Detailseite
Projekt Druckansicht

Verschmierte Grenzwertsätze

Fachliche Zuordnung Mathematik
Förderung Förderung von 2019 bis 2024
Projektkennung Deutsche Forschungsgemeinschaft (DFG) - Projektnummer 427894948
 
Erstellungsjahr 2024

Zusammenfassung der Projektergebnisse

In spite of limited funding and the pandemic hitting, with the new general CLT for Fréchet means on singular spaces (formulation and proof require approx. 100 pages or, equivalently, four papers), major progress has been achieved. It is the nature of this CLT that it has direct application for stickiness and this has spurred a separate investigation of various flavors (types) of stickiness and their relationship with respect to geometric properties of underlying spaces. In particular it led to the insight that, e.g. Hadamard manifolds cannot feature stickiness. Although not addressed in the project, we conjecture, that they can feature finite sample stickiness, however, making quantile based tests inappropriate. In fact, using the variance modulation’s sticky flavor allows for exact definition of finite sample stickiness and according statistical methods. Further investigation (e.g. discriminating different types as is the case with FSS) which may have direct impact on statistics of phylogenetic trees is left for further research, in particular exploiting stickiness as a blessing to identify hybridisation effects in a nascent collaboration with the Deutsche Primatenzentrum in Göttingen. Also on the side of smeariness – the original avenue of the project – major progress could be achieved. On the one hand, the variance modulation allowed not only for a precise definition of FSS – not available at the start of the project – but also for its statistical assessment. The connection between smeariness (begetting FSS) and compact manifolds is now much clearer, and so is smeariness as a phase transition between uniqueness and nonuniqueness, but, by no means, fully understood yet: For instance we have not been able to produce smeariness of arbitrary polynomial degree in spheres, let alone on general compact manifolds – although we conjecture even log-smeariness can be found there. Notably, log-smeariness has been recently established by Kroshnin (2021) on Wasserstein spaces. Also a precise formulation of an asymptotic measure on population Fréchet mean sets induced by sampling, and its consequences, e.g. for statistical testing, has not yet been addressed in the project and is left for future research. One major obstacle we frequently encountered is the very restrictive condition for the support of a probability distribution within a geodesic half ball in CAT(κ) spaces for κ > 0 to ensure uniqueness, cf. Afsari (2011). We expect that combining the many results of this project with a weakening of this restriction, e.g. in the spirit of Arnaudon and Miclo (2014) (the sample mean from a distribution with density w.r.t. Riemannian volume is a.s. unique) will lead to much stronger results; for instance that no Riemannian manifold may feature stickiness – thus far we could only show it for distributions restricted to geodesic half balls. Besides addressing the specific research goals, a number of overview articles could be published. Linking geometry and statistics with direct impact on applications in non-Euclidean statistics, we were able to answer a number of fundamental questions, a considerable number of questions remain open and many more new questions arose.

Projektbezogene Publikationen (Auswahl)

 
 

Zusatzinformationen

Textvergrößerung und Kontrastanpassung