|
Thema:
|
Estimation of barycenters in metric spaces with curvature bounded above - Zeige Abstract
Abstract: In non-Euclidean metric spaces, barycenters provide a canonical extension of means: they minimize the expected squared distance to a random point. We study the estimation of the barycenter of a probability distribution from iid data. We work under the assumption that the ambient metric space has curvature bounded above in Alexandrov’s sense: roughly speaking, this means that triangles in that space are no thicker than triangles with the same side lengths on a model space with constant curvature (e.g., hyperbolic plane, Euclidean plane or Euclidean sphere). We consider two estimators: the empirical barycenter of the data and a computationally efficient alternative particularly suited to streaming data. If the data distribution is supported on a small enough domain (e.g., a domain smaller than a hemisphere on a Euclidean sphere), we show that our estimators achieve a parametric rate, with error bounds both in expectation and with high probability. This is based on joint works with Jordan Serres (Université Paris-Sorbonne), Yassine Boukhateb (ENSAE) and Austin Stromme (ENSAE).
|