Convergence rate of $\ell^p$-energy minimization on graphs: sharp polynomial bounds and a phase transition at $p=3$

TL;DR

This paper analyzes the convergence rates of $ ext{ell}^p$-energy minimization dynamics on graphs, revealing a phase transition at $p=3$ and establishing sharp polynomial bounds that are proven to be optimal.

Contribution

It introduces sharp bounds for convergence times of nonlinear opinion dynamics on graphs, highlighting a novel phase transition at $p=3$ and proving the bounds' optimality.

Findings

Convergence time is at most $n^{eta_p}$ with $eta_p = ext{max}(rac{2p}{p-1},3)$.

The phase transition at $p=3$ significantly affects the convergence rate.

Matching upper and lower bounds are established for convergence time depending on $n$ and average degree.

Abstract

We consider the following dynamics on a connected graph $(V,E)$ with $n$ vertices. Given $p>1$ and an initial opinion profile $f_0:V \to [0,1]$, at each integer step $t \ge 1$ a uniformly random vertex $v=v_t$ is selected, and the opinion there is updated to the value $f_{t}(v)$ that minimizes the sum $\sum_{w \sim v} |f_t(v)-f_{t-1}(w)|^p$ over neighbours $w$ of $v$. The case $p=2$ yields linear averaging dynamics, but for all $p \ne 2$ the dynamics are nonlinear. In the limiting case $p=\infty$ (known as Lipschitz learning), $f_t(v)$ is the average of the largest and smallest values of $f_{t-1}(w)$ among the neighbours $w$ of $v$. We show that the number of steps needed to reduce the oscillation of $f_t$ below $\epsilon$ is at most $n^{\beta_p}$ (up to logarithmic factors in $n$ and $\epsilon$), where $\beta_p:=max(\frac{2p}{p-1},3)$; we prove that the exponent $\beta_p$ is optimal. The phase transition at $p=3$ is a new phenomenon. We also derive matching upper and lower bounds for convergence time as a function of $n$ and the average degree; these are the most challenging to prove.