Asymptotics for Push on the Complete Graph

Abstract

We study the popular randomized rumour spreading protocol Push. Initially, a node in a graph possesses some information, which is then spread in a round based manner. In each round, each informed node chooses uniformly at random one of its neighbours and passes the information to it. The central quantity to investigate is the Runtime, that is, the number of rounds needed until every node has received the information. The Push protocol and variations of it have been studied extensively. Here we study the case where the underlying graph is complete with $n$ nodes. Even in this most basic setting, specifying the limiting distribution of the runtime as well as determining related quantities, like its expectation, have remained open problems since the protocol was introduced. In our main result we describe the limiting distribution of the runtime. We show that it does not converge, and that it becomes, after the appropriate normalization, asymptotically periodic both on the $\log_2n$ as well as on the $\ln n$ scale. In particular, the limiting distribution converges only if we restrict ourselves to suitable subsequences of $\mathbb N$, where simultaneously $\log_2 n-\lfloor\log_2n\rfloor\to x$ and $\ln n-\lfloor\ln n\rfloor\to y$ for some fixed $x,y\in [0,1)$. On such subsequences we show that the expected runtime is $\log_2 n+\ln n+h(x,y)+o(1)$, where $h$ is explicitly given and numerically $|\sup h - \inf h| \approx 2\cdot 10^{-4}$.