Multidimensional probability density function approximations for detection, classification, and model order selection

This paper addresses the problem of calculating the multidimensional probability density functions (PDFs) of statistics derived from known many-to-one transformations of independent random variables (RVs) with known distributions. The statistics covered in the paper include reflection coefficients, autocorrelation estimates, cepstral coefficients, and general linear functions of independent RVs. Through PDF transformation, these results can be used for general PDF approximation, detection, classification, and model order selection. A model order selection example that shows significantly better performance than the Akaike and MDL method is included.