This fits in with my point: they imposed several restrictions about the model space: maximizing sparseness, and they also make several linearity assumptions.