Escaping Chaos in Random Multiplicative Functions

Technical Summary: Escaping Chaos in Random Multiplicative Functions

Problem Statement
The paper addresses a fundamental question regarding the behavior of Steinhaus random multiplicative functions, $f(n)$. While $f(n)$ is defined by independent, identically distributed values on the complex unit circle for prime arguments, the complete multiplicativity of $f(n)$ introduces strong global dependencies. Consequently, the full partial sums $\frac{1}{\sqrt{x}}\sum_{1 \le n \le x} f(n)$ do not obey a standard Central Limit Theorem (CLT); they fail to converge to a complex normal distribution $\mathcal{CN}(0, 1)$. Instead, the limiting distribution is non-Gaussian, characterized by a "random variance" $V_\infty$ and a "Gaussian Multiplicative Chaos" effect, with magnitudes scaling as $\sqrt{x}(\log \log x)^{1/4}$ rather than $\sqrt{x}$.

The central problem investigated is: What is the largest possible size of a subset $A \subset [1, N]$ such that the normalized sum over $A$ recovers the standard CLT? Specifically, under what conditions does the convergence
$$\frac{1}{\sqrt{|A|}} \sum_{n \in A} f(n) \xrightarrow{d} \mathcal{CN}(0, 1)$$
hold? The goal is to determine the threshold density of $A$ required to "escape" the chaotic global dependence inherent in $f(n)$.

Methodology
The author employs a combination of probabilistic heuristics, moment analysis, and martingale central limit theorems.

1. Heuristic Decomposition: The analysis begins by modeling $A$ as a random set with density $\rho$. By decomposing the indicator function $1_A(n)$ into a mean component $\rho$ and a fluctuation component $w(n) = 1_A(n) - \rho$, the sum is split into two parts:
   $$\sum_{n \in A} f(n) = \sum_{n \le N} w(n)f(n) + \rho \sum_{n \le N} f(n).$$
   Leveraging Harper's theorem, the second term (the global sum) is shown to be negligible relative to the normalization factor when $\rho$ is not too close to 1. The problem thus reduces to analyzing the weighted sum $\sum w(n)f(n)$.

2. Moment Convergence and Contradiction (Theorem 1): To prove that $A$ cannot have positive density for the standard normalization, the author analyzes the second moment of the normalized sum. If $|A| = \rho N$ with $\rho \gg 1$, the second moment of the normalized sum is shown to be $1-\rho$, which is strictly less than 1. Since convergence in distribution to $\mathcal{CN}(0, 1)$ generally requires the second moment to converge to 1 (or at least not be strictly bounded away from it in a way that contradicts the limit), this discrepancy proves that the standard CLT fails for sets of positive density.

3. Martingale Central Limit Theorem (Theorem 2): For the case where the density $\rho$ is small but non-zero, the author establishes the correct normalization. The proof utilizes a weighted version of McLeish's martingale central limit theorem (as established in Soundararajan and Xu [10]). The strategy involves:

   • Conditioning on the random weights $w(n)$.
   • Verifying specific "multiplicative energy" type conditions involving sums over quadruples $(n_1, n_2, m_1, m_2)$ satisfying $n_1 n_2 = m_1 m_2$.
   • Using concentration inequalities to show that these conditions hold for most deterministic sets $A$ with density $\rho$, rather than just random ones.

Key Results

• Theorem 1 (Non-existence of Positive Density): If $A \subset [1, N]$ is a sequence of sets such that the normalized sum converges to $\mathcal{CN}(0, 1)$, then the size of $A$ must be $o(N)$. In other words, a set with positive asymptotic density cannot satisfy the standard CLT with the standard normalization $\frac{1}{\sqrt{|A|}}$.
• Theorem 2 (Correct Normalization for Positive Density): For sets $A$ with density $\rho$ satisfying the condition $(1-\rho)^{-1} = o((\log \log N)^{1/2})$, the standard CLT holds if the normalization factor is adjusted. Specifically, for most sets $A$ with such density:
  $$\frac{1}{\sqrt{(1-\rho)|A|}} \sum_{n \in A} f(n) \xrightarrow{d} \mathcal{CN}(0, 1).$$
  The factor $\sqrt{1-\rho}$ is essential to compensate for the reduction in variance caused by the "mean subtraction" inherent in the decomposition when $\rho$ is non-negligible.

Significance and Claims
The paper claims to solve the problem of determining the maximal size of $A$ for which the "naive" CLT holds. The primary significance lies in establishing that:

1. Density Threshold: The threshold for the ideal CLT (with standard normalization) is strictly sub-linear density ($\rho = o(1)$).
2. Sharpness of Density: The condition $(1-\rho)^{-1} = o((\log \log N)^{1/2})$ is sharp. If the density is too high (closer to 1), the limiting distribution reverts to the non-Gaussian behavior observed in the full sum, regardless of normalization.
3. Correction Factor: The paper identifies and confirms the necessity of the $\sqrt{1-\rho}$ correction factor. This factor is trivial when $\rho \to 0$ but becomes a non-trivial multiplicative constant as $\rho$ increases, bridging the gap between the behavior of sparse sets and dense sets.

The work clarifies the boundary between "chaotic" global behavior and "Gaussian" local behavior in random multiplicative functions, demonstrating that while global dependence prevents a standard CLT for the full sum, carefully chosen subsets (or subsets with specific density adjustments) can effectively "escape" this chaos.