📊 statistics

Real-order moments, tail representations, and logarithmic means

This paper establishes a unified framework for real-order moments of arbitrary random variables by deriving general integral and series representations in terms of distribution functions that extend classical tail identities to cover positive, fractional, and negative moments, while also linking logarithmic moments to Laplace transforms and the Frullani identity.

Original authors: Roberto Vila, Eduardo Nakano

Published 2026-06-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Roberto Vila, Eduardo Nakano

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 trying to understand the "personality" of a mysterious character named Random Variable ( $X$ ). In the world of statistics, we usually try to understand this character by calculating their "moments."

Think of a moment like a snapshot of the character's weight or energy at a specific level.

The 1st moment is their average height (the mean).
The 2nd moment relates to how much they wiggle or vary (variance).
Usually, we only look at snapshots where the character is positive (standing up). But what if the character can also be negative (lying down) or take on weird, fractional shapes?

This paper by Roberto Vila and Eduardo Nakano is like a universal translator that finally allows us to take these snapshots for any character, no matter how weird they are, using a single, unified set of rules.

Here is the breakdown of their new "translation method" using simple analogies:

1. The Old Way vs. The New Way

The Old Way: Previously, if you wanted to know the "weight" (moment) of a character who could only be positive, you used a specific tool called a Tail-Integral. Imagine this as measuring how much "space" the character occupies as you look further and further out into the distance. It worked great for positive characters, but if the character could be negative or mixed, statisticians had to use different, messy tools for each case.

The New Way (The Paper's Contribution): The authors built a Master Key. They created one single formula that works for:

Continuous characters (smooth, flowing like water).
Discrete characters (stepped, like a staircase).
Mixed characters (a bit of both).
Positive, Negative, and Fractional moments (even if the "power" you are measuring is a weird number like 0.5 or -2).

They achieved this by looking at the character's Cumulative Distribution Function (CDF). Think of the CDF as a staircase map that shows the probability of the character being below a certain height. The paper shows that you can calculate the "weight" of the character simply by measuring the area above and below this staircase map.

2. The "Area" Analogy (Geometric Interpretation)

The paper explains that calculating a moment is like doing a tug-of-war between two areas on a graph.

The Red Area: This is the space above the staircase map (the "tail"). It represents the probability of the character being very large.
The Blue Area: This is the space below the staircase map. It represents the probability of the character being small or negative.

To find the character's "moment," you simply take the Red Area and subtract the Blue Area.

If the Red Area is huge, the character has a heavy positive moment.
If the Blue Area is huge, the character has a heavy negative moment.
If the areas are balanced, the moment is zero.

This works for discrete characters (like rolling a die) too. Instead of smooth areas, the paper shows you how to add up little "steps" in the staircase. It turns a complex calculus problem into a simple sum of probabilities.

3. The "Existence" Test (Will the Number Break?)

Sometimes, when you try to calculate a moment, the number explodes to infinity (it "breaks"). This usually happens if the character has a "heavy tail"—meaning they occasionally take on massive values that are hard to ignore.

The paper provides a simple litmus test:

Look at the very ends of the staircase (the tails).
If the "Red Area" and "Blue Area" are finite (they don't stretch to infinity), the moment exists.
If the tails are too "fat" (too much area), the moment doesn't exist.

They tested this on two famous characters:

The Zeta Distribution: A character known for having very heavy tails. The paper's method quickly confirmed the classic rule: "You can only measure this character's weight if the power you choose is small enough."
The Skellam Distribution: A character formed by subtracting two Poisson processes (like counting the difference between two types of events). The paper showed how to visualize their average behavior by looking at the areas under their specific staircase map.

4. The "Logarithmic" Mystery (The Zero Moment)

There is a special case in math called the Logarithmic Moment (related to $\log(X)$ ). It's like asking, "What is the character's weight if we zoom in infinitely close to zero?"

The authors discovered a clever trick to find this. They realized that if you take the "moment" formula and slowly turn the dial down to zero, it transforms into a new formula involving the Laplace Transform.

Think of the Laplace Transform as a "fingerprint" of the character. The paper shows that you can calculate the logarithmic moment by comparing this fingerprint against a standard "ghost" fingerprint (the exponential function).

They linked this to an old mathematical identity called Frullani's Identity (named after a mathematician from 1941).
The Result: If Character A has a "stronger" Laplace fingerprint than Character B, then Character A will have a smaller logarithmic moment. This gives statisticians a new way to compare characters without doing heavy calculations.

Summary

In short, this paper says:

Stop using different tools for different types of random variables.
Use the "Staircase Map" (CDF) to measure everything.
Calculate moments by measuring the Red Area minus the Blue Area.
Check for infinity by looking at how wide the tails of the staircase are.
Handle the tricky "Log" case by using the character's "Laplace Fingerprint."

It unifies the entire field of moments into one elegant, geometric framework, making it easier to see the shape of probability distributions, whether they are smooth, stepped, positive, or negative.

Technical Summary: Real-order Moments, Tail Representations, and Logarithmic Means

Problem Statement
While the classical tail-integral representation for moments of nonnegative random variables is a standard tool in probability theory (e.g., Feller, 1971; Shorack, 2000), corresponding formulas for arbitrary real-valued random variables are often treated in isolation. Specifically, existing literature frequently lacks a unified framework that simultaneously addresses continuous, discrete, and mixed distributions, as well as positive, fractional, and negative-order moments. Furthermore, the connection between logarithmic moments and Laplace transforms is often treated separately from general moment representations. This paper addresses the need for a cohesive theoretical structure that extends tail-integral identities to arbitrary random variables supported on the entire real line.

Methodology
The authors develop a unified framework based on the decomposition of the expectation $E[X^p]$ into positive and negative parts, utilizing the cumulative distribution function (CDF) $F_X$ and survival functions.

General Integral Representations: For an arbitrary random variable $X$ and real order $p > 0$ , the paper derives a general formula by splitting the domain into $X > 0$ and $X < 0$ . By applying Fubini's theorem to interchange the order of integration and utilizing the definition of the CDF, the authors establish that the $p$ -th moment can be expressed as:
$E[X^p] = p \int_0^\infty t^{p-1}[1 - F_X(t)] dt - p \int_{-\infty}^0 t^{p-1}F_X(t-) dt$
This formulation relies on the survival function for the positive part and the left-continuous CDF for the negative part.
Discrete Distributions: Exploiting the stepwise nature of the CDF for discrete variables, the authors convert the integral representations into explicit series. By partitioning the integration domain at the support points of the distribution, they derive summation formulas involving cumulative probabilities. For integer-valued variables, this yields specific series involving differences of powers of integers and tail probabilities.
Negative and Logarithmic Moments:
- Negative Moments: By applying the general representation to the reciprocal variable $Y = X^{-1}$ , the authors derive an integral expression for negative moments $E[X^{-q}]$ in terms of the CDF near the origin.
- Logarithmic Moments: The paper investigates the limiting behavior of real-order moments as $p \to 0$ . By differentiating the moment representation with respect to $p$ and utilizing the identity $1/t = \int_0^\infty e^{-tx} dx$ , the authors link the logarithmic moment $E[\log(X)]$ to the Laplace transform of $X$ .

Key Results and Contributions

Unified Moment Representation: The primary contribution is the derivation of a single integral identity (Equation 4) that applies to continuous, discrete, and mixed distributions on $\mathbb{R}$ . This identity generalizes the classical tail-integral formula for nonnegative variables to arbitrary real-valued variables.
Existence Criteria: The paper establishes necessary and sufficient conditions for the existence of real-order moments. Theorem 1 demonstrates that $E[|X|^p] < \infty$ if and only if the integrals of the tail probabilities (weighted by $t^{p-1}$ ) converge at both $+\infty$ and $-\infty$ .
Discrete Series Formulas: For discrete random variables, the paper provides explicit series representations (Equation 8) for moments in terms of cumulative probabilities. A specific case for integer-valued variables (Equation 9) recovers the well-known identity $E[X] = \sum_{k=0}^\infty P(X > k) - \sum_{k=1}^\infty P(X \leq -k)$ .
Applications to Specific Distributions:
- Zeta Distribution: The discrete series representation is applied to the Zeta distribution to recover the classical condition for the existence of power moments ( $p < \alpha - 1$ ), demonstrating how tail behavior dictates moment finiteness.
- Skellam Distribution: The framework is applied to the Skellam distribution (the difference of two Poisson variables), providing a geometric interpretation where the mean is represented as the difference between the area above and below the CDF.
- Skew-Normal Distribution: The continuous representation is illustrated using the skew-normal distribution, showing how the mean relates to the areas under the CDF.
Logarithmic Moments and Laplace Transforms: The paper derives the identity:
$E[\log(X)] = \int_0^\infty \frac{e^{-x} - L_X(x)}{x} dx$
where $L_X(x)$ is the Laplace transform of the positive random variable $X$ . This result connects logarithmic means directly to the Frullani identity and establishes that the Laplace-transform order ( $L_X(s) \geq L_Y(s)$ ) implies an ordering of logarithmic moments ( $E[\log(X)] \leq E[\log(Y)]$ ).

Significance and Claims
The paper claims to provide a "unified perspective" on moment representations. Its significance lies in:

Generality: It removes the restriction to nonnegative random variables and the requirement for probability density functions, accommodating discrete and mixed cases naturally.
Analytical Utility: The derived series for discrete distributions offer practical criteria for moment existence based on tail asymptotics, which is particularly useful for heavy-tailed models where densities may be difficult to handle.
Theoretical Connection: The derivation of the logarithmic moment representation bridges the gap between moment theory, Laplace transforms, and classical integral identities (Frullani), offering a new tool for comparing logarithmic moments via stochastic ordering.

The authors conclude that these results facilitate the study of tail behavior, risk measures, and stochastic modeling by providing a common treatment for positive, fractional, and negative moments. They suggest future extensions to truncated moments, conditional moments, and multivariate distributions but do not propose specific experimental validations or unmentioned applications.