Distribution of random multiplicative functions in short intervals, with proper normalization

Technical Summary: Distribution of Random Multiplicative Functions in Short Intervals

Problem Statement
The paper investigates the limiting distribution of partial sums of Steinhaus random multiplicative functions, $f(n)$, over short intervals $[x, x+y]$, where $y \to \infty$ but $y = o(x)$. Specifically, the authors study the behavior of the sum $S(x, y) = \sum_{x \le n \le x+y} f(n)$.

Previous work established that for very small $y$ (specifically $y \ll x/(\log x)^c$), the normalized sum $S(x, y)/\sqrt{y}$ converges to a standard complex Gaussian. However, as $y$ approaches $x$, the behavior changes. It was known that for $y$ sufficiently close to $x$ (e.g., $y \ge x/\exp((\log \log x)^{1/2-\epsilon})$), the standard normalization $1/\sqrt{y}$ yields a trivial limiting distribution due to "better than square-root cancellation" effects. The central open question was to determine the limiting distribution for the entire range $y \to \infty, y=o(x)$ and to identify the correct normalization factor that yields a non-degenerate Gaussian limit across this spectrum.

Methodology
The authors employ a sophisticated fusion of martingale techniques and delicate barrier arguments, building upon the framework of Harper, Soundararajan, and Xu. The proof strategy involves several key technical innovations:

1. Conditioning on Small Primes: Instead of applying the martingale central limit theorem (CLT) directly to the full sum, the authors condition on the values of $f(p)$ for primes $p \le z$, where $z = x^{1/\log \log \log x}$. This allows the sum to be decomposed into a sum over integers with large prime factors, weighted by coefficients determined by the small primes.
2. Conditional Variance Analysis: The core of the argument involves analyzing the conditional variance $V_f(x, y)$ of the sum given the small primes. The authors show that $V_f(x, y)$ concentrates around a deterministic quantity $V(x, y)$ with high probability.
3. Fourier Side Reformulation: Using Parseval's identity, the conditional variance is transformed from a physical space sum into an integral of the squared modulus of a random Euler product $F_z(1/2 + it)$ over a frequency range determined by $x/y$.
4. Barrier Events: To handle the heavy-tailed nature of the Euler products (which prevents standard moment methods from working when $y$ is large), the authors introduce "barrier events" $G(t)$ and a refined barrier $H(t)$. These events constrain the growth of partial Euler products at various scales.
   • $G(t)$ ensures the Euler products do not grow too large with high probability.
   • $H(t)$ imposes a stricter constraint, effectively suppressing rare extreme events that would otherwise dominate the variance.
   • The combination of these barriers allows for a concentration estimate of the conditional variance that holds even when $y$ is close to $x$.
5. Fourth Moment Estimates: To apply the martingale CLT, the authors must control the "off-diagonal" contributions to the fourth moment. They utilize contour integral representations and harmonic sums involving the random Euler products, again employing barrier events to manage the integrals and show that off-diagonal terms are negligible compared to the diagonal terms.

Key Contributions and Results

• Theorem 1.1 (Main Result): The authors prove that for any $y \to \infty$ with $y = o(x)$, the normalized sum
  $$\frac{1}{\sqrt{V(x, y)}} \sum_{x \le n \le x+y} f(n)$$
  converges in distribution to a standard complex Gaussian random variable.
• The Normalization Factor $V(x, y)$: A critical contribution is the determination of the scaling factor $V(x, y)$, which differs from the naive $\sqrt{y}$ when $y$ is close to $x$. The factor is given asymptotically by:
  $$V(x, y) \sim \frac{y}{\sqrt{2\pi}} \int_{-\kappa}^{\kappa} e^{-u^2/2} du, \quad \text{where } \kappa = \frac{\log(x/y)}{\sqrt{2 \log \log x}}.$$
  This implies:
  • When $y \ll x/(\log x)^c$, $V(x, y) \sim y$ (recovering the standard Gaussian behavior).
  • When $y$ is close to $x$ (specifically when $\log(x/y) \ll \sqrt{\log \log x}$), $V(x, y) \asymp y \frac{\log(x/y)}{\sqrt{\log \log x}}$.
  • The transition in the normalization factor reflects a "better than square-root cancellation" phenomenon driven by multiplicative chaos effects.
• Contrast with Long Sums: The paper contrasts this result with the behavior of long sums $\sum_{n \le x} f(n)$. While short intervals admit a deterministic normalization leading to a Gaussian limit, the authors (citing recent work by Gorodetsky and Wong) note that long sums do not converge to a Gaussian under any normalization. Instead, the limiting distribution for long sums is non-Gaussian and heavy-tailed, related to the total mass of critical multiplicative chaos.
• Proposition 1.2: The authors establish a lower bound on the tail probabilities of the long sums, showing that $P(|\sum_{n \le x} f(n)| \ge \lambda \sqrt{x}/(\log \log x)^{1/4}) \ge 1/\lambda^A$. This confirms that no normalization can yield a non-degenerate Gaussian limit for the full sum.

Significance
The paper claims to completely determine the distribution of random multiplicative functions over all short intervals $y = o(x)$, resolving the transitions in behavior that were previously only partially understood.

• Novel Normalization: This is the first instance where partial sums of random multiplicative functions are shown to have a Gaussian limiting distribution with a normalization factor that is non-obvious (i.e., not simply the standard deviation of the sum). The factor $V(x, y)$ emerges naturally from the concentration of the conditional variance.
• Structural Insight: The work highlights a fundamental difference between short intervals and long sums. In short intervals, the conditional variance concentrates around a deterministic value, allowing for a Gaussian limit. In long sums, the conditional variance itself is a random variable with a heavy-tailed distribution, preventing Gaussian convergence.
• Methodological Advance: The paper introduces a refined barrier method that successfully handles the heavy-tailedness of the conditional variance in the regime where $y$ is close to $x$. This technique overcomes the limitations of previous fourth-moment methods, which failed when $y$ was large enough that the fourth moment "blows up."

The authors note that while the proof is developed for Steinhaus functions (complex-valued), the strategy is expected to extend to Rademacher functions (real-valued), though some non-trivial technical adjustments regarding translation invariance in law would be required.