Phase Transitions in Gravitational Allocation

TL;DR

This paper investigates the geometric properties and large deviation behaviors of gravitational allocation cells in high-dimensional space, revealing phase transitions and precise tail decay rates for cell diameters and related quantities.

Contribution

It provides an exact characterization of large deviation functions and identifies phase transitions in the geometry of gravitational allocation cells across dimensions.

Findings

Large deviation probabilities decay at a stretched-exponential scale.

Identified phase transitions at specific gamma values in different dimensions.

Established matching lower bounds for cell diameter tail probabilities.

Abstract

Given a Poisson point process of unit masses (``stars'') in dimension d>=3, Newtonian gravity partitions space into domains of attraction (cells) of equal volume. In earlier work, we showed the diameters of these cells have exponential tails. Here we analyze the quantitative geometry of the cells and show that their large deviations occur at the stretched-exponential scale. More precisely, the probability that mass exp(-R^gamma) in a cell travels distance R decays like exp(-R^f_d(gamma)) where we identify the functions f_d exactly. These functions are piecewise smooth and the discontinuities of f_d' represent phase transitions. In dimension d=3, the large deviation is due to a ``distant attracting galaxy'' but a phase transition occurs when f_3(gamma)=1 (at that point, the fluctuations due to individual stars dominate). When d>=5, the large deviation is due to a thin tube (a ``wormhole'') along which the star density increases monotonically, until the point f_d(gamma)=1 (where again fluctuations due to individual stars dominate). In dimension 4 we find a double phase transition, where the transition between low-dimensional behavior (attracting galaxy) and high-dimensional behavior (wormhole) occurs at gamma=4/3. As consequences, we determine the tail behavior of the distance from a star to a uniform point in its cell, and prove a sharp lower bound for the tail probability of the cell's diameter, matching our earlier upper bound.