Consider 3 iid samples drawn from the uniform distribution $u(\theta, 2\theta)$, where $\theta$ is parameter. I want to find $$ \mathbb{E}\left[X_{(2)}| X_{(1)}, X_{(3)}\right] $$ where $X_{(i)}$ is order statistic $i$.

I would expect the result to be $$ \mathbb{E}\left[X_{(2)}| X_{(1)}, X_{(3)}\right] = \frac{X_{(1)}+ X_{(3)}}{2} $$ But the only way I can show this result seems to be too lengthy, I cannot come up with simple solution, am I missing something, is there some shortcut?

What I do is the following:

  • I find the conditional density

    $$ f(x_{(2)}| x_{(1)}, x_{(3)}) = \frac{ f(x_{(1)}, x_{(2)}, x_{(3)})}{f(x_{(1)}, x_{(3)})} $$

  • I integrate

$$ \mathbb{E}\left[X_{(2)}| X_{(1)}, X_{(3)}\right] = \int x f(x| x_{(1)}, x_{(3)}) dx $$


I adopt general formula for density of order statistic (with $\mathbb{I}_{\{A\}}$ an indicator of set $A$)

$$ f_{x_{(1)},\ldots , x_{(n)}}(x_1,\cdots, x_n) = n! \prod_{i=1}^n f_{x}(x_i)\mathbb{I}_{\{x_{(1)} \leq x_{(2)} \leq \cdots \leq x_{(n)}\}}(x_1,\cdots, x_n) $$

to obtain for my case

$$ f_{x_{(1)}, x_{(2)}, x_{(3)}}(x_1, x_2, x_3) = 3!\frac{1}{\theta^3}\mathbb{I}_{\{x_1 \leq x_2 \leq \cdots \leq x_n\}}(x_1,\cdots, x_3) $$

marginal of $f_{x_{(1)}, x_{(3)}}(u, v)$ is

$$f_{x_{(1)}, x_{(3)}}(u, v) = \int f_{x_{(1)}, x_{(2)}, x_{(3)}}(u, x_2, v) dx_2$$

that is

$$ f_{x_{(1)}, x_{(3)}}(u, v) = \int 3!\frac{1}{\theta^3}\mathbb{I}_{\{x_1 = u \leq x_2 \leq x_3 = v\}}(u, x, v) dx = 3!\frac{1}{\theta^3} [v-u] $$


$$ f(x_{(2)}| x_{(2)} = u, x_{(3)} = v) = \frac{ f(x_{(1)} = u, x_{(2)}, x_{(3)} = v)}{f(x_{(1)}= u, x_{(3)} = v)} = \frac{3!\frac{1}{\theta^3}\mathbb{I}_{u \leq x_2 \leq \cdots \leq v}(u,x_2, v) }{3!\frac{1}{\theta^3} [v-u]}= [v-u]^{-1}\mathbb{I}_{\{u<x_2<v\}} $$

which gives

$$ \mathbb{E}\left[X_{(2)}| X_{(1)} = u, X_{(3)} = v\right] = [v-u]^{-1}\int_{u}^{v} x dx = [v-u]^{-1}\frac{ [v^2 - u^2]}{2} = \frac{u+v}{2} $$

  • $\begingroup$ I didn't look at what you did, but you got an answer of $\frac{u - v}{2}$, not $\frac{u + v}{2}$ $\endgroup$ – Mark L. Stone Aug 18 '16 at 16:13
  • $\begingroup$ @MarkL.Stone you are right... I fixed that, the last line, integral of $\int x dx$ was incorrect. $\endgroup$ – them Aug 18 '16 at 16:18

Because the $X_i$ all have a uniform distribution, all (unordered) variables are assumed independent, and no other order statistic lies between $X_{(1)}$ and $X_{(3)}$, $X_{(2)}$ has a truncated uniform distribution supported on the interval $[X_{(1)}, X_{(3)}]$. Its mean obviously is $(X_{(1)}+X_{(3)})/2$, QED.

If you would like a formal demonstration, note that when the $X_i$ are iid with an absolutely continuous distribution $F$, the conditional density of $X_{(k)}$ (conditional on all the other order statistics) is $dF(x_k)/(F(x_{(k+1)}) - F(x_{(k-1)}))$, which is the truncated distribution. (When $k=1$, $F(x_{0})$ is taken to be $0$; and when $k=n$, $F(x_{n+1})$ is taken to be $1$.) This follows from Joint pdf of functions of order statistics, for instance, together with the definition of conditional densities.

  • $\begingroup$ whuber, when you write $dF(x_k)$ you refer to the probability density of X, am I right? $\endgroup$ – them Aug 24 '16 at 14:34
  • 1
    $\begingroup$ Yes, that is correct. By definition, $$dF(x)=\frac{d F}{d x}(x)\,dx.$$ (Technically, I should have called this the "probability element" rather than the "density".) $\endgroup$ – whuber Aug 24 '16 at 14:41

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.