xref: /original-bsd/lib/libm/common_source/gamma.c (revision 71eaf591)
1 /*-
2  * Copyright (c) 1992, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * %sccs.include.redist.c%
6  */
7 
8 #ifndef lint
9 static char sccsid[] = "@(#)gamma.c	8.1 (Berkeley) 06/04/93";
10 #endif /* not lint */
11 
12 /*
13  * This code by P. McIlroy, Oct 1992;
14  *
15  * The financial support of UUNET Communications Services is greatfully
16  * acknowledged.
17  */
18 
19 #include <math.h>
20 #include "mathimpl.h"
21 #include <errno.h>
22 
23 /* METHOD:
24  * x < 0: Use reflection formula, G(x) = pi/(sin(pi*x)*x*G(x))
25  * 	At negative integers, return +Inf, and set errno.
26  *
27  * x < 6.5:
28  *	Use argument reduction G(x+1) = xG(x) to reach the
29  *	range [1.066124,2.066124].  Use a rational
30  *	approximation centered at the minimum (x0+1) to
31  *	ensure monotonicity.
32  *
33  * x >= 6.5: Use the asymptotic approximation (Stirling's formula)
34  *	adjusted for equal-ripples:
35  *
36  *	log(G(x)) ~= (x-.5)*(log(x)-1) + .5(log(2*pi)-1) + 1/x*P(1/(x*x))
37  *
38  *	Keep extra precision in multiplying (x-.5)(log(x)-1), to
39  *	avoid premature round-off.
40  *
41  * Special values:
42  *	non-positive integer:	Set overflow trap; return +Inf;
43  *	x > 171.63:		Set overflow trap; return +Inf;
44  *	NaN: 			Set invalid trap;  return NaN
45  *
46  * Accuracy: Gamma(x) is accurate to within
47  *	x > 0:  error provably < 0.9ulp.
48  *	Maximum observed in 1,000,000 trials was .87ulp.
49  *	x < 0:
50  *	Maximum observed error < 4ulp in 1,000,000 trials.
51  */
52 
53 static double neg_gam __P((double));
54 static double small_gam __P((double));
55 static double smaller_gam __P((double));
56 static struct Double large_gam __P((double));
57 static struct Double ratfun_gam __P((double, double));
58 
59 /*
60  * Rational approximation, A0 + x*x*P(x)/Q(x), on the interval
61  * [1.066.., 2.066..] accurate to 4.25e-19.
62  */
63 #define LEFT -.3955078125	/* left boundary for rat. approx */
64 #define x0 .461632144968362356785	/* xmin - 1 */
65 
66 #define a0_hi 0.88560319441088874992
67 #define a0_lo -.00000000000000004996427036469019695
68 #define P0	 6.21389571821820863029017800727e-01
69 #define P1	 2.65757198651533466104979197553e-01
70 #define P2	 5.53859446429917461063308081748e-03
71 #define P3	 1.38456698304096573887145282811e-03
72 #define P4	 2.40659950032711365819348969808e-03
73 #define Q0	 1.45019531250000000000000000000e+00
74 #define Q1	 1.06258521948016171343454061571e+00
75 #define Q2	-2.07474561943859936441469926649e-01
76 #define Q3	-1.46734131782005422506287573015e-01
77 #define Q4	 3.07878176156175520361557573779e-02
78 #define Q5	 5.12449347980666221336054633184e-03
79 #define Q6	-1.76012741431666995019222898833e-03
80 #define Q7	 9.35021023573788935372153030556e-05
81 #define Q8	 6.13275507472443958924745652239e-06
82 /*
83  * Constants for large x approximation (x in [6, Inf])
84  * (Accurate to 2.8*10^-19 absolute)
85  */
86 #define lns2pi_hi 0.418945312500000
87 #define lns2pi_lo -.000006779295327258219670263595
88 #define Pa0	 8.33333333333333148296162562474e-02
89 #define Pa1	-2.77777777774548123579378966497e-03
90 #define Pa2	 7.93650778754435631476282786423e-04
91 #define Pa3	-5.95235082566672847950717262222e-04
92 #define Pa4	 8.41428560346653702135821806252e-04
93 #define Pa5	-1.89773526463879200348872089421e-03
94 #define Pa6	 5.69394463439411649408050664078e-03
95 #define Pa7	-1.44705562421428915453880392761e-02
96 
97 static const double zero = 0., one = 1.0, tiny = 1e-300;
98 static int endian;
99 /*
100  * TRUNC sets trailing bits in a floating-point number to zero.
101  * is a temporary variable.
102  */
103 #if defined(vax) || defined(tahoe)
104 #define _IEEE		0
105 #define TRUNC(x)	x = (double) (float) (x)
106 #else
107 #define _IEEE		1
108 #define TRUNC(x)	*(((int *) &x) + endian) &= 0xf8000000
109 #define infnan(x)	0.0
110 #endif
111 
112 double
113 gamma(x)
114 	double x;
115 {
116 	struct Double u;
117 	endian = (*(int *) &one) ? 1 : 0;
118 
119 	if (x >= 6) {
120 		if(x > 171.63)
121 			return(one/zero);
122 		u = large_gam(x);
123 		return(__exp__D(u.a, u.b));
124 	} else if (x >= 1.0 + LEFT + x0)
125 		return (small_gam(x));
126 	else if (x > 1.e-17)
127 		return (smaller_gam(x));
128 	else if (x > -1.e-17) {
129 		if (x == 0.0)
130 			if (!_IEEE) return (infnan(ERANGE));
131 			else return (one/x);
132 		one+1e-20;		/* Raise inexact flag. */
133 		return (one/x);
134 	} else if (!finite(x)) {
135 		if (_IEEE)		/* x = NaN, -Inf */
136 			return (x*x);
137 		else
138 			return (infnan(EDOM));
139 	 } else
140 		return (neg_gam(x));
141 }
142 /*
143  * Accurate to max(ulp(1/128) absolute, 2^-66 relative) error.
144  */
145 static struct Double
146 large_gam(x)
147 	double x;
148 {
149 	double z, p;
150 	int i;
151 	struct Double t, u, v;
152 
153 	z = one/(x*x);
154 	p = Pa0+z*(Pa1+z*(Pa2+z*(Pa3+z*(Pa4+z*(Pa5+z*(Pa6+z*Pa7))))));
155 	p = p/x;
156 
157 	u = __log__D(x);
158 	u.a -= one;
159 	v.a = (x -= .5);
160 	TRUNC(v.a);
161 	v.b = x - v.a;
162 	t.a = v.a*u.a;			/* t = (x-.5)*(log(x)-1) */
163 	t.b = v.b*u.a + x*u.b;
164 	/* return t.a + t.b + lns2pi_hi + lns2pi_lo + p */
165 	t.b += lns2pi_lo; t.b += p;
166 	u.a = lns2pi_hi + t.b; u.a += t.a;
167 	u.b = t.a - u.a;
168 	u.b += lns2pi_hi; u.b += t.b;
169 	return (u);
170 }
171 /*
172  * Good to < 1 ulp.  (provably .90 ulp; .87 ulp on 1,000,000 runs.)
173  * It also has correct monotonicity.
174  */
175 static double
176 small_gam(x)
177 	double x;
178 {
179 	double y, ym1, t, x1;
180 	struct Double yy, r;
181 	y = x - one;
182 	ym1 = y - one;
183 	if (y <= 1.0 + (LEFT + x0)) {
184 		yy = ratfun_gam(y - x0, 0);
185 		return (yy.a + yy.b);
186 	}
187 	r.a = y;
188 	TRUNC(r.a);
189 	yy.a = r.a - one;
190 	y = ym1;
191 	yy.b = r.b = y - yy.a;
192 	/* Argument reduction: G(x+1) = x*G(x) */
193 	for (ym1 = y-one; ym1 > LEFT + x0; y = ym1--, yy.a--) {
194 		t = r.a*yy.a;
195 		r.b = r.a*yy.b + y*r.b;
196 		r.a = t;
197 		TRUNC(r.a);
198 		r.b += (t - r.a);
199 	}
200 	/* Return r*gamma(y). */
201 	yy = ratfun_gam(y - x0, 0);
202 	y = r.b*(yy.a + yy.b) + r.a*yy.b;
203 	y += yy.a*r.a;
204 	return (y);
205 }
206 /*
207  * Good on (0, 1+x0+LEFT].  Accurate to 1ulp.
208  */
209 static double
210 smaller_gam(x)
211 	double x;
212 {
213 	double t, d;
214 	struct Double r, xx;
215 	if (x < x0 + LEFT) {
216 		t = x, TRUNC(t);
217 		d = (t+x)*(x-t);
218 		t *= t;
219 		xx.a = (t + x), TRUNC(xx.a);
220 		xx.b = x - xx.a; xx.b += t; xx.b += d;
221 		t = (one-x0); t += x;
222 		d = (one-x0); d -= t; d += x;
223 		x = xx.a + xx.b;
224 	} else {
225 		xx.a =  x, TRUNC(xx.a);
226 		xx.b = x - xx.a;
227 		t = x - x0;
228 		d = (-x0 -t); d += x;
229 	}
230 	r = ratfun_gam(t, d);
231 	d = r.a/x, TRUNC(d);
232 	r.a -= d*xx.a; r.a -= d*xx.b; r.a += r.b;
233 	return (d + r.a/x);
234 }
235 /*
236  * returns (z+c)^2 * P(z)/Q(z) + a0
237  */
238 static struct Double
239 ratfun_gam(z, c)
240 	double z, c;
241 {
242 	int i;
243 	double p, q;
244 	struct Double r, t;
245 
246 	q = Q0 +z*(Q1+z*(Q2+z*(Q3+z*(Q4+z*(Q5+z*(Q6+z*(Q7+z*Q8)))))));
247 	p = P0 + z*(P1 + z*(P2 + z*(P3 + z*P4)));
248 
249 	/* return r.a + r.b = a0 + (z+c)^2*p/q, with r.a truncated to 26 bits. */
250 	p = p/q;
251 	t.a = z, TRUNC(t.a);		/* t ~= z + c */
252 	t.b = (z - t.a) + c;
253 	t.b *= (t.a + z);
254 	q = (t.a *= t.a);		/* t = (z+c)^2 */
255 	TRUNC(t.a);
256 	t.b += (q - t.a);
257 	r.a = p, TRUNC(r.a);		/* r = P/Q */
258 	r.b = p - r.a;
259 	t.b = t.b*p + t.a*r.b + a0_lo;
260 	t.a *= r.a;			/* t = (z+c)^2*(P/Q) */
261 	r.a = t.a + a0_hi, TRUNC(r.a);
262 	r.b = ((a0_hi-r.a) + t.a) + t.b;
263 	return (r);			/* r = a0 + t */
264 }
265 
266 static double
267 neg_gam(x)
268 	double x;
269 {
270 	int sgn = 1;
271 	struct Double lg, lsine;
272 	double y, z;
273 
274 	y = floor(x + .5);
275 	if (y == x)		/* Negative integer. */
276 		if(!_IEEE)
277 			return (infnan(ERANGE));
278 		else
279 			return (one/zero);
280 	z = fabs(x - y);
281 	y = .5*ceil(x);
282 	if (y == ceil(y))
283 		sgn = -1;
284 	if (z < .25)
285 		z = sin(M_PI*z);
286 	else
287 		z = cos(M_PI*(0.5-z));
288 	/* Special case: G(1-x) = Inf; G(x) may be nonzero. */
289 	if (x < -170) {
290 		if (x < -190)
291 			return ((double)sgn*tiny*tiny);
292 		y = one - x;		/* exact: 128 < |x| < 255 */
293 		lg = large_gam(y);
294 		lsine = __log__D(M_PI/z);	/* = TRUNC(log(u)) + small */
295 		lg.a -= lsine.a;		/* exact (opposite signs) */
296 		lg.b -= lsine.b;
297 		y = -(lg.a + lg.b);
298 		z = (y + lg.a) + lg.b;
299 		y = __exp__D(y, z);
300 		if (sgn < 0) y = -y;
301 		return (y);
302 	}
303 	y = one-x;
304 	if (one-y == x)
305 		y = gamma(y);
306 	else		/* 1-x is inexact */
307 		y = -x*gamma(-x);
308 	if (sgn < 0) y = -y;
309 	return (M_PI / (y*z));
310 }
311