[Paper Review] Small-scale equidistribution for random spherical harmonics

Matthew de Courcy-Ireland|arXiv (Cornell University)|Nov 1, 2017

Analytic Number Theory Research20 references3 citations

TL;DR

This paper establishes small-scale equidistribution for random spherical harmonics of degree $m$ on $S^2$, showing that the $L^2$-mass of such functions equidistributes over spherical caps of radius $r$ as $m \to \infty$, provided $rm / \log m \to \infty$. The key result is that the discrepancy between the normalized $L^2$-mass and uniform measure converges to zero in probability at scales $r \gg \log m / m$, revealing a quantum mechanical limit at the Planck scale $1/m$. The proof combines variance estimates, tail bounds for quadratic forms in Gaussians, and a union bound over a fine grid on the sphere, leveraging exponential concentration to overcome combinatorial losses.

ABSTRACT

We study random spherical harmonics at shrinking scales. We compare the mass assigned to a small spherical cap with its area, and find the smallest possible scale at which, with high probability, the discrepancy between them is small simultaneously at every point on the sphere.

Motivation & Objective

• To understand the small-scale distribution of $L^2$-mass of random spherical harmonics on $S^2$.
• To determine the smallest scale $r$ at which the $L^2$-mass of a random harmonic is uniformly distributed across spherical caps of radius $r$, uniformly over all centers.
• To establish a sharp threshold in terms of $r$ and $m$ such that the discrepancy between the normalized mass and uniform measure vanishes in probability.
• To analyze the tail behavior of the random variable $X_z = \frac{1}{\operatorname{vol}(B_r)}\int_{B_r(z)} \phi^2$ at a fixed point $z$, and use this to control the supremum over all $z \in S^2$.

Proposed method

• Model random spherical harmonics $\phi$ as a linear combination of orthonormal degree-$m$ spherical harmonics with i.i.d. $N(0, 1/(2m+1))$ coefficients.
• Define the normalized local $L^2$-mass $X_z = \frac{1}{\operatorname{vol}(B_r)}\int_{B_r(z)} \phi^2$, with $\mathbb{E}[X_z] = 1/(4\pi)$.
• Express $X_z$ as a quadratic form $\sum \lambda_\nu \mathfrak{z}_\nu^2$ in independent standard Gaussians $\mathfrak{z}_\nu$, where $\lambda_\nu$ are eigenvalues derived from Bessel integrals.
• Establish exponential tail bounds for $X_z$ using characteristic function methods and contour deformation, showing $\mathbb{P}\left(\left|X_z - \frac{1}{4\pi}\right| > \epsilon\right) \leq C(\epsilon) e^{-c(\epsilon) rm}$ with $c(\epsilon) \sim \epsilon^2$.
• Use a fine grid on $S^2$ and a union bound over grid points to control the supremum discrepancy $D(r,m) = \sup_z |X_z - 1/(4\pi)|$, leveraging the exponential tail to absorb the $m^2$ grid points.
• Show that the metric $d(z,z') = \sqrt{\mathbb{E}[(X_z - X_{z'})^2]}$ decays with distance, enabling covering number estimates, and suggest that Dudley’s entropy method could yield an alternative proof.

Experimental results

Research questions

• RQ1 What is the smallest scale $r$ at which the $L^2$-mass of a random spherical harmonic of degree $m$ equidistributes uniformly over all spherical caps of radius $r$?
• RQ2 How fast must $r$ decay with $m$ for the discrepancy $D(r,m)$ to converge to zero in probability?
• RQ3 What is the precise tail behavior of the local $L^2$-mass $X_z$ at a fixed point $z$, and how does it depend on $r$ and $m$?
• RQ4 Can the supremum of $|X_z - \mathbb{E}[X_z]|$ over $z \in S^2$ be controlled using a union bound over a fine grid, given the tail decay of $X_z$?
• RQ5 Is the Planck scale $1/m$ the fundamental limit below which equidistribution fails, and why?

Key findings

• The discrepancy $D(r,m) = \sup_z \left| \frac{1}{\operatorname{vol}(B_r)}\int_{B_r(z)} \phi^2 - \frac{1}{4\pi} \right|$ converges to zero in probability as $m \to \infty$ provided $rm / \log m \to \infty$.
• The tail bound $\mathbb{P}\left(\left|X_z - \frac{1}{4\pi}\right| > \epsilon\right) \leq C(\epsilon) e^{-c(\epsilon) rm}$ holds with $c(\epsilon) \sim \epsilon^2$, showing exponential decay in $rm$.
• The variance of $X_z$ is of order $1/(rm)$, and the largest eigenvalue $\lambda_{\max}$ of the quadratic form is also $\sim 1/(rm)$, confirming the scaling.
• For $k \sim rm$, the eigenvalues $\lambda_k$ decay as $\sim (rm)^{-4/3 + \eta}$, and for large $k$, they decay super-exponentially.
• The union bound over a grid of size $\sim m^2$ is effective because the exponential tail $e^{-c(\epsilon) rm}$ dominates any polynomial loss, so long as $rm \gg \log m$.
• The result implies that equidistribution fails below the Planck scale $1/m$, as the mass is not uniformly distributed at scales $r \ll 1/m$, due to quantum effects.