[C.4] identity, max abs error over 1000 random policies: 1.78e-15
[C.5] lam=0.2: gain 0.3380  bound s*sqrt(2KL) 0.3380
[C.5] lam=1.0: gain 1.6900  bound s*sqrt(2KL) 1.6900
[C.5] lam=3.0: gain 5.0700  bound s*sqrt(2KL) 5.0700
[C.1/Lemma C.3] E||psi||^2 4.2442  exact sum_t (1 - |pi_t|^2) 4.2447  |cross term| 0.0002
[C.6] c=1.0: bias 0.0684 <= bound 0.3320
[C.6] c=2.0: bias 0.0512 <= bound 0.1915
[C.6] c=5.0: bias 0.0424 <= bound 0.0997
[C.7] J(q)-J(p) = -0.174699   sum d^q A^p = -0.174699
[C.3] true gradient [ 0.21  -0.174 -0.036]  LOO [ 0.21  -0.174 -0.036]  mean-centered [ 0.14  -0.116 -0.024]  ratio 0.6667 (1-1/G=0.6667)
all assertions passed
