Its auto regression. Think of a row of pixels as a signal y[n]. Assume y[n] = a y[n-1] + b y[n-2]. estimate a,b. Then store y[0], y[1], a, b, and the residual, then encode the four values and the residual which will mostly be zeros. You can vary the length and number of coefficients very cheaply, but a fixed two beat png in my tests. There are fast standard algorithms to find a,b and which along with some simple tricks to correct for the precision ensure lossless encoding. Its the mathematicians version of the ad hoc thing png tries to do.