def free_energy(eps,q,mu,samples=1000,use_annealed_approx=False):
"""Compute free energy for true distribution given best approximation. Return beta*free_energy,actually"""
# def Q(xs):
# return product(fd(ep,mu) if x else 1 - fd(ep,mu) for x,ep in zip(xs,eps))
Sq = -sum((p*log(p)+(1-p)*log(1-p) for p in probs(eps,mu)))
def E(xs):
"""Compute energy function for P,unnormalized"""
ff = falling_fac(q,sum(xs))
if ff == 0:
return 1000
else:
return log(ff) + sum(ep*x for (x,ep) in zip(xs,eps))
if use_annealed_approx: # to get around impossible states where E=\infty
mean_E = log(mean(exp(E(rstate(eps,mu))) for i in range(samples))) # take <E(xs)>_Q
else:
mean_E = mean(E(rstate(eps,mu)) for i in range(samples)) # take <E(xs)>_Q
return beta*mean_E - Sq
def proposal(ss):
#state = [int(random.random() < p) for _ in range(len(ss))]
state = rstate(eps,mu)
#print "proposed state with occ:",sum(state)
return state
