P (red) is fixed; Q (blue) is the same shape moved by θ and stretched by a width factor. The Jensen–Shannon divergence compares the two densities point by point, so once they stop overlapping it is stuck at ln 2 whatever the distance between them, and its gradient is zero. The Wasserstein-1 distance (the earth mover's distance of Section 17.1.2) is the least average distance the mass of Q must travel to become P; in one dimension it is the area between the two distribution functions, and it keeps growing with θ.
On a grid of 6001 points on [−15, 15]: JS = ½KL(P‖M) + ½KL(Q‖M), M = (P + Q)/2; W1 = ∫|FP − FQ| dx. The optimal critic f* has slope −1 where FP > FQ and +1 where FP < FQ, and attains W1 = EP[f*] − EQ[f*] (Kantorovich–Rubinstein). Descent uses central differences with step 0.001.
code/ch17_generative_adversarial_networks.py (shift_curves, shift_descent, optimal_critic); the numbers here equal the Python ones.