Means in complete manifolds: uniqueness and approximation

TL;DR

This paper proves the almost sure uniqueness of p-means in complete Riemannian manifolds and introduces a simulated annealing method for approximating these means, with convergence guarantees.

Contribution

It establishes the almost everywhere uniqueness of p-means in complete manifolds and develops a convergent simulated annealing approach for their approximation.

Findings

Almost everywhere uniqueness of p-means in complete manifolds.

A convergent simulated annealing algorithm for p-means.

Convergence to the set of minimizers of the distance integral.

Abstract

Let $M$ be a complete Riemannian manifold, $N\in \NN$ and $p\ge 1$. We prove that almost everywhere on $x=(x_1,...,x_N)\in M^N$ for Lebesgue measure in $M^N$, the measure $\di \mu(x)=\f1N\sum_{k=1}^N\d_{x_k}$ has a unique $p$-mean $e_p(x)$. As a consequence, if $X=(X_1,...,X_N)$ is a $M^N$-valued random variable with absolutely continuous law, then almost surely $\mu(X(\om))$ has a unique $p$-mean. In particular if $(X_n)_{n\ge 1}$ is an independent sample of an absolutely continuous law in $M$, then the process $e_{p,n}(\om)=e_p(X_1(\om),..., X_n(\om))$ is well-defined. Assume $M$ is compact and consider a probability measure $\nu$ in $M$. Using partial simulated annealing, we define a continuous semimartingale which converges to the set of minimizers of the integral of distance at power $p$ with respect to $\nu$. When the set is a singleton, it converges to the $p$-mean.