Solving learning problems numerically by discretizing the and variables allows in principle to deal with arbitrary non-Gaussian priors. Compared to Gaussian priors, however, the resulting stationarity ...