2
$\begingroup$

Problem: Find the UMVUE of $\frac1\theta$ for a random sample from the population distribution with density $$f(x;\theta)=\theta x^{\theta-1}$$ and show that its variance reaches the Cramér–Rao lower bound.

Progress: Knowing that the Cramér–Rao Lower Bound is $\frac{\gamma(\theta)^2}{nI(\theta)}$, where $\gamma(\theta)^2=\left(-\frac1{\theta^2}\right)^2=\frac1{\theta^4}$, I then compute the Fisher Information $I(\theta)$. This is where I get lost and become unsure about my work, mostly due to computation.

$$ \begin{align} I(\theta)&=\mathbb{E}\left(\frac{\partial\log{f(x;\theta)}}{\partial\theta}\right)^2\\ &=\int_0^1\left(\frac{\frac{\partial f(x;\theta)}{\partial \theta}}{f(x;\theta)}\right)^2f(x;\theta)\partial x \\ &=\int_0^1\frac{\left(\frac{\partial f(x;\theta)}{\partial \theta}\right)^2}{f(x;\theta)}\partial x \\ &=\int_0^1\left(\frac{1}{\theta}+\log{x})(x^{\theta-1}+(\theta^2-\theta)x^{\theta-2}\right)\partial x \\ &=\left[\log{x}\frac{x^\theta}{\theta}-\frac{x^{\theta-1}}{\theta-1}+\theta(\log{x})x^{\theta-1}\right]\biggr\rvert_0^1\\ &=\frac{-1}{\theta-1} \end{align} $$

I computed this and consistently got a negative answer, which doesn't make sense as plugging that into the CRLB formula yields a negative variance. I'd really appreciate if someone could quickly help me out on this!

Edit: I found out my mistake! It's a rather stupid one at the fourth equal sign: $$=\int_0^1\left(\frac{\partial\log{f(x;\theta)}}{\partial\theta}\right)\left(\frac{\partial f(x;\theta)}{\partial\theta}\right)$$ where I partially integrate the second term with respect to $x$ rather to to $\theta$.

$\endgroup$
2
  • 1
    $\begingroup$ You don't need to find the Cramér-Rao bound because the score function is of the form $\frac{\partial}{\partial\theta}\ln f(x;\theta)=k(\theta)\left(T(x)-\frac1{\theta}\right)$. This is the equality condition of Cramér-Rao equality. If $T$ is unbiased for $1/\theta$ then it must be UMVUE with its variance attaining the C-R bound. $\endgroup$ Apr 17 at 11:14
  • $\begingroup$ Hi @StubbornAtom, that's an interesting concept I think I either haven't learned or fully internalized yet. Are there any lecture notes or sources you can direct me to on this? $\endgroup$
    – mtcicero
    Apr 17 at 20:01
2
$\begingroup$

I can't follow your working after $\int_0^1\frac{\left(\frac{\partial f(x;\theta)}{\partial \theta}\right)^2}{f(x;\theta)}d x$, note that this expression is nonnegative as the function being integrated is nonnegative.

Let $u = 1 + \theta \log x, \frac{du}{dx}=\frac{\theta}x$,

\begin{align} &\int_0^1\frac{\left(\frac{\partial f(x;\theta)}{\partial \theta}\right)^2}{f(x;\theta)}d x \\ &=\int_0^1 \frac{\left( x^{\theta-1}+\theta x^{\theta -1} \log x\right)^2}{\theta x^{\theta -1}}\, dx \\ &=\frac1{\theta}\int_0^1 x^{\theta-1}(1+\theta \log x)^2 \, dx \\ &= \frac1{\theta} \int_{-\infty}^1 x^{\theta-1}u^2 \frac{x}{\theta}\, du \\ &= \frac1{\theta^2} \int_{-\infty}^1 x^\theta u^2 \, du\\ &= \frac1{\theta^2} \int_{-\infty}^1 e^{u-1}u^2 \, du\\ &= \frac1{\theta^2} \left[ e^{u-1}u^2|_{-\infty}^1 - 2\int_{-\infty}^1 ue^{u-1}\, du \right]\\ &=\frac1{\theta^2} \left[ 1-2\left[ e^{u-1}u|_{-\infty}^1-\int_{-\infty}^1 e^{u-1}\, du\right]\right]\\ &= \frac1{\theta^2} \ge 0 \end{align}

$\endgroup$

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.