The critical-window profile for $d_k$ in short intervals

Technical Summary: The Critical-Window Profile for the $k$-Fold Divisor Function in Almost All Short Intervals

Problem Statement
The paper addresses the problem of comparing local averages of the $k$-fold divisor function $d_k(n)$ with their long averages in short intervals. While the celebrated Matomäki–Radziwiłł theorem establishes that bounded multiplicative functions exhibit expected averages on almost all short intervals of length tending to infinity, unbounded functions like $d_k(n)$ present a different threshold phenomenon. The central difficulty lies in the "heavy tail" of $d_k(n)$, where integers with many prime factors carry disproportionately large weight, rendering bounded-function theories insufficient.

The specific goal is to determine the almost-all transition scale for $d_k(n)$. Previous work by Mangerel established that the long average is recovered when the interval length $h$ satisfies $h / (\log X)^{(k-1)^2} \to \infty$. This paper seeks to identify the precise critical scale and the transition profile below this threshold, specifically investigating intervals where $h$ is close to the critical length $D_k(X) = (\log X)^{k \log k - k + 1}$.

Methodology
The proof relies on two primary technical ingredients, combining probabilistic distributional results with analytic mean-value estimates:

1. Weighted Erdős–Kac Theorem: The author utilizes a central limit theorem for the function $\Omega(n)$ (the number of prime factors counted with multiplicity) under the probability measure weighted by $d_k(n)$. Under this measure, the normalized variable $\frac{\Omega(n) - k \log \log X}{\sqrt{k \log \log X}}$ converges to a standard Gaussian distribution. This provides the distributional input necessary to relate the interval length to the proportion of integers satisfying specific $\Omega(n)$ cutoffs.

2. $\Omega$-Truncated Matomäki–Radziwiłł Estimates: The analytic core involves establishing mean-square estimates for Dirichlet polynomials truncated by the condition $\Omega(n) \le L$.

   • Ramaré Decomposition: To handle the non-multiplicative nature of the truncation, the author employs a Ramaré decomposition. Crucially, they do not treat the $\Omega$-cutoff as a negligible tail in the critical range. Instead, they apply Cauchy's formula to the residual condition $\Omega(m) \le L-1$ only after the decomposition, converting the cutoff into multiplicative weights $z^{-\Omega(m)}$.
   • Halász Distance and Large Values: The remaining large-prime layer in the decomposition forces a uniform high-frequency Halász distance. This mechanism supplies the necessary logarithmic saving in the final large-value range, which a direct estimate of the tail would fail to provide.
   • Mean-Value Bounds: The proof utilizes a combination of Montgomery–Vaughan mean-value estimates, Shiu-type estimates for divisor functions, and Halász-type large-value bounds for prime polynomials to control the $L^2$ norm of the truncated sums.

Key Results
The paper establishes a precise transition profile for the average of $d_k(n)$ in short intervals $[x, x+h]$ for almost all $x \in [X, 2X]$.

• Critical Scale: The critical length is identified as $D_k(X) = (\log X)^{k \log k - k + 1}$.
• Transition Profile (Theorem 1.1): Let $\ell = \log \log X$ and define the normalized parameter $A(X) = \frac{1}{\sqrt{\ell}} \log \frac{h}{D_k(X)}$. If $A(X) = O(1)$, then for almost all $x$:
  $$\frac{1}{h} \sum_{x < n \le x+h} d_k(n) = \left( \Phi_G\left( \frac{A(X)}{\sqrt{k \log k}} \right) + o(1) \right) M_k(x)$$
  where $\Phi_G$ is the standard Gaussian distribution function and $M_k(x)$ is the long average. This formula describes the "finite part" of the transition, showing that the local average is a fraction of the global average determined by the Gaussian cumulative distribution function.
• Endpoint Regime (Corollary 1.2): If the interval length satisfies $\frac{\log(h/D_k(X))}{\sqrt{\ell}} \to +\infty$ (i.e., $A(X) \to +\infty$), the Gaussian factor tends to 1, and the full long average is recovered for almost all $x$.

Significance and Claims
The paper claims to provide a power-saving improvement over previous results (specifically Mangerel's exponent) regarding the length of intervals required to recover the long average of the divisor function. By identifying the exact transition profile, the work bridges the gap between the bounded-function theory and the behavior of unbounded multiplicative functions at polylogarithmic scales.

The author emphasizes that the transition is driven by the interplay between the interval length and the distribution of $\Omega(n)$ under the $d_k$-weight. The technical novelty lies in the handling of the $\Omega$-cutoff: rather than discarding it as a tail, the author integrates it into the analytic machinery via Ramaré decomposition and Cauchy's formula, allowing for a uniform high-frequency saving that is essential for the critical range. The results are presented as an "almost-all" transition theorem, consistent with the probabilistic nature of the Matomäki–Radziwiłł method.