An efficient implementation of a nonlinear predictor using a zero-memory nonlinearity followed by a second order Volterra filter

This paper is concerned with the development and analysis of a nonlinear predictor for a non-Gaussian process using a zero-memory nonlinearity (ZMNL) followed by a second order Volterra filter (SVF). The processor exploits partial statistical information such as marginal probability density function (PDF) and the covariance structure. The ZMNL transforms the process into a Gaussian process. The SVF can be implemented as a parallel combination of linear and quadratic filters. Another equivalent structure is to carry out a linear prediction in the transformed Gaussian domain and pass the linear predicted samples through a second order polynomial