11573Srgrimes/*-
21573Srgrimes * Copyright (c) 1992, 1993
31573Srgrimes *	The Regents of the University of California.  All rights reserved.
41573Srgrimes *
51573Srgrimes * Redistribution and use in source and binary forms, with or without
61573Srgrimes * modification, are permitted provided that the following conditions
71573Srgrimes * are met:
81573Srgrimes * 1. Redistributions of source code must retain the above copyright
91573Srgrimes *    notice, this list of conditions and the following disclaimer.
101573Srgrimes * 2. Redistributions in binary form must reproduce the above copyright
111573Srgrimes *    notice, this list of conditions and the following disclaimer in the
121573Srgrimes *    documentation and/or other materials provided with the distribution.
131573Srgrimes * 3. All advertising materials mentioning features or use of this software
141573Srgrimes *    must display the following acknowledgement:
151573Srgrimes *	This product includes software developed by the University of
161573Srgrimes *	California, Berkeley and its contributors.
171573Srgrimes * 4. Neither the name of the University nor the names of its contributors
181573Srgrimes *    may be used to endorse or promote products derived from this software
191573Srgrimes *    without specific prior written permission.
201573Srgrimes *
211573Srgrimes * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
221573Srgrimes * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
231573Srgrimes * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
241573Srgrimes * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
251573Srgrimes * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
261573Srgrimes * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
271573Srgrimes * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
281573Srgrimes * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
291573Srgrimes * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
301573Srgrimes * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
311573Srgrimes * SUCH DAMAGE.
321573Srgrimes */
331573Srgrimes
34176449Sdas/* @(#)gamma.c	8.1 (Berkeley) 6/4/93 */
3593211Sbde#include <sys/cdefs.h>
3692887Sobrien__FBSDID("$FreeBSD$");
371573Srgrimes
381573Srgrimes/*
391573Srgrimes * This code by P. McIlroy, Oct 1992;
401573Srgrimes *
411573Srgrimes * The financial support of UUNET Communications Services is greatfully
421573Srgrimes * acknowledged.
431573Srgrimes */
441573Srgrimes
451573Srgrimes#include <math.h>
461573Srgrimes#include "mathimpl.h"
471573Srgrimes
481573Srgrimes/* METHOD:
491573Srgrimes * x < 0: Use reflection formula, G(x) = pi/(sin(pi*x)*x*G(x))
50169212Sbde * 	At negative integers, return NaN and raise invalid.
511573Srgrimes *
521573Srgrimes * x < 6.5:
531573Srgrimes *	Use argument reduction G(x+1) = xG(x) to reach the
541573Srgrimes *	range [1.066124,2.066124].  Use a rational
551573Srgrimes *	approximation centered at the minimum (x0+1) to
561573Srgrimes *	ensure monotonicity.
571573Srgrimes *
581573Srgrimes * x >= 6.5: Use the asymptotic approximation (Stirling's formula)
591573Srgrimes *	adjusted for equal-ripples:
601573Srgrimes *
611573Srgrimes *	log(G(x)) ~= (x-.5)*(log(x)-1) + .5(log(2*pi)-1) + 1/x*P(1/(x*x))
621573Srgrimes *
631573Srgrimes *	Keep extra precision in multiplying (x-.5)(log(x)-1), to
641573Srgrimes *	avoid premature round-off.
651573Srgrimes *
661573Srgrimes * Special values:
67169212Sbde *	-Inf:			return NaN and raise invalid;
68169212Sbde *	negative integer:	return NaN and raise invalid;
69169209Sbde *	other x ~< 177.79:	return +-0 and raise underflow;
70169209Sbde *	+-0:			return +-Inf and raise divide-by-zero;
71169212Sbde *	finite x ~> 171.63:	return +Inf and raise overflow;
72169212Sbde *	+Inf:			return +Inf;
73169209Sbde *	NaN: 			return NaN.
741573Srgrimes *
75169209Sbde * Accuracy: tgamma(x) is accurate to within
761573Srgrimes *	x > 0:  error provably < 0.9ulp.
771573Srgrimes *	Maximum observed in 1,000,000 trials was .87ulp.
781573Srgrimes *	x < 0:
791573Srgrimes *	Maximum observed error < 4ulp in 1,000,000 trials.
801573Srgrimes */
811573Srgrimes
8292917Sobrienstatic double neg_gam(double);
8392917Sobrienstatic double small_gam(double);
8492917Sobrienstatic double smaller_gam(double);
8592917Sobrienstatic struct Double large_gam(double);
8692917Sobrienstatic struct Double ratfun_gam(double, double);
871573Srgrimes
881573Srgrimes/*
891573Srgrimes * Rational approximation, A0 + x*x*P(x)/Q(x), on the interval
901573Srgrimes * [1.066.., 2.066..] accurate to 4.25e-19.
911573Srgrimes */
921573Srgrimes#define LEFT -.3955078125	/* left boundary for rat. approx */
931573Srgrimes#define x0 .461632144968362356785	/* xmin - 1 */
941573Srgrimes
951573Srgrimes#define a0_hi 0.88560319441088874992
961573Srgrimes#define a0_lo -.00000000000000004996427036469019695
971573Srgrimes#define P0	 6.21389571821820863029017800727e-01
981573Srgrimes#define P1	 2.65757198651533466104979197553e-01
991573Srgrimes#define P2	 5.53859446429917461063308081748e-03
1001573Srgrimes#define P3	 1.38456698304096573887145282811e-03
1011573Srgrimes#define P4	 2.40659950032711365819348969808e-03
1021573Srgrimes#define Q0	 1.45019531250000000000000000000e+00
1031573Srgrimes#define Q1	 1.06258521948016171343454061571e+00
1041573Srgrimes#define Q2	-2.07474561943859936441469926649e-01
1051573Srgrimes#define Q3	-1.46734131782005422506287573015e-01
1061573Srgrimes#define Q4	 3.07878176156175520361557573779e-02
1071573Srgrimes#define Q5	 5.12449347980666221336054633184e-03
1081573Srgrimes#define Q6	-1.76012741431666995019222898833e-03
1091573Srgrimes#define Q7	 9.35021023573788935372153030556e-05
1101573Srgrimes#define Q8	 6.13275507472443958924745652239e-06
1111573Srgrimes/*
1121573Srgrimes * Constants for large x approximation (x in [6, Inf])
1131573Srgrimes * (Accurate to 2.8*10^-19 absolute)
1141573Srgrimes */
1151573Srgrimes#define lns2pi_hi 0.418945312500000
1161573Srgrimes#define lns2pi_lo -.000006779295327258219670263595
1171573Srgrimes#define Pa0	 8.33333333333333148296162562474e-02
1181573Srgrimes#define Pa1	-2.77777777774548123579378966497e-03
1191573Srgrimes#define Pa2	 7.93650778754435631476282786423e-04
1201573Srgrimes#define Pa3	-5.95235082566672847950717262222e-04
1211573Srgrimes#define Pa4	 8.41428560346653702135821806252e-04
1221573Srgrimes#define Pa5	-1.89773526463879200348872089421e-03
1231573Srgrimes#define Pa6	 5.69394463439411649408050664078e-03
1241573Srgrimes#define Pa7	-1.44705562421428915453880392761e-02
1251573Srgrimes
1261573Srgrimesstatic const double zero = 0., one = 1.0, tiny = 1e-300;
127138924Sdas
1281573Srgrimesdouble
12993211Sbdetgamma(x)
1301573Srgrimes	double x;
1311573Srgrimes{
1321573Srgrimes	struct Double u;
1331573Srgrimes
1341573Srgrimes	if (x >= 6) {
1351573Srgrimes		if(x > 171.63)
136169212Sbde			return (x / zero);
1371573Srgrimes		u = large_gam(x);
1381573Srgrimes		return(__exp__D(u.a, u.b));
1391573Srgrimes	} else if (x >= 1.0 + LEFT + x0)
1401573Srgrimes		return (small_gam(x));
1411573Srgrimes	else if (x > 1.e-17)
1421573Srgrimes		return (smaller_gam(x));
1431573Srgrimes	else if (x > -1.e-17) {
144169212Sbde		if (x != 0.0)
145169212Sbde			u.a = one - tiny;	/* raise inexact */
1461573Srgrimes		return (one/x);
147138924Sdas	} else if (!finite(x))
148169212Sbde		return (x - x);		/* x is NaN or -Inf */
149138924Sdas	else
1501573Srgrimes		return (neg_gam(x));
1511573Srgrimes}
1521573Srgrimes/*
1531573Srgrimes * Accurate to max(ulp(1/128) absolute, 2^-66 relative) error.
1541573Srgrimes */
1551573Srgrimesstatic struct Double
1561573Srgrimeslarge_gam(x)
1571573Srgrimes	double x;
1581573Srgrimes{
1591573Srgrimes	double z, p;
1601573Srgrimes	struct Double t, u, v;
1611573Srgrimes
1621573Srgrimes	z = one/(x*x);
1631573Srgrimes	p = Pa0+z*(Pa1+z*(Pa2+z*(Pa3+z*(Pa4+z*(Pa5+z*(Pa6+z*Pa7))))));
1641573Srgrimes	p = p/x;
1651573Srgrimes
1661573Srgrimes	u = __log__D(x);
1671573Srgrimes	u.a -= one;
1681573Srgrimes	v.a = (x -= .5);
1691573Srgrimes	TRUNC(v.a);
1701573Srgrimes	v.b = x - v.a;
1711573Srgrimes	t.a = v.a*u.a;			/* t = (x-.5)*(log(x)-1) */
1721573Srgrimes	t.b = v.b*u.a + x*u.b;
1731573Srgrimes	/* return t.a + t.b + lns2pi_hi + lns2pi_lo + p */
1741573Srgrimes	t.b += lns2pi_lo; t.b += p;
1751573Srgrimes	u.a = lns2pi_hi + t.b; u.a += t.a;
1761573Srgrimes	u.b = t.a - u.a;
1771573Srgrimes	u.b += lns2pi_hi; u.b += t.b;
1781573Srgrimes	return (u);
1791573Srgrimes}
1801573Srgrimes/*
1811573Srgrimes * Good to < 1 ulp.  (provably .90 ulp; .87 ulp on 1,000,000 runs.)
1821573Srgrimes * It also has correct monotonicity.
1831573Srgrimes */
1841573Srgrimesstatic double
1851573Srgrimessmall_gam(x)
1861573Srgrimes	double x;
1871573Srgrimes{
188138924Sdas	double y, ym1, t;
1891573Srgrimes	struct Double yy, r;
1901573Srgrimes	y = x - one;
1911573Srgrimes	ym1 = y - one;
1921573Srgrimes	if (y <= 1.0 + (LEFT + x0)) {
1931573Srgrimes		yy = ratfun_gam(y - x0, 0);
1941573Srgrimes		return (yy.a + yy.b);
1951573Srgrimes	}
1961573Srgrimes	r.a = y;
1971573Srgrimes	TRUNC(r.a);
1981573Srgrimes	yy.a = r.a - one;
1991573Srgrimes	y = ym1;
2001573Srgrimes	yy.b = r.b = y - yy.a;
2011573Srgrimes	/* Argument reduction: G(x+1) = x*G(x) */
2021573Srgrimes	for (ym1 = y-one; ym1 > LEFT + x0; y = ym1--, yy.a--) {
2031573Srgrimes		t = r.a*yy.a;
2041573Srgrimes		r.b = r.a*yy.b + y*r.b;
2051573Srgrimes		r.a = t;
2061573Srgrimes		TRUNC(r.a);
2071573Srgrimes		r.b += (t - r.a);
2081573Srgrimes	}
20993211Sbde	/* Return r*tgamma(y). */
2101573Srgrimes	yy = ratfun_gam(y - x0, 0);
2111573Srgrimes	y = r.b*(yy.a + yy.b) + r.a*yy.b;
2121573Srgrimes	y += yy.a*r.a;
2131573Srgrimes	return (y);
2141573Srgrimes}
2151573Srgrimes/*
2161573Srgrimes * Good on (0, 1+x0+LEFT].  Accurate to 1ulp.
2171573Srgrimes */
2181573Srgrimesstatic double
2191573Srgrimessmaller_gam(x)
2201573Srgrimes	double x;
2211573Srgrimes{
2221573Srgrimes	double t, d;
2231573Srgrimes	struct Double r, xx;
2241573Srgrimes	if (x < x0 + LEFT) {
2251573Srgrimes		t = x, TRUNC(t);
2261573Srgrimes		d = (t+x)*(x-t);
2271573Srgrimes		t *= t;
2281573Srgrimes		xx.a = (t + x), TRUNC(xx.a);
2291573Srgrimes		xx.b = x - xx.a; xx.b += t; xx.b += d;
2301573Srgrimes		t = (one-x0); t += x;
2311573Srgrimes		d = (one-x0); d -= t; d += x;
2321573Srgrimes		x = xx.a + xx.b;
2331573Srgrimes	} else {
2341573Srgrimes		xx.a =  x, TRUNC(xx.a);
2351573Srgrimes		xx.b = x - xx.a;
2361573Srgrimes		t = x - x0;
2371573Srgrimes		d = (-x0 -t); d += x;
2381573Srgrimes	}
2391573Srgrimes	r = ratfun_gam(t, d);
2401573Srgrimes	d = r.a/x, TRUNC(d);
2411573Srgrimes	r.a -= d*xx.a; r.a -= d*xx.b; r.a += r.b;
2421573Srgrimes	return (d + r.a/x);
2431573Srgrimes}
2441573Srgrimes/*
2451573Srgrimes * returns (z+c)^2 * P(z)/Q(z) + a0
2461573Srgrimes */
2471573Srgrimesstatic struct Double
2481573Srgrimesratfun_gam(z, c)
2491573Srgrimes	double z, c;
2501573Srgrimes{
2511573Srgrimes	double p, q;
2521573Srgrimes	struct Double r, t;
2531573Srgrimes
2541573Srgrimes	q = Q0 +z*(Q1+z*(Q2+z*(Q3+z*(Q4+z*(Q5+z*(Q6+z*(Q7+z*Q8)))))));
2551573Srgrimes	p = P0 + z*(P1 + z*(P2 + z*(P3 + z*P4)));
2561573Srgrimes
2571573Srgrimes	/* return r.a + r.b = a0 + (z+c)^2*p/q, with r.a truncated to 26 bits. */
2581573Srgrimes	p = p/q;
2591573Srgrimes	t.a = z, TRUNC(t.a);		/* t ~= z + c */
2601573Srgrimes	t.b = (z - t.a) + c;
2611573Srgrimes	t.b *= (t.a + z);
2621573Srgrimes	q = (t.a *= t.a);		/* t = (z+c)^2 */
2631573Srgrimes	TRUNC(t.a);
2641573Srgrimes	t.b += (q - t.a);
2651573Srgrimes	r.a = p, TRUNC(r.a);		/* r = P/Q */
2661573Srgrimes	r.b = p - r.a;
2671573Srgrimes	t.b = t.b*p + t.a*r.b + a0_lo;
2681573Srgrimes	t.a *= r.a;			/* t = (z+c)^2*(P/Q) */
2691573Srgrimes	r.a = t.a + a0_hi, TRUNC(r.a);
2701573Srgrimes	r.b = ((a0_hi-r.a) + t.a) + t.b;
2711573Srgrimes	return (r);			/* r = a0 + t */
2721573Srgrimes}
2731573Srgrimes
2741573Srgrimesstatic double
2751573Srgrimesneg_gam(x)
2761573Srgrimes	double x;
2771573Srgrimes{
2781573Srgrimes	int sgn = 1;
2791573Srgrimes	struct Double lg, lsine;
2801573Srgrimes	double y, z;
2811573Srgrimes
282169212Sbde	y = ceil(x);
2831573Srgrimes	if (y == x)		/* Negative integer. */
284169212Sbde		return ((x - x) / zero);
285169212Sbde	z = y - x;
286169212Sbde	if (z > 0.5)
287169212Sbde		z = one - z;
288169212Sbde	y = 0.5 * y;
2891573Srgrimes	if (y == ceil(y))
2901573Srgrimes		sgn = -1;
2911573Srgrimes	if (z < .25)
2921573Srgrimes		z = sin(M_PI*z);
2931573Srgrimes	else
2941573Srgrimes		z = cos(M_PI*(0.5-z));
2951573Srgrimes	/* Special case: G(1-x) = Inf; G(x) may be nonzero. */
2961573Srgrimes	if (x < -170) {
2971573Srgrimes		if (x < -190)
2981573Srgrimes			return ((double)sgn*tiny*tiny);
2991573Srgrimes		y = one - x;		/* exact: 128 < |x| < 255 */
3001573Srgrimes		lg = large_gam(y);
3011573Srgrimes		lsine = __log__D(M_PI/z);	/* = TRUNC(log(u)) + small */
3021573Srgrimes		lg.a -= lsine.a;		/* exact (opposite signs) */
3031573Srgrimes		lg.b -= lsine.b;
3041573Srgrimes		y = -(lg.a + lg.b);
3051573Srgrimes		z = (y + lg.a) + lg.b;
3061573Srgrimes		y = __exp__D(y, z);
3071573Srgrimes		if (sgn < 0) y = -y;
3081573Srgrimes		return (y);
3091573Srgrimes	}
3101573Srgrimes	y = one-x;
3111573Srgrimes	if (one-y == x)
31293211Sbde		y = tgamma(y);
3131573Srgrimes	else		/* 1-x is inexact */
31493211Sbde		y = -x*tgamma(-x);
3151573Srgrimes	if (sgn < 0) y = -y;
3161573Srgrimes	return (M_PI / (y*z));
3171573Srgrimes}
318