def test_weighted_ExtendedBinnedNLL():
def cdf(x, a, b):
return b * (1 - np.exp(-a * x))
xe = np.array([0, 1, 10])
n = np.diff(cdf(xe, 1, 100))
m1 = Minuit(ExtendedBinnedNLL(n, xe, cdf), 1, 100)
m1.migrad()
assert_allclose(m1.values, (1, 100), rtol=1e-2)
w = np.transpose((n, 4 * n))
m2 = Minuit(ExtendedBinnedNLL(w, xe, cdf), 1, 100)
m2.migrad()
assert_allclose(m2.values, (1, 100), rtol=1e-2)
assert m2.errors[0] == pytest.approx(2 * m1.errors[0], rel=1e-2)
def test_ExtendedBinnedNLL_mask():
c = ExtendedBinnedNLL([1, 1000, 2], [0, 1, 2, 3], expon_cdf)
assert c.ndata == 3
c_unmasked = c(2)
c.mask = np.arange(3) != 1
assert c(2) < c_unmasked
assert c.ndata == 2
def test_ExtendedBinnedNLL(binned, verbose):
mle, nx, xe = binned
def scaled_cdf(x, n, mu, sigma):
return n * norm(mu, sigma).cdf(x)
cost = ExtendedBinnedNLL(nx, xe, scaled_cdf, verbose=verbose)
m = Minuit(cost, n=mle[0], mu=0, sigma=1)
m.limits["n"] = (0, None)
m.limits["sigma"] = (0, None)
m.migrad()
# binning loses information compared to unbinned case
assert_allclose(m.values, mle, rtol=0.15)
assert m.errors["mu"] == pytest.approx(1000 ** -0.5, rel=0.05)
def test_ExtendedBinnedNLL_properties():
def cdf(x, a):
return 0
c = ExtendedBinnedNLL([1], [1, 2], cdf)
assert c.scaled_cdf is cdf
def test_ExtendedBinnedNLL_bad_input():
with pytest.raises(ValueError):
ExtendedBinnedNLL([1], [1], lambda x, a: 0)
def test_ExtendedBinnedNLL_mask():
c = ExtendedBinnedNLL([1, 1000, 2], [0, 1, 2, 3], lambda x, a: a * x)
assert c(2) == pytest.approx(-700, rel=0.1)
c.mask = np.arange(3) != 1
assert c(2) == pytest.approx(2, rel=0.1)