🔢 mathematics

Large fluctuations of sums of a random multiplicative function

This paper develops a general framework using quantitative martingale central limit theorems to characterize the large fluctuations of sums of Rademacher or Steinhaus random multiplicative functions over short intervals and polynomial sequences, establishing almost sure lower bounds and matching law of the iterated logarithm upper bounds that extend previous central limit theorem results.

Original authors: Besfort Shala

Published 2026-03-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Besfort Shala

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 understand the behavior of a mysterious, chaotic crowd. This crowd is made up of numbers, and each number has a hidden "personality" that is either positive (+1) or negative (-1), or perhaps a spinning arrow on a circle. These personalities are assigned randomly, but with a strict rule: if two numbers are "friends" (coprime), their personalities multiply together.

This is the world of Random Multiplicative Functions. Mathematicians have long been fascinated by them because they act like a bridge between the rigid, predictable world of prime numbers and the wild, unpredictable world of pure chance.

The paper you provided, written by Besfort Shala, is about figuring out how "loud" this crowd can get when we ask them to shout in unison. Specifically, the author asks: If we sum up these random personalities over a specific group of numbers, how big can the total sum get?

Here is the breakdown of the paper's discoveries using simple analogies:

1. The Main Character: The Random Crowd

Think of the function $f(n)$ as a coin flip for every number $n$ .

Rademacher: A standard coin flip (Heads = +1, Tails = -1).
Steinhaus: A spinning arrow that can point anywhere on a circle.

Usually, if you flip a coin $N$ times, the total sum hovers around zero, with a "typical" size of $\sqrt{N}$ . However, because these numbers are linked by multiplication rules (multiplicativity), they don't behave like independent coin flips. They have a hidden structure that sometimes cancels them out more than usual, making the typical sum smaller than expected.

2. The Two Big Questions

The paper investigates what happens when we look at the extreme outliers—the moments when the crowd gets incredibly loud (large fluctuations). The author focuses on two specific scenarios:

Scenario A: The Polynomial Party (Polynomial Images)

Imagine you only let numbers into the party if they are the result of a polynomial formula, like $n^2 + 1$ or $n^3 - 2$ .

The Old View: Previous research showed that for these specific numbers, the sum behaves somewhat like a normal bell curve (Gaussian distribution).
The New Discovery: Shala proves that even though the average behavior is calm, there are rare moments where the sum spikes up to a size of $\sqrt{N \log \log N}$ .
The Analogy: Imagine a calm lake (the average). Usually, the waves are small. But Shala proved that if you wait long enough, you will almost certainly see a massive, freak wave that is slightly taller than the standard "maximum wave" you'd expect from pure randomness.

Scenario B: The Short Interval Sprint (Short Intervals)

Instead of looking at numbers from 1 to $N$ , imagine looking at a very short sprint, like numbers from $N-H$ to $N$ (where $H$ is much smaller than $N$ ).

The Surprise: In these short sprints, the "noise" of the random function is actually louder relative to the size of the group than you might think.
The Discovery: The author shows that in these short intervals, the sum can fluctuate by a factor of $\sqrt{H \log(N/H)}$ .
The Analogy: Think of a short, crowded hallway. If you ask 10 people to shout, it's loud. If you ask 1,000 people in a stadium, it's louder, but the density of the noise in the hallway feels more intense and chaotic because the group is so small and the "randomness" hasn't had time to smooth itself out. Shala proved that these short bursts of noise are surprisingly powerful.

3. The Secret Weapon: The "Martingale" Ladder

How did Shala prove this? He didn't just count; he used a mathematical tool called a Martingale.

The Metaphor: Imagine climbing a ladder where each rung represents a prime number (2, 3, 5, 7...). As you climb, you add the "personality" of the numbers divisible by that prime.
Because the function is multiplicative, the value of the sum at any rung depends only on the previous rungs and the new prime you just added. This creates a "fair game" structure (a martingale).
Shala used a sophisticated version of the Central Limit Theorem (the math behind the bell curve) specifically designed for these ladders. He showed that if you look at many different "ladders" (different ranges of numbers) at the same time, they behave like a group of independent, dancing Gaussian clouds.

4. The "Kantorovich-Wasserstein" Distance

This is a fancy term for a way to measure how "far apart" two probability distributions are.

The Analogy: Imagine you have a pile of sand (the real data) and a perfect mold of a sandcastle (the ideal Gaussian curve). The "distance" is how much sand you have to move to turn the pile into the castle.
Shala used a new, sharper tool to measure this distance. This allowed him to prove that the real random sums are so close to the ideal mathematical curves that he could predict exactly how often the "freak waves" (large fluctuations) would happen.

5. Why Does This Matter?

For Math: It solves a puzzle about how "random" these number patterns really are. It shows that even with strict multiplication rules, the chaos eventually wins out, but in a very specific, measurable way.
For the Future: The methods developed here are like a new, high-powered microscope. They can be used to study other complex number patterns, like sums of squares or shifted primes, helping mathematicians understand the deep, hidden rhythms of the number system.

Summary

In simple terms, Besfort Shala's paper says: "Even in a world of random numbers with strict rules, if you wait long enough and look at the right groups, you will see massive, predictable spikes in the noise. We have built a new mathematical telescope to see exactly how big those spikes are and how often they happen."

1. Problem Statement and Context

The paper investigates the almost sure large fluctuations of partial sums of random multiplicative functions (RMFs) restricted to specific arithmetic subsets.

Random Multiplicative Functions: The study focuses on two types:
- Rademacher RMF: $f(p) \in \{-1, 1\}$ uniformly, extended multiplicatively to square-free integers.
- Steinhaus RMF: $f(p)$ is uniformly distributed on the unit circle $\mathbb{T}$ , extended multiplicatively.
The Core Question: While the Central Limit Theorem (CLT) for full sums $\sum_{n \le N} f(n)$ has been established (with non-standard normalization due to Harper's work showing typical size $\sqrt{N}/(\log \log N)^{1/4}$ ), this paper addresses the Law of the Iterated Logarithm (LIL) type behavior. Specifically, it seeks to determine the magnitude of the maximum fluctuations (lim sup) of these sums over specific subsets $A \subseteq [1, N]$ .
Target Subsets: The paper focuses on two primary cases where the distribution of sums is approximately Gaussian:
1. Polynomial Images: Sums of the form $\sum_{n \le N} f(P(n))$ .
2. Short Intervals: Sums of the form $\sum_{N-H < n \le N} f(n)$ .

2. Methodology

The author develops a general framework to establish almost sure lower bounds for large fluctuations, extending the CLT machinery of Soundararajan and Xu. The methodology relies on three main pillars:

A. Martingale Structure

The sums are viewed as martingales with respect to the filtration induced by the largest prime factor ( $P^+(n)$ ).
$\sum_{n \in A} f(n) = \sum_{p} f(p) \sum_{\substack{n \in A \\ P^+(n)=p}} f\left(\frac{n}{p}\right)$
This structure allows the application of quantitative martingale CLTs (specifically Hall and Heyde's bound on the Kolmogorov distance) to approximate the distribution of the sums by a Gaussian.

B. Quantitative Multivariate CLT via Kantorovich–Wasserstein Distance

To prove almost sure lower bounds, the author must show that the sum attains large values frequently. This requires analyzing the joint distribution of sums over multiple scales ( $N_1, N_2, \dots, N_k$ ).

Instead of using Stein's method (which struggles with complex dependencies between large primes), the paper employs a recent result by Bobkov and Götze.
This result bounds the Kantorovich–Wasserstein distance between a multivariate distribution and a Gaussian vector using the Kolmogorov distances of their linear combinations.
This allows the author to bootstrap univariate martingale CLTs into a quantitative multivariate CLT, establishing that the vector of normalized sums is close to a multivariate Gaussian with a specific covariance structure.

C. Arithmetic Input (Fourth Moment Equations)

The quality of the Gaussian approximation depends on controlling the number of "non-trivial" solutions to specific Diophantine equations (the fourth moment equations):

Rademacher: $n_1 n_2 n_3 n_4 = \square$ (perfect square).
Steinhaus: $n_1 n_2 = n_3 n_4$ .
The author must prove that for the specific subsets $A$ , the number of non-trivial solutions is significantly smaller than the trivial solutions (where variables pair up). This involves:
Polynomials: Using results from Klurman, Shkredov, Xu, and Chinis to bound solutions for $P(n_i)$ .
Short Intervals: Conditioning on small primes and using results on smooth integers to isolate large prime factors, thereby reducing the problem to counting solutions in a restricted range.

3. Key Contributions and Results

Theorem A: Polynomial Images

For a polynomial $P \in \mathbb{Z}[x]$ satisfying specific irreducibility or factorization conditions:

Upper Bound: Almost surely,
$\left| \sum_{n \le N} f(P(n)) \right| \ll \sqrt{N \log \log N}$
This is a new result for the upper bound in the Rademacher case and extends previous work for Steinhaus.
Lower Bound: Almost surely, there exists a sequence $N_k \to \infty$ such that
$\left| \sum_{n \le N_k} f(P(n)) \right| \gg \sqrt{N_k \log \log N_k}$
Significance: This extends previous work to polynomials that split into linear factors over $\mathbb{Q}$ (e.g., $P(n) = n(n+1)$ ), a case where previous combinatorial methods failed because all prime factors are small. The new method fully utilizes the martingale structure without combinatorial splitting.

Theorem B: Short Intervals

For a short interval $(N-H, N]$ where $H = H(N)$ satisfies $N^{11/15} \le H \le N/(\log N)^c$ (with $c > 2\log 2 - 1$ ):

Result: Almost surely, there exists a sequence $N_k \to \infty$ such that
$\left| \sum_{N_k - H_k < n \le N_k} f(n) \right| \gg \sqrt{H_k \log \left(\frac{N_k}{H_k}\right)}$
Significance: This establishes the correct order of magnitude for the fluctuations in short intervals. The factor $\sqrt{\log(N/H)}$ arises because short sums of length $H$ decorrelate faster; roughly $N/H$ independent sums fit in a window of size $N$ , leading to a maximum of size $\sqrt{\log(N/H)}$ rather than $\sqrt{\log N}$ .

4. Technical Innovations

Handling Linear Factors: Previous methods for polynomial sums relied on the existence of a large prime factor (degree $\ge 2$ ). This paper introduces a method that works even when $P(n)$ is a product of linear factors, where all prime factors are small ( $O(N)$ ).
Conditioning on Small Primes (Short Intervals): To handle the weak bounds on non-trivial solutions in short intervals, the author conditions on the values of $f(p)$ for small primes. This reduces the problem to a weighted sum of independent variables for large primes, bypassing the need for strong fourth-moment bounds on the entire sum.
Bobkov-Götze Application: The successful application of the Bobkov-Götze theorem to number-theoretic sums provides a robust alternative to Stein's method for establishing multivariate Gaussian approximations in the presence of complex arithmetic dependencies.

5. Significance and Impact

Resolution of LIL-type Questions: The paper provides the first almost sure lower bounds matching the expected order of magnitude for these specific arithmetic subsets, complementing existing upper bounds.
General Framework: The developed framework is flexible and can be applied to other subsets (e.g., shifted primes, sums of squares) provided the necessary arithmetic bounds on the fourth moment equations can be established.
Methodological Shift: By moving away from combinatorial splitting and towards a pure martingale approach combined with modern probability inequalities (Kantorovich–Wasserstein), the paper opens new avenues for studying random multiplicative functions in regimes where traditional independence assumptions fail.

In summary, Besfort Shala's work bridges the gap between the Central Limit Theorem and the Law of the Iterated Logarithm for random multiplicative functions, proving that despite their multiplicative structure, these sums exhibit Gaussian-like large fluctuations when restricted to "less structured" sets like short intervals or polynomial images.