📈 economics

Two Moments for Risk-Monotone Additive Statistics

This paper characterizes all additive statistics on probability laws with finite $p$ -th moments that are monotone with respect to mean-preserving spreads, showing that such statistics depend solely on the mean when $p < 2$ and on both the mean and variance when $p \ge 2$ .

Original authors: Mark Whitmeyer

Published 2026-08-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mark Whitmeyer

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve the mystery of "risk." In the world of economics and probability, risk isn't just a scary feeling; it's a mathematical shape that a set of numbers can take. Sometimes, a set of numbers is tight and predictable, like a group of friends standing close together. Other times, it's wild and spread out, like those same friends running in every direction. Mathematicians have a special tool called "convex order" to compare these shapes. If one shape is "riskier" than another, it means it has more extreme ups and downs, even if the average height of the group stays the same.

Now, imagine you have a magic calculator that gives you a single number to represent how "risky" a situation is. This calculator has two superpowers. First, it's additive: if you mix two independent risks together (like rolling two dice), the total risk score is just the sum of the individual scores. Second, it's monotone: if a situation gets riskier (more spread out), your calculator's number must go up. For decades, economists have wondered: what kind of calculator can do both? Is it just the average? Is it the variance (a measure of spread)? Or is it something weird and complex? This question matters because if we know exactly what these calculators look like, we can better understand how people make decisions when the future is uncertain.

The paper you are about to read, "Two Moments for Risk-Monotone Additive Statistics" by Mark Whitmeyer, acts as the final judge in this mystery. The author proves that the answer depends entirely on how "heavy" the tails of the risk are—specifically, whether the risk has a finite "second moment" (a fancy way of saying the variance exists and isn't infinite).

Here is the verdict:
If the risks you are looking at are "light" enough that their variance might not even exist (mathematically, when the moment $p$ is less than 2), the only possible calculator is one that ignores risk entirely. In this world, the only thing that matters is the average. Any attempt to add a "risk penalty" to the average breaks the rules of additivity. The paper proves that for these light risks, the statistic representing the riskiness itself (when the average is removed) must be zero. The total score is simply the average; there is no room for variance as a separate factor.

However, if the risks are "heavy" enough to guarantee a finite variance (when $p$ is 2 or greater), the story changes. In this world, the author proves that the only calculators that work are Mean-Variance calculators. This means your risk score is always the average plus (or minus) a multiple of the variance. The paper shows that variance is the only extra ingredient allowed. You can't invent a new, complex formula; you are stuck with the classic combination of "average" and "spread."

The author didn't just guess this; they built a rigorous mathematical proof. They showed that if you try to sneak in any other type of risk measurement, the math eventually explodes, leading to impossible contradictions. They also ruled out the idea that variance could exist as a factor for the "light" risks ( $p < 2$ ); the proof demonstrates that if you tried to use variance there, the numbers would blow up to infinity, which is impossible for a real-world statistic.

So, the paper draws a sharp line in the sand at the number 2. Below 2, risk is invisible to additive calculators (the score is just the mean); at 2 and above, risk is strictly a matter of variance. It's a clean, definitive answer to a question that has puzzled thinkers for years, confirming that the universe of risk statistics is much simpler than it appears: it's either just the average, or it's the average plus a bit of variance.

Technical Summary: Two Moments for Risk-Monotone Additive Statistics

Problem Statement
The paper investigates the characterization of real-valued statistics defined on probability laws (distributions) that satisfy two specific properties:

Additivity under convolution: The statistic of the sum of two independent risks equals the sum of their individual statistics ( $F(\mu * \nu) = F(\mu) + F(\nu)$ ).
Monotonicity in convex order: The statistic is non-decreasing with respect to the convex order ( $\mu \preceq_{cx} \nu \implies F(\mu) \leq F(\nu)$ ), which corresponds to monotonicity in mean-preserving spreads.

The central question is whether these two axioms uniquely identify the variance (or mean-variance combinations) as the only possible statistics, or if other forms exist. This inquiry is situated within the domain of probability measures with finite $p$ -th moments ( $L^p$ ), specifically examining how the threshold $p=2$ (the existence of a finite second moment) influences the characterization. The work addresses a question posed by Fritz et al. [1] regarding whether convex-order monotonicity singles out variance on the finite-second-moment domain, contrasting with the broader characterization of mean-variance certainty equivalents by Pomatto et al. [4] on domains with all finite moments.

Methodology
The proof strategy relies on a dichotomy based on the value of $p$ relative to 2. The analysis is conducted first on centered laws (mean zero) and then extended to arbitrary laws.

Reduction to Centered Laws: The author decomposes any random variable $X$ into a centered component $(X - E[X])$ and a constant mean. Since the convex order preserves the mean, the behavior of the statistic on the mean component is arbitrary (any additive function), while the behavior on the centered component is the focus of the characterization.
Bounded vs. Unbounded Risks: The proof proceeds by first establishing the form of the statistic on bounded centered risks and then extending these results to unbounded risks within $L^p$ $L^{p}$ .
- Bounded Risks: Using a result from Pomatto et al. [4], the author shows that if one bounded centered risk has a smaller variance than another, a common independent background risk can be added to make the first risk smaller in convex order. Combined with additivity, this forces the statistic to be non-decreasing in variance. Further, by considering sums of independent copies, the author proves that the statistic must be linear in variance for bounded risks.
- Unbounded Risks ( $p \geq 2$ ): For domains where variance is finite, the author defines a residual statistic $D(\mu) = R(\mu) - c\text{Var}(\mu)$ . The proof demonstrates that $D(\mu)$ must be non-negative (via truncation and Jensen's inequality) and that any positive value of $D$ can be "compressed" into arbitrarily small risks. By summing many such independent small risks, the additivity property would force the statistic to diverge to infinity, contradicting the assumption that the statistic is real-valued. Thus, $D(\mu) = 0$ .
- Unbounded Risks ( $p < 2$ ): In this range, the domain includes risks with finite $p$ -th moments but infinite variance. The author constructs a specific symmetric random variable with finite $p$ -th moment but infinite variance. Its bounded truncations have variances growing to infinity. If the coefficient $c$ on variance were positive, the monotonicity of the statistic would require the value of the unbounded risk to be infinite, a contradiction. Thus, $c$ must be zero.

Key Contributions and Results

The Dichotomy at $p=2$ : The paper establishes that $p=2$ is the exact threshold for the existence of non-trivial risk-monotone additive statistics beyond the mean.
- Case $1 \leq p < 2$ : On the space of laws with finite $p$ -th moments, the only centered statistic satisfying the axioms is the zero map. Consequently, any such statistic on the full domain depends only on an additive function of the mean.
- Case $2 \leq p < \infty$ : On the space of laws with finite second moments, the centered statistic must be a non-negative multiple of the variance ( $c \cdot \text{Var}(\mu)$ with $c \geq 0$ ). Consequently, any such statistic on the full domain is of the mean-variance form: $a(m(\mu)) + c \cdot \text{Var}(\mu)$ .
Characterization of Certainty Equivalents:
- When $p < 2$ , the only certainty equivalent (a statistic assigning a value to a lottery equal to a sure amount) that is additive and monotone in convex order is the expected value.
- When $p \geq 2$ , the class of such certainty equivalents expands to include all mean-variance forms ( $m(\mu) - c \cdot \text{Var}(\mu)$ with $c \geq 0$ ).
Resolution of Prior Questions: The paper answers the question raised by Fritz et al. [1] affirmatively: convex-order monotonicity does single out variance (up to a scalar) on the finite-second-moment domain, but only when the domain is restricted to $L^2$ or higher. It clarifies that the "room" for variance in these axioms exists exactly when the domain requires a second moment.

Significance
The paper contributes to the study of monotone additive statistics, a field pursued by Ruzsa and Székely [5], Mattner [2], and Mu et al. [3]. Its primary significance lies in identifying the precise moment condition ( $p=2$ ) that separates the uniqueness of the expected value from the possibility of mean-variance analysis. The proof technique is notable for combining the independent-background-risk comparison method (from Pomatto et al.) with the observation that failures of $L^p$ -continuity at zero can be amplified to arbitrary degrees by summing independent small risks. This demonstrates that the structural constraints of additivity and convex monotonicity are so rigid that they force the statistic to vanish entirely if the domain allows for infinite variance, and forces it to be purely variance-based if the domain restricts variance to be finite.

Technical Summary: Two Moments for Risk-Monotone Additive Statistics

More like this