1 2 /* @(#)e_lgamma_r.c 1.3 95/01/18 */ 3 /* 4 * ==================================================== 5 * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. 6 * 7 * Developed at SunSoft, a Sun Microsystems, Inc. business. 8 * Permission to use, copy, modify, and distribute this 9 * software is freely granted, provided that this notice 10 * is preserved. 11 * ==================================================== 12 * 13 */ 14 15 #include <sys/cdefs.h> 16 __FBSDID("$FreeBSD$"); 17 18 /* __ieee754_lgamma_r(x, signgamp) 19 * Reentrant version of the logarithm of the Gamma function 20 * with user provide pointer for the sign of Gamma(x). 21 * 22 * Method: 23 * 1. Argument Reduction for 0 < x <= 8 24 * Since gamma(1+s)=s*gamma(s), for x in [0,8], we may 25 * reduce x to a number in [1.5,2.5] by 26 * lgamma(1+s) = log(s) + lgamma(s) 27 * for example, 28 * lgamma(7.3) = log(6.3) + lgamma(6.3) 29 * = log(6.3*5.3) + lgamma(5.3) 30 * = log(6.3*5.3*4.3*3.3*2.3) + lgamma(2.3) 31 * 2. Polynomial approximation of lgamma around its 32 * minimun ymin=1.461632144968362245 to maintain monotonicity. 33 * On [ymin-0.23, ymin+0.27] (i.e., [1.23164,1.73163]), use 34 * Let z = x-ymin; 35 * lgamma(x) = -1.214862905358496078218 + z^2*poly(z) 36 * where 37 * poly(z) is a 14 degree polynomial. 38 * 2. Rational approximation in the primary interval [2,3] 39 * We use the following approximation: 40 * s = x-2.0; 41 * lgamma(x) = 0.5*s + s*P(s)/Q(s) 42 * with accuracy 43 * |P/Q - (lgamma(x)-0.5s)| < 2**-61.71 44 * Our algorithms are based on the following observation 45 * 46 * zeta(2)-1 2 zeta(3)-1 3 47 * lgamma(2+s) = s*(1-Euler) + --------- * s - --------- * s + ... 48 * 2 3 49 * 50 * where Euler = 0.5771... is the Euler constant, which is very 51 * close to 0.5. 52 * 53 * 3. For x>=8, we have 54 * lgamma(x)~(x-0.5)log(x)-x+0.5*log(2pi)+1/(12x)-1/(360x**3)+.... 55 * (better formula: 56 * lgamma(x)~(x-0.5)*(log(x)-1)-.5*(log(2pi)-1) + ...) 57 * Let z = 1/x, then we approximation 58 * f(z) = lgamma(x) - (x-0.5)(log(x)-1) 59 * by 60 * 3 5 11 61 * w = w0 + w1*z + w2*z + w3*z + ... + w6*z 62 * where 63 * |w - f(z)| < 2**-58.74 64 * 65 * 4. For negative x, since (G is gamma function) 66 * -x*G(-x)*G(x) = pi/sin(pi*x), 67 * we have 68 * G(x) = pi/(sin(pi*x)*(-x)*G(-x)) 69 * since G(-x) is positive, sign(G(x)) = sign(sin(pi*x)) for x<0 70 * Hence, for x<0, signgam = sign(sin(pi*x)) and 71 * lgamma(x) = log(|Gamma(x)|) 72 * = log(pi/(|x*sin(pi*x)|)) - lgamma(-x); 73 * Note: one should avoid compute pi*(-x) directly in the 74 * computation of sin(pi*(-x)). 75 * 76 * 5. Special Cases 77 * lgamma(2+s) ~ s*(1-Euler) for tiny s 78 * lgamma(1) = lgamma(2) = 0 79 * lgamma(x) ~ -log(|x|) for tiny x 80 * lgamma(0) = lgamma(neg.integer) = inf and raise divide-by-zero 81 * lgamma(inf) = inf 82 * lgamma(-inf) = inf (bug for bug compatible with C99!?) 83 * 84 */ 85 86 #include <float.h> 87 88 #include "math.h" 89 #include "math_private.h" 90 91 static const volatile double vzero = 0; 92 93 static const double 94 zero= 0.00000000000000000000e+00, 95 half= 5.00000000000000000000e-01, /* 0x3FE00000, 0x00000000 */ 96 one = 1.00000000000000000000e+00, /* 0x3FF00000, 0x00000000 */ 97 pi = 3.14159265358979311600e+00, /* 0x400921FB, 0x54442D18 */ 98 a0 = 7.72156649015328655494e-02, /* 0x3FB3C467, 0xE37DB0C8 */ 99 a1 = 3.22467033424113591611e-01, /* 0x3FD4A34C, 0xC4A60FAD */ 100 a2 = 6.73523010531292681824e-02, /* 0x3FB13E00, 0x1A5562A7 */ 101 a3 = 2.05808084325167332806e-02, /* 0x3F951322, 0xAC92547B */ 102 a4 = 7.38555086081402883957e-03, /* 0x3F7E404F, 0xB68FEFE8 */ 103 a5 = 2.89051383673415629091e-03, /* 0x3F67ADD8, 0xCCB7926B */ 104 a6 = 1.19270763183362067845e-03, /* 0x3F538A94, 0x116F3F5D */ 105 a7 = 5.10069792153511336608e-04, /* 0x3F40B6C6, 0x89B99C00 */ 106 a8 = 2.20862790713908385557e-04, /* 0x3F2CF2EC, 0xED10E54D */ 107 a9 = 1.08011567247583939954e-04, /* 0x3F1C5088, 0x987DFB07 */ 108 a10 = 2.52144565451257326939e-05, /* 0x3EFA7074, 0x428CFA52 */ 109 a11 = 4.48640949618915160150e-05, /* 0x3F07858E, 0x90A45837 */ 110 tc = 1.46163214496836224576e+00, /* 0x3FF762D8, 0x6356BE3F */ 111 tf = -1.21486290535849611461e-01, /* 0xBFBF19B9, 0xBCC38A42 */ 112 /* tt = -(tail of tf) */ 113 tt = -3.63867699703950536541e-18, /* 0xBC50C7CA, 0xA48A971F */ 114 t0 = 4.83836122723810047042e-01, /* 0x3FDEF72B, 0xC8EE38A2 */ 115 t1 = -1.47587722994593911752e-01, /* 0xBFC2E427, 0x8DC6C509 */ 116 t2 = 6.46249402391333854778e-02, /* 0x3FB08B42, 0x94D5419B */ 117 t3 = -3.27885410759859649565e-02, /* 0xBFA0C9A8, 0xDF35B713 */ 118 t4 = 1.79706750811820387126e-02, /* 0x3F9266E7, 0x970AF9EC */ 119 t5 = -1.03142241298341437450e-02, /* 0xBF851F9F, 0xBA91EC6A */ 120 t6 = 6.10053870246291332635e-03, /* 0x3F78FCE0, 0xE370E344 */ 121 t7 = -3.68452016781138256760e-03, /* 0xBF6E2EFF, 0xB3E914D7 */ 122 t8 = 2.25964780900612472250e-03, /* 0x3F6282D3, 0x2E15C915 */ 123 t9 = -1.40346469989232843813e-03, /* 0xBF56FE8E, 0xBF2D1AF1 */ 124 t10 = 8.81081882437654011382e-04, /* 0x3F4CDF0C, 0xEF61A8E9 */ 125 t11 = -5.38595305356740546715e-04, /* 0xBF41A610, 0x9C73E0EC */ 126 t12 = 3.15632070903625950361e-04, /* 0x3F34AF6D, 0x6C0EBBF7 */ 127 t13 = -3.12754168375120860518e-04, /* 0xBF347F24, 0xECC38C38 */ 128 t14 = 3.35529192635519073543e-04, /* 0x3F35FD3E, 0xE8C2D3F4 */ 129 u0 = -7.72156649015328655494e-02, /* 0xBFB3C467, 0xE37DB0C8 */ 130 u1 = 6.32827064025093366517e-01, /* 0x3FE4401E, 0x8B005DFF */ 131 u2 = 1.45492250137234768737e+00, /* 0x3FF7475C, 0xD119BD6F */ 132 u3 = 9.77717527963372745603e-01, /* 0x3FEF4976, 0x44EA8450 */ 133 u4 = 2.28963728064692451092e-01, /* 0x3FCD4EAE, 0xF6010924 */ 134 u5 = 1.33810918536787660377e-02, /* 0x3F8B678B, 0xBF2BAB09 */ 135 v1 = 2.45597793713041134822e+00, /* 0x4003A5D7, 0xC2BD619C */ 136 v2 = 2.12848976379893395361e+00, /* 0x40010725, 0xA42B18F5 */ 137 v3 = 7.69285150456672783825e-01, /* 0x3FE89DFB, 0xE45050AF */ 138 v4 = 1.04222645593369134254e-01, /* 0x3FBAAE55, 0xD6537C88 */ 139 v5 = 3.21709242282423911810e-03, /* 0x3F6A5ABB, 0x57D0CF61 */ 140 s0 = -7.72156649015328655494e-02, /* 0xBFB3C467, 0xE37DB0C8 */ 141 s1 = 2.14982415960608852501e-01, /* 0x3FCB848B, 0x36E20878 */ 142 s2 = 3.25778796408930981787e-01, /* 0x3FD4D98F, 0x4F139F59 */ 143 s3 = 1.46350472652464452805e-01, /* 0x3FC2BB9C, 0xBEE5F2F7 */ 144 s4 = 2.66422703033638609560e-02, /* 0x3F9B481C, 0x7E939961 */ 145 s5 = 1.84028451407337715652e-03, /* 0x3F5E26B6, 0x7368F239 */ 146 s6 = 3.19475326584100867617e-05, /* 0x3F00BFEC, 0xDD17E945 */ 147 r1 = 1.39200533467621045958e+00, /* 0x3FF645A7, 0x62C4AB74 */ 148 r2 = 7.21935547567138069525e-01, /* 0x3FE71A18, 0x93D3DCDC */ 149 r3 = 1.71933865632803078993e-01, /* 0x3FC601ED, 0xCCFBDF27 */ 150 r4 = 1.86459191715652901344e-02, /* 0x3F9317EA, 0x742ED475 */ 151 r5 = 7.77942496381893596434e-04, /* 0x3F497DDA, 0xCA41A95B */ 152 r6 = 7.32668430744625636189e-06, /* 0x3EDEBAF7, 0xA5B38140 */ 153 w0 = 4.18938533204672725052e-01, /* 0x3FDACFE3, 0x90C97D69 */ 154 w1 = 8.33333333333329678849e-02, /* 0x3FB55555, 0x5555553B */ 155 w2 = -2.77777777728775536470e-03, /* 0xBF66C16C, 0x16B02E5C */ 156 w3 = 7.93650558643019558500e-04, /* 0x3F4A019F, 0x98CF38B6 */ 157 w4 = -5.95187557450339963135e-04, /* 0xBF4380CB, 0x8C0FE741 */ 158 w5 = 8.36339918996282139126e-04, /* 0x3F4B67BA, 0x4CDAD5D1 */ 159 w6 = -1.63092934096575273989e-03; /* 0xBF5AB89D, 0x0B9E43E4 */ 160 161 /* 162 * Compute sin(pi*x) without actually doing the pi*x multiplication. 163 * sin_pi(x) is only called for x < 0 and |x| < 2**(p-1) where p is 164 * the precision of x. 165 */ 166 static double 167 sin_pi(double x) 168 { 169 volatile double vz; 170 double y,z; 171 int n; 172 173 y = -x; 174 175 vz = y+0x1p52; /* depend on 0 <= y < 0x1p52 */ 176 z = vz-0x1p52; /* rint(y) for the above range */ 177 if (z == y) 178 return zero; 179 180 vz = y+0x1p50; 181 GET_LOW_WORD(n,vz); /* bits for rounded y (units 0.25) */ 182 z = vz-0x1p50; /* y rounded to a multiple of 0.25 */ 183 if (z > y) { 184 z -= 0.25; /* adjust to round down */ 185 n--; 186 } 187 n &= 7; /* octant of y mod 2 */ 188 y = y - z + n * 0.25; /* y mod 2 */ 189 190 switch (n) { 191 case 0: y = __kernel_sin(pi*y,zero,0); break; 192 case 1: 193 case 2: y = __kernel_cos(pi*(0.5-y),zero); break; 194 case 3: 195 case 4: y = __kernel_sin(pi*(one-y),zero,0); break; 196 case 5: 197 case 6: y = -__kernel_cos(pi*(y-1.5),zero); break; 198 default: y = __kernel_sin(pi*(y-2.0),zero,0); break; 199 } 200 return -y; 201 } 202 203 204 double 205 __ieee754_lgamma_r(double x, int *signgamp) 206 { 207 double t,y,z,nadj,p,p1,p2,p3,q,r,w; 208 int32_t hx; 209 int i,ix,lx; 210 211 EXTRACT_WORDS(hx,lx,x); 212 213 /* purge off +-inf, NaN, +-0, tiny and negative arguments */ 214 *signgamp = 1; 215 ix = hx&0x7fffffff; 216 if(ix>=0x7ff00000) return x*x; 217 if((ix|lx)==0) { 218 if(hx<0) 219 *signgamp = -1; 220 return one/vzero; 221 } 222 if(ix<0x3b900000) { /* |x|<2**-70, return -log(|x|) */ 223 if(hx<0) { 224 *signgamp = -1; 225 return -__ieee754_log(-x); 226 } else return -__ieee754_log(x); 227 } 228 if(hx<0) { 229 if(ix>=0x43300000) /* |x|>=2**52, must be -integer */ 230 return one/vzero; 231 t = sin_pi(x); 232 if(t==zero) return one/vzero; /* -integer */ 233 nadj = __ieee754_log(pi/fabs(t*x)); 234 if(t<zero) *signgamp = -1; 235 x = -x; 236 } 237 238 /* purge off 1 and 2 */ 239 if((((ix-0x3ff00000)|lx)==0)||(((ix-0x40000000)|lx)==0)) r = 0; 240 /* for x < 2.0 */ 241 else if(ix<0x40000000) { 242 if(ix<=0x3feccccc) { /* lgamma(x) = lgamma(x+1)-log(x) */ 243 r = -__ieee754_log(x); 244 if(ix>=0x3FE76944) {y = one-x; i= 0;} 245 else if(ix>=0x3FCDA661) {y= x-(tc-one); i=1;} 246 else {y = x; i=2;} 247 } else { 248 r = zero; 249 if(ix>=0x3FFBB4C3) {y=2.0-x;i=0;} /* [1.7316,2] */ 250 else if(ix>=0x3FF3B4C4) {y=x-tc;i=1;} /* [1.23,1.73] */ 251 else {y=x-one;i=2;} 252 } 253 switch(i) { 254 case 0: 255 z = y*y; 256 p1 = a0+z*(a2+z*(a4+z*(a6+z*(a8+z*a10)))); 257 p2 = z*(a1+z*(a3+z*(a5+z*(a7+z*(a9+z*a11))))); 258 p = y*p1+p2; 259 r += (p-y/2); break; 260 case 1: 261 z = y*y; 262 w = z*y; 263 p1 = t0+w*(t3+w*(t6+w*(t9 +w*t12))); /* parallel comp */ 264 p2 = t1+w*(t4+w*(t7+w*(t10+w*t13))); 265 p3 = t2+w*(t5+w*(t8+w*(t11+w*t14))); 266 p = z*p1-(tt-w*(p2+y*p3)); 267 r += (tf + p); break; 268 case 2: 269 p1 = y*(u0+y*(u1+y*(u2+y*(u3+y*(u4+y*u5))))); 270 p2 = one+y*(v1+y*(v2+y*(v3+y*(v4+y*v5)))); 271 r += (-0.5*y + p1/p2); 272 } 273 } 274 else if(ix<0x40200000) { /* x < 8.0 */ 275 i = (int)x; 276 y = x-(double)i; 277 p = y*(s0+y*(s1+y*(s2+y*(s3+y*(s4+y*(s5+y*s6)))))); 278 q = one+y*(r1+y*(r2+y*(r3+y*(r4+y*(r5+y*r6))))); 279 r = half*y+p/q; 280 z = one; /* lgamma(1+s) = log(s) + lgamma(s) */ 281 switch(i) { 282 case 7: z *= (y+6); /* FALLTHRU */ 283 case 6: z *= (y+5); /* FALLTHRU */ 284 case 5: z *= (y+4); /* FALLTHRU */ 285 case 4: z *= (y+3); /* FALLTHRU */ 286 case 3: z *= (y+2); /* FALLTHRU */ 287 r += __ieee754_log(z); break; 288 } 289 /* 8.0 <= x < 2**58 */ 290 } else if (ix < 0x43900000) { 291 t = __ieee754_log(x); 292 z = one/x; 293 y = z*z; 294 w = w0+z*(w1+y*(w2+y*(w3+y*(w4+y*(w5+y*w6))))); 295 r = (x-half)*(t-one)+w; 296 } else 297 /* 2**58 <= x <= inf */ 298 r = x*(__ieee754_log(x)-one); 299 if(hx<0) r = nadj - r; 300 return r; 301 } 302 303 #if (LDBL_MANT_DIG == 53) 304 __weak_reference(lgamma_r, lgammal_r); 305 #endif 306 307