Phase Transitions for Binomial Sets Under Linear Forms

Abstract: We generalize the study of the sum and difference sets of a subset of $\mathbb{N}$ drawn from a binomial model to the following setting. Given $A \subseteq {0, 1, \dots, N}$, an integer $h \geq 2$, and a linear form $L: \mathbb{Z}^h \to \mathbb{Z}$ given by $$L(x_1, \dots, x_h) = u_1x_1 + \cdots + u_hx_h, \quad u_i \in \mathbb{Z}_{\neq 0} \text{ for all } i \in [h],$$ we study the size of $$L(A) = \left{u_1a_1 + \cdots + u_ha_h : a_i \in A \right}$$ and its complement $L(A)^c$ when each element of ${0, 1, \dots, N}$ is independently included in $A$ with probability $p(N)$. We identify two phase transition phenomena. The first global" phase transition concerns the relative sizes of $L(A)$ and $L(A)^c$, with $p(N) = N^{-\frac{h-1}{h}}$ as the threshold. Asymptotically almost surely, it holds below the threshold that almost all sums generated in $L(A)$ are distinct and almost all possible sums are in $L(A)^c$, and above the threshold that almost all possible sums are in $L(A)$. Our asymptotic formulae substantially extend work of Hegarty and Miller and completely settle, with appropriate corrections made to its statement, their conjecture from 2009. The secondlocal" phase transition concerns the asymptotic behavior of the number of distinct realizations in $L(A)$ of a given value, with $p(N) = N^{-\frac{h-2}{h-1}}$ as the threshold. Specifically, it identifies (in a sharp sense) when the number of such realizations obeys a Poisson limit. Our main tools are recent results concerning the asymptotic enumeration of partitions and weak compositions, classical theorems on Poisson approximation, and the martingale machinery of Kim and Vu.