1 /*- 2 * Copyright (c) 1992, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 3. All advertising materials mentioning features or use of this software 14 * must display the following acknowledgement: 15 * This product includes software developed by the University of 16 * California, Berkeley and its contributors. 17 * 4. Neither the name of the University nor the names of its contributors 18 * may be used to endorse or promote products derived from this software 19 * without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 */ 33 34 #ifndef lint 35 static char sccsid[] = "@(#)gamma.c 8.1 (Berkeley) 6/4/93"; 36 #endif /* not lint */ 37 #include <sys/cdefs.h> 38 __FBSDID("$FreeBSD$"); 39 40 /* 41 * This code by P. McIlroy, Oct 1992; 42 * 43 * The financial support of UUNET Communications Services is greatfully 44 * acknowledged. 45 */ 46 47 #include <math.h> 48 #include "mathimpl.h" 49 #include <errno.h> 50 51 /* METHOD: 52 * x < 0: Use reflection formula, G(x) = pi/(sin(pi*x)*x*G(x)) 53 * At negative integers, return +Inf, and set errno. 54 * 55 * x < 6.5: 56 * Use argument reduction G(x+1) = xG(x) to reach the 57 * range [1.066124,2.066124]. Use a rational 58 * approximation centered at the minimum (x0+1) to 59 * ensure monotonicity. 60 * 61 * x >= 6.5: Use the asymptotic approximation (Stirling's formula) 62 * adjusted for equal-ripples: 63 * 64 * log(G(x)) ~= (x-.5)*(log(x)-1) + .5(log(2*pi)-1) + 1/x*P(1/(x*x)) 65 * 66 * Keep extra precision in multiplying (x-.5)(log(x)-1), to 67 * avoid premature round-off. 68 * 69 * Special values: 70 * non-positive integer: Set overflow trap; return +Inf; 71 * x > 171.63: Set overflow trap; return +Inf; 72 * NaN: Set invalid trap; return NaN 73 * 74 * Accuracy: Gamma(x) is accurate to within 75 * x > 0: error provably < 0.9ulp. 76 * Maximum observed in 1,000,000 trials was .87ulp. 77 * x < 0: 78 * Maximum observed error < 4ulp in 1,000,000 trials. 79 */ 80 81 static double neg_gam(double); 82 static double small_gam(double); 83 static double smaller_gam(double); 84 static struct Double large_gam(double); 85 static struct Double ratfun_gam(double, double); 86 87 /* 88 * Rational approximation, A0 + x*x*P(x)/Q(x), on the interval 89 * [1.066.., 2.066..] accurate to 4.25e-19. 90 */ 91 #define LEFT -.3955078125 /* left boundary for rat. approx */ 92 #define x0 .461632144968362356785 /* xmin - 1 */ 93 94 #define a0_hi 0.88560319441088874992 95 #define a0_lo -.00000000000000004996427036469019695 96 #define P0 6.21389571821820863029017800727e-01 97 #define P1 2.65757198651533466104979197553e-01 98 #define P2 5.53859446429917461063308081748e-03 99 #define P3 1.38456698304096573887145282811e-03 100 #define P4 2.40659950032711365819348969808e-03 101 #define Q0 1.45019531250000000000000000000e+00 102 #define Q1 1.06258521948016171343454061571e+00 103 #define Q2 -2.07474561943859936441469926649e-01 104 #define Q3 -1.46734131782005422506287573015e-01 105 #define Q4 3.07878176156175520361557573779e-02 106 #define Q5 5.12449347980666221336054633184e-03 107 #define Q6 -1.76012741431666995019222898833e-03 108 #define Q7 9.35021023573788935372153030556e-05 109 #define Q8 6.13275507472443958924745652239e-06 110 /* 111 * Constants for large x approximation (x in [6, Inf]) 112 * (Accurate to 2.8*10^-19 absolute) 113 */ 114 #define lns2pi_hi 0.418945312500000 115 #define lns2pi_lo -.000006779295327258219670263595 116 #define Pa0 8.33333333333333148296162562474e-02 117 #define Pa1 -2.77777777774548123579378966497e-03 118 #define Pa2 7.93650778754435631476282786423e-04 119 #define Pa3 -5.95235082566672847950717262222e-04 120 #define Pa4 8.41428560346653702135821806252e-04 121 #define Pa5 -1.89773526463879200348872089421e-03 122 #define Pa6 5.69394463439411649408050664078e-03 123 #define Pa7 -1.44705562421428915453880392761e-02 124 125 static const double zero = 0., one = 1.0, tiny = 1e-300; 126 static int endian; 127 128 /* 129 * TRUNC sets trailing bits in a floating-point number to zero. 130 * is a temporary variable. 131 */ 132 #define TRUNC(x) *(((int *) &x) + endian) &= 0xf8000000 133 134 double 135 tgamma(x) 136 double x; 137 { 138 struct Double u; 139 endian = (*(int *) &one) ? 1 : 0; 140 141 if (x >= 6) { 142 if(x > 171.63) 143 return(one/zero); 144 u = large_gam(x); 145 return(__exp__D(u.a, u.b)); 146 } else if (x >= 1.0 + LEFT + x0) 147 return (small_gam(x)); 148 else if (x > 1.e-17) 149 return (smaller_gam(x)); 150 else if (x > -1.e-17) { 151 if (x == 0.0) 152 return (one/x); 153 one+1e-20; /* Raise inexact flag. */ 154 return (one/x); 155 } else if (!finite(x)) 156 return (x*x); /* x = NaN, -Inf */ 157 else 158 return (neg_gam(x)); 159 } 160 /* 161 * Accurate to max(ulp(1/128) absolute, 2^-66 relative) error. 162 */ 163 static struct Double 164 large_gam(x) 165 double x; 166 { 167 double z, p; 168 struct Double t, u, v; 169 170 z = one/(x*x); 171 p = Pa0+z*(Pa1+z*(Pa2+z*(Pa3+z*(Pa4+z*(Pa5+z*(Pa6+z*Pa7)))))); 172 p = p/x; 173 174 u = __log__D(x); 175 u.a -= one; 176 v.a = (x -= .5); 177 TRUNC(v.a); 178 v.b = x - v.a; 179 t.a = v.a*u.a; /* t = (x-.5)*(log(x)-1) */ 180 t.b = v.b*u.a + x*u.b; 181 /* return t.a + t.b + lns2pi_hi + lns2pi_lo + p */ 182 t.b += lns2pi_lo; t.b += p; 183 u.a = lns2pi_hi + t.b; u.a += t.a; 184 u.b = t.a - u.a; 185 u.b += lns2pi_hi; u.b += t.b; 186 return (u); 187 } 188 /* 189 * Good to < 1 ulp. (provably .90 ulp; .87 ulp on 1,000,000 runs.) 190 * It also has correct monotonicity. 191 */ 192 static double 193 small_gam(x) 194 double x; 195 { 196 double y, ym1, t; 197 struct Double yy, r; 198 y = x - one; 199 ym1 = y - one; 200 if (y <= 1.0 + (LEFT + x0)) { 201 yy = ratfun_gam(y - x0, 0); 202 return (yy.a + yy.b); 203 } 204 r.a = y; 205 TRUNC(r.a); 206 yy.a = r.a - one; 207 y = ym1; 208 yy.b = r.b = y - yy.a; 209 /* Argument reduction: G(x+1) = x*G(x) */ 210 for (ym1 = y-one; ym1 > LEFT + x0; y = ym1--, yy.a--) { 211 t = r.a*yy.a; 212 r.b = r.a*yy.b + y*r.b; 213 r.a = t; 214 TRUNC(r.a); 215 r.b += (t - r.a); 216 } 217 /* Return r*tgamma(y). */ 218 yy = ratfun_gam(y - x0, 0); 219 y = r.b*(yy.a + yy.b) + r.a*yy.b; 220 y += yy.a*r.a; 221 return (y); 222 } 223 /* 224 * Good on (0, 1+x0+LEFT]. Accurate to 1ulp. 225 */ 226 static double 227 smaller_gam(x) 228 double x; 229 { 230 double t, d; 231 struct Double r, xx; 232 if (x < x0 + LEFT) { 233 t = x, TRUNC(t); 234 d = (t+x)*(x-t); 235 t *= t; 236 xx.a = (t + x), TRUNC(xx.a); 237 xx.b = x - xx.a; xx.b += t; xx.b += d; 238 t = (one-x0); t += x; 239 d = (one-x0); d -= t; d += x; 240 x = xx.a + xx.b; 241 } else { 242 xx.a = x, TRUNC(xx.a); 243 xx.b = x - xx.a; 244 t = x - x0; 245 d = (-x0 -t); d += x; 246 } 247 r = ratfun_gam(t, d); 248 d = r.a/x, TRUNC(d); 249 r.a -= d*xx.a; r.a -= d*xx.b; r.a += r.b; 250 return (d + r.a/x); 251 } 252 /* 253 * returns (z+c)^2 * P(z)/Q(z) + a0 254 */ 255 static struct Double 256 ratfun_gam(z, c) 257 double z, c; 258 { 259 double p, q; 260 struct Double r, t; 261 262 q = Q0 +z*(Q1+z*(Q2+z*(Q3+z*(Q4+z*(Q5+z*(Q6+z*(Q7+z*Q8))))))); 263 p = P0 + z*(P1 + z*(P2 + z*(P3 + z*P4))); 264 265 /* return r.a + r.b = a0 + (z+c)^2*p/q, with r.a truncated to 26 bits. */ 266 p = p/q; 267 t.a = z, TRUNC(t.a); /* t ~= z + c */ 268 t.b = (z - t.a) + c; 269 t.b *= (t.a + z); 270 q = (t.a *= t.a); /* t = (z+c)^2 */ 271 TRUNC(t.a); 272 t.b += (q - t.a); 273 r.a = p, TRUNC(r.a); /* r = P/Q */ 274 r.b = p - r.a; 275 t.b = t.b*p + t.a*r.b + a0_lo; 276 t.a *= r.a; /* t = (z+c)^2*(P/Q) */ 277 r.a = t.a + a0_hi, TRUNC(r.a); 278 r.b = ((a0_hi-r.a) + t.a) + t.b; 279 return (r); /* r = a0 + t */ 280 } 281 282 static double 283 neg_gam(x) 284 double x; 285 { 286 int sgn = 1; 287 struct Double lg, lsine; 288 double y, z; 289 290 y = floor(x + .5); 291 if (y == x) /* Negative integer. */ 292 return (one/zero); 293 z = fabs(x - y); 294 y = .5*ceil(x); 295 if (y == ceil(y)) 296 sgn = -1; 297 if (z < .25) 298 z = sin(M_PI*z); 299 else 300 z = cos(M_PI*(0.5-z)); 301 /* Special case: G(1-x) = Inf; G(x) may be nonzero. */ 302 if (x < -170) { 303 if (x < -190) 304 return ((double)sgn*tiny*tiny); 305 y = one - x; /* exact: 128 < |x| < 255 */ 306 lg = large_gam(y); 307 lsine = __log__D(M_PI/z); /* = TRUNC(log(u)) + small */ 308 lg.a -= lsine.a; /* exact (opposite signs) */ 309 lg.b -= lsine.b; 310 y = -(lg.a + lg.b); 311 z = (y + lg.a) + lg.b; 312 y = __exp__D(y, z); 313 if (sgn < 0) y = -y; 314 return (y); 315 } 316 y = one-x; 317 if (one-y == x) 318 y = tgamma(y); 319 else /* 1-x is inexact */ 320 y = -x*tgamma(-x); 321 if (sgn < 0) y = -y; 322 return (M_PI / (y*z)); 323 } 324