I have found the maximum likelihood estimator $\hat{\sigma}$ of a iid r.vs $X_1, ..., X_n$ which all have normal distribution with known mean $\mu$ and unknown variance $\sigma^2$.

So $\hat{\sigma}$ turns out to be $\sqrt{\frac{1}{n} \displaystyle\sum_{i=1}^n (X_i - \mu)^2} $. Now if it wasn't for the square root sign I'd have no problem working out $\mathbb{E}[\hat{\sigma}]$. Could someone please help?


The random variable $Q:=\sum_{i=1}^n {\left(\frac{X_i - \mu}{\sigma}\right)}^2$ has a Chi-squared distribution with $n$ degrees of freedom. Denote by $f$ its pdf, which you can find on Wikipedia or many other places. It is given by $$ f(x) = C x^{\frac{n}{2}-1}\exp\bigl(-\frac{x}{2}\bigr) $$ where $C=\frac{1}{2^{\frac{n}{2}}\Gamma\bigl(\frac{n}{2}\bigr)}$ is a constant (not depending on $x$).

Your are looking for the expectation of the random variable $$ R= \sqrt{\frac{1}{n} \displaystyle\sum_{i=1}^n (X_i - \mu)^2}. $$ Once you get the pdf of $R$, say $g$, you can get the expectation of $R$ by calculating $E(R) = \int yg(y)\mathrm{d}y$.

The random variable $R$ is a function of $Q$, namely $R=h(Q)$ where $h(x)=\sigma \sqrt{\frac{1}{n}x}$. This function is one-to-one map from $[0, \infty[$ to $[0, \infty[$, therefore you can use the change of variables formula to get the pdf of $R$. Denoting by $g$ this pdf, it is given by $$ g(y) = {(h^{-1})}'(y)\times f\bigl(h^{-1}(y)\bigr). $$ The inverse of $h$ is $h^{-1}(y) = \frac{n}{\sigma^2}y^2$. Setting $\lambda=\frac{n}{\sigma^2}$ for notational simplicity, one gets (for $y > 0$) $$ \begin{align*} g(y) & = 2 \lambda y \times f\bigl(\lambda y^2\bigr) \\ & = 2 C\lambda y^{n-1} \exp \bigl(-\frac{\lambda y^2}{2}\bigr) \end{align*}, $$ and one finally has to calculate $$ E(R) = 2 C\lambda^{\frac{n}{2}} \int_0^\infty y^{n} \exp \bigl(-\frac{\lambda y^2}{2}\bigr)\mathrm{d}y. $$

By the change of variables $x=\alpha y^2$, $$ \begin{align*} \int_0^\infty y^{n} \exp \bigl(-\alpha y^2\bigr)\mathrm{d}y & = \frac{1}{2} \alpha^{-\frac{n+1}{2}} \int_0^\infty x^{\frac{n}{2}-1} \exp \bigl(-x\bigr)\mathrm{d}x \\ & = \frac{1}{2} \alpha^{-\frac{n+1}{2}} \Gamma\Bigl(\frac{n+1}{2} \Bigr). \end{align*} $$ Thus one finally gets $$ \begin{align*} E(R) & = \frac{1}{2^{\frac{n}{2}}\Gamma\bigl(\frac{n}{2}\bigr)} {\left(\frac{n}{\sigma^2}\right)}^{\frac{n}{2}} {\left(\frac{1}{2}\frac{n}{\sigma^2}\right)}^{-\frac{n+1}{2}}\Gamma\Bigl(\frac{n+1}{2} \Bigr) \\ & = \frac{\sqrt{2}\sigma}{\sqrt{n}}\frac{\Gamma\Bigl(\frac{n+1}{2} \Bigr)}{\Gamma\bigl(\frac{n}{2}\bigr)}. \end{align*} $$


> n <- 5
> sigma <- 2
> Q <- rchisq(10000, n)
> h <- function(x) sigma*sqrt(x/n)
> R <- h(Q)
> mean(R)
[1] 1.903887
> sqrt(2)*sigma/sqrt(n)*gamma((n+1)/2)/gamma(n/2)
[1] 1.903066

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.