Rapid Implementation of Score Matching Estimators

Abstract

Score matching (Hyvarinen, 2005) is an estimation technique that avoids normalising constants in model densities (i.e. it works on improper densities) and can thus be used in many cases that maximum likelihood estimation cannot.  Unfortunately, the implementation of score matching estimators often requires tedious calculus, especially for models of data that lie on multidimensional manifolds.  I have developed an R package that uses automatic differentiation to make implementation much faster.  The package already contains estimators for compositional data models and directional distributions.