Abstract
We show that a lower bound for covariance of $\min(X_1,X_2)$ and $\max(X_1,X_2)$ is $\text{cov}(X_1,X_2)$ and an upper bound for variance of $\min(X_2,\max(X,X_1))$ is $\text{var}(X) + \text{var}(X_1) + \text{var}(X_2)$, generalizing previous results. We also characterize the cases where these bounds are sharp.
Introduction
Truncation is a fundamental operation in statistics, probability, and operations research, occurring whenever a random variable is restricted within a specific range. In practical applications like inventory management, the actual quantity sold is often the minimum of demand and available stock. In queueing theory, waiting times are frequently modeled as the maximum of zero and a calculated delay.
While it is well-established that simple truncation typically reduces the variance of a single random variable, the mathematical behavior of multiple variables under truncation—especially when truncation points are themselves random—is more complex. This paper provides a rigorous exploration of these relationships, establishing new, sharp mathematical bounds for the covariance and variance of truncated random variables, generalizing previous work that was limited to identical distributions.
Theoretical Framework and Methods
The study approaches the problem through algebraic proofs based on the fundamental properties of expectations and variances. The authors focus on the relationship between two random variables, $X_1$ and $X_2$, and their order statistics: the minimum ($Y = \min(X_1, X_2)$) and the maximum ($Z = \max(X_1, X_2)$).
A central technique in the paper involves the exploitation of two key identities:
- $Y + Z = X_1 + X_2$
- $YZ = X_1X_2$
By manipulating these identities, the authors extend the findings of previous researchers (such as Papadatos) to cases where $X_1$ and $X_2$ follow different probability distributions. The research also identifies the specific conditions under which these inequalities become equalities, referred to as the "sharpness" of the bounds.
Key Insights and Theorems
The Covariance Floor
One of the primary contributions of the paper is the establishment of a lower bound for the covariance of order statistics.
Theorem 1: $\text{cov}(\min(X_1, X_2), \max(X_1, X_2)) \ge \text{cov}(X_1, X_2)$.
This result implies that the process of taking the minimum and maximum of two variables tends to preserve or even increase their linear relationship. This "covariance floor" holds regardless of whether the variables share the same distribution.
Variance Conservation and Double Truncation
The paper also explores how variance is distributed and bounded when variables are combined or truncated.
Theorem 2: $\text{var}(\min(X_1, X_2)) + \text{var}(\max(X_1, X_2)) \le \text{var}(X_1) + \text{var}(X_2)$.
This suggests a principle of variance conservation, where the combined variability of the order statistics is always less than or equal to the combined variability of the original components.
Extending this to "doubly truncated" variables—where a variable $X$ is bounded from below by $X_1$ and from above by $X_2$—the authors derive a definitive upper bound:
Corollary 2: For the doubly truncated variable $X(X_1, X_2) = \min(X_2, \max(X, X_1))$, the variance is bounded by: $$\text{var}(X(X_1, X_2)) \le \text{var}(X_2) + \text{var}(X_1) + \text{var}(X)$$
Sharpness Conditions
To provide a complete characterization, the authors identify the "if and only if" conditions for these bounds. For instance, the covariance bound in Theorem 1 is only sharp (reaching equality) if one variable is almost surely greater than the other (e.g., $X_1 \ge X_2$ or $X_2 \ge X_1$). Theorem 4 further details four specific sets of conditions involving constants $c_1$ and $c_2$ under which the double truncation variance bound reaches equality.
Conclusion and Future Directions
This research provides essential theoretical tools for understanding how truncation impacts the moments of random variables. By establishing general bounds that do not rely on identical distributions, the paper offers a more flexible framework for researchers in operations research and statistics.
However, the scope is limited to first and second moments (mean, variance, and covariance). Future research could extend these bounds to higher-order moments like skewness and kurtosis, which are critical for risk assessment. Additionally, while the paper is motivated by practical problems in inventory and queueing, the current work is purely theoretical; future studies could apply these bounds to numerical simulations or empirical case studies to demonstrate their utility in real-world engineering contexts.
Summary updated