Probability distribution of transformed random variables with application to nonlinear features extraction

summary:A method for estimation of probability distribution of transformed random variables is presented. The proposed approach admits an approximation of the transformation of the random variables. The approximate probability density function (pdf) is corrected to obtain a resulting pdf which incorporates a prior knowledge of approximation errors. The corrected pdf is not contaminated by any uncontrollable approximation. The method is applied to pattern recognition. It is shown that class conditional pdf of features can be easily computed even when the feature extraction was performed with nonlinear mapping of an input pattern