🔢 mathematics

Variance gamma approximation to sums of triplewise independent random variables

This paper establishes a connection between triplewise independent random variables and complete bipartite graphs to construct such variables, then derives variance gamma approximation bounds for their sums using Stein's method and the generalized zero-bias transformation.

Original authors: Aditi Panda, Kalyan Barman

Published 2026-08-14

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Aditi Panda, Kalyan Barman

Original paper licensed under CC BY 4.0 (https://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

The Great Statistical Party and the Rule of Three

Imagine a massive, chaotic party where thousands of guests are mingling. In the world of statistics, there is a famous rule called the Central Limit Theorem. Think of it as the universe's way of saying that if you gather enough people who are all acting completely independently of one another, their collective behavior will eventually smooth out into a perfect, predictable bell curve. It's like if you asked a million strangers to guess the weight of a pumpkin; even if their individual guesses are wild, the average of all their guesses will form a neat, symmetrical hill. This rule is the bedrock of modern science, finance, and polling.

But here is the twist: what happens if the guests aren't totally independent? What if they are only "sort of" independent? In statistics, we have different levels of friendship. "Pairwise independence" means any two guests don't influence each other. "Triplewise independence" is a stricter rule: it means that no matter which three guests you pick, they don't influence each other as a group. For a long time, mathematicians wondered: if we have a party where everyone is triplewise independent, will the crowd still form that perfect bell curve? The answer, surprisingly, is no. It turns out that being independent in groups of three isn't enough to guarantee the classic bell curve. Instead, the crowd might form a weird, lopsided shape that looks nothing like what we expect. This is where the story of this research begins.

The Paper's Journey: From Graphs to Gamma

In this article, authors Aditi Panda and Kalyan Barman from the National Institute of Technology Warangal dive deep into this mystery. They aren't just asking if the bell curve breaks; they are trying to figure out exactly what shape the crowd takes instead, and how to measure the distance between the messy reality and the perfect shape.

To visualize their setup, imagine a giant game board made of a "complete bipartite graph." Picture two teams of players, Team A and Team B, with $m$ players on each side. Every player on Team A is connected by a string to every player on Team B, creating a web of $m^2$ connections. The authors assign random numbers to the players and then look at the connections. If two players have the same number, a "link" is formed. They then count up all these links.

The authors discovered that if you add up these links from a triplewise independent setup, the result doesn't settle into a normal bell curve. Instead, it converges toward a specific, quirky shape called the Variance Gamma (VG) distribution. You can think of the VG distribution as a "super-charged" version of a bell curve that allows for more extreme jumps and a different kind of symmetry. It's the statistical equivalent of a rollercoaster that has a few extra loops compared to the standard one.

The main achievement of this paper is not just noticing this weird shape, but putting a ruler to it. Using a sophisticated mathematical toolkit called Stein's method (which is like a specialized measuring tape for probability distributions) and a technique called the generalized zero-bias transformation, the authors calculated precise error bounds. They determined exactly how close the sum of these triplewise independent variables gets to the Variance Gamma distribution as the number of players ( $m$ ) increases.

Their findings show that as the graph gets bigger (meaning $m$ grows larger), the difference between the actual sum and the Variance Gamma distribution shrinks. Specifically, they proved that the error decreases at a rate proportional to $1/\sqrt{m}$ . In simpler terms, if you double the size of your graph, the approximation gets significantly better, but it follows a predictable, mathematical path. They also provided specific formulas to calculate how far off the approximation might be, depending on how "smooth" the function you are measuring is.

This work is significant because it moves beyond just saying "the bell curve fails." It provides a concrete, mathematical description of what does happen in these triplewise independent scenarios. By establishing these bounds, the authors give statisticians and mathematicians a way to predict and quantify the behavior of these complex systems, ensuring that when the bell curve breaks, we know exactly what shape is taking its place.

Technical Summary: Variance Gamma Approximation to Sums of Triplewise Independent Random Variables

Problem Statement
The Central Limit Theorem (CLT) establishes that standardized partial sums of mutually independent and identically distributed (i.i.d.) random variables converge to a standard normal distribution. However, this convergence relies heavily on mutual independence. It is well-established that mutual independence cannot be relaxed to pairwise independence, nor even to triplewise independence, without potentially losing normality. Specifically, sequences of triplewise independent random variables can be constructed such that their standardized partial sums converge to a non-normal distribution.

Previous research, notably by Beaulieu et al., demonstrated that for a specific construction based on complete bipartite graphs, the standardized partial sums of triplewise independent random variables converge in law to a Variance Gamma (VG) distribution rather than a normal distribution. While this limiting behavior was known, there was a lack of quantitative error bounds measuring the rate of convergence for these approximations. This article addresses the problem of approximating the partial sums of triplewise independent random variables using the Variance Gamma distribution and deriving explicit error bounds for this approximation.

Methodology
The authors employ Stein's method, a probabilistic technique used to bound the distance between two probability distributions, combined with the generalized zero-bias transformation.

Graph-Theoretic Construction: The study utilizes a sequence of complete bipartite graphs $G_m$ with $m$ vertices in each partition. Random variables are assigned to the vertices, and edges are defined based on the equality of these vertex variables. This construction generates a sequence of triplewise independent random variables.
Stein's Method for VG Distribution: The authors utilize the Stein equation specific to the Variance Gamma distribution. For a target variable $Z \sim VG(1, 0, 1, 0)$ , the Stein operator is defined as $A_2f(x) = xf''(x) + f'(x) - xf(x)$ . The solution to this equation, $f_h$ , is analyzed to establish bounds on its derivatives.
Generalized Zero-Bias Transformation: The authors apply the concept of the zero-biased distribution of order $n$ (specifically order 2 in this context). They construct a coupling between the sum of random variables $W_l$ and its zero-biased counterpart $W_l^{*(2)}$ .
Bounding the Distance: By combining the properties of the Stein equation solution with the properties of the zero-bias transformation, the authors bound the Wasserstein and bounded Wasserstein distances between the partial sums and the limiting VG distribution.

Key Contributions and Results

First Error Bounds for Triplewise Independence: The paper claims to be the first work to derive explicit error bounds with respect to probability metrics (Wasserstein and bounded Wasserstein distances) for the approximation of sums of triplewise independent random variables.
Derivation of Bounds:
- The authors establish that for $W_l$ (the standardized sum) and $Z \sim VG(1, 0, 1, 0)$ , the bounded Wasserstein distance satisfies:
  $d_{bW}(W_l, Z) \leq \frac{5}{4\sqrt{l}m} \left( \|h'\| + \frac{9}{2}\|\tilde{h}(x)\| \right)$
- Consequently, the Wasserstein distance is bounded by:
  $d_W(W_l, Z) \leq \frac{495}{32\sqrt{l}m}$
- These results are extended to the case where the limit is $Z \sim VG(l-1, 0, 1, 0)$ , yielding bounds proportional to $\frac{\sqrt{l(l-1)}}{m}$ .
Rate of Convergence: The analysis reveals that the order of convergence for this limiting result is $O(m^{-1})$ (or equivalently $O(n^{-1/2})$ where $n=m^2$ is the number of edges).

Significance and Claims
The authors position this work as a novel contribution to the literature on limit theorems for dependent random variables. While Beaulieu et al. previously established the existence of the convergence to a Variance Gamma distribution for triplewise independent sums, this paper provides the quantitative rate of that convergence.

The paper explicitly states that obtaining the order of convergence $O(m^{-1})$ is a novel result. By providing these bounds, the work moves beyond qualitative counterexamples to the CLT, offering a rigorous framework to measure how closely the sums of triplewise independent variables approximate the Variance Gamma distribution. The methodology relies on the specific connection between triplewise independence and complete bipartite graphs, utilizing the generalized zero-bias transformation to handle the dependencies inherent in the construction.

The Great Statistical Party and the Rule of Three

The Paper's Journey: From Graphs to Gamma

Technical Summary: Variance Gamma Approximation to Sums of Triplewise Independent Random Variables

More like this