Consider for given mean $m\in \mathbb R$ the Gaussian product model $(\mathbb R^n, \mathcal{B}(\mathbb R^n),\mathcal{N}_{m,\theta}:\vartheta >0)$ where $\vartheta$ denotes the variance, and a statistic $T=\sqrt\frac{\pi}{2} \frac 1n \sum_{i=1}^n \mid X_i-m\mid$. Show that the variance of $T$ does not reach the Cramer-Rao-bound $G'(\vartheta)^2/I(\mu_\vartheta)$, where $G(\vartheta)=\sqrt\vartheta$ and $\mu_\vartheta=\cal N^{\otimes n}_{m,\vartheta}$

I have shown that $T$ is unbiased for $G(\vartheta)$. I am struggling to calculate the Fisher information, i.e.

$$I(\mu_\vartheta)=\operatorname{Var}\bigr(\frac{\partial}{\partial\vartheta}\log(\frac{1}{\sqrt {2\pi\theta}}\exp(\frac{(-x-m)^2)}{2\theta}) \bigl)\\=\operatorname{Var}\bigr( -\frac{\sqrt{2\pi}}{2\sqrt\theta}\frac{1}{\sqrt{2\pi\theta}}+\frac{(x-m)^2}{2\theta^2}\bigl)\\=\operatorname{Var}\bigr(\frac{-1}{2\theta}+\frac{(x-m)^2}{2\theta^2}\bigl)\qquad\qquad$$ I think $(x-m)^2$ is $\chi^2$ distributied, so my final result would be simply $$\operatorname{Var}\bigr(\frac{-1}{2\theta}+\frac{(x-m)^2}{2\theta^2}\bigl)=\operatorname{Var}\bigr(\frac{(x-m)^2}{2\theta^2} \bigl)=\frac{1}{2\theta^4}$$ is this reasoning correct?


Indeed, when $X\sim{\cal N}(m,\theta)$, $$Z=\dfrac{(X-m)^2}{\theta}\sim\chi^ 2_1$$which implies$$\mathbb{E}[Z^{1/2}]=\sqrt{2/\pi}\quad\mathbb{E}[Z]=1\quad\text{var}(Z)=2$$and hence $$\text{var}(Z/2\theta^2)=1/2\theta^2$$

  • 2
    $\begingroup$ yes I see.. perfect, I forgot to rescale the distribution, that's why I have had $1/2\theta^4$ instead of $1/ 2\theta^2$ $\endgroup$ – Joey Doey May 28 '18 at 10:03

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.