Assume X = $(X_1, ..., X_n)$ ~ $U(\theta, 2\theta)$, where $\theta \in \Bbb{R}^+$.

How does one calculate the conditional expectation of $E[X_1|X_{(1)},X_{(n)}]$, where $X_{(1)}$ and $X_{(n)}$ are the smallest and largest order statistics respectively?

My first thought would be that since the order statistics limit the range, it is simply $(X_{(1)}+X_{(n)})/2$, but im not sure if this is correct!


Consider the case of an iid sample $X_1, X_2, \ldots, X_n$ from a Uniform$(0,1)$ distribution. Scaling these variables by $\theta$ and translating them by $\theta$ endows them with a Uniform$(\theta, 2\theta)$ distribution. Everything relevant to this problem changes in the same way: the order statistics and the conditional expectations. Thus, the answer obtained in this special case will hold generally.

Let $1\lt k\lt n.$ By emulating the reasoning at https://stats.stackexchange.com/a/225990/919 (or elsewhere), find that the joint distribution of $(X_{(1)}, X_{(k)}, X_{(n)})$ has density function

$$f_{k;n}(x,y,z) = \mathcal{I}(0\le x\le y\le z \le 1) (y-x)^{k-2}(z-y)^{n-k-1}.$$

Fixing $(x,z)$ and viewing this as a function of $y,$ this is recognizable as a Beta$(k-1, n-k)$ distribution that has been scaled and translated into the interval $[x,z].$ Thus, the scale factor must be $z-x$ and the translation takes $0$ to $x.$

Since the expectation of a Beta$(k-1,n-k)$ distribution is $(k-1)/(n-1),$ we find that the conditional expectation of $X_{(k)}$ must be the scaled, translated expectation; namely,

$$\mathbb{E}\left(X_{(k)}\mid X_{(1)}, X_{(n)}\right) = X_{(1)} + \left(X_{(n)}-X_{(1)}\right) \frac{k-1}{n-1}.$$

The cases $k=1$ and $k=n$ are trivial: their conditional expectations are, respectively, $X_{(1)}$ and $X_{(k)}.$

Let's find the expectation of the sum of all order statistics:

$$\mathbb{E}\left(\sum_{k=1}^n X_{(k)}\right) = X_{(1)} + \sum_{k=2}^{n-1} \left(X_{(1)} + \left(X_{(n)}-X_{(1)}\right) \frac{k-1}{n-1}\right) + X_{(n)}.$$

The algebra comes down to obtaining the sum $$\sum_{k=2}^{n-1}(k-1) = (n-1)(n-2)/2.$$


$$\eqalign{ \mathbb{E}\left(\sum_{k=1}^n X_{(k)}\right) &= (n-1)X_{(1)} + \left(X_{(n)}-X_{(1)}\right) \frac{(n-1)(n-2)}{2(n-1)} + X_{(n)} \\ &= \frac{n}{2}\left(X_{(n)}+X_{(1)}\right). }$$

Finally, because the $X_i$ are identically distributed, they all have the same expectation, whence

$$\eqalign{n\mathbb{E}\left(X_1\mid X_{(1)}, X_{(n)}\right) &= \mathbb{E}\left(X_1\right) + \mathbb{E}\left(X_2\right) + \cdots + \mathbb{E}\left(X_n\right)\\ &= \mathbb{E}\left(X_{(1)}\right) + \mathbb{E}\left(X_{(2)}\right) + \cdots + \mathbb{E}\left(X_{(n)}\right) \\ &= \frac{n}{2}\left(X_{(n)}+X_{(1)}\right), }$$

with the unique solution

$$\mathbb{E}\left(X_1\mid X_{(1)}, X_{(n)}\right) = \left(X_{(n)}+X_{(1)}\right)/2.$$

It seems worth remarking that this result is not a sole consequence of the symmetry of the uniform distribution: it is particular to the uniform family of distributions. For some intuition, consider data drawn from a Beta$(a,a)$ distribution with $a \lt 1.$ This distribution's probabilities are concentrated near $0$ and $1$ (its density has a U or "bathtub" shape). When $X_{(n)}\lt 1/2,$ we can be sure most of the data are piled up close to $X_{(1)}$ and therefore will tend to have expectations less than the midpoint $(X_{(1)}+X_{(n)})/2;$ and when $X_{(1)}\gt 1/2,$ the opposite happens and most of the data are likely piled up close to $X_{(n)}.$


The following is not a proof but a verification of the desired result once you know that $(X_{(1)},X_{(n)})$ is a complete statistic for $\theta$ :

Joint pdf of $X_1,X_2,\ldots,X_n$ is

\begin{align} f_{\theta}(x_1,\ldots,x_n)&=\frac{1}{\theta^n}\mathbf1_{\theta<x_{(1)},x_{(n)}<2\theta} \\&=\frac{1}{\theta^n}\mathbf1_{\frac{1}{2}x_{(n)}<\theta<x_{(1)}}\quad,\,\theta\in\mathbb R^+ \end{align}

So $T=(X_{(1)},X_{(n)})$ is a sufficient statistic for $\theta$. It can be shown that $T$ is also a complete statistic by proceeding along these lines.

Then by Lehmann-Scheffe theorem, $E\,[X_1\mid T]$ is the UMVUE of $E(X_1)=\frac{3\theta}{2}$.

Now, $\frac{1}{\theta}(X_i-\theta)\stackrel{\text{i.i.d}}\sim \mathsf U(0,1)$, so that $\frac{1}{\theta}(X_{(n)}-\theta)\sim\mathsf{Beta}(n,1)$ and $\frac{1}{\theta}(X_{(1)}-\theta)\sim \mathsf{Beta}(1,n)$.

Therefore, $E(X_{(n)})=\frac{n\theta}{n+1}+\theta=\frac{(2n+1)\theta}{n+1}$ and $E(X_{(1)})=\frac{\theta}{n+1}+\theta=\frac{(n+2)\theta}{n+1}$.



This proves that $\frac{1}{2}(X_{(1)}+X_{(n)})$ is the UMVUE of $\frac{3\theta}{2}$ by Lehmann-Scheffe.

Since UMVUE is unique whenever it exists, it verifies the claim that $E\,[X_1\mid T]=\frac{1}{2}(X_{(1)}+X_{(n)})$.

  • $\begingroup$ +1 This answer is nice because it reveals a deeper way to understand the exercise and what it can teach us. $\endgroup$ – whuber Mar 6 '19 at 15:48

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.