xref: /netbsd/lib/libm/noieee_src/n_erf.c (revision bf9ec67e)
1 /*	$NetBSD: n_erf.c,v 1.4 1999/07/02 15:37:36 simonb Exp $	*/
2 /*-
3  * Copyright (c) 1992, 1993
4  *	The Regents of the University of California.  All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  * 3. All advertising materials mentioning features or use of this software
15  *    must display the following acknowledgement:
16  *	This product includes software developed by the University of
17  *	California, Berkeley and its contributors.
18  * 4. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 
35 #ifndef lint
36 #if 0
37 static char sccsid[] = "@(#)erf.c	8.1 (Berkeley) 6/4/93";
38 #endif
39 #endif /* not lint */
40 
41 #include "mathimpl.h"
42 
43 /* Modified Nov 30, 1992 P. McILROY:
44  *	Replaced expansions for x >= 1.25 (error 1.7ulp vs ~6ulp)
45  * Replaced even+odd with direct calculation for x < .84375,
46  * to avoid destructive cancellation.
47  *
48  * Performance of erfc(x):
49  * In 300000 trials in the range [.83, .84375] the
50  * maximum observed error was 3.6ulp.
51  *
52  * In [.84735,1.25] the maximum observed error was <2.5ulp in
53  * 100000 runs in the range [1.2, 1.25].
54  *
55  * In [1.25,26] (Not including subnormal results)
56  * the error is < 1.7ulp.
57  */
58 
59 /* double erf(double x)
60  * double erfc(double x)
61  *			     x
62  *		      2      |\
63  *     erf(x)  =  ---------  | exp(-t*t)dt
64  *		   sqrt(pi) \|
65  *			     0
66  *
67  *     erfc(x) =  1-erf(x)
68  *
69  * Method:
70  *      1. Reduce x to |x| by erf(-x) = -erf(x)
71  *	2. For x in [0, 0.84375]
72  *	    erf(x)  = x + x*P(x^2)
73  *          erfc(x) = 1 - erf(x)           if x<=0.25
74  *                  = 0.5 + ((0.5-x)-x*P)  if x in [0.25,0.84375]
75  *	   where
76  *			2		 2	  4		  20
77  *              P =  P(x ) = (p0 + p1 * x + p2 * x + ... + p10 * x  )
78  * 	   is an approximation to (erf(x)-x)/x with precision
79  *
80  *						 -56.45
81  *			| P - (erf(x)-x)/x | <= 2
82  *
83  *
84  *	   Remark. The formula is derived by noting
85  *          erf(x) = (2/sqrt(pi))*(x - x^3/3 + x^5/10 - x^7/42 + ....)
86  *	   and that
87  *          2/sqrt(pi) = 1.128379167095512573896158903121545171688
88  *	   is close to one. The interval is chosen because the fixed
89  *	   point of erf(x) is near 0.6174 (i.e., erf(x)=x when x is
90  *	   near 0.6174), and by some experiment, 0.84375 is chosen to
91  * 	   guarantee the error is less than one ulp for erf.
92  *
93  *      3. For x in [0.84375,1.25], let s = x - 1, and
94  *         c = 0.84506291151 rounded to single (24 bits)
95  *         	erf(x)  = c  + P1(s)/Q1(s)
96  *         	erfc(x) = (1-c)  - P1(s)/Q1(s)
97  *         	|P1/Q1 - (erf(x)-c)| <= 2**-59.06
98  *	   Remark: here we use the taylor series expansion at x=1.
99  *		erf(1+s) = erf(1) + s*Poly(s)
100  *			 = 0.845.. + P1(s)/Q1(s)
101  *	   That is, we use rational approximation to approximate
102  *			erf(1+s) - (c = (single)0.84506291151)
103  *	   Note that |P1/Q1|< 0.078 for x in [0.84375,1.25]
104  *	   where
105  *		P1(s) = degree 6 poly in s
106  *		Q1(s) = degree 6 poly in s
107  *
108  *	4. For x in [1.25, 2]; [2, 4]
109  *         	erf(x)  = 1.0 - tiny
110  *		erfc(x)	= (1/x)exp(-x*x-(.5*log(pi) -.5z + R(z)/S(z))
111  *
112  *	Where z = 1/(x*x), R is degree 9, and S is degree 3;
113  *
114  *      5. For x in [4,28]
115  *         	erf(x)  = 1.0 - tiny
116  *		erfc(x)	= (1/x)exp(-x*x-(.5*log(pi)+eps + zP(z))
117  *
118  *	Where P is degree 14 polynomial in 1/(x*x).
119  *
120  *      Notes:
121  *	   Here 4 and 5 make use of the asymptotic series
122  *			  exp(-x*x)
123  *		erfc(x) ~ ---------- * ( 1 + Poly(1/x^2) );
124  *			  x*sqrt(pi)
125  *
126  *		where for z = 1/(x*x)
127  *		P(z) ~ z/2*(-1 + z*3/2*(1 + z*5/2*(-1 + z*7/2*(1 +...))))
128  *
129  *	   Thus we use rational approximation to approximate
130  *              erfc*x*exp(x*x) ~ 1/sqrt(pi);
131  *
132  *		The error bound for the target function, G(z) for
133  *		the interval
134  *		[4, 28]:
135  * 		|eps + 1/(z)P(z) - G(z)| < 2**(-56.61)
136  *		for [2, 4]:
137  *      	|R(z)/S(z) - G(z)|	 < 2**(-58.24)
138  *		for [1.25, 2]:
139  *		|R(z)/S(z) - G(z)|	 < 2**(-58.12)
140  *
141  *      6. For inf > x >= 28
142  *         	erf(x)  = 1 - tiny  (raise inexact)
143  *         	erfc(x) = tiny*tiny (raise underflow)
144  *
145  *      7. Special cases:
146  *         	erf(0)  = 0, erf(inf)  = 1, erf(-inf) = -1,
147  *         	erfc(0) = 1, erfc(inf) = 0, erfc(-inf) = 2,
148  *	   	erfc/erf(NaN) is NaN
149  */
150 
151 #if defined(__vax__) || defined(tahoe)
152 #define _IEEE	0
153 #define TRUNC(x) (double)(x) = (float)(x)
154 #else
155 #define _IEEE	1
156 #define TRUNC(x) *(((int *) &x) + 1) &= 0xf8000000
157 #define infnan(x) 0.0
158 #endif
159 
160 #ifdef _IEEE_LIBM
161 /*
162  * redefining "___function" to "function" in _IEEE_LIBM mode
163  */
164 #include "ieee_libm.h"
165 #endif
166 
167 static double
168 tiny	    = 1e-300,
169 half	    = 0.5,
170 one	    = 1.0,
171 two	    = 2.0,
172 c 	    = 8.45062911510467529297e-01, /* (float)0.84506291151 */
173 /*
174  * Coefficients for approximation to erf in [0,0.84375]
175  */
176 p0t8 = 1.02703333676410051049867154944018394163280,
177 p0 =   1.283791670955125638123339436800229927041e-0001,
178 p1 =  -3.761263890318340796574473028946097022260e-0001,
179 p2 =   1.128379167093567004871858633779992337238e-0001,
180 p3 =  -2.686617064084433642889526516177508374437e-0002,
181 p4 =   5.223977576966219409445780927846432273191e-0003,
182 p5 =  -8.548323822001639515038738961618255438422e-0004,
183 p6 =   1.205520092530505090384383082516403772317e-0004,
184 p7 =  -1.492214100762529635365672665955239554276e-0005,
185 p8 =   1.640186161764254363152286358441771740838e-0006,
186 p9 =  -1.571599331700515057841960987689515895479e-0007,
187 p10=   1.073087585213621540635426191486561494058e-0008;
188 /*
189  * Coefficients for approximation to erf in [0.84375,1.25]
190  */
191 static double
192 pa0 =  -2.362118560752659485957248365514511540287e-0003,
193 pa1 =   4.148561186837483359654781492060070469522e-0001,
194 pa2 =  -3.722078760357013107593507594535478633044e-0001,
195 pa3 =   3.183466199011617316853636418691420262160e-0001,
196 pa4 =  -1.108946942823966771253985510891237782544e-0001,
197 pa5 =   3.547830432561823343969797140537411825179e-0002,
198 pa6 =  -2.166375594868790886906539848893221184820e-0003,
199 qa1 =   1.064208804008442270765369280952419863524e-0001,
200 qa2 =   5.403979177021710663441167681878575087235e-0001,
201 qa3 =   7.182865441419627066207655332170665812023e-0002,
202 qa4 =   1.261712198087616469108438860983447773726e-0001,
203 qa5 =   1.363708391202905087876983523620537833157e-0002,
204 qa6 =   1.198449984679910764099772682882189711364e-0002;
205 /*
206  * log(sqrt(pi)) for large x expansions.
207  * The tail (lsqrtPI_lo) is included in the rational
208  * approximations.
209 */
210 static double
211    lsqrtPI_hi = .5723649429247000819387380943226;
212 /*
213  * lsqrtPI_lo = .000000000000000005132975581353913;
214  *
215  * Coefficients for approximation to erfc in [2, 4]
216 */
217 static double
218 rb0  =	-1.5306508387410807582e-010,	/* includes lsqrtPI_lo */
219 rb1  =	 2.15592846101742183841910806188e-008,
220 rb2  =	 6.24998557732436510470108714799e-001,
221 rb3  =	 8.24849222231141787631258921465e+000,
222 rb4  =	 2.63974967372233173534823436057e+001,
223 rb5  =	 9.86383092541570505318304640241e+000,
224 rb6  =	-7.28024154841991322228977878694e+000,
225 rb7  =	 5.96303287280680116566600190708e+000,
226 rb8  =	-4.40070358507372993983608466806e+000,
227 rb9  =	 2.39923700182518073731330332521e+000,
228 rb10 =	-6.89257464785841156285073338950e-001,
229 sb1  =	 1.56641558965626774835300238919e+001,
230 sb2  =	 7.20522741000949622502957936376e+001,
231 sb3  =	 9.60121069770492994166488642804e+001;
232 /*
233  * Coefficients for approximation to erfc in [1.25, 2]
234 */
235 static double
236 rc0  =	-2.47925334685189288817e-007,	/* includes lsqrtPI_lo */
237 rc1  =	 1.28735722546372485255126993930e-005,
238 rc2  =	 6.24664954087883916855616917019e-001,
239 rc3  =	 4.69798884785807402408863708843e+000,
240 rc4  =	 7.61618295853929705430118701770e+000,
241 rc5  =	 9.15640208659364240872946538730e-001,
242 rc6  =	-3.59753040425048631334448145935e-001,
243 rc7  =	 1.42862267989304403403849619281e-001,
244 rc8  =	-4.74392758811439801958087514322e-002,
245 rc9  =	 1.09964787987580810135757047874e-002,
246 rc10 =	-1.28856240494889325194638463046e-003,
247 sc1  =	 9.97395106984001955652274773456e+000,
248 sc2  =	 2.80952153365721279953959310660e+001,
249 sc3  =	 2.19826478142545234106819407316e+001;
250 /*
251  * Coefficients for approximation to  erfc in [4,28]
252  */
253 static double
254 rd0  =	-2.1491361969012978677e-016,	/* includes lsqrtPI_lo */
255 rd1  =	-4.99999999999640086151350330820e-001,
256 rd2  =	 6.24999999772906433825880867516e-001,
257 rd3  =	-1.54166659428052432723177389562e+000,
258 rd4  =	 5.51561147405411844601985649206e+000,
259 rd5  =	-2.55046307982949826964613748714e+001,
260 rd6  =	 1.43631424382843846387913799845e+002,
261 rd7  =	-9.45789244999420134263345971704e+002,
262 rd8  =	 6.94834146607051206956384703517e+003,
263 rd9  =	-5.27176414235983393155038356781e+004,
264 rd10 =	 3.68530281128672766499221324921e+005,
265 rd11 =	-2.06466642800404317677021026611e+006,
266 rd12 =	 7.78293889471135381609201431274e+006,
267 rd13 =	-1.42821001129434127360582351685e+007;
268 
269 double erf(x)
270 	double x;
271 {
272 	double R,S,P,Q,ax,s,y,z,r;
273 	if(!finite(x)) {		/* erf(nan)=nan */
274 	    if (isnan(x))
275 		return(x);
276 	    return (x > 0 ? one : -one); /* erf(+/-inf)= +/-1 */
277 	}
278 	if ((ax = x) < 0)
279 		ax = - ax;
280 	if (ax < .84375) {
281 	    if (ax < 3.7e-09) {
282 		if (ax < 1.0e-308)
283 		    return 0.125*(8.0*x+p0t8*x);  /*avoid underflow */
284 		return x + p0*x;
285 	    }
286 	    y = x*x;
287 	    r = y*(p1+y*(p2+y*(p3+y*(p4+y*(p5+
288 			y*(p6+y*(p7+y*(p8+y*(p9+y*p10)))))))));
289 	    return x + x*(p0+r);
290 	}
291 	if (ax < 1.25) {		/* 0.84375 <= |x| < 1.25 */
292 	    s = fabs(x)-one;
293 	    P = pa0+s*(pa1+s*(pa2+s*(pa3+s*(pa4+s*(pa5+s*pa6)))));
294 	    Q = one+s*(qa1+s*(qa2+s*(qa3+s*(qa4+s*(qa5+s*qa6)))));
295 	    if (x>=0)
296 		return (c + P/Q);
297 	    else
298 		return (-c - P/Q);
299 	}
300 	if (ax >= 6.0) {		/* inf>|x|>=6 */
301 	    if (x >= 0.0)
302 		return (one-tiny);
303 	    else
304 		return (tiny-one);
305 	}
306     /* 1.25 <= |x| < 6 */
307 	z = -ax*ax;
308 	s = -one/z;
309 	if (ax < 2.0) {
310 		R = rc0+s*(rc1+s*(rc2+s*(rc3+s*(rc4+s*(rc5+
311 			s*(rc6+s*(rc7+s*(rc8+s*(rc9+s*rc10)))))))));
312 		S = one+s*(sc1+s*(sc2+s*sc3));
313 	} else {
314 		R = rb0+s*(rb1+s*(rb2+s*(rb3+s*(rb4+s*(rb5+
315 			s*(rb6+s*(rb7+s*(rb8+s*(rb9+s*rb10)))))))));
316 		S = one+s*(sb1+s*(sb2+s*sb3));
317 	}
318 	y = (R/S -.5*s) - lsqrtPI_hi;
319 	z += y;
320 	z = exp(z)/ax;
321 	if (x >= 0)
322 		return (one-z);
323 	else
324 		return (z-one);
325 }
326 
327 double erfc(x)
328 	double x;
329 {
330 	double R,S,P,Q,s,ax,y,z,r;
331 	if (!finite(x)) {
332 		if (isnan(x))		/* erfc(NaN) = NaN */
333 			return(x);
334 		else if (x > 0)		/* erfc(+-inf)=0,2 */
335 			return 0.0;
336 		else
337 			return 2.0;
338 	}
339 	if ((ax = x) < 0)
340 		ax = -ax;
341 	if (ax < .84375) {			/* |x|<0.84375 */
342 	    if (ax < 1.38777878078144568e-17)  	/* |x|<2**-56 */
343 		return one-x;
344 	    y = x*x;
345 	    r = y*(p1+y*(p2+y*(p3+y*(p4+y*(p5+
346 			y*(p6+y*(p7+y*(p8+y*(p9+y*p10)))))))));
347 	    if (ax < .0625) {  	/* |x|<2**-4 */
348 		return (one-(x+x*(p0+r)));
349 	    } else {
350 		r = x*(p0+r);
351 		r += (x-half);
352 	        return (half - r);
353 	    }
354 	}
355 	if (ax < 1.25) {		/* 0.84375 <= |x| < 1.25 */
356 	    s = ax-one;
357 	    P = pa0+s*(pa1+s*(pa2+s*(pa3+s*(pa4+s*(pa5+s*pa6)))));
358 	    Q = one+s*(qa1+s*(qa2+s*(qa3+s*(qa4+s*(qa5+s*qa6)))));
359 	    if (x>=0) {
360 	        z  = one-c; return z - P/Q;
361 	    } else {
362 		z = c+P/Q; return one+z;
363 	    }
364 	}
365 	if (ax >= 28) {	/* Out of range */
366  		if (x>0)
367 			return (tiny*tiny);
368 		else
369 			return (two-tiny);
370 	}
371 	z = ax;
372 	TRUNC(z);
373 	y = z - ax; y *= (ax+z);
374 	z *= -z;			/* Here z + y = -x^2 */
375 		s = one/(-z-y);		/* 1/(x*x) */
376 	if (ax >= 4) {			/* 6 <= ax */
377 		R = s*(rd1+s*(rd2+s*(rd3+s*(rd4+s*(rd5+
378 			s*(rd6+s*(rd7+s*(rd8+s*(rd9+s*(rd10
379 			+s*(rd11+s*(rd12+s*rd13))))))))))));
380 		y += rd0;
381 	} else if (ax >= 2) {
382 		R = rb0+s*(rb1+s*(rb2+s*(rb3+s*(rb4+s*(rb5+
383 			s*(rb6+s*(rb7+s*(rb8+s*(rb9+s*rb10)))))))));
384 		S = one+s*(sb1+s*(sb2+s*sb3));
385 		y += R/S;
386 		R = -.5*s;
387 	} else {
388 		R = rc0+s*(rc1+s*(rc2+s*(rc3+s*(rc4+s*(rc5+
389 			s*(rc6+s*(rc7+s*(rc8+s*(rc9+s*rc10)))))))));
390 		S = one+s*(sc1+s*(sc2+s*sc3));
391 		y += R/S;
392 		R = -.5*s;
393 	}
394 	/* return exp(-x^2 - lsqrtPI_hi + R + y)/x;	*/
395 	s = ((R + y) - lsqrtPI_hi) + z;
396 	y = (((z-s) - lsqrtPI_hi) + R) + y;
397 	r = __exp__D(s, y)/x;
398 	if (x>0)
399 		return r;
400 	else
401 		return two-r;
402 }
403