Restore Scipy-like precision for betainc

pytensor/scalar/c_code/incbet.c

@@ -16,10 +16,12 @@ Copyright 1984, 1995, 2000 by Stephen L. Moshier
 #include <numpy/npy_math.h>
 
 
-#define MINLOG     -170.0
-#define MAXLOG     +170.0
+// Constants borrowed from Scipy
+// https://github.com/scipy/scipy/blob/81c53d48a290b604ec5faa34c0a7d48537b487d6/scipy/special/special/cephes/const.h#L65-L78
+#define MINLOG     -7.451332191019412076235E2    // log 2**-1022
+#define MAXLOG     7.09782712893383996732E2      // log(DBL_MAX)
 #define MAXGAM     171.624376956302725
-#define EPSILON     2.2204460492503131e-16
+#define EPSILON     1.11022302462515654042e-16    // 2**-53
 
 DEVICE static double pseries(double, double, double);
 DEVICE static double incbcf(double, double, double);

pytensor/scalar/math.py

@@ -1497,7 +1497,7 @@ def c_code(self, node, name, inp, out, sub):
         raise NotImplementedError("type not supported", type)
 
     def c_code_cache_version(self):
-        return (1,)
+        return (2,)
 
 
 betainc = BetaInc(upgrade_to_float_no_complex, name="betainc")

tests/scalar/test_math.py

@@ -99,12 +99,17 @@ def test_gammau_nan_c():
     assert np.isnan(test_func(-1, -1))
 
 
-def test_betainc():
+@pytest.mark.parametrize("linker", ["py", "c"])
+def test_betainc(linker):
     a, b, x = pt.scalars("a", "b", "x")
     res = betainc(a, b, x)
-    test_func = function([a, b, x], res, mode=Mode("py"))
+    test_func = function([a, b, x], res, mode=Mode(linker=linker, optimizer="fast_run"))
     assert np.isclose(test_func(15, 10, 0.7), sp.betainc(15, 10, 0.7))
 
+    # Regression test for https://github.com/pymc-devs/pytensor/issues/906
+    if res.dtype == "float64":
+        assert test_func(100, 1.0, 0.1) > 0
+
 
 def test_betainc_derivative_nan():
     a, b, x = pt.scalars("a", "b", "x")