scipy 1.15.2__cp310-cp310-win_amd64.whl → 1.15.3__cp310-cp310-win_amd64.whl

scipy/special/_logsumexp.py

@@ -5,9 +5,9 @@ from scipy._lib._array_api import (
     array_namespace,
     xp_size,
     xp_broadcast_promote,
-    xp_real,
     xp_copy,
     xp_float_to_complex,
+    is_complex,
 )
 from scipy._lib import array_api_extra as xpx
 
@@ -219,25 +219,30 @@ def _logsumexp(a, b, axis, return_sign, xp):
     s = xp.where(s == 0, s, s/m)
 
     # Separate sign/magnitude information
-    sgn = None
-    if return_sign:
-        # Use the numpy>=2.0 convention for sign.
-        # When all array libraries agree, this can become sng = xp.sign(s).
-        sgn = _sign(s + 1, xp=xp) * _sign(m, xp=xp)
-
-        if xp.isdtype(s.dtype, "real floating"):
-            # The log functions need positive arguments
-            s = xp.where(s < -1, -s - 2, s)
-            m = xp.abs(m)
-        else:
-            # `a_max` can have a sign component for complex input
-            j = xp.asarray(1j, dtype=a_max.dtype)
-            sgn = sgn * xp.exp(xp.imag(a_max) * j)
+    # Originally, this was only performed if `return_sign=True`.
+    # However, this is also needed if any elements of `m < 0` or `s < -1`.
+    # An improvement would be to perform the calculations only on these entries.
+
+    # Use the numpy>=2.0 convention for sign.
+    # When all array libraries agree, this can become sng = xp.sign(s).
+    sgn = _sign(s + 1, xp=xp) * _sign(m, xp=xp)
+
+    if xp.isdtype(s.dtype, "real floating"):
+        # The log functions need positive arguments
+        s = xp.where(s < -1, -s - 2, s)
+        m = xp.abs(m)
+    else:
+        # `a_max` can have a sign component for complex input
+        sgn = sgn * xp.exp(xp.imag(a_max) * xp.asarray(1.0j, dtype=a_max.dtype))
 
     # Take log and undo shift
     out = xp.log1p(s) + xp.log(m) + a_max
 
-    out = xp_real(out) if return_sign else out
+    if return_sign:
+        if is_complex(out, xp):
+            out = xp.real(out)
+    elif xp.isdtype(out.dtype, 'real floating'):
+        out[sgn < 0] = xp.nan
 
     return out, sgn
 

scipy/special/tests/test_hyp2f1.py

@@ -2543,3 +2543,24 @@ class TestHyp2f1:
             if mark.name == 'parametrize'
             for case in mark.args[1]
         ]
+
+class TestHyp2f1ExtremeInputs:
+
+    @pytest.mark.parametrize("a", [1.0, 2.0, 3.0, -np.inf, np.inf])
+    @pytest.mark.parametrize("b", [3.0, 4.0, 5.0, -np.inf, np.inf])
+    @pytest.mark.parametrize("c", [3.0, 5.0, 6.0, 7.0])
+    @pytest.mark.parametrize("z", [4.0 + 1.0j])
+    def test_inf_a_b(self, a, b, c, z):
+        if np.any(np.isinf(np.asarray([a, b]))):
+            assert(np.isnan(hyp2f1(a, b, c, z)))
+
+    def test_large_a_b(self):
+        assert(np.isnan(hyp2f1(10**7, 1.0, 3.0, 4.0 + 1.0j)))
+        assert(np.isnan(hyp2f1(-10**7, 1.0, 3.0, 4.0 + 1.0j)))
+
+        assert(np.isnan(hyp2f1(1.0, 10**7, 3.0, 4.0 + 1.0j)))
+        assert(np.isnan(hyp2f1(1.0, -10**7, 3.0, 4.0 + 1.0j)))
+
+        # Already correct in main but testing for surety
+        assert(np.isnan(hyp2f1(np.inf, 1.0, 3.0, 4.0)))
+        assert(np.isnan(hyp2f1(1.0, np.inf, 3.0, 4.0)))

scipy/special/tests/test_logsumexp.py

@@ -257,6 +257,20 @@ class TestLogSumExp:
         xp_assert_close(xp.real(res), xp.real(ref))
         xp_assert_close(xp.imag(res), xp.imag(ref), atol=0, rtol=1e-15)
 
+    def test_gh22903(self, xp):
+        # gh-22903 reported that `logsumexp` produced NaN where the weight associated
+        # with the max magnitude element was negative and `return_sign=False`, even if
+        # the net result should be the log of a positive number.
+
+        # result is log of positive number
+        a = xp.asarray([3.06409428, 0.37251854, 3.87471931])
+        b = xp.asarray([1.88190708, 2.84174795, -0.85016884])
+        xp_assert_close(logsumexp(a, b=b), logsumexp(a, b=b, return_sign=True)[0])
+
+        # result is log of negative number
+        b = xp.asarray([1.88190708, 2.84174795, -3.85016884])
+        xp_assert_close(logsumexp(a, b=b), xp.asarray(xp.nan))
+
 
 class TestSoftmax:
     def test_softmax_fixtures(self):

scipy/special/xsf/hyp2f1.h

@@ -553,7 +553,7 @@ XSF_HOST_DEVICE inline std::complex<double> hyp2f1(double a, double b, double c,
      * the series at a or b of smaller magnitude. This is to ensure proper
      * handling of situations like a < c < b <= 0, a, b, c all non-positive
      * integers, where terminating at a would lead to a term of the form 0 / 0. */
-    std::uint64_t max_degree;
+    double max_degree;
     if (a_neg_int || b_neg_int) {
         if (a_neg_int && b_neg_int) {
             max_degree = a > b ? std::abs(a) : std::abs(b);
@@ -562,7 +562,7 @@ XSF_HOST_DEVICE inline std::complex<double> hyp2f1(double a, double b, double c,
         } else {
             max_degree = std::abs(b);
         }
-        if (max_degree <= UINT64_MAX) {
+        if (max_degree <= (double) UINT64_MAX) {
             auto series_generator = detail::HypergeometricSeriesGenerator(a, b, c, z);
             return detail::series_eval_fixed_length(series_generator, std::complex<double>{0.0, 0.0}, max_degree + 1);
         } else {
@@ -583,7 +583,7 @@ XSF_HOST_DEVICE inline std::complex<double> hyp2f1(double a, double b, double c,
      * (DLMF 15.8.1) */
     if (c_minus_a_neg_int || c_minus_b_neg_int) {
         max_degree = c_minus_b_neg_int ? std::abs(c - b) : std::abs(c - a);
-        if (max_degree <= UINT64_MAX) {
+        if (max_degree <= (double) UINT64_MAX) {
             result = std::pow(1.0 - z, c - a - b);
             auto series_generator = detail::HypergeometricSeriesGenerator(c - a, c - b, c, z);
             result *=

scipy/stats/_continuous_distns.py

@@ -3432,29 +3432,39 @@ class genextreme_gen(rv_continuous):
         g3 = g(3)
         g4 = g(4)
         g2mg12 = np.where(abs(c) < 1e-7, (c*np.pi)**2.0/6.0, g2-g1**2.0)
-        gam2k = np.where(abs(c) < 1e-7, np.pi**2.0/6.0,
-                         sc.expm1(sc.gammaln(2.0*c+1.0)-2*sc.gammaln(c + 1.0))/c**2.0)
+        def gam2k_f(c):
+            return sc.expm1(sc.gammaln(2.0*c+1.0)-2*sc.gammaln(c + 1.0))/c**2.0
+        gam2k = _lazywhere(abs(c) >= 1e-7, (c,), f=gam2k_f, fillvalue=np.pi**2.0/6.0)
         eps = 1e-14
-        gamk = np.where(abs(c) < eps, -_EULER, sc.expm1(sc.gammaln(c + 1))/c)
+        def gamk_f(c):
+            return sc.expm1(sc.gammaln(c + 1))/c
+        gamk = _lazywhere(abs(c) >= eps, (c,), f=gamk_f, fillvalue=-_EULER)
 
+        # mean
         m = np.where(c < -1.0, np.nan, -gamk)
+        
+        # variance
         v = np.where(c < -0.5, np.nan, g1**2.0*gam2k)
 
         # skewness
-        sk1 = _lazywhere(c >= -1./3,
-                         (c, g1, g2, g3, g2mg12),
-                         lambda c, g1, g2, g3, g2mg12:
-                             np.sign(c)*(-g3 + (g2 + 2*g2mg12)*g1)/g2mg12**1.5,
-                         fillvalue=np.nan)
-        sk = np.where(abs(c) <= eps**0.29, 12*np.sqrt(6)*_ZETA3/np.pi**3, sk1)
+        def sk1_eval(c, *args):
+            def sk1_eval_f(c, g1, g2, g3, g2mg12):
+                return np.sign(c)*(-g3 + (g2 + 2*g2mg12)*g1)/g2mg12**1.5
+            return _lazywhere(c >= -1./3, (c,)+args, f=sk1_eval_f, fillvalue=np.nan)
+
+        sk_fill = 12*np.sqrt(6)*_ZETA3/np.pi**3
+        args = (g1, g2, g3, g2mg12)
+        sk = _lazywhere(abs(c) > eps**0.29, (c,)+args, f=sk1_eval, fillvalue=sk_fill)
 
         # kurtosis
-        ku1 = _lazywhere(c >= -1./4,
-                         (g1, g2, g3, g4, g2mg12),
-                         lambda g1, g2, g3, g4, g2mg12:
-                             (g4 + (-4*g3 + 3*(g2 + g2mg12)*g1)*g1)/g2mg12**2,
-                         fillvalue=np.nan)
-        ku = np.where(abs(c) <= (eps)**0.23, 12.0/5.0, ku1-3.0)
+        def ku1_eval(c, *args):
+            def ku1_eval_f(g1, g2, g3, g4, g2mg12):
+                return (g4 + (-4*g3 + 3*(g2 + g2mg12)*g1)*g1)/g2mg12**2 - 3
+            return _lazywhere(c >= -1./4, args, ku1_eval_f, fillvalue=np.nan)
+
+        args = (g1, g2, g3, g4, g2mg12)
+        ku = _lazywhere(abs(c) > eps**0.23, (c,)+args, f=ku1_eval, fillvalue=12.0/5.0)
+
         return m, v, sk, ku
 
     def _fitstart(self, data):

scipy/stats/tests/test_distributions.py

@@ -9251,6 +9251,28 @@ def test_genextreme_give_no_warnings():
         assert_equal(number_of_warnings_thrown, 0)
 
 
+def test_moments_gh22400():
+    # Regression test for gh-22400
+    # Check for correct results at c=0 with no warnings. While we're at it,
+    # check that NaN and sufficiently negative input produce NaNs, and output
+    # with `c=1` also agrees with reference values.
+    res = np.asarray(stats.genextreme.stats([0.0, np.nan, 1, -1.5], moments='mvsk'))
+
+    # Reference values for c=0 (Wikipedia)
+    mean = np.euler_gamma
+    var = np.pi**2 / 6
+    skew = 12 * np.sqrt(6) * special.zeta(3) / np.pi**3
+    kurt = 12 / 5
+    ref_0 = [mean, var, skew, kurt]
+    ref_1 = ref_3 = [np.nan]*4
+    ref_2 = [0, 1, -2, 6]  # Wolfram Alpha, MaxStableDistribution[0, 1, -1]
+
+    assert_allclose(res[:, 0], ref_0, rtol=1e-14)
+    assert_equal(res[:, 1], ref_1)
+    assert_allclose(res[:, 2], ref_2, rtol=1e-14)
+    assert_equal(res[:, 3], ref_3)
+
+
 def test_genextreme_entropy():
     # regression test for gh-5181
     euler_gamma = 0.5772156649015329

scipy/version.py

@@ -2,10 +2,10 @@
 """
 Module to expose more detailed version info for the installed `scipy`
 """
-version = "1.15.2"
+version = "1.15.3"
 full_version = version
 short_version = version.split('.dev')[0]
-git_revision = "0f1fd4a7268b813fa2b844ca6038e4dfdf90084a"
+git_revision = "e29dcb65a2040f04819b426a04b60d44a8f69c04"
 release = 'dev' not in version and '+' not in version
 
 if not release:

scipy-1.15.3.dist-info/DELVEWHEEL

@@ -0,0 +1,2 @@
+Version: 1.10.1
+Arguments: ['C:\\Users\\runneradmin\\AppData\\Local\\Temp\\cibw-run-6w_ufqtf\\cp310-win_amd64\\build\\venv\\Scripts\\delvewheel', 'repair', '--add-path', 'C:/Users/runneradmin/AppData/Local/Temp/cibw-run-6w_ufqtf/cp310-win_amd64/build/venv/Lib/site-packages/scipy_openblas32/lib', '--no-dll', 'libsf_error_state.dll', '-w', 'C:\\Users\\runneradmin\\AppData\\Local\\Temp\\cibw-run-6w_ufqtf\\cp310-win_amd64\\repaired_wheel', 'C:\\Users\\runneradmin\\AppData\\Local\\Temp\\cibw-run-6w_ufqtf\\cp310-win_amd64\\built_wheel\\scipy-1.15.3-cp310-cp310-win_amd64.whl']

{scipy-1.15.2.dist-info → scipy-1.15.3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: scipy
-Version: 1.15.2
+Version: 1.15.3
 Summary: Fundamental algorithms for scientific computing in Python
 Maintainer-Email: SciPy Developers <scipy-dev@python.org>
 License: Copyright (c) 2001-2002 Enthought, Inc. 2003-2024, SciPy Developers.
@@ -968,7 +968,7 @@ Requires-Dist: numpydoc; extra == "doc"
 Requires-Dist: jupytext; extra == "doc"
 Requires-Dist: myst-nb; extra == "doc"
 Requires-Dist: pooch; extra == "doc"
-Requires-Dist: jupyterlite-sphinx>=0.16.5; extra == "doc"
+Requires-Dist: jupyterlite-sphinx>=0.19.1; extra == "doc"
 Requires-Dist: jupyterlite-pyodide-kernel; extra == "doc"
 Provides-Extra: dev
 Requires-Dist: mypy==1.10.0; extra == "dev"