Let $X_1,\ldots,X_n$ be a random sample from $N(\mu,\sigma^2)$ with both parameters unknown. How can I show that $(\bar{X}, S^2)$ is independent of $(X_{(n)}-\bar{X})/S$?


I am quite confident that I should use Basu's Theorem here. I can see that $(\bar{X}, S^2)$ is a complete sufficient statistic, but I am having trouble showing it. I assume that I will need to write $N(\mu,\sigma^2)$ as an exponential family to show this statistic is complete sufficient (for use in Basu's Theorem).

$$f(x\mid \mu,\sigma)=\frac{1}{\sqrt{2\pi\sigma^2}}\exp\left(\frac{-1}{2\sigma^2}(x-\mu)^2\right)=\frac{1}{\sqrt{2\pi\sigma^2}}\exp\left(\frac{-\mu^2}{2\sigma^2}\right)\exp\left(\frac{-x^2}{2\sigma^2} + \frac{x\mu}{\sigma^2}\right)$$

With this work, $T(\mathbf{X})=(\sum x_i, \sum x_i^2)$ is a complete statistic for $\mu, \sigma$, which is not $T(\mathbf{X})=(\bar{X}, S^2)$. I see that it closely resembles what I need, so how can I continue this work?

Additionally, we would need to show that $Y=(X_{(n)}-\bar{X})/S$ is ancillary, meaning that $Y$'s distribution is free of both $\mu$ and $\sigma$. However, I am not too sure on how to show this. Is this a case of applying a location-scale transformation?

  • 1
    $\begingroup$ One approach: what happens to the distribution of $(X_{(n)}-\bar X)/S$ when you change $\mu$ to $\mu + \lambda$? When you change $\sigma^2$ to $\lambda^2\sigma^2$? $\endgroup$
    – whuber
    Feb 6, 2020 at 5:15
  • 2
    $\begingroup$ How about you write $Y$ as $Y=\frac{\frac{X_{(n)}-\mu}{\sigma}-\frac{\bar X-\mu}{\sigma}}{S/\sigma}$? $\endgroup$ Feb 6, 2020 at 7:11
  • $\begingroup$ @StubbornAtom so you're approach is suggesting a scale transformation, whereas whuber's is suggesting a location-scale, correct? Is it typical to use location-scale transformations when showing a statistic is ancillary? It seems like the "direct" method of finding the joint pdf of these three variables and doing several transformations is not as efficient. So, I will attempt your scale transformation and get back to you. Thanks! $\endgroup$
    – Ron Snow
    Feb 6, 2020 at 17:20
  • $\begingroup$ @whuber if you make those suggested changes, then we have a location-scale transformation. From there, the typical line of attack is setting $Y_i = \lambda x_i + \lambda$, right? $\endgroup$
    – Ron Snow
    Feb 6, 2020 at 17:22

1 Answer 1


To expand on my comment, you can write $Y$ as $$Y=\frac{\frac{X_{(n)}-\mu}{\sigma}-\frac{\overline X-\mu}{\sigma}}{S/\sigma}$$

From the distributions of $\overline X$ and $S$ it should be clear that those of $\frac{\overline X-\mu}{\sigma}$ and $S/\sigma$ are free of $(\mu,\sigma^2)$. As for the order statistic, notice that $\frac{X_{(n)}-\mu}{\sigma}$ has the same distribution as $Z_{(n)}$ where $Z_1,\ldots,Z_n$ are i.i.d standard normal. Hence its distribution is also free of $(\mu,\sigma^2)$. You can show this from the distribution function if you want.

This shows $Y$ is ancillary for $(\mu,\sigma^2)$. Note that @whuber's suggestion is along the same lines.

And the pdf of $X_1,X_2,\ldots,X_n$ is indeed a member of a full-rank (regular) exponential family, which shows that $\left(\sum\limits_{i=1}^n X_i,\sum\limits_{i=1}^n X_i^2\right)$ and equivalently its one-to-one transform $(\overline X,S^2)$ is a complete sufficient statistic for $(\mu,\sigma^2)$.

  • $\begingroup$ Thank you for providing this detail. I understand the ancillary statistics for sure. I have two remaining questions for you. How can I show that $S^2$ is a one-to-one transformation of $\sum ^n_{i=1}X_i^2$? Also, how did you identify how to rewrite $Y$? $\endgroup$
    – Ron Snow
    Feb 7, 2020 at 1:40
  • 1
    $\begingroup$ $S^2=\frac1n\sum X_i^2-\overline X^2$ (does not matter if divisor is $n-1$) is a function of $\sum X_i^2$ and $\sum X_i$, making the pair $T_2=(\overline X,S^2)$ a function of $T_1=(\sum X_i,\sum X_i^2)$. The transformation $T_1\mapsto T_2$ is clearly one-to-one because there is no loss of information going from one pair to the other. You can also use the Factorization theorem to directly get $T_2$ instead of $T_1$. As for rewriting $Y$, it is just a location-scale transformation as you mentioned. $\endgroup$ Feb 9, 2020 at 15:50
  • $\begingroup$ Would it also be possible to let $X_i=\sigma Z_i + \mu$, so $(X_{(n)}-\bar{X})/S=\frac{\sqrt{n-1}(Z_{(n)}-\bar{Z})}{\sqrt{(n-1)S^2/\sigma^2}}$, whose distribution doesn't really on $(\mu,\sigma^2)$? $\endgroup$
    – Ron Snow
    Feb 12, 2020 at 2:54

Your Answer

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

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