Random Pairwise Gossip on $CAT(\kappa)$ Metric Spaces

TL;DR

This paper demonstrates that gossip algorithms can be effectively applied to $CAT(\kappa)$ metric spaces, broadening their applicability beyond Riemannian manifolds, with proven convergence and speed results validated by simulations.

Contribution

The paper introduces a simple metric property enabling gossip algorithms to converge on $CAT(\kappa)$ spaces, extending beyond Riemannian manifolds and providing convergence rate analysis.

Findings

Convergence is guaranteed on $CAT(\kappa)$ spaces.

Linear convergence rates are established for $CAT(0)$ spaces.

Local linear rates are shown for $CAT(\kappa)$ spaces with $\kappa > 0$.

Abstract

In the context of sensor networks, gossip algorithms are a popular, well esthablished technique for achieving consensus when sensor data is encoded in linear spaces. Gossip algorithms also have several extensions to non linear data spaces. Most of these extensions deal with Riemannian manifolds and use Riemannian gradient descent. This paper, instead, exhibits a very simple metric property that do not rely on any differential structure. This property strongly suggests that gossip algorithms could be studied on a broader family than Riemannian manifolds. And it turns out that, indeed, (local) convergence is guaranteed as soon as the data space is a mere $CAT(\kappa)$ metric space. We also study convergence speed in this setting and establish linear rates for $CAT(0)$ spaces, and local linear rates for $CAT(\kappa)$ spaces with $\kappa > 0$. Numerical simulations on several scenarii, with corresponding state spaces that are either Riemannian manifolds -- as in the problem of positive definite matrices consensus -- or bare metric spaces -- as in the problem of arms consensus -- validate the results. This shows that not only does our metric approach allows for a simpler and more general mathematical analysis but also paves the way for new kinds of applications that go beyond the Riemannian setting.