def test_correct_values(self) -> None:
for rho in [0.0, 0.01, 0.2, 0.8]:
with self.subTest(rho=rho):
# MI for two correlated unit variance Gaussians
mi = -0.5 * np.log(1 - rho**2)
self.assertAlmostEqual(normalize_mi(mi), rho, delta=0.001)
def test_list_is_handled_elementwise(self) -> None:
mi = [-1, 0, 1]
cor = normalize_mi(mi)
self.assertEqual(cor.shape, (3, )) # type: ignore
self.assertAlmostEqual(cor[0], -1.0, delta=0.001)
self.assertAlmostEqual(cor[1], 0.0, delta=0.001)
self.assertAlmostEqual(cor[2], 0.93, delta=0.03)
def test_pandas_structure_is_preserved(self) -> None:
mi = np.asarray([[0.1, 0.5], [0, -1]])
mi = pd.DataFrame(mi, columns=["A", "B"], index=[14, 52])
cor = normalize_mi(mi) # type: pd.DataFrame
self.assertAlmostEqual(cor.loc[14, "A"], 0.4, delta=0.05)
self.assertAlmostEqual(cor.loc[14, "B"], 0.8, delta=0.05)
self.assertAlmostEqual(cor.loc[52, "A"], 0.0, delta=0.001)
self.assertAlmostEqual(cor.loc[52, "B"], -1.0, delta=0.001)
def test_array_is_handled_elementwise(self) -> None:
mi = np.asarray([[0.1, 0.5], [0, -1]])
cor = normalize_mi(mi)
self.assertEqual(cor.shape, (2, 2))
# Large deltas because I looked these up manually - the order is what matters
self.assertAlmostEqual(cor[0, 0], 0.4, delta=0.05)
self.assertAlmostEqual(cor[0, 1], 0.8, delta=0.05)
self.assertAlmostEqual(cor[1, 0], 0.0, delta=0.001)
self.assertAlmostEqual(cor[1, 1], -1.0, delta=0.001)
def test_negative_values_are_passed_as_is(self) -> None:
for value in [-0.01, -0.1, -1, -np.inf]:
with self.subTest(value=value):
self.assertAlmostEqual(normalize_mi(value), value, delta=0.001)