CMStatistics 2021: Start Registration
View Submission - CMStatistics
Title: Parameter estimators for non-normalized statistical models based on Stein characterizations Authors:  Steffen Betsch - Karlsruhe Institute of Technology (Germany) [presenting]
Bruno Ebner - Karlsruhe Institute of Technology (Germany)
Bernhard Klar - Karlsruhe institute of Technology (Germany)
Abstract: A new estimation method for the parameters of non-normalized models consisting of smooth density functions on the real line is presented. For everyone to be on the same page, a short summary of the issues with non-normalized densities is given. The eventual estimation procedure is based on a recently introduced characterization for the respective probability distributions that is derived from a classical identity which underlies some applications of Stein's method. These characterizations are stated and used to construct minimum distance parameter estimators. The new method is compared, in simulations, with different approaches from the machine learning literature that tackle the same problem. Moreover, connections to inferential methods based on Stein discrepancies are discussed.