def __init__(self, mu=0.0, sigmasq=1.0, lb=-np.Inf, ub=np.Inf):
assert np.all(sigmasq) >= 0
# Broadcast arrays to be of the same shape
self.mu, self.sigmasq, self.lb, self.ub \
= np.broadcast_arrays(mu, sigmasq, lb, ub)
# Precompute the normalizers
self.zlb = (self.lb-self.mu) / np.sqrt(self.sigmasq)
self.zub = (self.ub-self.mu) / np.sqrt(self.sigmasq)
self.Z = normal_cdf(self.zub) - normal_cdf(self.zlb)
def mf_expected_log_notp(self):
"""
Compute the expected log probability of no connection under Z
:return:
"""
mu = self.mf_mu_Z
return np.log(normal_cdf(0, mu=mu, sigma=1.0))
def mf_expected_log_p(self):
"""
Compute the expected log probability of a connection under Z
:return:
"""
mu = self.mf_mu_Z
# return np.nan_to_num(np.log(1.0 - normal_cdf(0, mu=mu, sigma=1.0)))
return np.log(1.0 - normal_cdf(0, mu=mu, sigma=1.0))
def mf_expected_log_p_mc(self):
N_samples = 100
E_mu = self.mf_expected_mu()[None, :, :]
std_mu = np.sqrt(self.mf_variance_mu()[None, :, :])
mus = E_mu + std_mu * np.random.randn(N_samples, self.N, self.N)
# ps = 1.0 - normal_cdf(0, mu=mus, sigma=1.0)
# log_ps = np.log(ps)
# log_notps = np.log(1-ps)
# Approximate log(1-p) for p~= 1
u = normal_cdf(0, mu=mus, sigma=1.0)
log_ps = np.log1p(-u)
log_notps = np.log1p(-1+u)
return log_ps.mean(0), log_notps.mean(0)
