A simple procedure for improving linear prediction is presented. A zero mem
ory nonlinear step at the output of the linear predictor may be adequate to
obtain significant improvements. Two techniques are proposed for designing
the nonlinearity, the object of each being to achieve the conditional mean
of the nonlinear prediction given the linear one. The technique is applied
to improve the signal-to-noise ratio in the automatic detection of lire by
infrared signal processing.