For 1/f^0.5 targets, too-small sampled-frequency scale yields underfitting and slow convergence, with training and test errors similar.

来源：paper:Figure 4 p. 7; Figure 13 p. 16; §B.2 pp. 15–16