Dr. Paul Marriott

Department of Statistics and Actuarial Science

University of Waterloo

Title: "Local and Global Mixture Geometry"

Abstract:
This talk looks at some geometric ideas which have a found a rich application in statistical
inference. In general the inference problem with mixture models can be very challenging and
very different from the gold-standard of inference in exponential families. However a simplifying
modelling assumption, which often holds in practical situations, exploits the underlying geometry
of the mixture model and ensures that the resulting inference problem is tractable. This
simplification gives rise to a class of models which I have called local mixtures. The talk
describes these local mixture models and then investigates if the idea of local mixtures can
be extended to much more general families.