generator           JSD      V(D*, G)   -log4 + 2 JSD   V(D=1/2, G)
uniform             0.0529   -1.280510  -1.280510        -1.386294
shifted by one      0.0516   -1.283000  -1.283000        -1.386294
two middle points   0.1639   -1.058501  -1.058501        -1.386294
the data itself     0.0000   -1.386294  -1.386294        -1.386294
D* for the data itself: 0.50 0.50 0.50 0.50 0.50 0.50   -log 4 = -1.386294

logit a   D = sigmoid(a)   |d log(1-D) / da|   |d (-log D) / da|
 -8.0     0.00034          0.00034              0.99966
 -4.0     0.01799          0.01799              0.98201
 -2.0     0.11920          0.11920              0.88080
 +0.0     0.50000          0.50000              0.50000

after 300 discriminator steps against an untrained generator, mean D(G(z)) = 0.0154
generator gradient norm with loss log(1 - D): 0.4386
generator gradient norm with loss -log D    : 13.0146
ratio 29.7

seed 0 while it trains (the real data has 0.700 of its mass in the right hump)
step   D(real)  D(fake)  W1     right hump
500    0.517    0.517    0.432  0.715
1000   0.503    0.498    0.328  0.702
1500   0.494    0.491    0.194  0.693
2000   0.507    0.502    0.117  0.698

after 2000 steps, four seeds
seed  W1      right hump  share in the gap  discriminator accuracy
0     0.113   0.694       0.022             0.501
1     0.338   0.705       0.029             0.461
2     0.234   0.718       0.019             0.515
3     0.161   0.710       0.021             0.493
two real samples of this size differ by W1 0.016; the gap holds 0.016 of the real data; 0.500 is a discriminator that cannot tell
