Bayesian model based parameter estimation and model selection in impulsive noise environments

An effective Monte Carlo method using importance sampling is presented to estimate Bayesian model evidence. This is applied to Bayesian model selection in non-Gaussian noise, determining the appropriate signal model and noise statistics simultaneously. The authors also discuss the resolution of two closely spaced frequencies in impulsive noise. The resolution obtained assuming the correct noise statistics is contrasted with that obtained using the fast Fourier transform (FFT) and Gaussian noise assumption. Examples of parameter estimation and model selection in impulsive noise environments are included.<<ETX>>