2
$\begingroup$

I'm having trouble understanding example 6.2.5 of Casella and Berger:

Let $X_1,\ldots, X_n$ be iid from a pdf $f$, where we are unable to specify any more information about the pdf. It then follows that the sample density is given by $$f(\mathbf{x}) = \prod_{i=1}^n f(x_i) = \prod_{i=1}^n f(x_{(i)}), $$ where $x_{(1)} \leq \ldots\leq x_{(n)}$ are the order statistics. By Theorem 6.2.2, we can show that the order statistics are a sufficient statistic.

So the identity for the density should hold when $X_1, X_2 \sim Unif(0,1)$ are iid, but then $f(\mathbf{x}) = f(x_1)\cdot f(x_2) = 1$ while $$\prod_{i=1}^n f(x_{(i)}) = f(x_{(1)})\cdot f(x_{(2)}) = 2(1-x_{(1)})\cdot 2x_{(2)} $$ since $X_{(j)} \sim Beta(j, n-j+1)$.

Is there an error in the book or am I misunderstanding something?

$\endgroup$
5
$\begingroup$

Your confusion stems from the loose notations I presume: when writing $$f(\mathbf{x}) = \prod_{i=1}^n f(x_i) = \prod_{i=1}^n f(x_{(i)}),$$ George Casella and Roger Berger first use $f$ for the density of the vector, $f(\mathbf{x})$, then second as the density of the components, $f(x_i)$, and then again as the density of the components, $f(x_i)$, rather than as the densities of the order statistics. Distinguishing between all three as in \begin{align*}\mathbf X &\sim f_{1:n}(\mathbf x) \\ X_i &\sim f_1(x_i)\quad i=1,\ldots,n\\ X_{(i)} &\sim f_{(i)}(x_{(i)}) \end{align*} the identity looks like $$\underbrace{f_{1:n}(\mathbf{x})}_{\text{joint}} = \prod_{i=1}^n \underbrace{f_1(x_i)}_{\text{individual}} = \overbrace{\prod_{i=1}^n \underbrace{f_1(x_{(i)})}_{\text{individual}}}^{\text{reordering}\\\text{by rank}} \ne \overbrace{\prod_{i=1}^n\underbrace{f_{(i)}(x_{(i)})}_{\text{ordered}}}^{\text{incorrect due}\\\text{to dependence}},$$

$\endgroup$
  • $\begingroup$ Ooh I see. And the joint distribution $q(T(x)|\theta)$ of the order statistics is $n! \prod f_1(x_{(i)}) = n! \prod f_1(x_{i})$ so that $f(x|\theta) / q(T(x)|\theta) = 1/n!$ which is independent of $\theta$! Thank you! $\endgroup$ – Alain Sep 26 '17 at 5:57

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.