For 1/f^1.5 targets, too-large sampled-frequency scale yields overfitting: test error increases while training error approaches machine precision.

来源：paper:Figure 4 p. 7; Figure 13 and caption p. 16; §B.2