An nonlinear digital signal processing filter (100, 200, 1100, 1308, 1310, 1312, 1346, 1604) maintains a magnitude ordering for successive windows of signal samples. A set of filter density generator values [f.sub.1, f.sub.2, f.sub.3 . . . f.sub.j . . . f.sub.ndensities] are used according to the ordering in a recursion relation that computes successive values of a set function over the set of filter density generator values. The recursion relation involves an adjustable nonlinearity defining parameter .lamda.. The values are normalized by dividing by a largest of the values, and differences between successive values are taken. An inner product between each window of signal values (used in order according to magnitude) and the adaptive differences is a filtered signal sample.