrong for the calculation.

Thank you very much.

$\endgroup$
1
  • $\begingroup$ Maybe you've N(theta,sigma-squared) distribution, so please edit it $\endgroup$ May 3, 2015 at 9:01

1 Answer 1

1
$\begingroup$

Thanks you guy's responses. I found where the problem was.

I should use $f(X;\theta)$ to calculate the $I(\theta)$, but not $f(\bar{X};\theta)$

Now $I(\theta)=\frac{1}{\sigma^2}$ and $I(\theta^2)=\frac{1}{4\theta^2\sigma^2}$

So Efficiency=$\frac{\frac{1}{nI(\theta)}}{Var(\bar{X}^2+\frac{\sigma^2}{n})}=\frac{4\theta^2\sigma^2}{n}.\frac{n^2}{2\sigma^4+4\theta^2\sigma^2n}=\frac{n}{\frac{\sigma^2}{2\theta^2}+n}=\frac{1}{1+\frac{\sigma^2}{2n\theta^2}}$

When n becomes bigger efficiency close to 1.

Now I think the solution is correct.

$\endgroup$

Your Answer

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

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