Abstract: For an order statistic $(X_{1:n}, \dots, X_{n:n})$ of a collection of independent but not necessarily identically distributed random variables, and any $i \in \{1, \dots, n\}$, the conditional distribution of $(X_{i+1:n}, \dots, X_{n:n})$ given $X_{i:n} > s$ is shown to be stochastically increasing in $s$. This answers a question by Hu and Xie (2006).
Introduction
In probability theory and reliability analysis, order statistics—the values of a sample sorted in non-decreasing order ($X_{1:n} \le X_{2:n} \le \dots \le X_{n:n}$)—are fundamental tools. Because these statistics are derived from the same underlying set of random variables, they naturally exhibit positive dependence. If one order statistic is known to be large, it generally suggests that others in the set are likely to be large as well.
In 2006, researchers Hu and Xie posed a specific question regarding the strength of this dependence: If we condition on the $i$-th order statistic being greater than a threshold $s$ ($X_{i:n} > s$), does the resulting conditional distribution stochastically dominate the distribution conditioned on a smaller threshold $s'$? While intuition suggests that forcing one part of the sorted sequence upward should shift the entire sequence upward, this paper proves that the reality is more complex.
The Problem and Intuition
The core challenge addressed in this research is determining the exact boundaries of stochastic domination within order statistics. Stochastic domination ($A \ge_{st} B$) implies that for any increasing function, the expected value under distribution $A$ is greater than or equal to that under $B$.
While it seems intuitive that $X_{i:n} > s$ should make the entire vector $(X_{1:n}, \dots, X_{n:n})$ "larger," the authors demonstrate that this intuition is mathematically false for the full vector. Specifically, the lower order statistics (those with indices less than $i$) do not necessarily follow this trend. The paper therefore focuses on proving the refined conjecture: that stochastic domination holds specifically for the upper order statistics—those at or above the index $i$.
Methodology
The authors utilize the coupling method, a powerful probabilistic technique used to prove inequalities between distributions by constructing different random variables on a single probability space. The proof follows a rigorous four-step progression:
- Simplification: The authors initially assume the random variables have finite, completely disjoint supports. This removes the complexities of overlapping values and ties.
- Indicator Variables: The problem is translated into the study of Bernoulli indicator variables (where a variable is 1 if it exceeds a threshold and 0 otherwise). This allows the authors to leverage known lemmas regarding the sums of independent Bernoulli variables.
- Coupling Construction: A joint distribution is constructed such that the variables conditioned on a higher threshold $s$ are guaranteed to be point-wise greater than or equal to those conditioned on a lower threshold $s'$.
- Limiting Arguments: To ensure the theorem applies to all independent random variables (including continuous ones), the authors use limiting arguments. They shift variables by infinitesimal amounts ($\epsilon$) and apply discrete grids that become infinitely fine, ensuring the discrete proof generalizes to the continuous case.
Key Results
Theorem 1: The Main Result
The primary contribution of the paper is the formal proof of the following: For a set of independent random variables and any index $i$, the conditional distribution of the upper order statistics $[(X_{i:n}, \dots, X_{n:n}) \mid X_{i:n} > s]$ is stochastically increasing in $s$. That is, if $s < s'$, then:
$$[(X_{i:n}, \dots, X_{n:n}) \mid X_{i:n} > s'] \ge_{st} [(X_{i:n}, \dots, X_{n:n}) \mid X_{i:n} > s]$$
Disproving the General Case
The paper definitively shows that the property does not extend to the entire vector. Through an elegant counterexample involving two discrete random variables, the authors prove that conditioning on $X_{i:n} > s$ does not necessarily increase the distribution of order statistics below index $i$. This highlights a critical distinction between "orthant properties" (which Hu and Xie previously proved) and full stochastic domination.
Symmetry and Extensions
By applying a transformation (substituting $-X$ for $X$), the authors note that a symmetric property holds for the lower tail: conditioning on $X_{i:n} < s$ stochastically decreases the distribution of all smaller order statistics ($X_{1:n}$ through $X_{i:n}$).
Conclusion
This research provides a definitive answer to the 2006 conjecture by Hu and Xie. While it confirms that the "upper" portion of a sorted sequence behaves as intuition suggests under conditioning, it also serves as a cautionary tale about the limits of stochastic dependence. Interestingly, the authors' coupling construction reveals that the failed general case is "not terribly far from being true," as the domination only fails at a specific index. This suggests that while strict stochastic domination is too high a bar for the full vector, other forms of positive dependence may yet be discovered in the relationship between lower and upper order statistics.
Summary updated