Detection In Alpha-stable Noise Environments Based On Nonlinear Prediction

In this paper', we consider detection of signals in a mixture of Gaussian noise and impulsive noise modeled as an alpha-stable process. Since our noise model has infinite variance, in order to use a minimum meansquared error (MMSE) criterion, we apply zero memory nonlinearity (ZMNL) to the information-bearing signal, in such a way that the variance of the noise is limited and the inform* tion signal is not distorted. We generalize the class of detectors which are based on a noise estimation-cancellation technique. In particular, by exploiting the past decisions as well as the past received samples, a nonlinear MMSE estimate of the transformed noise is made and subsequently canceled. We optimize the performance of the system with respect to the ZMNL at the input of the receiver. Our objective is to use predictors of the lowest complexity which give satisfactory estimation accuracy. The proposed subop t imd receivers are designed and analyzed in the context of Partial Response Signaling (PRS). The effects of the predictor order, the number of exploited samples and filtering allocation, on the system performance are examined.