We consider the description of a nonlinear stochastic transduction in terms of its input/output distribution. We construct a sequence of approximating maximum-entropy estimates from a finite set of input/output observations. This procedure extends the Wiener theory to the analysis of nonlinear stochastic transducers and to the analysis of transducers with multiple outputs but an inaccessible input.
We also show that for a deterministic transduction, the Wiener kernels represent maximum-entropy approximations.