Given is a a random sample $X_1,\ldots,X_n$ from a distribution with a pdf $f(x;\theta) = 2\theta^{-1} x^{(2/\theta)-1}$ for $0<x<1$, zero otherwise. We know that maximum likelihood estimator is $\hat{\theta} = -\dfrac 2 n \sum_i \log X_i$. Determine $c$ such that $c\left(\dfrac{n}{-2\sum_i \log X_i} - \dfrac{1}{\theta}\right) \xrightarrow{\text{d}} Z \sim N(0,1)$, i.e. converges in distribution.

So in the correction sheet it says the answer is $\dfrac{1}{\sqrt{CRLB}}$ where CRLB stands for the Cramer Rao Lower Bound. I don't really understand this. I have found in my book that for large $n$ $\hat{\theta}_n \sim N(\theta, CRLB)$, but how can we normalise this result to a $N(0,1)$?

  • $\begingroup$ If $X_n \to X$ in distribution, $a X_n + b \to a X + b$ in distribution. This is a consequence of slutsky's theorem, though easy enough to prove from first principles. Also, the asymptotics of the MLE are usually $\sqrt{n} (\hat{\theta}_{MLE} - \theta) \to N(0, \sigma_{\theta}^2)$ where $\sigma_\theta^2$ is the asymptotic variance, which is 1/ fisher information. So, check the statement in your book more carefully You'd expect a $\sqrt{n}$ normalizing factor for Gaussianity after the $c$ in the problem. $\endgroup$ – Batman Jun 8 '17 at 23:35

$X_1, X_2, \ldots, X_n$ is a sample from distribution with pdf $f(x; \theta) = 2\theta^{-1}x^{(2/\theta)-1}$ for $x \in (0,1)$ and zero otherwise.

Now, the distribution of $Y=-2\log X$ is derived as follows.


$$\frac{dX}{dY} = -\frac{1}{2}e^{-\frac{Y}{2}}$$

The pdf of $Y$ is given by,

$$g(y;\theta) = f(e^{-\frac{Y}{2}};\theta)\left | \frac{dX}{dY} \right | = \frac{1}{\theta}e^{-\frac{y}{\theta}}$$

Therefore, $Y_1, Y_2, \ldots, Y_n$ is transformed sample with $Y_i = -2\log X_i$ with $EY_i = \theta$ and $VarY_i = \theta^2$.

Now, one can use central limits theorem to get,

$$\sqrt{n}\frac{(\overline{Y_n} - \theta)}{\theta} \rightarrow \mathcal{N}(0,1)$$

Let $h(z) = \frac{1}{z}$. Using the first order taylor approximation to estimate the limiting distribution of $h(\overline{Y_n})$ (which is $\frac{1}{\hat{\theta}}$ as per your notation), we get,

$$h(\overline{Y_n}) \approx h(\theta) + h^{'}(\theta)(\overline{Y_n}-\theta)$$

$$\sqrt{n}\frac{(h(\overline{Y_n}) - h(\theta))}{|h^{'}(\theta)|\theta} \approx \sqrt{n}\frac{(\overline{Y_n} - \theta)}{\theta} \rightarrow \mathcal{N}(0,1)$$

$$\sqrt{n}\theta(h(\overline{Y_n}) - h(\theta)) = \sqrt{n}\theta(\frac{1}{\hat{\theta}} - \frac{1}{\theta})\rightarrow \mathcal{N}(0,1)$$

Note that I used $\frac{|h^{'}(\theta)|\theta}{\sqrt{n}}$ in the denominator because it is the standard deviation of $h(\overline{Y_n})$ which is the square root of the variance $\frac{h^{'}(\theta)^2\theta^2}{n}$. That's why || came up.

So the $c$ you are looking for is $\sqrt{n}\theta.$

Edit 1. If $X \sim \mathcal{N}(\mu,\sigma^2)$ then $\frac{X-\mu}{\sigma} \sim \mathcal{N}(0,1)$


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.