UW STAT 512: Introduction to Statistical Inference (2022 Autumn)
Lecture 9: Order Statistics from a Continuous Univariate Distribution

Yen-Chi Chen (University of Washington)

Let \(X_1,\cdots, X_n\) be independent and identically distributed continuous random variables with a probability density function \(p_X(x)\) and a cumulative distribution function \(F_X(x)\). Since they are continuous random variables, they assume distinct values almost surely. The order statistics \(Y_1<Y_2<\cdots<Y_n\) are the ordered versions of these \(n\) random variables such that \(Y_j\) is the \(j\)-th smallest value among \(\{X_1,\cdots, X_n\}\). Thus, \[\begin{align*} Y_1 &= \min\{X_1,\cdots, X_n\}\\ Y_n &= \max\{X_1,\cdots, X_n\}. \end{align*}\] Sometimes, we use the notation \(X_{(j)} = Y_j\). Note that the mapping \[(X_1,\cdots, X_n)\rightarrow (Y_1,\cdots, Y_n)\] is not one-to-one but \(n!\)-to-one. This is due to the fact that any of the \(n!\) permutations of \(X_1,\cdots, X_n\) will lead to the same order statistics.

The probability density functions of the \(Y_i\)’s are related to the probability density functions of the \(X_i\)’s as follows:

Case study: uniform distribution

Consider the case where \(X_1,\cdots, X_n\) are IID from \({\sf Uni}[0,1]\). Then \(p_X(x)=1\) and \(F_X(x) = x\) when \(x\in[0,1]\). Thus, \[p_{Y_j}(y) = \frac{n!}{(j-1)!(n-j)!} y^{j-1} (1-y)^{n-j},\] which is the PDF of \({\sf Beta}(j, n-j+1)\).

Here is an interesting note regarding the variance. The variance of \(Y_j\) is \[{\sf Var}(Y_j) = \frac{j(n-j+1)}{(n+1)^2 (n+2)},\] which is maximized when \(j= \frac{n+1}{2}\), assuming \(n\) is an odd number. The value \(j= \frac{n+1}{2}\) corresponds to the median of \(\{X_1,\cdots, X_n\}\). Thus, the median has the highest variability. In this case, \[{\sf Var}(Y_{\frac{n+1}{2}}) = \frac{1}{4(n+2)} = O(n^{-1}).\] On the other hand, the maximum or minimum value has the lowest variance: \[{\sf Var}(Y_{1}) = {\sf Var}(Y_{n}) = \frac{n}{(n+1)^2 (n+2)} = O(n^{-2}).\]

Now we consider another way to look at the order statistics. Let \(W_1,\cdots, W_{n}, W_{n+1}\) be the ‘spacing’ between consecutive order statistics: \[\begin{align*} W_1 & = Y_1 - 0\\ W_2 & = Y_2 - Y_1 \\ W_3 & = Y_3-Y_2\\ \vdots\\ W_n & = Y_n-Y_{n-1}\\ W_{n+1} & = 1-Y_n. \end{align*}\] It is easy to see that \(W_i\in[0,1]\) and \(W_1+W_2+\cdots +W_{n+1}=1\). Also, we can reparameterize \(Y_j\) via the \(W_i\)’s: \[Y_j = W_1+W_2+\cdots+W_j.\] Since the \(X_i\)’s are uniform over \([0,1]\), the joint probability density function of \(Y_1,\cdots, Y_n\) is \[p(y_1,\cdots, y_n) = n!\] whenever \(0<y_1<\cdots<y_n<1\). By the Jacobian method and the fact that \(\left|{\sf det}\left(\frac{dY}{dW}\right)\right| = 1\), we conclude that \[p(w_1,\cdots, w_n) = n!\] whenever \(w_i\in[0,1]\) and \(w_1+\cdots+w_n <1\). One can easily see that \(p(w_1,\cdots, w_n)\) is invariant under permutations of \(W_1,\cdots, W_n\) (i.e., they are exchangeable), so the marginal distribution of \(W_i\) is the same as the marginal distribution of \(W_j\) for all \(i,j=1,\cdots, n\). Because \(W_1 = Y_1\) follows a \({\sf Beta}(1, n)\) distribution, we conclude that \(W_j\) is a \({\sf Beta}(1, n)\) random variable.

Note that \(W_i\) and \(W_j\) are dependent for \(i\neq j\). Due to the exchangeability property, the joint distribution of \((W_i,W_j)\) is the same as the joint distribution of \((W_1,W_2)\), so \[\begin{align*} {\sf Cov}(W_i,W_j) &= {\sf Cov}(W_1,W_2)\\ & = \frac{1}{2}\left({\sf Var}(W_1+W_2) -{\sf Var}(W_1) -{\sf Var}(W_2)\right)\\ & = \frac{1}{2}\left({\sf Var}(Y_2) -2{\sf Var}(Y_1)\right)\\ &= \frac{1}{2}\left(\frac{2(n-1)}{(n+1)^2 (n+2)} - 2 \frac{n)}{(n+1)^2 (n+2)}\right)\\ & = \frac{-1}{(n+1)^2 (n+2)}<0. \end{align*}\]