Changeset 97f9b46 in sasmodels

Timestamp:

Oct 20, 2016 3:58:39 PM (8 years ago)

Author:

Paul Kienzle <pkienzle@…>

Branches:

master, core_shell_microgels, costrafo411, magnetic_model, ticket-1257-vesicle-product, ticket_1156, ticket_1265_superball, ticket_822_more_unit_tests

Children:

aea515c

Parents:

ca9e54e

Message:

tgamma: better handling of gamma(x) for x<0

File:

• sasmodels/models/lib/sas_gamma.c (modified) (2 diffs)

sasmodels/models/lib/sas_gamma.c

@@ -5 +5 @@
 We use gamma definition Gamma(t + 1) = t * Gamma(t) to compute
 to function for values lower than 1.0. Namely Gamma(t) = 1/t * Gamma(t + 1)
+For t < 0, we use Gamma(t) = pi / ( Gamma(1 - t) * sin(pi * t) )
 */
 
@@ -137 +138 @@
 
 
-inline double sas_gamma(double x) {
-#if 1
-    // Fast reliable gamma in [-n,171], where n is a small integer
-    double norm = 1.;
-    while (x < 1.) {
-        norm *= x;
-        x += 1.0;
-    }
-    return tgamma(x)/norm;
-#else
-    // Fast reliable gamma in [0,170]
-    return tgamma(x+1)/x;
-#endif
+inline double sas_gamma(double x)
+{
+    // Note: the builtin tgamma can give slow and unreliable results for x<1.
+    // The following transform extends it to zero and to negative values.
+    // It should return NaN for zero and negative integers but doesn't.
+    // The accuracy is okay but not wonderful for negative numbers, maybe
+    // one or two digits lost in the calculation. If higher accuracy is
+    // needed, you could test the following loop:
+    //    double norm = 1.;
+    //    while (x<1.) { norm*=x; x+=1.; }
+    //    return tgamma(x)/norm;
+    return (x<0. ? M_PI/tgamma(1.-x)/sin(M_PI*x) : tgamma(x+1)/x);
 }