For 1/f^0.5 targets, the best sparse random-feature models match the performance of the same MLP using dense Fourier features.

来源：paper:Figure 4(a) and caption, p. 7; §5 p. 6