scipy 1.16.2__cp311-cp311-win_arm64.whl

scipy/special/tests/test_hypergeometric.py

@@ -0,0 +1,234 @@
+import pytest
+import numpy as np
+from numpy.testing import assert_allclose, assert_equal
+import scipy.special as sc
+
+
+class TestHyperu:
+
+    def test_negative_x(self):
+        a, b, x = np.meshgrid(
+            [-1, -0.5, 0, 0.5, 1],
+            [-1, -0.5, 0, 0.5, 1],
+            np.linspace(-100, -1, 10),
+        )
+        assert np.all(np.isnan(sc.hyperu(a, b, x)))
+
+    def test_special_cases(self):
+        assert sc.hyperu(0, 1, 1) == 1.0
+
+    @pytest.mark.parametrize('a', [0.5, 1, np.nan])
+    @pytest.mark.parametrize('b', [1, 2, np.nan])
+    @pytest.mark.parametrize('x', [0.25, 3, np.nan])
+    def test_nan_inputs(self, a, b, x):
+        assert np.isnan(sc.hyperu(a, b, x)) == np.any(np.isnan([a, b, x]))
+
+    @pytest.mark.parametrize(
+        'a,b,x,expected',
+        [(0.21581740448533887, 1.0, 1e-05, 3.6030558839391325),
+         (0.21581740448533887, 1.0, 0.00021544346900318823, 2.8783254988948976),
+         (0.21581740448533887, 1.0, 0.004641588833612777, 2.154928216691109),
+         (0.21581740448533887, 1.0, 0.1, 1.446546638718792),
+         (0.0030949064301273865, 1.0, 1e-05, 1.0356696454116199),
+         (0.0030949064301273865, 1.0, 0.00021544346900318823, 1.0261510362481985),
+         (0.0030949064301273865, 1.0, 0.004641588833612777, 1.0166326903402296),
+         (0.0030949064301273865, 1.0, 0.1, 1.0071174207698674),
+         (0.1509924314279033, 1.0, 1e-05, 2.806173846998948),
+         (0.1509924314279033, 1.0, 0.00021544346900318823, 2.3092158526816124),
+         (0.1509924314279033, 1.0, 0.004641588833612777, 1.812905980588048),
+         (0.1509924314279033, 1.0, 0.1, 1.3239738117634872),
+         (-0.010678995342969011, 1.0, 1e-05, 0.8775194903781114),
+         (-0.010678995342969011, 1.0, 0.00021544346900318823, 0.9101008998540128),
+         (-0.010678995342969011, 1.0, 0.004641588833612777, 0.9426854294058609),
+         (-0.010678995342969011, 1.0, 0.1, 0.9753065150174902),
+         (-0.06556622211831487, 1.0, 1e-05, 0.26435429752668904),
+         (-0.06556622211831487, 1.0, 0.00021544346900318823, 0.4574756033875781),
+         (-0.06556622211831487, 1.0, 0.004641588833612777, 0.6507121093358457),
+         (-0.06556622211831487, 1.0, 0.1, 0.8453129788602187),
+         (-0.21628242470175185, 1.0, 1e-05, -1.2318314201114489),
+         (-0.21628242470175185, 1.0, 0.00021544346900318823, -0.6704694233529538),
+         (-0.21628242470175185, 1.0, 0.004641588833612777, -0.10795098653682857),
+         (-0.21628242470175185, 1.0, 0.1, 0.4687227684115524)]
+    )
+    def test_gh_15650_mp(self, a, b, x, expected):
+        # See https://github.com/scipy/scipy/issues/15650
+        # b == 1, |a| < 0.25, 0 < x < 1
+        #
+        # This purpose of this test is to check the accuracy of results
+        # in the region that was impacted by gh-15650.
+        #
+        # Reference values computed with mpmath using the script:
+        #
+        # import itertools as it
+        # import numpy as np
+        #
+        # from mpmath import mp
+        #
+        # rng = np.random.default_rng(1234)
+        #
+        # cases = []
+        # for a, x in it.product(
+        #         np.random.uniform(-0.25, 0.25, size=6),
+        #         np.logspace(-5, -1, 4),
+        # ):
+        #     with mp.workdps(100):
+        #         cases.append((float(a), 1.0, float(x), float(mp.hyperu(a, 1.0, x))))
+        assert_allclose(sc.hyperu(a, b, x), expected, rtol=1e-13)
+
+    def test_gh_15650_sanity(self):
+        # The purpose of this test is to sanity check hyperu in the region that
+        # was impacted by gh-15650 by making sure there are no excessively large
+        # results, as were reported there.
+        a = np.linspace(-0.5, 0.5, 500)
+        x = np.linspace(1e-6, 1e-1, 500)
+        a, x = np.meshgrid(a, x)
+        results = sc.hyperu(a, 1.0, x)
+        assert np.all(np.abs(results) < 1e3)
+
+
+class TestHyp1f1:
+
+    @pytest.mark.parametrize('a, b, x', [
+        (np.nan, 1, 1),
+        (1, np.nan, 1),
+        (1, 1, np.nan)
+    ])
+    def test_nan_inputs(self, a, b, x):
+        assert np.isnan(sc.hyp1f1(a, b, x))
+
+    def test_poles(self):
+        assert_equal(sc.hyp1f1(1, [0, -1, -2, -3, -4], 0.5), np.inf)
+
+    @pytest.mark.parametrize('a, b, x, result', [
+        (-1, 1, 0.5, 0.5),
+        (1, 1, 0.5, 1.6487212707001281468),
+        (2, 1, 0.5, 2.4730819060501922203),
+        (1, 2, 0.5, 1.2974425414002562937),
+        (-10, 1, 0.5, -0.38937441413785204475)
+    ])
+    def test_special_cases(self, a, b, x, result):
+        # Hit all the special case branches at the beginning of the
+        # function. Desired answers computed using Mpmath.
+        assert_allclose(sc.hyp1f1(a, b, x), result, atol=0, rtol=1e-15)
+
+    @pytest.mark.parametrize('a, b, x, result', [
+        (1, 1, 0.44, 1.5527072185113360455),
+        (-1, 1, 0.44, 0.55999999999999999778),
+        (100, 100, 0.89, 2.4351296512898745592),
+        (-100, 100, 0.89, 0.40739062490768104667),
+        (1.5, 100, 59.99, 3.8073513625965598107),
+        (-1.5, 100, 59.99, 0.25099240047125826943)
+    ])
+    def test_geometric_convergence(self, a, b, x, result):
+        # Test the region where we are relying on the ratio of
+        #
+        # (|a| + 1) * |x| / |b|
+        #
+        # being small. Desired answers computed using Mpmath
+        assert_allclose(sc.hyp1f1(a, b, x), result, atol=0, rtol=1e-15)
+
+    @pytest.mark.parametrize('a, b, x, result', [
+        (-1, 1, 1.5, -0.5),
+        (-10, 1, 1.5, 0.41801777430943080357),
+        (-25, 1, 1.5, 0.25114491646037839809),
+        (-50, 1, 1.5, -0.25683643975194756115),
+        (-80, 1, 1.5, -0.24554329325751503601),
+        (-150, 1, 1.5, -0.173364795515420454496),
+    ])
+    def test_a_negative_integer(self, a, b, x, result):
+        # Desired answers computed using Mpmath.
+        assert_allclose(sc.hyp1f1(a, b, x), result, atol=0, rtol=2e-14)
+
+    @pytest.mark.parametrize('a, b, x, expected', [
+        (0.01, 150, -4, 0.99973683897677527773),        # gh-3492
+        (1, 5, 0.01, 1.0020033381011970966),            # gh-3593
+        (50, 100, 0.01, 1.0050126452421463411),         # gh-3593
+        (1, 0.3, -1e3, -7.011932249442947651455e-04),   # gh-14149
+        (1, 0.3, -1e4, -7.001190321418937164734e-05),   # gh-14149
+        (9, 8.5, -350, -5.224090831922378361082e-20),   # gh-17120
+        (9, 8.5, -355, -4.595407159813368193322e-20),   # gh-17120
+        (75, -123.5, 15, 3.425753920814889017493e+06),
+    ])
+    def test_assorted_cases(self, a, b, x, expected):
+        # Expected values were computed with mpmath.hyp1f1(a, b, x).
+        assert_allclose(sc.hyp1f1(a, b, x), expected, atol=0, rtol=1e-14)
+
+    def test_a_neg_int_and_b_equal_x(self):
+        # This is a case where the Boost wrapper will call hypergeometric_pFq
+        # instead of hypergeometric_1F1.  When we use a version of Boost in
+        # which https://github.com/boostorg/math/issues/833 is fixed, this
+        # test case can probably be moved into test_assorted_cases.
+        # The expected value was computed with mpmath.hyp1f1(a, b, x).
+        a = -10.0
+        b = 2.5
+        x = 2.5
+        expected = 0.0365323664364104338721
+        computed = sc.hyp1f1(a, b, x)
+        assert_allclose(computed, expected, atol=0, rtol=1e-13)
+
+    @pytest.mark.parametrize('a, b, x, desired', [
+        (-1, -2, 2, 2),
+        (-1, -4, 10, 3.5),
+        (-2, -2, 1, 2.5)
+    ])
+    def test_gh_11099(self, a, b, x, desired):
+        # All desired results computed using Mpmath
+        assert sc.hyp1f1(a, b, x) == desired
+
+    @pytest.mark.parametrize('a', [-3, -2])
+    def test_x_zero_a_and_b_neg_ints_and_a_ge_b(self, a):
+        assert sc.hyp1f1(a, -3, 0) == 1
+
+    # In the following tests with complex z, the reference values
+    # were computed with mpmath.hyp1f1(a, b, z), and verified with
+    # Wolfram Alpha Hypergeometric1F1(a, b, z), except for the
+    # case a=0.1, b=1, z=7-24j, where Wolfram Alpha reported
+    # "Standard computation time exceeded".  That reference value
+    # was confirmed in an online Matlab session, with the commands
+    #
+    #  > format long
+    #  > hypergeom(0.1, 1, 7-24i)
+    #  ans =
+    #   -3.712349651834209 + 4.554636556672912i
+    #
+    @pytest.mark.parametrize(
+        'a, b, z, ref',
+        [(-0.25, 0.5, 1+2j, 1.1814553180903435-1.2792130661292984j),
+         (0.25, 0.5, 1+2j, 0.24636797405707597+1.293434354945675j),
+         (25, 1.5, -2j, -516.1771262822523+407.04142751922024j),
+         (12, -1.5, -10+20j, -5098507.422706547-1341962.8043508842j),
+         pytest.param(
+             10, 250, 10-15j, 1.1985998416598884-0.8613474402403436j,
+             marks=pytest.mark.xfail,
+         ),
+         pytest.param(
+             0.1, 1, 7-24j, -3.712349651834209+4.554636556672913j,
+             marks=pytest.mark.xfail,
+         )
+         ],
+    )
+    def test_complex_z(self, a, b, z, ref):
+        h = sc.hyp1f1(a, b, z)
+        assert_allclose(h, ref, rtol=4e-15)
+
+    # The "legacy edge cases" mentioned in the comments in the following
+    # tests refers to the behavior of hyp1f1(a, b, x) when b is a nonpositive
+    # integer.  In some subcases, the behavior of SciPy does not match that
+    # of Boost (1.81+), mpmath and Mathematica (via Wolfram Alpha online).
+    # If the handling of these edges cases is changed to agree with those
+    # libraries, these test will have to be updated.
+
+    @pytest.mark.parametrize('b', [0, -1, -5])
+    def test_legacy_case1(self, b):
+        # Test results of hyp1f1(0, n, x) for n <= 0.
+        # This is a legacy edge case.
+        # Boost (versions greater than 1.80), Mathematica (via Wolfram Alpha
+        # online) and mpmath all return 1 in this case, but SciPy's hyp1f1
+        # returns inf.
+        assert_equal(sc.hyp1f1(0, b, [-1.5, 0, 1.5]), [np.inf, np.inf, np.inf])
+
+    def test_legacy_case2(self):
+        # This is a legacy edge case.
+        # In software such as boost (1.81+), mpmath and Mathematica,
+        # the value is 1.
+        assert sc.hyp1f1(-4, -3, 0) == np.inf

scipy/special/tests/test_iv_ratio.py

@@ -0,0 +1,249 @@
+# This file contains unit tests for iv_ratio() and related functions.
+
+import pytest
+import numpy as np
+from numpy.testing import assert_equal, assert_allclose
+from scipy.special._ufuncs import (  # type: ignore[attr-defined]
+    _iv_ratio as iv_ratio,
+    _iv_ratio_c as iv_ratio_c,
+)
+
+
+class TestIvRatio:
+
+    @pytest.mark.parametrize('v,x,r', [
+        (0.5, 0.16666666666666666, 0.16514041292462933),
+        (0.5, 0.3333333333333333, 0.32151273753163434),
+        (0.5, 0.5, 0.46211715726000974),
+        (0.5, 0.6666666666666666, 0.5827829453479101),
+        (0.5, 0.8333333333333335, 0.6822617902381698),
+        (1, 0.3380952380952381, 0.1666773049170313),
+        (1, 0.7083333333333333, 0.33366443586989925),
+        (1, 1.1666666666666667, 0.5023355231537423),
+        (1, 1.8666666666666665, 0.674616572252164),
+        (1, 3.560606060606061, 0.844207659503163),
+        (2.34, 0.7975238095238094, 0.16704903081553285),
+        (2.34, 1.7133333333333334, 0.3360215931268845),
+        (2.34, 2.953333333333333, 0.50681909317803),
+        (2.34, 5.0826666666666656, 0.6755252698800679),
+        (2.34, 10.869696969696973, 0.8379351104498762),
+        (56.789, 19.46575238095238, 0.1667020505391409),
+        (56.789, 42.55008333333333, 0.33353809996933026),
+        (56.789, 75.552, 0.5003932381177826),
+        (56.789, 135.76026666666667, 0.6670528221946127),
+        (56.789, 307.8642424242425, 0.8334999441460798),
+    ])
+    def test_against_reference_values(self, v, x, r):
+        """The reference values are computed using mpmath as follows.
+
+        from mpmath import mp
+        mp.dps = 100
+
+        def iv_ratio_mp(v, x):
+            return mp.besseli(v, x) / mp.besseli(v - 1, x)
+
+        def _sample(n, *, v):
+            '''Return n positive real numbers x such that iv_ratio(v, x) are
+            roughly evenly spaced over (0, 1).  The formula is taken from [1].
+
+            [1] Banerjee A., Dhillon, I. S., Ghosh, J., Sra, S. (2005).
+                "Clustering on the Unit Hypersphere using von Mises-Fisher
+                Distributions."  Journal of Machine Learning Research,
+                6(46):1345-1382.
+            '''
+            r = np.arange(1, n+1) / (n+1)
+            return r * (2*v-r*r) / (1-r*r)
+
+        for v in (0.5, 1, 2.34, 56.789):
+            xs = _sample(5, v=v)
+            for x in xs:
+                print(f"({v}, {x}, {float(iv_ratio_mp(v,x))}),")
+        """
+        assert_allclose(iv_ratio(v, x), r, rtol=4e-16, atol=0)
+
+    @pytest.mark.parametrize('v,x,r', [
+        (1, np.inf, 1),
+        (np.inf, 1, 0),
+    ])
+    def test_inf(self, v, x, r):
+        """If exactly one of v or x is inf and the other is within domain,
+        should return 0 or 1 accordingly."""
+        assert_equal(iv_ratio(v, x), r)
+
+    @pytest.mark.parametrize('v', [0.49, -np.inf, np.nan, np.inf])
+    @pytest.mark.parametrize('x', [-np.finfo(float).smallest_normal,
+                                   -np.finfo(float).smallest_subnormal,
+                                   -np.inf, np.nan, np.inf])
+    def test_nan(self, v, x):
+        """If at least one argument is out of domain, or if v = x = inf,
+        the function should return nan."""
+        assert_equal(iv_ratio(v, x), np.nan)
+
+    @pytest.mark.parametrize('v', [0.5, 1, np.finfo(float).max, np.inf])
+    def test_zero_x(self, v):
+        """If x is +/-0.0, return x to ensure iv_ratio is an odd function."""
+        assert_equal(iv_ratio(v, 0.0), 0.0)
+        assert_equal(iv_ratio(v, -0.0), -0.0)
+
+    @pytest.mark.parametrize('v,x', [
+        (1, np.finfo(float).smallest_normal),
+        (1, np.finfo(float).smallest_subnormal),
+        (1, np.finfo(float).smallest_subnormal*2),
+        (1e20, 123),
+        (np.finfo(float).max, 1),
+        (np.finfo(float).max, np.sqrt(np.finfo(float).max)),
+    ])
+    def test_tiny_x(self, v, x):
+        """If x is much less than v, the bounds
+
+                    x                                 x
+        --------------------------- <= R <= -----------------------
+        v-0.5+sqrt(x**2+(v+0.5)**2)         v-1+sqrt(x**2+(v+1)**2)
+
+        collapses to R ~= x/2v.  Test against this asymptotic expression.
+        """
+        assert_equal(iv_ratio(v, x), (0.5*x)/v)
+
+    @pytest.mark.parametrize('v,x', [
+        (1, 1e16),
+        (1e20, 1e40),
+        (np.sqrt(np.finfo(float).max), np.finfo(float).max),
+    ])
+    def test_huge_x(self, v, x):
+        """If x is much greater than v, the bounds
+
+                    x                                 x
+        --------------------------- <= R <= ---------------------------
+        v-0.5+sqrt(x**2+(v+0.5)**2)         v-0.5+sqrt(x**2+(v-0.5)**2)
+
+        collapses to R ~= 1.  Test against this asymptotic expression.
+        """
+        assert_equal(iv_ratio(v, x), 1.0)
+
+    @pytest.mark.parametrize('v,x', [
+        (np.finfo(float).max, np.finfo(float).max),
+        (np.finfo(float).max / 3, np.finfo(float).max),
+        (np.finfo(float).max, np.finfo(float).max / 3),
+    ])
+    def test_huge_v_x(self, v, x):
+        """If both x and v are very large, the bounds
+
+                    x                                 x
+        --------------------------- <= R <= -----------------------
+        v-0.5+sqrt(x**2+(v+0.5)**2)         v-1+sqrt(x**2+(v+1)**2)
+
+        collapses to R ~= x/(v+sqrt(x**2+v**2).  Test against this asymptotic
+        expression, and in particular that no numerical overflow occurs during
+        intermediate calculations.
+        """
+        t = x / v
+        expected = t / (1 + np.hypot(1, t))
+        assert_allclose(iv_ratio(v, x), expected, rtol=4e-16, atol=0)
+
+
+class TestIvRatioC:
+
+    @pytest.mark.parametrize('v,x,r', [
+        (0.5, 0.16666666666666666, 0.8348595870753707),
+        (0.5, 0.3333333333333333, 0.6784872624683657),
+        (0.5, 0.5, 0.5378828427399902),
+        (0.5, 0.6666666666666666, 0.4172170546520899),
+        (0.5, 0.8333333333333335, 0.3177382097618302),
+        (1, 0.3380952380952381, 0.8333226950829686),
+        (1, 0.7083333333333333, 0.6663355641301008),
+        (1, 1.1666666666666667, 0.4976644768462577),
+        (1, 1.8666666666666665, 0.325383427747836),
+        (1, 3.560606060606061, 0.155792340496837),
+        (2.34, 0.7975238095238094, 0.8329509691844672),
+        (2.34, 1.7133333333333334, 0.6639784068731155),
+        (2.34, 2.953333333333333, 0.49318090682197),
+        (2.34, 5.0826666666666656, 0.3244747301199321),
+        (2.34, 10.869696969696973, 0.16206488955012377),
+        (56.789, 19.46575238095238, 0.8332979494608591),
+        (56.789, 42.55008333333333, 0.6664619000306697),
+        (56.789, 75.552, 0.4996067618822174),
+        (56.789, 135.76026666666667, 0.3329471778053873),
+        (56.789, 307.8642424242425, 0.16650005585392025),
+    ])
+    def test_against_reference_values(self, v, x, r):
+        """The reference values are one minus those of TestIvRatio."""
+        assert_allclose(iv_ratio_c(v, x), r, rtol=1e-15, atol=0)
+
+    @pytest.mark.parametrize('v,x,r', [
+        (1, np.inf, 0),
+        (np.inf, 1, 1),
+    ])
+    def test_inf(self, v, x, r):
+        """If exactly one of v or x is inf and the other is within domain,
+        should return 0 or 1 accordingly."""
+        assert_equal(iv_ratio_c(v, x), r)
+
+    @pytest.mark.parametrize('v', [0.49, -np.inf, np.nan, np.inf])
+    @pytest.mark.parametrize('x', [-np.finfo(float).smallest_normal,
+                                   -np.finfo(float).smallest_subnormal,
+                                   -np.inf, np.nan, np.inf])
+    def test_nan(self, v, x):
+        """If at least one argument is out of domain, or if v = x = inf,
+        the function should return nan."""
+        assert_equal(iv_ratio_c(v, x), np.nan)
+
+    @pytest.mark.parametrize('v', [0.5, 1, np.finfo(float).max, np.inf])
+    def test_zero_x(self, v):
+        """If x is +/-0.0, return 1."""
+        assert_equal(iv_ratio_c(v, 0.0), 1.0)
+        assert_equal(iv_ratio_c(v, -0.0), 1.0)
+
+    @pytest.mark.parametrize('v,x', [
+        (1, np.finfo(float).smallest_normal),
+        (1, np.finfo(float).smallest_subnormal),
+        (1, np.finfo(float).smallest_subnormal*2),
+        (1e20, 123),
+        (np.finfo(float).max, 1),
+        (np.finfo(float).max, np.sqrt(np.finfo(float).max)),
+    ])
+    def test_tiny_x(self, v, x):
+        """If x is much less than v, the bounds
+
+                    x                                 x
+        --------------------------- <= R <= -----------------------
+        v-0.5+sqrt(x**2+(v+0.5)**2)         v-1+sqrt(x**2+(v+1)**2)
+
+        collapses to 1-R ~= 1-x/2v.  Test against this asymptotic expression.
+        """
+        assert_equal(iv_ratio_c(v, x), 1.0-(0.5*x)/v)
+
+    @pytest.mark.parametrize('v,x', [
+        (1, 1e16),
+        (1e20, 1e40),
+        (np.sqrt(np.finfo(float).max), np.finfo(float).max),
+    ])
+    def test_huge_x(self, v, x):
+        """If x is much greater than v, the bounds
+
+                    x                                 x
+        --------------------------- <= R <= ---------------------------
+        v-0.5+sqrt(x**2+(v+0.5)**2)         v-0.5+sqrt(x**2+(v-0.5)**2)
+
+        collapses to 1-R ~= (v-0.5)/x.  Test against this asymptotic expression.
+        """
+        assert_allclose(iv_ratio_c(v, x), (v-0.5)/x, rtol=1e-15, atol=0)
+
+    @pytest.mark.parametrize('v,x', [
+        (np.finfo(float).max, np.finfo(float).max),
+        (np.finfo(float).max / 3, np.finfo(float).max),
+        (np.finfo(float).max, np.finfo(float).max / 3),
+    ])
+    def test_huge_v_x(self, v, x):
+        """If both x and v are very large, the bounds
+
+                    x                                 x
+        --------------------------- <= R <= -----------------------
+        v-0.5+sqrt(x**2+(v+0.5)**2)         v-1+sqrt(x**2+(v+1)**2)
+
+        collapses to 1 - R ~= 1 - x/(v+sqrt(x**2+v**2).  Test against this
+        asymptotic expression, and in particular that no numerical overflow
+        occurs during intermediate calculations.
+        """
+        t = x / v
+        expected = 1 - t / (1 + np.hypot(1, t))
+        assert_allclose(iv_ratio_c(v, x), expected, rtol=4e-16, atol=0)