DependencyModelsbasedonGeneralizedGaussianScale.PDFVIP

J. A. Palmer, K. Kreutz-Delgado, and S. Makeig ∗ September 7, 2011 This paper presents a framework for modeling dependency among multiple random variables based on scale mixtures and location-scale mixtures, drawing on the work of Barndorff-Nielsen and others. A simple way to construct a non-Gaussian multivariate density is to take an arbitrary scalar function of a quadratic form, subject to integrability constraints. This makes the probability density in the neighborhood of a vector a function of the level ellipsoid on which it lies. We thus have the or densities. 1 2 A simple extension of this idea is to make the probability density a function of an arbitrary “gauge-like” [14] function with concentric level sets. For example we might ∑ take an arbitrary norm ( |x | ) instead of the radially symmetric two-norm. We might generalize the idea of p-norm gauges by taking the sum of general nonlinear functions of the components,