def test_product_with_badrvs():
"""
Test product_distribution() with overlapping rvs specification.
"""
d = dit.example_dists.Xor()
with pytest.raises(Exception):
dit.product_distribution(d, [[0, 1], [0]])
def test_product_nonjoint():
"""
Test product_distribution() from a ScalarDistribution.
"""
d = dit.ScalarDistribution([.5, .5])
with pytest.raises(Exception):
dit.product_distribution(d)
def test_product_with_badrvs():
"""
Test product_distribution() with overlapping rvs specification.
"""
d = dit.example_dists.Xor()
with pytest.raises(Exception):
dit.product_distribution(d, [[0,1], [0]])
def test_product_nonjoint():
"""
Test product_distribution() from a ScalarDistribution.
"""
d = dit.ScalarDistribution([.5, .5])
with pytest.raises(Exception):
dit.product_distribution(d)
def marginal_maxent_dists(dist,
k_max=None,
maxiters=1000,
tol=1e-3,
verbose=False):
"""
Return the marginal-constrained maximum entropy distributions.
Parameters
----------
dist : distribution
The distribution used to constrain the maxent distributions.
k_max : int
The maximum order to calculate.
"""
dist = prepare_dist(dist)
n_variables = dist.outcome_length()
symbols = dist.alphabet[0]
if k_max is None:
k_max = n_variables
outcomes = list(dist._sample_space)
# Optimization for the k=0 and k=1 cases are slow since you have to optimize
# the full space. We also know the answer in these cases.
# This is safe since the distribution must be dense.
k0 = dit.Distribution(outcomes, [1] * len(outcomes),
base='linear',
validate=False)
k0.normalize()
k1 = dit.product_distribution(dist)
dists = [k0, k1]
for k in range(k_max + 1):
if verbose:
print(
"Constraining maxent dist to match {0}-way marginals.".format(
k))
if k in [0, 1, n_variables]:
continue
kwargs = {'maxiters': maxiters, 'tol': tol, 'verbose': verbose}
pmf_opt, opt = marginal_maxent(dist, k, **kwargs)
d = dit.Distribution(outcomes, pmf_opt)
d.make_sparse()
dists.append(d)
# To match the all-way marginal is to match itself. Again, this is a time
# savings decision, even though the optimization should be fast.
if k_max == n_variables:
dists.append(dist)
return dists
def test_product_with_rvs2():
"""
Test product_distribution() with an rvs specification.
"""
d = dit.example_dists.Xor()
d_iid = dit.product_distribution(d, [[0, 1]])
d_truth = dit.uniform_distribution(2, ['01'])
d_truth = dit.modify_outcomes(d_truth, lambda x: ''.join(x))
assert d_truth.is_approx_equal(d_iid)
def test_product():
"""
Smoke test for product_distribution().
"""
d = dit.example_dists.Xor()
d_iid = dit.product_distribution(d)
d_truth = dit.uniform_distribution(3, ['01'])
d_truth = dit.modify_outcomes(d_truth, lambda x: ''.join(x))
assert d_truth.is_approx_equal(d_iid)
def test_product_with_rvs2():
"""
Test product_distribution() with an rvs specification.
"""
d = dit.example_dists.Xor()
d_iid = dit.product_distribution(d, [[0,1]])
d_truth = dit.uniform_distribution(2, ['01'])
d_truth = dit.modify_outcomes(d_truth, lambda x: ''.join(x))
assert_true(d_truth.is_approx_equal(d_iid))
def test_product():
"""
Smoke test for product_distribution().
"""
d = dit.example_dists.Xor()
d_iid = dit.product_distribution(d)
d_truth = dit.uniform_distribution(3, ['01'])
d_truth = dit.modify_outcomes(d_truth, lambda x: ''.join(x))
assert_true(d_truth.is_approx_equal(d_iid))
def marginal_maxent_dists(dist, k_max=None, maxiters=1000, tol=1e-3, verbose=False):
"""
Return the marginal-constrained maximum entropy distributions.
Parameters
----------
dist : distribution
The distribution used to constrain the maxent distributions.
k_max : int
The maximum order to calculate.
"""
dist = prepare_dist(dist)
n_variables = dist.outcome_length()
symbols = dist.alphabet[0]
if k_max is None:
k_max = n_variables
outcomes = list(dist._sample_space)
# Optimization for the k=0 and k=1 cases are slow since you have to optimize
# the full space. We also know the answer in these cases.
# This is safe since the distribution must be dense.
k0 = dit.Distribution(outcomes, [1]*len(outcomes), base='linear', validate=False)
k0.normalize()
k1 = dit.product_distribution(dist)
dists = [k0, k1]
for k in range(k_max + 1):
if verbose:
print("Constraining maxent dist to match {0}-way marginals.".format(k))
if k in [0, 1, n_variables]:
continue
kwargs = {'maxiters': maxiters, 'tol': tol, 'verbose': verbose}
pmf_opt, opt = marginal_maxent(dist, k, **kwargs)
d = dit.Distribution(outcomes, pmf_opt)
d.make_sparse()
dists.append(d)
# To match the all-way marginal is to match itself. Again, this is a time
# savings decision, even though the optimization should be fast.
if k_max == n_variables:
dists.append(dist)
return dists