We investigate the neural computations underlying pattern processing with stimuli based on textures balanced for spatial frequency content (and second-order correlations) but not for higher-order correlations (Julesz et al. 1978). Interchange between two such isodipole textures produces a robust human visual evoked potential (VEP). The difference in population activity driven by two isodipole textures is quantified by the antisymmetric component of the VEP. Statistical properties of the textures eliminate contributions from linear mechanisms to the antisymmetric VEP.

The dependence of the antisymmetric VEP on check size and fourth-order correlation statistics is used to test nonlinear models for the underlying neural computations. Linear summation, followed by a simple nonlinearity (such as rectification, saturation, or threshold), is inconsistent with the data. More elaborate models, in which a second nonlinear stage combines the output of local nonlinear mechanisms, are consistent with the data, provided that an appropriate spatial scale is chosen for the second stage of processing. For checks 4 min or smaller, the deduced interaction length is 10–15 min. For checks larger than 4 min, the interaction length is proportional to check size.