For $X_i \sim$ iid random variables:

For $1\le r_1 < ..<r_k \le n$ integers, I am trying to find the joint pdf of: $$ (X_{(r_1)},...,X_{(r_n)}) $$ where $X_{(r_1)}$ is the $r_1$th largest observation. I am wondering if anyone has seen the solution to this problem somewhere online? My current attempt:

Choosing $\epsilon$ small enough such that only one observation falls in an interval of width $\epsilon$

\begin{align*} &P(X_{(r_1)} \in (x_1 - \epsilon,x_1+\epsilon),.......,X_{(r_k)}\in (x_k - \epsilon,x_k+\epsilon))\\ &=P( n-k ~\text{of}~ X_1,....,X_n \in (-\infty, x_1-\epsilon),\\ &1 ~\text{of}~ X_1,....,X_n \in (x_1-\epsilon,x_1+\epsilon),\\ &....\\ &1 ~\text{of}~ X_1,....,X_n \in (x_k-\epsilon,x_k+\epsilon),)\\ &+\\ &P( n-k-1 ~\text{of}~ X_1,....,X_n \in (-\infty, x_1-\epsilon),\\ &1 ~\text{of}~ X_1,....,X_n \in (x_1-\epsilon,x_1+\epsilon),\\ &....\\ &1 ~\text{of}~ X_1,....,X_n \in (x_k-\epsilon,x_k+\epsilon),\\ &1 ~\text{of}~ X_1,....,X_n \in (x_k+\epsilon + \infty),\\ &+.... \end{align*} and so on and so forth accounting for all distributions of the remaining n-k observations between the area before $r_1$ and the area after $r_k$.

Each one of these is a multinomial, and has corresponding expressions in terms of the CDFs (dividing and taking epsilons to zero). After all the working, I get to the point:

$$ f_{(X_(r_1),....,X_(r_k)} (x_1,...,x_k)= $$ $$ n! \prod_{i=1}^{k} f(x_i) \left[ \frac{F(x_1)^{n-k}}{(n-k)!} +\frac{F(x_1)^{n-k-1} (1-F(x_k))}{(n-k-1)!} +......+\frac{ (1-F(x_k))^{n-k}}{(n-k)!} \right] $$

Not sure if I am on the right track, if anyone has seen this distribution before could you let me know if my attempt is so far correct?

  • 2
    $\begingroup$ i think you can refer to this question for (a) stats.stackexchange.com/questions/161145/… $\endgroup$
    – Deep North
    Aug 1, 2015 at 6:25
  • $\begingroup$ $=n!f(y_1)f(y_2)...f(y_r)\frac{[1-F(y_r)]^{n-r}}{(n-r)!}$, $y_1$ equal $X_{(r1)}$ for your case $\endgroup$
    – Deep North
    Aug 1, 2015 at 6:34
  • 2
    $\begingroup$ @DeepNorth the very last statement of this paper : google.com.au/… , has an expression for the joint pdf, I'm not sure if this agrees with the result you figured out in the question you linked ? $\endgroup$ Aug 1, 2015 at 11:10
  • $\begingroup$ @DeepNorth also, your answer in (a) assumes $r_1 = 1$, whereas what im trying to figure out doesn't necessarily have that? $\endgroup$ Aug 1, 2015 at 11:13
  • 1
    $\begingroup$ in this question two points are missing, first you did not tell us anything about $X_i,i=1,\ldots$ and their inter correlations and also type of variables. Under general situations that variables are iid, the answer is $f_{X_{(1)},\ldots,X_{(n)}}=n!\prod _1^n f_X(x_i)$ see here en.wikipedia.org/wiki/Order_statistic $\endgroup$
    – TPArrow
    Aug 1, 2015 at 11:51

1 Answer 1


Since$$f_{X_{(1)},\ldots,X_{(n)}}(x_1,\ldots,x_n)=n!\prod_{i=1}^n f_X(x_i)\mathbb{I}_{x_1\le x_2\le\ldots\le x_n}$$the marginal of $(X_{(r_1)},\ldots,X_{(r_k)})$ is obtained by integration (with some abuses of notation, see e.g. the integral bounds): \begin{align} f_{X_{(r_1)},\ldots,X_{(r_k)}}(x_{r_1},\ldots,x_{r_k}) &=\int f_{X_{(1)},\ldots,X_{(n)}}(x_1,\ldots,x_n)\mathbb{I}_{x_{r_1}\le x_{r_2}\le\ldots\le x_{r_k}}\,\prod_{i\notin\{r_1,\ldots,r_k\}}\text{d}x_i\\ &=\int n!\prod_{i=1}^n f_X(x_i)\mathbb{I}_{x_1\le x_2\le\ldots\le x_n}\prod_{i\notin\{r_1,\ldots,r_k\}}\text{d}x_i\\ &=n!\int \prod_{i=1}^{r_1-1} f_X(x_i)f_X(x_{r_1})\prod_{i=r_1+1}^{r_2-1} f_X(x_i)\cdots \\ &\qquad\cdots f_X(x_{r_k})\prod_{i=r_k+1}^{n}f_X(x_i)\mathbb{I}_{x_1\le x_2\le\ldots\le x_n}\prod_{i\notin\{r_1,\ldots,r_k\}}\text{d}x_i\\ &=n!\prod_{i=1}^{r_1-1}\int_{x_{i-1}}^{x_{i+1}} f_X(x_i)\text{d}x_i\,f_X(x_{r_1})\prod_{i=r_1+1}^{r_2-1}\int_{x_{i-1}}^{x_{i+1}}\, f_X(x_i)\text{d}x_i\,f_X(x_{r_2})\cdots\\ &\quad\cdots f_X(x_{r_k})\,\prod_{i=r_k+1}^{n}\int_{x_{i-1}}^{x_{i+1}} f_X(x_i)\text{d}x_i\mathbb{I}_{x_{r_1}\le x_{r_2}\le\ldots\le x_{r_k}}\\ &=n!\frac{F_X(x_{r_1})^{r_1-1}}{(r_1-1)!}\frac{[F_X(x_{r_2})-F_X(x_{r_1})]^{r_2-r_1-1}}{(r_2-r_1-1)!}\cdots\\ &\qquad\cdots\frac{[1-F_X(x_{r_k})]^{n-r_k-1}}{(n-r_k-1)!}\prod_{i=1}^k f_X(x_{r_i})\mathbb{I}_{x_{r_1}\le x_{r_2}\le\ldots\le x_{r_k}}\\ \end{align} which is the result produced in the reference (except for the use of $(x_1,...,x_k)$ as the argument of the density). The last integral follows from repeated integrations of $f_X(x)[F_X(x)-F_X(x_j)]^\alpha$; for instance, the first group of integrals leads to \begin{align*} \int_{x_{1}\le\ldots\le x_{r_1}}\prod_{i=1}^{r_1-1} f_X(x_i)\text{d}x_i &=\int_{x_2\le\ldots\le x_{r_1}}\prod_{i=2}^{r_1-1} f_X(x_i)\left\{\int_{-\infty}^{x_2} f_X(x_1)\right\}\text{d}x_1\prod_{i=2}^{r_1-1}\text{d}x_i\\ &=\int_{x_2\le\ldots\le x_{r_1}}\prod_{i=2}^{r_1-1}f_X(x_i)F_X(x_2)\text{d}x_i\\ &=\int_{x_3\le\ldots\le x_{r_1}}\prod_{i=2}^{r_1-1}f_X(x_i)\left\{\int_{-\infty}^{x_3} f_X(x_2)F_X(x_2)\text{d}x_2\right\}\prod_{i=3}^{r_1-1}\text{d}x_i\\ &=\int_{x_3\le\ldots\le x_{r_1}}\prod_{i=3}^{r_1-1}f_X(x_i)\frac{F_X(x_3)^2}{2!}\text{d}x_i\\ &=\ldots \end{align*}

  • $\begingroup$ one last issue I am having, when evaluating one of the expressions like: $\prod_{i=1}^{r_1-1}\int_{x_{i-1}}^{x_{i+1}} f_X(x_i)\text{d}x_i\,f_X(x_{r_1})$, would this not evaluate out to be: \begin{align} \prod_{i=1}^{r_1-1} \int_{x_{i-1}}^{x_{i+1}}f(x_i)dx_i &=\prod_{i=2}^{r_1-1} \int_{x_{i-1}}^{x_{i+1}}f(x_i)dx_i \left (\int_{-\infty}^{x_2}f(x_2)dx_2 \right)\\ &=\prod_{i=2}^{r_1-1} \int_{x_{i-1}}^{x_{i+1}}f(x_i)F(x_2)dx_i\\ &=\prod_{i=3}^{r_1-1} \int_{x_{i-1}}^{x_{i+1}}f(x_i) dx_i\left(\int_{x_1}^{x_3}f(x_2)F(x_2)dx_2 \right)\\ &=\cdots \end{align} $\endgroup$ Aug 3, 2015 at 3:02
  • $\begingroup$ As indicated in my reply, the notation $\int_{x_{i-1}}^{x_{i+1}}f$ is abusive. The correct presentation is the one of my second set of equation. In your first equation, the dummy variable of the inner integral should be $x_1$ not $x_2$. And the lower bound of the inner integral of the last equation should be $-\infty$, not $x_1$ which has already been integrated out. $\endgroup$
    – Xi'an
    Aug 3, 2015 at 6:21

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.