Asymptotic Tail of the Product of Independent Poisson Random Variables

This paper derives an explicit Laplace-type asymptotic approximation with vanishing relative error for the tail probability of the product of independent Poisson random variables, utilizing Stirling's approximation, a constrained saddle-point method, and the Lambert $W$ function.

The Gist

Imagine you are trying to predict the weather, but instead of rain or sunshine, you are tracking the "luck" of a group of people.

This paper is about a specific mathematical puzzle: What happens when you multiply the "luck" of several independent people together?

In math, "luck" is often modeled by something called a Poisson distribution. Think of a Poisson variable as a counter for random events. For example:

• How many emails you get in an hour.
• How many cars pass your house in 10 minutes.
• How many typos a writer makes in a day.

These events happen randomly, but on average, they happen at a steady rate.

The Problem: The "Multiplication Monster"

Usually, in probability, we are good at adding things up. If you add the number of emails from three different people, the result is predictable and follows a nice, bell-shaped curve (the Central Limit Theorem).

But this paper looks at multiplying them.
If Person A gets 5 emails, Person B gets 5, and Person C gets 5, the product is $5 \times 5 \times 5 = 125$.
But what if Person A gets 100? Then the product explodes to $100 \times 5 \times 5 = 2,500$.

The authors wanted to answer a very specific question: If we multiply $m$ of these random counters together, how likely is it that the final number is huge?

In everyday language: If I multiply the daily typos of 5 different writers, what are the odds that the total product of their typos is a number so big it would make a mathematician sweat?

The Discovery: The "One Big Fish" Effect

The authors found something surprising. When you multiply random numbers, the result is usually dominated by one single, unusually large number.

Imagine a school of fish. If you multiply their sizes, the total "fishiness" is mostly determined by the one giant whale in the group, not the thousands of tiny minnows.

Because of this, the "tail" of the probability (the chance of getting a massive number) is much "heavier" than you would expect. It's not just a little bit more likely; it's a completely different beast.

The Solution: The "Saddle-Point" Hike

To figure out exactly how heavy this tail is, the authors had to use some very fancy mathematical tools. They didn't just guess; they went on a mathematical hike.

1. The Landscape (Stirling's Approximation): They first smoothed out the jagged, bumpy terrain of the math (using something called Stirling's approximation) so they could see the big picture.
2. The Saddle (The Saddle-Point Method): Imagine a mountain pass (a saddle) between two peaks. To find the most likely way to get a huge product, you have to find the "highest point" on the path where the numbers multiply to equal your target.
   • The authors realized that to get a product of $n$, the individual numbers usually need to be roughly equal (like $\sqrt{n}$ if you have two numbers).
   • They used a tool called the Lambert W function (think of it as a special key that unlocks complex equations) to find the exact coordinates of this "saddle point."
3. The Gaussian Prefactor (The Crowd): Once they found the peak, they had to count how many "paths" led there. This is like counting how many hikers are standing on that specific mountain pass. This gave them a precise multiplier for their answer.

The Big Result

The paper gives a formula that tells you the probability of this massive product happening.

• For 2 people: The chance of the product being huge drops off like $e^{-\sqrt{n} \log n}$.
  • Analogy: If you want to know the odds of two random numbers multiplying to a trillion, the odds are incredibly small, but not impossibly small. They drop off slower than you'd think.
• For $m$ people: The formula changes to $e^{-n^{1/m} \log n}$.
  • The Twist: As you add more people to the multiplication, the "tail" gets even heavier. The probability of a massive number happening actually increases relative to the single-variable case. The more variables you multiply, the more likely it is that one of them will be a "giant" and blow up the whole product.

Why Does This Matter?

You might ask, "Who multiplies random typos together?"

This isn't just about typos. This math applies to:

• Finance: If you have a chain of investments where each one has a random return, the final wealth is a product of all those returns. Understanding the "tail" helps you understand the risk of a massive crash or a massive windfall.
• Physics: In particle physics, sometimes you multiply probabilities of independent events.
• Risk Management: Understanding that "one big factor" can dominate a system helps engineers and bankers build better safety nets.

The Takeaway

The authors showed that when you multiply random things, you can't just use the old rules of addition. You have to look for the "Saddle Point"—the specific balance where the numbers align to create a giant product.

They proved that while getting a massive product is rare, it's much more common than standard intuition suggests, because nature loves to throw in one "giant fish" that dominates the whole school. Their formula allows us to calculate exactly how rare (or common) that giant fish is.

Technical Summary

Here is a detailed technical summary of the paper "Asymptotic Tail of the Product of Independent Poisson Random Variables" by Džiugas Chvoinikov and Jonas Šiaulys.

1. Problem Statement

The paper addresses a gap in probability theory regarding the tail behavior of products of independent discrete random variables. While the asymptotic behavior of sums is well-understood (via the Central Limit Theorem and Large Deviation Theory), products are significantly more complex because they do not preserve the structure of the original distributions and often lack closed-form expressions.

Specifically, the authors investigate the right-tail probability:
$$P(Z_m \geq n) = P\left(\prod_{j=1}^m X_j \geq n\right)$$
where $X_1, \dots, X_m$ are independent Poisson random variables with parameters $\lambda_1, \dots, \lambda_m > 0$. The goal is to derive an explicit, precise asymptotic approximation for this probability as $n \to \infty$, with a relative error that tends to zero.

2. Methodology

The authors employ a sophisticated combination of analytical tools to handle the discrete nature of the Poisson distribution and the multiplicative constraint:

• Stirling's Logarithmic Approximation: The Poisson probability mass function is approximated using the refined Stirling formula:
  $$\log k! = k \log k - k + \frac{1}{2}\log(2\pi k) + O(1/k)$$
  This transforms the discrete summation into a continuous optimization problem involving a function $T(k, \ell)$.
• Constrained Saddle-Point Method: The tail probability is dominated by the maximum of the joint log-probability subject to the constraint $k_1 \dots k_m = n$. The authors use Lagrange multipliers to find the saddle point $(k^*, \ell^*)$ (or $k^*_1, \dots, k^*_m$ for the general case).
• Lambert W Function: The system of equations derived from the Lagrange multipliers is solved explicitly using the Lambert W function (the inverse of $f(y) = ye^y$). This allows for an exact representation of the saddle-point coordinates.
• Multidimensional Laplace Approximation: Once the saddle point is identified, the sum is approximated by an integral over the constraint manifold. The authors calculate the Gaussian prefactor by evaluating the restricted Hessian (the second derivative matrix projected onto the tangent space of the constraint).
• Regime Decomposition: For the two-variable case ($m=2$), the authors rigorously prove that the dominant contribution comes from the "balanced" regime where $k \approx \ell \approx \sqrt{n}$. They show that "unbalanced" regimes (where one variable is small and the other is very large) contribute negligibly to the tail probability.

3. Key Contributions and Results

A. The Two-Variable Case ($m=2$)

The paper provides two main theorems for $Z = XY$ where $X \sim \text{Pois}(\lambda_1)$ and $Y \sim \text{Pois}(\lambda_2)$.

1. Refined Asymptotic Formula (Theorem 2.1):
   The authors derive an exact Laplace-type approximation:
   $$P(XY \geq n) \sim \sqrt{2\pi} n^{1/4} \exp\left( T(k^*_n, \ell^*_n) \right)$$
   where $T$ is the log-probability function evaluated at the unique saddle point $(k^*_n, \ell^*_n)$, which is defined implicitly via the Lambert W function. This formula includes all lower-order terms necessary for high precision.

2. Leading-Order Behavior (Theorem 2.2):
   A simplified asymptotic form is derived:
   $$P(XY \geq n) = \sqrt{2\pi} n^{1/4} \exp\left( -\sqrt{n} \log n + O(\sqrt{n}) \right)$$
   Crucially, the logarithm of the tail probability behaves as:
   $$\log P(XY \geq n) = -\sqrt{n} \log n + O(\sqrt{n})$$
   This indicates a stretched-exponential decay, which is significantly heavier than the tail of a single Poisson variable (which decays exponentially).

3. Necessity of Exact Saddle Points:
   A critical theoretical finding is that finite truncations of the saddle-point expansion (e.g., approximating $k^*$ simply as $\sqrt{n}$) result in a non-vanishing error in the exponent. To achieve an $o(1)$ error in the log-probability, the exact saddle point (solved via Lambert W) must be used.

B. Generalization to $m$ Variables

The results are extended to the product of $m$ independent Poisson variables ($Z_m = \prod_{j=1}^m X_j$).

• General Asymptotic Rate (Theorem 5.1):
  The tail probability decays as:
  $$\log P(Z_m \geq n) \sim -n^{1/m} \log n$$
  The prefactor scales as $(2\pi)^{(m-1)/2} n^{(m-1)/(2m)}$.
• Dimensional Effect: As the dimension $m$ increases, the decay rate slows down ($n^{1/m}$ becomes smaller), resulting in a "heavier" tail.

4. Numerical Validation

The authors validate their theoretical findings through numerical experiments:

• Accuracy: The Laplace approximation evaluated at the numerically computed saddle point matches the exact tail probability (computed via direct summation) with high accuracy for moderate to large $n$.
• Efficiency: The saddle-point method is computationally more efficient than direct summation for large $n$.
• Truncation Analysis: Numerical comparisons show that while the leading term $-\sqrt{n}\log n$ is insufficient, keeping the first two terms (the dominant term and the $\sqrt{n}$ correction) provides a highly accurate approximation for practical purposes.

5. Significance

• Filling a Literature Gap: Prior to this work, asymptotic results for products of independent discrete light-tailed variables were nonexistent. Previous literature focused on continuous distributions (Gaussian, Gamma, Cauchy).
• Tail Behavior Insight: The paper demonstrates that multiplying light-tailed variables (like Poisson) fundamentally alters the tail class, creating a heavy-tailed (stretched-exponential) distribution. This has implications for risk modeling where multiplicative interactions occur.
• Methodological Rigor: The paper provides a rigorous framework for handling discrete constraints in saddle-point analysis, specifically addressing the subtleties of the Lambert W function and the necessity of exact saddle-point evaluation to avoid exponential errors.

In summary, this paper establishes a precise mathematical framework for analyzing the extreme tails of products of Poisson variables, revealing that their decay is governed by a stretched-exponential law determined by the dimension of the product.