1 |
|
2 /* |
|
3 * Copyright (c) 1998, 2001, Oracle and/or its affiliates. All rights reserved. |
|
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
|
5 * |
|
6 * This code is free software; you can redistribute it and/or modify it |
|
7 * under the terms of the GNU General Public License version 2 only, as |
|
8 * published by the Free Software Foundation. Oracle designates this |
|
9 * particular file as subject to the "Classpath" exception as provided |
|
10 * by Oracle in the LICENSE file that accompanied this code. |
|
11 * |
|
12 * This code is distributed in the hope that it will be useful, but WITHOUT |
|
13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
|
14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
|
15 * version 2 for more details (a copy is included in the LICENSE file that |
|
16 * accompanied this code). |
|
17 * |
|
18 * You should have received a copy of the GNU General Public License version |
|
19 * 2 along with this work; if not, write to the Free Software Foundation, |
|
20 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
|
21 * |
|
22 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
|
23 * or visit www.oracle.com if you need additional information or have any |
|
24 * questions. |
|
25 */ |
|
26 |
|
27 /* double erf(double x) |
|
28 * double erfc(double x) |
|
29 * x |
|
30 * 2 |\ |
|
31 * erf(x) = --------- | exp(-t*t)dt |
|
32 * sqrt(pi) \| |
|
33 * 0 |
|
34 * |
|
35 * erfc(x) = 1-erf(x) |
|
36 * Note that |
|
37 * erf(-x) = -erf(x) |
|
38 * erfc(-x) = 2 - erfc(x) |
|
39 * |
|
40 * Method: |
|
41 * 1. For |x| in [0, 0.84375] |
|
42 * erf(x) = x + x*R(x^2) |
|
43 * erfc(x) = 1 - erf(x) if x in [-.84375,0.25] |
|
44 * = 0.5 + ((0.5-x)-x*R) if x in [0.25,0.84375] |
|
45 * where R = P/Q where P is an odd poly of degree 8 and |
|
46 * Q is an odd poly of degree 10. |
|
47 * -57.90 |
|
48 * | R - (erf(x)-x)/x | <= 2 |
|
49 * |
|
50 * |
|
51 * Remark. The formula is derived by noting |
|
52 * erf(x) = (2/sqrt(pi))*(x - x^3/3 + x^5/10 - x^7/42 + ....) |
|
53 * and that |
|
54 * 2/sqrt(pi) = 1.128379167095512573896158903121545171688 |
|
55 * is close to one. The interval is chosen because the fix |
|
56 * point of erf(x) is near 0.6174 (i.e., erf(x)=x when x is |
|
57 * near 0.6174), and by some experiment, 0.84375 is chosen to |
|
58 * guarantee the error is less than one ulp for erf. |
|
59 * |
|
60 * 2. For |x| in [0.84375,1.25], let s = |x| - 1, and |
|
61 * c = 0.84506291151 rounded to single (24 bits) |
|
62 * erf(x) = sign(x) * (c + P1(s)/Q1(s)) |
|
63 * erfc(x) = (1-c) - P1(s)/Q1(s) if x > 0 |
|
64 * 1+(c+P1(s)/Q1(s)) if x < 0 |
|
65 * |P1/Q1 - (erf(|x|)-c)| <= 2**-59.06 |
|
66 * Remark: here we use the taylor series expansion at x=1. |
|
67 * erf(1+s) = erf(1) + s*Poly(s) |
|
68 * = 0.845.. + P1(s)/Q1(s) |
|
69 * That is, we use rational approximation to approximate |
|
70 * erf(1+s) - (c = (single)0.84506291151) |
|
71 * Note that |P1/Q1|< 0.078 for x in [0.84375,1.25] |
|
72 * where |
|
73 * P1(s) = degree 6 poly in s |
|
74 * Q1(s) = degree 6 poly in s |
|
75 * |
|
76 * 3. For x in [1.25,1/0.35(~2.857143)], |
|
77 * erfc(x) = (1/x)*exp(-x*x-0.5625+R1/S1) |
|
78 * erf(x) = 1 - erfc(x) |
|
79 * where |
|
80 * R1(z) = degree 7 poly in z, (z=1/x^2) |
|
81 * S1(z) = degree 8 poly in z |
|
82 * |
|
83 * 4. For x in [1/0.35,28] |
|
84 * erfc(x) = (1/x)*exp(-x*x-0.5625+R2/S2) if x > 0 |
|
85 * = 2.0 - (1/x)*exp(-x*x-0.5625+R2/S2) if -6<x<0 |
|
86 * = 2.0 - tiny (if x <= -6) |
|
87 * erf(x) = sign(x)*(1.0 - erfc(x)) if x < 6, else |
|
88 * erf(x) = sign(x)*(1.0 - tiny) |
|
89 * where |
|
90 * R2(z) = degree 6 poly in z, (z=1/x^2) |
|
91 * S2(z) = degree 7 poly in z |
|
92 * |
|
93 * Note1: |
|
94 * To compute exp(-x*x-0.5625+R/S), let s be a single |
|
95 * precision number and s := x; then |
|
96 * -x*x = -s*s + (s-x)*(s+x) |
|
97 * exp(-x*x-0.5626+R/S) = |
|
98 * exp(-s*s-0.5625)*exp((s-x)*(s+x)+R/S); |
|
99 * Note2: |
|
100 * Here 4 and 5 make use of the asymptotic series |
|
101 * exp(-x*x) |
|
102 * erfc(x) ~ ---------- * ( 1 + Poly(1/x^2) ) |
|
103 * x*sqrt(pi) |
|
104 * We use rational approximation to approximate |
|
105 * g(s)=f(1/x^2) = log(erfc(x)*x) - x*x + 0.5625 |
|
106 * Here is the error bound for R1/S1 and R2/S2 |
|
107 * |R1/S1 - f(x)| < 2**(-62.57) |
|
108 * |R2/S2 - f(x)| < 2**(-61.52) |
|
109 * |
|
110 * 5. For inf > x >= 28 |
|
111 * erf(x) = sign(x) *(1 - tiny) (raise inexact) |
|
112 * erfc(x) = tiny*tiny (raise underflow) if x > 0 |
|
113 * = 2 - tiny if x<0 |
|
114 * |
|
115 * 7. Special case: |
|
116 * erf(0) = 0, erf(inf) = 1, erf(-inf) = -1, |
|
117 * erfc(0) = 1, erfc(inf) = 0, erfc(-inf) = 2, |
|
118 * erfc/erf(NaN) is NaN |
|
119 */ |
|
120 |
|
121 |
|
122 #include "fdlibm.h" |
|
123 |
|
124 #ifdef __STDC__ |
|
125 static const double |
|
126 #else |
|
127 static double |
|
128 #endif |
|
129 tiny = 1e-300, |
|
130 half= 5.00000000000000000000e-01, /* 0x3FE00000, 0x00000000 */ |
|
131 one = 1.00000000000000000000e+00, /* 0x3FF00000, 0x00000000 */ |
|
132 two = 2.00000000000000000000e+00, /* 0x40000000, 0x00000000 */ |
|
133 /* c = (float)0.84506291151 */ |
|
134 erx = 8.45062911510467529297e-01, /* 0x3FEB0AC1, 0x60000000 */ |
|
135 /* |
|
136 * Coefficients for approximation to erf on [0,0.84375] |
|
137 */ |
|
138 efx = 1.28379167095512586316e-01, /* 0x3FC06EBA, 0x8214DB69 */ |
|
139 efx8= 1.02703333676410069053e+00, /* 0x3FF06EBA, 0x8214DB69 */ |
|
140 pp0 = 1.28379167095512558561e-01, /* 0x3FC06EBA, 0x8214DB68 */ |
|
141 pp1 = -3.25042107247001499370e-01, /* 0xBFD4CD7D, 0x691CB913 */ |
|
142 pp2 = -2.84817495755985104766e-02, /* 0xBF9D2A51, 0xDBD7194F */ |
|
143 pp3 = -5.77027029648944159157e-03, /* 0xBF77A291, 0x236668E4 */ |
|
144 pp4 = -2.37630166566501626084e-05, /* 0xBEF8EAD6, 0x120016AC */ |
|
145 qq1 = 3.97917223959155352819e-01, /* 0x3FD97779, 0xCDDADC09 */ |
|
146 qq2 = 6.50222499887672944485e-02, /* 0x3FB0A54C, 0x5536CEBA */ |
|
147 qq3 = 5.08130628187576562776e-03, /* 0x3F74D022, 0xC4D36B0F */ |
|
148 qq4 = 1.32494738004321644526e-04, /* 0x3F215DC9, 0x221C1A10 */ |
|
149 qq5 = -3.96022827877536812320e-06, /* 0xBED09C43, 0x42A26120 */ |
|
150 /* |
|
151 * Coefficients for approximation to erf in [0.84375,1.25] |
|
152 */ |
|
153 pa0 = -2.36211856075265944077e-03, /* 0xBF6359B8, 0xBEF77538 */ |
|
154 pa1 = 4.14856118683748331666e-01, /* 0x3FDA8D00, 0xAD92B34D */ |
|
155 pa2 = -3.72207876035701323847e-01, /* 0xBFD7D240, 0xFBB8C3F1 */ |
|
156 pa3 = 3.18346619901161753674e-01, /* 0x3FD45FCA, 0x805120E4 */ |
|
157 pa4 = -1.10894694282396677476e-01, /* 0xBFBC6398, 0x3D3E28EC */ |
|
158 pa5 = 3.54783043256182359371e-02, /* 0x3FA22A36, 0x599795EB */ |
|
159 pa6 = -2.16637559486879084300e-03, /* 0xBF61BF38, 0x0A96073F */ |
|
160 qa1 = 1.06420880400844228286e-01, /* 0x3FBB3E66, 0x18EEE323 */ |
|
161 qa2 = 5.40397917702171048937e-01, /* 0x3FE14AF0, 0x92EB6F33 */ |
|
162 qa3 = 7.18286544141962662868e-02, /* 0x3FB2635C, 0xD99FE9A7 */ |
|
163 qa4 = 1.26171219808761642112e-01, /* 0x3FC02660, 0xE763351F */ |
|
164 qa5 = 1.36370839120290507362e-02, /* 0x3F8BEDC2, 0x6B51DD1C */ |
|
165 qa6 = 1.19844998467991074170e-02, /* 0x3F888B54, 0x5735151D */ |
|
166 /* |
|
167 * Coefficients for approximation to erfc in [1.25,1/0.35] |
|
168 */ |
|
169 ra0 = -9.86494403484714822705e-03, /* 0xBF843412, 0x600D6435 */ |
|
170 ra1 = -6.93858572707181764372e-01, /* 0xBFE63416, 0xE4BA7360 */ |
|
171 ra2 = -1.05586262253232909814e+01, /* 0xC0251E04, 0x41B0E726 */ |
|
172 ra3 = -6.23753324503260060396e+01, /* 0xC04F300A, 0xE4CBA38D */ |
|
173 ra4 = -1.62396669462573470355e+02, /* 0xC0644CB1, 0x84282266 */ |
|
174 ra5 = -1.84605092906711035994e+02, /* 0xC067135C, 0xEBCCABB2 */ |
|
175 ra6 = -8.12874355063065934246e+01, /* 0xC0545265, 0x57E4D2F2 */ |
|
176 ra7 = -9.81432934416914548592e+00, /* 0xC023A0EF, 0xC69AC25C */ |
|
177 sa1 = 1.96512716674392571292e+01, /* 0x4033A6B9, 0xBD707687 */ |
|
178 sa2 = 1.37657754143519042600e+02, /* 0x4061350C, 0x526AE721 */ |
|
179 sa3 = 4.34565877475229228821e+02, /* 0x407B290D, 0xD58A1A71 */ |
|
180 sa4 = 6.45387271733267880336e+02, /* 0x40842B19, 0x21EC2868 */ |
|
181 sa5 = 4.29008140027567833386e+02, /* 0x407AD021, 0x57700314 */ |
|
182 sa6 = 1.08635005541779435134e+02, /* 0x405B28A3, 0xEE48AE2C */ |
|
183 sa7 = 6.57024977031928170135e+00, /* 0x401A47EF, 0x8E484A93 */ |
|
184 sa8 = -6.04244152148580987438e-02, /* 0xBFAEEFF2, 0xEE749A62 */ |
|
185 /* |
|
186 * Coefficients for approximation to erfc in [1/.35,28] |
|
187 */ |
|
188 rb0 = -9.86494292470009928597e-03, /* 0xBF843412, 0x39E86F4A */ |
|
189 rb1 = -7.99283237680523006574e-01, /* 0xBFE993BA, 0x70C285DE */ |
|
190 rb2 = -1.77579549177547519889e+01, /* 0xC031C209, 0x555F995A */ |
|
191 rb3 = -1.60636384855821916062e+02, /* 0xC064145D, 0x43C5ED98 */ |
|
192 rb4 = -6.37566443368389627722e+02, /* 0xC083EC88, 0x1375F228 */ |
|
193 rb5 = -1.02509513161107724954e+03, /* 0xC0900461, 0x6A2E5992 */ |
|
194 rb6 = -4.83519191608651397019e+02, /* 0xC07E384E, 0x9BDC383F */ |
|
195 sb1 = 3.03380607434824582924e+01, /* 0x403E568B, 0x261D5190 */ |
|
196 sb2 = 3.25792512996573918826e+02, /* 0x40745CAE, 0x221B9F0A */ |
|
197 sb3 = 1.53672958608443695994e+03, /* 0x409802EB, 0x189D5118 */ |
|
198 sb4 = 3.19985821950859553908e+03, /* 0x40A8FFB7, 0x688C246A */ |
|
199 sb5 = 2.55305040643316442583e+03, /* 0x40A3F219, 0xCEDF3BE6 */ |
|
200 sb6 = 4.74528541206955367215e+02, /* 0x407DA874, 0xE79FE763 */ |
|
201 sb7 = -2.24409524465858183362e+01; /* 0xC03670E2, 0x42712D62 */ |
|
202 |
|
203 #ifdef __STDC__ |
|
204 double erf(double x) |
|
205 #else |
|
206 double erf(x) |
|
207 double x; |
|
208 #endif |
|
209 { |
|
210 int hx,ix,i; |
|
211 double R,S,P,Q,s,y,z,r; |
|
212 hx = __HI(x); |
|
213 ix = hx&0x7fffffff; |
|
214 if(ix>=0x7ff00000) { /* erf(nan)=nan */ |
|
215 i = ((unsigned)hx>>31)<<1; |
|
216 return (double)(1-i)+one/x; /* erf(+-inf)=+-1 */ |
|
217 } |
|
218 |
|
219 if(ix < 0x3feb0000) { /* |x|<0.84375 */ |
|
220 if(ix < 0x3e300000) { /* |x|<2**-28 */ |
|
221 if (ix < 0x00800000) |
|
222 return 0.125*(8.0*x+efx8*x); /*avoid underflow */ |
|
223 return x + efx*x; |
|
224 } |
|
225 z = x*x; |
|
226 r = pp0+z*(pp1+z*(pp2+z*(pp3+z*pp4))); |
|
227 s = one+z*(qq1+z*(qq2+z*(qq3+z*(qq4+z*qq5)))); |
|
228 y = r/s; |
|
229 return x + x*y; |
|
230 } |
|
231 if(ix < 0x3ff40000) { /* 0.84375 <= |x| < 1.25 */ |
|
232 s = fabs(x)-one; |
|
233 P = pa0+s*(pa1+s*(pa2+s*(pa3+s*(pa4+s*(pa5+s*pa6))))); |
|
234 Q = one+s*(qa1+s*(qa2+s*(qa3+s*(qa4+s*(qa5+s*qa6))))); |
|
235 if(hx>=0) return erx + P/Q; else return -erx - P/Q; |
|
236 } |
|
237 if (ix >= 0x40180000) { /* inf>|x|>=6 */ |
|
238 if(hx>=0) return one-tiny; else return tiny-one; |
|
239 } |
|
240 x = fabs(x); |
|
241 s = one/(x*x); |
|
242 if(ix< 0x4006DB6E) { /* |x| < 1/0.35 */ |
|
243 R=ra0+s*(ra1+s*(ra2+s*(ra3+s*(ra4+s*( |
|
244 ra5+s*(ra6+s*ra7)))))); |
|
245 S=one+s*(sa1+s*(sa2+s*(sa3+s*(sa4+s*( |
|
246 sa5+s*(sa6+s*(sa7+s*sa8))))))); |
|
247 } else { /* |x| >= 1/0.35 */ |
|
248 R=rb0+s*(rb1+s*(rb2+s*(rb3+s*(rb4+s*( |
|
249 rb5+s*rb6))))); |
|
250 S=one+s*(sb1+s*(sb2+s*(sb3+s*(sb4+s*( |
|
251 sb5+s*(sb6+s*sb7)))))); |
|
252 } |
|
253 z = x; |
|
254 __LO(z) = 0; |
|
255 r = __ieee754_exp(-z*z-0.5625)*__ieee754_exp((z-x)*(z+x)+R/S); |
|
256 if(hx>=0) return one-r/x; else return r/x-one; |
|
257 } |
|
258 |
|
259 #ifdef __STDC__ |
|
260 double erfc(double x) |
|
261 #else |
|
262 double erfc(x) |
|
263 double x; |
|
264 #endif |
|
265 { |
|
266 int hx,ix; |
|
267 double R,S,P,Q,s,y,z,r; |
|
268 hx = __HI(x); |
|
269 ix = hx&0x7fffffff; |
|
270 if(ix>=0x7ff00000) { /* erfc(nan)=nan */ |
|
271 /* erfc(+-inf)=0,2 */ |
|
272 return (double)(((unsigned)hx>>31)<<1)+one/x; |
|
273 } |
|
274 |
|
275 if(ix < 0x3feb0000) { /* |x|<0.84375 */ |
|
276 if(ix < 0x3c700000) /* |x|<2**-56 */ |
|
277 return one-x; |
|
278 z = x*x; |
|
279 r = pp0+z*(pp1+z*(pp2+z*(pp3+z*pp4))); |
|
280 s = one+z*(qq1+z*(qq2+z*(qq3+z*(qq4+z*qq5)))); |
|
281 y = r/s; |
|
282 if(hx < 0x3fd00000) { /* x<1/4 */ |
|
283 return one-(x+x*y); |
|
284 } else { |
|
285 r = x*y; |
|
286 r += (x-half); |
|
287 return half - r ; |
|
288 } |
|
289 } |
|
290 if(ix < 0x3ff40000) { /* 0.84375 <= |x| < 1.25 */ |
|
291 s = fabs(x)-one; |
|
292 P = pa0+s*(pa1+s*(pa2+s*(pa3+s*(pa4+s*(pa5+s*pa6))))); |
|
293 Q = one+s*(qa1+s*(qa2+s*(qa3+s*(qa4+s*(qa5+s*qa6))))); |
|
294 if(hx>=0) { |
|
295 z = one-erx; return z - P/Q; |
|
296 } else { |
|
297 z = erx+P/Q; return one+z; |
|
298 } |
|
299 } |
|
300 if (ix < 0x403c0000) { /* |x|<28 */ |
|
301 x = fabs(x); |
|
302 s = one/(x*x); |
|
303 if(ix< 0x4006DB6D) { /* |x| < 1/.35 ~ 2.857143*/ |
|
304 R=ra0+s*(ra1+s*(ra2+s*(ra3+s*(ra4+s*( |
|
305 ra5+s*(ra6+s*ra7)))))); |
|
306 S=one+s*(sa1+s*(sa2+s*(sa3+s*(sa4+s*( |
|
307 sa5+s*(sa6+s*(sa7+s*sa8))))))); |
|
308 } else { /* |x| >= 1/.35 ~ 2.857143 */ |
|
309 if(hx<0&&ix>=0x40180000) return two-tiny;/* x < -6 */ |
|
310 R=rb0+s*(rb1+s*(rb2+s*(rb3+s*(rb4+s*( |
|
311 rb5+s*rb6))))); |
|
312 S=one+s*(sb1+s*(sb2+s*(sb3+s*(sb4+s*( |
|
313 sb5+s*(sb6+s*sb7)))))); |
|
314 } |
|
315 z = x; |
|
316 __LO(z) = 0; |
|
317 r = __ieee754_exp(-z*z-0.5625)* |
|
318 __ieee754_exp((z-x)*(z+x)+R/S); |
|
319 if(hx>0) return r/x; else return two-r/x; |
|
320 } else { |
|
321 if(hx>0) return tiny*tiny; else return two-tiny; |
|
322 } |
|
323 } |
|